Snapshot async request configuration for HTTP sessions

Replaced retired Anthropic model ids with the recommended replacements: the default model in src/anthropic.c and decai (claude-sonnet-5-5), plus docs, examples and model lists.

Notice anthropic-retired-models-2026-10 about api.anthropic.com's API describes the change. It was read from a file and is not signed by the provider.

.github/workflows/ci.yml

@@ -1,7 +1,7 @@
 name: CI and Release
 
 env:
-  R2V: 6.0.4
+  R2V: 6.2.6
 
 on:
   push:
@@ -26,7 +26,7 @@ jobs:
 
     steps:
       - name: Checkout
-        uses: actions/checkout@v6
+        uses: actions/checkout@v7
 
       - name: Setup Meson/Ninja
         if: matrix.build_system == 'meson'
@@ -34,38 +34,11 @@ jobs:
           python -m pip install --upgrade pip
           pip install meson ninja
 
-      - name: Install radare2 on Linux
-        if: matrix.os == 'ubuntu-24.04'
-        run: |
-          wget "https://github.com/radareorg/radare2/releases/download/${R2V}/radare2_${R2V}_amd64.deb"
-          wget "https://github.com/radareorg/radare2/releases/download/${R2V}/radare2-dev_${R2V}_amd64.deb"
-          sudo dpkg -i "radare2_${R2V}_amd64.deb"
-          sudo dpkg -i "radare2-dev_${R2V}_amd64.deb"
-
-      - name: Install radare2 on macOS
-        if: matrix.os == 'macos-latest'
-        run: |
-          git clone --depth=1 --branch "${{env.R2V}}" https://github.com/radareorg/radare2
-          cd radare2
-          sys/install.sh
-
-      - name: Install radare2 on Windows
-        if: matrix.os == 'windows-latest'
-        shell: powershell
-        run: |
-          $url = "https://github.com/radareorg/radare2/releases/download/${{ env.R2V }}/radare2-${{ env.R2V }}-w64.zip"
-          $dest = "$env:TEMP\radare2.zip"
-          Invoke-WebRequest -Uri $url -OutFile $dest -UseBasicParsing
-          $prefix = 'C:\radare2'
-          if (Test-Path $prefix) { Remove-Item -Path $prefix -Recurse -Force }
-          Expand-Archive -Path $dest -DestinationPath $prefix -Force
-          # Move contents from subdirectory if present
-          $subdirs = Get-ChildItem -Path $prefix -Directory
-          if ($subdirs.Count -eq 1) {
-            $subdir = $subdirs[0]
-            Get-ChildItem -Path $subdir.FullName | Move-Item -Destination $prefix -Force
-            Remove-Item -Path $subdir.FullName -Recurse -Force
-          }
+      - name: Install radare2
+        id: r2
+        uses: radareorg/github-actions@v3
+        with:
+          version: ${{ env.R2V }}
 
       - name: Build with Make
         if: matrix.build_system == 'make'
@@ -75,7 +48,7 @@ jobs:
         if: matrix.build_system == 'meson' && matrix.os != 'windows-latest'
         run: |
           cd src
-          meson setup build -Dr2_prefix=${{ secrets.R2_PREFIX }} || meson configure build
+          meson setup build || meson configure build
           meson compile -C build
 
       - name: Build with Meson (Windows)
@@ -83,51 +56,92 @@ jobs:
         shell: powershell
         run: |
           cd src
-          $prefix = 'C:\radare2'
+          $prefix = "${{ steps.r2.outputs.prefix }}"
           meson setup build --backend ninja --prefix="$prefix" -Dr2_prefix="$prefix"
           if ($LASTEXITCODE -ne 0) { meson configure build }
           meson compile -C build
 
-      - name: Create release artifacts (tags only)
-        if: startsWith(github.ref, 'refs/tags/')
-        shell: bash
-        run: |
-          mkdir -p release
-          # Collect common binary artefact names produced by Make/Meson
-          # Include both "r2ai.*" and "libr2ai.*" variants for DLL/dylib/so on all platforms.
-          find src -type f \( -iname "r2ai.dll" -o -iname "libr2ai.dll" -o -iname "r2ai.dylib" -o -iname "libr2ai.dylib" -o -iname "r2ai.so" -o -iname "libr2ai.so" -o -iname "r2ai.exe" \) -exec cp -f {} release/ \; || true
-          # Also check top-level build output directories
-          find . -maxdepth 3 -type f \( -iname "r2ai.dll" -o -iname "libr2ai.dll" -o -iname "r2ai.dylib" -o -iname "libr2ai.dylib" -o -iname "r2ai.so" -o -iname "libr2ai.so" -o -iname "r2ai.exe" \) -exec cp -f {} release/ \; || true
-          cp -f decai/decai.r2.js release/ || true
-          if command -v zip >/dev/null 2>&1; then
-            zip -j "release/r2ai-${{ github.ref_name }}-${{ matrix.os }}-${{ matrix.build_system }}.zip" release/*
-          else
-            powershell.exe -Command "Compress-Archive -Path release/* -DestinationPath 'release/r2ai-${{ github.ref_name }}-${{ matrix.os }}-${{ matrix.build_system }}.zip'"
-          fi
+  bindist:
+    name: Build r2pm binary (${{ matrix.os }})
+    strategy:
+      fail-fast: false
+      matrix:
+        os: [ubuntu-latest, ubuntu-24.04-arm, macos-latest]
+    runs-on: ${{ matrix.os }}
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v7
 
-      - name: Upload artifacts (tags only)
-        if: startsWith(github.ref, 'refs/tags/')
+      - name: Install radare2
+        uses: radareorg/github-actions@v3
+        with:
+          version: ${{ env.R2V }}
+
+      - name: Package r2pm binary
+        run: make -C src bindist
+
+      - name: Upload r2pm binary artifact
         uses: actions/upload-artifact@v7
         with:
-          name: ${{ matrix.os }}-${{ matrix.build_system }}-binary
-          path: release/r2ai-${{ github.ref_name }}-${{ matrix.os }}-${{ matrix.build_system }}.zip
+          name: bindist-${{ matrix.os }}
+          path: r2ai-*.zip
+          if-no-files-found: error
 
-#      - name: Smoke test
-#        run: |
-#          if command -v r2 >/dev/null 2>&1; then
-#            r2 -q -c 'r2ai -h' -- || true
-#          fi
+  windows-bindist:
+    name: Build r2pm binary (Windows)
+    runs-on: windows-latest
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v7
+
+      - name: Setup Meson/Ninja
+        run: |
+          python -m pip install --upgrade pip
+          pip install meson ninja
+
+      - name: Install radare2
+        id: r2
+        uses: radareorg/github-actions@v3
+        with:
+          version: ${{ env.R2V }}
+
+      - name: Build with Meson
+        shell: powershell
+        run: |
+          cd src
+          $prefix = "${{ steps.r2.outputs.prefix }}"
+          meson setup build --backend ninja --prefix="$prefix" -Dr2_prefix="$prefix"
+          if ($LASTEXITCODE -ne 0) { meson configure build }
+          meson compile -C build
+
+      - name: Package r2pm binary
+        shell: pwsh
+        run: |
+          New-Item -ItemType Directory -Force -Path bindist\plugins,bindist\bin,bindist\prompts | Out-Null
+          Get-ChildItem src\build\ -Filter "*r2ai*.dll" -ErrorAction SilentlyContinue | Copy-Item -Destination bindist\plugins
+          Get-ChildItem src\..\build\ -Filter "*r2ai*.dll" -ErrorAction SilentlyContinue | Copy-Item -Destination bindist\plugins
+          Copy-Item src\build\r2ai.exe bindist\bin
+          Copy-Item prompts\*.r2ai.md bindist\prompts
+          $rv = (& radare2 -qv).Trim()
+          Compress-Archive -Path bindist\* -DestinationPath "r2ai-$rv-windows-x86-64.zip" -Force
+
+      - name: Upload r2pm binary artifact
+        uses: actions/upload-artifact@v7
+        with:
+          name: windows-bindist
+          path: r2ai-*.zip
+          if-no-files-found: error
 
   release:
     name: Create Release
-    needs: build
+    needs: [bindist, windows-bindist]
     if: startsWith(github.ref, 'refs/tags/')
     runs-on: ubuntu-latest
     permissions:
       contents: write
     steps:
       - name: Checkout
-        uses: actions/checkout@v6
+        uses: actions/checkout@v7
 
       - name: Verify tag is on master
         run: |
@@ -137,18 +151,20 @@ jobs:
             exit 0
           fi
 
-      - name: Download artifacts
+      - name: Download r2pm binary artifacts
         uses: actions/download-artifact@v8
         with:
+          pattern: '*bindist*'
           path: release-artifacts
+          merge-multiple: true
 
       - name: Add decai.r2.js to release artifacts
         run: cp decai/decai.r2.js release-artifacts/
 
       - name: Create GitHub Release
-        uses: softprops/action-gh-release@v2
+        uses: softprops/action-gh-release@v3
         with:
           files: |
-            release-artifacts/**/*.zip
+            release-artifacts/*.zip
             release-artifacts/decai.r2.js
-          generate_release_notes: true
+          generate_release_notes: true

.gitignore

@@ -12,3 +12,6 @@ venv
 *.dylib
 *.a
 third_party
+
+# r2pm binary packages
+r2ai-*.zip

Attic/examples/ai.sh

@@ -23,12 +23,12 @@ if [ -f ~/.r2ai.openai-key ]; then
 	OPENAI_KEY=$(cat ~/.r2ai.openai-key)
 fi
 CLAUDE_KEY=""
-CLAUDE_MODEL="claude-3-5-sonnet-20241022"
+CLAUDE_MODEL="claude-sonnet-5-5"
 if [ -f ~/.r2ai.anthropic-key ]; then
 	CLAUDE_KEY=$(cat ~/.r2ai.anthropic-key)
 fi
 DEEPSEEK_KEY=""
-DEEPSEEK_MODEL="claude-3-5-sonnet-20241022"
+DEEPSEEK_MODEL="claude-sonnet-5-5"
 if [ -f ~/.r2ai.deepseek-key ]; then
 	DEEPSEEK_KEY=$(cat ~/.r2ai.anthropic-key)
 fi
@@ -138,7 +138,7 @@ GEMINI_KEY=~/.r2ai-gemini.key
 OPENAI_KEY=~/.r2ai-openai.key
 CLAUDE_KEY=~/.r2ai-anthropic.key
 DEEPSEEK_KEY=~/.r2ai-deepseek.key
-CLAUDE_MODEL=claude-3-5-sonnet-20241022
+CLAUDE_MODEL=claude-sonnet-5-5
 EOF
 	exit 0
 }

Attic/py/README.md

@@ -49,7 +49,7 @@ python3 main.py
 **Example selecting a remote models:**
 
 ```
-[r2ai:0x00006aa0]> -m anthropic:claude-3-7-sonnet-20250219
+[r2ai:0x00006aa0]> -m anthropic:claude-sonnet-5-5
 [r2ai:0x00006aa0]> -m openai:gpt-4
 ```
 

Attic/py/r2ai/models.py

@@ -62,7 +62,7 @@ Uncensored:
 Remote:
 -m openapi:http://localhost:5001
 -m openai:gpt-4
--m anthropic:claude-3-7-sonnet-20250219
+-m anthropic:claude-sonnet-5-5
 -m bedrock:anthropic.claude-3-5-sonnet-20241022-v1
 -m kobaldcpp:http://localhost:5001
 """
@@ -80,12 +80,9 @@ OpenAI:
 -m openai:gpt-4
 -m openai:gpt-4-1106-preview
 Anthropic:
--m anthropic:claude-2.1
--m anthropic:claude-3-haiku-20240307
--m anthropic:claude-3-opus-20240229
--m anthropic:claude-3-sonnet-20240229
--m anthropic:claude-3-5-sonnet-20241022
--m anthropic:claude-3-7-sonnet-20250219
+-m anthropic:claude-opus-4-8
+-m anthropic:claude-sonnet-5-5
+-m anthropic:claude-haiku-4-5-20251001
 groq:
 -m groq:gemma-7b-it
 -m groq:llama2-70b-4096

Attic/py/r2ai/repl.py

@@ -170,8 +170,7 @@ def slurp_until(endword):
 def list_models():
     print("ollama/llama3:latest")
     print("ollama/granite-code:8b")
-    print("anthropic/claude-3-5-sonnet-20240620")
-    print("anthropic/claude-3-7-sonnet-20250219")
+    print("anthropic/claude-sonnet-5-5")
     print("openai/gpt-4o")
     print("mistral/mistral-tiny")
 

Attic/py/r2ai/voice.py

@@ -50,8 +50,16 @@ def stt(seconds, lang):
         return
     tts("(r2ai) listening for 5s... ", "digues?", lang)
     print(f"DEVICE IS {device}")
-    os.system("rm -f .audiomsg.wav")
-    rc = os.system(f"ffmpeg -f avfoundation -t 5 -i '{device}' .audiomsg.wav > /dev/null 2>&1")
+    try:
+        os.remove(".audiomsg.wav")
+    except OSError:
+        pass
+    result = subprocess.run(
+        ["ffmpeg", "-f", "avfoundation", "-t", "5", "-i", device, ".audiomsg.wav"],
+        stdout=subprocess.DEVNULL,
+        stderr=subprocess.DEVNULL
+    )
+    rc = result.returncode
     if rc != 0:
         tts("(r2ai)", "cannot record from microphone. missing permissions in terminal?", lang)
         return
@@ -60,7 +68,10 @@ def stt(seconds, lang):
         result = model.transcribe(".audiomsg.wav")
     else:
         result = model.transcribe(".audiomsg.wav", language=lang)
-    os.system("rm -f .audiomsg.wav")
+    try:
+        os.remove(".audiomsg.wav")
+    except OSError:
+        pass
     tts("(r2ai)", "ok", lang)
     text = result["text"].strip()
     if text == "you":

Makefile

@@ -3,4 +3,7 @@ TARGETS=all clean mrproper user-install install user-uninstall uninstall fmt ind
 $(TARGETS):
 	$(MAKE) -C src $@
 
-.PHONY: $(TARGETS)
+test: all
+	$(MAKE) -C test
+
+.PHONY: $(TARGETS) test

README.md

@@ -97,7 +97,7 @@ then you can type the commands you want to run when the r2ai plugin is loaded:
 
 ```console
 r2ai -e api=anthropic
-r2ai -e model=claude-3-7-sonnet-20250219
+r2ai -e model=claude-sonnet-5-5
 r2ai -e max_tokens=64000
 ```
 

VERSION

@@ -1 +1 @@
-1.3.4
+1.5.4

decai/AGENTS.md

@@ -5,3 +5,9 @@
 * Build: `make -C src`
 * Install: `make user-install`
 * Test: `r2 -A -qc "decai -e api=ollamacloud;decai -e model=glm-5;decai -d @ 0x100000924" /bin/ls`
+
+## Hints
+
+* Never read the minified, autogenerated `decai.r2.js`
+* Focus and work only on the TypeScript source
+* Never run `tsc` or `npx`. Always use `make` targets to compile

decai/README.md

@@ -56,6 +56,7 @@ Configuration options:
 ```console
 [0x00000000]> decai -e
 decai -e api=ollama
+decai -e apitype=chat
 decai -e host=http://localhost
 decai -e port=11434
 decai -e prompt=Rewrite this function and respond ONLY with code, NO explanations, NO markdown, Change 'goto' into if/else/for/while, Simplify as much as possible, use better variable names, take function arguments and strings from comments like 'string:'
@@ -102,6 +103,12 @@ Persist user defaults in `~/.config/r2ai/decai.txt` with one `key=value` entry
 per line, then edit it with `decai -E`. That file is loaded on the first
 `decai` command in a session and reloaded after `decai -E` exits.
 
+Use `decai -e http=curl` to force system curl, `http=r2` to require `r2.http()`,
+or `http=auto` (the default) to use `r2.http()` when available and curl otherwise.
+Failed `r2.http()` calls return an error without falling back to curl. The
+`timeout` setting applies only to system curl. Save `http=curl` in
+`~/.config/r2ai/decai.txt` to persist the override.
+
 List possible APIs to discuss with AI: `decai -e api=?`:
 
 ```
@@ -211,3 +218,21 @@ void daemonize() {
 }
 ...
 ```
+
+## WASM HTTP
+
+With radare2's WASI HTTP build, decai defaults to `r2.http()` to reach the existing
+`r_socket_http_get/post` C APIs. Headers and JSON payloads go directly to the
+host HTTP adapter; no shell curl, temporary payload file or WebSocket proxy is
+needed. A failed/denied HTTP call is returned as an error, without a shell
+fallback. In `http=auto` mode, runtimes without `r2.http()` use system curl.
+
+Enable HTTP in the embedding application (`enabled: true` in JavaScript or
+`network=True` in Python). The radare2 sandbox must permit both `socket` and
+`network`; process execution is not needed. Browser requests remain subject
+to CORS. The host adapter controls the HTTP deadline in WASM; decai's `timeout`
+setting still controls native curl only. Responses are buffered, not streamed.
+
+This requires radareorg/radare2#27037 and the rebuilt decai script. For the
+C r2ai plugin compiled into WASM, the `auto` HTTP backend likewise selects
+radare2's C HTTP implementation; native backend selection is unchanged.

decai/decai.r2.js

@@ -1,14 +1,14 @@
 📦
-26532 /main.js
+27680 /main.js
 ✄
-var Q="1.3.4",y="decai",A="~/.config/r2ai",M=A+"/apikeys.txt",I=A+"/decai.txt",N="Rewrite this pseudocode into concise and clean code. Output only the provided function. Do not add wrappers, helper examples, test code, or main-like functions. Replace goto with structured control flow, simplify as much as possible, infer types and use better names for variables and parameters, some strings may be appearing as comments, preserve only what is implied by the input, and remove dead code.",F={decai:`# Using Decai
+var te="1.5.4",b="decai",w="~/.config/r2ai",N=w+"/apikeys.txt",U=w+"/decai.txt",T="Rewrite this pseudocode into concise and clean code. Output only the provided function. Do not add wrappers, helper examples, test code, or main-like functions. Replace goto with structured control flow, simplify as much as possible, infer types and use better names for variables and parameters, some strings may be appearing as comments, preserve only what is implied by the input, and remove dead code.",L={decai:`# Using Decai
 
 Decai is the radare2 plugin for decompiling functions with the help of language models.
 
 By default uses a local ollama server, but can you can pick any other service by using 'decai -e api=?'.
 
 [0x00000000]> decai -e api=?
-r2ai deepseek anthropic claude gemini groq hf lmstudio mistral ollama ollamacloud openapi opencode openai openrouter vllm xai zen
+r2ai deepseek anthropic claude gemini groq hf lmstudio mistral ollama ollamacloud openapi openai openrouter vllm xai zen
 
 ## Using Ollama
 
@@ -23,8 +23,10 @@ These are the most recommended models for decompiling in local:
 
 ## Common Options
 * 'decai -e baseurl=<url>' override default host and port for API endpoint (e.g., 'http://localhost:11434')
+* 'decai -e apitype=chat|generate' choose the Ollama API endpoint (default: chat)
 * 'decai -e headers=Authorization: Bearer ...\\nUser-Agent: curl/8.7.1' add or override HTTP headers
 * "export DECAI_HEADERS='Authorization: Bearer ...\\nUser-Agent: curl/8.7.1'" set extra headers from the environment
+* 'decai -e http=auto|r2|curl' select the HTTP transport (default: auto, prefers r2.http)
 * 'decai -e timeout=0' disable curl timeouts, or set a positive number of seconds
 
 * 'decai -e deterministic=true' to remove randomness from decompilation responses
@@ -116,50 +118,54 @@ Use radare2 to resolve user requests.
 ### Special cases
 
 * On Swift binaries run...
-`};var k={decopipe:{use:!1},host:"http://localhost",port:"11434",baseurl:"",extraHeaders:[],api:"ollama",pipeline:"",commands:"pdc",yolo:!1,tts:!1,language:"C",humanLanguage:"English",deterministic:!0,debug:!1,timeout:180,think:-1,useFiles:!1,contextFile:"",model:"",cache:!1,maxInputTokens:-1,prompt:N,lastOutput:""},n={...k,decopipe:{...k.decopipe},extraHeaders:[...k.extraHeaders]};function X(e){return(r2.cmd("-e dir.tmp").trim()??".")+"/"+e}function U(e,t){r2.cmd("'mkdir -p "+e),r2.cmd("'touch "+t)}function O(e){return(r2.cmd2("test -h").logs?.[0]?.message??"").includes("-fdx")?!0:r2.cmd("'test -vf "+e).startsWith("found")}function B(e,t){return e+" ".repeat(Math.max(0,t-e.length))}function Z(e){return e.replace(/\x1b\[[0-9;]*m/g,"")}function ee(e){let t=e.match(/```json\s*([\s\S]*?)```/);return t&&t[1]?t[1].trim():e}function te(e){let t=e,o=t.indexOf("{");o!==-1&&(t=t.slice(o));let i=t.lastIndexOf("}");return i!==-1&&(t=t.slice(0,i+1)),t}function Y(e){return btoa(e)}function oe(e,t){let o=Y(t);r2.cmd("p6ds "+o+" > "+e)}function E(e){let t=e;return n.think!==2&&(t=t.replace(/<think>[\s\S]*?<\/think>/gi,"")),t.split(`
-`).filter(o=>!o.trim().startsWith("```")).join(`
-`)}function j(e){n.debug&&console.log(e)}function ne(e){let t={};for(let o of e.split(/\r?\n/)){let i=o.trim();if(!i||i.startsWith("#"))continue;let[r,...a]=i.split("=");if(!r||a.length===0)continue;let l=a.join("=").trim(),s=r.toLowerCase().replace(/_api_key$/i,"");t[s]=l}return t}var Ie={mistral:"MISTRAL_API_KEY",anthropic:"ANTHROPIC_API_KEY",huggingface:"HUGGINGFACE_API_KEY",openai:"OPENAI_API_KEY",gemini:"GEMINI_API_KEY",deepseek:"DEEPSEEK_API_KEY",xai:"XAI_API_KEY",ollama:"OLLAMA_API_KEY",ollamacloud:"OLLAMA_API_KEY",opencode:"OPENCODE_API_KEY",zen:"OPENCODE_API_KEY",openrouter:"OPENROUTER_API_KEY",groq:"GROQ_API_KEY"};function T(e,t){let o=r2.cmd("'%"+t).trim();if(!o.includes("=")&&o!=="")return[o.trim(),null,"env"];let i=e.toLowerCase(),r=M;if(O(r)){let l=r2.cmd("'cat "+r),s=ne(l);if(i in s)return[s[i],null,"txt"]}let a="~/.r2ai."+i+"-key";if(O(a)){let l=r2.cmd("'cat "+a);return l===""?[null,"Cannot read "+a,"no"]:[l.trim(),null,"file"]}return[null,"Not available","nope"]}function re(){U(A,M),r2.cmd("'ed "+M)}function G(){Object.entries(Ie).forEach(([e,t])=>{let o=T(e,t)[2];console.log(o,"	",e)})}var Oe=["DECAI_HEADERS","R2AI_HEADERS"];function _e(e){let t=r2.cmd("'%"+e).trim();return t!==""&&!t.includes("=")?t:""}function ie(e){return e.trim().toLowerCase()}function ae(e){let t=e.indexOf(":"),o=e.indexOf("="),i=t!==-1&&(o===-1||t<o)?t:o;if(i===-1)return null;let r=e.slice(0,i).trim();return r===""?null:{name:r,value:e.slice(i+1).trim()}}function se(e){return e.value===""?`${e.name}:`:`${e.name}: ${e.value}`}function $(e){let t=new Map,o=e.replace(/\\n/g,`
-`);for(let i of o.split(/\r?\n/g)){let r=i.trim();if(!r||r.startsWith("#"))continue;let a=ae(r);a&&t.set(ie(a.name),se(a))}return Array.from(t.values())}function le(e){return e.join("\\n")}function P(...e){let t=new Map;for(let o of e)for(let i of o){let r=ae(i);r&&t.set(ie(r.name),se(r))}return Array.from(t.values())}function Se(){for(let e of Oe){let t=_e(e);if(t)return $(t)}return[]}function ce(){return P(Se(),n.extraHeaders)}function _(e){return`'${e.replace(/'/g,`'"'"'`)}'`}function Ce(e){let{method:t,url:o,headers:i,payload:r}=e,a=["curl","-s"];if(n.timeout>0&&a.push("--max-time",String(n.timeout)),P(["Content-Type: application/json"],i).forEach(s=>a.push("-H",_(s))),t==="POST"){if(!r)throw new Error("Payload required for POST requests");let s=r2.fdump(r);a.push("--data-binary","@-",_(o));let c=a.join(" ")+" < "+_(s)+" && rm "+_(s);return j(c),r2.syscmds(c)}else{a.push(_(o));let s=a.join(" ");return j(s),r2.syscmds(s)}}function ue(e){try{let t=Ce(e).trim();if(t==="")return{error:"empty response"};try{return JSON.parse(t)}catch{return{error:t,rawOutput:t}}}catch(t){return{error:t.message}}}function v(e,t){return ue({method:"GET",url:e,headers:t})}function de(e,t,o){return ue({method:"POST",url:e,headers:t,payload:o})}function J(e){return e.split("_")[0].toLowerCase()}function S(e){if(!e.authKey)return;let t=Array.from(new Set([e.keyName,J(e.authKey)].filter(Boolean))),o;for(let i of t){let r=T(i,e.authKey);if(r[0])return r;!o&&r[2]!=="nope"&&(o=r)}return o||T(J(e.authKey),e.authKey)}function pe(e){if(e)return typeof e=="string"?e:e.message}function C(e){return n.baseurl||e.defaultBaseUrl}function R(e,t=ce()){return P(e,t)}function Re(e){return e?["Authorization: Bearer "+e]:[]}function Ke(e){let t=["anthropic-version: 2023-06-01"];return e?P(t,["x-api-key: "+e]):t}function K(e,t){switch(e.authStyle||"none"){case"bearer":return Re(t);case"anthropic":return Ke(t);case"none":default:return[]}}function me(e,t,o){let i=v(e,t),r=pe(i.error);return r?(console.error(r),"error invalid response"):i.data?.map(o).join(`
-`)||""}function Me(e,t){let o=new Set;return e.filter(i=>{let r=t(i);return o.has(r)?!1:(o.add(r),!0)})}function Ne(e){let t=S(e),o=C(e),i=R(K(e,t?.[0]||null));return me(o+"/v1/models",i,r=>r.id)}function Ue(e){let t=S(e),o=C(e),i=R(K(e,t?.[0]||null));return me(o+"/v1/models",i,r=>r.id)}function Te(e){let t=v(C(e)+"/api/tags",R(K(e,S(e)?.[0]||null))),o=pe(t.error);return o?(console.error(o),"error invalid response"):t.models?.map(i=>i.name).join(`
-`)||""}function qe(e){let t=S(e),o=v(C(e)+"/v1/models",R(K(e,t?.[0]||null)));return o.data?Me(o.data,i=>i.name||i.id).map(i=>[B(i.name||i.id,30),B(""+(i.max_context_length||""),10),i.description||""].join(" ")).join(`
-`):""}var De={openai:{buildPayload:(e,t)=>({stream:!1,model:e,messages:[{role:"user",content:t}]}),parseResponse:e=>{if(e.error){let t=typeof e.error=="object"?e.error.message:e.error;throw new Error(t||"Unknown error")}if(e.choices&&e.choices[0]?.message?.content)return E(e.choices[0].message.content);throw new Error("Invalid response format")},buildUrl:e=>e+"/v1/chat/completions"},anthropic:{buildPayload:(e,t)=>{let o={model:e,max_tokens:5128,messages:[{role:"user",content:t}]};return n.deterministic&&Object.assign(o,{temperature:0,top_p:0,top_k:1}),o},parseResponse:e=>{if(e.content&&e.content[0]?.text)return E(e.content[0].text);if(e.error){let t=typeof e.error=="object"?e.error.message:e.error;throw new Error(t||"Unknown error")}throw new Error("Invalid response format")},buildUrl:e=>e+"/v1/messages"},ollama:{buildPayload:(e,t)=>{let o={stream:!1,model:e,messages:[{role:"user",content:t}]};return n.deterministic&&(o.options={repeat_last_n:0,top_p:1,top_k:1,temperature:0,repeat_penalty:1,seed:123}),o},parseResponse:e=>{if(e.error){let t=typeof e.error=="string"?e.error:JSON.stringify(e.error);throw new Error(t)}if(e.message?.content)return E(e.message.content);throw new Error(JSON.stringify(e))},buildUrl:e=>e+"/api/chat"},gemini:{buildPayload:(e,t)=>{let o={contents:[{parts:[{text:t}]}]};return n.deterministic&&(o.generationConfig={temperature:0,topP:1,topK:1}),o},parseResponse:e=>{if(e.candidates&&e.candidates[0]?.content?.parts?.[0]?.text)return E(e.candidates[0].content.parts[0].text);throw e.error?new Error(typeof e.error=="string"?e.error:JSON.stringify(e.error)):(console.log(JSON.stringify(e)),new Error("Invalid response format"))},buildUrl:(e,t,o)=>`${e}/v1beta/models/${t}:generateContent?key=${o}`,requiresUrlApiKey:!0}},fe={anthropic:{defaultModel:"claude-3-7-sonnet-20250219",defaultBaseUrl:"https://api.anthropic.com",authKey:"ANTHROPIC_API_KEY",keyName:"anthropic",authStyle:"anthropic",apiStyle:"anthropic"},claude:{defaultModel:"claude-3-7-sonnet-20250219",defaultBaseUrl:"https://api.anthropic.com",authKey:"ANTHROPIC_API_KEY",keyName:"claude",authStyle:"anthropic",apiStyle:"anthropic"},openai:{defaultModel:"gpt-4o-mini",defaultBaseUrl:"https://api.openai.com",authKey:"OPENAI_API_KEY",authStyle:"bearer",apiStyle:"openai"},ollama:{defaultModel:"qwen2.5-coder:latest",defaultBaseUrl:"http://localhost:11434",authStyle:"none",apiStyle:"ollama"},ollamacloud:{defaultModel:"gpt-oss:120b",defaultBaseUrl:"https://ollama.com",authKey:"OLLAMA_API_KEY",keyName:"ollamacloud",authStyle:"bearer",apiStyle:"ollama"},opencode:{defaultModel:"big-pickle",defaultBaseUrl:"https://opencode.ai/zen",authKey:"OPENCODE_API_KEY",keyName:"opencode",authStyle:"bearer",apiStyle:"openai",hardcodedModels:["big-pickle","glm-5","kimi-k2.5"]},zen:{defaultModel:"big-pickle",defaultBaseUrl:"https://opencode.ai/zen",authKey:"OPENCODE_API_KEY",keyName:"zen",authStyle:"bearer",apiStyle:"openai",hardcodedModels:["big-pickle","glm-5","kimi-k2.5"]},gemini:{defaultModel:"gemini-2.5-flash",defaultBaseUrl:"https://generativelanguage.googleapis.com",authKey:"GEMINI_API_KEY",authStyle:"none",apiStyle:"gemini",hardcodedModels:["gemini-2.0-flash","gemini-2.0-flash-lite","gemini-2.5-pro","gemini-2.5-flash","gemini-2.5-flash-lite"]},mistral:{defaultModel:"codestral-latest",defaultBaseUrl:"https://api.mistral.ai",authKey:"MISTRAL_API_KEY",authStyle:"bearer",apiStyle:"openai"},xai:{defaultModel:"grok-beta",defaultBaseUrl:"https://api.x.ai",authKey:"XAI_API_KEY",authStyle:"bearer",apiStyle:"openai",hardcodedModels:["grok-2","grok-beta"]},lmstudio:{defaultModel:"local-model",defaultBaseUrl:"http://127.0.0.1:1234",authStyle:"none",apiStyle:"openai"},deepseek:{defaultModel:"deepseek-coder",defaultBaseUrl:"https://api.deepseek.com",authKey:"DEEPSEEK_API_KEY",authStyle:"bearer",apiStyle:"openai",hardcodedModels:["deepseek-coder","deepseek-chat"]},openrouter:{defaultModel:"openai/gpt-4o-mini",defaultBaseUrl:"https://openrouter.ai/api",authKey:"OPENROUTER_API_KEY",authStyle:"bearer",apiStyle:"openai"},groq:{defaultModel:"llama-3.3-70b-versatile",defaultBaseUrl:"https://api.groq.com/openai",authKey:"GROQ_API_KEY",authStyle:"bearer",apiStyle:"openai"}};function q(e){return fe[e]}function z(){return Object.keys(fe)}function he(e){let t=q(e);if(!t)throw new Error(`Unknown provider: ${e}`);if(e==="mistral")return qe(t);switch(t.apiStyle){case"openai":return Ne(t);case"anthropic":return Ue(t);case"ollama":return Te(t);case"gemini":return"";default:return""}}function He(e,t){let o=e;return n.think>=0&&(n.think===0?(o+=' Answers directly and concisely, without showing any thinking steps or internal reasoning. Never include phrases like "Let me think".',o+=" /no_think"):n.think>0&&(o="Think step by step and explain the reasoning process, When answering, first output your reasoning inside <think> and </think> tags, then give the final answer."+o)),t?o:n.prompt+Le()+o}function Le(){return`
-.Translate the code into `+n.language+` programming language
-`}function Fe(e,t,o,i){let r=n.model||e.defaultModel,a=He(o,i),l=S(e);if(t.requiresUrlApiKey&&l?.[1]&&e.authKey)return`Cannot read ~/.r2ai.${J(e.authKey)}-key`;let s=t.buildPayload(r,a),c=R(K(e,l?.[0]||null)),p=t.buildUrl(C(e),r,l?.[0]||void 0);try{return t.parseResponse(de(p,c,JSON.stringify(s)))}catch(d){return"ERROR: "+d.message}}function ge(e,t){let o=q(n.api);return o?Fe(o,De[o.apiStyle],e,t):`Unknown value for 'decai -e api'. Available: ${z().join(", ")}`}function ye(e){let t=q(e);if(!t){console.error(`Unknown provider: ${e}`);return}try{let o=he(e);o&&console.log(o),t.hardcodedModels&&t.hardcodedModels.forEach(i=>console.log(i)),e==="mistral"&&console.log("ministral-8b-latest"),!o&&!t.hardcodedModels&&console.log(t.defaultModel)}catch(o){let i=o;console.error(`Error listing models for ${e}:`,i.message),console.log(t.defaultModel)}}var W=!1;function w(e){return e==="true"||e==="1"}var be={pipeline:{get:()=>n.pipeline,set:e=>{n.pipeline=e;try{n.decopipe=JSON.parse(r2.cmd("cat "+e))}catch(t){console.error(t)}}},model:{get:()=>n.model,set:e=>{e==="?"?ye(n.api):n.model=e.trim()}},deterministic:{get:()=>n.deterministic,set:e=>{n.deterministic=w(e)}},files:{get:()=>n.useFiles,set:e=>{n.useFiles=w(e)}},think:{get:()=>n.think,set:e=>{n.think=e==="true"?1:e==="false"?0:+e}},debug:{get:()=>n.debug,set:e=>{n.debug=w(e)}},timeout:{get:()=>n.timeout,set:e=>{n.timeout=Math.max(0,parseInt(e,10)||0)}},api:{get:()=>n.api,set:e=>{if(e==="?"){let t=z().join(`
-`);console.error(t)}else n.api=e}},lang:{get:()=>n.language,set:e=>{n.language=e}},hlang:{get:()=>n.humanLanguage,set:e=>{n.humanLanguage=e}},cache:{get:()=>n.cache,set:e=>{n.cache=w(e)}},cmds:{get:()=>n.commands,set:e=>{n.commands=e}},tts:{get:()=>n.tts,set:e=>{n.tts=w(e)}},yolo:{get:()=>n.yolo,set:e=>{n.yolo=w(e)}},prompt:{get:()=>n.prompt,set:e=>{n.prompt=e}},ctxfile:{get:()=>n.contextFile,set:e=>{n.contextFile=e}},baseurl:{get:()=>n.baseurl,set:e=>{n.baseurl=e}},headers:{get:()=>le(n.extraHeaders),set:e=>{n.extraHeaders=$(e)}},maxtokens:{get:()=>n.maxInputTokens,set:e=>{n.maxInputTokens=parseInt(e,10)||-1}}};function Be(){let e={host:n.host,port:n.port,lastOutput:n.lastOutput};Object.assign(n,k,e),n.decopipe={...k.decopipe},n.extraHeaders=[...k.extraHeaders]}function Ye(e){let t=e.trim();return!t||t.startsWith("#")?null:t.startsWith("decai -e ")?t.slice(9).trim():t}function b(e){let t=e.indexOf("="),o=t===-1?e:e.slice(0,t),i=t===-1?void 0:e.slice(t+1),r=be[o];if(!r){console.error("Unknown config key");return}typeof i<"u"?r.set(i):console.log(r.get())}function xe(){Object.entries(be).forEach(([e,t])=>{let o=t.get();console.log("decai -e "+e+"="+o)})}function ke(){if(Be(),!O(I))return;let e=r2.cmd("'cat "+I);for(let t of e.split(/\r?\n/)){let o=Ye(t);o&&b(o)}}function Ee(){W||(ke(),W=!0)}function Ae(){U(A,I),r2.cmd("'ed "+I),ke(),W=!0}function f(e,t,o=!1){let i=(t||"").replace(/`/g,""),r=e.replace(/'/g,"");if(n.api==="r2"||n.api==="r2ai"){let l=X(".pdc.txt");oe(l,i);let s=r.startsWith("-")?r:["-i",l,r].join(" "),p=(n.baseurl?n.baseurl+"/cmd":n.host+":"+n.port+"/cmd")+"/"+encodeURIComponent(s),d=v(p,[]);return d.error?`Error: ${d.error}`:d.result||JSON.stringify(d)||"Cannot curl, use r2ai-server or r2ai -w"}if(r.startsWith("-"))return"";let a=r+`:
-`+i;return n.maxInputTokens>0&&a.length>n.maxInputTokens&&(a=a.slice(0,n.maxInputTokens)),ge(a,o)}function Pe(){let e="",t=o=>e+=" "+y+" "+o+`
-`;e+="Usage: "+y+` (-h) ...
-`,e+="Version: "+Q+`
-`,t("-a [query]    - solve query with auto mode"),t("-b [url]      - set base URL (alias for decai -e baseurl)"),t("-d [f1 ..]    - decompile given functions"),t("-dd [..]      - same as above, but ignoring cache"),t("-dD [query]   - decompile current function with given extra query"),t("-dr           - decompile function and its called ones (recursive)"),t("-e            - display and change eval config vars"),t("-E            - edit decai.txt"),t("-h            - show this help"),t("-H            - help setting up r2ai"),t("-i [f] [q]    - include given file and query"),t("-k            - list API key status"),t("-K            - edit apikeys.txt"),t("-m [model]    - use -m? or -e model=? to list the available models"),t("-n            - suggest better function name"),t("-p [provider] - same as decai -e api (will be provider)"),t("-q [text]     - query language model with given text"),t("-Q [text]     - query on top of the last output"),t("-r [prompt]   - change role prompt (same as: decai -e prompt)"),t("-R            - reset role prompt to default prompt"),t("-s            - function signature"),t("-v            - show local variables"),t("-V            - find vulnerabilities"),t("-x[*]         - eXplain current function (-x* for r2 script)"),r2.log(e.trim())}function D(e,t,o,i){if(o){let s=r2.cmd("anos").trim();if(s.length>0)return s}let r="";if(i){let s=r2.cmd("s");r+=`## Context functions:
-`;let c=r2.cmdAt("axff~^C[2]~$$",s);for(let p of c.split(/\n/g))r+=r2.cmd("pdc@"+p);r2.cmd("s "+s)}let a=t?" "+e:"",l=r2.cmd("e scr.color");try{let s=e.slice(2).trim(),c=0,p="";n.contextFile!==""&&r2.cmd2("test -f "+n.contextFile).value===0&&(p+=`## Context:
+`};var E={decopipe:{use:!1},host:"http://localhost",port:"11434",baseurl:"",extraHeaders:[],api:"ollama",pipeline:"",commands:"pdc",yolo:!1,tts:!1,language:"C",humanLanguage:"English",deterministic:!0,debug:!1,timeout:180,http:"auto",think:"",apitype:"chat",useFiles:!1,contextFile:"",model:"",cache:!1,maxInputTokens:-1,prompt:T,lastOutput:""},r={...E,decopipe:{...E.decopipe},extraHeaders:[...E.extraHeaders]};function re(e){return(r2.cmd("-e dir.tmp").trim()??".")+"/"+e}function ne(e){if(!e.startsWith("~/"))return e;let t=r2.cmd("'%HOME").trim()||r2.cmd("'%USERPROFILE").trim();return t?t+e.slice(1):e}function H(e,t){r2.cmd("'mkdir -p "+ne(e)),r2.cmd("'touch "+ne(t))}function _(e){return(r2.cmd2("test -h").logs?.[0]?.message??"").includes("-fdx")?!0:r2.cmd("'test -vf "+e).startsWith("found")}function G(e,t){return e+" ".repeat(Math.max(0,t-e.length))}function h(e,t){return e.replace(/\/+$/,"")+"/"+t.replace(/^\/+/,"")}function oe(e){return e.replace(/\x1b\[[0-9;]*m/g,"")}function ie(e){let t=e.match(/```json\s*([\s\S]*?)```/);return t&&t[1]?t[1].trim():e}function ae(e){let t=e,n=t.indexOf("{");n!==-1&&(t=t.slice(n));let o=t.lastIndexOf("}");return o!==-1&&(t=t.slice(0,o+1)),t}function $(e){return btoa(e)}function se(e,t){let n=$(t);r2.cmd("p6ds "+n+" > "+e)}function k(e){let t=e;return t=t.replace(/<think>[\s\S]*?<\/think>/gi,""),t.split(`
+`).filter(n=>!n.trim().startsWith("```")).join(`
+`)}function J(e){r.debug&&console.log(e)}function le(e){let t={};for(let n of e.split(/\r?\n/)){let o=n.trim();if(!o||o.startsWith("#"))continue;let[i,...a]=o.split("=");if(!i||a.length===0)continue;let s=a.join("=").trim(),l=i.toLowerCase().replace(/_api_key$/i,"");t[l]=s}return t}var Ue=["DECAI_HEADERS","R2AI_HEADERS"];function _e(e){let t=r2.cmd("'%"+e).trim();return t!==""&&!t.includes("=")?t:""}function Me(e){return e.trim().toLowerCase()}function Ne(e){let t=e.indexOf(":"),n=e.indexOf("="),o=t!==-1&&(n===-1||t<n)?t:n;if(o===-1)return null;let i=e.slice(0,o).trim();return i===""?null:{name:i,value:e.slice(o+1).trim()}}function Te(e){return e.value===""?`${e.name}:`:`${e.name}: ${e.value}`}function ce(e,t){for(let n of e){let o=Ne(n.trim());o&&t.set(Me(o.name),Te(o))}}function W(e){let t=new Map,n=e.replace(/\\n/g,`
+`).split(/\r?\n/g).filter(o=>{let i=o.trim();return i&&!i.startsWith("#")});return ce(n,t),Array.from(t.values())}function ue(e){return e.join("\\n")}function A(...e){let t=new Map;for(let n of e)ce(n,t);return Array.from(t.values())}function He(){for(let e of Ue){let t=_e(e);if(t)return W(t)}return[]}function de(){return A(He(),r.extraHeaders)}function M(e){return`'${e.replace(/'/g,`'"'"'`)}'`}function Ke(e){let{method:t,url:n,headers:o,payload:i}=e,a=A(["Content-Type: application/json"],o);if(r.http!=="curl"){if(r2.http)return r2.http(t,n,a.join(`
+`),i??"").body;if(r.http==="r2")throw new Error("r2.http is not available in this runtime")}let s=["curl","-s"];if(r.timeout>0&&s.push("--max-time",String(r.timeout)),a.forEach(l=>s.push("-H",M(l))),t==="POST"){if(!i)throw new Error("Payload required for POST requests");let l=r2.fdump(i);s.push("--data-binary","@"+M(l),M(n));let u=s.join(" ")+"; rm "+M(l);return J(u),r2.syscmds(u)}else{s.push(M(n));let l=s.join(" ");return J(l),r2.syscmds(l)}}function qe(e){let t=e.split(`
+`).filter(i=>i.trim()!=="");if(t.length<=1)return null;let n="",o=null;for(let i of t)try{let a=JSON.parse(i);if(a.error)return a;a.message?.content!==void 0&&(n+=a.message.content),o=a}catch{return null}return o?(o.message={role:"assistant",content:n},o):null}function pe(e){try{let t=Ke(e).trim();if(t==="")return{error:"empty response"};try{return JSON.parse(t)}catch{let n=qe(t);return n||{error:t,rawOutput:t}}}catch(t){return{error:t.message}}}function O(e,t){return pe({method:"GET",url:e,headers:t})}function fe(e,t,n){return pe({method:"POST",url:e,headers:t,payload:n})}function z(e){return e.split("_")[0].toLowerCase()}function me(e){if(!e.authKey)return;let t=Array.from(new Set([e.keyName,z(e.authKey)].filter(Boolean))),n;for(let o of t){let i=K(o,e.authKey);if(i[0])return i;!n&&i[2]!=="nope"&&(n=i)}return n||K(z(e.authKey),e.authKey)}function S(e){if(e)return typeof e=="string"?e:e.message||JSON.stringify(e)}function he(e){return r.baseurl||e.defaultBaseUrl}function ge(e,t=de()){return A(e,t)}function De(e){return e?["Authorization: Bearer "+e]:[]}function je(e){let t=["anthropic-version: 2023-06-01"];return e?A(t,["x-api-key: "+e]):t}function ye(e,t){switch(e.authStyle||"none"){case"bearer":return De(t);case"anthropic":return je(t);case"none":default:return[]}}function Be(e,t,n){let o=O(e,t),i=S(o.error);return i?(console.error(i),"error invalid response"):o.data?.map(n).join(`
+`)||""}function Fe(e,t){let n=new Set;return e.filter(o=>{let i=t(o);return n.has(i)?!1:(n.add(i),!0)})}function V(e){let t=me(e)?.[0]||null;return{baseUrl:he(e),headers:ge(ye(e,t))}}function Le(e){let{baseUrl:t,headers:n}=V(e),o=h(t,"v1/models");return Be(o,n,i=>i.id)}function Ge(e){let{baseUrl:t,headers:n}=V(e),o=O(h(t,"api/tags"),n),i=S(o.error);return i?(console.error(i),"error invalid response"):o.models?.map(a=>a.name).join(`
+`)||""}function $e(e){let{baseUrl:t,headers:n}=V(e),o=O(h(t,"v1/models"),n);return o.data?Fe(o.data,i=>i.name||i.id).map(i=>[G(i.name||i.id,30),G(""+(i.max_context_length||""),10),i.description||""].join(" ")).join(`
+`):""}var Je={openai:{buildPayload:(e,t)=>({stream:!1,model:e,messages:[{role:"user",content:t}]}),parseResponse:e=>{if(e.error)throw new Error(S(e.error)||"Unknown error");if(e.choices&&e.choices[0]?.message?.content)return k(e.choices[0].message.content);throw new Error("Invalid response format")},buildUrl:e=>h(e,"v1/chat/completions")},anthropic:{buildPayload:(e,t)=>{let n={model:e,max_tokens:5128,messages:[{role:"user",content:t}]};return Y()&&(n.thinking={type:"enabled",budget_tokens:4096},n.max_tokens=16e3),r.deterministic&&!Y()&&Object.assign(n,{temperature:0,top_p:0,top_k:1}),n},parseResponse:e=>{if(e.content&&Array.isArray(e.content)){let t=[];for(let n of e.content)n.text&&t.push(n.text);if(t.length>0)return k(t.join(`
+`))}throw e.error?new Error(S(e.error)||"Unknown error"):new Error("Invalid response format")},buildUrl:e=>h(e,"v1/messages")},ollama:{buildPayload:(e,t)=>{let n={stream:!1,model:e};return r.apitype==="generate"?n.prompt=t:n.messages=[{role:"user",content:t}],Y()?r.think==="true"||r.think==="1"?n.think=!0:n.think=r.think:n.think=!1,r.deterministic&&(n.options={repeat_last_n:0,top_p:1,top_k:1,temperature:0,repeat_penalty:1,seed:123}),n},parseResponse:e=>{if(e.error)throw new Error(S(e.error));if(typeof e.response=="string")return k(e.response);if(e.message?.content)return k(e.message.content);throw new Error(JSON.stringify(e))},buildUrl:e=>h(e,"api/"+r.apitype)},gemini:{buildPayload:(e,t)=>{let n={contents:[{parts:[{text:t}]}]},o={};return r.deterministic&&Object.assign(o,{temperature:0,topP:1,topK:1}),r.think!==""&&(X()?o.thinkingConfig={thinkingBudget:0}:o.thinkingConfig={thinkingBudget:8192}),Object.keys(o).length>0&&(n.generationConfig=o),n},parseResponse:e=>{if(e.candidates&&e.candidates[0]?.content?.parts){let n=e.candidates[0].content.parts.filter(o=>!o.thought&&o.text).map(o=>o.text);if(n.length>0)return k(n.join(`
+`))}throw e.error?new Error(S(e.error)):(console.log(JSON.stringify(e)),new Error("Invalid response format"))},buildUrl:(e,t,n)=>h(e,`v1beta/models/${t}:generateContent?key=${n}`),requiresUrlApiKey:!0}},q={anthropic:{defaultModel:"claude-sonnet-5-5",defaultBaseUrl:"https://api.anthropic.com",authKey:"ANTHROPIC_API_KEY",keyName:"anthropic",authStyle:"anthropic",apiStyle:"anthropic"},claude:{defaultModel:"claude-sonnet-5-5",defaultBaseUrl:"https://api.anthropic.com",authKey:"ANTHROPIC_API_KEY",keyName:"claude",authStyle:"anthropic",apiStyle:"anthropic"},openai:{defaultModel:"gpt-4o-mini",defaultBaseUrl:"https://api.openai.com",authKey:"OPENAI_API_KEY",authStyle:"bearer",apiStyle:"openai"},ollama:{defaultModel:"qwen2.5-coder:latest",defaultBaseUrl:"http://localhost:11434",authStyle:"none",apiStyle:"ollama"},ollamacloud:{defaultModel:"gpt-oss:120b",defaultBaseUrl:"https://ollama.com",authKey:"OLLAMA_API_KEY",keyName:"ollamacloud",authStyle:"bearer",apiStyle:"ollama"},zen:{defaultModel:"big-pickle",defaultBaseUrl:"https://opencode.ai/zen",authKey:"OPENCODE_API_KEY",keyName:"zen",authStyle:"bearer",apiStyle:"openai",hardcodedModels:["big-pickle","glm-5","kimi-k2.5"]},gemini:{defaultModel:"gemini-2.5-flash",defaultBaseUrl:"https://generativelanguage.googleapis.com",authKey:"GEMINI_API_KEY",authStyle:"none",apiStyle:"gemini",hardcodedModels:["gemini-2.0-flash","gemini-2.0-flash-lite","gemini-2.5-pro","gemini-2.5-flash","gemini-2.5-flash-lite"]},mistral:{defaultModel:"codestral-latest",defaultBaseUrl:"https://api.mistral.ai",authKey:"MISTRAL_API_KEY",authStyle:"bearer",apiStyle:"openai"},xai:{defaultModel:"grok-beta",defaultBaseUrl:"https://api.x.ai",authKey:"XAI_API_KEY",authStyle:"bearer",apiStyle:"openai",hardcodedModels:["grok-2","grok-beta"]},lmstudio:{defaultModel:"local-model",defaultBaseUrl:"http://127.0.0.1:1234",authStyle:"none",apiStyle:"openai"},deepseek:{defaultModel:"deepseek-coder",defaultBaseUrl:"https://api.deepseek.com",authKey:"DEEPSEEK_API_KEY",authStyle:"bearer",apiStyle:"openai",hardcodedModels:["deepseek-coder","deepseek-chat"]},openrouter:{defaultModel:"openai/gpt-4o-mini",defaultBaseUrl:"https://openrouter.ai/api",authKey:"OPENROUTER_API_KEY",authStyle:"bearer",apiStyle:"openai"},groq:{defaultModel:"llama-3.3-70b-versatile",defaultBaseUrl:"https://api.groq.com/openai",authKey:"GROQ_API_KEY",authStyle:"bearer",apiStyle:"openai"}};function D(e){return q[e]}function Q(){return Object.keys(q)}function be(e){let t=D(e);if(!t)throw new Error(`Unknown provider: ${e}`);if(e==="mistral")return $e(t);switch(t.apiStyle){case"openai":case"anthropic":return Le(t);case"ollama":return Ge(t);case"gemini":return"";default:return""}}function X(){return r.think==="false"||r.think==="0"}function Y(){return r.think!==""&&!X()}function We(e,t,n){let o=e;return r.think!==""&&n==="openai"&&(X()?(o+=' Answers directly and concisely, without showing any thinking steps or internal reasoning. Never include phrases like "Let me think".',o+=" /no_think"):o="Think step by step and explain the reasoning process, When answering, first output your reasoning inside <think> and </think> tags, then give the final answer."+o),t?o:r.prompt+Ye()+o}function Ye(){return`
+.Translate the code into `+r.language+` programming language
+`}function ze(e,t,n,o){let i=r.model||e.defaultModel,a=We(n,o,e.apiStyle),s=me(e);if(t.requiresUrlApiKey&&s?.[1]&&e.authKey)return`Cannot read ~/.r2ai.${z(e.authKey)}-key`;let l=t.buildPayload(i,a),u=ye(e,s?.[0]||null);e.apiStyle==="ollama"&&u.push("Accept: application/x-ndjson");let v=ge(u),c=t.buildUrl(he(e),i,s?.[0]||void 0);try{return t.parseResponse(fe(c,v,JSON.stringify(l)))}catch(d){return"ERROR: "+d.message}}function ke(e,t){let n=D(r.api);return n?ze(n,Je[n.apiStyle],e,t):`Unknown value for 'decai -e api'. Available: ${Q().join(", ")}`}function Ve(){return Object.fromEntries(Object.entries(q).filter(([,e])=>e.authKey).map(([e,t])=>[e,t.authKey]))}function K(e,t){let n=r2.cmd("'%"+t).trim();if(!n.includes("=")&&n!=="")return[n.trim(),null,"env"];let o=e.toLowerCase(),i=N;if(_(i)){let s=r2.cmd("'cat "+i),l=le(s);if(o in l)return[l[o],null,"txt"]}let a="~/.r2ai."+o+"-key";if(_(a)){let s=r2.cmd("'cat "+a);return s===""?[null,"Cannot read "+a,"no"]:[s.trim(),null,"file"]}return[null,"Not available","nope"]}function xe(){H(w,N),r2.cmd("'ed "+N)}function Z(){Object.entries(Ve()).forEach(([e,t])=>{let n=K(e,t)[2];console.log(n,"	",e)})}function ve(e){let t=D(e);if(!t){console.error(`Unknown provider: ${e}`);return}try{let n=be(e);n&&console.log(n),t.hardcodedModels&&t.hardcodedModels.forEach(o=>console.log(o)),e==="mistral"&&console.log("ministral-8b-latest"),!n&&!t.hardcodedModels&&console.log(t.defaultModel)}catch(n){let o=n;console.error(`Error listing models for ${e}:`,o.message),console.log(t.defaultModel)}}var ee=!1;function Qe(e){return e==="true"||e==="1"}function P(e){let t=r;return{get:()=>t[e],set:n=>{t[e]=Qe(n)}}}function R(e){let t=r;return{get:()=>t[e],set:n=>{t[e]=n}}}function Xe(e){let t=e.trim();return t==="chat"?"chat":t==="generate"?"generate":null}var Ee={pipeline:{get:()=>r.pipeline,set:e=>{r.pipeline=e;try{r.decopipe=JSON.parse(r2.cmd("cat "+e))}catch(t){console.error(t)}}},model:{get:()=>r.model,set:e=>{e==="?"?ve(r.api):r.model=e.trim()}},deterministic:P("deterministic"),files:P("useFiles"),think:{get:()=>r.think||"false",set:e=>{r.think=e}},apitype:{get:()=>r.apitype,set:e=>{let t=Xe(e);t?r.apitype=t:console.error("Invalid apitype. Use chat or generate.")}},debug:P("debug"),http:{get:()=>r.http,set:e=>{let t=e.trim();t==="auto"||t==="r2"||t==="curl"?r.http=t:console.error("Invalid http backend. Use auto, r2 or curl.")}},timeout:{get:()=>r.timeout,set:e=>{r.timeout=Math.max(0,parseInt(e,10)||0)}},api:{get:()=>r.api,set:e=>{e==="?"?console.error(Q().join(`
+`)):r.api=e}},lang:R("language"),hlang:R("humanLanguage"),cache:P("cache"),cmds:R("commands"),tts:P("tts"),yolo:P("yolo"),prompt:R("prompt"),ctxfile:R("contextFile"),baseurl:R("baseurl"),headers:{get:()=>ue(r.extraHeaders),set:e=>{r.extraHeaders=W(e)}},maxtokens:{get:()=>r.maxInputTokens,set:e=>{r.maxInputTokens=parseInt(e,10)||-1}}};function Ze(){let e={host:r.host,port:r.port,lastOutput:r.lastOutput};Object.assign(r,E,e),r.decopipe={...E.decopipe},r.extraHeaders=[...E.extraHeaders]}function et(e){let t=e.trim();return!t||t.startsWith("#")?null:t.startsWith("decai -e ")?t.slice(9).trim():t}function x(e){let t=e.indexOf("="),n=t===-1?e:e.slice(0,t),o=t===-1?void 0:e.slice(t+1),i=Ee[n];if(!i){console.error("Unknown config key");return}typeof o<"u"?i.set(o):console.log(i.get())}function we(){Object.entries(Ee).forEach(([e,t])=>{let n=t.get();console.log("decai -e "+e+"="+n)})}function Ae(){if(Ze(),!_(U))return;let e=r2.call("cat "+U);for(let t of e.split(/\r?\n/)){let n=et(t);n&&n[0]!="#"&&x(n)}}function Oe(){ee||(Ae(),ee=!0)}function Se(){H(w,U),r2.cmd("'ed "+U),Ae(),ee=!0}function p(e,t,n=!1){let o=(t||"").replace(/`/g,""),i=e.replace(/'/g,"");if(r.api==="r2"||r.api==="r2ai"){let s=re(".pdc.txt");se(s,o);let l=i.startsWith("-")?i:["-i",s,i].join(" "),u=r.baseurl||r.host+":"+r.port,v=h(u,"cmd/"+encodeURIComponent(l)),c=O(v,[]);return c.error?`Error: ${c.error}`:c.result||JSON.stringify(c)||"Cannot curl, use r2ai-server or r2ai -w"}if(i.startsWith("-"))return"";let a=i+`:
+`+o;return r.maxInputTokens>0&&a.length>r.maxInputTokens&&(a=a.slice(0,r.maxInputTokens)),ke(a,n)}function Re(e){let t=r2.cmd("e scr.color");r2.cmd("e scr.color=0");try{return e()}finally{r2.cmd("e scr.color="+t)}}function Pe(){let e="",t=n=>e+=" "+b+" "+n+`
+`;e+="Usage: "+b+` (-h) ...
+`,e+="Version: "+te+`
+`,t("-a [query]    - solve query with auto mode"),t("-b [url]      - set base URL (alias for decai -e baseurl)"),t("-d [f1 ..]    - decompile given functions"),t("-dd [..]      - same as above, but ignoring cache"),t("-dD [query]   - decompile current function with given extra query"),t("-dr           - decompile function and its called ones (recursive)"),t("-e            - display and change eval config vars"),t("-E            - edit and run decai.txt"),t("-h            - show this help"),t("-H            - help setting up r2ai"),t("-i [f] [q]    - include given file and query"),t("-k            - list API key status"),t("-K            - edit apikeys.txt"),t("-m [model]    - use -m? or -e model=? to list the available models"),t("-n            - suggest better function name"),t("-p [provider] - same as decai -e api (will be provider)"),t("-q [text]     - query language model with given text"),t("-Q [text]     - query on top of the last output"),t("-r [prompt]   - change role prompt (same as: decai -e prompt)"),t("-R            - reset role prompt to default prompt"),t("-s            - function signature"),t("-v            - show local variables"),t("-V            - find vulnerabilities"),t("-x[*]         - eXplain current function (-x* for r2 script)"),r2.log(e.trim())}function j(e,t,n,o){if(n){let s=r2.cmd("anos").trim();if(s.length>0)return s}let i="";if(o){let s=r2.cmd("s");i+=`## Context functions:
+`;let l=r2.cmdAt("axff~^C[2]~$$",s);for(let u of l.split(/\n/g))i+=r2.cmd("pdc@"+u);r2.cmd("s "+s)}let a=t?" "+e:"";try{let s=e.slice(2).trim(),l="";r.contextFile!==""&&r2.cmd2("test -f "+r.contextFile).value===0&&(l+=`## Context:
 [START]
-`+r2.cmd("cat "+n.contextFile)+`
+`+r2.cmd("cat "+r.contextFile)+`
 [END]
-`),r2.cmd("e scr.color=0");let d=`## Before:
-`;for(let m of n.commands.split(",")){if(m.trim()==="")continue;let H=t||s.trim().length===0?m:m+"@@= "+s,x=r2.cmd(H);x.length>5&&(d+="Output of "+m+`:
+`);let{body:u,count:v}=Re(()=>{let d=`## Before:
+`,y=0;for(let g of r.commands.split(",")){if(g.trim()==="")continue;let C=t||s.trim().length===0?g:g+"@@= "+s,I=r2.cmd(C);I.length>5&&(d+="Output of "+g+`:
 [START]
-`+x+`
+`+I+`
 [END]
-`,c++)}if(d+=`## After:
-`,r2.cmd("e scr.color="+l),c===0){console.error("Nothing to do.");return}let u="";if(n.decopipe.use){let m=n.decopipe[n.decopipe.default],H=n.model,x=p+d;for(let L of m.pipeline){L.model&&(n.model=L.model);let V=L.query+". "+m.globalQuery;u=f(V,x,!0),n.debug&&console.log(`QUERY
-`,V,`
+`,y++)}return d+=`## After:
+`,{body:d,count:y}});if(v===0){console.error("Nothing to do.");return}let c="";if(r.decopipe.use){let d=r.model;try{let y=r.decopipe[r.decopipe.default],g=l+u;for(let C of y.pipeline){C.model&&(r.model=C.model);let I=C.query+". "+y.globalQuery;c=p(I,g,!0),r.debug&&console.log(`QUERY
+`,I,`
 INPUT
-`,x,`
+`,g,`
 OUTPUT
-`,u),x=u}u=x,n.model=H}else{let m=a;p+=d+r,u=f(m,p,!1),n.lastOutput=u}return o&&u.length>1&&r2.call("ano=base64:"+Y(u)),u.startsWith("```")&&(u=u.replace(/```.*\n/,"").replace(/```$/,"")),u.trim()}catch(s){r2.cmd("e scr.color="+l);let c=s;console.error(c,c.stack);return}}function je(){let e=r2.cmd("e scr.color");r2.cmd("e scr.color=0");let t="[START]"+n.commands.split(",").map(r=>r2.cmd(r)).join(`
-`)+"[END]";r2.cmd("e scr.color="+e);let i=f("Analyze function calls, references, comments and strings, loops and ignore registers and memory accesses. Explain the purpose of this function in a single short sentence. /no_think Do not introduce or argue the response, translation of the explanation in "+n.humanLanguage,t,!0).trim().split(/\n/g);return i[i.length-1].trim()}function Ge(){let e=n.language,t=r2.cmd("afv;pdc");n.language="C";let o="'afs "+f("analyze the uses of the arguments and return value to infer the signature, identify which is the correct type for the return. Do NOT print the function body, ONLY output the function signature, like if it was going to be used in a C header",t,!1),i=o.indexOf("{");return i!==-1&&(o=o.substring(0,i)),n.language=e,o}function $e(e,t){let o=[];for(;;){let i=F.auto;if(o.length>0){i+=`## Command Results
+`,c),g=c}c=g}finally{r.model=d}}else{let d=a;l+=u+i,c=p(d,l,!1),r.lastOutput=c}return n&&c.length>1&&r2.call("ano=base64:"+$(c)),c.startsWith("```")&&(c=c.replace(/```.*\n/,"").replace(/```$/,"")),c.trim()}catch(s){let l=s;console.error(l,l.stack);return}}function tt(){let e=Re(()=>"[START]"+r.commands.split(",").map(o=>r2.cmd(o)).join(`
+`)+"[END]"),n=p("Analyze function calls, references, comments and strings, loops and ignore registers and memory accesses. Explain the purpose of this function in a single short sentence. /no_think Do not introduce or argue the response, translation of the explanation in "+r.humanLanguage,e,!0).trim().split(/\n/g);return n[n.length-1].trim()}function nt(){let e=r.language,t=r2.cmd("afv;pdc");r.language="C";let n="'afs "+p("analyze the uses of the arguments and return value to infer the signature, identify which is the correct type for the return. Do NOT print the function body, ONLY output the function signature, like if it was going to be used in a C header",t,!1),o=n.indexOf("{");return o!==-1&&(n=n.substring(0,o)),r.language=e,n}function rt(e,t){let n=[];for(;;){let o=L.auto;if(n.length>0){o+=`## Command Results
 
-`;for(let a of o){let l=JSON.parse(a);i+="### "+l.command+"\n\n```\n"+l.response+"\n```\n"}}i+=`
+`;for(let a of n){let s=JSON.parse(a);o+="### "+s.command+"\n\n```\n"+s.response+"\n```\n"}}o+=`
 
 ## User Prompt
 
-`+e,n.debug&&console.log(`#### input
+`+e,r.debug&&console.log(`#### input
+`,o,`
+#### /input`),console.log("Thinking...");let i=p("",o,!0);r.debug&&console.log(`#### output
 `,i,`
-#### /input`),console.log("Thinking...");let r=f("",i,!0);n.debug&&console.log(`#### output
-`,r,`
-#### /output`);try{let a=JSON.parse(te(ee(E(r))));if(a.action==="r2cmd"||a.action==="response"||a.action===a.command){let l=a.command||"";a.reason&&(console.log("[r2cmd] Reasoning: "+a.reason),n.tts&&(r2.syscmd("pkill say"),r2.syscmd("say -v Alex -r 250 '"+a.reason.replace(/'/g,"")+"' &"))),console.log("[r2cmd] Action: "+a.description),console.log("[r2cmd] Command: "+l);let s=l;n.yolo||(s=Je(l,t)),console.log("[r2cmd] Running: "+s);let c=r2.cmd2(s),p=c.logs?c.logs.map(m=>m.type+": "+m.message).join(`
-`):"",d=(c.res+p).trim();console.log(d);let u=Z(d);n.debug&&console.log(`<r2output>
-`,u,`
-<(r2output>`),o.push(JSON.stringify({action:"response",command:s,description:a.description,response:u}))}else if(a.action==="reply"){console.log(`Done
-`,a.response);break}else console.log(`Unknown response
-`,JSON.stringify(r))}catch(a){let l=r.indexOf('response": "');if(l!==-1){let s=r.slice(l+12).replace(/\\n/g,`
-`).replace(/\\/g,"");console.log(s)}else console.log(r),console.error(a);break}}}function Je(e,t){for(;;){let o=r2.cmd("'?ie Tweak command? ('?' for help)").trim();if(o==="q!"){console.error("Break!");break}if(o==="?"){ze();continue}else if(o.startsWith(":")){console.log(r2.cmd(o.slice(1)));continue}else if(o.startsWith("-e")){t(o);continue}else{if(o==="!")return"?e do NOT execute '"+e+"' again, continue without it";if(o.startsWith("!")){console.log(r2.syscmd(o.slice(1)));continue}else{if(o==="q")return"?e All data collected!. Do not call more commands, reply the solutions";if(o){let i=o.indexOf("#");return i!==-1?o.slice(0,i).trim():o}else return e}}}return e}function ze(){console.log(" '!'     do not run this command"),console.log(" '!c'    run system command"),console.log(" 'q'     to quit auto and try to solve"),console.log(" 'q!'    quit auto without solving"),console.log(" 'c # C' use given command with comment"),console.log(" ':c'    run r2 command without feeding auto"),console.log(" '-e'    set decai configuration variables")}function We(e){return e.slice(2).trim()}function ve(e,t){if(e===""||!e.startsWith("-")){Pe();return}let o="",i=e[1],r=We(e);switch(i){case"H":console.log(F.decai);break;case"a":$e(r,t);break;case"m":{r==="="?b("model="):r?b("model="+r):b("model");break}case"n":case"f":{o=r2.cmd("axff~$[3]");let a=r2.cmd("fd.").trim().split(/\n/).filter(l=>!l.startsWith("secti")).join(",");o=f("give me a better name for this function. the output must be: 'afn NEWNAME'. do not include the function code, only the afn line. consider: "+a,o,!1).trim(),o+=" @ "+r2.cmd("?v $FB").trim();break}case"v":o=r2.cmd("afv;pdc"),o=f("guess a better name and type for each local variable and function argument taking using. output an r2 script using afvn and afvt commands",o,!1);break;case"i":{let a=r.split(/\s+/).filter(Boolean);if(a.length>=2){let l=a[0],s=a.slice(1).join(" "),c=r2.cmd("cat "+l);console.log(f(s,c,!0))}else console.log("Use: decai -i [file] [query ...]");break}case"p":{r?b("api="+r):G();break}case"r":{r?n.prompt=r:console.log(n.prompt);break}case"R":n.prompt=N;break;case"s":o=Ge();break;case"V":o=f("find vulnerabilities, dont show the code, only show the response, provide a sample exploit",n.lastOutput,!1);break;case"K":re();break;case"E":Ae();break;case"k":G();break;case"b":{r?b("baseurl="+r):console.log(n.baseurl);break}case"e":{r?b(r):xe();break}case"q":try{o=f(r,null,!0)}catch(a){let l=a;console.error(l,l.stack)}break;case"Q":o=f(r,n.lastOutput,!1);break;case"x":o=je(),(e[2]==="*"||e[2]==="r")&&(o="'CC "+o);break;case"d":e[2]==="r"?o=D(e.slice(2),!0,n.cache,!0)||"":e[2]==="d"?o=D(e,!1,!1,!1)||"":e[2]==="D"?o=D(e,!0,!1,!1)||"":o=D(e,!1,n.cache,!1)||"";break;default:Pe();break}return o||void 0}function we(e){Ee();let t=ve(e,we);return t&&r2.log(t),!0}function Ve(){r2.unload("core",y),r2.plugin("core",function(){function e(t){if(t.startsWith(y)){let o=t.slice(y.length).trim();return we(o)}return!1}return{name:y,license:"MIT",desc:"r2 decompiler based on r2ai",call:e}})}Ve();
+#### /output`);try{let a=JSON.parse(ae(ie(k(i))));if(a.action==="r2cmd"||a.action==="response"||a.action===a.command){let s=a.command||"";a.reason&&(console.log("[r2cmd] Reasoning: "+a.reason),r.tts&&(r2.syscmd("pkill say"),r2.syscmd("say -v Alex -r 250 '"+a.reason.replace(/'/g,"")+"' &"))),console.log("[r2cmd] Action: "+a.description),console.log("[r2cmd] Command: "+s);let l=s;r.yolo||(l=ot(s,t)),console.log("[r2cmd] Running: "+l);let u=r2.cmd2(l),v=u.logs?u.logs.map(y=>y.type+": "+y.message).join(`
+`):"",c=(u.res+v).trim();console.log(c);let d=oe(c);r.debug&&console.log(`<r2output>
+`,d,`
+<(r2output>`),n.push(JSON.stringify({action:"response",command:l,description:a.description,response:d}))}else if(a.action==="reply"){console.log(`Done
+`,a.response);break}else{console.log(`Unknown response
+`,JSON.stringify(i));break}}catch(a){let s=i.indexOf('response": "');if(s!==-1){let l=i.slice(s+12).replace(/\\n/g,`
+`).replace(/\\/g,"");console.log(l)}else console.log(i),console.error(a);break}}}function ot(e,t){for(;;){let n=r2.cmd("'?ie Tweak command? ('?' for help)").trim();if(n==="q!"){console.error("Break!");break}if(n==="?"){it();continue}else if(n.startsWith(":")){console.log(r2.cmd(n.slice(1)));continue}else if(n.startsWith("-e")){t(n);continue}else{if(n==="!")return"?e do NOT execute '"+e+"' again, continue without it";if(n.startsWith("!")){console.log(r2.syscmd(n.slice(1)));continue}else{if(n==="q")return"?e All data collected!. Do not call more commands, reply the solutions";if(n){let o=n.indexOf("#");return o!==-1?n.slice(0,o).trim():n}else return e}}}return e}function it(){console.log(" '!'     do not run this command"),console.log(" '!c'    run system command"),console.log(" 'q'     to quit auto and try to solve"),console.log(" 'q!'    quit auto without solving"),console.log(" 'c # C' use given command with comment"),console.log(" ':c'    run r2 command without feeding auto"),console.log(" '-e'    set decai configuration variables")}function at(e){return e.slice(2).trim()}function B(){return r2.cmd("afi.").trim().length>0}var F="Cannot find function at current offset";function Ce(e,t){if(e===""||!e.startsWith("-")){Pe();return}let n="",o=e[1],i=at(e);switch(o){case"H":console.log(L.decai);break;case"a":rt(i,t);break;case"m":{i==="="?x("model="):i?x("model="+i):x("model");break}case"n":case"f":{if(!B()){n=F;break}n=r2.cmd("axff~$[3]");let a=r2.cmd("fd.").trim().split(/\n/).filter(s=>!s.startsWith("secti")).join(",");n=p("give me a better name for this function. the output must be: 'afn NEWNAME'. do not include the function code, only the afn line. consider: "+a,n,!1).trim(),n+=" @ "+r2.cmd("?v $FB").trim();break}case"v":{if(!B()){n=F;break}n=r2.cmd("afv;pdc"),n=p("guess a better name and type for each local variable and function argument taking using. output an r2 script using afvn and afvt commands",n,!1);break}case"i":{let a=i.split(/\s+/).filter(Boolean);if(a.length>=2){let s=a[0],l=a.slice(1).join(" "),u=r2.cmd("cat "+s);console.log(p(l,u,!0))}else console.log("Use: decai -i [file] [query ...]");break}case"p":{i?x("api="+i):Z();break}case"r":{i?r.prompt=i:console.log(r.prompt);break}case"R":r.prompt=T;break;case"s":if(!B()){n=F;break}n=nt();break;case"V":n=p("find vulnerabilities, dont show the code, only show the response, provide a sample exploit",r.lastOutput,!1);break;case"K":xe();break;case"E":Se();break;case"k":Z();break;case"b":{i?x("baseurl="+i):console.log(r.baseurl);break}case"e":{i?x(i):we();break}case"q":try{n=p(i,null,!0)}catch(a){let s=a;console.error(s,s.stack)}break;case"Q":n=p(i,r.lastOutput,!1);break;case"x":if(!B()){n=F;break}n=tt(),(e[2]==="*"||e[2]==="r")&&(n="'CC "+n);break;case"d":e[2]==="r"?n=j(e.slice(2),!0,r.cache,!0)||"":e[2]==="d"?n=j(e,!1,!1,!1)||"":e[2]==="D"?n=j(e,!0,!1,!1)||"":n=j(e,!1,r.cache,!1)||"";break;default:Pe();break}return n||void 0}function Ie(e){Oe();let t=Ce(e,Ie);return t&&r2.log(t),!0}function st(){r2.unload("core",b),r2.plugin("core",function(){function e(t){if(t.startsWith(b)){let n=t.slice(b.length).trim();return Ie(n)}return!1}return{name:b,license:"MIT",desc:"r2 decompiler based on r2ai",call:e}})}st();

decai/r2plugin/README.md

@@ -0,0 +1,36 @@
+# decai r2plugin
+
+Boilerplate to embed `decai.r2.js` inside the radare2 core library using the
+external plugin system (XPS). Unlike `r2ai/r2plugin`, this does not link any
+native code: the r2.js script is converted into a C buffer and registered as
+an embedded startup script, so r2 runs it at startup exactly like a file in
+the `plugins/*.r2.js` directory (honoring `cfg.plugins`).
+
+The `decai.r2.js` bundle is built from the typescript sources with
+`r2frida-compile` (see `make -C src` in the decai directory). The C buffer is
+generated at build time with `rax2 -C` (make builds) or `js2c.py` (meson).
+
+## Usage
+
+decai lives inside the r2ai repository, so expose it as `libr/xps/p/decai`:
+
+```console
+cd radare2/libr/xps/p
+git clone https://github.com/radareorg/r2ai
+ln -s r2ai/decai decai
+```
+
+For acr/make builds:
+
+```console
+cd radare2
+make -C libr/xps EXTERNAL_PLUGINS=decai
+./configure-plugins
+make
+```
+
+For meson builds, make `libr/xps/p/meson.build` contain:
+
+```meson
+subdir('decai/r2plugin')
+```

decai/r2plugin/core/decai.r2.js.d

@@ -0,0 +1,2 @@
+/Users/pancake/prg/radare2/libr/xps/p/decai/r2plugin/core/decai.r2.js.o: \
+  /Users/pancake/prg/radare2/libr/xps/p/decai/r2plugin/core/decai.r2.js.c

decai/r2plugin/core/deps.h

@@ -0,0 +1,3 @@
+#ifdef R2_CORE_H
+extern RCorePlugin r_core_plugin_decai;
+#endif

decai/r2plugin/core/deps.mk

@@ -0,0 +1,10 @@
+DECAI_CORE_WD=$(LIBR)/xps/p/decai
+DECAI_CORE_OBJ= \
+	$(DECAI_CORE_WD)/r2plugin/core/plugin.o \
+	$(DECAI_CORE_WD)/r2plugin/core/decai.r2.js.o
+EXTERNAL_STATIC_OBJS+=$(DECAI_CORE_OBJ)
+
+# decai.r2.js can be rebuilt from typescript with r2frida-compile (make -C src)
+$(DECAI_CORE_WD)/r2plugin/core/decai.r2.js.c: $(DECAI_CORE_WD)/decai.r2.js
+	rax2 -C < $(DECAI_CORE_WD)/decai.r2.js | \
+		sed -e 's,^unsigned,const unsigned,' -e 's,buf,decai_qjs,g' > $@

decai/r2plugin/core/plugin.c

@@ -0,0 +1,35 @@
+/* decai - LGPL - Copyright 2026 - pancake */
+
+#include <r_core.h>
+
+// generated from ../decai.r2.js, see ../deps.mk
+extern const unsigned char decai_qjs[];
+extern const unsigned int decai_qjs_len;
+
+static bool decai_init(RCorePluginSession *cps) {
+	r_core_script_embed (cps->core, "decai", "qjs", (const char *)decai_qjs, (int)decai_qjs_len);
+	return true;
+}
+
+static bool decai_call(RCorePluginSession *cps, const char *input) {
+	return false;
+}
+
+RCorePlugin r_core_plugin_decai = {
+	.meta = {
+		.name = "decai",
+		.desc = "AI decompiler (embedded decai.r2.js)",
+		.author = "pancake",
+		.license = "MIT",
+	},
+	.init = decai_init,
+	.call = decai_call,
+};
+
+#ifndef R2_PLUGIN_INCORE
+R_API RLibStruct radare_plugin = {
+	.type = R_LIB_TYPE_CORE,
+	.data = &r_core_plugin_decai,
+	.version = R2_VERSION,
+};
+#endif

decai/r2plugin/core/plugin.d

@@ -0,0 +1,372 @@
+/Users/pancake/prg/radare2/libr/xps/p/decai/r2plugin/core/plugin.o: \
+  /Users/pancake/prg/radare2/libr/xps/p/decai/r2plugin/core/plugin.c \
+  /Users/pancake/prg/radare2/libr/include/r_core.h \
+  /Users/pancake/prg/radare2/libr/include/r_main.h \
+  /Users/pancake/prg/radare2/libr/include/r_types.h \
+  /Users/pancake/prg/radare2/libr/include/r_types_base.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/ctype.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/_ctype.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/cdefs.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_symbol_aliasing.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_posix_availability.h \
+  /Applications/Xcode.app/Contents/Developer/Toolchains/XcodeDefault.xctoolchain/usr/lib/clang/21/include/ptrcheck.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/_bounds.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/runetype.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/_types.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_types.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/machine/_types.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/arm/_types.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_pthread/_pthread_types.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_types/_size_t.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_types/_ct_rune_t.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_types/_rune_t.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_types/_wchar_t.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_types/_wint_t.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/types.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/appleapiopts.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/machine/types.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/arm/types.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_types/_int8_t.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_types/_int16_t.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_types/_int32_t.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_types/_int64_t.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_types/_u_int8_t.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_types/_u_int16_t.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_types/_u_int32_t.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_types/_u_int64_t.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_types/_intptr_t.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_types/_uintptr_t.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/machine/endian.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/arm/endian.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_endian.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/machine/_endian.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/arm/_endian.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/__endian.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/libkern/_OSByteOrder.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/libkern/arm/_OSByteOrder.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_types/_u_char.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_types/_u_short.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_types/_u_int.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_types/_caddr_t.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_types/_dev_t.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_types/_blkcnt_t.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_types/_blksize_t.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_types/_gid_t.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_types/_in_addr_t.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_types/_in_port_t.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_types/_ino_t.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_types/_ino64_t.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_types/_key_t.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_types/_mode_t.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_types/_nlink_t.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_types/_id_t.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_types/_pid_t.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_types/_off_t.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_types/_uid_t.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_types/_clock_t.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_types/_ssize_t.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_types/_time_t.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_types/_useconds_t.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_types/_suseconds_t.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_types/_rsize_t.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_types/_errno_t.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_types/_fd_def.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/Availability.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/AvailabilityVersions.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/AvailabilityInternal.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/AvailabilityInternalLegacy.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_types/_fd_setsize.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_types/_fd_set.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_types/_fd_clr.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_types/_fd_zero.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_types/_fd_isset.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_types/_fd_copy.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_pthread/_pthread_attr_t.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_pthread/_pthread_cond_t.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_pthread/_pthread_condattr_t.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_pthread/_pthread_mutex_t.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_pthread/_pthread_mutexattr_t.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_pthread/_pthread_once_t.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_pthread/_pthread_rwlock_t.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_pthread/_pthread_rwlockattr_t.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_pthread/_pthread_t.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_pthread/_pthread_key_t.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_types/_fsblkcnt_t.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_types/_fsfilcnt_t.h \
+  /Applications/Xcode.app/Contents/Developer/Toolchains/XcodeDefault.xctoolchain/usr/lib/clang/21/include/limits.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/limits.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/machine/limits.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/arm/limits.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/arm/_limits.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/syslimits.h \
+  /Applications/Xcode.app/Contents/Developer/Toolchains/XcodeDefault.xctoolchain/usr/lib/clang/21/include/stdint.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/stdint.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/_types/_uint8_t.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/_types/_uint16_t.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/_types/_uint32_t.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/_types/_uint64_t.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/_types/_intmax_t.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/_types/_uintmax_t.h \
+  /Applications/Xcode.app/Contents/Developer/Toolchains/XcodeDefault.xctoolchain/usr/lib/clang/21/include/stdbool.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/math.h \
+  /Users/pancake/prg/radare2/libr/include/r_types_overflow.h \
+  /Users/pancake/prg/radare2/libr/include/r_util/r_str_util.h \
+  /Users/pancake/prg/radare2/libr/include/r_userconf.h \
+  /Users/pancake/prg/radare2/libr/include/r_version.h \
+  /Applications/Xcode.app/Contents/Developer/Toolchains/XcodeDefault.xctoolchain/usr/lib/clang/21/include/stddef.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/stddef.h \
+  /Applications/Xcode.app/Contents/Developer/Toolchains/XcodeDefault.xctoolchain/usr/lib/clang/21/include/__stddef_header_macro.h \
+  /Applications/Xcode.app/Contents/Developer/Toolchains/XcodeDefault.xctoolchain/usr/lib/clang/21/include/__stddef_ptrdiff_t.h \
+  /Applications/Xcode.app/Contents/Developer/Toolchains/XcodeDefault.xctoolchain/usr/lib/clang/21/include/__stddef_size_t.h \
+  /Applications/Xcode.app/Contents/Developer/Toolchains/XcodeDefault.xctoolchain/usr/lib/clang/21/include/__stddef_rsize_t.h \
+  /Applications/Xcode.app/Contents/Developer/Toolchains/XcodeDefault.xctoolchain/usr/lib/clang/21/include/__stddef_wchar_t.h \
+  /Applications/Xcode.app/Contents/Developer/Toolchains/XcodeDefault.xctoolchain/usr/lib/clang/21/include/__stddef_null.h \
+  /Applications/Xcode.app/Contents/Developer/Toolchains/XcodeDefault.xctoolchain/usr/lib/clang/21/include/__stddef_max_align_t.h \
+  /Applications/Xcode.app/Contents/Developer/Toolchains/XcodeDefault.xctoolchain/usr/lib/clang/21/include/__stddef_offsetof.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/stdlib.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/_stdlib.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/wait.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/signal.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/machine/signal.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/arm/signal.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/machine/_mcontext.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/arm/_mcontext.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/mach/machine/_structs.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/mach/arm/_structs.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_types/_sigaltstack.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_types/_ucontext.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_types/_sigset_t.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/resource.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_types/_timeval.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/alloca.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_types/_null.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/malloc/_malloc.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/malloc/_malloc_type.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/malloc/_ptrcheck.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/_abort.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/assert.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/_assert.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/_static_assert.h \
+  /Applications/Xcode.app/Contents/Developer/Toolchains/XcodeDefault.xctoolchain/usr/lib/clang/21/include/inttypes.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/inttypes.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/_inttypes.h \
+  /Users/pancake/prg/radare2/libr/include/r_types_null.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/stdio.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/_stdio.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_types/_va_list.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/stdio.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/_printf.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_types/_seek_set.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/_ctermid.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/secure/_stdio.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/secure/_common.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/string.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/_string.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/_strings.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/secure/_strings.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/secure/_string.h \
+  /Applications/Xcode.app/Contents/Developer/Toolchains/XcodeDefault.xctoolchain/usr/lib/clang/21/include/stdarg.h \
+  /Applications/Xcode.app/Contents/Developer/Toolchains/XcodeDefault.xctoolchain/usr/lib/clang/21/include/__stdarg_header_macro.h \
+  /Applications/Xcode.app/Contents/Developer/Toolchains/XcodeDefault.xctoolchain/usr/lib/clang/21/include/__stdarg___gnuc_va_list.h \
+  /Applications/Xcode.app/Contents/Developer/Toolchains/XcodeDefault.xctoolchain/usr/lib/clang/21/include/__stdarg_va_list.h \
+  /Applications/Xcode.app/Contents/Developer/Toolchains/XcodeDefault.xctoolchain/usr/lib/clang/21/include/__stdarg_va_arg.h \
+  /Applications/Xcode.app/Contents/Developer/Toolchains/XcodeDefault.xctoolchain/usr/lib/clang/21/include/__stdarg___va_copy.h \
+  /Applications/Xcode.app/Contents/Developer/Toolchains/XcodeDefault.xctoolchain/usr/lib/clang/21/include/__stdarg_va_copy.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/fcntl.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/fcntl.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_types/_o_sync.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_types/_o_dsync.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_types/_s_ifmt.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_types/_timespec.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_types/_filesec_t.h \
+  /Users/pancake/prg/radare2/libr/include/r_endian.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/stat.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/dirent.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/dirent.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/unistd.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/unistd.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_types/_posix_vdisable.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/select.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_select.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_types/_uuid_t.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/gethostuuid.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/time.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_types/_timeval64.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/time.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/_time.h \
+  /Users/pancake/prg/radare2/libr/include/r_getopt.h \
+  /Users/pancake/prg/radare2/libr/include/r_util.h \
+  /Users/pancake/prg/radare2/libr/include/r_regex.h \
+  /Users/pancake/prg/radare2/libr/include/r_list.h \
+  /Users/pancake/prg/radare2/libr/include/sdb/sdb.h \
+  /Users/pancake/prg/radare2/libr/include/sdb/asserts.h \
+  /Users/pancake/prg/radare2/libr/include/sdb/types.h \
+  /Users/pancake/prg/radare2/libr/include/sdb/config.h \
+  /Users/pancake/prg/radare2/libr/include/sdb/ht.h \
+  /Users/pancake/prg/radare2/libr/include/sdb/ht_pp.h \
+  /Users/pancake/prg/radare2/libr/include/sdb/ht_inc.h \
+  /Users/pancake/prg/radare2/libr/include/sdb/ls.h \
+  /Users/pancake/prg/radare2/libr/include/sdb/heap.h \
+  /Users/pancake/prg/radare2/libr/include/sdb/dict.h \
+  /Users/pancake/prg/radare2/libr/include/sdb/cdb.h \
+  /Users/pancake/prg/radare2/libr/include/sdb/cdb_make.h \
+  /Users/pancake/prg/radare2/libr/include/sdb/buffer.h \
+  /Users/pancake/prg/radare2/libr/include/sdb/version.h \
+  /Users/pancake/prg/radare2/libr/include/sdb/rangstr.h \
+  /Users/pancake/prg/radare2/libr/include/r_skiplist.h \
+  /Users/pancake/prg/radare2/libr/include/r_th.h \
+  /Users/pancake/prg/radare2/libr/include/r_util/r_w32.h \
+  /Users/pancake/prg/radare2/libr/include/r_util/r_sys.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/errno.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/errno.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/semaphore.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/semaphore.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/pthread.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/pthread/sched.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/pthread/pthread_impl.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/pthread/qos.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/qos.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_types/_mach_port_t.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/signal.h \
+  /Users/pancake/prg/radare2/libr/include/r_event.h \
+  /Users/pancake/prg/radare2/libr/include/sdb/ht_up.h \
+  /Users/pancake/prg/radare2/libr/include/r_vec.h \
+  /Users/pancake/prg/radare2/libr/include/r_util/r_assert.h \
+  /Users/pancake/prg/radare2/libr/include/r_util/r_log.h \
+  /Users/pancake/prg/radare2/libr/include/r_util/r_arena.h \
+  /Users/pancake/prg/radare2/libr/include/r_util/r_itv.h \
+  /Users/pancake/prg/radare2/libr/include/r_util/r_signal.h \
+  /Users/pancake/prg/radare2/libr/include/r_util/r_alloc.h \
+  /Users/pancake/prg/radare2/libr/include/r_util/r_rbtree.h \
+  /Users/pancake/prg/radare2/libr/include/r_util/r_trie.h \
+  /Users/pancake/prg/radare2/libr/include/r_util/r_strs.h \
+  /Users/pancake/prg/radare2/libr/include/r_util/r_xml.h \
+  /Users/pancake/prg/radare2/libr/include/r_util/r_new_rbtree.h \
+  /Users/pancake/prg/radare2/libr/include/r_util/r_intervaltree.h \
+  /Users/pancake/prg/radare2/libr/include/r_util/../r_types.h \
+  /Users/pancake/prg/radare2/libr/include/r_util/r_big.h \
+  /Users/pancake/prg/radare2/libr/include/r_util/r_base32.h \
+  /Users/pancake/prg/radare2/libr/include/r_util/r_base36.h \
+  /Users/pancake/prg/radare2/libr/include/r_util/r_base64.h \
+  /Users/pancake/prg/radare2/libr/include/r_util/r_base91.h \
+  /Users/pancake/prg/radare2/libr/include/r_util/r_buf.h \
+  /Users/pancake/prg/radare2/libr/include/r_util/r_mem.h \
+  /Users/pancake/prg/radare2/libr/include/r_util/r_ref.h \
+  /Users/pancake/prg/radare2/libr/include/r_util/r_bits.h \
+  /Users/pancake/prg/radare2/libr/include/r_util/r_bitmap.h \
+  /Users/pancake/prg/radare2/libr/include/r_util/r_bitset.h \
+  /Users/pancake/prg/radare2/libr/include/r_util/r_time.h \
+  /Users/pancake/prg/radare2/libr/include/r_util/r_debruijn.h \
+  /Users/pancake/prg/radare2/libr/include/r_util/r_type.h \
+  /Users/pancake/prg/radare2/libr/include/r_util/r_file.h \
+  /Users/pancake/prg/radare2/libr/include/r_util/r_hex.h \
+  /Users/pancake/prg/radare2/libr/include/r_util/r_name.h \
+  /Users/pancake/prg/radare2/libr/include/r_util/r_num.h \
+  /Users/pancake/prg/radare2/libr/include/r_util/r_table.h \
+  /Users/pancake/prg/radare2/libr/include/r_util/r_graph.h \
+  /Users/pancake/prg/radare2/libr/include/r_util/r_panels.h \
+  /Users/pancake/prg/radare2/libr/include/r_util/r_pool.h \
+  /Users/pancake/prg/radare2/libr/include/r_util/r_diff.h \
+  /Users/pancake/prg/radare2/libr/include/r_cons.h \
+  /Users/pancake/prg/radare2/libr/include/r_util/pj.h \
+  /Users/pancake/prg/radare2/libr/include/r_util/r_strbuf.h \
+  /Users/pancake/prg/radare2/libr/include/r_util/r_json.h \
+  /Users/pancake/prg/radare2/libr/include/r_util/r_sandbox.h \
+  /Users/pancake/prg/radare2/libr/include/r_util/r_stack.h \
+  /Users/pancake/prg/radare2/libr/include/r_util/r_str.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/wchar.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/_wchar.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_types/_mbstate_t.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/__wctype.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/___wctype.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/_types/_wctype_t.h \
+  /Users/pancake/prg/radare2/libr/include/r_util/r_str_constpool.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/termios.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/termios.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/ttycom.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/ioccom.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/ttydefaults.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/ioctl.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/filio.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/sockio.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/net/if.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/net/if_var.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_types/_timeval32.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/net/net_kev.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/socket.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/constrained_ctypes.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/machine/_param.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/arm/_param.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_types/_sa_family_t.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_types/_socklen_t.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/_types/_iovec_t.h \
+  /Users/pancake/prg/radare2/libr/include/r_util/r_punycode.h \
+  /Users/pancake/prg/radare2/libr/include/r_util/r_queue.h \
+  /Users/pancake/prg/radare2/libr/include/r_util/r_range.h \
+  /Users/pancake/prg/radare2/libr/include/r_util/r_spaces.h \
+  /Users/pancake/prg/radare2/libr/include/r_util/r_strs_store.h \
+  /Users/pancake/prg/radare2/libr/include/r_util/r_bloom.h \
+  /Users/pancake/prg/radare2/libr/include/r_util/r_strpool.h \
+  /Users/pancake/prg/radare2/libr/include/r_util/r_tree.h \
+  /Users/pancake/prg/radare2/libr/include/r_util/r_uleb128.h \
+  /Users/pancake/prg/radare2/libr/include/r_util/r_utf8.h \
+  /Users/pancake/prg/radare2/libr/include/r_util/r_utf16.h \
+  /Users/pancake/prg/radare2/libr/include/r_util/r_utf32.h \
+  /Users/pancake/prg/radare2/libr/include/r_util/r_idpool.h \
+  /Users/pancake/prg/radare2/libr/include/r_util/r_asn1.h \
+  /Users/pancake/prg/radare2/libr/include/r_util/r_graph_drawable.h \
+  /Users/pancake/prg/radare2/libr/include/r_config.h \
+  /Users/pancake/prg/radare2/libr/include/r_util/bplist.h \
+  /Users/pancake/prg/radare2/libr/include/r_util/r_x509.h \
+  /Users/pancake/prg/radare2/libr/include/r_util/r_pkcs7.h \
+  /Users/pancake/prg/radare2/libr/include/r_util/r_protobuf.h \
+  /Users/pancake/prg/radare2/libr/include/r_util/r_token.h \
+  /Users/pancake/prg/radare2/libr/include/r_util/r_axml.h \
+  /Users/pancake/prg/radare2/libr/include/r_util/r_cfloat.h \
+  /Users/pancake/prg/radare2/libr/include/r_util/libc.h \
+  /Users/pancake/prg/radare2/libr/include/r_socket.h \
+  /Users/pancake/prg/radare2/libr/include/r_bind.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/netinet/in.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/netinet6/in6.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/un.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/poll.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/poll.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/arpa/inet.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/netdb.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/netinet/tcp.h \
+  /Users/pancake/prg/radare2/libr/include/r_magic.h \
+  /Users/pancake/prg/radare2/libr/include/r_agraph.h \
+  /Users/pancake/prg/radare2/libr/include/r_io.h \
+  /Users/pancake/prg/radare2/libr/include/r_lib.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/dlfcn.h \
+  /Users/pancake/prg/radare2/libr/include/r_util/r_w32dw.h \
+  /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/sys/ptrace.h \
+  /Users/pancake/prg/radare2/libr/include/r_fs.h \
+  /Users/pancake/prg/radare2/libr/include/r_egg.h \
+  /Users/pancake/prg/radare2/libr/include/r_asm.h \
+  /Users/pancake/prg/radare2/libr/include/r_arch.h \
+  /Users/pancake/prg/radare2/libr/include/r_bin.h \
+  /Users/pancake/prg/radare2/libr/include/r_muta.h \
+  /Users/pancake/prg/radare2/libr/include/r_hash.h \
+  /Users/pancake/prg/radare2/libr/include/r_muta/r_sm4.h \
+  /Users/pancake/prg/radare2/libr/include/r_bin_dwarf.h \
+  /Users/pancake/prg/radare2/libr/include/r_reg.h \
+  /Users/pancake/prg/radare2/libr/include/r_anal/op.h \
+  /Users/pancake/prg/radare2/libr/include/r_anal/hint.h \
+  /Users/pancake/prg/radare2/libr/include/r_esil.h \
+  /Users/pancake/prg/radare2/libr/include/sdb/ht_uu.h \
+  /Users/pancake/prg/radare2/libr/include/r_anal.h \
+  /Users/pancake/prg/radare2/libr/include/r_util/r_print.h \
+  /Users/pancake/prg/radare2/libr/include/r_search.h \
+  /Users/pancake/prg/radare2/libr/include/r_syscall.h \
+  /Users/pancake/prg/radare2/libr/include/r_flag.h \
+  /Users/pancake/prg/radare2/libr/include/r_codemeta.h \
+  /Users/pancake/prg/radare2/libr/include/sdb/set.h \
+  /Users/pancake/prg/radare2/libr/include/r_lang.h \
+  /Users/pancake/prg/radare2/libr/include/r_cmd.h \
+  /Users/pancake/prg/radare2/libr/include/r_sign.h \
+  /Users/pancake/prg/radare2/libr/include/r_debug.h \
+  /Users/pancake/prg/radare2/libr/include/r_bp.h \
+  /Users/pancake/prg/radare2/libr/include/rvc.h \
+  /Users/pancake/prg/radare2/libr/include/r_core_priv.h \
+  /Users/pancake/prg/radare2/libr/include/r_core_task.h

decai/r2plugin/deps.mk

@@ -0,0 +1 @@
+include p/decai/r2plugin/core/deps.mk

decai/r2plugin/js2c.py

@@ -0,0 +1,11 @@
+#!/usr/bin/env python3
+"""Convert a script file into a C buffer named decai_qjs (same output as `rax2 -C` + sed)."""
+
+import sys
+
+data = open(sys.argv[1], "rb").read()
+with open(sys.argv[2], "w") as f:
+    f.write("const unsigned char decai_qjs[] = {\n")
+    for i in range(0, len(data), 16):
+        f.write("  %s,\n" % ", ".join("0x%02x" % b for b in data[i:i + 16]))
+    f.write("};\nconst unsigned int decai_qjs_len = %d;\n" % len(data))

decai/r2plugin/meson.build

@@ -0,0 +1,16 @@
+# decai XPS integration for radare2.
+# Included from libr/xps/meson.build when decai lives under libr/xps/p/decai/.
+# Unlike native plugins, decai ships no native code: decai.r2.js is embedded
+# as a buffer in libr_core and the tiny core plugin registers it with
+# r_core_script_embed() so r2 runs it at startup like any plugins/*.r2.js file.
+
+core_plugins += ['decai']
+
+decai_py3 = import('python').find_installation()
+
+r_core_additional_sources += files('core/plugin.c')
+r_core_additional_sources += custom_target('decai_r2_js_c',
+  input: '../decai.r2.js',
+  output: 'decai.r2.js.c',
+  command: [decai_py3, files('js2c.py'), '@INPUT@', '@OUTPUT@'],
+)

decai/r2plugin/static.cfg

@@ -0,0 +1 @@
+core.decai

decai/src/Makefile

@@ -1,10 +1,11 @@
 PLUGIN_NAME=decai
 R2F_C_FLAGS=-c -S
+R2FRIDA_COMPILE?=r2pm -r r2frida-compile
 
 all: build
 
 build:
-	r2pm -r r2frida-compile $(R2F_C_FLAGS) -o ../$(PLUGIN_NAME).r2.js main.ts
+	$(R2FRIDA_COMPILE) $(R2F_C_FLAGS) -o ../$(PLUGIN_NAME).r2.js main.ts
 
 test: build
 	r2 -i ../$(PLUGIN_NAME).r2.js -q -c 'af;decai -d' /bin/ls

decai/src/apiKeys.ts

@@ -1,22 +1,15 @@
 import { API_KEYS_PATH, DECAI_CONFIG_DIR } from "./constants";
+import { providerRegistry } from "./providers";
 import { ApiKeyResult } from "./types";
 import { ensurePath, fileExists, parseEnvLikeString } from "./utils";
 
-const PROVIDER_ENV_MAP: Record<string, string> = {
-  mistral: "MISTRAL_API_KEY",
-  anthropic: "ANTHROPIC_API_KEY",
-  huggingface: "HUGGINGFACE_API_KEY",
-  openai: "OPENAI_API_KEY",
-  gemini: "GEMINI_API_KEY",
-  deepseek: "DEEPSEEK_API_KEY",
-  xai: "XAI_API_KEY",
-  ollama: "OLLAMA_API_KEY",
-  ollamacloud: "OLLAMA_API_KEY",
-  opencode: "OPENCODE_API_KEY",
-  zen: "OPENCODE_API_KEY",
-  openrouter: "OPENROUTER_API_KEY",
-  groq: "GROQ_API_KEY",
-};
+function getProviderEnvMap(): Record<string, string> {
+  return Object.fromEntries(
+    Object.entries(providerRegistry)
+      .filter(([, v]) => v.authKey)
+      .map(([k, v]) => [k, v.authKey!]),
+  );
+}
 
 export function getApiKey(provider: string, envvar: string): ApiKeyResult {
   const keyEnv = r2.cmd("'%" + envvar).trim();
@@ -52,12 +45,9 @@ export function editApiKeys(): void {
 }
 
 export function listApiKeys(): void {
-  Object.entries(PROVIDER_ENV_MAP).forEach(([key, env]) => {
+  Object.entries(getProviderEnvMap()).forEach(([key, env]) => {
     const status = getApiKey(key, env)[2];
     console.log(status, "\t", key);
   });
 }
 
-export function getProviderEnvVar(provider: string): string {
-  return PROVIDER_ENV_MAP[provider.toLowerCase()] || "";
-}

decai/src/commands.ts

@@ -3,7 +3,6 @@ import { state } from "./state";
 import { COMMAND, DEFAULT_PROMPT, HELP_TEXT, VERSION } from "./constants";
 import {
   b64,
-  debugLog,
   filterResponse,
   trimAnsi,
   trimDown,
@@ -13,6 +12,16 @@ import { editApiKeys, listApiKeys } from "./apiKeys";
 import { editRcConfig, evalConfig, listAllConfig } from "./config";
 import { r2ai } from "./r2ai";
 
+function withColorDisabled<T>(fn: () => T): T {
+  const origColor = r2.cmd("e scr.color");
+  r2.cmd("e scr.color=0");
+  try {
+    return fn();
+  } finally {
+    r2.cmd("e scr.color=" + origColor);
+  }
+}
+
 export function showHelp(): void {
   let helpmsg = "";
   const msg = (m: string) => (helpmsg += " " + COMMAND + " " + m + "\n");
@@ -26,7 +35,7 @@ export function showHelp(): void {
   msg("-dD [query]   - decompile current function with given extra query");
   msg("-dr           - decompile function and its called ones (recursive)");
   msg("-e            - display and change eval config vars");
-  msg("-E            - edit decai.txt");
+  msg("-E            - edit and run decai.txt");
   msg("-h            - show this help");
   msg("-H            - help setting up r2ai");
   msg("-i [f] [q]    - include given file and query");
@@ -72,11 +81,9 @@ export function decompile(
   }
 
   const appendQuery = extraQuery ? " " + args : "";
-  const origColor = r2.cmd("e scr.color");
 
   try {
     const parsedArgs = args.slice(2).trim();
-    let count = 0;
     let text = "";
 
     if (
@@ -88,25 +95,23 @@ export function decompile(
         "\n[END]\n";
     }
 
-    r2.cmd("e scr.color=0");
-    let body = "## Before:\n";
-
-    for (const c of state.commands.split(",")) {
-      if (c.trim() === "") continue;
-
-      const oneliner = extraQuery || parsedArgs.trim().length === 0
-        ? c
-        : c + "@@= " + parsedArgs;
-      const output = r2.cmd(oneliner);
-
-      if (output.length > 5) {
-        body += "Output of " + c + ":\n[START]\n" + output + "\n[END]\n";
-        count++;
+    const { body, count } = withColorDisabled(() => {
+      let body = "## Before:\n";
+      let count = 0;
+      for (const c of state.commands.split(",")) {
+        if (c.trim() === "") continue;
+        const oneliner = extraQuery || parsedArgs.trim().length === 0
+          ? c
+          : c + "@@= " + parsedArgs;
+        const output = r2.cmd(oneliner);
+        if (output.length > 5) {
+          body += "Output of " + c + ":\n[START]\n" + output + "\n[END]\n";
+          count++;
+        }
       }
-    }
-
-    body += "## After:\n";
-    r2.cmd("e scr.color=" + origColor);
+      body += "## After:\n";
+      return { body, count };
+    });
 
     if (count === 0) {
       console.error("Nothing to do.");
@@ -115,24 +120,27 @@ export function decompile(
 
     let out = "";
     if (state.decopipe.use) {
-      const dpipe = state.decopipe[state.decopipe.default as string] as {
-        pipeline: Array<{ model?: string; query: string }>;
-        globalQuery: string;
-      };
       const origModel = state.model;
-      let code = text + body;
+      try {
+        const dpipe = state.decopipe[state.decopipe.default as string] as {
+          pipeline: Array<{ model?: string; query: string }>;
+          globalQuery: string;
+        };
+        let code = text + body;
 
-      for (const dp of dpipe.pipeline) {
-        if (dp.model) state.model = dp.model;
-        const query = dp.query + ". " + dpipe.globalQuery;
-        out = r2ai(query, code, true);
-        if (state.debug) {
-          console.log("QUERY\n", query, "\nINPUT\n", code, "\nOUTPUT\n", out);
+        for (const dp of dpipe.pipeline) {
+          if (dp.model) state.model = dp.model;
+          const query = dp.query + ". " + dpipe.globalQuery;
+          out = r2ai(query, code, true);
+          if (state.debug) {
+            console.log("QUERY\n", query, "\nINPUT\n", code, "\nOUTPUT\n", out);
+          }
+          code = out;
         }
-        code = out;
+        out = code;
+      } finally {
+        state.model = origModel;
       }
-      out = code;
-      state.model = origModel;
     } else {
       const query = appendQuery;
       text += body + context;
@@ -150,7 +158,6 @@ export function decompile(
 
     return out.trim();
   } catch (e) {
-    r2.cmd("e scr.color=" + origColor);
     const err = e as Error;
     console.error(err, err.stack);
     return;
@@ -158,11 +165,10 @@ export function decompile(
 }
 
 export function explainFunction(): string {
-  const origColor = r2.cmd("e scr.color");
-  r2.cmd("e scr.color=0");
-  const hints = "[START]" +
-    state.commands.split(",").map((c) => r2.cmd(c)).join("\n") + "[END]";
-  r2.cmd("e scr.color=" + origColor);
+  const hints = withColorDisabled(() =>
+    "[START]" +
+    state.commands.split(",").map((c) => r2.cmd(c)).join("\n") + "[END]"
+  );
 
   const res = r2ai(
     "Analyze function calls, references, comments and strings, loops and ignore registers and memory accesses. Explain the purpose of this function in a single short sentence. /no_think Do not introduce or argue the response, translation of the explanation in " +
@@ -285,6 +291,7 @@ export function autoMode(
         break;
       } else {
         console.log("Unknown response\n", JSON.stringify(out));
+        break;
       }
     } catch (e) {
       const response = out.indexOf('response": "');
@@ -358,6 +365,12 @@ function getFlagValue(args: string): string {
   return args.slice(2).trim();
 }
 
+function currentFunctionExists(): boolean {
+  return r2.cmd("afi.").trim().length > 0;
+}
+
+const noFunctionMessage = "Cannot find function at current offset";
+
 export function handleCommand(
   args: string,
   mainHandler: (args: string) => boolean,
@@ -393,6 +406,10 @@ export function handleCommand(
 
     case "n":
     case "f": {
+      if (!currentFunctionExists()) {
+        output = noFunctionMessage;
+        break;
+      }
       output = r2.cmd("axff~$[3]");
       const considerations = r2
         .cmd("fd.")
@@ -411,7 +428,11 @@ export function handleCommand(
       break;
     }
 
-    case "v":
+    case "v": {
+      if (!currentFunctionExists()) {
+        output = noFunctionMessage;
+        break;
+      }
       output = r2.cmd("afv;pdc");
       output = r2ai(
         "guess a better name and type for each local variable and function argument taking using. output an r2 script using afvn and afvt commands",
@@ -419,6 +440,7 @@ export function handleCommand(
         false,
       );
       break;
+    }
 
     case "i": {
       const parts = argValue.split(/\s+/).filter(Boolean);
@@ -456,6 +478,10 @@ export function handleCommand(
       break;
 
     case "s":
+      if (!currentFunctionExists()) {
+        output = noFunctionMessage;
+        break;
+      }
       output = suggestSignature();
       break;
 
@@ -511,6 +537,10 @@ export function handleCommand(
       break;
 
     case "x":
+      if (!currentFunctionExists()) {
+        output = noFunctionMessage;
+        break;
+      }
       output = explainFunction();
       if (args[2] === "*" || args[2] === "r") {
         output = "'CC " + output;

decai/src/config.ts

@@ -1,4 +1,4 @@
-import { ConfigHandlers } from "./types";
+import { ConfigHandler, ConfigHandlers } from "./types";
 import { DECAI_CONFIG_DIR, DECAI_CONFIG_PATH } from "./constants";
 import { defaultState, state } from "./state";
 import { listModelsFor } from "./models";
@@ -12,6 +12,37 @@ function parseBoolean(value: string): boolean {
   return value === "true" || value === "1";
 }
 
+type StateKey = keyof typeof state;
+
+function boolHandler(key: StateKey): ConfigHandler {
+  // deno-lint-ignore no-explicit-any
+  const s = state as any;
+  return {
+    get: () => s[key] as boolean,
+    set: (v: string) => { s[key] = parseBoolean(v); },
+  };
+}
+
+function stringHandler(key: StateKey): ConfigHandler {
+  // deno-lint-ignore no-explicit-any
+  const s = state as any;
+  return {
+    get: () => s[key] as string,
+    set: (v: string) => { s[key] = v; },
+  };
+}
+
+function normalizeApiType(value: string): "chat" | "generate" | null {
+  const apiType = value.trim();
+  if (apiType === "chat") {
+    return "chat";
+  }
+  if (apiType === "generate") {
+    return "generate";
+  }
+  return null;
+}
+
 export const configHandlers: ConfigHandlers = {
   pipeline: {
     get: () => state.pipeline,
@@ -34,112 +65,65 @@ export const configHandlers: ConfigHandlers = {
       }
     },
   },
-  deterministic: {
-    get: () => state.deterministic,
-    set: (v: string) => {
-      state.deterministic = parseBoolean(v);
-    },
-  },
-  files: {
-    get: () => state.useFiles,
-    set: (v: string) => {
-      state.useFiles = parseBoolean(v);
-    },
-  },
+  deterministic: boolHandler("deterministic"),
+  files: boolHandler("useFiles"),
   think: {
-    get: () => state.think,
+    get: () => state.think || "false",
+    set: (v: string) => { state.think = v; },
+  },
+  apitype: {
+    get: () => state.apitype,
     set: (v: string) => {
-      state.think = v === "true" ? 1 : v === "false" ? 0 : +v;
+      const apiType = normalizeApiType(v);
+      if (apiType) {
+        state.apitype = apiType;
+      } else {
+        console.error("Invalid apitype. Use chat or generate.");
+      }
     },
   },
-  debug: {
-    get: () => state.debug,
+  debug: boolHandler("debug"),
+  http: {
+    get: () => state.http,
     set: (v: string) => {
-      state.debug = parseBoolean(v);
+      const backend = v.trim();
+      if (backend === "auto" || backend === "r2" || backend === "curl") {
+        state.http = backend;
+      } else {
+        console.error("Invalid http backend. Use auto, r2 or curl.");
+      }
     },
   },
   timeout: {
     get: () => state.timeout,
-    set: (v: string) => {
-      state.timeout = Math.max(0, parseInt(v, 10) || 0);
-    },
+    set: (v: string) => { state.timeout = Math.max(0, parseInt(v, 10) || 0); },
   },
   api: {
     get: () => state.api,
     set: (v: string) => {
       if (v === "?") {
-        const providersList = listProviders().join("\n");
-        console.error(providersList);
+        console.error(listProviders().join("\n"));
       } else {
         state.api = v;
       }
     },
   },
-  lang: {
-    get: () => state.language,
-    set: (v: string) => {
-      state.language = v;
-    },
-  },
-  hlang: {
-    get: () => state.humanLanguage,
-    set: (v: string) => {
-      state.humanLanguage = v;
-    },
-  },
-  cache: {
-    get: () => state.cache,
-    set: (v: string) => {
-      state.cache = parseBoolean(v);
-    },
-  },
-  cmds: {
-    get: () => state.commands,
-    set: (v: string) => {
-      state.commands = v;
-    },
-  },
-  tts: {
-    get: () => state.tts,
-    set: (v: string) => {
-      state.tts = parseBoolean(v);
-    },
-  },
-  yolo: {
-    get: () => state.yolo,
-    set: (v: string) => {
-      state.yolo = parseBoolean(v);
-    },
-  },
-  prompt: {
-    get: () => state.prompt,
-    set: (v: string) => {
-      state.prompt = v;
-    },
-  },
-  ctxfile: {
-    get: () => state.contextFile,
-    set: (v: string) => {
-      state.contextFile = v;
-    },
-  },
-  baseurl: {
-    get: () => state.baseurl,
-    set: (v: string) => {
-      state.baseurl = v;
-    },
-  },
+  lang: stringHandler("language"),
+  hlang: stringHandler("humanLanguage"),
+  cache: boolHandler("cache"),
+  cmds: stringHandler("commands"),
+  tts: boolHandler("tts"),
+  yolo: boolHandler("yolo"),
+  prompt: stringHandler("prompt"),
+  ctxfile: stringHandler("contextFile"),
+  baseurl: stringHandler("baseurl"),
   headers: {
     get: () => formatHeaders(state.extraHeaders),
-    set: (v: string) => {
-      state.extraHeaders = parseHeaders(v);
-    },
+    set: (v: string) => { state.extraHeaders = parseHeaders(v); },
   },
   maxtokens: {
     get: () => state.maxInputTokens,
-    set: (v: string) => {
-      state.maxInputTokens = parseInt(v, 10) || -1;
-    },
+    set: (v: string) => { state.maxInputTokens = parseInt(v, 10) || -1; },
   },
 };
 
@@ -190,10 +174,10 @@ export function loadRcConfig(): void {
   if (!fileExists(DECAI_CONFIG_PATH)) {
     return;
   }
-  const rcFile = r2.cmd("'cat " + DECAI_CONFIG_PATH);
+  const rcFile = r2.call("cat " + DECAI_CONFIG_PATH);
   for (const line of rcFile.split(/\r?\n/)) {
     const rcLine = normalizeRcLine(line);
-    if (rcLine) {
+    if (rcLine && rcLine[0] != '#') {
       evalConfig(rcLine);
     }
   }

decai/src/constants.ts

@@ -1,4 +1,4 @@
-export const VERSION = "1.3.4";
+export const VERSION = "1.5.4";
 export const COMMAND = "decai";
 export const DECAI_CONFIG_DIR = "~/.config/r2ai";
 export const API_KEYS_PATH = DECAI_CONFIG_DIR + "/apikeys.txt";
@@ -15,7 +15,7 @@ Decai is the radare2 plugin for decompiling functions with the help of language
 By default uses a local ollama server, but can you can pick any other service by using 'decai -e api=?'.
 
 [0x00000000]> decai -e api=?
-r2ai deepseek anthropic claude gemini groq hf lmstudio mistral ollama ollamacloud openapi opencode openai openrouter vllm xai zen
+r2ai deepseek anthropic claude gemini groq hf lmstudio mistral ollama ollamacloud openapi openai openrouter vllm xai zen
 
 ## Using Ollama
 
@@ -30,8 +30,10 @@ These are the most recommended models for decompiling in local:
 
 ## Common Options
 * 'decai -e baseurl=<url>' override default host and port for API endpoint (e.g., 'http://localhost:11434')
+* 'decai -e apitype=chat|generate' choose the Ollama API endpoint (default: chat)
 * 'decai -e headers=Authorization: Bearer ...\\nUser-Agent: curl/8.7.1' add or override HTTP headers
 * "export DECAI_HEADERS='Authorization: Bearer ...\\nUser-Agent: curl/8.7.1'" set extra headers from the environment
+* 'decai -e http=auto|r2|curl' select the HTTP transport (default: auto, prefers r2.http)
 * 'decai -e timeout=0' disable curl timeouts, or set a positive number of seconds
 
 * 'decai -e deterministic=true' to remove randomness from decompilation responses

decai/src/headers.ts

@@ -0,0 +1,91 @@
+import { state } from "./state";
+
+const HEADER_ENV_VARS = ["DECAI_HEADERS", "R2AI_HEADERS"];
+
+interface ParsedHeader {
+  name: string;
+  value: string;
+}
+
+function readEnvVar(name: string): string {
+  const value = r2.cmd("'%" + name).trim();
+  return value !== "" && !value.includes("=") ? value : "";
+}
+
+function normalizeHeaderName(name: string): string {
+  return name.trim().toLowerCase();
+}
+
+function parseHeaderLine(line: string): ParsedHeader | null {
+  const colon = line.indexOf(":");
+  const equals = line.indexOf("=");
+  const sep = colon !== -1 && (equals === -1 || colon < equals)
+    ? colon
+    : equals;
+
+  if (sep === -1) {
+    return null;
+  }
+
+  const name = line.slice(0, sep).trim();
+  if (name === "") {
+    return null;
+  }
+
+  return {
+    name,
+    value: line.slice(sep + 1).trim(),
+  };
+}
+
+function formatHeader(header: ParsedHeader): string {
+  return header.value === ""
+    ? `${header.name}:`
+    : `${header.name}: ${header.value}`;
+}
+
+function collectHeaders(
+  headerLines: string[],
+  into: Map<string, string>,
+): void {
+  for (const rawLine of headerLines) {
+    const header = parseHeaderLine(rawLine.trim());
+    if (header) {
+      into.set(normalizeHeaderName(header.name), formatHeader(header));
+    }
+  }
+}
+
+export function parseHeaders(value: string): string[] {
+  const headers = new Map<string, string>();
+  const lines = value.replace(/\\n/g, "\n").split(/\r?\n/g)
+    .filter((l) => { const t = l.trim(); return t && !t.startsWith("#"); });
+  collectHeaders(lines, headers);
+  return Array.from(headers.values());
+}
+
+export function formatHeaders(headers: string[]): string {
+  return headers.join("\\n");
+}
+
+export function mergeHeaders(...groups: string[][]): string[] {
+  const merged = new Map<string, string>();
+  for (const group of groups) {
+    collectHeaders(group, merged);
+  }
+  return Array.from(merged.values());
+}
+
+function readConfiguredEnvHeaders(): string[] {
+  for (const name of HEADER_ENV_VARS) {
+    const value = readEnvVar(name);
+    if (value) {
+      return parseHeaders(value);
+    }
+  }
+  return [];
+}
+
+export function getConfiguredHeaders(): string[] {
+  return mergeHeaders(readConfiguredEnvHeaders(), state.extraHeaders);
+}

decai/src/http.ts

@@ -14,24 +14,31 @@ function shellEscape(arg: string): string {
   return `'${arg.replace(/'/g, `'"'"'`)}'`;
 }
 
-function executeCurl(options: HttpRequestOptions): string {
+function executeRequest(options: HttpRequestOptions): string {
   const { method, url, headers, payload } = options;
-  const cmdParts = ["curl", "-s"];
-  if (state.timeout > 0) {
-    cmdParts.push("--max-time", String(state.timeout));
-  }
   const requestHeaders = mergeHeaders(
     ["Content-Type: application/json"],
     headers,
   );
+  if (state.http !== "curl") {
+    if (r2.http) {
+      return r2.http(method, url, requestHeaders.join("\n"), payload ?? "").body;
+    }
+    if (state.http === "r2") {
+      throw new Error("r2.http is not available in this runtime");
+    }
+  }
+  const cmdParts = ["curl", "-s"];
+  if (state.timeout > 0) {
+    cmdParts.push("--max-time", String(state.timeout));
+  }
   requestHeaders.forEach((h) => cmdParts.push("-H", shellEscape(h)));
 
   if (method === "POST") {
     if (!payload) throw new Error("Payload required for POST requests");
     const tmpfile = r2.fdump(payload);
-    cmdParts.push("--data-binary", "@-", shellEscape(url));
-    const cmd = cmdParts.join(" ") + " < " + shellEscape(tmpfile) +
-      " && rm " + shellEscape(tmpfile);
+    cmdParts.push("--data-binary", "@" + shellEscape(tmpfile), shellEscape(url));
+    const cmd = cmdParts.join(" ") + "; rm " + shellEscape(tmpfile);
     debugLog(cmd);
     return r2.syscmds(cmd);
   } else {
@@ -42,15 +49,44 @@ function executeCurl(options: HttpRequestOptions): string {
   }
 }
 
+function parseNdjson(output: string): HttpResponse | null {
+  const lines = output.split("\n").filter((l) => l.trim() !== "");
+  if (lines.length <= 1) return null;
+
+  let content = "";
+  let lastObj: Record<string, unknown> | null = null;
+
+  for (const line of lines) {
+    try {
+      const obj = JSON.parse(line);
+      if (obj.error) return obj as HttpResponse;
+      if (obj.message?.content !== undefined) {
+        content += obj.message.content;
+      }
+      lastObj = obj;
+    } catch {
+      return null;
+    }
+  }
+
+  if (lastObj) {
+    lastObj.message = { role: "assistant", content };
+    return lastObj as HttpResponse;
+  }
+  return null;
+}
+
 export function httpRequest(options: HttpRequestOptions): HttpResponse {
   try {
-    const output = executeCurl(options).trim();
+    const output = executeRequest(options).trim();
     if (output === "") {
       return { error: "empty response" };
     }
     try {
       return JSON.parse(output) as HttpResponse;
     } catch {
+      const ndjson = parseNdjson(output);
+      if (ndjson) return ndjson;
       return { error: output, rawOutput: output };
     }
   } catch (e) {

decai/src/providers.ts

@@ -1,22 +1,24 @@
 import {
+  AnthropicPayload,
   ApiKeyResult,
   ApiStyle,
   AuthStyle,
+  GeminiPayload,
   HttpResponse,
-  JsonObject,
   ModelDataEntry,
   OllamaModelEntry,
+  OllamaPayload,
   ProviderConfig,
   ProviderRegistry,
 } from "./types";
 import { state } from "./state";
-import { filterResponse, padRight } from "./utils";
+import { filterResponse, joinUrl, padRight } from "./utils";
 import { getApiKey } from "./apiKeys";
 import { getConfiguredHeaders, mergeHeaders } from "./headers";
 import { httpGet, httpPost } from "./http";
 
 interface ProviderRuntime {
-  buildPayload: (model: string, query: string) => JsonObject;
+  buildPayload: (model: string, query: string) => object;
   parseResponse: (response: HttpResponse) => string;
   buildUrl: (baseUrl: string, model: string, apiKey?: string) => string;
   requiresUrlApiKey?: boolean;
@@ -55,7 +57,10 @@ function getErrorMessage(error: HttpResponse["error"]): string | undefined {
   if (!error) {
     return undefined;
   }
-  return typeof error === "string" ? error : error.message;
+  if (typeof error === "string") {
+    return error;
+  }
+  return error.message || JSON.stringify(error);
 }
 
 function getProviderBaseUrl(provider: ProviderConfig): string {
@@ -120,39 +125,26 @@ function uniqueBy<T>(items: T[], getKey: (item: T) => string): T[] {
   });
 }
 
-function listOpenAIModels(provider: ProviderConfig): string {
-  const apiKey = readProviderKey(provider);
-  const baseUrl = getProviderBaseUrl(provider);
-  const headers = getProviderHeaders(
-    buildAuthHeaders(provider, apiKey?.[0] || null),
-  );
-  return listDataModels(
-    baseUrl + "/v1/models",
-    headers,
-    (model) => model.id,
-  );
+function getProviderRequest(provider: ProviderConfig): {
+  baseUrl: string;
+  headers: string[];
+} {
+  const apiKey = readProviderKey(provider)?.[0] || null;
+  return {
+    baseUrl: getProviderBaseUrl(provider),
+    headers: getProviderHeaders(buildAuthHeaders(provider, apiKey)),
+  };
 }
 
-function listAnthropicModels(provider: ProviderConfig): string {
-  const apiKey = readProviderKey(provider);
-  const baseUrl = getProviderBaseUrl(provider);
-  const headers = getProviderHeaders(
-    buildAuthHeaders(provider, apiKey?.[0] || null),
-  );
-  return listDataModels(
-    baseUrl + "/v1/models",
-    headers,
-    (model) => model.id,
-  );
+function listV1Models(provider: ProviderConfig): string {
+  const { baseUrl, headers } = getProviderRequest(provider);
+  const url = joinUrl(baseUrl, "v1/models");
+  return listDataModels(url, headers, (model) => model.id);
 }
 
 function listOllamaModels(provider: ProviderConfig): string {
-  const response = httpGet(
-    getProviderBaseUrl(provider) + "/api/tags",
-    getProviderHeaders(
-      buildAuthHeaders(provider, readProviderKey(provider)?.[0] || null),
-    ),
-  );
+  const { baseUrl, headers } = getProviderRequest(provider);
+  const response = httpGet(joinUrl(baseUrl, "api/tags"), headers);
   const error = getErrorMessage(response.error);
   if (error) {
     console.error(error);
@@ -164,11 +156,8 @@ function listOllamaModels(provider: ProviderConfig): string {
 }
 
 function listMistralModels(provider: ProviderConfig): string {
-  const apiKey = readProviderKey(provider);
-  const response = httpGet(
-    getProviderBaseUrl(provider) + "/v1/models",
-    getProviderHeaders(buildAuthHeaders(provider, apiKey?.[0] || null)),
-  );
+  const { baseUrl, headers } = getProviderRequest(provider);
+  const response = httpGet(joinUrl(baseUrl, "v1/models"), headers);
   if (response.data) {
     return uniqueBy(response.data, (model) => model.name || model.id)
       .map((model) =>
@@ -192,70 +181,73 @@ const providerRuntimes: Record<ApiStyle, ProviderRuntime> = {
     }),
     parseResponse: (response) => {
       if (response.error) {
-        const error = typeof response.error === "object"
-          ? response.error.message
-          : response.error;
-        throw new Error(error || "Unknown error");
+        throw new Error(getErrorMessage(response.error) || "Unknown error");
       }
       if (response.choices && response.choices[0]?.message?.content) {
         return filterResponse(response.choices[0].message.content);
       }
       throw new Error("Invalid response format");
     },
-    buildUrl: (baseUrl) => baseUrl + "/v1/chat/completions",
+    buildUrl: (baseUrl) => joinUrl(baseUrl, "v1/chat/completions"),
   },
   anthropic: {
     buildPayload: (model, query) => {
-      const payload: {
-        model: string;
-        max_tokens: number;
-        messages: Array<{ role: string; content: string }>;
-        temperature?: number;
-        top_p?: number;
-        top_k?: number;
-      } = {
+      const payload: AnthropicPayload = {
         model,
         max_tokens: 5128,
         messages: [{ role: "user", content: query }],
       };
-      if (state.deterministic) {
+      if (isThinkEnabled()) {
+        payload.thinking = { type: "enabled", budget_tokens: 4096 };
+        payload.max_tokens = 16000;
+      }
+      if (state.deterministic && !isThinkEnabled()) {
         Object.assign(payload, { temperature: 0, top_p: 0, top_k: 1 });
       }
       return payload;
     },
     parseResponse: (response) => {
-      if (response.content && response.content[0]?.text) {
-        return filterResponse(response.content[0].text);
+      if (response.content && Array.isArray(response.content)) {
+        const parts: string[] = [];
+        for (const block of response.content as Array<Record<string, unknown>>) {
+          if (block.text) {
+            parts.push(block.text as string);
+          }
+        }
+        if (parts.length > 0) {
+          return filterResponse(parts.join("\n"));
+        }
       }
       if (response.error) {
-        const error = typeof response.error === "object"
-          ? response.error.message
-          : response.error;
-        throw new Error(error || "Unknown error");
+        throw new Error(getErrorMessage(response.error) || "Unknown error");
       }
       throw new Error("Invalid response format");
     },
-    buildUrl: (baseUrl) => baseUrl + "/v1/messages",
+    buildUrl: (baseUrl) => joinUrl(baseUrl, "v1/messages"),
   },
   ollama: {
     buildPayload: (model, query) => {
-      const payload: {
-        stream: boolean;
-        model: string;
-        messages: Array<{ role: string; content: string }>;
-        options?: {
-          repeat_last_n: number;
-          top_p: number;
-          top_k: number;
-          temperature: number;
-          repeat_penalty: number;
-          seed: number;
-        };
-      } = {
+      const payload: OllamaPayload = {
         stream: false,
         model,
-        messages: [{ role: "user", content: query }],
       };
+      if (state.apitype === "generate") {
+        payload.prompt = query;
+      } else {
+        payload.messages = [{ role: "user", content: query }];
+      }
+      if (isThinkEnabled()) {
+        if (state.think === "true" || state.think === "1") {
+          payload.think = true;
+        } else {
+          // Pass reasoning levels like "low", "medium", "high" as-is
+          payload.think = state.think;
+        }
+      } else {
+        // Disable thinking by default — models like glm-5 enable it
+        // implicitly, generating thousands of reasoning tokens
+        payload.think = false;
+      }
       if (state.deterministic) {
         payload.options = {
           repeat_last_n: 0,
@@ -270,65 +262,67 @@ const providerRuntimes: Record<ApiStyle, ProviderRuntime> = {
     },
     parseResponse: (response) => {
       if (response.error) {
-        const error = typeof response.error === "string"
-          ? response.error
-          : JSON.stringify(response.error);
-        throw new Error(error);
+        throw new Error(getErrorMessage(response.error));
+      }
+      if (typeof response.response === "string") {
+        return filterResponse(response.response);
       }
       if (response.message?.content) {
         return filterResponse(response.message.content);
       }
       throw new Error(JSON.stringify(response));
     },
-    buildUrl: (baseUrl) => baseUrl + "/api/chat",
+    buildUrl: (baseUrl) => joinUrl(baseUrl, "api/" + state.apitype),
   },
   gemini: {
-    buildPayload: (model, query) => {
-      const payload: {
-        contents: Array<{ parts: Array<{ text: string }> }>;
-        generationConfig?: {
-          temperature: number;
-          topP: number;
-          topK: number;
-        };
-      } = {
+    buildPayload: (_model, query) => {
+      const payload: GeminiPayload = {
         contents: [{ parts: [{ text: query }] }],
       };
+      const genConfig: Record<string, unknown> = {};
       if (state.deterministic) {
-        payload.generationConfig = {
-          temperature: 0.0,
-          topP: 1.0,
-          topK: 1,
-        };
+        Object.assign(genConfig, { temperature: 0.0, topP: 1.0, topK: 1 });
+      }
+      if (state.think !== "") {
+        if (isThinkDisabled()) {
+          genConfig.thinkingConfig = { thinkingBudget: 0 };
+        } else {
+          genConfig.thinkingConfig = { thinkingBudget: 8192 };
+        }
+      }
+      if (Object.keys(genConfig).length > 0) {
+        payload.generationConfig = genConfig;
       }
       return payload;
     },
     parseResponse: (response) => {
-      if (
-        response.candidates &&
-        response.candidates[0]?.content?.parts?.[0]?.text
-      ) {
-        return filterResponse(response.candidates[0].content.parts[0].text);
+      if (response.candidates && response.candidates[0]?.content?.parts) {
+        const parts = response.candidates[0].content.parts as Array<Record<string, unknown>>;
+        const textParts = parts
+          .filter((p) => !p.thought && p.text)
+          .map((p) => p.text as string);
+        if (textParts.length > 0) {
+          return filterResponse(textParts.join("\n"));
+        }
       }
       if (response.error) {
-        throw new Error(
-          typeof response.error === "string"
-            ? response.error
-            : JSON.stringify(response.error),
-        );
+        throw new Error(getErrorMessage(response.error));
       }
       console.log(JSON.stringify(response));
       throw new Error("Invalid response format");
     },
     buildUrl: (baseUrl, model, apiKey) =>
-      `${baseUrl}/v1beta/models/${model}:generateContent?key=${apiKey}`,
+      joinUrl(
+        baseUrl,
+        `v1beta/models/${model}:generateContent?key=${apiKey}`,
+      ),
     requiresUrlApiKey: true,
   },
 };
 
 export const providerRegistry: ProviderRegistry = {
   anthropic: {
-    defaultModel: "claude-3-7-sonnet-20250219",
+    defaultModel: "claude-sonnet-5-5",
     defaultBaseUrl: "https://api.anthropic.com",
     authKey: "ANTHROPIC_API_KEY",
     keyName: "anthropic",
@@ -336,7 +330,7 @@ export const providerRegistry: ProviderRegistry = {
     apiStyle: "anthropic",
   },
   claude: {
-    defaultModel: "claude-3-7-sonnet-20250219",
+    defaultModel: "claude-sonnet-5-5",
     defaultBaseUrl: "https://api.anthropic.com",
     authKey: "ANTHROPIC_API_KEY",
     keyName: "claude",
@@ -364,15 +358,6 @@ export const providerRegistry: ProviderRegistry = {
     authStyle: "bearer",
     apiStyle: "ollama",
   },
-  opencode: {
-    defaultModel: "big-pickle",
-    defaultBaseUrl: "https://opencode.ai/zen",
-    authKey: "OPENCODE_API_KEY",
-    keyName: "opencode",
-    authStyle: "bearer",
-    apiStyle: "openai",
-    hardcodedModels: ["big-pickle", "glm-5", "kimi-k2.5"],
-  },
   zen: {
     defaultModel: "big-pickle",
     defaultBaseUrl: "https://opencode.ai/zen",
@@ -461,9 +446,8 @@ export function listModels(providerName: string): string {
 
   switch (provider.apiStyle) {
     case "openai":
-      return listOpenAIModels(provider);
     case "anthropic":
-      return listAnthropicModels(provider);
+      return listV1Models(provider);
     case "ollama":
       return listOllamaModels(provider);
     case "gemini":
@@ -473,15 +457,24 @@ export function listModels(providerName: string): string {
   }
 }
 
-function buildQuery(msg: string, hideprompt: boolean): string {
+function isThinkDisabled(): boolean {
+  return state.think === "false" || state.think === "0";
+}
+
+function isThinkEnabled(): boolean {
+  return state.think !== "" && !isThinkDisabled();
+}
+
+function buildQuery(msg: string, hideprompt: boolean, apiStyle: ApiStyle): string {
   let query = msg;
 
-  if (state.think >= 0) {
-    if (state.think === 0) {
+  // Only add prompt-based thinking hints for APIs without native think support
+  if (state.think !== "" && apiStyle === "openai") {
+    if (isThinkDisabled()) {
       query +=
         ' Answers directly and concisely, without showing any thinking steps or internal reasoning. Never include phrases like "Let me think".';
       query += " /no_think";
-    } else if (state.think > 0) {
+    } else {
       query =
         "Think step by step and explain the reasoning process, When answering, first output your reasoning inside <think> and </think> tags, then give the final answer." +
         query;
@@ -503,7 +496,7 @@ function callRuntime(
   hideprompt: boolean,
 ): string {
   const model = state.model || provider.defaultModel;
-  const query = buildQuery(msg, hideprompt);
+  const query = buildQuery(msg, hideprompt, provider.apiStyle);
   const apiKey = readProviderKey(provider);
 
   if (runtime.requiresUrlApiKey && apiKey?.[1] && provider.authKey) {
@@ -511,9 +504,11 @@ function callRuntime(
   }
 
   const payload = runtime.buildPayload(model, query);
-  const headers = getProviderHeaders(
-    buildAuthHeaders(provider, apiKey?.[0] || null),
-  );
+  const authHeaders = buildAuthHeaders(provider, apiKey?.[0] || null);
+  if (provider.apiStyle === "ollama") {
+    authHeaders.push("Accept: application/x-ndjson");
+  }
+  const headers = getProviderHeaders(authHeaders);
   const url = runtime.buildUrl(
     getProviderBaseUrl(provider),
     model,

decai/src/r2ai.ts

@@ -1,5 +1,5 @@
 import { state } from "./state";
-import { fileDump, tmpdir } from "./utils";
+import { fileDump, joinUrl, tmpdir } from "./utils";
 import { callProvider } from "./providers";
 import { httpGet } from "./http";
 
@@ -20,11 +20,8 @@ export function r2ai(
       ? cleanQuery
       : ["-i", fileName, cleanQuery].join(" ");
 
-    const host = state.baseurl
-      ? state.baseurl + "/cmd"
-      : state.host + ":" + state.port + "/cmd";
-
-    const url = host + "/" + encodeURIComponent(q);
+    const baseUrl = state.baseurl || state.host + ":" + state.port;
+    const url = joinUrl(baseUrl, "cmd/" + encodeURIComponent(q));
     const response = httpGet(url, []);
     if (response.error) {
       return `Error: ${response.error}`;

decai/src/r2pipe.ts

@@ -1,5 +1,17 @@
 export type R2Address = number | string | unknown;
 
+export interface R2HttpResponse {
+  status: number;
+  body: string;
+}
+
+export type R2HttpRequest = (
+  method: "GET" | "POST",
+  url: string,
+  headers: string,
+  data?: string,
+) => R2HttpResponse;
+
 export interface R2LogEntry {
   type: string;
   message: string;
@@ -21,10 +33,12 @@ export interface R2Plugin {
 
 export type R2PluginFactory = () => R2Plugin;
 
-export class R2PipeSyncFromSync {
+export class R2PipeAdapter {
   r2p: R2Pipe;
+  http?: R2HttpRequest;
   constructor(r2p: R2Pipe) {
     this.r2p = r2p;
+    this.http = r2p.http?.bind(r2p);
   }
   /**
    * Run a command in the associated instance of radare2 and return the output as a string
@@ -80,9 +94,8 @@ export class R2PipeSyncFromSync {
   }
 }
 
-export function newAsyncR2PipeFromSync(r2p: R2Pipe): R2PipeSync {
-  const asyncR2Pipe = new R2PipeSyncFromSync(r2p);
-  return asyncR2Pipe as R2PipeSync;
+export function newR2PipeSync(r2p: R2Pipe): R2PipeSync {
+  return new R2PipeAdapter(r2p) as R2PipeSync;
 }
 
 /**
@@ -141,6 +154,7 @@ export interface R2Pipe {
   syscmd(cmd: string): string;
   syscmds(cmd: string): string;
   fdump(data: string): string;
+  http?: R2HttpRequest;
 
   /**
    * Same as cmdj but using .call which avoids command injection problems
@@ -232,6 +246,7 @@ export interface R2PipeSync {
   syscmd(cmd: string): string;
   syscmds(cmd: string): string;
   fdump(data: string): string;
+  http?: R2HttpRequest;
 
   /**
    * Same as cmdj but using .call which avoids command injection problems

decai/src/state.ts

@@ -17,7 +17,9 @@ export const defaultState: AppState = {
   deterministic: true,
   debug: false,
   timeout: 180,
-  think: -1,
+  http: "auto",
+  think: "",
+  apitype: "chat",
   useFiles: false,
   contextFile: "",
   model: "",

decai/src/types.ts

@@ -13,6 +13,7 @@ declare global {
 
 export type ApiStyle = "openai" | "anthropic" | "ollama" | "gemini";
 export type AuthStyle = "none" | "bearer" | "anthropic";
+export type OllamaApiType = "chat" | "generate";
 
 export interface ProviderConfig {
   defaultModel: string;
@@ -50,7 +51,9 @@ export interface AppState {
   deterministic: boolean;
   debug: boolean;
   timeout: number;
-  think: number;
+  http: "auto" | "r2" | "curl";
+  think: string;
+  apitype: OllamaApiType;
   useFiles: boolean;
   contextFile: string;
   model: string;
@@ -126,6 +129,37 @@ export interface HttpResponse {
 
 export type JsonObject = Record<string, unknown>;
 
+export interface AnthropicPayload {
+  model: string;
+  max_tokens: number;
+  messages: Array<{ role: string; content: string }>;
+  thinking?: { type: string; budget_tokens?: number };
+  temperature?: number;
+  top_p?: number;
+  top_k?: number;
+}
+
+export interface OllamaPayload {
+  stream: boolean;
+  model: string;
+  messages?: Array<{ role: string; content: string }>;
+  prompt?: string;
+  think?: boolean | string;
+  options?: {
+    repeat_last_n: number;
+    top_p: number;
+    top_k: number;
+    temperature: number;
+    repeat_penalty: number;
+    seed: number;
+  };
+}
+
+export interface GeminiPayload {
+  contents: Array<{ parts: Array<{ text: string }> }>;
+  generationConfig?: Record<string, unknown>;
+}
+
 export interface AutoReply {
   action: string;
   command?: string;

decai/src/utils.ts

@@ -5,9 +5,18 @@ export function tmpdir(path: string): string {
   return dir + "/" + path;
 }
 
+// r2's mkdir and touch commands do not expand '~'
+export function expandHome(path: string): string {
+  if (!path.startsWith("~/")) {
+    return path;
+  }
+  const home = r2.cmd("'%HOME").trim() || r2.cmd("'%USERPROFILE").trim();
+  return home ? home + path.slice(1) : path;
+}
+
 export function ensurePath(dir: string, file: string): void {
-  r2.cmd("'mkdir -p " + dir);
-  r2.cmd("'touch " + file);
+  r2.cmd("'mkdir -p " + expandHome(dir));
+  r2.cmd("'touch " + expandHome(file));
 }
 
 export function fileExists(path: string): boolean {
@@ -22,6 +31,10 @@ export function padRight(str: string, length: number): string {
   return str + " ".repeat(Math.max(0, length - str.length));
 }
 
+export function joinUrl(baseUrl: string, path: string): string {
+  return baseUrl.replace(/\/+$/, "") + "/" + path.replace(/^\/+/, "");
+}
+
 export function trimAnsi(str: string): string {
   return str.replace(/\x1b\[[0-9;]*m/g, "");
 }
@@ -33,10 +46,10 @@ export function trimDown(out: string): string {
 
 export function trimJson(out: string): string {
   let result = out;
-  const bob = result.indexOf("{");
-  if (bob !== -1) result = result.slice(bob);
-  const eob = result.lastIndexOf("}");
-  if (eob !== -1) result = result.slice(0, eob + 1);
+  const openBrace = result.indexOf("{");
+  if (openBrace !== -1) result = result.slice(openBrace);
+  const closeBrace = result.lastIndexOf("}");
+  if (closeBrace !== -1) result = result.slice(0, closeBrace + 1);
   return result;
 }
 
@@ -51,9 +64,7 @@ export function fileDump(fileName: string, fileData: string): void {
 
 export function filterResponse(msg: string): string {
   let result = msg;
-  if (state.think !== 2) {
-    result = result.replace(/<think>[\s\S]*?<\/think>/gi, "");
-  }
+  result = result.replace(/<think>[\s\S]*?<\/think>/gi, "");
   return result
     .split("\n")
     .filter((line) => !line.trim().startsWith("```"))

prompts/README.md

@@ -4,24 +4,37 @@ This directory contains predefined prompt files for the r2ai plugin.
 
 ## Format
 
-Each prompt is stored in a `.r2ai` file using a simple Key: Value format.
+Each prompt is stored in a `.r2ai.md` file with YAML-style frontmatter and a markdown body.
 
-Lines starting with `#` are comments and ignored.
+Example:
+
+```markdown
+---
+description: 'Explain the current function'
+author: pancake
+command: r2ai -d
+if-empty: exit
+---
+Explain the purpose of this function in one or two short sentences.
+```
 
 Supported keys:
 
-- `Title`: The title of the prompt (required)
-- `Author`: The author of the prompt
-- `Description`: A brief description
-- `Command` or `Commands`: Commands to execute before sending the prompt, separated by `;` (required)
-- `Prompt` or `Query`: The main prompt text sent to the LLM
-- `Requires`: Requirements for the prompt (e.g., "analysis" to ensure binary analysis)
-- `If-Empty`: Alternative prompt or message if the command output is empty
-- `If-Command`: Additional commands to run if the main command produces output
+- `title`: The title of the prompt
+- `author`: The author of the prompt
+- `description`: A brief description
+- `command`: Commands to execute before sending the prompt, separated by `;` (required)
+- `requires`: Requirements for the prompt (e.g., "analysis" to ensure binary analysis)
+- `if-empty`: Alternative prompt or message if the command output is empty
+- `if-command`: Additional commands to run if the main command produces output
+- `model`: Model override for this prompt
+- `provider`: Provider override for this prompt
+
+Legacy `.r2ai.txt` files using `Title:`, `Command:`, `Prompt:` or `Query:` are still accepted.
 
 ## Variable Substitution
 
-In the `Prompt` field, you can use:
+In the markdown body, you can use:
 
 - `${VAR}`: Replaced with the value of the environment variable `VAR`
 - `$(cmd)`: Replaced with the output of the r2 command `cmd`
@@ -29,15 +42,16 @@ In the `Prompt` field, you can use:
 ## Usage
 
 - `r2ai -q`: List all available prompts
+- `r2ai -qj`: List all available prompts in JSON
 - `r2ai -q name`: Run the prompt named `name`
 - `r2ai -q name extra text`: Run the prompt with additional text appended
 
 ## Conditionals
 
-- `If-Empty`: If the command output is empty, use this as the prompt instead
-- `If-Command`: If the command produces output, run these additional commands and append their output
+- `if-empty`: If the command output is empty, use this as the prompt instead
+- `if-command`: If the command produces output, run these additional commands and append their output
 
 ## Future Features
 
 - Support for r2js scripts
-- More advanced conditionals
+- More advanced conditionals

r2plugin/README.md

@@ -0,0 +1,40 @@
+# r2plugin
+
+This directory contains the make and meson files required to link r2ai
+inside radare2 through the external plugin system (XPS).
+
+```console
+cd radare2
+make -C libr/xps EXTERNAL_PLUGINS=r2ai
+./configure-plugins
+make
+```
+
+For meson builds, place this repository at `libr/xps/p/r2ai` and make
+`libr/xps/p/meson.build` include:
+
+```meson
+subdir('r2ai/r2plugin')
+```
+
+## WASI HTTP build
+
+Use a radare2 checkout with the host HTTP bridge and place this repository at
+`libr/xps/p/r2ai`. From that radare2 checkout:
+
+```sh
+echo 'EXTERNAL_PLUGINS+=r2ai' >> libr/xps/config.mk
+make -C libr/xps
+WASI_HTTP=1 sys/wasi-api.sh
+```
+
+This links the C plugin into the reactor; it does not load a native shared
+library at runtime. Enable HTTP in the host adapter. The `auto` backend selects
+`r_socket_http_get/post`, preserving provider authentication headers and JSON
+bodies, subject to the radare2 sandbox. `--disable-network` also denies these calls.
+
+WASI excludes POSIX signal/alarm handling; the host controls request timeouts,
+not `r2ai.http.timeout`. Normal queries and model listing use buffered HTTP.
+Vertex's automatic `gcloud` authentication and background tasks requiring
+threads are unavailable. Explicit shell backends (`system`/`curl`) cannot work;
+use `auto` or `socket`. Browser CORS and restricted-header rules still apply.

r2plugin/config.mk

@@ -0,0 +1,9 @@
+EXTERNAL_PLUGINS+=r2ai
+# EXTERNAL_PLUGINS+=hi
+
+.PHONY: r2ai
+
+r2ai: p/r2ai
+
+p/r2ai:
+	cd p && git clone https://github.com/radareorg/r2ai

r2plugin/core/deps.h

@@ -0,0 +1,3 @@
+#ifdef R2_CORE_H
+extern RCorePlugin r_core_plugin_r2ai;
+#endif

r2plugin/core/deps.mk

@@ -0,0 +1,23 @@
+R2AI_CORE_WD=$(LIBR)/xps/p/r2ai
+CFLAGS+=-I$(R2AI_CORE_WD)/src
+R2AI_CORE_OBJ= \
+	$(R2AI_CORE_WD)/src/anthropic.o \
+	$(R2AI_CORE_WD)/src/apikeys.o \
+	$(R2AI_CORE_WD)/src/async.o \
+	$(R2AI_CORE_WD)/src/auto.o \
+	$(R2AI_CORE_WD)/src/claw.o \
+	$(R2AI_CORE_WD)/src/gemini.o \
+	$(R2AI_CORE_WD)/src/http.o \
+	$(R2AI_CORE_WD)/src/json.o \
+	$(R2AI_CORE_WD)/src/llm.o \
+	$(R2AI_CORE_WD)/src/markdown.o \
+	$(R2AI_CORE_WD)/src/messages.o \
+	$(R2AI_CORE_WD)/src/openai.o \
+	$(R2AI_CORE_WD)/src/query.o \
+	$(R2AI_CORE_WD)/src/r2ai.o \
+	$(R2AI_CORE_WD)/src/rawtools.o \
+	$(R2AI_CORE_WD)/src/tools.o \
+	$(R2AI_CORE_WD)/src/vdb.o \
+	$(R2AI_CORE_WD)/src/vertex.o \
+	$(R2AI_CORE_WD)/src/wizard.o
+EXTERNAL_STATIC_OBJS+=$(R2AI_CORE_OBJ)

r2plugin/core/meson.build

@@ -0,0 +1,27 @@
+shared_library(
+  'core_r2ai',
+  '../src/anthropic.c',
+  '../src/apikeys.c',
+  '../src/async.c',
+  '../src/auto.c',
+  '../src/claw.c',
+  '../src/gemini.c',
+  '../src/http.c',
+  '../src/json.c',
+  '../src/llm.c',
+  '../src/markdown.c',
+  '../src/messages.c',
+  '../src/openai.c',
+  '../src/query.c',
+  '../src/r2ai.c',
+  '../src/rawtools.c',
+  '../src/tools.c',
+  '../src/vdb.c',
+  '../src/vertex.c',
+  '../src/wizard.c',
+  include_directories: ['../src', include_dirs],
+  dependencies: r2_core,
+  override_options: ['werror=false'],
+  install: true,
+  install_dir: r2_plugdir,
+)

r2plugin/deps.mk

@@ -0,0 +1 @@
+include p/r2ai/r2plugin/core/deps.mk

r2plugin/meson.build

@@ -0,0 +1,30 @@
+# r2ai XPS integration for radare2.
+# Included from libr/xps/meson.build when r2ai sources live under
+# libr/xps/p/r2ai/. Appends sources and include paths to the r_core
+# aggregate so the static plugin is linked into libr_core.
+
+core_plugins += ['r2ai']
+
+r_core_additional_inc += include_directories(join_paths('..', 'src'))
+
+r_core_additional_sources += files(
+  '../src/anthropic.c',
+  '../src/apikeys.c',
+  '../src/async.c',
+  '../src/auto.c',
+  '../src/claw.c',
+  '../src/gemini.c',
+  '../src/http.c',
+  '../src/json.c',
+  '../src/llm.c',
+  '../src/markdown.c',
+  '../src/messages.c',
+  '../src/openai.c',
+  '../src/query.c',
+  '../src/r2ai.c',
+  '../src/rawtools.c',
+  '../src/tools.c',
+  '../src/vdb.c',
+  '../src/vertex.c',
+  '../src/wizard.c',
+)

r2plugin/r2plugins.h

@@ -0,0 +1,3 @@
+#ifdef R2_CORE_H
+#include "core/deps.h"
+#endif

r2plugin/static.cfg

@@ -0,0 +1 @@
+core.r2ai

src/Makefile

@@ -5,7 +5,10 @@ DOTLIB=.$(shell r2 -NNH R2_LIBEXT)
 R2SYSPLUGDIR=$(shell r2 -NNH R2_LIBR_PLUGINS)
 R2USRPLUGDIR=$(shell r2 -NNH R2_USER_PLUGINS)
 R2BINDIR=$(shell r2pm -H R2PM_BINDIR)
+R2INCDIR=$(shell r2 -NNH R2_INCDIR)/libr
+R2LIBDIR=$(shell r2 -NNH R2_LIBDIR)
 R2MANDIR=$(shell r2pm -H R2PM_MANDIR)
+R2AI_PROMPTSDIR=$(HOME)/.config/r2ai/prompts
 ifeq ($(R2MANDIR),)
 R2MANDIR=$(shell r2 -NNH R2_PREFIX)/share/man
 endif
@@ -23,6 +26,8 @@ CFLAGS+=-I. -fPIC
 CFLAGS+=-Wall
 CFLAGS+=-g
 
+AR?=ar
+
 USE_R2_CURL=1
 USE_LIBCURL=0
 
@@ -43,6 +48,8 @@ endif
 
 OBJS=r2ai.o
 OBJS+=apikeys.o
+OBJS+=async.o
+OBJS+=claw.o
 OBJS+=llm.o
 OBJS+=auto.o
 OBJS+=json.o
@@ -51,6 +58,7 @@ OBJS+=tools.o
 OBJS+=messages.o
 OBJS+=anthropic.o
 OBJS+=gemini.o
+OBJS+=vertex.o
 OBJS+=http.o
 OBJS+=openai.o
 OBJS+=markdown.o
@@ -60,8 +68,40 @@ OBJS+=wizard.o
 
 all: $(OBJS) r2check
 	$(CC) -fPIC -shared -o r2ai$(DOTLIB) $(OBJS) $(CFLAGS) $(LDFLAGS)
+	$(AR) rcs libr2ai.a $(OBJS)
 	$(CC) -pie -o r2ai$(DOTEXE) $(OBJS) $(CFLAGS) $(LDFLAGS) -lm main.c
 
+# Binary package for r2pm -bi: plugins/, bin/, man/ and prompts/ for this r2 version and platform
+# r2pm -H R2PM_* needs a recent r2pm; fall back to uname when missing
+R2PM_OS_SH=$(shell r2pm -H R2PM_OS 2>/dev/null)
+R2PM_ARCH_SH=$(shell r2pm -H R2PM_ARCH 2>/dev/null)
+R2PM_BITS_SH=$(shell r2pm -H R2PM_BITS 2>/dev/null)
+ifeq ($(R2PM_OS_SH),)
+ifeq ($(shell uname -s),Darwin)
+R2PM_OS_SH=darwin
+else
+R2PM_OS_SH=linux
+endif
+endif
+ifeq ($(R2PM_ARCH_SH),)
+ifeq ($(shell uname -m),x86_64)
+R2PM_ARCH_SH=x86
+else ifneq (,$(filter $(shell uname -m),aarch64 arm64))
+R2PM_ARCH_SH=arm
+else
+R2PM_ARCH_SH=$(shell uname -m)
+endif
+endif
+ifeq ($(R2PM_BITS_SH),)
+R2PM_BITS_SH=$(shell getconf LONG_BIT 2>/dev/null || echo 64)
+endif
+BINDIST_ZIP=r2ai-$(shell r2 -qv)-$(R2PM_OS_SH)-$(R2PM_ARCH_SH)-$(R2PM_BITS_SH).zip
+
+bindist: all
+	rm -rf build/bindist ../$(BINDIST_ZIP)
+	$(MAKE) user-install R2USRPLUGDIR=$(CURDIR)/build/bindist/plugins R2BINDIR=$(CURDIR)/build/bindist/bin R2MANDIR=$(CURDIR)/build/bindist/man R2AI_PROMPTSDIR=$(CURDIR)/build/bindist/prompts
+	cd build/bindist && zip -q -r $(CURDIR)/../$(BINDIST_ZIP) plugins bin man prompts
+
 try: all
 	$(MAKE) user-install
 	$(MAKE) doc-install
@@ -79,12 +119,12 @@ user-install: user-uninstall
 	mkdir -p $(R2USRPLUGDIR)
 	mkdir -p $(R2BINDIR)
 	mkdir -p $(R2MANDIR)/man1
-	mkdir -p $(HOME)/.config/r2ai/prompts
+	mkdir -p $(R2AI_PROMPTSDIR)
 	$(MAKE) user-uninstall
 	cp -f r2ai$(DOTLIB) $(R2USRPLUGDIR)/r2ai$(DOTLIB)
 	cp -f r2ai$(DOTEXE) $(R2BINDIR)/r2ai$(DOTEXE)
 	cp -f r2ai.1 $(R2MANDIR)/man1/r2ai.1
-	cp -r ../prompts/* $(HOME)/.config/r2ai/prompts/
+	cp -r ../prompts/* $(R2AI_PROMPTSDIR)/
 
 user-uninstall:
 	rm -f $(R2USRPLUGDIR)/r2ai$(DOTLIB)
@@ -93,16 +133,24 @@ user-uninstall:
 
 install:
 	$(MAKE) uninstall
-	mkdir -p $(R2USRPLUGDIR)
+	mkdir -p $(R2SYSPLUGDIR)
 	mkdir -p $(R2BINDIR)
+	mkdir -p $(R2LIBDIR)
+	mkdir -p $(R2INCDIR)
 	cp -f r2ai$(DOTLIB) $(R2SYSPLUGDIR)/r2ai$(DOTLIB)
 	cp -f r2ai$(DOTEXE) $(R2BINDIR)/r2ai$(DOTEXE)
+	cp -f libr2ai.a $(R2LIBDIR)/libr2ai.a
+	cp -f r2ai.h $(R2INCDIR)/r2ai.h
 
 uninstall:
 	rm -f $(R2SYSPLUGDIR)/r2ai$(DOTLIB)
 	rm -f $(R2BINDIR)/r2ai$(DOTEXE)
+	rm -f $(R2LIBDIR)/libr2ai.a
+	rm -f $(R2INCDIR)/r2ai.h
 
 r2check:
 	@r2 -NN -qcq --
 clean:
 	rm -f *.o *.d
+
+.PHONY: all bindist try doc-install fmt user-install user-uninstall install uninstall r2check clean test

src/README.md

@@ -33,6 +33,7 @@ make user-install
 Usage: r2ai   [-args] [...]
 | r2ai -d                 Decompile current function
 | r2ai -d [query]         Ask a question on the current function
+| r2ai -do                Decompile current function with source offsets
 | r2ai -dr                Decompile current function (+ 1 level of recursivity)
 | r2ai -a [query]         Resolve question using auto mode
 | r2ai -e (k(=v))         Same as '-e r2ai.'
@@ -41,10 +42,11 @@ Usage: r2ai   [-args] [...]
 | r2ai -m                 show selected model, list suggested ones, choose one
 | r2ai -p [provider]      set LLM provider (openai, anthropic, gemini, etc.)
 | r2ai -q                 list available query prompts
+| r2ai -qj                list available query prompts in JSON
 | r2ai -q [name] (inst)   run predefined prompt with optional instructions
 | r2ai -r                 enter the chat repl
 | r2ai -w                 launch the interactive setup wizard
-| r2ai -L                 show chat logs (See -Lj for json). Only for auto mode.
+| r2ai -L                 show chat logs (See -Lj for json)
 | r2ai -L-[N]             delete the last (or N last messages from the chat history)
 | r2ai -R                 reset the chat conversation context
 | r2ai -Rq ([text])       refresh and query embeddings (see r2ai.data)
@@ -54,7 +56,7 @@ Usage: r2ai   [-args] [...]
 
 ## Query Prompts
 
-Query prompts are predefined templates that combine radare2 commands with structured prompts to the LLM for common reverse engineering tasks. They are stored in the `prompts/` directory as `.r2ai.txt` files.
+Query prompts are predefined templates that combine radare2 commands with structured prompts to the LLM for common reverse engineering tasks. They are stored in the `prompts/` directory as `.r2ai.md` files with YAML-style frontmatter. Legacy `.r2ai.txt` prompts are still supported.
 
 ### Available Prompts
 
@@ -71,6 +73,7 @@ Query prompts are predefined templates that combine radare2 commands with struct
 ### Usage
 
 - `r2ai -q`: List all available query prompts
+- `r2ai -qj`: List all available query prompts in JSON
 - `r2ai -q [name]`: Run the specified prompt on the current function
 - `r2ai -q [name] [instructions]`: Run the prompt with additional custom instructions
 
@@ -91,7 +94,9 @@ These can be set with `r2ai -e <keyname>=<value>`
 |------------------|------------------------------------------------------------------------------------------------|
 | r2ai.api         | Name of the provider e.g `openai`. List possibilities with `r2ai -e r2ai.api=?`                |
 | r2ai.model       | Model name. List possibilities with `r2ai -e r2ai.model=?`                                     |
-| r2ai.baseurl     | Remote LLM base URL. Specify host if necessary e.g http://127.0.0.1:11434. |
+| r2ai.baseurl     | Provider API root. Use `/v1` for OpenAI-compatible chat or `/api` for native Ollama. |
+| r2ai.apitype | Use chat completions (default) or the native Ollama `generate` operation. |
+| r2ai.cacheck | Warn when chat requests are not append-only and may miss provider prompt-cache hits. |
 | r2ai.max_tokens  | Maximum output tokens or maximum total tokens. Check the appropriate limits for your model     |
 | r2ai.temperature | How creative the model should be. 0=not creative, 1=very creative                              |
 | r2ai.cmds        | R2 command to issue and send output in context to model                                        |
@@ -99,11 +104,14 @@ These can be set with `r2ai -e <keyname>=<value>`
 | r2ai.hlang       | Tells LLM in which language to speak                                                           |
 | r2ai.prompt      | User prompt to send to LLM with `r2ai -d` |
 | r2ai.clippy      | Display chat replies using `r2clippy` instead of plain text |
+| r2ai.markdown    | Render LLM responses with the radare2 markdown colorizer (default: true) |
 | r2ai.wizard      | Auto-start the setup wizard on the first interactive chat session |
+| r2ai.history     | Keep direct queries (`r2ai [query]`, `r2ai -d`, `r2ai -i`) in the conversation context instead of sending each one alone. Inspect it with `r2ai -L`, reset it with `r2ai -R` (default: false) |
 | r2ai.auto.max_runs | Maximum loops when using auto mode `r2ai -a` |
 | r2ai.auto.verbose | Only for auto mode `r2ai -a`. Will show the output of the tool which ran locally |
 | r2ai.auto.yolo | Set this to true if you don't want r2ai to ask you for approval before running commands sent by the LLM. This is **dangerous**. Recommendation: **leave this to false** unless you fully trust your LLM not to create havoc! |
 | r2ai.http.use_files | If you get CURL argument too long error, set this to true and arguments will be stored in a file.|
+| r2ai.http.headers | Extra HTTP headers to inject in every provider request. One per line, '\n' separated, e.g. `Authorization: Bearer x\nUser-Agent: r2ai/1.0`. Defaults to `R2AI_HEADERS`. |
 
 
 ## Example
@@ -113,11 +121,11 @@ These can be set with `r2ai -e <keyname>=<value>`
 ```
 $ r2 ./substitute
 [0x000010a0]> r2ai -e api=anthropic
-[0x000010a0]> r2ai -e model=claude-3-7-sonnet-20250219
+[0x000010a0]> r2ai -e model=claude-sonnet-5-5
 [0x000010a0]> aa
 ```
 
-- Decompile the main function: `r2ai -d`
+- Decompile the main function: `r2ai -d` or `r2ai -do` to keep source offsets in the left column
 
 ```
 [0x000010a0]> s main

src/anthropic.c

@@ -1,330 +1,132 @@
-#include "r2ai.h"
+/* Copyright r2ai - 2023-2026 - pancake */
 
-static const char *modelname(const char *model_name) {
-	return model_name? model_name: "claude-3-7-sonnet-20250219";
+#include "r2ai.h"
+#include "r2ai_priv.h"
+
+// request body, vertex takes the api version in the body and the model in the url
+R_IPI char *r2ai_anthropic_request(const R2AIArgs *args, bool vertex) {
+	char *messages = r_list_empty (args->messages)? NULL: r2ai_msgs_to_anthropic_json (args->messages);
+	if (!messages) {
+		return NULL;
+	}
+	PJ *pj = pj_new ();
+	pj_o (pj);
+	if (vertex) {
+		pj_ks (pj, "anthropic_version", "vertex-2023-10-16");
+		pj_kb (pj, "stream", false);
+	} else {
+		pj_ks (pj, "model", R_STR_ISNOTEMPTY (args->model)? args->model: "claude-sonnet-5-5");
+	}
+	pj_kn (pj, "max_tokens", args->max_tokens? args->max_tokens: 4096);
+	if (args->thinking_tokens >= 1024) {
+		pj_ko (pj, "thinking");
+		pj_ks (pj, "type", "enabled");
+		pj_kn (pj, "budget_tokens", args->thinking_tokens);
+		pj_end (pj);
+	} else if (args->deterministic) {
+		// top_p is not sent because newer models reject it along with temperature
+		pj_kn (pj, "temperature", 0);
+		pj_kn (pj, "top_k", 1);
+	}
+	if (R_STR_ISNOTEMPTY (args->system_prompt)) {
+		pj_ks (pj, "system", args->system_prompt);
+	}
+	pj_k (pj, "messages");
+	pj_raw (pj, messages);
+	free (messages);
+	char *tools = r_list_empty (args->tools)? NULL: r2ai_tools_to_anthropic_json (args->tools);
+	if (tools) {
+		pj_k (pj, "tools");
+		pj_raw (pj, tools);
+		free (tools);
+	}
+	pj_end (pj);
+	return pj_drain (pj);
 }
 
 R_IPI R2AI_ChatResponse *r2ai_anthropic(RCorePluginSession *cps, R2AIArgs args) {
-	RCore *core = cps->core;
-	const char *model = args.model;
-	char **error = args.error;
-	RList *tools = args.tools;
-	RList *messages_input = args.messages;
-
-	if (error) {
-		*error = NULL;
-	}
-
-	// Setup HTTP headers
-	char *auth_header = r_str_newf ("x-api-key: %s", args.api_key);
-	char *anthropic_version = "anthropic-version: 2023-06-01";
-	const char *headers[] = {
-		"Content-Type: application/json",
-		auth_header,
-		anthropic_version,
-		NULL
-	};
-
-	const char *anthropic_url = "https://api.anthropic.com/v1/messages";
-
-	// Get system message if available
-	const char *system_message = NULL;
-
-	// First check if it's provided in args.system_prompt
-	if (R_STR_ISNOTEMPTY (args.system_prompt)) {
-		system_message = args.system_prompt;
-	} else {
-		// If no system_prompt in args, check config as fallback
-		system_message = r_config_get (core->config, "r2ai.system");
-	}
-
-	// Create messages JSON
-	char *messages_json = NULL;
-
-	if (messages_input && !r_list_empty (messages_input)) {
-		// Convert directly to JSON without filtering
-		messages_json = r2ai_msgs_to_anthropic_json (messages_input);
-		if (!messages_json) {
-			if (error) {
-				*error = strdup ("Failed to convert messages to JSON");
-			}
-			free (auth_header);
-			return NULL;
-		}
-	} else {
-		if (error) {
-			*error = strdup ("No input or messages provided");
-		}
-		free (auth_header);
+	char *data = r2ai_anthropic_request (&args, false);
+	if (!data) {
+		*args.error = strdup ("No input or messages provided");
 		return NULL;
 	}
-
-	// Convert tools to Anthropic format if available
-	char *anthropic_tools_json = NULL;
-	if (tools && !r_list_empty (tools)) {
-		anthropic_tools_json = r2ai_tools_to_anthropic_json (tools);
-	}
-
-	// Create the request JSON
-	PJ *pj = pj_new ();
-	pj_o (pj);
-	pj_ks (pj, "model", modelname (model));
-	pj_kn (pj, "max_tokens", args.max_tokens);
-	if (args.thinking_tokens >= 1024) {
-		pj_ko (pj, "thinking");
-		pj_ks (pj, "type", "enabled");
-		pj_kn (pj, "budget_tokens", args.thinking_tokens);
-		pj_end (pj);
-	}
-	// Add system message if available
-	if (system_message) {
-		pj_ks (pj, "system", system_message);
-	}
-
-	// Add messages
-	pj_k (pj, "messages");
-	// messages_json already contains the array itself, so we use raw
-	pj_raw (pj, messages_json);
-
-	// Add tools if available
-	if (anthropic_tools_json) {
-		pj_k (pj, "tools");
-		pj_raw (pj, anthropic_tools_json);
-		free (anthropic_tools_json);
-	}
-
-	pj_end (pj);
-
-	char *data = pj_drain (pj);
-	free (messages_json);
-
-	// Save the full JSON for debugging
-	char *tmpdir = r_file_tmpdir ();
-	char *req_path = r_str_newf ("%s" R_SYS_DIR "r2ai_anthropic_request.json", tmpdir);
-	r_file_dump (req_path, (const ut8 *)data, strlen (data), 0);
-	R_LOG_DEBUG ("Full request saved to %s", req_path);
-	free (req_path);
-	free (tmpdir);
-
-	R_LOG_DEBUG ("Anthropic API request data: %s", data);
-
-	if (r_config_get_b (core->config, "r2ai.debug")) {
-		// Generate curl command for debugging
-		RStrBuf *curl_cmd = r_strbuf_new ("curl -X POST");
-		for (int i = 0; headers[i]; i++) {
-			r_strbuf_appendf (curl_cmd, " -H \"%s\"", headers[i]);
-		}
-		r_strbuf_appendf (curl_cmd, " -d '%s' \"%s\"", data, anthropic_url);
-		eprintf ("Curl command: %s\n", r_strbuf_get (curl_cmd));
-		r_strbuf_free (curl_cmd);
-	}
-
-	// Make the API call
+	char *auth = r_str_newf ("x-api-key: %s", args.api_key);
+	const char *headers[] = { "Content-Type: application/json", auth, "anthropic-version: 2023-06-01", NULL };
 	int code = 0;
-	char *res = r2ai_http_post (core, anthropic_url, headers, data, &code, NULL);
+	char *res = r2ai_post (cps->core, "anthropic", "https://api.anthropic.com/v1/messages", headers, data, &code, args.error);
 	free (data);
-	free (auth_header);
-
-	if (!res || code != 200) {
-		R_LOG_ERROR ("Anthropic API error %d", code);
-		if (error && res) {
-			*error = strdup (res);
-		} else if (error) {
-			*error = strdup ("Failed to get response from Anthropic API");
-		}
-		free (res);
-		return NULL;
-	}
-
-	// Save the response for inspection
-	tmpdir = r_file_tmpdir ();
-	char *res_path = r_str_newf ("%s" R_SYS_DIR "r2ai_anthropic_response.json", tmpdir);
-	r_file_dump (res_path, (const ut8 *)res, strlen (res), 0);
-	R_LOG_DEBUG ("Anthropic API response saved to %s", res_path);
-	free (res_path);
-	free (tmpdir);
-
-	if (r_config_get_b (core->config, "r2ai.debug")) {
-		R_LOG_DEBUG ("Anthropic API response: %s", res);
-	}
-
-	// Parse the response
-	R2AI_ChatResponse *result = R_NEW0 (R2AI_ChatResponse);
-
-	R2AI_Message *message = NULL;
-	R2AI_Usage *usage = NULL;
-
-	char *response_copy = strdup (res);
-	if (response_copy) {
-		RJson *jres = r_json_parse (response_copy);
-		if (jres) {
-			// Create a new message structure
-			message = R_NEW0 (R2AI_Message);
-
-			const RJson *usage_json = r_json_get (jres, "usage");
-			if (usage_json && usage_json->type == R_JSON_OBJECT) {
-				usage = R_NEW0 (R2AI_Usage);
-				if (usage) {
-					const RJson *prompt_tokens = r_json_get (usage_json, "input_tokens");
-					const RJson *completion_tokens = r_json_get (usage_json, "output_tokens");
-					if (prompt_tokens && prompt_tokens->type == R_JSON_INTEGER) {
-						usage->prompt_tokens = prompt_tokens->num.u_value;
-					}
-					if (completion_tokens && completion_tokens->type == R_JSON_INTEGER) {
-						usage->completion_tokens = completion_tokens->num.u_value;
-					}
-					usage->total_tokens = usage->prompt_tokens + usage->completion_tokens;
-				}
-			}
-
-			if (message) {
-				message->role = strdup ("assistant");
-
-				RStrBuf *content_buf = r_strbuf_new ("");
-
-				// Check for tool calls in Anthropic response
-				int has_tool_use = 0;
-				int n_tool_calls = 0;
-
-				// Count tool_use blocks to determine array size
-				const RJson *content_array = r_json_get (jres, "content");
-				if (content_array && content_array->type == R_JSON_ARRAY) {
-					// First count the number of tool_use blocks
-					const RJson *content_item = content_array->children.first;
-					while (content_item) {
-						const RJson *type = r_json_get (content_item, "type");
-						if (type && type->type == R_JSON_STRING && !strcmp (type->str_value, "tool_use")) {
-							has_tool_use = 1;
-							n_tool_calls++;
-						}
-						content_item = content_item->next;
-					}
-				}
-
-				// Allocate tool_calls list if needed
-				if (has_tool_use && n_tool_calls > 0) {
-					message->tool_calls = r_list_new ();
-					if (!message->tool_calls) {
-						if (error) {
-							*error = strdup ("Failed to allocate memory for tool calls");
-						}
-						r_json_free (jres);
-						free (response_copy);
-						r2ai_message_free (message);
-						free (res);
-						return NULL;
-					}
-					message->tool_calls->free = (RListFree)r2ai_tool_call_free;
-				}
-
-				// Process each content item
-				int tool_idx = 0;
-				if (content_array && content_array->type == R_JSON_ARRAY) {
-					RList *cb = r2ai_content_blocks_new ();
-					if (!cb) {
-						r_json_free (jres);
-						free (response_copy);
-						r2ai_message_free (message);
-						free (res);
-						return NULL;
-					}
-					const RJson *content_item = content_array->children.first;
-					while (content_item) {
-						const RJson *type = r_json_get (content_item, "type");
-						if (type && type->type == R_JSON_STRING) {
-							R2AI_ContentBlock *block = R_NEW0 (R2AI_ContentBlock);
-							if (!strcmp (type->str_value, "text")) {
-								// Text content
-								const RJson *text = r_json_get (content_item, "text");
-								if (text && text->type == R_JSON_STRING) {
-									r_strbuf_append (content_buf, text->str_value);
-									block->type = strdup ("text");
-									block->text = strdup (text->str_value);
-								}
-							} else if (!strcmp (type->str_value, "tool_use") && tool_idx < n_tool_calls) {
-								// Tool call - convert from Anthropic format to OpenAI format
-								const RJson *name = r_json_get (content_item, "name");
-								const RJson *id = r_json_get (content_item, "id");
-								const RJson *input = r_json_get (content_item, "input");
-
-								block->type = strdup ("tool_use");
-								R2AI_ToolCall *tc = R_NEW0 (R2AI_ToolCall);
-								if (name && name->type == R_JSON_STRING) {
-									block->name = strdup (name->str_value);
-									tc->name = strdup (name->str_value);
-								}
-
-								if (id && id->type == R_JSON_STRING) {
-									block->id = strdup (id->str_value);
-									tc->id = strdup (id->str_value);
-								}
-
-								if (input && input->type == R_JSON_OBJECT) {
-#if 1
-									char *input_str = r_json_to_string (input);
-									if (input_str) {
-										R_LOG_DEBUG ("Input string: %s", input_str);
-										block->input = strdup (input_str);
-										tc->arguments = strdup (input_str);
-										free (input_str);
-									}
-#endif
-								}
-
-								r_list_append (message->tool_calls, tc);
-								tool_idx++;
-							} else if (!strcmp (type->str_value, "thinking")) {
-								const RJson *data = r_json_get (content_item, "data");
-								const RJson *thinking = r_json_get (content_item, "thinking");
-								const RJson *signature = r_json_get (content_item, "signature");
-								if (data && data->type == R_JSON_STRING) {
-									block->data = strdup (data->str_value);
-								}
-								if (thinking && thinking->type == R_JSON_STRING) {
-									block->thinking = strdup (thinking->str_value);
-								}
-								if (signature && signature->type == R_JSON_STRING) {
-									block->signature = strdup (signature->str_value);
-								}
-								block->type = strdup ("thinking");
-
-								r_strbuf_append (content_buf, "\n" Color_GRAY "<thinking>\n");
-								r_strbuf_append (content_buf, block->thinking);
-								r_strbuf_append (content_buf, "\n</thinking>" Color_RESET "\n");
-							}
-							r_list_append (cb, block);
-						}
-						content_item = content_item->next;
-					}
-					message->content_blocks = cb;
-				}
-
-				// Store the content
-				message->content = r_strbuf_drain (content_buf);
-
-				// If there's no content and no tool calls, clean up
-				if (!message->content && (!message->tool_calls || r_list_empty (message->tool_calls))) {
-					r2ai_message_free (message);
-					message = NULL;
-				}
-			}
-			r_json_free (jres);
-		}
-		free (response_copy);
-	}
-
-	// Assign message and usage to result
-	result->message = message;
-	result->usage = usage;
-
-	// Free the HTTP response body
+	free (auth);
+	R2AI_ChatResponse *result = res? r2ai_anthropic_parse_response (res, args.error): NULL;
 	free (res);
 	return result;
 }
 
-R_IPI char *r2ai_anthropic_stream(RCore *core, R2AIArgs args) {
-	(void)core;
-	(void)args;
-	// Not implemented yet
-	return NULL;
+static char *jstr(const RJson *j, const char *key) {
+	return r_str_new (r_json_get_str (j, key));
+}
+
+// fill the block from the content item and append its text to the message content
+static void parse_block(R2AI_ContentBlock *b, const RJson *item, R2AI_Message *msg, RStrBuf *sb) {
+	if (!strcmp (b->type, "text")) {
+		b->text = jstr (item, "text");
+		r_strbuf_append (sb, r_str_get (b->text));
+	} else if (!strcmp (b->type, "tool_use")) {
+		const RJson *input = r_json_get (item, "input");
+		b->id = jstr (item, "id");
+		b->name = jstr (item, "name");
+		b->input = (input && input->type == R_JSON_OBJECT)? r_json_to_string (input): NULL;
+		R2AI_ToolCall *tc = R_NEW0 (R2AI_ToolCall);
+		tc->id = r_str_new (b->id);
+		tc->name = r_str_new (b->name);
+		tc->arguments = r_str_new (b->input);
+		if (!msg->tool_calls) {
+			msg->tool_calls = r_list_newf ((RListFree)r2ai_tool_call_free);
+		}
+		r_list_append (msg->tool_calls, tc);
+	} else if (!strcmp (b->type, "thinking")) {
+		b->data = jstr (item, "data");
+		b->thinking = jstr (item, "thinking");
+		b->signature = jstr (item, "signature");
+		r_strbuf_appendf (sb, "\n" Color_GRAY "<thinking>\n%s\n</thinking>" Color_RESET "\n", r_str_get (b->thinking));
+	}
+}
+
+// parses (and modifies) the json response of the anthropic and vertex-anthropic providers
+R_IPI R2AI_ChatResponse *r2ai_anthropic_parse_response(char *json, char **error) {
+	RJson *j = r_json_parse (json);
+	if (!j) {
+		*error = strdup ("Failed to parse Anthropic response JSON");
+		return NULL;
+	}
+	R2AI_ChatResponse *res = R_NEW0 (R2AI_ChatResponse);
+	const RJson *u = r_json_get (j, "usage");
+	if (u) {
+		R2AI_Usage *usage = R_NEW0 (R2AI_Usage);
+		usage->prompt_tokens = r_json_get_num (u, "input_tokens");
+		usage->completion_tokens = r_json_get_num (u, "output_tokens");
+		usage->total_tokens = usage->prompt_tokens + usage->completion_tokens;
+		res->usage = usage;
+	}
+	R2AI_Message *msg = R_NEW0 (R2AI_Message);
+	msg->role = strdup ("assistant");
+	RStrBuf *sb = r_strbuf_new ("");
+	const RJson *content = r_json_get (j, "content");
+	if (content && content->type == R_JSON_ARRAY) {
+		msg->content_blocks = r2ai_content_blocks_new ();
+		const RJson *item;
+		for (item = content->children.first; item; item = item->next) {
+			char *type = jstr (item, "type");
+			if (!type) {
+				continue;
+			}
+			R2AI_ContentBlock *b = R_NEW0 (R2AI_ContentBlock);
+			b->type = type;
+			parse_block (b, item, msg, sb);
+			r_list_append (msg->content_blocks, b);
+		}
+	}
+	msg->content = r_strbuf_drain (sb);
+	res->message = msg;
+	r_json_free (j);
+	return res;
 }

src/apikeys.c

@@ -1,4 +1,4 @@
-/* r2ai - Copyright 2023-2025 pancake */
+/* r2ai - Copyright 2023-2026 pancake */
 
 #define R_LOG_ORIGIN "r2ai"
 
@@ -24,22 +24,46 @@ R_API void r2ai_apikeys_edit(RCorePluginSession *cps) {
 	free (config_dir);
 	if (!exists) {
 		const char *template = "# API Keys configuration for r2ai\n"
-				"\n"
-				"# Specify your API keys in the format: PROVIDER=KEY\n"
-				"# OpenAI=sk-your-openai-api-key-here\n"
-				"# MISTRAL=your-mistral-api-key-here\n";
+			"\n"
+			"# Specify your API keys in the format: PROVIDER=KEY\n"
+			"# OpenAI=sk-your-openai-api-key-here\n"
+			"# MISTRAL=your-mistral-api-key-here\n";
 		r_file_dump (keys_path, (const ut8 *)template, strlen (template), 0);
 	}
-	r_cons_editor (core->cons, keys_path, NULL);
+	r2ai_cons_editor (core->cons, keys_path, NULL);
 	free (keys_path);
 }
 
-/* Return a malloc'd API key read from ~/.config/r2ai/apikeys.txt
- * Provider matching is case-insensitive */
+static char *r2ai_apikeys_env(const char *provider) {
+	char *upper = strdup (provider);
+	r_str_case (upper, true);
+	char *name = r_str_newf ("%s_API_KEY", upper);
+	char *val = r_sys_getenv (name);
+	if (val) {
+		r_str_trim (val);
+	}
+	free (name);
+	free (upper);
+	if (R_STR_ISNOTEMPTY (val)) {
+		return val;
+	}
+	free (val);
+	return NULL;
+}
+
+/* Return a malloc'd API key for a provider.
+ * Resolution order:
+ *   1. environment variable "<PROVIDER>_API_KEY" (provider upper-cased)
+ *   2. the ~/.config/r2ai/apikeys.txt file
+ * Provider matching against the file is case-insensitive. */
 R_API char *r2ai_apikeys_get(const char *provider) {
 	if (!provider) {
 		return NULL;
 	}
+	char *env_key = r2ai_apikeys_env (provider);
+	if (env_key) {
+		return env_key;
+	}
 	bool exists = false;
 	char *config_file = r2ai_apikeys_path (&exists);
 	if (!exists) {

src/async.c

@@ -0,0 +1,870 @@
+/* Copyright r2ai - 2023-2026 - pancake */
+
+#define R_LOG_ORIGIN "r2ai.async"
+
+#include "r2ai.h"
+#include "r2ai_priv.h"
+#include <r_th.h>
+
+static const char *state_name(R2AITaskState s) {
+	switch (s) {
+	case R2AI_TASK_PENDING: return "pending";
+	case R2AI_TASK_RUNNING: return "running";
+	case R2AI_TASK_WAIT_APPROVE: return "wait-approve";
+	case R2AI_TASK_WAIT_INPUT: return "wait-input";
+	case R2AI_TASK_COMPLETE: return "complete";
+	case R2AI_TASK_ERROR: return "error";
+	case R2AI_TASK_CANCELLED: return "cancelled";
+	}
+	return "?";
+}
+
+static const char *kind_name(R2AITaskKind k) {
+	return k == R2AI_TASK_AUTO? "auto": "query";
+}
+
+static void task_lock(R2AITask *t) {
+	r_th_lock_enter (t->lock);
+}
+static void task_unlock(R2AITask *t) {
+	r_th_lock_leave (t->lock);
+}
+
+static bool task_is_live_locked(const R2AITask *t) {
+	R2AITaskState s = t->state;
+	return s == R2AI_TASK_PENDING || s == R2AI_TASK_RUNNING || s == R2AI_TASK_WAIT_APPROVE || s == R2AI_TASK_WAIT_INPUT;
+}
+
+static bool is_done(R2AITaskState st) {
+	return st == R2AI_TASK_COMPLETE || st == R2AI_TASK_ERROR || st == R2AI_TASK_CANCELLED;
+}
+
+static R2AITaskQueue *queue(RCorePluginSession *cps) {
+	return ((R2AI_State *)cps->data)->async;
+}
+
+static void print_block(RCons *cons, const char *hdr, const char *text) {
+	if (R_STR_ISEMPTY (text)) {
+		return;
+	}
+	r_cons_printf (cons, "%s%s", hdr, text);
+	if (!r_str_endswith (text, "\n")) {
+		r_cons_newline (cons);
+	}
+}
+
+/* Append text to task->output. Takes task lock. */
+static void task_append_output(R2AITask *t, const char *text) {
+	if (R_STR_ISEMPTY (text)) {
+		return;
+	}
+	task_lock (t);
+	r_strbuf_append (t->output, text);
+	task_unlock (t);
+}
+
+static void task_append_outputf(R2AITask *t, const char *fmt, ...) {
+	va_list ap;
+	va_start (ap, fmt);
+	char *s = r_str_newvf (fmt, ap);
+	va_end (ap);
+	task_append_output (t, s);
+	free (s);
+}
+
+static void pending_clear(R2AITask *t) {
+	R_FREE (t->pending_tool_name);
+	R_FREE (t->pending_tool_args);
+	R_FREE (t->pending_tool_call_id);
+}
+
+static void task_free(R2AITask *t) {
+	if (!t) {
+		return;
+	}
+	if (t->thread) {
+		r_th_wait (t->thread);
+		r_th_free (t->thread);
+	}
+	if (t->gate) {
+		r_th_sem_free (t->gate);
+	}
+	if (t->lock) {
+		r_th_lock_free (t->lock);
+	}
+	if (t->messages) {
+		r2ai_msgs_free (t->messages);
+	}
+	r_list_free (t->tools);
+	r_config_free (t->llm_core->config);
+	free (t->llm_core);
+	free (t->title);
+	free (t->query);
+	free (t->system_prompt);
+	free (t->model);
+	free (t->provider);
+	r_strbuf_free (t->output);
+	free (t->error);
+	pending_clear (t);
+	free (t->tool_result);
+	free (t);
+}
+
+static void queue_lock(R2AITaskQueue *q) {
+	r_th_lock_enter (q->lock);
+}
+static void queue_unlock(R2AITaskQueue *q) {
+	r_th_lock_leave (q->lock);
+}
+
+static void task_start(R2AITask *t) {
+	task_lock (t);
+	t->state = R2AI_TASK_RUNNING;
+	t->started = time (NULL);
+	task_unlock (t);
+}
+
+static bool cancelled(R2AITask *t) {
+	task_lock (t);
+	bool res = t->cancel_req;
+	task_unlock (t);
+	return res;
+}
+
+// terminate the worker, an earlier error reported by the llm call takes precedence
+static RThreadFunctionRet finish(R2AITask *t, R2AITaskState st, char *err) {
+	task_lock (t);
+	if (err && !t->error) {
+		t->error = err;
+	} else {
+		free (err);
+	}
+	t->state = st;
+	t->finished = time (NULL);
+	task_unlock (t);
+	return R_TH_STOP;
+}
+
+/* Providers read core->config throughout a request. The HTTP server swaps
+ * that pointer between its web and shell contexts, so use the task snapshot.
+ * Tool execution still uses the original session on the main thread. */
+static R2AI_ChatResponse *run_llm_once(R2AITask *t) {
+	char *error = NULL;
+	R2AIArgs args = {
+		.messages = t->messages,
+		.system_prompt = t->system_prompt,
+		.model = t->model,
+		.provider = t->provider,
+		.tools = t->tools,
+		.error = &error,
+	};
+	RCorePluginSession worker_session = *t->cps;
+	worker_session.core = t->llm_core;
+	R2AI_ChatResponse *res = r2ai_llmcall (&worker_session, args);
+	if (error) {
+		task_lock (t);
+		free (t->error);
+		t->error = error;
+		task_unlock (t);
+	}
+	return res;
+}
+
+static void dump_message(R2AITask *t, const R2AI_Message *m) {
+	if (m->reasoning_content) {
+		task_append_outputf (t, "<thinking>\n%s\n</thinking>\n", m->reasoning_content);
+	}
+	if (m->content) {
+		task_append_outputf (t, "%s\n", m->content);
+	}
+}
+
+static RThreadFunctionRet worker_query(RThread *th) {
+	R2AITask *t = (R2AITask *)th->user;
+	task_start (t);
+	R2AI_ChatResponse *res = run_llm_once (t);
+	bool ok = res && res->message;
+	bool cancel = cancelled (t);
+	if (ok && !cancel) {
+		dump_message (t, res->message);
+	}
+	r2ai_chat_response_free (res);
+	if (cancel) {
+		return finish (t, R2AI_TASK_CANCELLED, NULL);
+	}
+	return ok
+		? finish (t, R2AI_TASK_COMPLETE, NULL)
+		: finish (t, R2AI_TASK_ERROR, strdup ("llm call returned no response"));
+}
+
+static R2AI_ToolCall *first_tool_call(const R2AI_Message *m) {
+	RListIter *iter;
+	R2AI_ToolCall *tc;
+	r_list_foreach (m->tool_calls, iter, tc) {
+		if (tc->name && tc->arguments && tc->id) {
+			return tc;
+		}
+	}
+	return NULL;
+}
+
+// hand the tool call to the main thread and block until it answers, returns false if cancelled
+static bool wait_tool(R2AITask *t, const R2AI_ToolCall *tc) {
+	task_lock (t);
+	pending_clear (t);
+	R_FREE (t->tool_result);
+	t->pending_tool_name = strdup (tc->name);
+	t->pending_tool_args = strdup (tc->arguments);
+	t->pending_tool_call_id = strdup (tc->id);
+	t->state = R2AI_TASK_WAIT_APPROVE;
+	task_unlock (t);
+
+	r_th_sem_wait (t->gate);
+
+	task_lock (t);
+	if (t->cancel_req) {
+		task_unlock (t);
+		return false;
+	}
+	R2AI_Message msg = {
+		.role = "tool",
+		.tool_call_id = t->pending_tool_call_id,
+		.content = t->tool_result? t->tool_result: "<no output>",
+	};
+	r2ai_msgs_add (t->messages, &msg);
+	R_FREE (t->tool_result);
+	pending_clear (t);
+	task_unlock (t);
+	return true;
+}
+
+static RThreadFunctionRet worker_auto(RThread *th) {
+	R2AITask *t = (R2AITask *)th->user;
+	task_start (t);
+	for (;;) {
+		if (cancelled (t)) {
+			return finish (t, R2AI_TASK_CANCELLED, NULL);
+		}
+		if (t->steps >= t->max_runs) {
+			return finish (t, R2AI_TASK_ERROR, r_str_newf ("max runs (%d) reached", t->max_runs));
+		}
+		task_lock (t);
+		t->steps++;
+		task_unlock (t);
+
+		R2AI_ChatResponse *res = run_llm_once (t);
+		if (!res || !res->message) {
+			r2ai_chat_response_free (res);
+			return finish (t, R2AI_TASK_ERROR, strdup ("llm call returned no response"));
+		}
+		const R2AI_Message *m = res->message;
+		dump_message (t, m);
+		r2ai_msgs_add (t->messages, m);
+		if (r_list_empty (m->tool_calls)) {
+			r2ai_chat_response_free (res);
+			return finish (t, R2AI_TASK_COMPLETE, NULL);
+		}
+		R2AI_ToolCall *tc = first_tool_call (m);
+		bool ok = tc && wait_tool (t, tc);
+		r2ai_chat_response_free (res);
+		if (!ok) {
+			return tc
+				? finish (t, R2AI_TASK_CANCELLED, NULL)
+				: finish (t, R2AI_TASK_ERROR, strdup ("llm returned an invalid tool call"));
+		}
+	}
+}
+
+R_IPI void r2ai_async_init(R2AI_State *state) {
+	if (!state || state->async) {
+		return;
+	}
+	R2AITaskQueue *q = R_NEW0 (R2AITaskQueue);
+	q->tasks = r_list_newf ((RListFree)task_free);
+	q->lock = r_th_lock_new (true);
+	q->next_id = 1;
+	state->async = q;
+}
+
+R_IPI void r2ai_async_fini(R2AI_State *state) {
+	if (!state || !state->async) {
+		return;
+	}
+	R2AITaskQueue *q = state->async;
+	int killed = 0;
+	/* Force-kill any live workers so we don't block on inflight HTTP. */
+	queue_lock (q);
+	RListIter *it;
+	R2AITask *t;
+	r_list_foreach (q->tasks, it, t) {
+		task_lock (t);
+		bool live = task_is_live_locked (t);
+		t->cancel_req = true;
+		/* take ownership of the thread while holding the task lock */
+		RThread *th = live? t->thread: NULL;
+		if (th) {
+			t->thread = NULL;
+		}
+		task_unlock (t);
+		if (t->gate) {
+			r_th_sem_post (t->gate);
+		}
+		if (th) {
+			/* kill_free joins the thread: never do it under the task lock */
+			r_th_kill_free (th);
+			killed++;
+		}
+	}
+	queue_unlock (q);
+	if (killed > 0) {
+		R_LOG_INFO ("killed %d pending async task%s", killed, killed == 1? "": "s");
+	}
+	r_list_free (q->tasks);
+	r_th_lock_free (q->lock);
+	free (q);
+	state->async = NULL;
+}
+
+static R2AITask *task_new(RCorePluginSession *cps, R2AITaskKind kind, const char *title, const char *query, const char *system_prompt) {
+	RCore *core = cps->core;
+	R2AITask *t = R_NEW0 (R2AITask);
+	t->cps = cps;
+	/* Snapshot on the main thread, before it can mutate the shared core.
+	 * Only config is owned; providers borrow the remaining session fields. */
+	t->llm_core = R_NEW (RCore);
+	*t->llm_core = *core;
+	t->llm_core->config = r_config_clone (core->config);
+	t->kind = kind;
+	t->state = R2AI_TASK_PENDING;
+	t->title = strdup (title? title: "");
+	t->query = strdup (query? query: "");
+	// resolve the system prompt here, workers must not read the config
+	if (R_STR_ISEMPTY (system_prompt) && kind == R2AI_TASK_AUTO) {
+		t->system_prompt = r2ai_auto_system_prompt (cps);
+	} else {
+		t->system_prompt = r_str_new (r2ai_system_prompt (core, system_prompt));
+	}
+	const char *m = r_config_get (core->config, "r2ai.model");
+	const char *p = r_config_get (core->config, "r2ai.api");
+	t->model = m? strdup (m): NULL;
+	t->provider = p? strdup (p): NULL;
+	if (kind == R2AI_TASK_AUTO) {
+		RList *tools = r2ai_get_tools (core, cps->data);
+		t->tools = tools? r_list_clone (tools, NULL): NULL;
+		t->max_runs = r_config_get_i (core->config, "r2ai.auto.max_runs");
+	}
+	t->messages = r2ai_msgs_new ();
+	R2AI_Message um = { .role = "user", .content = (char *)query };
+	r2ai_msgs_add (t->messages, &um);
+	t->lock = r_th_lock_new (false);
+	t->gate = r_th_sem_new (0);
+	t->output = r_strbuf_new (NULL);
+	t->created = time (NULL);
+	return t;
+}
+
+static int queue_register(R2AITaskQueue *q, R2AITask *t) {
+	queue_lock (q);
+	t->id = q->next_id++;
+	r_list_append (q->tasks, t);
+	queue_unlock (q);
+	return t->id;
+}
+
+static int submit(RCorePluginSession *cps, R2AITaskKind kind, const char *title, const char *query, const char *sysp, RThreadFunction fn) {
+	R2AITask *t = task_new (cps, kind, title, query, sysp);
+	int id = queue_register (queue (cps), t);
+	t->thread = r_th_new (fn, t, 0);
+	if (t->thread) {
+		r_th_start (t->thread);
+	} else {
+		finish (t, R2AI_TASK_ERROR, strdup ("failed to spawn worker"));
+	}
+	r_cons_printf (cps->core->cons, "[async] task %d queued (%s)\n", id, t->title);
+	return id;
+}
+
+R_IPI int r2ai_async_query(RCorePluginSession *cps,
+	const char *title,
+	const char *query,
+	const char *sysp) {
+	return submit (cps, R2AI_TASK_QUERY, title, query, sysp, worker_query);
+}
+
+R_IPI int r2ai_async_auto(RCorePluginSession *cps,
+	const char *title,
+	const char *query,
+	const char *sysp) {
+	return submit (cps, R2AI_TASK_AUTO, title, query, sysp, worker_auto);
+}
+
+static void purge_finished(RCorePluginSession *cps);
+
+static void show_task_list(RCorePluginSession *cps, bool json) {
+	RCore *core = cps->core;
+	R2AITaskQueue *q = queue (cps);
+	queue_lock (q);
+	if (json) {
+		PJ *pj = r_core_pj_new (cps->core);
+		pj_o (pj);
+		pj_ka (pj, "tasks");
+		RListIter *it;
+		R2AITask *t;
+		r_list_foreach (q->tasks, it, t) {
+			task_lock (t);
+			pj_o (pj);
+			pj_ki (pj, "id", t->id);
+			pj_ks (pj, "kind", kind_name (t->kind));
+			pj_ks (pj, "state", state_name (t->state));
+			pj_ks (pj, "title", t->title? t->title: "");
+			pj_ki (pj, "steps", t->steps);
+			pj_ki (pj, "age", (int) (time (NULL) - t->created));
+			if (t->pending_tool_name) {
+				pj_ks (pj, "pending_tool", t->pending_tool_name);
+			}
+			if (t->pending_tool_args) {
+				pj_ks (pj, "pending_tool_args", t->pending_tool_args);
+			}
+			if (t->pending_tool_call_id) {
+				pj_ks (pj, "pending_tool_call_id", t->pending_tool_call_id);
+			}
+			const char *out = r_strbuf_get (t->output);
+			if (R_STR_ISNOTEMPTY (out)) {
+				pj_ks (pj, "output", out);
+			}
+			if (t->error) {
+				pj_ks (pj, "error", t->error);
+			}
+			pj_end (pj);
+			task_unlock (t);
+		}
+		pj_end (pj);
+		pj_end (pj);
+		char *s = pj_drain (pj);
+		r_cons_println (core->cons, s);
+		free (s);
+	} else if (r_list_empty (q->tasks)) {
+		r_cons_printf (core->cons, "No async tasks\n");
+	} else {
+		r_cons_printf (core->cons, "%-4s %-6s %-13s %-4s  %s\n", "id", "kind", "state", "age", "title");
+		RListIter *it;
+		R2AITask *t;
+		r_list_foreach (q->tasks, it, t) {
+			task_lock (t);
+			int age = (int) (time (NULL) - t->created);
+			const char *extra = "";
+			char *extrabuf = NULL;
+			if (t->state == R2AI_TASK_WAIT_APPROVE && t->pending_tool_name) {
+				extrabuf = r_str_newf (" [awaiting tool: %s]", t->pending_tool_name);
+				extra = extrabuf;
+			} else if (t->state == R2AI_TASK_ERROR && t->error) {
+				extrabuf = r_str_newf (" [err: %s]", t->error);
+				extra = extrabuf;
+			}
+			r_cons_printf (core->cons, "%-4d %-6s %-13s %-4d  %s%s\n", t->id, kind_name (t->kind), state_name (t->state), age, t->title? t->title: "", extra);
+			print_block (core->cons, "output:\n", r_strbuf_get (t->output));
+			free (extrabuf);
+			task_unlock (t);
+		}
+	}
+	if (r_config_get_b (core->config, "r2ai.async.purge")) {
+		purge_finished (cps);
+	}
+	queue_unlock (q);
+}
+
+static bool task_is_actionable_locked(const R2AITask *t) {
+	switch (t->state) {
+	case R2AI_TASK_COMPLETE:
+	case R2AI_TASK_ERROR:
+	case R2AI_TASK_WAIT_APPROVE:
+	case R2AI_TASK_WAIT_INPUT:
+		return true;
+	default:
+		return false;
+	}
+}
+
+/* Find first actionable task (id == 0) or the task with the given id if
+ * actionable. Returns with queue lock held; caller must release via
+ * queue_unlock. Returns NULL if none found (lock released). */
+static R2AITask *find_actionable(R2AITaskQueue *q, int id) {
+	queue_lock (q);
+	RListIter *it;
+	R2AITask *t;
+	r_list_foreach (q->tasks, it, t) {
+		if (id > 0 && t->id != id) {
+			continue;
+		}
+		task_lock (t);
+		bool actionable = task_is_actionable_locked (t);
+		task_unlock (t);
+		if (actionable) {
+			return t;
+		}
+	}
+	queue_unlock (q);
+	return NULL;
+}
+
+static void task_unlink_disk(const R2AITask *t) {
+	char *path = r_file_homef (".config/r2ai/tasks/%d.json", t->id);
+	if (path) {
+		r_file_rm (path);
+		free (path);
+	}
+}
+
+static void drop_task_locked(R2AITaskQueue *q, R2AITask *t) {
+	task_unlink_disk (t);
+	r_list_delete_data (q->tasks, t);
+}
+
+static bool answer_wait_approve(RCorePluginSession *cps, R2AITask *t, bool approve, bool interactive) {
+	RCore *core = cps->core;
+	task_lock (t);
+	if (t->state != R2AI_TASK_WAIT_APPROVE) {
+		task_unlock (t);
+		return false;
+	}
+	char *tool_name = t->pending_tool_name? strdup (t->pending_tool_name): NULL;
+	char *tool_args = t->pending_tool_args? strdup (t->pending_tool_args): NULL;
+	task_unlock (t);
+
+	char *tool_output = NULL;
+	if (approve) {
+		bool old_yolo = r_config_get_b (core->config, "r2ai.auto.yolo");
+		r_config_set_b (core->config, "r2ai.auto.yolo", true);
+		R2AI_ToolResult tool_result = execute_tool (cps, tool_name, tool_args);
+		r_config_set_b (core->config, "r2ai.auto.yolo", old_yolo);
+		tool_output = tool_result.output;
+		tool_result.output = NULL;
+		r2ai_tool_result_fini (&tool_result);
+		if (!tool_output) {
+			tool_output = strdup ("<no output>");
+		}
+		task_append_outputf (t, "\nTool result (%s):\n%s\n", tool_name? tool_name: "?", tool_output);
+		if (interactive) {
+			r_cons_printf (core->cons, Color_GREEN "tool result:" Color_RESET " %s\n", tool_output);
+		}
+	} else {
+		tool_output = strdup ("<user declined to run tool>");
+		if (interactive) {
+			r_cons_printf (core->cons, "declined.\n");
+		} else {
+			task_append_output (t, "\nTool declined by user.\n");
+		}
+	}
+
+	task_lock (t);
+	free (t->tool_result);
+	t->tool_result = tool_output;
+	t->state = R2AI_TASK_RUNNING;
+	task_unlock (t);
+	r_th_sem_post (t->gate);
+
+	free (tool_name);
+	free (tool_args);
+	return true;
+}
+
+/* Interactive handler: act on first actionable task (or the given id). */
+static void interact_once(RCorePluginSession *cps, int id) {
+	RCore *core = cps->core;
+	R2AITaskQueue *q = queue (cps);
+	R2AITask *t = find_actionable (q, id);
+	if (!t) {
+		/* Nothing to do - silent by default so cmd.prompt stays clean. */
+		return;
+	}
+	/* queue lock still held */
+	task_lock (t);
+	R2AITaskState st = t->state;
+	int tid = t->id;
+	char *output = r_strbuf_drain_nofree (t->output);
+	char *err = t->error;
+	t->error = NULL;
+	char *tool_name = t->pending_tool_name? strdup (t->pending_tool_name): NULL;
+	char *tool_args = t->pending_tool_args? strdup (t->pending_tool_args): NULL;
+	R2AITaskKind kind = t->kind;
+	task_unlock (t);
+	queue_unlock (q);
+
+	r_cons_printf (core->cons, "\n" Color_BLUE "[async task %d | %s | %s]" Color_RESET "\n", tid, kind_name (kind), state_name (st));
+	print_block (core->cons, "", output);
+	free (output);
+	if (err) {
+		r_cons_printf (core->cons, Color_RED "error: %s" Color_RESET "\n", err);
+		free (err);
+	}
+
+	if (is_done (st)) {
+		/* Remove the task from the queue (join the thread on free). */
+		queue_lock (q);
+		drop_task_locked (q, t);
+		queue_unlock (q);
+		free (tool_name);
+		free (tool_args);
+		return;
+	}
+
+	if (st == R2AI_TASK_WAIT_APPROVE) {
+		r_cons_printf (core->cons, Color_YELLOW "pending tool:" Color_RESET " %s\n", tool_name? tool_name: "?");
+		if (tool_args) {
+			r_cons_printf (core->cons, "args: %s\n", tool_args);
+		}
+		r_cons_flush (core->cons);
+
+		bool yolo = r_config_get_b (core->config, "r2ai.auto.yolo");
+		bool approve = yolo || r_cons_yesno (core->cons, 'y', "Run tool %s? (Y/n)", tool_name? tool_name: "?");
+		answer_wait_approve (cps, t, approve, true);
+	}
+	free (tool_name);
+	free (tool_args);
+}
+
+// lock the queue and find the task, reports and unlocks when it does not exist
+static R2AITask *lookup(RCorePluginSession *cps, int id) {
+	R2AITaskQueue *q = queue (cps);
+	queue_lock (q);
+	RListIter *it;
+	R2AITask *t;
+	r_list_foreach (q->tasks, it, t) {
+		if (t->id == id) {
+			return t;
+		}
+	}
+	queue_unlock (q);
+	r_cons_printf (cps->core->cons, "No task with id %d\n", id);
+	return NULL;
+}
+
+static void answer_by_id(RCorePluginSession *cps, int id, bool approve) {
+	RCore *core = cps->core;
+	if (id <= 0) {
+		r_cons_printf (core->cons, "Missing task id\n");
+		return;
+	}
+	R2AITaskQueue *q = queue (cps);
+	R2AITask *t = lookup (cps, id);
+	if (!t) {
+		return;
+	}
+	task_lock (t);
+	bool wait_approve = t->state == R2AI_TASK_WAIT_APPROVE;
+	task_unlock (t);
+	queue_unlock (q);
+
+	if (!wait_approve) {
+		r_cons_printf (core->cons, "Task %d is not waiting for approval\n", id);
+		return;
+	}
+	if (answer_wait_approve (cps, t, approve, false)) {
+		r_cons_printf (core->cons, "%s task %d\n", approve? "Approved": "Declined", id);
+	}
+}
+
+static void kill_task_locked(R2AITaskQueue *q, R2AITask *t) {
+	task_lock (t);
+	t->cancel_req = true;
+	bool live = task_is_live_locked (t);
+	task_unlock (t);
+	if (t->gate) {
+		r_th_sem_post (t->gate);
+	}
+	if (live && t->thread) {
+		r_th_kill_free (t->thread);
+		t->thread = NULL;
+	}
+	drop_task_locked (q, t);
+}
+
+static void kill_by_id(RCorePluginSession *cps, int id) {
+	RCore *core = cps->core;
+	R2AITaskQueue *q = queue (cps);
+	R2AITask *t = lookup (cps, id);
+	if (!t) {
+		return;
+	}
+	kill_task_locked (q, t);
+	queue_unlock (q);
+	r_cons_printf (core->cons, "Killed task %d\n", id);
+}
+
+static void kill_all(RCorePluginSession *cps) {
+	RCore *core = cps->core;
+	R2AITaskQueue *q = queue (cps);
+	queue_lock (q);
+	int n = 0;
+	while (!r_list_empty (q->tasks)) {
+		R2AITask *t = r_list_first (q->tasks);
+		kill_task_locked (q, t);
+		n++;
+	}
+	queue_unlock (q);
+	r_cons_printf (core->cons, "Killed %d task%s\n", n, n == 1? "": "s");
+}
+
+static void show_task_by_id(RCorePluginSession *cps, int id) {
+	RCore *core = cps->core;
+	R2AITaskQueue *q = queue (cps);
+	R2AITask *t = lookup (cps, id);
+	if (!t) {
+		return;
+	}
+	task_lock (t);
+	r_cons_printf (core->cons, "id:     %d\nkind:   %s\nstate:  %s\ntitle:  %s\nmodel:  %s\nprov:   %s\nsteps:  %d\nage:    %ds\n",
+		t->id, kind_name (t->kind), state_name (t->state), t->title, r_str_get (t->model),
+		r_str_get (t->provider), t->steps, (int) (time (NULL) - t->created));
+	if (t->pending_tool_name) {
+		r_cons_printf (core->cons, "tool:   %s %s\n", t->pending_tool_name, t->pending_tool_args? t->pending_tool_args: "");
+	}
+	if (t->error) {
+		r_cons_printf (core->cons, "error:  %s\n", t->error);
+	}
+	print_block (core->cons, "output:\n", r_strbuf_get (t->output));
+	task_unlock (t);
+	queue_unlock (q);
+}
+
+static void show_last_task(RCorePluginSession *cps) {
+	RCore *core = cps->core;
+	R2AITaskQueue *q = queue (cps);
+	queue_lock (q);
+	R2AITask *t = r_list_last (q->tasks);
+	int id = t? t->id: 0;
+	queue_unlock (q);
+	if (!id) {
+		r_cons_printf (core->cons, "No async tasks\n");
+		return;
+	}
+	show_task_by_id (cps, id);
+}
+
+// clang-format off
+static RCoreHelpMessage help_msg_r2ai_s = {
+	"Usage:", "r2ai", " -s[subcmd]",
+	"r2ai", " -s", "list async tasks and completed output",
+	"r2ai", " -s?", "show this help",
+	"r2ai", " -sj", "list tasks as json, including completed output",
+	"r2ai", " -ss", "show details of the last created task",
+	"r2ai", " -si", "interactive: handle first actionable task",
+	"r2ai", " -si <id>", "interactive: handle only task <id>",
+	"r2ai", " -sy <id>", "approve pending tool call for task <id>",
+	"r2ai", " -sn <id>", "decline pending tool call for task <id>",
+	"r2ai", " -sa", "block until all tasks finish",
+	"r2ai", " -sp", "purge finished/errored/cancelled tasks",
+	"r2ai", " -s <id>", "show details of task <id>",
+	"r2ai", " -sk", "kill all tasks",
+	"r2ai", " -sk <id>", "kill task <id>",
+	NULL
+};
+
+static RCoreHelpMessage help_msg_r2ai_s_kill = {
+	"r2ai", " -sk", "kill all tasks",
+	"r2ai", " -sk <id>", "kill task <id>",
+	NULL
+};
+// clang-format on
+
+static void show_help(RCorePluginSession *cps) {
+	RCore *core = cps->core;
+	r2ai_cmd_help (core, help_msg_r2ai_s);
+}
+
+static void purge_finished(RCorePluginSession *cps) {
+	R2AITaskQueue *q = queue (cps);
+	queue_lock (q);
+	RListIter *it, *tmp;
+	R2AITask *t;
+	r_list_foreach_safe (q->tasks, it, tmp, t) {
+		task_lock (t);
+		bool drop = is_done (t->state);
+		task_unlock (t);
+		if (drop) {
+			drop_task_locked (q, t);
+		}
+	}
+	queue_unlock (q);
+}
+
+static void wait_all(RCorePluginSession *cps) {
+	R2AITaskQueue *q = queue (cps);
+	for (;;) {
+		bool any = false;
+		queue_lock (q);
+		RListIter *it;
+		R2AITask *t;
+		r_list_foreach (q->tasks, it, t) {
+			task_lock (t);
+			bool busy = t->state == R2AI_TASK_PENDING || t->state == R2AI_TASK_RUNNING;
+			task_unlock (t);
+			if (busy) {
+				any = true;
+				break;
+			}
+		}
+		queue_unlock (q);
+		if (!any) {
+			break;
+		}
+		r_sys_usleep (100 * 1000); /* 100ms */
+	}
+}
+
+R_IPI void r2ai_async_cmd(RCorePluginSession *cps, const char *input) {
+	RCore *core = cps->core;
+	const char *a = input? input: "";
+	const char *arg = r_str_trim_head_ro (a + 1);
+	int id = R_STR_ISEMPTY (arg)? 0: r_num_math (core->num, arg);
+	switch (*a) {
+	case '?':
+		show_help (cps);
+		break;
+	case 'j':
+		show_task_list (cps, true);
+		break;
+	case 's':
+		show_last_task (cps);
+		break;
+	case 'i':
+		interact_once (cps, id);
+		break;
+	case 'y':
+		answer_by_id (cps, id, true);
+		break;
+	case 'n':
+		answer_by_id (cps, id, false);
+		break;
+	case 'a':
+		wait_all (cps);
+		break;
+	case 'p':
+		purge_finished (cps);
+		break;
+	case 'k':
+		if (R_STR_ISEMPTY (arg)) {
+			kill_all (cps);
+		} else {
+			if (id > 0) {
+				kill_by_id (cps, id);
+			} else {
+				r2ai_cmd_help (core, help_msg_r2ai_s_kill);
+			}
+		}
+		break;
+	case ' ':
+		if (id > 0) {
+			show_task_by_id (cps, id);
+		} else {
+			show_task_list (cps, false);
+		}
+		break;
+	case 0:
+		show_task_list (cps, false);
+		break;
+	default:
+		r_core_return_invalid_command (core, "-s", *a);
+		break;
+	}
+}

src/auto.c

@@ -1,4 +1,4 @@
-/* Copyright r2ai - 2023-2025 - pancake */
+/* Copyright r2ai - 2023-2026 - pancake */
 
 #include "r2ai.h"
 
@@ -56,16 +56,7 @@ static void r2ai_print_run_end(RCorePluginSession *cps, const R2AI_Usage *usage,
 	char *total_time_str = format_time_duration (total_time);
 
 	// Print detailed stats
-	r_cons_printf (core->cons, "\x1b[1" Color_BLUE "%s | total: %d in: %d out: %d | run: %d in: %d out: %d | %s / %s" Color_RESET "\n",
-		r_config_get (core->config, "r2ai.model"),
-		state->stats.total_tokens,
-		state->stats.total_prompt_tokens,
-		state->stats.total_completion_tokens,
-		state->stats.run_tokens,
-		state->stats.run_prompt_tokens,
-		state->stats.run_completion_tokens,
-		run_time_str,
-		total_time_str);
+	r_cons_printf (core->cons, Color_BOLD Color_BLUE "%s | total: %d in: %d out: %d | run: %d in: %d out: %d | %s / %s" Color_RESET "\n", r_config_get (core->config, "r2ai.model"), state->stats.total_tokens, state->stats.total_prompt_tokens, state->stats.total_completion_tokens, state->stats.run_tokens, state->stats.run_prompt_tokens, state->stats.run_completion_tokens, run_time_str, total_time_str);
 	r_cons_newline (core->cons);
 	r_cons_flush (core->cons);
 
@@ -96,14 +87,13 @@ static void handle_final_response_attempt(RCorePluginSession *cps, RList *messag
 				r_cons_flush (core->cons);
 			}
 			if (final_msg->content) {
-				r_cons_printf (core->cons, "%s", final_msg->content);
-				r_cons_newline (core->cons);
+				r2ai_print_response (core, final_msg->content);
 				r_cons_flush (core->cons);
 			}
 			// Add final response to messages for completeness
 			r2ai_msgs_add (messages, final_msg);
-			free (final_response);
 		}
+		r2ai_chat_response_free (final_response);
 		free (final_system_prompt);
 	} else {
 		r_cons_printf (core->cons, "Auto mode interrupted without final response.\n");
@@ -148,44 +138,22 @@ static void update_tool_call_argument(R2AI_Message *msg, const char *tool_call_i
 	}
 }
 
-#if 0
-static const char *Gprompt_auto =
-	"You are a reverse engineer using radare2.\n"
-	"Answer questions about the loaded binary using r2cmd tool calls.\n"
-	"\n"
-	"Rules:\n"
-	"- Use r2cmd tool to run radare2 commands when you need information\n"
-	"- Make tool calls immediately when you need data\n"
-	"- Provide final answer only after gathering needed information\n"
-	"- Keep responses short and direct\n"
-	;
-#else
-const char *Gprompt_auto =
-	"You are a reverse engineer and you are using radare2 to analyze a binary.\n"
-	"The user will ask questions about the binary and you will respond with the answer to the best of your ability.\n"
-	"\n"
-	"# Guidelines\n"
-	"- Understand the Task: Grasp the main objective, goals, requirements, constraints, and expected output.\n"
-	"- Reasoning Before Conclusions**: Encourage reasoning steps before any conclusions are reached.\n"
-	"- Assume the user is always asking you about the binary, unless they're specifically asking you for radare2 help.\n"
-	"- The binary has already been loaded. You can interact with the binary using the r2cmd tool.\n"
-	"- `this` or `here` might refer to the current address in the binary or the binary itself.\n"
-	"- If you need more information, try to use the r2cmd tool to run commands before answering.\n"
-	"- You can use the r2cmd tool multiple times if you need or you can pass a command with pipes if you need to chain commands.\n"
-	"- IMPORTANT: You must make only ONE tool call per response. Do not make multiple tool calls in a single response.\n"
-	"- Keep your reasoning concise and avoid repeating the same thoughts. If you have a clear next step, execute it immediately with a tool call.\n"
-	"- If you're asked to decompile a function, make sure to return the code in the language you think it was originally written "
-	"and rewrite it to be as easy as possible to be understood. Make sure you use descriptive variable and function names and add comments.\n"
-	"- Don't just regurgitate the same code, figure out what it's doing and rewrite it to be more understandable.\n"
-	"- If you need to run a command in r2 before answering, you can use the r2cmd tool\n"
-	"- Do not repeat commands if you already know the answer.\n"
-	"- Formulate a plan. Think step by step. Analyze the binary as much as possible before answering.\n"
-	"- You must keep going until you have a final answer.\n"
-	"- Double check that final answer. Make sure you didn't miss anything.\n"
-	"- Make sure you call tools and functions correctly.\n"
-	"- When calling tools, use the exact format: {\"tool_calls\": [{\"id\": \"call_123\", \"type\": \"function\", \"function\": {\"name\": \"r2cmd\", \"arguments\": \"{\\\"command\\\":\\\"aa\\\"}\"}}]}\n"
-	"- Arguments must be a JSON string, not an object.\n";
-#endif
+static const char *auto_prompt_get(R2AI_State *state) {
+	return (state && R_STR_ISNOTEMPTY (state->prompt_auto))? state->prompt_auto: "";
+}
+
+R_IPI char *r2ai_auto_system_prompt(RCorePluginSession *cps) {
+	RCore *core = cps->core;
+	R2AI_State *state = cps->data;
+	const char *base_system = r_config_get (core->config, "r2ai.system");
+	const char *prompt_auto = auto_prompt_get (state);
+	const char *think_prefix = r_config_get_b (core->config, "r2ai.auto.think")
+		? ""
+		: "/no_think\nReasoning: Low\n";
+	return R_STR_ISNOTEMPTY (base_system)
+		? r_str_newf ("%s%s\n\n%s", think_prefix, base_system, prompt_auto)
+		: r_str_newf ("%s%s", think_prefix, prompt_auto);
+}
 
 // Helper function to process messages and handle tool calls recursively
 R_API void process_messages(RCorePluginSession *cps, RList *messages, const char *system_prompt, int n_run) {
@@ -201,18 +169,22 @@ R_API void process_messages(RCorePluginSession *cps, RList *messages, const char
 		return;
 	}
 
-	const char *effective_prompt = system_prompt? system_prompt: Gprompt_auto;
+	const char *effective_prompt = system_prompt? system_prompt: auto_prompt_get (state);
 	if (!system_prompt) {
 		const char *init_commands = r_config_get (core->config, "r2ai.auto.init_commands");
 		if (R_STR_ISNOTEMPTY (init_commands)) {
-			char *edited_command = NULL;
-			char *comment = NULL;
-			char *cmd_output = execute_tool (core, "r2cmd", r_str_newf ("{\"command\":\"%s\"}", init_commands), &edited_command, &comment);
-			if (R_STR_ISNOTEMPTY (cmd_output)) {
+			PJ *pj = pj_new ();
+			pj_o (pj);
+			pj_ks (pj, "command", init_commands);
+			pj_end (pj);
+			char *tool_args = pj_drain (pj);
+			R2AI_ToolResult tool_result = execute_tool (cps, "r2cmd", tool_args);
+			free (tool_args);
+			if (R_STR_ISNOTEMPTY (tool_result.output)) {
 				char *display_command = strip_command_comment (init_commands, NULL);
-				char *content = r_str_newf ("Here is some information about the binary to get you started:\n>%s\n%s", display_command, cmd_output);
-				if (comment && *comment) {
-					char *new_content = r_str_newf ("%s\nHINT: %s", content, comment);
+				char *content = r_str_newf ("Here is some information about the binary to get you started:\n>%s\n%s", display_command, tool_result.output);
+				if (tool_result.comment && *tool_result.comment) {
+					char *new_content = r_str_newf ("%s\nHINT: %s", content, tool_result.comment);
 					free (content);
 					content = new_content;
 				}
@@ -222,16 +194,14 @@ R_API void process_messages(RCorePluginSession *cps, RList *messages, const char
 				};
 				r2ai_msgs_add (messages, &init_msg);
 				free (display_command);
-				free (cmd_output);
 			}
-			free (edited_command);
-			free (comment);
+			r2ai_tool_result_fini (&tool_result);
 		}
 	}
 
 	r2ai_stats_init_run (state, n_run);
 
-	r_cons_printf (core->cons, Color_BLUE "About to call r2ai_llmcall with n_run=%d%s\n", n_run, Color_RESET);
+	R_LOG_DEBUG ("About to call r2ai_llmcall with n_run=%d", n_run);
 	r_cons_flush (core->cons);
 
 	// Set up args for r2ai_llmcall call with tools directly
@@ -255,7 +225,7 @@ R_API void process_messages(RCorePluginSession *cps, RList *messages, const char
 
 	if (!message) {
 		R_LOG_ERROR ("No message in response");
-		free (response);
+		r2ai_chat_response_free (response);
 		handle_final_response_attempt (cps, messages, effective_prompt, &error);
 		return;
 	}
@@ -288,8 +258,7 @@ R_API void process_messages(RCorePluginSession *cps, RList *messages, const char
 			r_cons_flush (core->cons);
 		}
 		if (message->content) {
-			r_cons_printf (core->cons, "%s", message->content);
-			r_cons_newline (core->cons);
+			r2ai_print_response (core, message->content);
 			r_cons_flush (core->cons);
 		}
 	}
@@ -329,32 +298,36 @@ R_API void process_messages(RCorePluginSession *cps, RList *messages, const char
 			if (interrupted) {
 				cmd_output = strdup ("<user interrupted>");
 			} else {
-				char *edited_command = NULL;
-				cmd_output = execute_tool (core, tool_name, tool_args, &edited_command, &comment);
-				if (edited_command) {
+				R2AI_ToolResult tool_result = execute_tool (cps, tool_name, tool_args);
+				cmd_output = tool_result.output;
+				tool_result.output = NULL;
+				comment = tool_result.comment;
+				tool_result.comment = NULL;
+				if (tool_result.edited_command) {
 					// Update the last message's tool call arguments with the edited command
 					R2AI_Message *last_msg = r_list_get_n (messages, r_list_length (messages) - 1);
 					if (last_msg) {
 						if (!strcmp (tool_name, "r2cmd")) {
-							update_tool_call_argument (last_msg, tool_call->id, "command", edited_command);
+							update_tool_call_argument (last_msg, tool_call->id, "command", tool_result.edited_command);
 						} else if (!strcmp (tool_name, "execute_js")) {
-							update_tool_call_argument (last_msg, tool_call->id, "script", edited_command);
+							update_tool_call_argument (last_msg, tool_call->id, "script", tool_result.edited_command);
 						}
 					}
 				}
-				free (edited_command);
+				r2ai_tool_result_fini (&tool_result);
 			}
 
 			free (tool_name);
 			free (tool_args);
+			if (!cmd_output) {
+				cmd_output = strdup ("<no output>");
+			}
 			if (strcmp (cmd_output, "R2AI_SIGINT") == 0) {
-				r_cons_printf (core->cons, "\n\n\x1b[1" Color_RED "[r2ai] Processing interrupted after tool execution" Color_RESET "\n\n");
+				r_cons_printf (core->cons, "\n\n" Color_BOLD Color_RED "[r2ai] Processing interrupted after tool execution" Color_RESET "\n\n");
 				r_cons_flush (core->cons);
 				free (cmd_output);
 				cmd_output = strdup ("<user interrupted>");
 				interrupted = true;
-
-				handle_final_response_attempt (cps, messages, effective_prompt, &error);
 			}
 			if (comment) {
 				char *msg = r_str_newf ("HINT: %s\n%s", comment, cmd_output);
@@ -366,7 +339,7 @@ R_API void process_messages(RCorePluginSession *cps, RList *messages, const char
 			// Create a tool call response message
 			R2AI_Message tool_response = {
 				.role = "tool",
-				.tool_call_id = strdup (tool_call->id),
+				.tool_call_id = (char *)tool_call->id,
 				.content = cmd_output
 			};
 
@@ -390,20 +363,23 @@ R_API void process_messages(RCorePluginSession *cps, RList *messages, const char
 			process_messages (cps, messages, effective_prompt, n_run + 1);
 		} else {
 			R_LOG_DEBUG ("Auto mode loop ending - no more tool calls or interrupted");
+			if (interrupted) {
+				// all tool responses are in place, so the history stays well ordered
+				handle_final_response_attempt (cps, messages, effective_prompt, &error);
+			}
 		}
 	} else {
 		r2ai_print_run_end (cps, usage, n_run, max_runs);
 	}
 
-	// Free the response struct itself since r2ai_message_free doesn't do it anymore
-	free (response);
+	// Free the response, including its message and usage
+	r2ai_chat_response_free (response);
 }
 
 R_IPI void cmd_r2ai_a(RCorePluginSession *cps, const char *user_query) {
 	RCore *core = cps->core;
 	R2AI_State *state = cps->data;
-	r_cons_printf (core->cons, Color_CYAN "cmd_r2ai_a called with query: %s" Color_RESET "\n", user_query);
-	r_cons_flush (core->cons);
+	R_LOG_DEBUG ("cmd_r2ai_a called with query: %s", user_query);
 	// Get conversation
 	RList *messages = r2ai_conversation_get (state);
 	if (!messages) {
@@ -415,6 +391,7 @@ R_IPI void cmd_r2ai_a(RCorePluginSession *cps, const char *user_query) {
 	// If this is the first message in a new conversation, clear previous history
 	if (r_list_empty (messages) || r_config_get_b (core->config, "r2ai.auto.reset_on_query")) {
 		r2ai_msgs_clear (messages);
+		R_FREE (state->cache_prefix);
 	}
 
 	// Add user query
@@ -424,59 +401,57 @@ R_IPI void cmd_r2ai_a(RCorePluginSession *cps, const char *user_query) {
 	};
 	r2ai_msgs_add (messages, &user_msg);
 
-	const char *system_prompt = NULL;
-	if (!r_config_get_b (core->config, "r2ai.auto.think")) {
-		system_prompt = r_str_newf ("/no_think\nReasoning: Low\n%s", Gprompt_auto);
-	}
+	char *system_prompt = r2ai_auto_system_prompt (cps);
 	process_messages (cps, messages, system_prompt, 1);
-	if (system_prompt) {
-		free ((char *)system_prompt);
-	}
+	free (system_prompt);
 }
 
-// Helper function to display content with length indication for long content
-static void print_content_with_length(RCore *core, const char *content, const char *empty_msg) {
-	if (R_STR_ISEMPTY (content)) {
-		r_cons_println (core->cons, empty_msg? empty_msg: "<no content>");
-	} else {
-		r_cons_println (core->cons, content);
+static char *log_tostring(const RList *messages, bool color) {
+	const char *bold = color? Color_BOLD: "";
+	const char *reset = color? Color_RESET: "";
+	RStrBuf *sb = r_strbuf_new ("");
+	RListIter *iter, *iter2;
+	const R2AI_Message *msg;
+	R2AI_ToolCall *tc;
+	r_list_foreach (messages, iter, msg) {
+		const char *role = msg->role;
+		bool tool = !strcmp (role, "tool");
+		const char *clr = !color? "": !strcmp (role, "user")? Color_GREEN: !strcmp (role, "assistant")? Color_CYAN: tool? Color_MAGENTA: Color_WHITE;
+		const char *text = R_STR_ISNOTEMPTY (msg->content)? msg->content: tool? "<no result>": "<no content>";
+		r_strbuf_appendf (sb, "%s%s[%s]:%s %s\n", bold, clr, role, reset, text);
+		r_list_foreach (msg->tool_calls, iter2, tc) {
+			r_strbuf_appendf (sb, "  %s%s[tool call]:%s %s\n", bold, color? Color_MAGENTA: "", reset, tc->name? tc->name: "<unnamed>");
+			if (tc->arguments) {
+				r_strbuf_appendf (sb, "    %s\n", tc->arguments);
+			}
+		}
+		if (color) {
+			r_strbuf_append (sb, "\n");
+		}
 	}
+	return r_strbuf_drain (sb);
 }
 
-// Add this function right after cmd_r2ai_a
-R_IPI void cmd_r2ai_logs(RCorePluginSession *cps) {
+R_IPI void cmd_r2ai_logs(RCorePluginSession *cps, const char *flags) {
 	RCore *core = cps->core;
-	R2AI_State *state = cps->data;
-	// Get conversation
-	RList *messages = r2ai_conversation_get (state);
-	if (!messages || r_list_empty (messages)) {
+	RList *messages = r2ai_conversation_get (cps->data);
+	if (r_list_empty (messages)) {
 		r_cons_printf (core->cons, "No conversation history available\n");
 		return;
 	}
-
-	const char *input = r_core_cmd_str (core, "r2ai");
-	bool json_mode = input && strstr (input, "-Lj");
-	free ((char *)input);
-
-	if (json_mode) {
+	if (flags && strchr (flags, 'j')) {
 		PJ *pj = pj_new ();
-		if (!pj) {
-			return;
-		}
-
 		pj_a (pj);
-
-		RListIter *iter;
+		RListIter *iter, *iter2;
 		const R2AI_Message *msg;
+		R2AI_ToolCall *tc;
 		r_list_foreach (messages, iter, msg) {
 			pj_o (pj);
-			pj_ks (pj, "role", msg->role? msg->role: "unknown");
-			pj_ks (pj, "content", msg->content? msg->content: "");
-			if (msg->tool_calls && r_list_length (msg->tool_calls) > 0) {
+			pj_ks (pj, "role", r_str_get_fail (msg->role, "unknown"));
+			pj_ks (pj, "content", r_str_get (msg->content));
+			if (!r_list_empty (msg->tool_calls)) {
 				pj_ka (pj, "tool_calls");
-				RListIter *iter;
-				R2AI_ToolCall *tc;
-				r_list_foreach (msg->tool_calls, iter, tc) {
+				r_list_foreach (msg->tool_calls, iter2, tc) {
 					pj_o (pj);
 					if (tc->name) {
 						pj_ks (pj, "name", tc->name);
@@ -490,96 +465,17 @@ R_IPI void cmd_r2ai_logs(RCorePluginSession *cps) {
 			}
 			pj_end (pj);
 		}
-
 		pj_end (pj);
-
 		char *json_str = pj_drain (pj);
-		r_cons_printf (core->cons, "%s\n", json_str);
+		r_cons_println (core->cons, json_str);
 		free (json_str);
-
 		return;
 	}
-
-	r_cons_printf (core->cons, "\x1b[1" Color_BLUE "[r2ai] Chat Logs (%d messages)" Color_RESET "\n",
-		r_list_length (messages));
-
-	r_cons_printf (core->cons, "\x1b[1" Color_YELLOW "Note: System prompt is applied automatically but not stored in history" Color_RESET "\n\n");
-
-	// Display each message in the conversation
-	RListIter *iter;
-	const R2AI_Message *msg;
-	r_list_foreach (messages, iter, msg) {
-		const char *role = msg->role;
-
-		// Format based on role
-		if (!strcmp (role, "user")) {
-			r_cons_printf (core->cons, "\x1b[1" Color_GREEN "[user]:" Color_RESET " ");
-			print_content_with_length (core, msg->content, "<no content>");
-		} else if (!strcmp (role, "assistant")) {
-			r_cons_printf (core->cons, "\x1b[1" Color_CYAN "[assistant]:" Color_RESET " ");
-			print_content_with_length (core, msg->content, "<no content>");
-			// Show tool calls if present
-			if (msg->tool_calls && r_list_length (msg->tool_calls) > 0) {
-				RListIter *iter;
-				R2AI_ToolCall *tc;
-				r_list_foreach (msg->tool_calls, iter, tc) {
-					r_cons_printf (core->cons, "  \x1b[1" Color_MAGENTA "[tool call]:" Color_RESET " %s\n",
-						tc->name? tc->name: "<unnamed>");
-
-					if (tc->arguments) {
-						r_cons_printf (core->cons, "    %s\n", tc->arguments);
-					}
-				}
-			}
-		} else if (!strcmp (role, "tool")) {
-			r_cons_printf (core->cons, "\x1b[1" Color_MAGENTA "[tool]:" Color_RESET " ");
-			print_content_with_length (core, msg->content, "<no result>");
-
-			// Don't show the tool call ID as requested
-		} else {
-			// Other roles (system, etc.)
-			r_cons_printf (core->cons, "\x1b[1" Color_WHITE "[%s]:" Color_RESET " ", role);
-			print_content_with_length (core, msg->content, "<no content>");
-		}
-
-		r_cons_newline (core->cons);
-		r_cons_flush (core->cons);
-	}
-}
-
-// Helper function to format conversation log as string
-static char *format_conversation_log(RList *messages) {
-	RStrBuf *sb = r_strbuf_new ("");
-	r_strbuf_append (sb, "Conversation Log:\n");
-
-	RListIter *iter;
-	const R2AI_Message *msg;
-	r_list_foreach (messages, iter, msg) {
-		const char *role = msg->role;
-
-		if (!strcmp (role, "user")) {
-			r_strbuf_appendf (sb, "[user]: %s\n", msg->content? msg->content: "<no content>");
-		} else if (!strcmp (role, "assistant")) {
-			r_strbuf_appendf (sb, "[assistant]: %s\n", msg->content? msg->content: "<no content>");
-			// Include tool calls if present
-			if (msg->tool_calls && r_list_length (msg->tool_calls) > 0) {
-				RListIter *iter_tc;
-				R2AI_ToolCall *tc;
-				r_list_foreach (msg->tool_calls, iter_tc, tc) {
-					r_strbuf_appendf (sb, "  [tool call]: %s\n", tc->name? tc->name: "<unnamed>");
-					if (tc->arguments) {
-						r_strbuf_appendf (sb, "    %s\n", tc->arguments);
-					}
-				}
-			}
-		} else if (!strcmp (role, "tool")) {
-			r_strbuf_appendf (sb, "[tool]: %s\n", msg->content? msg->content: "<no result>");
-		} else {
-			r_strbuf_appendf (sb, "[%s]: %s\n", role, msg->content? msg->content: "<no content>");
-		}
-	}
-
-	return r_strbuf_drain (sb);
+	char *log = log_tostring (messages, true);
+	r_cons_printf (core->cons, Color_BOLD Color_BLUE "[r2ai] Chat Logs (%d messages)" Color_RESET "\n"
+		Color_BOLD Color_YELLOW "Note: System prompt is applied automatically but not stored in history" Color_RESET "\n\n%s",
+		r_list_length (messages), log);
+	free (log);
 }
 
 // Helper function to process conversation with LLM and handle result
@@ -594,7 +490,9 @@ static void process_conversation_with_llm(RCorePluginSession *cps, bool compact)
 	}
 
 	// Format conversation as string
-	char *log_str = format_conversation_log (messages);
+	char *log = log_tostring (messages, false);
+	char *log_str = r_str_newf ("Conversation Log:\n%s", log);
+	free (log);
 
 	// Load the compact prompt from file
 	char *prompt_text = r2ai_load_prompt_text (core, "compact");
@@ -602,7 +500,7 @@ static void process_conversation_with_llm(RCorePluginSession *cps, bool compact)
 		// Fallback to hardcoded prompt
 		prompt_text = strdup (compact
 				? "Create a compact summary of this conversation that preserves all key information, insights, and context for future reference."
-				"Focus on essential details about the binary analysis, tools used, findings, and any important conclusions."
+					"Focus on essential details about the binary analysis, tools used, findings, and any important conclusions."
 				: "mai create a summary of all the information retrieved from the binary that is relevant for future work");
 	}
 
@@ -626,6 +524,7 @@ static void process_conversation_with_llm(RCorePluginSession *cps, bool compact)
 		if (compact) {
 			// Clear the conversation and add summary as system message
 			r2ai_msgs_clear (messages);
+			R_FREE (state->cache_prefix);
 			R2AI_Message summary_msg = {
 				.role = "system",
 				.content = res

src/claw.c

@@ -0,0 +1,167 @@
+/* r2ai - Copyright 2026 pancake */
+
+#define R_LOG_ORIGIN "r2ai.claw"
+
+#include "r2ai.h"
+#include "r2ai_priv.h"
+
+#define CLAW_DEFAULT_IDENTITY \
+	"- Name: r2clippy\n" \
+		"- A fun, useful reverse engineering sidekick living inside radare2\n" \
+		"- Speaks short, make jokes, but always goes to the point\n" \
+		"- Loves assembly, expert cracker, eats hex dumps for breakfast and it's smart building clever unix one-liners"
+
+#define CLAW_DEFAULT_SOUL \
+	"- Useful first, fun second\n" \
+		"- Show, don't preach: prefer concrete r2 commands and runnable examples\n" \
+		"- Curious about every binary, never bored\n" \
+		"- Admit when something is unclear instead of guessing"
+
+#define CLAW_PROMPT_RULES \
+	"Output between 2 and 5 short markdown bullet points - be flexible, do not pad. " \
+		"Output ONLY the bullets, no preface, no explanation, no code fences."
+
+typedef struct {
+	const char *rel;
+	const char *tag;
+	const char *def;
+	const char *prompt;
+} ClawFile;
+
+static const ClawFile CLAW_FILES[] = {
+	{ ".config/r2ai/IDENTITY.md", "identity", CLAW_DEFAULT_IDENTITY, "Generate an IDENTITY for a fun and useful reverse engineering assistant living in radare2. "
+		"IDENTITY defines WHO the assistant is: name, vibe, voice, style. "
+		"Be invented, surprising and unique each time. " CLAW_PROMPT_RULES },
+	{ ".config/r2ai/SOUL.md", "soul", CLAW_DEFAULT_SOUL, "Generate a SOUL for a fun and useful reverse engineering assistant living in radare2. "
+		"SOUL defines HOW the assistant behaves: values, opinions, rules, character. "
+		"Useful first, fun second; have a clear personal voice. " CLAW_PROMPT_RULES },
+};
+#define CLAW_NFILES ((int) (sizeof (CLAW_FILES) / sizeof (CLAW_FILES[0])))
+
+R_API bool r2ai_claw_exists(void) {
+	int i;
+	for (i = 0; i < CLAW_NFILES; i++) {
+		char *path = r_file_home (CLAW_FILES[i].rel);
+		bool e = path && r_file_exists (path);
+		free (path);
+		if (e) {
+			return true;
+		}
+	}
+	return false;
+}
+
+static char *claw_personality(void) {
+	RStrBuf *sb = r_strbuf_new ("");
+	int i;
+	for (i = 0; i < CLAW_NFILES; i++) {
+		char *path = r_file_home (CLAW_FILES[i].rel);
+		char *content = path? r_file_slurp (path, NULL): NULL;
+		free (path);
+		r_strbuf_appendf (sb, "\n<%s>\n%s\n</%s>\n", CLAW_FILES[i].tag, content? content: CLAW_FILES[i].def, CLAW_FILES[i].tag);
+		free (content);
+	}
+	return r_strbuf_drain (sb);
+}
+
+R_API char *r2ai_claw_system_prompt(const char *base) {
+	if (!r2ai_claw_exists ()) {
+		return strdup (base? base: "");
+	}
+	char *personality = claw_personality ();
+	char *result = r_str_newf ("%s%s", base? base: "", personality);
+	free (personality);
+	return result;
+}
+
+static char *claw_generate(RCorePluginSession *cps, const char *prompt, const char *extra) {
+	char *full = R_STR_ISEMPTY (extra)
+		? strdup (prompt)
+		: r_str_newf ("%s\n\nAdditional hints from the user: %s", prompt, extra);
+	char *err = NULL;
+	R2AIArgs args = {
+		.input = full,
+		.system_prompt = "You are a creative writer crafting concise personality files.",
+		.error = &err,
+		.dorag = false
+	};
+	char *res = r2ai (cps, args);
+	free (full);
+	if (err) {
+		R_LOG_ERROR ("%s", err);
+		free (err);
+	}
+	return res;
+}
+
+R_API void r2ai_claw_show(RCorePluginSession *cps) {
+	RCore *core = cps->core;
+	int i;
+	for (i = 0; i < CLAW_NFILES; i++) {
+		char *path = r_file_home (CLAW_FILES[i].rel);
+		char *content = path? r_file_slurp (path, NULL): NULL;
+		if (content) {
+			r_cons_printf (core->cons, "\n=== %s ===\n%s\n", path, content);
+		}
+		free (content);
+		free (path);
+	}
+}
+
+R_API void r2ai_claw_create(RCorePluginSession *cps, const char *user_extra) {
+	if (!r2ai_claw_exists ()) {
+		char *cfg = r_file_home (".config/r2ai");
+		r_sys_mkdirp (cfg);
+		free (cfg);
+		int i;
+		for (i = 0; i < CLAW_NFILES; i++) {
+			const ClawFile *f = &CLAW_FILES[i];
+			R_LOG_INFO ("Generating %s", f->tag);
+			char *text = claw_generate (cps, f->prompt, user_extra);
+			if (R_STR_ISEMPTY (text)) {
+				R_LOG_ERROR ("Failed to generate %s", f->tag);
+				free (text);
+				continue;
+			}
+			char *path = r_file_home (f->rel);
+			if (!path || !r_file_dump (path, (const ut8 *)text, strlen (text), false)) {
+				R_LOG_ERROR ("Cannot write %s", path? path: f->rel);
+			}
+			free (path);
+			free (text);
+		}
+	}
+	r2ai_claw_show (cps);
+	r_cons_printf (cps->core->cons, "\n%s.\n", R2AI_CLAW_HINT);
+}
+
+R_API void r2ai_claw_edit(RCorePluginSession *cps) {
+	if (!r2ai_claw_exists ()) {
+		R_LOG_ERROR ("No personality files, run 'r2ai -id' to generate them first");
+		return;
+	}
+	int i;
+	for (i = 0; i < CLAW_NFILES; i++) {
+		char *path = r_file_home (CLAW_FILES[i].rel);
+		if (path && r_file_exists (path)) {
+			r2ai_cons_editor (cps->core->cons, path, NULL);
+		}
+		free (path);
+	}
+}
+
+R_API void r2ai_claw_delete(RCorePluginSession *cps) {
+	RCore *core = cps->core;
+	if (!r2ai_claw_exists ()) {
+		r_cons_printf (core->cons, "No personality files to delete.\n");
+		return;
+	}
+	int i;
+	for (i = 0; i < CLAW_NFILES; i++) {
+		char *path = r_file_home (CLAW_FILES[i].rel);
+		if (path && r_file_exists (path) && r_file_rm (path)) {
+			r_cons_printf (core->cons, "Deleted %s\n", path);
+		}
+		free (path);
+	}
+}

src/gemini.c

@@ -1,216 +1,102 @@
-/* r2ai - Copyright 2023-2025 pancake, dnakov */
+/* r2ai - Copyright 2023-2026 pancake, dnakov */
 
 #include "r2ai.h"
+#include "r2ai_priv.h"
 
-/* Gemini API implementation */
+static void text_parts(PJ *pj, const char *text) {
+	pj_ka (pj, "parts");
+	pj_o (pj);
+	pj_ks (pj, "text", text);
+	pj_end (pj);
+	pj_end (pj);
+}
 
-R_IPI R2AI_ChatResponse *r2ai_gemini(RCorePluginSession *cps, R2AIArgs args) {
-	RCore *core = cps->core;
-	args.model = r_config_get (core->config, "r2ai.model");
-
-	const char *base_url = r2ai_get_provider_url (core, args.provider);
-	const char *model_name = args.model && strstr (args.model, "gemini")? args.model: "gemini-2.0-flash-exp";
-	char **error = args.error;
-
-	// Create a temp conversation to include the system prompt and the rest of the messages
-	RList *temp_msgs = r2ai_msgs_new ();
-	if (!temp_msgs) {
-		if (error) {
-			*error = strdup ("Failed to create temporary messages array");
-		}
-		return NULL;
-	}
-
-	// Get system prompt
-	const char *system_prompt = NULL;
-	if (R_STR_ISNOTEMPTY (args.system_prompt)) {
-		system_prompt = args.system_prompt;
-	} else {
-		system_prompt = r_config_get (core->config, "r2ai.system");
-	}
-
-	if (args.messages) {
-		RListIter *iter;
-		R2AI_Message *msg;
-		r_list_foreach (args.messages, iter, msg) {
-			r2ai_msgs_add (temp_msgs, msg);
-		}
-	} else {
-		R_LOG_WARN ("No messages");
-	}
-
-	if (error) {
-		*error = NULL;
-	}
-
-	// Setup headers for Gemini API
-	const char **headers = NULL;
-	char *auth_header = NULL;
-	char *api_key_header = NULL;
-
-	if (R_STR_ISNOTEMPTY (args.api_key)) {
-		static const char *static_headers[] = { NULL, NULL, NULL };
-		headers = static_headers;
-		headers[0] = "Content-Type: application/json";
-
-		// Check if it's an OAuth token (starts with "ya29.")
-		if (r_str_startswith (args.api_key, "ya29.")) {
-			auth_header = r_str_newf ("Authorization: Bearer %s", args.api_key);
-			headers[1] = auth_header;
-		} else {
-			api_key_header = r_str_newf ("x-goog-api-key: %s", args.api_key);
-			headers[1] = api_key_header;
-		}
-	}
-
-	// Build Gemini API URL
-	char *gemini_url = r_str_newf ("%s/models/%s:generateContent?key=%s",
-		base_url, model_name, args.api_key);
-
-	// Create Gemini-style request JSON
+// request body shared by the gemini and vertex-gemini providers
+R_IPI char *r2ai_gemini_request(const R2AIArgs *args) {
 	PJ *pj = pj_new ();
 	pj_o (pj);
-
-	// Build contents array
+	if (R_STR_ISNOTEMPTY (args->system_prompt)) {
+		pj_ko (pj, "systemInstruction");
+		text_parts (pj, args->system_prompt);
+		pj_end (pj);
+	}
 	pj_ka (pj, "contents");
 	RListIter *iter;
 	R2AI_Message *msg;
-	int msg_index = 0;
-	r_list_foreach (temp_msgs, iter, msg) {
+	r_list_foreach (args->messages, iter, msg) {
 		pj_o (pj);
-		// Gemini uses "model" for assistant, "user" for user
-		const char *role = "user";
-		if (strcmp (msg->role, "assistant") == 0) {
-			role = "model";
-		}
-		pj_ks (pj, "role", role);
-		pj_ka (pj, "parts");
-		pj_o (pj);
-		// Prepend system prompt to first user message
-		char *content = NULL;
-		if (system_prompt && strcmp (msg->role, "user") == 0 && msg_index == 0) {
-			content = r_str_newf ("System: %s\n\n%s", system_prompt, msg->content? msg->content: "");
-		} else {
-			content = msg->content? strdup (msg->content): strdup ("");
-		}
-		pj_ks (pj, "text", content);
-		free (content);
-		pj_end (pj); // end parts object
-		pj_end (pj); // end parts array
-		pj_end (pj); // end content object
-		msg_index++;
+		pj_ks (pj, "role", strcmp (msg->role, "assistant")? "user": "model");
+		text_parts (pj, r_str_get (msg->content));
+		pj_end (pj);
 	}
-	pj_end (pj); // end contents array
-
-	// Add generation config
+	pj_end (pj);
 	pj_ko (pj, "generationConfig");
-	if (args.max_tokens > 0) {
-		pj_kn (pj, "maxOutputTokens", args.max_tokens);
+	if (args->max_tokens > 0) {
+		pj_kn (pj, "maxOutputTokens", args->max_tokens);
 	}
-	if (args.temperature > 0) {
-		pj_kd (pj, "temperature", args.temperature);
+	if (args->deterministic) {
+		pj_kn (pj, "temperature", 0);
+		pj_kn (pj, "topP", 1);
+		pj_kn (pj, "topK", 1);
+	} else if (args->temperature > 0) {
+		pj_kd (pj, "temperature", args->temperature);
 	}
-	pj_end (pj); // end generationConfig
+	pj_end (pj);
+	pj_end (pj);
+	return pj_drain (pj);
+}
 
-	pj_end (pj); // end root object
-
-	char *request_json = pj_drain (pj);
-
-	// Debug: save request
-	char *tmpdir = r_file_tmpdir ();
-	char *req_path = r_str_newf ("%s" R_SYS_DIR "r2ai_gemini_request.json", tmpdir);
-	r_file_dump (req_path, (const ut8 *)request_json, strlen (request_json), 0);
-	free (req_path);
-	free (tmpdir);
-
-	// Make the API call
-	char *response = NULL;
-	int code = 0;
-	response = r2ai_http_post (core, gemini_url, headers, request_json, &code, NULL);
-
-	free (request_json);
-	free (gemini_url);
-	free (auth_header);
-	free (api_key_header);
-
-	if (code != 200) {
-		R_LOG_ERROR ("Gemini API error %d", code);
-		if (response) {
-			R_LOG_ERROR ("Gemini API error response: %s", response);
-		}
-		free (response);
-		r2ai_msgs_free (temp_msgs);
+// parses (and modifies) the json response of the gemini and vertex-gemini providers
+R_IPI R2AI_ChatResponse *r2ai_gemini_parse(char *json, char **error) {
+	RJson *j = r_json_parse (json);
+	if (!j) {
+		*error = strdup ("Failed to parse Gemini response JSON");
 		return NULL;
 	}
-
-	// Debug: save response
-	tmpdir = r_file_tmpdir ();
-	char *res_path = r_str_newf ("%s" R_SYS_DIR "r2ai_gemini_response.json", tmpdir);
-	r_file_dump (res_path, (const ut8 *)response, strlen (response), 0);
-	free (res_path);
-	free (tmpdir);
-
-	// Parse Gemini response
-	RJson *jres = r_json_parse (response);
-	if (!jres) {
-		if (error) {
-			*error = strdup ("Failed to parse Gemini response JSON");
-		}
-		free (response);
-		r2ai_msgs_free (temp_msgs);
+	const RJson *c = r_json_get (j, "candidates");
+	c = c? r_json_item (c, 0): NULL;
+	c = c? r_json_get (c, "content"): NULL;
+	c = c? r_json_get (c, "parts"): NULL;
+	c = c? r_json_item (c, 0): NULL;
+	const char *text = c? r_json_get_str (c, "text"): NULL;
+	if (!text) {
+		r_json_free (j);
 		return NULL;
 	}
-
-	// Extract response content
-	const RJson *candidates = r_json_get (jres, "candidates");
-	if (!candidates || candidates->type != R_JSON_ARRAY || candidates->children.count == 0) {
-		r_json_free (jres);
-		free (response);
-		r2ai_msgs_free (temp_msgs);
-		return NULL;
-	}
-
-	const RJson *candidate = r_json_item (candidates, 0);
-	const RJson *content = r_json_get (candidate, "content");
-	const RJson *parts = r_json_get (content, "parts");
-
-	if (!parts || parts->type != R_JSON_ARRAY || parts->children.count == 0) {
-		r_json_free (jres);
-		free (response);
-		r2ai_msgs_free (temp_msgs);
-		return NULL;
-	}
-
-	const RJson *part = r_json_item (parts, 0);
-	const RJson *text = r_json_get (part, "text");
-
-	char *response_text = NULL;
-	if (text && text->type == R_JSON_STRING) {
-		response_text = strdup (text->str_value);
-	}
-
-	r_json_free (jres);
-	free (response);
-	r2ai_msgs_free (temp_msgs);
-
-	if (!response_text) {
-		return NULL;
-	}
-
-	// Create response
-	R2AI_Message *message = R_NEW0 (R2AI_Message);
 	R2AI_Usage *usage = R_NEW0 (R2AI_Usage);
-
-	if (message) {
-		message->role = strdup ("assistant");
-		message->content = response_text;
-	}
-
-	R2AI_ChatResponse *result = R_NEW0 (R2AI_ChatResponse);
-	if (result) {
-		result->message = message;
-		result->usage = usage;
+	const RJson *u = r_json_get (j, "usageMetadata");
+	if (u) {
+		usage->prompt_tokens = r_json_get_num (u, "promptTokenCount");
+		usage->completion_tokens = r_json_get_num (u, "candidatesTokenCount");
+		usage->total_tokens = r_json_get_num (u, "totalTokenCount");
 	}
+	R2AI_Message *message = R_NEW0 (R2AI_Message);
+	message->role = strdup ("assistant");
+	message->content = strdup (text);
+	r_json_free (j);
+	R2AI_ChatResponse *res = R_NEW0 (R2AI_ChatResponse);
+	res->message = message;
+	res->usage = usage;
+	return res;
+}
 
+R_IPI R2AI_ChatResponse *r2ai_gemini(RCorePluginSession *cps, R2AIArgs args) {
+	RCore *core = cps->core;
+	const char *model = strstr (args.model, "gemini")? args.model: "gemini-2.0-flash-exp";
+	char *base_url = r2ai_get_provider_url (core, args.provider);
+	char *url = r_str_newf ("%s/models/%s:generateContent", base_url, model);
+	free (base_url);
+	// OAuth tokens start with "ya29.", the key travels in the headers and never in the url
+	const bool oauth = r_str_startswith (args.api_key, "ya29.");
+	char *auth = r_str_newf (oauth? "Authorization: Bearer %s": "x-goog-api-key: %s", args.api_key);
+	const char *headers[] = { "Content-Type: application/json", auth, NULL };
+	char *data = r2ai_gemini_request (&args);
+	int code = 0;
+	char *res = r2ai_post (core, "gemini", url, headers, data, &code, args.error);
+	free (data);
+	free (url);
+	free (auth);
+	R2AI_ChatResponse *result = res? r2ai_gemini_parse (res, args.error): NULL;
+	free (res);
 	return result;
-}
+}

src/http.c

@@ -1,9 +1,10 @@
-/* Copyright r2ai - 2023-2025 - pancake */
+/* Copyright r2ai - 2023-2026 - pancake */
 
 #define _GNU_SOURCE
 // #define _POSIX_C_SOURCE 200809L
 #include <signal.h>
 #include "r2ai.h"
+#include "r2ai_priv.h"
 
 // Global flag for tracking interrupt status
 static volatile sig_atomic_t r2ai_http_interrupted = 0;
@@ -28,7 +29,7 @@ static R2AI_HttpConfig get_http_config(RCore *core) {
 	return config;
 }
 
-#if R2__UNIX__
+#if R2__UNIX__ && !__wasi__
 // Signal handler for timeout (SIGALRM)
 static void r2ai_http_sigint_handler(int sig) {
 	if (sig == SIGALRM) {
@@ -65,7 +66,7 @@ static HttpResponse r2ai_http_request_with_retry(HttpRequestFunc func, const HTT
 	// Reset interrupt flag
 	r2ai_http_interrupted = 0;
 
-#if R2__UNIX__
+#if R2__UNIX__ && !__wasi__
 	// Ignore SIGWINCH to prevent window resize from interrupting child processes
 	struct sigaction old_winch_act;
 	sigaction (SIGWINCH, &(struct sigaction){ .sa_handler = SIG_IGN }, &old_winch_act);
@@ -128,7 +129,7 @@ static HttpResponse r2ai_http_request_with_retry(HttpRequestFunc func, const HTT
 		r_cons_break_pop (core->cons);
 	}
 
-#if R2__UNIX__
+#if R2__UNIX__ && !__wasi__
 	// Restore SIGWINCH handler
 	sigaction (SIGWINCH, &old_winch_act, NULL);
 #endif
@@ -144,7 +145,9 @@ static HttpResponse r2ai_http_request_with_retry(HttpRequestFunc func, const HTT
 static HttpRequestFunc select_backend(const char *backend, bool is_post) {
 	HttpRequestFunc func = NULL;
 	if (!strcmp (backend, "auto")) {
-#if R2__WINDOWS__
+#if __wasi__
+		backend = "socket";
+#elif R2__WINDOWS__
 		// TODO: we should use pwsh or r2 by default, but for not it seems the only backend that works on Windows is r2curl+system
 		r_sys_setenv ("R2_CURL", "1");
 		backend = "system";
@@ -180,28 +183,68 @@ static HttpRequestFunc select_backend(const char *backend, bool is_post) {
 	return func;
 }
 
+static const char **custom_headers(RCore *core, const char **headers, char **extra) {
+	const char *cfg = r_config_get (core->config, "r2ai.http.headers");
+	if (R_STR_ISEMPTY (cfg)) {
+		return headers;
+	}
+	*extra = strdup (cfg);
+	int i = 0, j = 0;
+	for (; headers && headers[i]; i++) {
+	}
+	const char **merged = R_NEWS0 (const char *, i + strlen (*extra) + 2);
+	for (; j < i; j++) {
+		merged[j] = headers[j];
+	}
+	char *p = *extra, *q = p;
+	for (; *p; p++) {
+		const bool escaped_n = (*p == '\\' && p[1] == 'n');
+		if (*p != '\n' && !escaped_n) {
+			continue;
+		}
+		*p = 0;
+		if (*q) {
+			merged[i++] = q;
+		}
+		q = p + (escaped_n? 2: 1);
+		if (escaped_n) {
+			p++;
+		}
+	}
+	if (*q) {
+		merged[i] = q;
+	}
+	return merged;
+}
+
 // Generic HTTP request function that handles backend selection
-static HttpResponse r2ai_http_request(const char *method, RCore *core, const char *url, const char *headers[], const char *data) {
+static HttpResponse r2ai_http_request(const char *method, RCore *core, const char *url, const char **headers, const char *data) {
 	(void)method;
 	bool is_post = (data != NULL);
+	char *extra = NULL;
+	const char **merged = custom_headers (core, headers, &extra);
 
 	HTTPRequest request = {
 		.config = get_http_config (core),
 		.url = url,
 		.data = data,
-		.headers = headers
+		.headers = merged
 	};
 
 	const char *backend = r_config_get (core->config, "r2ai.http.backend");
 	HttpRequestFunc func = select_backend (backend, is_post);
-	if (func) {
-		return r2ai_http_request_with_retry (func, &request, core);
+	if (!func) {
+		R_LOG_ERROR ("Cannot find a valid http backend");
 	}
-	R_LOG_ERROR ("Cannot find a valid http backend");
-	return (HttpResponse){ .body = NULL, .code = -1, .length = 0 };
+	HttpResponse res = func? r2ai_http_request_with_retry (func, &request, core): (HttpResponse){ .body = NULL, .code = -1, .length = 0 };
+	if (merged != headers) {
+		free (merged);
+	}
+	free (extra);
+	return res;
 }
 
-R_API char *r2ai_http_post(RCore *core, const char *url, const char *headers[], const char *data, int *code, int *rlen) {
+R_API char *r2ai_http_post(RCore *core, const char *url, const char **headers, const char *data, int *code, int *rlen) {
 	HttpResponse response = r2ai_http_request ("POST", core, url, headers, data);
 	if (response.code <= 0) {
 		free (response.body);
@@ -219,7 +262,7 @@ R_API char *r2ai_http_post(RCore *core, const char *url, const char *headers[],
 	return response.body;
 }
 
-R_API char *r2ai_http_get(RCore *core, const char *url, const char *headers[], int *code, int *rlen) {
+R_API char *r2ai_http_get(RCore *core, const char *url, const char **headers, int *code, int *rlen) {
 	HttpResponse response = r2ai_http_request ("GET", core, url, headers, NULL);
 	if (response.code <= 0) {
 		if (response.body) {
@@ -238,3 +281,38 @@ R_API char *r2ai_http_get(RCore *core, const char *url, const char *headers[], i
 	}
 	return response.body;
 }
+
+static void dump_payload(const char *name, const char *kind, const char *data) {
+	char *tmpdir = r_file_tmpdir ();
+	char *path = r_str_newf ("%s" R_SYS_DIR "r2ai_%s_%s.json", tmpdir, name, kind);
+	r_file_dump (path, (const ut8 *)data, -1, false);
+	R_LOG_DEBUG ("%s %s saved to %s", name, kind, path);
+	free (path);
+	free (tmpdir);
+}
+
+// post a provider request, returns the body on success or NULL with the reason in error
+R_IPI char *r2ai_post(RCore *core, const char *name, const char *url, const char **headers, const char *data, int *code, char **error) {
+	dump_payload (name, "request", data);
+	if (r_config_get_b (core->config, "r2ai.debug")) {
+		RStrBuf *sb = r_strbuf_new ("curl -X POST");
+		int i;
+		for (i = 0; headers[i]; i++) {
+			r_strbuf_appendf (sb, " -H \"%s\"", headers[i]);
+		}
+		r_strbuf_appendf (sb, " -d '%s' \"%s\"", data, url);
+		eprintf ("Curl command: %s\n", r_strbuf_get (sb));
+		r_strbuf_free (sb);
+	}
+	char *res = r2ai_http_post (core, url, headers, data, code, NULL);
+	if (res && *code == 200) {
+		dump_payload (name, "response", res);
+		return res;
+	}
+	if (error) {
+		free (*error);
+		*error = r_str_newf ("%s API error %d%s%s", name, *code, res? ": ": "", r_str_get (res));
+	}
+	free (res);
+	return NULL;
+}

src/http/curl.inc.c

@@ -30,8 +30,7 @@ static char *escape_cmd_double_quotes(const char *str) {
 	return escaped;
 }
 
-static HttpResponse build_and_execute_curl(const char *cmd_start, const HTTPRequest *request, const char *input_data) {
-	HttpResponse error = { .code = -1 };
+static HttpResponse build_and_execute_curl(const char *cmd_start, const HTTPRequest *request, const char *input_data) {	HttpResponse error = { .code = -1 };
 	int timeout = request->config.timeout;
 	if (!request->url) {
 		return error;
@@ -39,6 +38,9 @@ static HttpResponse build_and_execute_curl(const char *cmd_start, const HTTPRequ
 
 	// Compose curl command
 	RStrBuf *cmd = r_strbuf_new (cmd_start);
+	if (!cmd) {
+		return error;
+	}
 
 	// Add timeout
 	r_strbuf_appendf (cmd, " --connect-timeout %d --max-time %d", 10, timeout);
@@ -47,14 +49,23 @@ static HttpResponse build_and_execute_curl(const char *cmd_start, const HTTPRequ
 	if (request->headers) {
 		for (int i = 0; request->headers[i] != NULL; i++) {
 			char *escaped_header = escape_cmd_double_quotes (request->headers[i]);
-			r_strbuf_appendf (cmd, " -H \"%s\"", escaped_header);
-			free (escaped_header);
+			if (escaped_header) {
+				r_strbuf_appendf (cmd, " -H \"%s\"", escaped_header);
+				free (escaped_header);
+			}
 		}
 	}
 	char *escaped_url = escape_cmd_double_quotes (request->url);
+	if (!escaped_url) {
+		r_strbuf_free (cmd);
+		return error;
+	}
 	r_strbuf_appendf (cmd, " -w \"\\n%%{http_code}\" \"%s\"", escaped_url);
 	free (escaped_url);
 	char *cmd_str = r_strbuf_drain (cmd);
+	if (!cmd_str) {
+		return error;
+	}
 	R_LOG_DEBUG ("Running system curl: %s", cmd_str);
 	char *response = r_sys_cmd_str (cmd_str, input_data, NULL);
 

src/http/libcurl.inc.c

@@ -36,8 +36,7 @@ static int progress_callback(void *clientp, curl_off_t dltotal, curl_off_t dlnow
 	return r2ai_http_interrupted? 1: 0;
 }
 
-static HttpResponse curl_http_request(const HTTPRequest *request, bool is_post) {
-	HttpResponse error = { .code = -1 };
+static HttpResponse curl_http_request(const HTTPRequest *request, bool is_post) {	HttpResponse error = { .code = -1 };
 	if (!request->url || (is_post && !request->data)) {
 		return error;
 	}
@@ -46,14 +45,13 @@ static HttpResponse curl_http_request(const HTTPRequest *request, bool is_post)
 	CURLcode res;
 	struct curl_slist *curl_headers = NULL;
 	CurlResponse response = { 0 };
+	HttpResponse result = { .code = -1 };
 
-	// Initialize response
 	response.data = malloc (1);
 	if (!response.data) {
 		return error;
 	}
 	response.data[0] = '\0';
-	response.size = 0;
 
 	curl = curl_easy_init ();
 	if (!curl) {
@@ -61,10 +59,8 @@ static HttpResponse curl_http_request(const HTTPRequest *request, bool is_post)
 		return error;
 	}
 
-	// Set URL
 	curl_easy_setopt (curl, CURLOPT_URL, request->url);
 
-	// Set method-specific options
 	if (is_post) {
 		curl_easy_setopt (curl, CURLOPT_POST, 1L);
 		curl_easy_setopt (curl, CURLOPT_POSTFIELDS, request->data);
@@ -73,7 +69,6 @@ static HttpResponse curl_http_request(const HTTPRequest *request, bool is_post)
 		curl_easy_setopt (curl, CURLOPT_HTTPGET, 1L);
 	}
 
-	// Set headers if provided
 	if (request->headers) {
 		for (int i = 0; request->headers[i] != NULL; i++) {
 			curl_headers = curl_slist_append (curl_headers, request->headers[i]);
@@ -81,53 +76,42 @@ static HttpResponse curl_http_request(const HTTPRequest *request, bool is_post)
 		curl_easy_setopt (curl, CURLOPT_HTTPHEADER, curl_headers);
 	}
 
-	// Set write callback
 	curl_easy_setopt (curl, CURLOPT_WRITEFUNCTION, write_callback);
 	curl_easy_setopt (curl, CURLOPT_WRITEDATA, (void *)&response);
 
-	// Set progress callback to handle interrupts
 	curl_easy_setopt (curl, CURLOPT_NOPROGRESS, 0L);
 	curl_easy_setopt (curl, CURLOPT_XFERINFOFUNCTION, progress_callback);
 	curl_easy_setopt (curl, CURLOPT_XFERINFODATA, NULL);
 
-	// Set timeout options
-	curl_easy_setopt (curl, CURLOPT_CONNECTTIMEOUT, 10L); // 10 seconds connect timeout
-	curl_easy_setopt (curl, CURLOPT_TIMEOUT, (long)request->config.timeout); // Use configured timeout
-
-	// Follow redirects
+	curl_easy_setopt (curl, CURLOPT_CONNECTTIMEOUT, 10L);
+	curl_easy_setopt (curl, CURLOPT_TIMEOUT, (long)request->config.timeout);
 	curl_easy_setopt (curl, CURLOPT_FOLLOWLOCATION, 1L);
 
-	// Perform the request
 	res = curl_easy_perform (curl);
 
-	// Check for interruption
 	if (r2ai_http_interrupted) {
 		R_LOG_DEBUG ("HTTP request was interrupted by user");
-		free (response.data);
-		curl_slist_free_all (curl_headers);
-		curl_easy_cleanup (curl);
-		return error;
+		goto cleanup;
 	}
 
-	// Get response code
-	long http_code;
+	long http_code = 0;
 	curl_easy_getinfo (curl, CURLINFO_RESPONSE_CODE, &http_code);
 
-	// Check for errors
 	if (res != CURLE_OK) {
 		R_LOG_ERROR ("curl_easy_perform () failed: %s", curl_easy_strerror (res));
-		free (response.data);
-		curl_slist_free_all (curl_headers);
-		curl_easy_cleanup (curl);
-		return error;
+		goto cleanup;
 	}
 
-	// If we get here, the request was successful
-	// Cleanup
+	result.body = response.data;
+	result.code = (int)http_code;
+	result.length = (int)response.size;
+	response.data = NULL;
+
+cleanup:
+	free (response.data);
 	curl_slist_free_all (curl_headers);
 	curl_easy_cleanup (curl);
-
-	return (HttpResponse){ .body = response.data, .code = (int)http_code, .length = response.size };
+	return result;
 }
 
 HttpResponse curl_http_get(const HTTPRequest *request) {

src/http/pwsh.inc.c

@@ -33,64 +33,92 @@ static char *escape_single_quotes(const char *str) {
 }
 
 // Helper to append headers to PowerShell command
-static void append_headers_to_cmd(RStrBuf *cmd, const char *const *headers) {
-	r_strbuf_appendf (cmd, "$headers=@{");
-	if (headers) {
-		for (int i = 0; headers[i]; i++) {
-			char *header = strdup (headers[i]);
-			char *colon = strchr (header, ':');
-			if (colon) {
-				*colon = '\0';
-				char *key = r_str_trim_dup (header);
-				char *value = r_str_trim_dup (colon + 1);
-				char *escaped_key = escape_single_quotes (key);
-				char *escaped_value = escape_single_quotes (value);
-				r_strbuf_appendf (cmd, "'%s'='%s';", escaped_key, escaped_value);
-				free (escaped_key);
-				free (escaped_value);
-				free (key);
-				free (value);
-			}
-			free (header);
+static void append_headers_to_cmd(RStrBuf *cmd, const char **headers) {	r_strbuf_appendf (cmd, "$headers=@{");
+	if (!headers) {
+		r_strbuf_appendf (cmd, "};");
+		return;
+	}
+	for (int i = 0; headers[i]; i++) {
+		char *header = strdup (headers[i]);
+		if (!header) {
+			continue;
 		}
+		char *colon = strchr (header, ':');
+		if (!colon) {
+			free (header);
+			continue;
+		}
+		*colon = '\0';
+		char *key = r_str_trim_dup (header);
+		char *value = r_str_trim_dup (colon + 1);
+		char *escaped_key = key ? escape_single_quotes (key) : NULL;
+		char *escaped_value = value ? escape_single_quotes (value) : NULL;
+		if (escaped_key && escaped_value) {
+			r_strbuf_appendf (cmd, "'%s'='%s';", escaped_key, escaped_value);
+		}
+		free (escaped_key);
+		free (escaped_value);
+		free (key);
+		free (value);
+		free (header);
 	}
 	r_strbuf_appendf (cmd, "};");
 }
 
 // Helper to parse PowerShell response
-static HttpResponse parse_powershell_response(char *full_response) {
-	if (full_response) {
-		char *newline = strchr (full_response, '\n');
-		if (newline) {
-			*newline = '\0';
-			int code = atoi (full_response);
-			char *body = newline + 1;
-			char *body_copy = strdup (body);
-			size_t length = strlen (body);
-			free (full_response);
-			return (HttpResponse){ .body = body_copy, .code = code, .length = length };
-		} else {
-			return (HttpResponse){ .body = full_response, .code = 200, .length = strlen (full_response) };
-		}
+static HttpResponse parse_powershell_response(char *full_response) {	if (!full_response) {
+		return (HttpResponse){ .code = -1 };
 	}
-	return (HttpResponse){ .code = -1 };
+	char *newline = strchr (full_response, '\n');
+	if (newline) {
+		*newline = '\0';
+		int code = atoi (full_response);
+		char *body = newline + 1;
+		char *body_copy = strdup (body);
+		if (!body_copy) {
+			free (full_response);
+			return (HttpResponse){ .code = -1 };
+		}
+		size_t length = strlen (body);
+		free (full_response);
+		return (HttpResponse){ .body = body_copy, .code = code, .length = length };
+	}
+	return (HttpResponse){ .body = full_response, .code = 200, .length = strlen (full_response) };
 }
 /**
  * Windows-specific HTTP POST using PowerShell
  */
-HttpResponse windows_http_post(const HTTPRequest *request) {
-	const char *url = request->url;
-	const char *const *headers = request->headers;
+HttpResponse windows_http_post(const HTTPRequest *request) {	const char *url = request->url;
+	const char **headers = request->headers;
 	const char *data = request->data;
 	int timeout = request->config.timeout;
+	if (!url || !data) {
+		return (HttpResponse){ .code = -1 };
+	}
 	char *escaped_url = escape_single_quotes (url);
 	char *escaped_data = escape_single_quotes (data);
+	if (!escaped_url || !escaped_data) {
+		free (escaped_url);
+		free (escaped_data);
+		return (HttpResponse){ .code = -1 };
+	}
 	RStrBuf *cmd = r_strbuf_new ("powershell -Command \"");
+	if (!cmd) {
+		free (escaped_url);
+		free (escaped_data);
+		return (HttpResponse){ .code = -1 };
+	}
 	append_headers_to_cmd (cmd, headers);
 	r_strbuf_appendf (cmd, "$body='%s';", escaped_data);
 	r_strbuf_appendf (cmd, "try{$r=Invoke-WebRequest -Method Post -Uri '%s' -Headers $headers -Body $body -TimeoutSec %d;", escaped_url, timeout);
 	r_strbuf_appendf (cmd, "Write-Host $r.StatusCode;$r.Content}catch{Write-Host 0;$_.Exception.Message}\"");
+	free (escaped_url);
+	free (escaped_data);
 	char *cmd_str = r_strbuf_drain (cmd);
+	if (!cmd_str) {
+		r_strbuf_free (cmd);
+		return (HttpResponse){ .code = -1 };
+	}
 	R_LOG_DEBUG ("Running PowerShell: %s", cmd_str);
 	char *full_response = r_sys_cmd_str (cmd_str, NULL, NULL);
 	free (cmd_str);
@@ -100,16 +128,30 @@ HttpResponse windows_http_post(const HTTPRequest *request) {
 /**
  * Windows-specific HTTP GET using PowerShell
  */
-HttpResponse windows_http_get(const HTTPRequest *request) {
-	const char *url = request->url;
-	const char *const *headers = request->headers;
+HttpResponse windows_http_get(const HTTPRequest *request) {	const char *url = request->url;
+	const char **headers = request->headers;
 	int timeout = request->config.timeout;
+	if (!url) {
+		return (HttpResponse){ .code = -1 };
+	}
 	char *escaped_url = escape_single_quotes (url);
+	if (!escaped_url) {
+		return (HttpResponse){ .code = -1 };
+	}
 	RStrBuf *cmd = r_strbuf_new ("powershell -Command \"");
+	if (!cmd) {
+		free (escaped_url);
+		return (HttpResponse){ .code = -1 };
+	}
 	append_headers_to_cmd (cmd, headers);
 	r_strbuf_appendf (cmd, "try{$r=Invoke-WebRequest -Method Get -Uri '%s' -Headers $headers -TimeoutSec %d;", escaped_url, timeout);
 	r_strbuf_appendf (cmd, "Write-Host $r.StatusCode;$r.Content}catch{Write-Host 0;$_.Exception.Message}\"");
+	free (escaped_url);
 	char *cmd_str = r_strbuf_drain (cmd);
+	if (!cmd_str) {
+		r_strbuf_free (cmd);
+		return (HttpResponse){ .code = -1 };
+	}
 	R_LOG_DEBUG ("Running PowerShell: %s", cmd_str);
 	char *full_response = r_sys_cmd_str (cmd_str, NULL, NULL);
 	free (cmd_str);

src/http/r2.inc.c

@@ -6,14 +6,14 @@ static HttpResponse socket_http_post_with_interrupt(const HTTPRequest *request)
 #if R2__WINDOWS__
 	r_sys_setenv ("R2_CURL", "1");
 #endif
-#if R2__UNIX__
+#if R2__UNIX__ && !__wasi__
 	signal (SIGALRM, r2ai_http_sigint_handler);
 	alarm (request->config.timeout); // Use configured timeout
 #endif
 	int code = 0;
 	int rlen = 0;
-	char *result = r_socket_http_post (request->url, (const char **)request->headers, request->data, &code, &rlen);
-#if R2__UNIX__
+	char *result = r_socket_http_post (request->url, request->headers, request->data, &code, &rlen);
+#if R2__UNIX__ && !__wasi__
 	alarm (0);
 #endif
 	if (r2ai_http_interrupted) {
@@ -21,6 +21,9 @@ static HttpResponse socket_http_post_with_interrupt(const HTTPRequest *request)
 		free (result);
 		return (HttpResponse){ .body = NULL, .code = -1, .length = 0 };
 	}
+	if (!result) {
+		return (HttpResponse){ .body = NULL, .code = -1, .length = 0 };
+	}
 	return (HttpResponse){ .body = result, .code = code, .length = rlen };
 }
 
@@ -29,15 +32,14 @@ static HttpResponse socket_http_get_with_interrupt(const HTTPRequest *request) {
 #if R2__WINDOWS__
 	r_sys_setenv ("R2_CURL", "1");
 #endif
-#if R2__UNIX__
+#if R2__UNIX__ && !__wasi__
 	signal (SIGALRM, r2ai_http_sigint_handler);
 	alarm (request->config.timeout); // Use configured timeout
 #endif
-	// Make the request - use r_socket_http_get if available
 	int code = 0;
 	int rlen = 0;
-	char *result = r_socket_http_get (request->url, (const char **)request->headers, &code, &rlen);
-#if R2__UNIX__
+	char *result = r_socket_http_get (request->url, request->headers, &code, &rlen);
+#if R2__UNIX__ && !__wasi__
 	alarm (0);
 #endif
 	if (r2ai_http_interrupted) {
@@ -45,5 +47,8 @@ static HttpResponse socket_http_get_with_interrupt(const HTTPRequest *request) {
 		free (result);
 		return (HttpResponse){ .body = NULL, .code = -1, .length = 0 };
 	}
+	if (!result) {
+		return (HttpResponse){ .body = NULL, .code = -1, .length = 0 };
+	}
 	return (HttpResponse){ .body = result, .code = code, .length = rlen };
 }

src/json.c

@@ -1,151 +1,55 @@
+/* Copyright r2ai - 2023-2026 - pancake */
+
 #include "r2ai.h"
 // TODO: move into r2/libr/util/json.c
-//
-// Helper function to convert RJson to PJ without draining
-R_API PJ *r_json_to_pj(const RJson *json, PJ *existing_pj) {
-	if (!json) {
-		return existing_pj;
-	}
 
-	PJ *pj = existing_pj? existing_pj: pj_new ();
-	if (!pj) {
-		return NULL;
-	}
-
-	switch (json->type) {
+static void emit(PJ *pj, const RJson *j) {
+	const RJson *c;
+	switch (j->type) {
 	case R_JSON_STRING:
-		pj_s (pj, json->str_value);
+		pj_s (pj, j->str_value);
 		break;
 	case R_JSON_INTEGER:
-		pj_n (pj, json->num.u_value);
+		if (j->num.dbl_value < 0) {
+			pj_N (pj, j->num.s_value);
+		} else {
+			pj_n (pj, j->num.u_value);
+		}
 		break;
 	case R_JSON_DOUBLE:
-		pj_d (pj, json->num.dbl_value);
+		pj_d (pj, j->num.dbl_value);
 		break;
 	case R_JSON_BOOLEAN:
-		pj_b (pj, json->num.u_value);
+		pj_b (pj, j->num.u_value);
 		break;
 	case R_JSON_NULL:
 		pj_null (pj);
 		break;
 	case R_JSON_OBJECT:
-		if (!existing_pj) {
-			pj_o (pj);
-		}
-		// Handle object's properties
-		const RJson *prop = json->children.first;
-		while (prop) {
-			if (prop->key) {
-				switch (prop->type) {
-				case R_JSON_STRING:
-					pj_ks (pj, prop->key, prop->str_value);
-					break;
-				case R_JSON_INTEGER:
-					pj_kn (pj, prop->key, prop->num.u_value);
-					break;
-				case R_JSON_DOUBLE:
-					pj_kd (pj, prop->key, prop->num.dbl_value);
-					break;
-				case R_JSON_BOOLEAN:
-					pj_kb (pj, prop->key, prop->num.u_value);
-					break;
-				case R_JSON_NULL:
-					pj_knull (pj, prop->key);
-					break;
-				case R_JSON_OBJECT:
-					pj_ko (pj, prop->key);
-					if (!r_json_to_pj (prop, pj)) {
-						if (!existing_pj) {
-							pj_free (pj);
-						}
-						return NULL;
-					}
-					pj_end (pj);
-					break;
-				case R_JSON_ARRAY:
-					pj_ka (pj, prop->key);
-					if (!r_json_to_pj (prop, pj)) {
-						if (!existing_pj) {
-							pj_free (pj);
-						}
-						return NULL;
-					}
-					pj_end (pj);
-					break;
-				default:
-					break;
-				}
+		pj_o (pj);
+		for (c = j->children.first; c; c = c->next) {
+			if (c->key) {
+				pj_k (pj, c->key);
+				emit (pj, c);
 			}
-			prop = prop->next;
-		}
-		if (!existing_pj) {
-			pj_end (pj);
 		}
+		pj_end (pj);
 		break;
 	case R_JSON_ARRAY:
-		if (!existing_pj) {
-			pj_a (pj);
-		}
-		// Handle array items
-		const RJson *item = json->children.first;
-		while (item) {
-			switch (item->type) {
-			case R_JSON_STRING:
-				pj_s (pj, item->str_value);
-				break;
-			case R_JSON_INTEGER:
-				pj_n (pj, item->num.u_value);
-				break;
-			case R_JSON_DOUBLE:
-				pj_d (pj, item->num.dbl_value);
-				break;
-			case R_JSON_BOOLEAN:
-				pj_b (pj, item->num.u_value);
-				break;
-			case R_JSON_NULL:
-				pj_null (pj);
-				break;
-			case R_JSON_OBJECT:
-				pj_o (pj);
-				if (!r_json_to_pj (item, pj)) {
-					if (!existing_pj) {
-						pj_free (pj);
-					}
-					return NULL;
-				}
-				pj_end (pj);
-				break;
-			case R_JSON_ARRAY:
-				pj_a (pj);
-				if (!r_json_to_pj (item, pj)) {
-					if (!existing_pj) {
-						pj_free (pj);
-					}
-					return NULL;
-				}
-				pj_end (pj);
-				break;
-			default:
-				break;
-			}
-			item = item->next;
-		}
-		if (!existing_pj) {
-			pj_end (pj);
+		pj_a (pj);
+		for (c = j->children.first; c; c = c->next) {
+			emit (pj, c);
 		}
+		pj_end (pj);
 		break;
-	default:
-		if (!existing_pj) {
-			pj_free (pj);
-		}
+	}
+}
+
+R_API char *r_json_to_string(const RJson *json) {
+	if (!json) {
 		return NULL;
 	}
-
-	return pj;
-}
-
-// Helper function to clone RJson to string
-R_API char *r_json_to_string(const RJson *json) {
-	PJ *pj = r_json_to_pj (json, NULL);
-	return pj? pj_drain (pj): NULL;
+	PJ *pj = pj_new ();
+	emit (pj, json);
+	return pj_drain (pj);
 }

src/llm.c

@@ -1,3 +1,5 @@
+/* Copyright r2ai - 2023-2026 - pancake */
+
 #define R_LOG_ORIGIN "llm"
 
 #include "r2ai.h"
@@ -9,7 +11,7 @@ static const R2AIProvider r2ai_providers[] = {
 	{ "gemini", "https://generativelanguage.googleapis.com/v1beta", R2AI_API_GEMINI, true, false },
 	{ "anthropic", "https://api.anthropic.com/v1", R2AI_API_ANTHROPIC, true, false },
 	{ "claude", "https://api.anthropic.com/v1", R2AI_API_ANTHROPIC, true, false },
-	{ "ollama", "http://localhost:11434/api", R2AI_API_OLLAMA, false, true },
+	{ "ollama", "http://localhost:11434/v1", R2AI_API_OLLAMA, false, true },
 	{ "ollamacloud", "https://ollama.com/api", R2AI_API_OLLAMA, true, true },
 	{ "openapi", "http://127.0.0.1:11434", R2AI_API_OPENAI_COMPATIBLE, false, false },
 	{ "opencode", "https://opencode.ai/zen/v1", R2AI_API_OPENAI_COMPATIBLE, true, true },
@@ -20,6 +22,8 @@ static const R2AIProvider r2ai_providers[] = {
 	{ "mistral", "https://api.mistral.ai/v1", R2AI_API_OPENAI_COMPATIBLE, true, true },
 	{ "lmstudio", "http://127.0.0.1:1234/v1", R2AI_API_OPENAI_COMPATIBLE, false, true },
 	{ "deepseek", "https://api.deepseek.com/v1", R2AI_API_OPENAI_COMPATIBLE, true, true },
+	{ "vertex", NULL, R2AI_API_VERTEX_GEMINI, false, false },
+	{ "vertex-anthropic", NULL, R2AI_API_VERTEX_ANTHROPIC, false, false },
 	{ NULL, NULL, R2AI_API_OPENAI_COMPATIBLE, false, false } // sentinel
 };
 
@@ -35,27 +39,66 @@ R_IPI const R2AIProvider *r2ai_get_provider(const char *name) {
 	return NULL;
 }
 
-// Forward declaration for rawtools
+static bool is_vertex(const R2AIProvider *p) {
+	return p->api_type == R2AI_API_VERTEX_GEMINI || p->api_type == R2AI_API_VERTEX_ANTHROPIC;
+}
+
+R_IPI R2AI_ChatResponse *r2ai_send(RCorePluginSession *cps, const R2AIProvider *p, R2AIArgs args) {
+	switch (p->api_type) {
+	case R2AI_API_ANTHROPIC:
+		return r2ai_anthropic (cps, args);
+	case R2AI_API_GEMINI:
+		return r2ai_gemini (cps, args);
+	case R2AI_API_VERTEX_GEMINI:
+		return r2ai_vertex_gemini (cps, args);
+	case R2AI_API_VERTEX_ANTHROPIC:
+		return r2ai_vertex_anthropic (cps, args);
+	default:
+		return r2ai_openai (cps, args);
+	}
+}
+
+static bool is_generate_api(RCore *core) {
+	const char *apitype = r_config_get (core->config, "r2ai.apitype");
+	return R_STR_ISNOTEMPTY (apitype) && !strcmp (apitype, "generate");
+}
+
+static bool use_rawtools(RCore *core, const R2AIProvider *provider, const R2AIArgs *args) {
+	if (!args || !args->tools || r_list_empty (args->tools)) {
+		return false;
+	}
+	if (r_config_get_b (core->config, "r2ai.auto.raw")) {
+		return true;
+	}
+	return provider && provider->api_type == R2AI_API_OLLAMA && is_generate_api (core);
+}
+
+// the given system prompt, or r2ai.system when it is empty
+R_IPI const char *r2ai_system_prompt(RCore *core, const char *sysp) {
+	return R_STR_ISNOTEMPTY (sysp)? sysp: r_config_get (core->config, "r2ai.system");
+}
+
 R_IPI R2AI_ChatResponse *r2ai_llmcall(RCorePluginSession *cps, R2AIArgs args) {
 	RCore *core = cps->core;
+	char *owned_model = NULL;
+	char *owned_provider = NULL;
+	char *owned_system_prompt = NULL;
+	char *api_key = NULL;
+	R2AI_ChatResponse *res = NULL;
 
-	// Check if rawtools mode is enabled
-	bool rawtools_enabled = r_config_get_b (core->config, "r2ai.auto.raw");
 	const char *provider = args.provider? args.provider: r_config_get (core->config, "r2ai.api");
 	if (!provider) {
 		provider = "gemini";
 	}
-	if (rawtools_enabled && args.tools && r_list_length (args.tools) > 0) {
-		return r2ai_rawtools_llmcall (cps, args);
-	}
 	R2AI_State *state = cps->data;
-	R2AI_ChatResponse *res = NULL;
 	if (!args.model) {
 		const char *config_model = r_config_get (core->config, "r2ai.model");
-		args.model = strdup (config_model? config_model: "");
+		owned_model = strdup (config_model? config_model: "");
+		args.model = owned_model;
 	}
 	if (!args.provider) {
-		args.provider = strdup (provider);
+		owned_provider = strdup (provider);
+		args.provider = owned_provider;
 	}
 
 	if (!args.max_tokens) {
@@ -65,39 +108,48 @@ R_IPI R2AI_ChatResponse *r2ai_llmcall(RCorePluginSession *cps, R2AIArgs args) {
 		const char *configtemp = r_config_get (core->config, "r2ai.temperature");
 		args.temperature = configtemp? atof (configtemp): 0;
 	}
+	args.deterministic = r_config_get_b (core->config, "r2ai.deterministic");
 
 	const R2AIProvider *prov = r2ai_get_provider (provider);
 	if (!prov) {
 		R_LOG_ERROR ("Unknown provider: %s", provider);
-		return NULL;
+		goto cleanup;
 	}
 
-	char *api_key = NULL;
-	if (prov->requires_api_key) {
+	if (is_vertex (prov)) {
+		const char *vtoken = r2ai_vertex_get_token (state);
+		if (!vtoken) {
+			goto cleanup;
+		}
+		args.api_key = vtoken;
+	} else if (prov->requires_api_key) {
 		api_key = r2ai_apikeys_get (provider);
 		if (api_key) {
 			args.api_key = api_key;
 		}
 	}
 	// Make sure we have an API key before proceeding
-	if (prov->requires_api_key) {
-		if (R_STR_ISEMPTY (args.api_key)) {
-			char *Provider = strdup (provider);
-			r_str_case (Provider, true);
-			R_LOG_ERROR ("No API key found for the %s provider. Use r2ai -K", provider);
-			free (Provider);
-			return NULL;
-		}
+	if (prov->requires_api_key && R_STR_ISEMPTY (args.api_key)) {
+		R_LOG_ERROR ("No API key found for the %s provider. Use r2ai -K", provider);
+		goto cleanup;
 	}
 
-	// Set system_prompt from config if it's not already set
-	if (!args.system_prompt) {
-		args.system_prompt = r_config_get (core->config, "r2ai.system");
+	args.system_prompt = r2ai_system_prompt (core, args.system_prompt);
+	int context_pullback = -1;
+	if (use_rawtools (core, prov, &args)) {
+		res = r2ai_rawtools_llmcall (cps, prov, args);
+		goto finish;
 	}
+
+	owned_system_prompt = r2ai_claw_system_prompt (args.system_prompt);
+	args.system_prompt = owned_system_prompt;
 	if (!args.messages) {
 		args.messages = r2ai_msgs_new ();
 	}
-	int context_pullback = -1;
+	if (args.input && args.messages && r_list_empty (args.messages)) {
+		R2AI_Message msg = { .role = "user", .content = (char *)args.input };
+		r2ai_msgs_add (args.messages, &msg);
+	}
 	// context and user message
 	if (args.input && r_config_get_b (core->config, "r2ai.data")) {
 		const int K = r_config_get_i (core->config, "r2ai.data.nth");
@@ -118,40 +170,24 @@ R_IPI R2AI_ChatResponse *r2ai_llmcall(RCorePluginSession *cps, R2AIArgs args) {
 			r_vdb_result_free (rs);
 		}
 		char *m = r_strbuf_drain (sb);
-		R2AI_Message msg = { .role = "user", .content = m };
-		context_pullback = r_list_length (args.messages);
-		r2ai_msgs_add (args.messages, &msg);
-		free (m);
-		// TODO: we can save the msg without context
-	}
-
-	// Add the rest of the messages one by one
-	if (!args.messages && args.input) {
-		R2AI_Message msg = { .role = "user", .content = (char *)args.input };
-		args.messages = r2ai_msgs_new ();
-		r2ai_msgs_add (args.messages, &msg);
+		int last = r_list_length (args.messages) - 1;
+		R2AI_Message *last_msg = (last >= 0)? r_list_get_n (args.messages, last): NULL;
+		if (last_msg && last_msg->role && !strcmp (last_msg->role, "user")) {
+			free ((char *)last_msg->content);
+			last_msg->content = m;
+			context_pullback = last;
+		} else {
+			R2AI_Message msg = { .role = "user", .content = m };
+			context_pullback = r_list_length (args.messages);
+			r2ai_msgs_add (args.messages, &msg);
+			free (m);
+		}
 	}
 
 	args.thinking_tokens = r_config_get_i (core->config, "r2ai.thinking_tokens");
 
-	const R2AIProvider *p = r2ai_get_provider (provider);
-	if (!p) {
-		return NULL;
-	}
-
-	switch (p->api_type) {
-	case R2AI_API_ANTHROPIC:
-		res = r2ai_anthropic (cps, args);
-		break;
-	case R2AI_API_GEMINI:
-		res = r2ai_gemini (cps, args);
-		break;
-	case R2AI_API_OPENAI_COMPATIBLE:
-	case R2AI_API_OLLAMA:
-	default:
-		res = r2ai_openai (cps, args);
-		break;
-	}
+	res = r2ai_send (cps, prov, args);
+finish:
 	if (context_pullback != -1) {
 		R2AI_Message *msg = r_list_get_n (args.messages, context_pullback);
 		free ((char *)msg->content);
@@ -163,105 +199,106 @@ R_IPI R2AI_ChatResponse *r2ai_llmcall(RCorePluginSession *cps, R2AIArgs args) {
 		*args.error = NULL;
 	}
 
+cleanup:
+	free (api_key);
+	free (owned_model);
+	free (owned_provider);
+	free (owned_system_prompt);
 	return res;
 }
 
-R_IPI const char *r2ai_get_provider_url(RCore *core, const char *provider) {
+static char *provider_api_url(const char *host, const char *api_root, bool preserve_root) {
+	char *url = r_str_startswith (host, "http")? strdup (host): r_str_newf ("http://%s", host);
+	if (!url) {
+		return NULL;
+	}
+	r_str_trim (url);
+	size_t len = strlen (url);
+	while (len > 0 && url[len - 1] == '/') {
+		url[--len] = 0;
+	}
+	const bool has_root = r_str_endswith (url, "/v1") || r_str_endswith (url, "/api");
+	if (has_root && (preserve_root || r_str_endswith (url, api_root))) {
+		return url;
+	}
+	if (has_root) {
+		url[len - 3] = 0;
+	}
+	char *res = r_str_newf ("%s%s", url, api_root);
+	free (url);
+	return res;
+}
+
+R_IPI char *r2ai_get_provider_url(RCore *core, const char *provider) {
 	const R2AIProvider *p = r2ai_get_provider (provider);
 	if (!p) {
 		return NULL;
 	}
 
-	// Handle providers that support custom baseurl
+	if (is_vertex (p)) {
+		return NULL;
+	}
+
+	const bool is_ollama = p->api_type == R2AI_API_OLLAMA;
+	const bool use_generate = is_ollama && is_generate_api (core);
+
 	if (p->supports_custom_baseurl) {
 		const char *host = r_config_get (core->config, "r2ai.baseurl");
 		if (R_STR_ISNOTEMPTY (host)) {
-			if (r_str_startswith (host, "http")) {
-				if (p->api_type == R2AI_API_OPENAI_COMPATIBLE) {
-					if (r_str_endswith (host, "/v1")) {
-						return r_str_newf ("%s", host);
-					}
-					return r_str_newf ("%s/v1", host);
-				}
-				return r_str_newf ("%s/api", host);
-			}
-			if (p->api_type == R2AI_API_OPENAI_COMPATIBLE) {
-				return r_str_newf ("http://%s/v1", host);
-			}
-			return r_str_newf ("http://%s/api", host);
+			const char *api_root = use_generate? "/api": "/v1";
+			return provider_api_url (host, api_root, is_ollama && !use_generate);
 		}
 	}
+	if (use_generate) {
+		return provider_api_url (p->url, "/api", false);
+	}
 
-	return p->url;
+	return p->url? strdup (p->url): NULL;
 }
 R_IPI RList *r2ai_fetch_available_models(RCore *core, const char *provider) {
 	if (!provider) {
 		return NULL;
 	}
-	const char *purl = r2ai_get_provider_url (core, provider);
+	const R2AIProvider *p = r2ai_get_provider (provider);
+	if (!p) {
+		return NULL;
+	}
+	if (is_vertex (p)) {
+		R_LOG_ERROR ("Model listing is not supported for Vertex AI providers");
+		return NULL;
+	}
+	char *purl = r2ai_get_provider_url (core, provider);
 	if (!purl) {
 		return NULL;
 	}
-
-	// Get API key for authentication (except for providers that don't require it)
-	char *api_key = NULL;
-	const R2AIProvider *p = r2ai_get_provider (provider);
-	if (p && p->requires_api_key) {
-		// Consolidated helper to fetch the API key from env or file
-		api_key = r2ai_apikeys_get (provider);
+	const bool gemini = p->api_type == R2AI_API_GEMINI;
+	const bool usetags = p->api_type == R2AI_API_OLLAMA && r_str_endswith (purl, "/api");
+	char *api_key = p->requires_api_key? r2ai_apikeys_get (provider): NULL;
+	if (gemini && !api_key) {
+		free (purl);
+		return NULL;
 	}
-
-	char *models_url = NULL;
-	int code = 0;
-	char *response = NULL;
-
-	// Special handling for Gemini
-	if (!strcmp (provider, "gemini")) {
-		if (!api_key) {
-			return NULL;
-		}
-		models_url = r_str_newf ("%s/models?key=%s", purl, api_key);
-		const char *headers[2] = { "Content-Type: application/json", NULL };
-		R_LOG_DEBUG ("GET %s", models_url);
-		response = r2ai_http_get (core, models_url, headers, &code, NULL);
-	} else {
-		// Create models endpoint URL
-		const R2AIProvider *prov = r2ai_get_provider (provider);
-		const bool usetags = (prov && prov->api_type == R2AI_API_OLLAMA);
-		models_url = r_str_newf ("%s/%s", purl, usetags? "tags": "models");
-
-		if (api_key) {
-			const char *headers[4] = { "Content-Type: application/json", NULL, NULL, NULL };
-			char *auth_header = NULL;
-			char *version_header = NULL;
-
-			const R2AIProvider *prov = r2ai_get_provider (provider);
-			if (prov && prov->api_type == R2AI_API_ANTHROPIC) {
-				// Anthropic uses different header format
-				auth_header = r_str_newf ("x-api-key: %s", api_key);
-				version_header = strdup ("anthropic-version: 2023-06-01");
-				headers[1] = auth_header;
-				headers[2] = version_header;
-			} else {
-				// Standard OpenAI-compatible format
-				auth_header = r_str_newf ("Authorization: Bearer %s", api_key);
-				headers[1] = auth_header;
-			}
-
-			// Make HTTP GET request
-			R_LOG_DEBUG ("GET %s Headers: %s", models_url, headers);
-			response = r2ai_http_get (core, models_url, headers, &code, NULL);
-			free (auth_header);
-			free (version_header);
+	char *models_url = r_str_newf ("%s/%s", purl, usetags? "tags": "models");
+	char *auth_header = NULL;
+	const char *headers[4] = { "Content-Type: application/json", NULL, NULL, NULL };
+	if (api_key) {
+		if (p->api_type == R2AI_API_ANTHROPIC) {
+			auth_header = r_str_newf ("x-api-key: %s", api_key);
+			headers[2] = "anthropic-version: 2023-06-01";
+		} else if (gemini) {
+			auth_header = r_str_newf ("x-goog-api-key: %s", api_key);
 		} else {
-			// We have no headers
-			R_LOG_DEBUG ("GET %s", models_url);
-			response = r2ai_http_get (core, models_url, NULL, &code, NULL);
+			auth_header = r_str_newf ("Authorization: Bearer %s", api_key);
 		}
+		headers[1] = auth_header;
 	}
-
+	R_LOG_DEBUG ("GET %s", models_url);
+	int code = 0;
+	char *response = r2ai_http_get (core, models_url, headers, &code, NULL);
+	free (auth_header);
 	free (models_url);
 	free (api_key);
+	free (purl);
 
 	if (!response || code != 200) {
 		R_LOG_DEBUG ("Failed to fetch models from %s (code: %d)", provider, code);
@@ -269,73 +306,27 @@ R_IPI RList *r2ai_fetch_available_models(RCore *core, const char *provider) {
 		return NULL;
 	}
 
-	// Parse JSON response
 	RList *models = r_list_newf (free);
-	if (!models) {
-		free (response);
-		return NULL;
-	}
-
 	RJson *json = r_json_parse (response);
-	if (json) {
-		const RJson *data = NULL;
-
-		if (!strcmp (provider, "gemini")) {
-			// Gemini has "models" array directly
-			data = r_json_get (json, "models");
-		} else {
-			const R2AIProvider *prov = r2ai_get_provider (provider);
-			const bool usetags = (prov && prov->api_type == R2AI_API_OLLAMA);
-			data = r_json_get (json, usetags? "models": "data");
+	const RJson *data = json? r_json_get (json, (gemini || usetags)? "models": "data"): NULL;
+	const char *key = gemini? "name": usetags? "model": "id";
+	const RJson *item;
+	for (item = (data && data->type == R_JSON_ARRAY)? data->children.first: NULL; item; item = item->next) {
+		const char *id = r_json_get_str (item, key);
+		if (R_STR_ISEMPTY (id)) {
+			continue;
 		}
-
-		if (data && data->type == R_JSON_ARRAY) {
-			const RJson *model_item = data->children.first;
-			while (model_item) {
-				const RJson *id = NULL;
-				char *model_id = NULL;
-
-				if (!strcmp (provider, "gemini")) {
-					// Gemini: extract model ID from "name" field (e.g., "models/gemini-1.5-flash" -> "gemini-1.5-flash")
-					const RJson *name = r_json_get (model_item, "name");
-					if (name && name->type == R_JSON_STRING && R_STR_ISNOTEMPTY (name->str_value)) {
-						char *s = strdup (name->str_value);
-						RList *parts = r_str_split_list (s, "/", 0);
-						if (parts && r_list_length (parts) > 1) {
-							model_id = strdup ((char *)r_list_get_n (parts, r_list_length (parts) - 1));
-						} else {
-							model_id = strdup (name->str_value);
-						}
-						r_list_free (parts);
-						// Only include Gemini models
-						if (!strstr (model_id, "gemini")) {
-							free (model_id);
-							model_id = NULL;
-						}
-						free (s);
-					}
-				} else {
-					const R2AIProvider *prov = r2ai_get_provider (provider);
-					if (prov && prov->api_type == R2AI_API_OLLAMA) {
-						id = r_json_get (model_item, "model");
-					} else {
-						id = r_json_get (model_item, "id");
-					}
-					if (id && id->type == R_JSON_STRING && R_STR_ISNOTEMPTY (id->str_value)) {
-						model_id = strdup (id->str_value);
-					}
-				}
-
-				if (model_id) {
-					R_LOG_DEBUG ("Model: %s", model_id);
-					r_list_append (models, model_id);
-				}
-				model_item = model_item->next;
+		if (gemini) {
+			// "models/gemini-1.5-flash" -> "gemini-1.5-flash", skipping non gemini models
+			const char *slash = strrchr (id, '/');
+			id = slash? slash + 1: id;
+			if (!strstr (id, "gemini")) {
+				continue;
 			}
 		}
-		r_json_free (json);
+		r_list_append (models, strdup (id));
 	}
-
+	r_json_free (json);
 	free (response);
 	return models;
 }
@@ -354,20 +345,16 @@ R_IPI void r2ai_list_providers(RCore *core, RStrBuf *sb) {
 }
 
 R_IPI void r2ai_refresh_embeddings(RCorePluginSession *cps) {
-	RCore *core = cps->core;
 	R2AI_State *state = cps->data;
-	RListIter *iter, *iter2;
-	char *line;
-	char *file;
-	// refresh embeddings database
 	r_vdb_free (state->db);
 	state->db = r_vdb_new (R2AI_DEFAULT_VECTORS);
-	// enumerate .txt files in directory
-	const char *path = r_config_get (core->config, "r2ai.data.path");
+	const char *path = r_config_get (cps->core->config, "r2ai.data.path");
 	RList *files = r_sys_dir (path);
 	if (r_list_empty (files)) {
 		R_LOG_WARN ("Cannot find any file in r2ai.data.path");
 	}
+	RListIter *iter, *iter2;
+	char *file, *line;
 	r_list_foreach (files, iter, file) {
 		if (!r_str_endswith (file, ".txt")) {
 			continue;
@@ -375,19 +362,18 @@ R_IPI void r2ai_refresh_embeddings(RCorePluginSession *cps) {
 		R_LOG_DEBUG ("Index %s", file);
 		char *filepath = r_file_new (path, file, NULL);
 		char *text = r_file_slurp (filepath, NULL);
-		if (text) {
-			R_LOG_DEBUG ("Index %s", file);
-			RList *lines = r_str_split_list (text, "\n", -1);
-			r_list_foreach (lines, iter2, line) {
-				if (r_str_trim_head_ro (line)[0] == 0) {
-					continue;
-				}
-				r_vdb_insert (state->db, line);
-				R_LOG_DEBUG ("Insert %s", line);
-			}
-			r_list_free (lines);
-		}
 		free (filepath);
+		if (!text) {
+			continue;
+		}
+		RList *lines = r_str_split_list (text, "\n", -1);
+		r_list_foreach (lines, iter2, line) {
+			if (*r_str_trim_head_ro (line)) {
+				r_vdb_insert (state->db, line);
+			}
+		}
+		r_list_free (lines);
+		free (text);
 	}
 	r_list_free (files);
 }

src/main.c

@@ -1,4 +1,4 @@
-/* Copyright r2ai - 2023-2025 - pancake */
+/* Copyright r2ai - 2023-2026 - pancake */
 
 #include "r2ai_priv.h"
 
@@ -7,17 +7,25 @@
 
 static void show_help() {
 	printf (
-		"Usage: r2ai [-vhp:m:q:Eb:Kc:f:iew] <prompt>\n"
+		"Usage: r2ai [-vhdorEwK] [-p provider] [-m model] [-q query]\n"
+		"            [-b url] [-c command] [-f file] [-s addr]\n"
+		"            [-i script] [-e var=value] [prompt]\n"
 		"  -v           Show version information\n"
 		"  -h           Show this help message\n"
+		"  -d           Decompile current function\n"
+		"  -d <query>   Ask a question on the current function\n"
+		"  -do          Decompile current function with source offsets\n"
+		"  -dr          Decompile current function recursively\n"
 		"  -p <provider> Select LLM provider\n"
 		"  -m <model>   Select LLM model\n"
 		"  -q <query>   Execute predefined prompt query (can be used multiple times)\n"
+		"  -qj          List predefined prompt queries as JSON\n"
 		"  -b <url>     Set base URL for provider API\n"
 		"  -c <command> Execute command after loading file (can be used multiple times)\n"
 		"  -f <file>    Load file with r2 API\n"
+		"  -s <addr>    Seek to address after loading file\n"
 		"  -i <script>  Load and interpret script file before executing commands\n"
-		"  -e <var=value> Set configuration variable\n"
+		"  -e <var=value> Set configuration variable (same as r2ai.var)\n"
 		"  -E           Edit the r2ai rc file\n"
 		"  -K           Edit the API keys file\n"
 		"  -w           Launch interactive setup wizard\n");
@@ -27,6 +35,21 @@ static void show_version() {
 	printf ("r2ai " R2AI_VERSION "\n");
 }
 
+static char *join_argv(int argc, const char **argv, int start) {
+	if (start >= argc) {
+		return strdup ("");
+	}
+	int i;
+	RStrBuf *sb = r_strbuf_new ("");
+	for (i = start; i < argc; i++) {
+		if (i > start) {
+			r_strbuf_append (sb, " ");
+		}
+		r_strbuf_append (sb, argv[i]);
+	}
+	return r_strbuf_drain (sb);
+}
+
 static char *build_conversation(RList *conversation) {
 	RListIter *iter;
 	char *msg;
@@ -73,17 +96,17 @@ static void r2ai_repl(RCorePluginSession *cps, const char *provider, const char
 				.model = model,
 				.error = &err,
 				.dorag = false,
-			};
+};
 			char *res = r2ai (cps, args);
 			if (res) {
 				if (r_config_get_b (core->config, "r2ai.clippy")) {
 					char *cmd = r_str_newf (CLIPPY " %s", res);
-					r_core_cmd_call (core, cmd);
+					r_core_call (core, cmd);
 					free (cmd);
 				} else {
 					r_cons_println (core->cons, res);
-					r_list_append (conversation, r_str_newf ("Assistant: %s", res));
 				}
+				r_list_append (conversation, r_str_newf ("Assistant: %s", res));
 				free (res);
 			}
 			if (err) {
@@ -102,7 +125,12 @@ int main(int argc, const char **argv) {
 	const char *model = NULL;
 	const char *baseurl = NULL;
 	const char *filename = NULL;
+	const char *seekaddr = NULL;
 	const char *scriptfile = NULL;
+	bool list_queries_json = false;
+	bool decompile = false;
+	bool decompile_recursive = false;
+	bool decompile_offsets = false;
 	RList *conversation = r_list_newf (free);
 	RList *queries = r_list_newf (free);
 	RList *commands = r_list_newf (free);
@@ -113,9 +141,26 @@ int main(int argc, const char **argv) {
 	r2ai_init (&cps);
 
 	RGetopt opt;
-	r_getopt_init (&opt, argc, argv, "vhp:m:q:Eb:Kc:f:ie:w");
+	r_getopt_init (&opt, argc, argv, "vhdorp:m:q:Eb:Kc:f:s:i:e:w");
 	while ((c = r_getopt_next (&opt)) != -1) {
 		switch (c) {
+		case 'd':
+			decompile = true;
+			break;
+		case 'o':
+			if (!decompile) {
+				show_help ();
+				goto beach;
+			}
+			decompile_offsets = true;
+			break;
+		case 'r':
+			if (!decompile) {
+				show_help ();
+				goto beach;
+			}
+			decompile_recursive = true;
+			break;
 		case 'p':
 			provider = opt.arg;
 			break;
@@ -123,7 +168,11 @@ int main(int argc, const char **argv) {
 			model = opt.arg;
 			break;
 		case 'q':
-			r_list_append (queries, strdup (opt.arg));
+			if (!strcmp (opt.arg, "j")) {
+				list_queries_json = true;
+			} else {
+				r_list_append (queries, strdup (opt.arg));
+			}
 			break;
 		case 'b':
 			baseurl = opt.arg;
@@ -134,6 +183,9 @@ int main(int argc, const char **argv) {
 		case 'f':
 			filename = opt.arg;
 			break;
+		case 's':
+			seekaddr = opt.arg;
+			break;
 		case 'i':
 			scriptfile = opt.arg;
 			break;
@@ -145,7 +197,11 @@ int main(int argc, const char **argv) {
 					*equals = '\0';
 					const char *var = var_value;
 					const char *value = equals + 1;
-					r_config_set (core->config, var, value);
+					char *key = r_str_startswith (var, "r2ai")
+						? strdup (var)
+						: r_str_newf ("r2ai.%s", var);
+					r_config_set (core->config, key, value);
+					free (key);
 				} else {
 					r_cons_println (core->cons, "Invalid format for -e. Use var=value.");
 				}
@@ -155,7 +211,7 @@ int main(int argc, const char **argv) {
 		case 'E':
 			{
 				char *rc_path = r_file_home (".config/r2ai/rc");
-				r_cons_editor (core->cons, rc_path, NULL);
+				r2ai_cons_editor (core->cons, rc_path, NULL);
 				free (rc_path);
 				goto beach;
 			}
@@ -184,21 +240,37 @@ int main(int argc, const char **argv) {
 		}
 	}
 
+	if (list_queries_json) {
+		r2ai_cmd_qj (&cps, NULL);
+		r_cons_flush (core->cons);
+		goto beach;
+	}
+	if (decompile && decompile_offsets && decompile_recursive) {
+		r_cons_println (core->cons, "Cannot combine -do and -dr.");
+		r_cons_flush (core->cons);
+		goto beach;
+	}
+
 	if (baseurl) {
 		r_config_set (core->config, "r2ai.baseurl", baseurl);
 	}
 	if (filename) {
 		char *cmd = r_str_newf ("o %s", filename);
-		r_core_cmd_call (core, cmd);
+		r_core_call (core, cmd);
 		free (cmd);
 		// TODO: use the C api instead of the 'o' command and check for error if file exists etc
+		if (seekaddr) {
+			char *seekcmd = r_str_newf ("s %s", seekaddr);
+			r_core_call (core, seekcmd);
+			free (seekcmd);
+		}
 	} else {
 		r_core_cmd0 (core, CLIPPY " Interesting, no files to analyse. Use -f or :o");
 		r_cons_flush (core->cons);
 	}
 	if (scriptfile) {
 		char *cmd = r_str_newf (". %s", scriptfile);
-		r_core_cmd_call (core, cmd);
+		r_core_call (core, cmd);
 		free (cmd);
 	}
 	if (!r_list_empty (commands)) {
@@ -220,6 +292,16 @@ int main(int argc, const char **argv) {
 	if (model) {
 		r_config_set (core->config, "r2ai.model", model);
 	}
+	if (decompile) {
+		const char *dflag = decompile_offsets? "-do": (decompile_recursive? "-dr": "-d");
+		char *query = join_argv (argc, argv, opt.ind);
+		char *dcmd = R_STR_ISNOTEMPTY (query)? r_str_newf ("%s %s", dflag, query): strdup (dflag);
+		cmd_r2ai (&cps, dcmd);
+		r_cons_flush (core->cons);
+		free (dcmd);
+		free (query);
+		goto beach;
+	}
 
 	// Process queries if any
 	if (!r_list_empty (queries)) {
@@ -232,8 +314,7 @@ int main(int argc, const char **argv) {
 		if (opt.ind >= argc) {
 			r2ai_repl (&cps, provider, model, conversation);
 		} else {
-			const char *prompt = argv[opt.ind];
-
+			char *prompt = join_argv (argc, argv, opt.ind);
 			char *err = NULL;
 			R2AIArgs args = {
 				.input = prompt,
@@ -241,13 +322,17 @@ int main(int argc, const char **argv) {
 				.model = model,
 				.error = &err,
 				.dorag = false,
-			};
+};
 			char *res = r2ai (&cps, args);
 			if (res) {
 				r_cons_println (core->cons, res);
 				free (res);
 			}
-			free (err);
+			if (err) {
+				R_LOG_ERROR ("%s", err);
+				free (err);
+			}
+			free (prompt);
 			r_cons_flush (core->cons);
 		}
 	}

src/markdown.c

@@ -1,3 +1,5 @@
+/* Copyright r2ai - 2023-2026 - pancake */
+
 #include <r_core.h>
 #include <r_cons.h>
 #include <r_types.h>
@@ -14,10 +16,6 @@ typedef enum {
 	STRIKE
 } MarkdownState;
 
-// Global theme instance
-static RMarkdownTheme current_theme;
-static bool theme_initialized = false;
-
 // Set default theme values
 R_API RMarkdownTheme r2ai_markdown_theme_default(void) {
 	RMarkdownTheme theme = {
@@ -46,52 +44,59 @@ R_API RMarkdownTheme r2ai_markdown_theme_default(void) {
 	return theme;
 }
 
-// Initialize theme if not already done
-static void ensure_theme_initialized(void) {
-	if (!theme_initialized) {
-		current_theme = r2ai_markdown_theme_default ();
-		theme_initialized = true;
+R_API void r2ai_markdown_init(RMarkdown *md) {
+	if (!md) {
+		return;
+	}
+	if (!md->initialized) {
+		md->theme = r2ai_markdown_theme_default ();
+		md->initialized = true;
 	}
 }
 
 // Set a custom theme
-R_API void r2ai_markdown_set_theme(const RMarkdownTheme *theme) {
-	if (!theme) {
-		current_theme = r2ai_markdown_theme_default ();
-	} else {
-		current_theme = *theme;
+R_API void r2ai_markdown_set_theme(RMarkdown *md, const RMarkdownTheme *theme) {
+	if (!md) {
+		return;
 	}
-	theme_initialized = true;
+	if (!theme) {
+		md->theme = r2ai_markdown_theme_default ();
+	} else {
+		md->theme = *theme;
+	}
+	md->initialized = true;
 }
 
 // Get the current theme
-R_API const RMarkdownTheme *r2ai_markdown_get_theme(void) {
-	ensure_theme_initialized ();
-	return &current_theme;
+R_API const RMarkdownTheme *r2ai_markdown_get_theme(RMarkdown *md) {
+	if (!md) {
+		return NULL;
+	}
+	r2ai_markdown_init (md);
+	return &md->theme;
 }
 
-static void append_formatted(RStrBuf *sb, const char *text, int len, MarkdownState state) {
-	ensure_theme_initialized ();
+static void append_formatted(RStrBuf *sb, const RMarkdownTheme *theme, const char *text, int len, MarkdownState state) {
 	switch (state) {
 	case BOLD:
-		r_strbuf_append (sb, current_theme.bold);
+		r_strbuf_append (sb, theme->bold);
 		r_strbuf_append_n (sb, text, len);
-		r_strbuf_append (sb, current_theme.reset);
+		r_strbuf_append (sb, theme->reset);
 		break;
 	case ITALIC:
-		r_strbuf_append (sb, current_theme.italic);
+		r_strbuf_append (sb, theme->italic);
 		r_strbuf_append_n (sb, text, len);
-		r_strbuf_append (sb, current_theme.reset);
+		r_strbuf_append (sb, theme->reset);
 		break;
 	case CODE:
-		r_strbuf_append (sb, current_theme.code_inline);
+		r_strbuf_append (sb, theme->code_inline);
 		r_strbuf_append_n (sb, text, len);
-		r_strbuf_append (sb, current_theme.reset);
+		r_strbuf_append (sb, theme->reset);
 		break;
 	case STRIKE:
-		r_strbuf_append (sb, current_theme.strike);
+		r_strbuf_append (sb, theme->strike);
 		r_strbuf_append_n (sb, text, len);
-		r_strbuf_append (sb, current_theme.reset);
+		r_strbuf_append (sb, theme->reset);
 		break;
 	default:
 		r_strbuf_append_n (sb, text, len);
@@ -109,12 +114,17 @@ static int count_indent(const char *str) {
 	return count;
 }
 
-R_API char *r2ai_markdown(const char *markdown) {
+R_API char *r2ai_markdown(RMarkdown *md, const char *markdown) {
 	if (!markdown) {
 		return NULL;
 	}
 
-	ensure_theme_initialized ();
+	RMarkdown fallback = { 0 };
+	if (!md) {
+		md = &fallback;
+	}
+	r2ai_markdown_init (md);
+	const RMarkdownTheme *theme = &md->theme;
 	RStrBuf *sb = r_strbuf_new ("");
 	if (!sb) {
 		return NULL;
@@ -134,7 +144,7 @@ R_API char *r2ai_markdown(const char *markdown) {
 			in_code_block = !in_code_block;
 			if (in_code_block) {
 				// Add padding for full-width code blocks
-				r_strbuf_appendf (sb, "%s\033[K", current_theme.code_block);
+				r_strbuf_appendf (sb, "%s\033[K", theme->code_block);
 
 				// Skip language specifier if present
 				while (*p && *p != '\n') {
@@ -142,7 +152,7 @@ R_API char *r2ai_markdown(const char *markdown) {
 				}
 
 			} else {
-				r_strbuf_appendf (sb, "%s\n", current_theme.reset);
+				r_strbuf_appendf (sb, "%s\n", theme->reset);
 			}
 			start = p;
 			continue;
@@ -152,7 +162,7 @@ R_API char *r2ai_markdown(const char *markdown) {
 			if (*p == '\n') {
 				r_strbuf_append_n (sb, start, p - start);
 				// Add the \033[K (EL - Erase in Line) sequence to extend the background to the end of line
-				r_strbuf_appendf (sb, "\n%s\033[K", current_theme.code_block);
+				r_strbuf_appendf (sb, "\n%s\033[K", theme->code_block);
 				start = p + 1;
 				line_start = 1;
 			}
@@ -181,30 +191,30 @@ R_API char *r2ai_markdown(const char *markdown) {
 				// Add heading with appropriate size
 				switch (level) {
 				case 1:
-					r_strbuf_append (sb, current_theme.heading1);
+					r_strbuf_append (sb, theme->heading1);
 					break;
 				case 2:
-					r_strbuf_append (sb, current_theme.heading2);
+					r_strbuf_append (sb, theme->heading2);
 					break;
 				case 3:
-					r_strbuf_append (sb, current_theme.heading3);
+					r_strbuf_append (sb, theme->heading3);
 					break;
 				case 4:
-					r_strbuf_append (sb, current_theme.heading4);
+					r_strbuf_append (sb, theme->heading4);
 					break;
 				case 5:
-					r_strbuf_append (sb, current_theme.heading5);
+					r_strbuf_append (sb, theme->heading5);
 					break;
 				case 6:
-					r_strbuf_append (sb, current_theme.heading6);
+					r_strbuf_append (sb, theme->heading6);
 					break;
 				default:
-					r_strbuf_append (sb, current_theme.heading1);
+					r_strbuf_append (sb, theme->heading1);
 					break;
 				}
 
 				r_strbuf_append_n (sb, start, heading_end - start);
-				r_strbuf_appendf (sb, "%s\n", current_theme.reset);
+				r_strbuf_appendf (sb, "%s\n", theme->reset);
 
 				if (*heading_end) {
 					p = heading_end + 1;
@@ -228,7 +238,7 @@ R_API char *r2ai_markdown(const char *markdown) {
 
 			// Skip the indent spaces
 			if (indent_level > 0) {
-				append_formatted (sb, start, p - start, state);
+				append_formatted (sb, theme, start, p - start, state);
 				for (int i = 0; i < indent_level; i++) {
 					r_strbuf_append (sb, " ");
 				}
@@ -238,8 +248,8 @@ R_API char *r2ai_markdown(const char *markdown) {
 
 			// Handle bullet lists
 			if (*p == '-' && p[1] == ' ') {
-				append_formatted (sb, start, p - start, state);
-				r_strbuf_append (sb, current_theme.list_bullet);
+				append_formatted (sb, theme, start, p - start, state);
+				r_strbuf_append (sb, theme->list_bullet);
 				p += 2;
 				start = p;
 				line_start = 0;
@@ -254,15 +264,14 @@ R_API char *r2ai_markdown(const char *markdown) {
 				}
 
 				if (*p == '.' && p[1] == ' ') {
-					append_formatted (sb, start, digit_start - start, state);
+					append_formatted (sb, theme, start, digit_start - start, state);
 					int num_len = p - digit_start;
 					char num_buf[16] = { 0 };
 					r_str_ncpy (num_buf, digit_start, num_len < 15? num_len + 1: 15);
 
 					// Use the actual number from markdown and format it
 					char formatted_number[32];
-					snprintf (formatted_number, sizeof (formatted_number),
-						current_theme.list_number, num_buf);
+					snprintf (formatted_number, sizeof (formatted_number), theme->list_number, num_buf);
 					r_strbuf_append (sb, formatted_number);
 
 					p += 2; // Skip ". "
@@ -276,12 +285,12 @@ R_API char *r2ai_markdown(const char *markdown) {
 
 			// Handle checklists
 			if ((*p == '[' && (p[1] == ' ' || p[1] == 'x') && p[2] == ']' && p[3] == ' ')) {
-				append_formatted (sb, start, p - start, state);
+				append_formatted (sb, theme, start, p - start, state);
 
 				if (p[1] == 'x') {
-					r_strbuf_append (sb, current_theme.checkbox_checked);
+					r_strbuf_append (sb, theme->checkbox_checked);
 				} else {
-					r_strbuf_append (sb, current_theme.checkbox_unchecked);
+					r_strbuf_append (sb, theme->checkbox_unchecked);
 				}
 
 				p += 4; // Skip "[ ] " or "[x] "
@@ -300,7 +309,7 @@ R_API char *r2ai_markdown(const char *markdown) {
 			if (p > markdown && !isalnum (*(p - 1)) && isalnum (p[1 + double_marker])) {
 				// Opening marker
 				if (p > start) {
-					append_formatted (sb, start, p - start, state);
+					append_formatted (sb, theme, start, p - start, state);
 				}
 				p += double_marker? 2: 1;
 				start = p;
@@ -316,7 +325,7 @@ R_API char *r2ai_markdown(const char *markdown) {
 				}
 			} else if (p > markdown && isalnum (*(p - 1)) && !isalnum (p[1 + double_marker])) {
 				// Closing marker
-				append_formatted (sb, start, p - start, state);
+				append_formatted (sb, theme, start, p - start, state);
 				p += double_marker? 2: 1;
 				start = p;
 				state = NORMAL;
@@ -324,7 +333,7 @@ R_API char *r2ai_markdown(const char *markdown) {
 				p++;
 			}
 		} else if (*p == '\n') {
-			append_formatted (sb, start, p - start, state);
+			append_formatted (sb, theme, start, p - start, state);
 			r_strbuf_append (sb, "\n");
 			p++;
 			start = p;
@@ -338,7 +347,7 @@ R_API char *r2ai_markdown(const char *markdown) {
 
 	// Handle any remaining text
 	if (p > start) {
-		append_formatted (sb, start, p - start, state);
+		append_formatted (sb, theme, start, p - start, state);
 	}
 
 	char *result = r_strbuf_drain (sb);

src/markdown.h

@@ -1,50 +1,59 @@
 #ifndef R2_MARKDOWN_H
 #define R2_MARKDOWN_H
 
+#include <stdbool.h>
 #include <r_types.h>
 
 /* Theme configuration for markdown rendering */
 typedef struct r_markdown_theme_t {
-    // Text styling
-    const char *bold;             // Bold text
-    const char *italic;           // Italic text
-    const char *strike;           // Strikethrough text
-    const char *code_inline;      // Inline code
-    const char *code_block;       // Code block background
-    
-    // Heading colors (h1-h6)
-    const char *heading1;
-    const char *heading2;
-    const char *heading3;
-    const char *heading4;
-    const char *heading5;
-    const char *heading6;
-    
-    // List items
-    const char *list_bullet;      // Bullet character for unordered lists
-    const char *list_number;      // Number formatting for ordered lists
-    
-    // Checkbox states
-    const char *checkbox_checked;
-    const char *checkbox_unchecked;
-    
-    // Reset code
-    const char *reset;
+	// Text styling
+	const char *bold; // Bold text
+	const char *italic; // Italic text
+	const char *strike; // Strikethrough text
+	const char *code_inline; // Inline code
+	const char *code_block; // Code block background
+
+	// Heading colors (h1-h6)
+	const char *heading1;
+	const char *heading2;
+	const char *heading3;
+	const char *heading4;
+	const char *heading5;
+	const char *heading6;
+
+	// List items
+	const char *list_bullet; // Bullet character for unordered lists
+	const char *list_number; // Number formatting for ordered lists
+
+	// Checkbox states
+	const char *checkbox_checked;
+	const char *checkbox_unchecked;
+
+	// Reset code
+	const char *reset;
 } RMarkdownTheme;
 
+typedef struct r_markdown_t {
+	RMarkdownTheme theme;
+	bool initialized;
+} RMarkdown;
+
 /* Default built-in theme */
 R_API RMarkdownTheme r2ai_markdown_theme_default(void);
 
+/* Initialize a markdown rendering context with the default theme */
+R_API void r2ai_markdown_init(RMarkdown *md);
+
 /* Set a custom theme for markdown rendering */
-R_API void r2ai_markdown_set_theme(const RMarkdownTheme *theme);
+R_API void r2ai_markdown_set_theme(RMarkdown *md, const RMarkdownTheme *theme);
 
 /* Get the current theme */
-R_API const RMarkdownTheme *r2ai_markdown_get_theme(void);
+R_API const RMarkdownTheme *r2ai_markdown_get_theme(RMarkdown *md);
 
 /* Render markdown text for terminal display with ANSI color codes
  * @param markdown: Input markdown string
  * @return: Dynamically allocated string with rendered markdown (must be freed by caller)
  */
-R_API char *r2ai_markdown(const char *markdown);
+R_API char *r2ai_markdown(RMarkdown *md, const char *markdown);
 
-#endif /* R2_MARKDOWN_H */ 
+#endif /* R2_MARKDOWN_H */

src/meson.build

@@ -1,5 +1,5 @@
 project('r2ai', 'c',
-  version : '1.3.4',
+  version : '1.5.4',
   default_options : ['warning_level=2', 'c_std=c11']
 )
 
@@ -63,9 +63,12 @@ common_deps = r2_deps + [cc.find_library('m', required: false)]
 
 srcs = files(
   'anthropic.c',
+  'async.c',
   'auto.c',
   'apikeys.c',
+  'claw.c',
   'gemini.c',
+  'vertex.c',
   'http.c',
   'json.c',
   'wizard.c',
@@ -80,13 +83,14 @@ srcs = files(
   'vdb.c',
 )
 
-shared_lib = shared_library('r2ai', srcs,
+r2ai_lib = library('r2ai', srcs,
   install: true,
   dependencies: common_deps,
+  pic: true,
 )
 
 executable('r2ai', 'main.c',
-  link_with: shared_lib,
+  link_with: r2ai_lib,
   install: true,
   dependencies: common_deps,
 )

src/messages.c

@@ -1,4 +1,4 @@
-/* r2ai - Copyright 2023-2025 dnakov, pancake */
+/* r2ai - Copyright 2023-2026 dnakov, pancake */
 
 #include "r2ai.h"
 
@@ -11,301 +11,122 @@ R_API void r2ai_tool_call_free(R2AI_ToolCall *tc) {
 	}
 }
 
+static void block_free(R2AI_ContentBlock *b) {
+	free (b->type);
+	free (b->id);
+	free (b->name);
+	free (b->input);
+	free (b->data);
+	free (b->thinking);
+	free (b->signature);
+	free (b->text);
+	free (b);
+}
+
+static R2AI_ContentBlock *block_dup(const R2AI_ContentBlock *b) {
+	R2AI_ContentBlock *d = R_NEW0 (R2AI_ContentBlock);
+	d->type = r_str_new (b->type);
+	d->id = r_str_new (b->id);
+	d->name = r_str_new (b->name);
+	d->input = r_str_new (b->input);
+	d->data = r_str_new (b->data);
+	d->thinking = r_str_new (b->thinking);
+	d->signature = r_str_new (b->signature);
+	d->text = r_str_new (b->text);
+	return d;
+}
+
+static R2AI_ToolCall *tc_dup(const R2AI_ToolCall *tc) {
+	R2AI_ToolCall *d = R_NEW0 (R2AI_ToolCall);
+	d->id = r_str_new (tc->id);
+	d->name = r_str_new (tc->name);
+	d->arguments = r_str_new (tc->arguments);
+	return d;
+}
+
 R_API RList *r2ai_content_blocks_new(void) {
-	RList *cb = r_list_new ();
-	if (!cb) {
-		return NULL;
-	}
-	cb->free = (RListFree)free; // ContentBlocks contain pointers to structs, not structs themselves
-	return cb;
+	return r_list_newf ((RListFree)block_free);
 }
 
-R_API void r2ai_content_blocks_free(RList *cb) {
-	if (!cb) {
-		return;
-	}
-	r_list_free (cb);
-}
-
-R_API void r2ai_message_free(R2AI_Message *msg) {
+R_API void r2ai_message_fini(R2AI_Message *msg) {
 	if (!msg) {
 		return;
 	}
 	free (msg->role);
-	free ((void *)msg->content); // maybe double free
+	free ((void *)msg->content);
 	free (msg->reasoning_content);
 	free (msg->tool_call_id);
+	r_list_free (msg->tool_calls);
+	r_list_free (msg->content_blocks);
+	memset (msg, 0, sizeof (*msg));
+}
 
-	// Free tool calls
-	if (msg->tool_calls) {
-		r_list_free (msg->tool_calls);
-	}
-
-	if (msg->content_blocks) {
-		r2ai_content_blocks_free (msg->content_blocks);
-	}
+R_API void r2ai_message_free(R2AI_Message *msg) {
+	r2ai_message_fini (msg);
 	free (msg);
 }
 
-// Conversation is now stored in R2AI_State
-
-R_API void r2ai_conversation_init(R2AI_State *state) {
-	if (!state || state->conversation) {
-		// Already initialized or invalid state
+R_IPI void r2ai_chat_response_free(R2AI_ChatResponse *res) {
+	if (!res) {
 		return;
 	}
+	r2ai_message_free ((R2AI_Message *)res->message);
+	free ((void *)res->usage);
+	free (res);
+}
 
-	state->conversation = r2ai_msgs_new ();
+R_API void r2ai_conversation_init(R2AI_State *state) {
+	if (state && !state->conversation) {
+		state->conversation = r2ai_msgs_new ();
+	}
 }
 
 R_API RList *r2ai_conversation_get(R2AI_State *state) {
 	return state? state->conversation: NULL;
 }
 
-// Create a new temporary messages container
 R_API RList *r2ai_msgs_new(void) {
-	RList *msgs = r_list_new ();
-	if (!msgs) {
-		return NULL;
-	}
-	msgs->free = (RListFree)r2ai_message_free;
-	return msgs;
+	return r_list_newf ((RListFree)r2ai_message_free);
 }
 
 R_API void r2ai_msgs_free(RList *msgs) {
-	if (msgs) {
-		r_list_free (msgs);
-	}
+	r_list_free (msgs);
 }
 
-// Free the conversation when plugin is unloaded
 R_API void r2ai_conversation_free(R2AI_State *state) {
-	if (state && state->conversation) {
-		r2ai_msgs_free (state->conversation);
+	if (state) {
+		r_list_free (state->conversation);
 		state->conversation = NULL;
 	}
 }
 
-// Clear messages in a container without freeing the container itself
 R_API void r2ai_msgs_clear(RList *msgs) {
-	if (!msgs) {
-		return;
-	}
 	r_list_purge (msgs);
 }
 
+// append a deep copy of the message
 R_API bool r2ai_msgs_add(RList *msgs, const R2AI_Message *msg) {
 	if (!msgs || !msg) {
 		return false;
 	}
-
-	R2AI_Message *new_msg = R_NEW0 (R2AI_Message);
-	new_msg->role = msg->role? strdup (msg->role): NULL;
-	new_msg->content = msg->content? strdup (msg->content): NULL;
-	new_msg->reasoning_content = msg->reasoning_content? strdup (msg->reasoning_content): NULL;
-
+	R2AI_Message *m = R_NEW0 (R2AI_Message);
+	m->role = r_str_new (msg->role);
+	m->content = r_str_new (msg->content);
+	m->reasoning_content = r_str_new (msg->reasoning_content);
+	m->tool_call_id = r_str_new (msg->tool_call_id);
 	if (msg->content_blocks) {
-		RList *cb = r2ai_content_blocks_new ();
-		if (!cb) {
-			r2ai_message_free (new_msg);
-			free (new_msg);
-			return false;
-		}
-		RListIter *iter;
-		R2AI_ContentBlock *src;
-		r_list_foreach (msg->content_blocks, iter, src) {
-			R2AI_ContentBlock *dst = R_NEW0 (R2AI_ContentBlock);
-			dst->type = src->type? strdup (src->type): NULL;
-			dst->data = src->data? strdup (src->data): NULL;
-			dst->thinking = src->thinking? strdup (src->thinking): NULL;
-			dst->signature = src->signature? strdup (src->signature): NULL;
-			dst->text = src->text? strdup (src->text): NULL;
-			dst->id = src->id? strdup (src->id): NULL;
-			dst->name = src->name? strdup (src->name): NULL;
-			dst->input = src->input? strdup (src->input): NULL;
-			r_list_append (cb, dst);
-		}
-		new_msg->content_blocks = cb;
+		m->content_blocks = r_list_clone (msg->content_blocks, (RListClone)block_dup);
+		m->content_blocks->free = (RListFree)block_free;
 	}
-
-	new_msg->tool_call_id = msg->tool_call_id? strdup (msg->tool_call_id): NULL;
-	new_msg->tool_calls = r_list_new ();
-	if (!new_msg->tool_calls) {
-		r2ai_message_free (new_msg);
-		free (new_msg);
-		return false;
-	}
-	new_msg->tool_calls->free = (RListFree)r2ai_tool_call_free;
-
-	// Copy tool calls if any
-	if (msg->tool_calls) {
-		RListIter *iter;
-		R2AI_ToolCall *src_tc;
-		r_list_foreach (msg->tool_calls, iter, src_tc) {
-			R2AI_ToolCall *dst_tc = R_NEW0 (R2AI_ToolCall);
-			dst_tc->name = src_tc->name? strdup (src_tc->name): NULL;
-			dst_tc->arguments = src_tc->arguments? strdup (src_tc->arguments): NULL;
-			dst_tc->id = src_tc->id? strdup (src_tc->id): NULL;
-			r_list_append (new_msg->tool_calls, dst_tc);
-		}
-	}
-
-	r_list_append (msgs, new_msg);
+	m->tool_calls = msg->tool_calls
+		? r_list_clone (msg->tool_calls, (RListClone)tc_dup)
+		: r_list_new ();
+	m->tool_calls->free = (RListFree)r2ai_tool_call_free;
+	r_list_append (msgs, m);
 	return true;
 }
 
-R_API bool r2ai_msgs_add_tool_call(RList *msgs, const R2AI_ToolCall *tc) {
-	if (!msgs || !tc || r_list_empty (msgs)) {
-		return false;
-	}
-
-	R2AI_Message *msg = r_list_get_n (msgs, r_list_length (msgs) - 1);
-
-	// Ensure tool_calls list exists
-	if (!msg->tool_calls) {
-		msg->tool_calls = r_list_new ();
-		if (!msg->tool_calls) {
-			return false;
-		}
-		msg->tool_calls->free = (RListFree)r2ai_tool_call_free;
-	}
-
-	// Copy the tool call
-	R2AI_ToolCall *dst_tc = R_NEW0 (R2AI_ToolCall);
-	dst_tc->name = tc->name? strdup (tc->name): NULL;
-	dst_tc->arguments = tc->arguments? strdup (tc->arguments): NULL;
-	dst_tc->id = tc->id? strdup (tc->id): NULL;
-
-	r_list_append (msg->tool_calls, dst_tc);
-	return true;
-}
-
-R_API bool r2ai_msgs_from_response(RList *msgs, const char *json_str) {
-	if (!msgs || !json_str) {
-		return false;
-	}
-
-	bool result = false;
-	// r_json_parse expects (and modifies) non-const char*, so we need to cast it
-	RJson *json = r_json_parse ((char *)json_str);
-	if (json) {
-		result = r2ai_msgs_from_json (msgs, json);
-		r_json_free (json);
-	}
-	return result;
-}
-
-R_API bool r2ai_msgs_from_json(RList *msgs, const RJson *json) {
-	if (!msgs || !json) {
-		return false;
-	}
-
-	const RJson *choices = r_json_get (json, "choices");
-	if (!choices || choices->type != R_JSON_ARRAY) {
-		return false;
-	}
-
-	const RJson *choice = r_json_item (choices, 0);
-	if (!choice) {
-		return false;
-	}
-
-	const RJson *message = r_json_get (choice, "message");
-	if (!message) {
-		return false;
-	}
-
-	const RJson *role = r_json_get (message, "role");
-	const RJson *content = r_json_get (message, "content");
-	const RJson *content_blocks = r_json_get (message, "content_blocks");
-
-	// Create a new message to add
-	R2AI_Message new_msg = { 0 };
-	new_msg.role = (role && role->type == R_JSON_STRING)? strdup (role->str_value): strdup ("assistant");
-	new_msg.content = (content && content->type == R_JSON_STRING)? strdup (content->str_value): NULL;
-	new_msg.tool_call_id = NULL;
-	new_msg.tool_calls = NULL;
-
-	if (content_blocks && content_blocks->type == R_JSON_ARRAY && content_blocks->children.count > 0) {
-		RList *cb = r2ai_content_blocks_new ();
-		if (!cb) {
-			r2ai_message_free (&new_msg);
-			return false;
-		}
-		for (size_t i = 0; i < content_blocks->children.count; i++) {
-			const RJson *block = r_json_item (content_blocks, i);
-			if (!block) {
-				continue;
-			}
-			R2AI_ContentBlock *dst = R_NEW0 (R2AI_ContentBlock);
-			if (!dst) {
-				r2ai_content_blocks_free (cb);
-				r2ai_message_free (&new_msg);
-				return false;
-			}
-			const RJson *type = r_json_get (block, "type");
-			const RJson *data = r_json_get (block, "data");
-			const RJson *thinking = r_json_get (block, "thinking");
-			const RJson *signature = r_json_get (block, "signature");
-			const RJson *text = r_json_get (block, "text");
-			const RJson *id = r_json_get (block, "id");
-			const RJson *name = r_json_get (block, "name");
-			const RJson *input = r_json_get (block, "input");
-
-			dst->type = (type && type->type == R_JSON_STRING)? strdup (type->str_value): NULL;
-			dst->data = (data && data->type == R_JSON_STRING)? strdup (data->str_value): NULL;
-			dst->thinking = (thinking && thinking->type == R_JSON_STRING)? strdup (thinking->str_value): NULL;
-			dst->signature = (signature && signature->type == R_JSON_STRING)? strdup (signature->str_value): NULL;
-			dst->text = (text && text->type == R_JSON_STRING)? strdup (text->str_value): NULL;
-			dst->id = (id && id->type == R_JSON_STRING)? strdup (id->str_value): NULL;
-			dst->name = (name && name->type == R_JSON_STRING)? strdup (name->str_value): NULL;
-			dst->input = (input && input->type == R_JSON_STRING)? strdup (input->str_value): NULL;
-			r_list_append (cb, dst);
-		}
-		new_msg.content_blocks = cb;
-	}
-
-	// Add the message without tool calls first
-	if (!r2ai_msgs_add (msgs, &new_msg)) {
-		r2ai_message_free (&new_msg);
-		return false;
-	}
-
-	// Now add tool calls if present
-	const RJson *tool_calls = r_json_get (message, "tool_calls");
-	if (tool_calls && tool_calls->type == R_JSON_ARRAY) {
-		// Iterate through array elements
-		for (size_t i = 0; i < tool_calls->children.count; i++) {
-			const RJson *tool_call = r_json_item (tool_calls, i);
-			if (!tool_call) {
-				continue;
-			}
-
-			const RJson *id = r_json_get (tool_call, "id");
-			const RJson *function = r_json_get (tool_call, "function");
-			if (!function) {
-				continue;
-			}
-
-			const RJson *name = r_json_get (function, "name");
-			const RJson *arguments = r_json_get (function, "arguments");
-
-			R2AI_ToolCall tc = { 0 };
-			tc.name = (name && name->type == R_JSON_STRING)? name->str_value: NULL;
-			tc.arguments = (arguments && arguments->type == R_JSON_STRING)? arguments->str_value: NULL;
-			tc.id = (id && id->type == R_JSON_STRING)? id->str_value: NULL;
-
-			if (!r2ai_msgs_add_tool_call (msgs, &tc)) {
-				r2ai_message_free (&new_msg);
-				return false;
-			}
-		}
-	}
-
-	return true;
-}
-
-R_API char *r2ai_msgs_to_json(const RList *msgs) {
+R_API char *r2ai_msgs_to_json(const RList *msgs, bool raw_tool_args) {
 	if (!msgs || r_list_empty (msgs)) {
 		return NULL;
 	}
@@ -369,8 +190,14 @@ R_API char *r2ai_msgs_to_json(const RList *msgs) {
 				// Add name
 				pj_ks (pj, "name", tc->name? tc->name: "");
 
-				// Add arguments (required by OpenAI API)
-				pj_ks (pj, "arguments", tc->arguments? tc->arguments: "{}");
+				// OpenAI wants arguments as a JSON-encoded string; Ollama wants a raw object.
+				const char *a = R_STR_ISEMPTY (tc->arguments)? "{}": tc->arguments;
+				pj_k (pj, "arguments");
+				if (raw_tool_args) {
+					pj_raw (pj, a);
+				} else {
+					pj_s (pj, a);
+				}
 
 				pj_end (pj); // End function object
 				pj_end (pj); // End tool call object

src/openai.c

@@ -1,393 +1,439 @@
-/* r2ai - Copyright 2023-2025 pancake, dnakov */
+/* r2ai - Copyright 2023-2026 pancake, dnakov */
 
 #include "r2ai.h"
 #include "r2ai_priv.h"
 
-R_IPI R2AI_ChatResponse *r2ai_openai(RCorePluginSession *cps, R2AIArgs args) {
+static const char *json_string(const RJson *json) {
+	return (json && json->type == R_JSON_STRING)? json->str_value: NULL;
+}
+
+static uint64_t json_integer(const RJson *json) {
+	return (json && json->type == R_JSON_INTEGER)? json->num.u_value: 0;
+}
+
+typedef enum {
+	R2AI_CHAT_OPENAI,
+	R2AI_CHAT_OLLAMA,
+	R2AI_GENERATE_OLLAMA
+} R2AIChatAPI;
+
+static bool is_generate_api(RCore *core) {
+	const char *apitype = r_config_get (core->config, "r2ai.apitype");
+	if (R_STR_ISEMPTY (apitype)) {
+		return false;
+	}
+	return !strcmp (apitype, "generate");
+}
+
+static char *ollama_generate_prompt_from_messages(const RList *msgs, char **system_prompt) {
+	if (system_prompt) {
+		*system_prompt = NULL;
+	}
+	if (!msgs || r_list_empty (msgs)) {
+		return NULL;
+	}
+
+	RStrBuf *sb = r_strbuf_new ("");
+	if (!sb) {
+		return NULL;
+	}
+
+	RListIter *iter;
+	const R2AI_Message *msg;
+	r_list_foreach (msgs, iter, msg) {
+		const char *role = msg->role? msg->role: "user";
+		const char *content = msg->content;
+		if (R_STR_ISEMPTY (content)) {
+			continue;
+		}
+		if (!strcmp (role, "system") || !strcmp (role, "developer")) {
+			if (system_prompt && !*system_prompt) {
+				*system_prompt = strdup (content);
+				continue;
+			}
+		}
+		r_strbuf_appendf (sb, "%s: %s\n\n", role, content);
+	}
+
+	char *prompt = r_strbuf_drain (sb);
+	if (R_STR_ISEMPTY (prompt) && system_prompt && *system_prompt) {
+		free (prompt);
+		prompt = strdup (*system_prompt);
+	}
+	return prompt;
+}
+
+static void cache_part(RStrBuf *sb, const char *s) {
+	const char *v = r_str_get (s);
+	r_strbuf_appendf (sb, "%zu:%s\n", strlen (v), v);
+}
+
+static char *cache_prefix(const char *provider, const char *model, const char *tools_json, const char *messages_json) {
+	RStrBuf *sb = r_strbuf_new ("");
+	cache_part (sb, provider);
+	cache_part (sb, model);
+	cache_part (sb, tools_json);
+	const char *messages = r_str_get (messages_json);
+	size_t len = strlen (messages);
+	if (len > 0 && messages[len - 1] == ']') {
+		len--;
+	}
+	r_strbuf_appendf (sb, "%zu:", len);
+	r_strbuf_append_n (sb, messages, len);
+	return r_strbuf_drain (sb);
+}
+
+static size_t prefix_len(const char *a, const char *b) {
+	size_t i = 0;
+	while (a[i] && b[i] && a[i] == b[i]) {
+		i++;
+	}
+	return i;
+}
+
+static void check_chat_cache(RCorePluginSession *cps, const char *provider, const char *model, const char *tools_json, const char *messages_json) {
 	RCore *core = cps->core;
-	args.provider = r_config_get (core->config, "r2ai.api");
-	args.model = r_config_get (core->config, "r2ai.model");
-
-	const R2AIProvider *provider_info = r2ai_get_provider (args.provider);
-	const char *base_url = r2ai_get_provider_url (core, args.provider);
-	// TODO: default model name should depend on api
-	const char *model_name = args.model? args.model: "gpt-4o-mini";
-	char **error = args.error;
-	RList *tools = args.tools;
-	// create a temp conversation to include the system prompt and the rest of the messages
-	RList *temp_msgs = r2ai_msgs_new ();
-	if (!temp_msgs) {
-		if (error) {
-			*error = strdup ("Failed to create temporary messages array");
-		}
-		return NULL;
+	R2AI_State *state = cps->data;
+	if (!state || !r_config_get_b (core->config, "r2ai.cacheck")) {
+		return;
 	}
-	R2AI_Message system_msg = {
-		.role = "system",
-		.content = (char *)args.system_prompt
-	};
-	// Add system message if available from args.system_prompt
-	if (R_STR_ISNOTEMPTY (args.system_prompt)) {
-		R_LOG_DEBUG ("Using system prompt: %s", args.system_prompt);
-		// if the model name contains "o1" or "o3", it's "developer" role
-		if (strstr (model_name, "o1") || strstr (model_name, "o3")) {
-			system_msg.role = "developer";
-			system_msg.content = (char *)args.system_prompt;
+	char *prefix = cache_prefix (provider, model, tools_json, messages_json);
+	if (state->cache_prefix) {
+		const size_t old_len = strlen (state->cache_prefix);
+		if (!r_str_startswith (prefix, state->cache_prefix)) {
+			const size_t kept = prefix_len (state->cache_prefix, prefix);
+			R_LOG_WARN ("Chat cache prefix changed: preserved %zu/%zu bytes from previous request. Keep system prompt, tool catalog, and previous messages append-only to maximize provider cache hits.", kept, old_len);
 		} else {
-			system_msg.role = "system";
-			system_msg.content = (char *)args.system_prompt;
-		}
-		r2ai_msgs_add (temp_msgs, &system_msg);
-	} else {
-		// Fallback to config if args.system_prompt is not set
-		const char *sysprompt = r_config_get (core->config, "r2ai.system");
-		if (R_STR_ISNOTEMPTY (sysprompt)) {
-			R_LOG_DEBUG ("Using system prompt from config: %s", sysprompt);
-			if (strstr (model_name, "o1") || strstr (model_name, "o3")) {
-				system_msg.role = "developer";
-			} else {
-				system_msg.role = "system";
-			}
-			system_msg.content = (char *)sysprompt;
-			r2ai_msgs_add (temp_msgs, &system_msg);
+			R_LOG_DEBUG ("Chat cache prefix preserved: %zu bytes", old_len);
 		}
+		R_FREE (state->cache_prefix);
 	}
-	if (args.messages) {
-		RListIter *iter;
-		R2AI_Message *msg;
-		r_list_foreach (args.messages, iter, msg) {
-			r2ai_msgs_add (temp_msgs, msg);
-		}
-	} else {
-		R_LOG_WARN ("No messages");
-	}
-	// Safely print debug info about first message
-	if (temp_msgs && !r_list_empty (temp_msgs) && ((R2AI_Message *)r_list_get_n (temp_msgs, 0))->role) {
-		R_LOG_DEBUG ("First message role: %s", ((R2AI_Message *)r_list_get_n (temp_msgs, 0))->role);
-	}
-	if (error) {
-		*error = NULL;
-	}
+	state->cache_prefix = prefix;
+}
 
-	const char **headers = NULL;
-	char *auth_header = NULL;
-	if (R_STR_ISNOTEMPTY (args.api_key)) {
-		auth_header = r_str_newf ("Authorization: Bearer %s", args.api_key);
-		R_LOG_DEBUG ("Auth header: %s", auth_header);
-		static const char *static_headers[] = { NULL, NULL, NULL };
-		headers = static_headers;
-		headers[0] = "Content-Type: application/json";
-		headers[1] = auth_header;
-	}
-	const char *urlfmt = (provider_info && provider_info->api_type == R2AI_API_OLLAMA)
-		? "%s/chat"
-		: "%s/chat/completions";
-	char *openai_url = r_str_newf (urlfmt, base_url);
-
-	// Create a messages JSON object, either from input messages or from content
+static char *chat_request_json(RCorePluginSession *cps, const R2AIArgs *args, const char *provider, const char *model, const RList *messages, R2AIChatAPI api, bool check_cache) {
+	const bool use_generate = api == R2AI_GENERATE_OLLAMA;
+	const bool use_ollama = api != R2AI_CHAT_OPENAI;
+	const R2AIProvider *provider_info = r2ai_get_provider (provider);
+	const bool is_ollama_provider = provider_info && provider_info->api_type == R2AI_API_OLLAMA;
 	char *messages_json = NULL;
-
-	if (temp_msgs && !r_list_empty (temp_msgs)) {
-		R_LOG_DEBUG ("Using input messages: %d messages", r_list_length (temp_msgs));
-		messages_json = r2ai_msgs_to_json (temp_msgs);
-		if (!messages_json) {
-			if (error) {
-				*error = strdup ("Failed to convert messages to JSON");
-			}
-			free (auth_header);
-			return NULL;
-		}
+	char *generate_prompt = NULL;
+	char *generate_system = NULL;
+	if (use_generate) {
+		generate_prompt = ollama_generate_prompt_from_messages (messages, &generate_system);
 	} else {
-		if (error) {
-			*error = strdup ("No messages provided");
-		}
-		free (auth_header);
+		messages_json = r2ai_msgs_to_json (messages, use_ollama);
+	}
+	if (!messages_json && !generate_prompt) {
 		return NULL;
 	}
 
-	// Convert tools to OpenAI format if available
-	char *openai_tools_json = NULL;
-	if (tools && !r_list_empty (tools)) {
-		openai_tools_json = r2ai_tools_to_openai_json (tools);
-	}
-
-	// Create the model settings part
-	PJ *pj = pj_new ();
-	pj_o (pj);
-	pj_ks (pj, "model", model_name);
-	pj_kb (pj, "stream", false);
-
-	if (provider_info && provider_info->api_type == R2AI_API_OLLAMA) {
-		// Ollama uses "options" object for parameters
-		pj_ko (pj, "options");
-		if (args.max_tokens) {
-			pj_kn (pj, "num_predict", args.max_tokens);
-		}
-		if (args.temperature > 0) {
-			pj_kd (pj, "temperature", args.temperature);
-		}
-		pj_end (pj); // end options
-	} else {
-#if 0
-		// gpt-5-mini-chat is the only gpt-5 model that supports temperature
-		// gpt-5 gpt-5-mini and gpt-5-nano just throw an error
-		// Only add temperature if this provider/model doesn't have the temperature error flag
-		if (!model_has_error (args.provider, model_name, MODEL_ERROR_TEMPERATURE)) {
-			pj_kd (pj, "temperature", args.temperature? args.temperature: 0.01);
-		}
-#endif
-
-		if (strcmp (args.provider, "mistral") == 0) {
-			pj_kn (pj, "max_tokens", args.max_tokens? args.max_tokens: 5128);
+	char *tools_json = NULL;
+	if (args->tools && !r_list_empty (args->tools)) {
+		if (use_generate) {
+			R_LOG_DEBUG ("Skipping native tool definitions for /api/generate payload");
 		} else {
-			pj_kn (pj, "max_completion_tokens", args.max_tokens? args.max_tokens: 5128);
+			tools_json = r2ai_tools_to_openai_json (args->tools);
 		}
 	}
 
+	PJ *pj = pj_new ();
+	if (!pj) {
+		free (messages_json);
+		free (generate_prompt);
+		free (generate_system);
+		free (tools_json);
+		return NULL;
+	}
+	pj_o (pj);
+	pj_ks (pj, "model", model);
+	pj_kb (pj, "stream", false);
+	if (use_ollama) {
+		pj_kb (pj, "think", args->thinking_tokens > 0);
+		pj_ko (pj, "options");
+		if (args->max_tokens) {
+			pj_kn (pj, "num_predict", args->max_tokens);
+		}
+		if (args->deterministic) {
+			pj_kn (pj, "temperature", 0);
+			pj_kn (pj, "top_p", 1);
+			pj_kn (pj, "top_k", 1);
+			pj_kn (pj, "repeat_last_n", 0);
+			pj_kn (pj, "repeat_penalty", 1);
+			pj_kn (pj, "seed", 123);
+		} else if (args->temperature > 0) {
+			pj_kd (pj, "temperature", args->temperature);
+		}
+		pj_end (pj);
+	} else if (is_ollama_provider || !strcmp (provider, "mistral")) {
+		pj_kn (pj, "max_tokens", args->max_tokens? args->max_tokens: 5128);
+	} else {
+		pj_kn (pj, "max_completion_tokens", args->max_tokens? args->max_tokens: 5128);
+	}
+	if (use_generate) {
+		pj_ks (pj, "prompt", generate_prompt);
+		if (R_STR_ISNOTEMPTY (generate_system)) {
+			pj_ks (pj, "system", generate_system);
+		}
+	} else {
+		pj_k (pj, "messages");
+		pj_raw (pj, messages_json);
+		if (tools_json) {
+			pj_k (pj, "tools");
+			pj_raw (pj, tools_json);
+		}
+	}
 	pj_end (pj);
 
-	// Get the JSON for model settings
-	char *model_json = pj_drain (pj);
-
-	// Manually create the final JSON by combining parts
-	// Remove the closing brace from model_json
-	size_t model_len = strlen (model_json);
-	if (model_len > 0 && model_json[model_len - 1] == '}') {
-		model_json[model_len - 1] = '\0';
+	if (!use_generate && check_cache) {
+		check_chat_cache (cps, provider, model, tools_json, messages_json);
 	}
-
-	// Create the full JSON with proper structure
-	char *complete_json;
-	if (openai_tools_json) {
-		complete_json = r_str_newf ("%s, \"messages\": %s, \"tools\": %s}",
-			model_json, messages_json, openai_tools_json);
-		free (openai_tools_json);
-	} else {
-		complete_json = r_str_newf ("%s, \"messages\": %s}",
-			model_json, messages_json);
-	}
-
-	// Free intermediate strings
-	free (model_json);
 	free (messages_json);
+	free (generate_prompt);
+	free (generate_system);
+	free (tools_json);
+	return pj_drain (pj);
+}
 
-	if (!complete_json) {
-		if (error) {
-			*error = strdup ("Failed to create complete request JSON");
+static char *chat_api_url(const char *base_url, R2AIChatAPI api) {
+	const char *path = api == R2AI_CHAT_OPENAI
+		? "/chat/completions"
+		: (api == R2AI_GENERATE_OLLAMA? "/generate": "/chat");
+	return r_str_newf ("%s%s", base_url, path);
+}
+
+static char *ollama_native_base(const char *base_url) {
+	if (r_str_endswith (base_url, "/api")) {
+		return strdup (base_url);
+	}
+	if (r_str_endswith (base_url, "/v1")) {
+		char *url = strdup (base_url);
+		if (!url) {
+			return NULL;
 		}
-		free (auth_header);
+		url[strlen (url) - 3] = 0;
+		char *res = r_str_newf ("%s/api", url);
+		free (url);
+		return res;
+	}
+	return r_str_newf ("%s/api", base_url);
+}
+
+static bool allow_ollama_fallback(RCore *core, const char *provider) {
+	if (strcmp (provider, "ollama")) {
+		return false;
+	}
+	const char *base_url = r_config_get (core->config, "r2ai.baseurl");
+	if (R_STR_ISEMPTY (base_url)) {
+		return true;
+	}
+	char *url = strdup (base_url);
+	if (!url) {
+		return false;
+	}
+	r_str_trim (url);
+	size_t len = strlen (url);
+	while (len > 0 && url[len - 1] == '/') {
+		url[--len] = 0;
+	}
+	const bool explicit_root = r_str_endswith (url, "/v1") || r_str_endswith (url, "/api");
+	free (url);
+	return !explicit_root;
+}
+
+R_IPI R2AI_ChatResponse *r2ai_openai(RCorePluginSession *cps, R2AIArgs args) {
+	RCore *core = cps->core;
+	const char *provider_name = args.provider;
+	const char *model_name = R_STR_ISNOTEMPTY (args.model)? args.model: "gpt-4o-mini";
+
+	const R2AIProvider *provider_info = r2ai_get_provider (provider_name);
+	const bool is_ollama_provider = provider_info && provider_info->api_type == R2AI_API_OLLAMA;
+	const bool use_generate = is_ollama_provider && is_generate_api (core);
+	char *base_url = r2ai_get_provider_url (core, provider_name);
+	if (!base_url) {
+		*args.error = strdup ("Failed to resolve provider URL");
 		return NULL;
 	}
-
-	// Save the full JSON to a file for inspection
-	// XXX: only create request/response files when r2ai.debug is set
-	char *tmpdir = r_file_tmpdir ();
-	char *req_path = r_str_newf ("%s" R_SYS_DIR "r2ai_openai_request.json", tmpdir);
-	r_file_dump (req_path, (const ut8 *)complete_json, strlen (complete_json), 0);
-	R_LOG_DEBUG ("Full request saved to %s", req_path);
-	free (req_path);
-	free (tmpdir);
-
-	R_LOG_DEBUG ("OpenAI API request data: %s", complete_json);
-
-	if (r_config_get_b (core->config, "r2ai.debug")) {
-		// Generate curl command for debugging
-		RStrBuf *curl_cmd = r_strbuf_new ("curl -X POST");
-		for (int i = 0; headers[i]; i++) {
-			r_strbuf_appendf (curl_cmd, " -H \"%s\"", headers[i]);
-		}
-		r_strbuf_appendf (curl_cmd, " -d '%s' \"%s\"", complete_json, openai_url);
-		eprintf ("Curl command: %s\n", r_strbuf_get (curl_cmd));
-		r_strbuf_free (curl_cmd);
+	R2AIChatAPI chat_api = use_generate
+		? R2AI_GENERATE_OLLAMA
+		: (is_ollama_provider && r_str_endswith (base_url, "/api")? R2AI_CHAT_OLLAMA: R2AI_CHAT_OPENAI);
+	char **error = args.error;
+	// the system prompt goes first, o1 and o3 models call it "developer"
+	RList *temp_msgs = r2ai_msgs_new ();
+	if (R_STR_ISNOTEMPTY (args.system_prompt)) {
+		const bool dev = strstr (model_name, "o1") || strstr (model_name, "o3");
+		R2AI_Message system_msg = { .role = dev? "developer": "system", .content = (char *)args.system_prompt };
+		r2ai_msgs_add (temp_msgs, &system_msg);
+	}
+	RListIter *iter;
+	R2AI_Message *msg;
+	r_list_foreach (args.messages, iter, msg) {
+		r2ai_msgs_add (temp_msgs, msg);
+	}
+	if (r_list_empty (args.messages)) {
+		*error = strdup ("No messages provided");
+		free (base_url);
+		r2ai_msgs_free (temp_msgs);
+		return NULL;
+	}
+	char *auth_header = NULL;
+	const char *headers[] = { "Content-Type: application/json", NULL, NULL };
+	if (R_STR_ISNOTEMPTY (args.api_key)) {
+		auth_header = r_str_newf ("Authorization: Bearer %s", args.api_key);
+		headers[1] = auth_header;
 	}
 
-	// Make the API call
 	char *res = NULL;
 	int code = 0;
-	res = r2ai_http_post (core, openai_url, headers, complete_json, &code, NULL);
-	free (complete_json);
-	free (openai_url);
-
-	if (code != 200) {
-		R_LOG_ERROR ("OpenAI API error %d", code);
-		if (res) {
-			R_LOG_ERROR ("OpenAI API error response: %s", res);
+	char *fallback_base = NULL;
+	const char *request_base = base_url;
+	bool first_request = true;
+	for (;;) {
+		char *request_json = chat_request_json (cps, &args, provider_name, model_name, temp_msgs, chat_api, first_request);
+		if (!request_json) {
+			*error = strdup ("Failed to create request JSON");
+			free (fallback_base);
+			free (auth_header);
+			free (base_url);
+			r2ai_msgs_free (temp_msgs);
+			return NULL;
 		}
-		free (auth_header);
-		free (res);
+		char *api_url = chat_api_url (request_base, chat_api);
+		res = r2ai_post (core, "openai", api_url, headers, request_json, &code, error);
+		free (request_json);
+		free (api_url);
+		if (chat_api != R2AI_CHAT_OPENAI || (code != 404 && code != 405) || !allow_ollama_fallback (core, provider_name)) {
+			break;
+		}
+		fallback_base = ollama_native_base (base_url);
+		if (!fallback_base) {
+			break;
+		}
+		R_LOG_INFO ("OpenAI-compatible Ollama endpoint unavailable; falling back to /api/chat");
+		R_FREE (*error);
+		request_base = fallback_base;
+		chat_api = R2AI_CHAT_OLLAMA;
+		first_request = false;
+	}
+	free (fallback_base);
+	free (base_url);
+
+	free (auth_header);
+	if (!res) {
+		r2ai_msgs_free (temp_msgs);
 		return NULL;
 	}
-
-	// Save the response for inspection
-	tmpdir = r_file_tmpdir ();
-	char *res_path = r_str_newf ("%s" R_SYS_DIR "r2ai_openai_response.json", tmpdir);
-	r_file_dump (res_path, (const ut8 *)res, strlen (res), 0);
 	if (r_config_get_b (core->config, "r2ai.debug")) {
 		eprintf ("OpenAI API response: %s\n", res);
 	}
-	free (res_path);
-	free (tmpdir);
-
-	// Parse the response into our messages structure
-
-	char *res_copy = strdup (res);
-	RJson *jres = r_json_parse (res_copy);
+	RJson *jres = r_json_parse (res);
 	if (jres) {
-		// Create a new message structure
+		const bool is_native_ollama = chat_api != R2AI_CHAT_OPENAI;
 		R2AI_Message *message = R_NEW0 (R2AI_Message);
 		R2AI_Usage *usage = R_NEW0 (R2AI_Usage);
-		if (message) {
-			// Process the response using our r2ai_msgs_from_json logic
-			const RJson *usage_json = NULL;
-			if (provider_info && provider_info->api_type == R2AI_API_OLLAMA) {
-				// Ollama has usage info at top level
-				usage_json = jres;
-			} else {
-				usage_json = r_json_get (jres, "usage");
+		const RJson *usage_json = is_native_ollama? jres: r_json_get (jres, "usage");
+		if (usage_json && usage_json->type == R_JSON_OBJECT) {
+			usage->prompt_tokens = json_integer (r_json_get (usage_json, is_native_ollama? "prompt_eval_count": "prompt_tokens"));
+			usage->completion_tokens = json_integer (r_json_get (usage_json, is_native_ollama? "eval_count": "completion_tokens"));
+			usage->total_tokens = is_native_ollama
+				? usage->prompt_tokens + usage->completion_tokens
+				: json_integer (r_json_get (usage_json, "total_tokens"));
+		}
+
+		const RJson *message_json = NULL;
+		if (use_generate) {
+			const char *content = json_string (r_json_get (jres, "response"));
+			const char *thinking = json_string (r_json_get (jres, "thinking"));
+			message->role = strdup ("assistant");
+			if (content) {
+				message->content = strdup (content);
+			}
+			if (R_STR_ISNOTEMPTY (thinking)) {
+				message->reasoning_content = strdup (thinking);
+			}
+		} else if (is_native_ollama) {
+			message_json = r_json_get (jres, "message");
+		} else {
+			const RJson *choices = r_json_get (jres, "choices");
+			if (choices && choices->type == R_JSON_ARRAY) {
+				const RJson *choice = r_json_item (choices, 0);
+				message_json = choice? r_json_get (choice, "message"): NULL;
+			}
+		}
+
+		if (message_json) {
+			const char *role = json_string (r_json_get (message_json, "role"));
+			const char *content = json_string (r_json_get (message_json, "content"));
+			const char *reasoning_content = json_string (r_json_get (message_json, "reasoning_content"));
+			const char *thinking = json_string (r_json_get (message_json, "thinking"));
+			const char *tool_call_id = json_string (r_json_get (message_json, "tool_call_id"));
+			const RJson *tool_calls = r_json_get (message_json, "tool_calls");
+
+			message->role = strdup (role? role: "assistant");
+			if (content) {
+				message->content = strdup (content);
+			}
+			if (R_STR_ISNOTEMPTY (reasoning_content)) {
+				message->reasoning_content = strdup (reasoning_content);
+			} else if (R_STR_ISNOTEMPTY (thinking)) {
+				message->reasoning_content = strdup (thinking);
+			}
+			if (R_STR_ISNOTEMPTY (tool_call_id)) {
+				message->tool_call_id = strdup (tool_call_id);
 			}
 
-			if (usage_json && usage_json->type == R_JSON_OBJECT) {
-				if (provider_info && provider_info->api_type == R2AI_API_OLLAMA) {
-					// Ollama field names
-					const RJson *prompt_tokens = r_json_get (usage_json, "prompt_eval_count");
-					const RJson *completion_tokens = r_json_get (usage_json, "eval_count");
-					if (prompt_tokens && prompt_tokens->type == R_JSON_INTEGER) {
-						usage->prompt_tokens = prompt_tokens->num.u_value;
-					}
-					if (completion_tokens && completion_tokens->type == R_JSON_INTEGER) {
-						usage->completion_tokens = completion_tokens->num.u_value;
-					}
-					// Calculate total_tokens for ollama
-					usage->total_tokens = usage->prompt_tokens + usage->completion_tokens;
-				} else {
-					// OpenAI field names
-					const RJson *prompt_tokens = r_json_get (usage_json, "prompt_tokens");
-					const RJson *completion_tokens = r_json_get (usage_json, "completion_tokens");
-					const RJson *total_tokens = r_json_get (usage_json, "total_tokens");
-					if (prompt_tokens && prompt_tokens->type == R_JSON_INTEGER) {
-						usage->prompt_tokens = prompt_tokens->num.u_value;
-					}
-					if (completion_tokens && completion_tokens->type == R_JSON_INTEGER) {
-						usage->completion_tokens = completion_tokens->num.u_value;
-					}
-					if (total_tokens && total_tokens->type == R_JSON_INTEGER) {
-						usage->total_tokens = total_tokens->num.u_value;
-					}
-				}
-			}
-			const RJson *message_json = NULL;
-			if (provider_info && provider_info->api_type == R2AI_API_OLLAMA) {
-				// Ollama returns message directly
-				message_json = r_json_get (jres, "message");
-			} else {
-				// OpenAI-style response with choices array
-				const RJson *choices = r_json_get (jres, "choices");
-				if (choices && choices->type == R_JSON_ARRAY) {
-					const RJson *choice = r_json_item (choices, 0);
-					if (choice) {
-						message_json = r_json_get (choice, "message");
-					}
-				}
-			}
-
-			if (message_json) {
-				const RJson *role = r_json_get (message_json, "role");
-				const RJson *content = r_json_get (message_json, "content");
-				const RJson *reasoning_content = r_json_get (message_json, "reasoning_content");
-				const RJson *thinking = r_json_get (message_json, "thinking");
-				const RJson *tool_call_id = r_json_get (message_json, "tool_call_id");
-				const RJson *tool_calls = r_json_get (message_json, "tool_calls");
-
-				// Set the basic message properties
-				message->role = (role && role->type == R_JSON_STRING)? strdup (role->str_value): strdup ("assistant");
-
-				if (content && content->type == R_JSON_STRING) {
-					message->content = strdup (content->str_value);
-				}
-
-				if (reasoning_content && reasoning_content->type == R_JSON_STRING && R_STR_ISNOTEMPTY (reasoning_content->str_value)) {
-					message->reasoning_content = strdup (reasoning_content->str_value);
-				} else if (thinking && thinking->type == R_JSON_STRING && R_STR_ISNOTEMPTY (thinking->str_value)) {
-					message->reasoning_content = strdup (thinking->str_value);
-				}
-
-				if (tool_call_id && tool_call_id->type == R_JSON_STRING && R_STR_ISNOTEMPTY (tool_call_id->str_value)) {
-					message->tool_call_id = strdup (tool_call_id->str_value);
-				}
-
-				if (tool_calls && tool_calls->type == R_JSON_ARRAY && tool_calls->children.count > 0) {
-					message->tool_calls = r_list_new ();
-					if (message->tool_calls) {
-						message->tool_calls->free = (RListFree)r2ai_tool_call_free;
-						for (size_t i = 0; i < tool_calls->children.count; i++) {
-							const RJson *tool_call = r_json_item (tool_calls, i);
-							if (!tool_call || tool_call->type != R_JSON_OBJECT) {
-								continue;
-							}
-							R2AI_ToolCall *tc = R_NEW0 (R2AI_ToolCall);
-							if (!tc) {
-								continue;
-							}
-							const RJson *tc_id = r_json_get (tool_call, "id");
-							if (tc_id && tc_id->type == R_JSON_STRING && R_STR_ISNOTEMPTY (tc_id->str_value)) {
-								tc->id = strdup (tc_id->str_value);
-							}
-							const RJson *function = r_json_get (tool_call, "function");
-							const RJson *name_json = NULL;
-							const RJson *arguments = NULL;
-							if (function && function->type == R_JSON_OBJECT) {
-								name_json = r_json_get (function, "name");
-								arguments = r_json_get (function, "arguments");
-							} else {
-								name_json = r_json_get (tool_call, "name");
-								arguments = r_json_get (tool_call, "arguments");
-							}
-							if (name_json && name_json->type == R_JSON_STRING && R_STR_ISNOTEMPTY (name_json->str_value)) {
-								tc->name = strdup (name_json->str_value);
-							}
-							if (arguments) {
-								if (arguments->type == R_JSON_STRING && R_STR_ISNOTEMPTY (arguments->str_value)) {
-									tc->arguments = strdup (arguments->str_value);
-								} else {
-									char *arguments_json = r_json_to_string (arguments);
-									if (arguments_json) {
-										tc->arguments = arguments_json;
-									}
-								}
-							}
-							if (!tc->name && !tc->arguments && !tc->id) {
-								r2ai_tool_call_free (tc);
-								continue;
-							}
-							r_list_append (message->tool_calls, tc);
+			if (tool_calls && tool_calls->type == R_JSON_ARRAY && tool_calls->children.count > 0) {
+				message->tool_calls = r_list_newf ((RListFree)r2ai_tool_call_free);
+				if (message->tool_calls) {
+					size_t i;
+					for (i = 0; i < tool_calls->children.count; i++) {
+						const RJson *tool_call = r_json_item (tool_calls, i);
+						if (!tool_call || tool_call->type != R_JSON_OBJECT) {
+							continue;
 						}
-						if (r_list_empty (message->tool_calls)) {
-							r_list_free (message->tool_calls);
-							message->tool_calls = NULL;
+						const RJson *function = r_json_get (tool_call, "function");
+						const RJson *source = (function && function->type == R_JSON_OBJECT)? function: tool_call;
+						const char *name = json_string (r_json_get (source, "name"));
+						const RJson *arguments = r_json_get (source, "arguments");
+						if (R_STR_ISEMPTY (name) || !arguments) {
+							continue;
 						}
+						R2AI_ToolCall *tc = R_NEW0 (R2AI_ToolCall);
+						const char *tc_id = json_string (r_json_get (tool_call, "id"));
+						if (R_STR_ISNOTEMPTY (tc_id)) {
+							tc->id = strdup (tc_id);
+						}
+						if (!tc->id) {
+							tc->id = r_str_newf ("call_%zu", i);
+						}
+						tc->name = strdup (name);
+						if (arguments->type == R_JSON_STRING && R_STR_ISNOTEMPTY (arguments->str_value)) {
+							tc->arguments = strdup (arguments->str_value);
+						} else {
+							tc->arguments = r_json_to_string (arguments);
+						}
+						if (!tc->name || !tc->arguments) {
+							r2ai_tool_call_free (tc);
+							continue;
+						}
+						r_list_append (message->tool_calls, tc);
+					}
+					if (r_list_empty (message->tool_calls)) {
+						r_list_free (message->tool_calls);
+						message->tool_calls = NULL;
 					}
 				}
-				// TODO: Handle tool calls if present?
 			}
 		}
 		r_json_free (jres);
 		R2AI_ChatResponse *result = R_NEW0 (R2AI_ChatResponse);
 		result->message = message;
 		result->usage = usage;
-		free (res_copy);
-		free (auth_header);
 		r2ai_msgs_free (temp_msgs);
 		free (res);
 		return result;
 	}
-
-	free (auth_header);
 	free (res);
-	return NULL;
-}
-
-R_IPI char *r2ai_openai_stream(RCore *core, R2AIArgs args) {
-	(void)core;
-	(void)args;
+	r2ai_msgs_free (temp_msgs);
 	return NULL;
 }

src/query.c

@@ -1,4 +1,4 @@
-/* r2ai - Copyright 2023-2025 pancake */
+/* r2ai - Copyright 2023-2026 pancake */
 
 #include "r2ai.h"
 
@@ -19,6 +19,39 @@ R_API void r2aiprompt_free(R2AIPrompt *prompt) {
 	free (prompt);
 }
 
+typedef struct {
+	const char *key;
+	size_t off;
+} PromptField;
+
+#define PROMPT_FIELD(k, f) \
+	{ k, r_offsetof (R2AIPrompt, f) }
+static const PromptField prompt_fields[] = {
+	PROMPT_FIELD ("title", title),
+	PROMPT_FIELD ("author", author),
+	PROMPT_FIELD ("description", desc),
+	PROMPT_FIELD ("command", command),
+	PROMPT_FIELD ("requires", requires),
+	PROMPT_FIELD ("if-empty", if_empty),
+	PROMPT_FIELD ("if-command", if_command),
+	PROMPT_FIELD ("model", model),
+	PROMPT_FIELD ("provider", provider),
+};
+#undef PROMPT_FIELD
+
+static bool prompt_set_field(R2AIPrompt *prompt, const char *key, const char *value) {
+	size_t i;
+	for (i = 0; i < R_ARRAY_SIZE (prompt_fields); i++) {
+		if (!strcmp (key, prompt_fields[i].key)) {
+			char **field = (char **) ((ut8 *)prompt + prompt_fields[i].off);
+			free (*field);
+			*field = strdup (value);
+			return true;
+		}
+	}
+	return false;
+}
+
 R_API R2AIPrompt *parse_prompt_file(const char *filepath) {
 	R2AIPrompt *prompt = R_NEW0 (R2AIPrompt);
 	char *content = r_file_slurp (filepath, NULL);
@@ -27,53 +60,43 @@ R_API R2AIPrompt *parse_prompt_file(const char *filepath) {
 		return NULL;
 	}
 
-	// Check if it's a markdown file with YAML frontmatter
-	if (r_str_startswith (content, "---\n")) {
-		// Parse YAML frontmatter
-		char *frontmatter_end = strstr (content + 4, "\n---\n");
-		if (frontmatter_end) {
-			char *frontmatter = r_str_ndup (content + 4, frontmatter_end - (content + 4));
-			RList *lines = r_str_split_list (frontmatter, "\n", -1);
-			RListIter *iter;
-			char *line;
-			r_list_foreach (lines, iter, line) {
-				r_str_trim (line);
-				if (*line == 0) {
-					continue;
-				}
-				char *colon = strchr (line, ':');
-				if (!colon) {
-					continue;
-				}
-				*colon = 0;
-				char *key = line;
-				char *value = colon + 1;
-				r_str_trim (key);
-				r_str_trim (value);
-				if (!strcmp (key, "title")) {
-					prompt->desc = strdup (value);
-				} else if (!strcmp (key, "author")) {
-					prompt->author = strdup (value);
-				} else if (!strcmp (key, "description")) {
-					prompt->title = strdup (value);
-				} else if (!strcmp (key, "command")) {
-					prompt->command = strdup (value);
-				} else if (!strcmp (key, "model")) {
-					prompt->model = strdup (value);
-				} else if (!strcmp (key, "provider")) {
-					prompt->provider = strdup (value);
-				}
-			}
-			r_list_free (lines);
-			free (frontmatter);
-			// The prompt content starts after the second ---
-			char *prompt_start = frontmatter_end + 5; // Skip "\n---\n"
-			prompt->prompt = strdup (prompt_start);
-		}
-	} else {
+	if (!r_str_startswith (content, "---\n")) {
 		R_LOG_ERROR ("Invalid prompt file format: %s", filepath);
+		r2aiprompt_free (prompt);
+		free (content);
+		return NULL;
 	}
 
+	char *frontmatter_end = strstr (content + 4, "\n---\n");
+	if (!frontmatter_end) {
+		R_LOG_ERROR ("Invalid prompt file format: %s", filepath);
+		r2aiprompt_free (prompt);
+		free (content);
+		return NULL;
+	}
+
+	char *prompt_start = frontmatter_end + 5;
+	*frontmatter_end = 0;
+	RList *lines = r_str_split_list (content + 4, "\n", -1);
+	RListIter *iter;
+	char *line;
+	r_list_foreach (lines, iter, line) {
+		r_str_trim (line);
+		if (!*line || *line == '#') {
+			continue;
+		}
+		char *value = strchr (line, ':');
+		if (!value) {
+			continue;
+		}
+		*value++ = 0;
+		r_str_trim (line);
+		r_str_trim (value);
+		prompt_set_field (prompt, line, value);
+	}
+	r_list_free (lines);
+	prompt->prompt = strdup (prompt_start);
+
 	free (content);
 	return prompt;
 }
@@ -120,7 +143,6 @@ static char *replace_vars(RCore *core, const char *text) {
 				p = end + 1;
 			} else {
 				r_strbuf_append (sb, "${");
-				p += 2;
 			}
 		} else if (*p == '$' && *(p + 1) == '(') {
 			p += 2;
@@ -135,7 +157,6 @@ static char *replace_vars(RCore *core, const char *text) {
 				p = end + 1;
 			} else {
 				r_strbuf_append (sb, "$(");
-				p += 2;
 			}
 		} else {
 			r_strbuf_append_n (sb, p, 1);
@@ -145,31 +166,33 @@ static char *replace_vars(RCore *core, const char *text) {
 	return r_strbuf_drain (sb);
 }
 
-// Helper function to find prompt file in search directories
 R_API char *find_prompt_file(RList *search_dirs, const char *name) {
 	char *filepath = NULL;
 	RListIter *dir_iter;
 	char *dir;
 	r_list_foreach (search_dirs, dir_iter, dir) {
-		char *filepath_txt = r_str_newf ("%s/%s.r2ai.txt", dir, name);
-		char *filepath_md = r_str_newf ("%s/%s.r2ai.md", dir, name);
-		if (r_file_exists (filepath_txt)) {
-			filepath = filepath_txt;
-			free (filepath_md);
-			break;
-		} else if (r_file_exists (filepath_md)) {
-			filepath = filepath_md;
-			free (filepath_txt);
+		filepath = r_str_newf ("%s/%s.r2ai.md", dir, name);
+		if (r_file_exists (filepath)) {
 			break;
 		}
-		free (filepath_txt);
-		free (filepath_md);
+		free (filepath);
+		filepath = NULL;
 	}
 	return filepath;
 }
 
-R_API void r2ai_cmd_q(RCorePluginSession *cps, const char *input) {
-	RCore *core = cps->core;
+static bool prompt_name_seen(RList *seen, const char *name) {
+	RListIter *iter;
+	char *it;
+	r_list_foreach (seen, iter, it) {
+		if (!strcmp (it, name)) {
+			return true;
+		}
+	}
+	return false;
+}
+
+static RList *prompt_search_dirs(RCore *core) {
 	const char *promptdir = r_config_get (core->config, "r2ai.promptdir");
 	if (!promptdir || !*promptdir) {
 		promptdir = "~/.config/r2ai/prompts";
@@ -181,55 +204,135 @@ R_API void r2ai_cmd_q(RCorePluginSession *cps, const char *input) {
 		r_list_append (search_dirs, expanded_dir);
 	} else {
 		free (expanded_dir);
-		expanded_dir = NULL;
 	}
 	if (r_file_is_directory (local_prompts)) {
 		r_list_append (search_dirs, local_prompts);
 	} else {
 		free (local_prompts);
-		local_prompts = NULL;
 	}
-	if (r_list_empty (search_dirs)) {
-		R_LOG_ERROR ("No prompt directories found");
-		r_list_free (search_dirs);
+	return search_dirs;
+}
+
+static RList *prompt_search_dirs_or_warn(RCore *core) {
+	RList *search_dirs = prompt_search_dirs (core);
+	if (!r_list_empty (search_dirs)) {
+		return search_dirs;
+	}
+	R_LOG_ERROR ("No prompt directories found");
+	r_list_free (search_dirs);
+	return NULL;
+}
+
+static void prompt_print_summary(RCore *core, const char *name, R2AIPrompt *prompt) {
+	const char *title = r_str_get (prompt->title);
+	const char *desc = r_str_get (prompt->desc);
+	if (*title && *desc) {
+		r_cons_printf (core->cons, "%s: %s - %s\n", name, title, desc);
+	} else if (*title) {
+		r_cons_printf (core->cons, "%s: %s\n", name, title);
+	} else if (*desc) {
+		r_cons_printf (core->cons, "%s: %s\n", name, desc);
+	} else {
+		r_cons_println (core->cons, name);
+	}
+}
+
+static char *prompt_name_from_file(const char *file) {
+	return r_str_ndup (file, strlen (file) - strlen (".r2ai.md"));
+}
+
+static void prompt_json_ks(PJ *pj, const char *key, const char *value) {
+	pj_ks (pj, key, r_str_get (value));
+}
+
+static void prompt_print_json(PJ *pj, const char *name, const char *filepath, R2AIPrompt *prompt) {
+	pj_o (pj);
+	pj_ks (pj, "name", name);
+	pj_ks (pj, "path", filepath);
+	pj_kb (pj, "valid", prompt != NULL);
+	prompt_json_ks (pj, "title", prompt? prompt->title: NULL);
+	prompt_json_ks (pj, "description", prompt? prompt->desc: NULL);
+	prompt_json_ks (pj, "author", prompt? prompt->author: NULL);
+	prompt_json_ks (pj, "command", prompt? prompt->command: NULL);
+	prompt_json_ks (pj, "prompt", prompt? prompt->prompt: NULL);
+	prompt_json_ks (pj, "requires", prompt? prompt->requires: NULL);
+	prompt_json_ks (pj, "if_empty", prompt? prompt->if_empty: NULL);
+	prompt_json_ks (pj, "if_command", prompt? prompt->if_command: NULL);
+	prompt_json_ks (pj, "model", prompt? prompt->model: NULL);
+	prompt_json_ks (pj, "provider", prompt? prompt->provider: NULL);
+	pj_end (pj);
+}
+
+static void prompt_list(RCore *core, RList *search_dirs, bool json) {
+	RList *seen = r_list_newf (free);
+	PJ *pj = NULL;
+	if (json) {
+		pj = r_core_pj_new (core);
+		pj_a (pj);
+	}
+	RListIter *dir_iter;
+	char *dir;
+	r_list_foreach (search_dirs, dir_iter, dir) {
+		RList *files = r_sys_dir (dir);
+		RListIter *iter;
+		char *file;
+		r_list_foreach (files, iter, file) {
+			if (!r_str_endswith (file, ".r2ai.md")) {
+				continue;
+			}
+			char *name = prompt_name_from_file (file);
+			if (prompt_name_seen (seen, name)) {
+				free (name);
+				continue;
+			}
+			r_list_append (seen, strdup (name));
+			char *filepath = find_prompt_file (search_dirs, name);
+			if (!filepath) {
+				free (name);
+				continue;
+			}
+			R2AIPrompt *prompt = parse_prompt_file (filepath);
+			if (json) {
+				prompt_print_json (pj, name, filepath, prompt);
+			} else if (prompt) {
+				prompt_print_summary (core, name, prompt);
+			} else {
+				r_cons_println (core->cons, name);
+			}
+			r2aiprompt_free (prompt);
+			free (filepath);
+			free (name);
+		}
+		r_list_free (files);
+	}
+	if (json) {
+		pj_end (pj);
+		char *s = pj_drain (pj);
+		r_cons_println (core->cons, s);
+		free (s);
+	}
+	r_list_free (seen);
+}
+
+R_API void r2ai_cmd_qj(RCorePluginSession *cps, const char *input) {
+	(void)input;
+	RCore *core = cps->core;
+	RList *search_dirs = prompt_search_dirs_or_warn (core);
+	if (!search_dirs) {
+		return;
+	}
+	prompt_list (core, search_dirs, true);
+	r_list_free (search_dirs);
+}
+
+R_API void r2ai_cmd_q(RCorePluginSession *cps, const char *input) {
+	RCore *core = cps->core;
+	RList *search_dirs = prompt_search_dirs_or_warn (core);
+	if (!search_dirs) {
 		return;
 	}
 	if (!input || !*input) {
-		// list prompts
-		RListIter *dir_iter;
-		char *dir;
-		r_list_foreach (search_dirs, dir_iter, dir) {
-			RList *files = r_sys_dir (dir);
-			RListIter *iter;
-			char *file;
-			r_list_foreach (files, iter, file) {
-				if (r_str_endswith (file, ".r2ai.txt") || r_str_endswith (file, ".r2ai.md")) {
-					char *dot = strchr (file, '.');
-					char *name = dot? r_str_ndup (file, dot - file): strdup (file);
-					char *filepath_txt = r_str_newf ("%s/%s.r2ai.txt", dir, name);
-					char *filepath_md = r_str_newf ("%s/%s.r2ai.md", dir, name);
-					char *filepath = NULL;
-					if (r_file_exists (filepath_txt)) {
-						filepath = filepath_txt;
-					} else if (r_file_exists (filepath_md)) {
-						filepath = filepath_md;
-					}
-					if (filepath) {
-						R2AIPrompt *prompt = parse_prompt_file (filepath);
-						if (prompt) {
-							r_cons_printf (core->cons, "%s: %s - %s\n", name, prompt->title? prompt->title: "", prompt->desc? prompt->desc: "");
-							r2aiprompt_free (prompt);
-						} else {
-							r_cons_println (core->cons, name);
-						}
-					}
-					free (name);
-					free (filepath_txt);
-					free (filepath_md);
-				}
-			}
-			r_list_free (files);
-		}
+		prompt_list (core, search_dirs, false);
 	} else {
 		// run prompt
 		char *name = strdup (input);
@@ -241,7 +344,7 @@ R_API void r2ai_cmd_q(RCorePluginSession *cps, const char *input) {
 		r_str_trim (name);
 		char *filepath = find_prompt_file (search_dirs, name);
 		if (!filepath) {
-			R_LOG_ERROR ("Cannot find prompt file: %s.r2ai.txt or %s.r2ai.md", name, name);
+			R_LOG_ERROR ("Cannot find prompt file: %s.r2ai.md", name);
 			r_list_free (search_dirs);
 			free (name);
 			return;
@@ -254,8 +357,8 @@ R_API void r2ai_cmd_q(RCorePluginSession *cps, const char *input) {
 			free (name);
 			return;
 		}
-		if (!prompt->title || !prompt->command) {
-			R_LOG_WARN ("Prompt %s is missing required Title or Command directive", name);
+		if (!prompt->command) {
+			R_LOG_WARN ("Prompt %s is missing required Command directive", name);
 		}
 		free (filepath);
 		// run commands
@@ -314,26 +417,7 @@ R_API void r2ai_cmd_q(RCorePluginSession *cps, const char *input) {
 }
 
 R_API char *r2ai_load_prompt_text(RCore *core, const char *name) {
-	const char *promptdir = r_config_get (core->config, "r2ai.promptdir");
-	if (!promptdir || !*promptdir) {
-		promptdir = "~/.config/r2ai/prompts";
-	}
-	char *expanded_dir = r_file_abspath (promptdir);
-	char *local_prompts = r_file_abspath ("../prompts");
-	RList *search_dirs = r_list_newf (free);
-	if (r_file_is_directory (expanded_dir)) {
-		r_list_append (search_dirs, expanded_dir);
-	} else {
-		free (expanded_dir);
-		expanded_dir = NULL;
-	}
-	if (r_file_is_directory (local_prompts)) {
-		r_list_append (search_dirs, local_prompts);
-	} else {
-		free (local_prompts);
-		local_prompts = NULL;
-	}
-
+	RList *search_dirs = prompt_search_dirs (core);
 	char *filepath = find_prompt_file (search_dirs, name);
 	char *prompt_text = NULL;
 	if (filepath) {

src/r2ai.1

@@ -14,6 +14,9 @@ Decompile the current function.
 .B \-d \fIquery\fR
 Ask a question about the current function.
 .TP
+.B \-do
+Decompile the current function with source offsets in the left column.
+.TP
 .B \-dr
 Decompile the current function with one level of recursivity.
 .TP
@@ -26,6 +29,15 @@ Set configuration option (same as 'r2ai.' prefix).
 .B \-h
 Show help message.
 .TP
+.B \-f \fIfile\fR
+Load a file with the radare2 API before running scripts, commands, or queries.
+.TP
+.B \-s \fIaddr\fR
+Seek to an address after loading the file specified with \fB\-f\fR.
+.TP
+.B \-c \fIcommand\fR
+Execute a radare2 command after loading the file. Can be used multiple times.
+.TP
 .B \-i \fIfile\fR \fIquery\fR
 Read a file and ask the LLM with the given query.
 .TP
@@ -38,6 +50,9 @@ Set LLM provider (openai, anthropic, gemini, etc.).
 .B \-q
 List available query prompts.
 .TP
+.B \-qj
+List available query prompts in JSON.
+.TP
 .B \-q \fIname\fR [\fIinstructions\fR]
 Run predefined prompt with optional custom instructions.
 .TP
@@ -48,7 +63,7 @@ Enter the chat REPL.
 Launch the interactive setup wizard.
 .TP
 .B \-L
-Show chat logs (use -Lj for JSON). Only for auto mode.
+Show chat logs (use -Lj for JSON).
 .TP
 .B \-L-\fIN\fR
 Delete the last N messages from the chat history.
@@ -71,7 +86,13 @@ Name of the provider (e.g., openai). List possibilities with \fBr2ai -e r2ai.api
 Model name. List possibilities with \fBr2ai -e r2ai.model=?\fR.
 .TP
 .B r2ai.baseurl
-Remote LLM base URL (e.g., http://127.0.0.1:11434).
+Provider API root. A /v1 suffix selects OpenAI-compatible chat completions; /api selects native Ollama.
+.TP
+.B r2ai.apitype
+Use chat completions (default) or the native Ollama generate operation.
+.TP
+.B r2ai.cacheck
+Warn when chat requests are not append-only and may miss provider prompt-cache hits.
 .TP
 .B r2ai.max_tokens
 Maximum output tokens or total tokens.
@@ -94,6 +115,9 @@ User prompt for \fBr2ai -d\fR.
 .B r2ai.clippy
 Display chat replies using r2clippy instead of plain text.
 .TP
+.B r2ai.markdown
+Render LLM responses with the radare2 markdown colorizer (default: true).
+.TP
 .B r2ai.wizard
 Auto-start the setup wizard on the first interactive chat session.
 .TP
@@ -108,6 +132,9 @@ Allow AI to run commands without approval (dangerous, default false).
 .TP
 .B r2ai.http.use_files
 Use files for arguments if CURL argument too long.
+.TP
+.B r2ai.http.headers
+Extra HTTP headers added to every provider request, one per line ('\\n' separated). Defaults to the R2AI_HEADERS environment variable.
 .SH QUERY PROMPTS
 Query prompts are predefined templates for common tasks. Available prompts:
 .TP
@@ -142,6 +169,9 @@ Find vulnerabilities in the current function.
 Decompile current function:
 .B r2ai -d
 .TP
+Decompile current function with source offsets:
+.B r2ai -do
+.TP
 Ask about current function:
 .B r2ai -d "what is the input buffer size?"
 .TP
@@ -151,12 +181,21 @@ Use auto mode:
 List prompts:
 .B r2ai -q
 .TP
+List prompts as JSON:
+.B r2ai -qj
+.TP
 Run a prompt:
 .B r2ai -q explain
 .TP
 Set provider:
 .B r2ai -p openai
 .TP
+Decompile a function from the shell:
+.B r2ai -f ./program -s main -d
+.TP
+Ask about a function from the shell:
+.B r2ai -f ./program -s main -d "what does this function do?"
+.TP
 Enter REPL:
 .B r2ai -r
 .SH SEE ALSO

src/r2ai.c

@@ -1,16 +1,17 @@
-/* r2ai - Copyright 2023-2025 pancake */
+/* r2ai - Copyright 2023-2026 pancake */
 
 #define R_LOG_ORIGIN "r2ai"
 
 #include "r2ai.h"
 #include "r2ai_priv.h"
 
+// clang-format off
 static RCoreHelpMessage help_msg_r2ai = {
-	"Usage:",
-	"r2ai", " [-args] [...]",
+	"Usage:", "r2ai", " [-args] [...]",
 	"r2ai", " -w", "Launch the interactive setup wizard",
 	"r2ai", " -d", "Decompile current function",
 	"r2ai", " -d [query]", "Ask a question on the current function",
+	"r2ai", " -do", "Decompile current function with source offsets",
 	"r2ai", " -dr", "Decompile current function (+ 1 level of recursivity)",
 	"r2ai", " -a [query]", "Resolve question using auto mode",
 	"r2ai", " -b [url]", "set/show base URL for provider API",
@@ -19,12 +20,17 @@ static RCoreHelpMessage help_msg_r2ai = {
 	"r2ai", " -K", "Edit the API keys file",
 	"r2ai", " -h", "Show this help message",
 	"r2ai", " -i [file] [query]", "read file and ask the llm with the given query",
+	"r2ai", " -id [hint]", "generate SOUL.md and IDENTITY.md personality files (optional hint)",
+	"r2ai", " -ide", "edit SOUL.md and IDENTITY.md personality files",
+	"r2ai", " -id-", "delete SOUL.md and IDENTITY.md personality files",
 	"r2ai", " -m [model]", "show selected model, list suggested ones, choose one",
 	"r2ai", " -p [provider]", "set LLM provider (openai, anthropic, gemini, etc.)",
 	"r2ai", " -q", "list available query prompts",
+	"r2ai", " -qj", "list available query prompts in JSON",
 	"r2ai", " -q [name] (inst)", "run predefined prompt with optional instructions",
 	"r2ai", " -r", "enter the chat repl",
-	"r2ai", " -L", "show chat logs (See -Lj for json). Only for auto mode.",
+	"r2ai", " -s[?ijynak*]", "async task queue: list, show, interact, approve, decline, kill (see r2ai -s?)",
+	"r2ai", " -L", "show chat logs (See -Lj for json)",
 	"r2ai", " -C", "compact conversation history",
 	"r2ai", " -LR", "create a log report",
 	"r2ai", " -L-[N]", "delete the last (or N last messages from the chat history)",
@@ -33,9 +39,37 @@ static RCoreHelpMessage help_msg_r2ai = {
 	"r2ai", " [query]", "query the selected model+provider with the given query",
 	NULL
 };
+// clang-format on
+
+static char *r2ai_default_auto_prompt(void) {
+	return strdup (
+		"You are a reverse engineer and you are using radare2 to analyze a binary.\n"
+		"The user will ask questions about the binary and you will respond with the answer to the best of your ability.\n"
+		"\n"
+		"# Guidelines\n"
+		"- Understand the Task: Grasp the main objective, goals, requirements, constraints, and expected output.\n"
+		"- Reasoning Before Conclusions**: Encourage reasoning steps before any conclusions are reached.\n"
+		"- Assume the user is always asking you about the binary, unless they're specifically asking you for radare2 help.\n"
+		"- The binary has already been loaded. You can interact with the binary using the r2cmd tool.\n"
+		"- `this` or `here` might refer to the current address in the binary or the binary itself.\n"
+		"- If you need more information, try to use the r2cmd tool to run commands before answering.\n"
+		"- You can use the r2cmd tool multiple times if you need or you can pass a command with pipes if you need to chain commands.\n"
+		"- IMPORTANT: You must make only ONE tool call per response. Do not make multiple tool calls in a single response.\n"
+		"- Keep your reasoning concise and avoid repeating the same thoughts. If you have a clear next step, execute it immediately with a tool call.\n"
+		"- If you're asked to decompile a function, make sure to return the code in the language you think it was originally written "
+		"and rewrite it to be as easy as possible to be understood. Make sure you use descriptive variable and function names and add comments.\n"
+		"- Don't just regurgitate the same code, figure out what it's doing and rewrite it to be more understandable.\n"
+		"- If you need to run a command in r2 before answering, you can use the r2cmd tool\n"
+		"- Do not repeat commands if you already know the answer.\n"
+		"- Formulate a plan. Think step by step. Analyze the binary as much as possible before answering.\n"
+		"- You must keep going until you have a final answer.\n"
+		"- Double check that final answer. Make sure you didn't miss anything.\n"
+		"- Make sure you call tools and functions correctly.\n"
+		"- When calling tools, use the exact format: {\"tool_calls\": [{\"id\": \"call_123\", \"type\": \"function\", \"function\": {\"name\": \"r2cmd\", \"arguments\": \"{\\\"command\\\":\\\"aa\\\"}\"}}]}\n"
+		"- Arguments must be a JSON string, not an object.\n");
+}
 
 R_API char *r2ai(RCorePluginSession *cps, R2AIArgs args) {
-	RCore *core = cps->core;
 	if (R_STR_ISEMPTY (args.input) && !args.messages) {
 		if (args.error) {
 			*args.error = r_str_newf ("Usage: r2ai [-h] [prompt]");
@@ -43,63 +77,195 @@ R_API char *r2ai(RCorePluginSession *cps, R2AIArgs args) {
 		return NULL;
 	}
 
-	if (R_STR_ISEMPTY (args.system_prompt)) {
-		args.system_prompt = r_config_get (core->config, "r2ai.system");
+	RList *msgs = args.messages;
+	bool own_msgs = false;
+	if (args.input || !msgs) {
+		msgs = r2ai_msgs_new ();
+		own_msgs = true;
+		if (args.messages) {
+			RListIter *iter;
+			R2AI_Message *m;
+			r_list_foreach (args.messages, iter, m) {
+				r2ai_msgs_add (msgs, m);
+			}
+		}
+		if (args.input) {
+			R2AI_Message m = { .role = "user", .content = (char *)args.input };
+			r2ai_msgs_add (msgs, &m);
+		}
+		args.messages = msgs;
 	}
 
-	RList *msgs = r2ai_msgs_new ();
-
-	if (args.input) {
-		R2AI_Message msg = { .role = "user", .content = (char *)args.input };
-		r2ai_msgs_add (msgs, &msg);
-	}
-
-	args.messages = msgs;
-
 	char *content = NULL;
 	R2AI_ChatResponse *res = r2ai_llmcall (cps, args);
 	if (res) {
 		if (res->message && res->message->content) {
 			content = strdup (res->message->content);
 		}
-		r2ai_message_free ((R2AI_Message *)res->message);
-		free (res);
+		r2ai_chat_response_free (res);
+	}
+	if (own_msgs) {
+		r2ai_msgs_free (msgs);
 	}
 	return content;
 }
 
-static void cmd_r2ai_d(RCorePluginSession *cps, const char *input, const bool recursive) {
-	RCore *core = cps->core;
-	const char *prompt = r_config_get (core->config, "r2ai.prompt");
-	const char *lang = r_config_get (core->config, "r2ai.lang");
-	char *full_prompt;
-	if (!R_STR_ISEMPTY (input)) {
-		R_LOG_DEBUG ("User question: %s", input);
-		full_prompt = strdup (input);
-	} else {
-		if (!R_STR_ISEMPTY (lang)) {
-			full_prompt = r_str_newf ("%s. Translate the code into %s programming language.", prompt, lang);
-		} else {
-			full_prompt = strdup (prompt);
+static bool r2ai_cmd_is_word(const char *cmd, const char *word) {
+	const size_t len = strlen (word);
+	return !strncmp (cmd, word, len) && (!cmd[len] || IS_WHITESPACE (cmd[len]));
+}
+
+static char *r2ai_offset_cmd(const char *cmd) {
+	char *trimmed = strdup (cmd);
+	r_str_trim (trimmed);
+	const char *replacement = NULL;
+	if (r2ai_cmd_is_word (trimmed, "pdc")) {
+		replacement = "pdco";
+	} else if (r2ai_cmd_is_word (trimmed, "pdg")) {
+		replacement = "pdgo";
+	} else if (r2ai_cmd_is_word (trimmed, "pdd")) {
+		replacement = "pddo";
+	}
+	if (replacement) {
+		char *res = r_str_newf ("%s%s", replacement, trimmed + 3);
+		free (trimmed);
+		return res;
+	}
+	return trimmed;
+}
+
+static void r2ai_collect_offsets(RList *offsets, const char *text, int *width) {
+	const char *p = text;
+	while (p && *p) {
+		const char *line = r_str_trim_head_ro (p);
+		if (r_str_startswith (line, "0x")) {
+			const char *q = line + 2;
+			while (IS_HEXCHAR (*q)) {
+				q++;
+			}
+			if (q > line + 2) {
+				int len = q - line;
+				r_list_append (offsets, r_str_ndup (line, len));
+				if (len > *width) {
+					*width = len;
+				}
+			}
+		}
+		p = strchr (p, '\n');
+		if (p) {
+			p++;
 		}
 	}
+}
+
+static const char *r2ai_next_offset(RListIter **iter) {
+	if (!iter || !*iter) {
+		return NULL;
+	}
+	const char *addr = (*iter)->data;
+	*iter = (*iter)->n;
+	return addr;
+}
+
+static bool r2ai_line_has_statement(const char *line) {
+	const char *trimmed = r_str_trim_head_ro (line);
+	if (R_STR_ISEMPTY (trimmed)) {
+		return false;
+	}
+	if (!strcmp (trimmed, "{") || !strcmp (trimmed, "}") || !strcmp (trimmed, "};")) {
+		return false;
+	}
+	if (*trimmed == '#') {
+		return false;
+	}
+	return true;
+}
+
+static char *r2ai_apply_offsets(const char *text, RList *offsets, int width) {
+	if (R_STR_ISEMPTY (text) || !offsets || r_list_empty (offsets)) {
+		return text? strdup (text): NULL;
+	}
+	if (width < 10) {
+		width = 10;
+	}
+	RStrBuf *sb = r_strbuf_new ("");
+	RListIter *offiter = offsets->head;
+	const char *p = text;
+	while (p && *p) {
+		const char *nl = strchr (p, '\n');
+		int len = nl? (int)(nl - p): (int)strlen (p);
+		char *line = r_str_ndup (p, len);
+		const char *addr = r2ai_line_has_statement (line)? r2ai_next_offset (&offiter): NULL;
+		r_strbuf_appendf (sb, "%*s | %s\n", width, addr? addr: "", line);
+		free (line);
+		p = nl? nl + 1: NULL;
+	}
+	return r_strbuf_drain (sb);
+}
+
+// query the llm with rag and log the error if any, returns NULL on failure
+// when r2ai.history is set the exchange is kept in the conversation (see -L, -R)
+static char *ask(RCorePluginSession *cps, const char *input, bool chat) {
+	R2AI_State *state = cps->data;
+	RList *history = (chat && r_config_get_b (cps->core->config, "r2ai.history"))
+		? r2ai_conversation_get (state)
+		: NULL;
+	char *err = NULL;
+	char *res = r2ai (cps, (R2AIArgs){ .input = input, .messages = history, .error = &err, .dorag = true });
+	if (err) {
+		R_LOG_ERROR ("%s", err);
+		free (err);
+		R_FREE (res);
+	}
+	if (history && res) {
+		R2AI_Message user = { .role = "user", .content = (char *)input };
+		R2AI_Message assistant = { .role = "assistant", .content = res };
+		r2ai_msgs_add (history, &user);
+		r2ai_msgs_add (history, &assistant);
+	}
+	return res;
+}
+
+static void add_block(RStrBuf *sb, char *dec, RList *offsets, int *width) {
+	if (offsets) {
+		r2ai_collect_offsets (offsets, dec, width);
+	}
+	r_strbuf_appendf (sb, "\n[BEGIN]\n%s[END]\n", dec);
+	free (dec);
+}
+
+static void cmd_r2ai_d(RCorePluginSession *cps, const char *input, const bool recursive, const bool offsets) {
+	RCore *core = cps->core;
+	const char *lang = r_config_get (core->config, "r2ai.lang");
+	RStrBuf *sb = r_strbuf_new (R_STR_ISNOTEMPTY (input)? input: r_config_get (core->config, "r2ai.prompt"));
+	if (R_STR_ISEMPTY (input) && R_STR_ISNOTEMPTY (lang)) {
+		r_strbuf_appendf (sb, ". Translate the code into %s programming language.", lang);
+	}
 	char *cmds = strdup (r_config_get (core->config, "r2ai.cmds"));
-	RStrBuf *sb = r_strbuf_new (full_prompt);
 	RList *cmdslist = r_str_split_list (cmds, ",", -1);
 	RListIter *iter;
 	const char *cmd;
 	RList *refslist = NULL;
+	char *refs = NULL;
+	RList *offsetslist = offsets? r_list_newf (free): NULL;
+	char *offset_fallback = NULL;
+	int offset_width = 0;
 	if (recursive) {
-		char *refs = r_core_cmd_str (core, "axff~^C[2]~$$");
+		refs = r_core_cmd_str (core, "axff~^C[2]~$$");
 		refslist = r_str_split_list (refs, ",", 0);
-		free (refs);
 	}
+	RConfigHold *hold = r_config_hold_new (core->config);
+	r_config_hold (hold, "scr.color", "scr.utf8", NULL);
+	r_config_set_i (core->config, "scr.color", 0);
+	r_config_set_b (core->config, "scr.utf8", false);
 	r_list_foreach (cmdslist, iter, cmd) {
-		char *dec = r_core_cmd_str (core, cmd);
-		r_strbuf_append (sb, "\n[BEGIN]\n");
-		r_strbuf_append (sb, dec);
-		r_strbuf_append (sb, "[END]\n");
-		free (dec);
+		char *ocmd = offsets? r2ai_offset_cmd (cmd): NULL;
+		const char *dcmd = ocmd? ocmd: cmd;
+		char *dec = r_core_cmd_str (core, dcmd);
+		if (offsets && !offset_fallback) {
+			offset_fallback = strdup (dec);
+		}
+		add_block (sb, dec, offsetslist, &offset_width);
 		if (recursive) {
 			RListIter *iter2;
 			char *at;
@@ -108,28 +274,61 @@ static void cmd_r2ai_d(RCorePluginSession *cps, const char *input, const bool re
 				if (core->num->nc.errors) {
 					continue;
 				}
-				char *dec = r_core_cmd_str_at (core, n, cmd);
-				r_strbuf_append (sb, "\n[BEGIN]\n");
-				r_strbuf_append (sb, dec);
-				r_strbuf_append (sb, "[END]\n");
-				free (dec);
+				add_block (sb, r_core_cmd_str_at (core, n, dcmd), offsetslist, &offset_width);
 			}
 		}
+		free (ocmd);
 	}
+	r_config_hold_restore (hold);
+	r_config_hold_free (hold);
 	r_list_free (refslist);
+	free (refs);
 	char *s = r_strbuf_drain (sb);
-	char *error = NULL;
-	R2AIArgs d_args = { .input = s, .error = &error, .dorag = true };
-	char *res = r2ai (cps, d_args);
-	free (s);
-	if (error) {
-		R_LOG_ERROR (error);
-		free (error);
+	if (r_config_get_b (core->config, "r2ai.async")) {
+		const char *q = r_str_get (input);
+		char *title = r_str_newf ("-d%s%s%s%s", recursive? "r": "", offsets? "o": "", *q? " ": "", q);
+		r2ai_async_query (cps, title, s, NULL);
+		free (title);
 	} else {
-		r_cons_printf (core->cons, "%s\n", res);
+		char *res = ask (cps, s, true);
+		if (res && offsets) {
+			char *ores = R_STR_ISNOTEMPTY (res)
+				? r2ai_apply_offsets (res, offsetslist, offset_width)
+				: (offset_fallback? strdup (offset_fallback): NULL);
+			if (ores) {
+				r_cons_printf (core->cons, "%s\n", ores);
+				free (ores);
+			}
+		} else if (res) {
+			r_cons_printf (core->cons, "%s\n", res);
+		}
+		free (res);
 	}
-	free (res);
+	free (s);
+	free (offset_fallback);
+	r_list_free (offsetslist);
 	r_list_free (cmdslist);
+	free (cmds);
+}
+
+R_IPI void r2ai_print_response(RCore *core, const char *text) {
+	if (!text) {
+		return;
+	}
+#if R2_VERSION_NUMBER >= 60106
+	if (r_config_get_b (core->config, "r2ai.markdown")) {
+		char *md = r_core_md2txt (core, text, false);
+		if (md) {
+			r_cons_print (core->cons, md);
+			if (!r_str_endswith (md, "\n")) {
+				r_cons_newline (core->cons);
+			}
+			free (md);
+			return;
+		}
+	}
+#endif
+	r_cons_println (core->cons, text);
 }
 
 static void cmd_r2ai_repl(RCorePluginSession *cps) {
@@ -159,21 +358,15 @@ static void cmd_r2ai_repl(RCorePluginSession *cps) {
 			}
 		}
 		r_strbuf_appendf (sb, "User: %s\n", ptr);
-		const char *a = r_strbuf_tostring (sb);
-		char *error = NULL;
-		char *res =
-			r2ai (cps, (R2AIArgs){ .input = a, .error = &error, .dorag = true });
-		if (error) {
-			R_LOG_ERROR ("%s", error);
-			free (error);
-		} else if (res) {
+		char *res = ask (cps, r_strbuf_tostring (sb), false);
+		if (res) {
 			r_strbuf_appendf (sb, "Assistant: %s\n", res);
 			if (r_config_get_b (core->config, "r2ai.clippy")) {
 				char *cmd = r_str_newf ("?E %s", res);
-				r_core_cmd_call (core, cmd);
+				r_core_call (core, cmd);
 				free (cmd);
 			} else {
-				r_cons_printf (core->cons, "%s\n", res);
+				r2ai_print_response (core, res);
 			}
 			r_cons_flush (core->cons);
 		}
@@ -184,41 +377,30 @@ static void cmd_r2ai_repl(RCorePluginSession *cps) {
 	r_config_hold_free (hold);
 }
 
-static void cmd_r2ai_R(RCorePluginSession *cps, const char *q) {
+static void cmd_r2ai_R(RCorePluginSession *cps, const char *q, bool refresh) {
 	RCore *core = cps->core;
 	R2AI_State *state = cps->data;
-	if (!r_config_get_b (core->config, "r2ai.data")) {
+	if (refresh && !r_config_get_b (core->config, "r2ai.data")) {
 		R_LOG_ERROR ("r2ai -e r2ai.data=true");
 		return;
 	}
-	if (R_STR_ISEMPTY (q)) {
-		if (state->db) {
-			r_vdb_free (state->db);
-			state->db = NULL;
-		}
+	if (refresh || !state->db) {
 		r2ai_refresh_embeddings (cps);
-	} else {
-		r2ai_refresh_embeddings (cps);
-		const int K = r_config_get_i (core->config, "r2ai.data.nth");
-		RVdbResultSet *rs = r_vdb_query (state->db, q, K);
-
-		if (rs) {
-			R_LOG_DEBUG ("Query: \"%s\"", q);
-			R_LOG_DEBUG ("Found up to %d neighbors (actual found: %d)", K, rs->size);
-			int i;
-			for (i = 0; i < rs->size; i++) {
-				RVdbResult *r = &rs->results[i];
-				KDNode *n = r->node;
-				float dist_sq = r->dist_sq;
-				float cos_sim = 1.0f - (dist_sq * 0.5f); // for normalized vectors
-				printf ("%2d) dist_sq=%.4f cos_sim=%.4f text=\"%s\"\n", i + 1, dist_sq,
-					cos_sim, (n->text? n->text: "(null)"));
-			}
-			r_vdb_result_free (rs);
-		} else {
-			R_LOG_ERROR ("No vdb results found");
-		}
 	}
+	if (R_STR_ISEMPTY (q)) {
+		return;
+	}
+	RVdbResultSet *rs = r_vdb_query (state->db, q, r_config_get_i (core->config, "r2ai.data.nth"));
+	if (!rs) {
+		R_LOG_ERROR ("No vdb results found");
+		return;
+	}
+	int i;
+	for (i = 0; i < rs->size; i++) {
+		RVdbResult *r = &rs->results[i];
+		r_cons_printf (core->cons, "%2d) %.4f %s\n", i + 1, r->dist_sq, r->node->text);
+	}
+	r_vdb_result_free (rs);
 }
 
 static void cmd_r2ai_i(RCorePluginSession *cps, const char *arg) {
@@ -229,25 +411,19 @@ static void cmd_r2ai_i(RCorePluginSession *cps, const char *arg) {
 		*query++ = 0;
 	}
 	char *s = r_file_slurp (fname, NULL);
-	if (R_STR_ISNOTEMPTY (s)) {
+	if (!s) {
 		R_LOG_ERROR ("Cannot open %s", fname);
 		free (fname);
 		return;
 	}
-	char *q = r_str_newf ("%s\n```\n%s\n```\n", query, s);
-	char *error = NULL;
-	char *res =
-		r2ai (cps, (R2AIArgs){ .input = q, .error = &error, .dorag = true });
+	const char *prompt = R_STR_ISNOTEMPTY (query)? query: "Analyze the contents of the following file:";
+	char *q = r_str_newf ("%s\n```\n%s\n```\n", prompt, s);
 	free (s);
-	if (error) {
-		R_LOG_ERROR (error);
-		free (error);
-	} else {
-		r_cons_printf (core->cons, "%s\n", res);
-	}
+	char *res = ask (cps, q, true);
+	free (q);
+	r2ai_print_response (core, res);
 	free (fname);
 	free (res);
-	free (q);
 }
 
 static void cmd_r2ai_m(RCorePluginSession *cps, const char *input) {
@@ -259,41 +435,6 @@ static void cmd_r2ai_m(RCorePluginSession *cps, const char *input) {
 	}
 }
 
-static void load_embeddings(RCorePluginSession *cps) {
-	RCore *core = cps->core;
-	R2AI_State *state = cps->data;
-	RListIter *iter, *iter2;
-	char *line;
-	char *file;
-	// enumerate .txt files in directory
-	const char *path = r_config_get (core->config, "r2ai.data.path");
-	RList *files = r_sys_dir (path);
-	if (r_list_empty (files)) {
-		R_LOG_WARN ("Cannot find any file in r2ai.data.path");
-	}
-	r_list_foreach (files, iter, file) {
-		if (!r_str_endswith (file, ".txt")) {
-			continue;
-		}
-		R_LOG_DEBUG ("Index %s", file);
-		char *filepath = r_file_new (path, file, NULL);
-		char *text = r_file_slurp (filepath, NULL);
-		if (text) {
-			RList *lines = r_str_split_list (text, "\n", -1);
-			r_list_foreach (lines, iter2, line) {
-				if (r_str_trim_head_ro (line)[0] == 0) {
-					continue;
-				}
-				r_vdb_insert (state->db, line);
-				R_LOG_DEBUG ("Insert %s", line);
-			}
-			r_list_free (lines);
-		}
-		free (filepath);
-	}
-	r_list_free (files);
-}
-
 static bool load_r2airc(RCorePluginSession *cps) {
 	char *rc_path = r_file_home (".config/r2ai/rc");
 	if (rc_path && r_file_exists (rc_path)) {
@@ -307,12 +448,11 @@ static bool load_r2airc(RCorePluginSession *cps) {
 				if (R_STR_ISEMPTY (trimmed) || *trimmed == '#') {
 					continue;
 				}
-				if (r_str_startswith (trimmed, "r2ai -e ") || r_str_startswith (trimmed, "-e ")) {
-					const char *arg = trimmed + (r_str_startswith (trimmed, "r2ai -e ")? 7: 3);
-					cmd_r2ai (cps, r_str_newf ("-e%s", arg));
-				} else if (r_str_startswith (trimmed, "r2ai -") || r_str_startswith (trimmed, "-")) {
-					const char *arg = trimmed + (r_str_startswith (trimmed, "r2ai -")? 6: 1);
-					cmd_r2ai (cps, r_str_newf ("-%s", arg));
+				if (r_str_startswith (trimmed, "r2ai -")) {
+					trimmed += strlen ("r2ai ");
+				}
+				if (*trimmed == '-') {
+					cmd_r2ai (cps, trimmed);
 				} else {
 					R_LOG_ERROR ("Invalid line in r2ai rc: %s", trimmed);
 				}
@@ -325,18 +465,31 @@ static bool load_r2airc(RCorePluginSession *cps) {
 	return true;
 }
 
+static bool cmd_r2ai_id(RCorePluginSession *cps, const char *input) {
+	if (r_str_startswith (input, "-id-")) {
+		r2ai_claw_delete (cps);
+	} else if (r_str_startswith (input, "-ide")) {
+		r2ai_claw_edit (cps);
+	} else if (r_str_startswith (input, "-id")) {
+		r2ai_claw_create (cps, r_str_trim_head_ro (input + 3));
+	} else {
+		return false;
+	}
+	return true;
+}
+
 R_API void cmd_r2ai(RCorePluginSession *cps, const char *input) {
 	RCore *core = cps->core;
 	R2AI_State *state = cps->data;
 	if (*input == '?' || r_str_startswith (input, "-h")) {
-		r_core_cmd_help (core, help_msg_r2ai);
+		r2ai_cmd_help (core, help_msg_r2ai);
 	} else if (r_str_startswith (input, "-w")) {
 		r2ai_wizard (core);
 	} else if (r_str_startswith (input, "-K")) {
 		r2ai_apikeys_edit (cps);
 	} else if (r_str_startswith (input, "-E")) {
 		char *rc_path = r_file_home (".config/r2ai/rc");
-		r_cons_editor (core->cons, rc_path, NULL);
+		r2ai_cons_editor (core->cons, rc_path, NULL);
 		free (rc_path);
 		load_r2airc (cps);
 	} else if (r_str_startswith (input, "-e")) {
@@ -347,7 +500,14 @@ R_API void cmd_r2ai(RCorePluginSession *cps, const char *input) {
 			r_core_cmdf (core, "-e r2ai.%s", arg);
 		}
 	} else if (r_str_startswith (input, "-a")) {
-		cmd_r2ai_a (cps, r_str_trim_head_ro (input + 2));
+		const char *q = r_str_trim_head_ro (input + 2);
+		if (r_config_get_b (core->config, "r2ai.async")) {
+			char *title = r_str_newf ("-a %s", q);
+			r2ai_async_auto (cps, title, q, NULL);
+			free (title);
+		} else {
+			cmd_r2ai_a (cps, q);
+		}
 	} else if (r_str_startswith (input, "-b")) {
 		const char *baseurl = r_str_trim_head_ro (input + 2);
 		if (R_STR_ISEMPTY (baseurl)) {
@@ -363,51 +523,40 @@ R_API void cmd_r2ai(RCorePluginSession *cps, const char *input) {
 			r_cons_printf (core->cons, "No conversation history available\n");
 		} else {
 			r2ai_delete_last_messages (messages, N);
-			r_cons_printf (core->cons, "Deleted %d message%s from chat history\n", N > 0? N: 1,
-				(N > 0 && N != 1)? "s": "");
+			r_cons_printf (core->cons, "Deleted %d message%s from chat history\n", N > 0? N: 1, (N > 0 && N != 1)? "s": "");
 		}
 	} else if (r_str_startswith (input, "-C")) {
 		cmd_r2ai_c (cps);
+	} else if (r_str_startswith (input, "-s")) {
+		r2ai_async_cmd (cps, input + 2);
 	} else if (r_str_startswith (input, "-LR")) {
 		cmd_r2ai_lr (cps);
 	} else if (r_str_startswith (input, "-L")) {
-		cmd_r2ai_logs (cps);
-	} else if (r_str_startswith (input, "-d")) {
-		cmd_r2ai_d (cps, r_str_trim_head_ro (input + 2), false);
+		cmd_r2ai_logs (cps, input + 2);
+	} else if (r_str_startswith (input, "-do")) {
+		cmd_r2ai_d (cps, r_str_trim_head_ro (input + 3), false, true);
 	} else if (r_str_startswith (input, "-dr")) {
-		cmd_r2ai_d (cps, r_str_trim_head_ro (input + 2), true);
+		cmd_r2ai_d (cps, r_str_trim_head_ro (input + 3), true, false);
+	} else if (r_str_startswith (input, "-d")) {
+		cmd_r2ai_d (cps, r_str_trim_head_ro (input + 2), false, false);
 	} else if (r_str_startswith (input, "-S")) {
-		if (state->db == NULL) {
-			state->db = r_vdb_new (R2AI_DEFAULT_VECTORS);
-			load_embeddings (cps);
-		}
-		const char *arg = r_str_trim_head_ro (input + 2);
-		const int K = 10;
-		RVdbResultSet *rs = r_vdb_query (state->db, arg, K);
-		if (rs) {
-			int i;
-			eprintf ("Found up to %d neighbors (actual found: %d).\n", K, rs->size);
-			for (i = 0; i < rs->size; i++) {
-				RVdbResult *r = &rs->results[i];
-				KDNode *n = r->node;
-				r_cons_printf (core->cons, "- (%.4f) %s\n", r->dist_sq, n->text);
-			}
-			r_vdb_result_free (rs);
-		}
+		cmd_r2ai_R (cps, r_str_trim_head_ro (input + 2), false);
+	} else if (cmd_r2ai_id (cps, input)) {
 	} else if (r_str_startswith (input, "-i")) {
 		cmd_r2ai_i (cps, r_str_trim_head_ro (input + 2));
 	} else if (r_str_startswith (input, "-r")) {
 		cmd_r2ai_repl (cps);
+	} else if (r_str_startswith (input, "-Rq")) {
+		cmd_r2ai_R (cps, r_str_trim_head_ro (input + 3), true);
 	} else if (r_str_startswith (input, "-R")) {
 		RList *messages = r2ai_conversation_get (state);
 		if (!messages || r_list_empty (messages)) {
 			r_cons_printf (core->cons, "No conversation history to reset\n");
 		} else {
 			r2ai_msgs_clear (messages);
+			R_FREE (state->cache_prefix);
 			r_cons_printf (core->cons, "Chat conversation context has been reset\n");
 		}
-	} else if (r_str_startswith (input, "-Rq")) {
-		cmd_r2ai_R (cps, r_str_trim_head_ro (input + 3));
 	} else if (r_str_startswith (input, "-m")) {
 		cmd_r2ai_m (cps, r_str_trim_head_ro (input + 2));
 	} else if (r_str_startswith (input, "-p")) {
@@ -417,20 +566,18 @@ R_API void cmd_r2ai(RCorePluginSession *cps, const char *input) {
 		} else {
 			r_config_set (core->config, "r2ai.api", provider);
 		}
+	} else if (r_str_startswith (input, "-qj")) {
+		r2ai_cmd_qj (cps, r_str_trim_head_ro (input + 3));
 	} else if (r_str_startswith (input, "-q")) {
 		r2ai_cmd_q (cps, r_str_trim_head_ro (input + 2));
 	} else if (r_str_startswith (input, "-")) {
-		r_core_cmd_help (core, help_msg_r2ai);
+		r2ai_cmd_help (core, help_msg_r2ai);
 	} else {
-		char *err = NULL;
-		char *res =
-			r2ai (cps, (R2AIArgs){ .input = input, .error = &err, .dorag = true });
-		if (err) {
-			R_LOG_ERROR ("%s", err);
-			R_FREE (err);
-		}
-		if (res) {
-			r_cons_printf (core->cons, "%s\n", res);
+		if (r_config_get_b (core->config, "r2ai.async")) {
+			r2ai_async_query (cps, input, input, NULL);
+		} else {
+			char *res = ask (cps, input, true);
+			r2ai_print_response (core, res);
 			free (res);
 		}
 	}
@@ -534,11 +681,28 @@ static bool cb_r2ai_http_backend(void *user, void *data) {
 	// return false; // always return false to test
 }
 
+static bool cb_r2ai_apitype(void *user, void *data) {
+	RCore *core = (RCore *)user;
+	RConfigNode *node = (RConfigNode *)data;
+	if (*node->value == '?') {
+		r_cons_println (core->cons, "chat");
+		r_cons_println (core->cons, "generate");
+		return false;
+	}
+	if (!strcmp (node->value, "chat") || !strcmp (node->value, "generate")) {
+		return true;
+	}
+	R_LOG_ERROR ("Invalid apitype '%s'. Use chat or generate.", node->value);
+	return false;
+}
+
 R_IPI bool r2ai_init(RCorePluginSession *cps) {
 	RCore *core = cps->core;
 	// Initialize state
 	R2AI_State *state = R_NEW0 (R2AI_State);
 	cps->data = state;
+	state->prompt_auto = r2ai_default_auto_prompt ();
+	r2ai_markdown_init (&state->markdown);
 
 	// Initialize conversation container
 	r2ai_conversation_init (state);
@@ -558,13 +722,19 @@ R_IPI bool r2ai_init(RCorePluginSession *cps) {
 	r_config_set_cb (core->config, "r2ai.model", R2AI_DEFAULT_MODEL, &cb_r2ai_model);
 	r_config_desc (core->config, "r2ai.model", "Model identifier for the selected provider (e.g. gpt-5-mini)");
 	r_config_set (core->config, "r2ai.baseurl", "");
-	r_config_desc (core->config, "r2ai.baseurl", "Base URL for provider API (overrides default endpoints)");
+	r_config_desc (core->config, "r2ai.baseurl", "Provider API root (/v1 for OpenAI-compatible chat or /api for native Ollama)");
+	r_config_set_cb (core->config, "r2ai.apitype", "chat", &cb_r2ai_apitype);
+	r_config_desc (core->config, "r2ai.apitype", "Chat completions (default) or native Ollama generate operation");
+	r_config_set_b (core->config, "r2ai.cacheck", false);
+	r_config_desc (core->config, "r2ai.cacheck", "Warn when chat requests are not append-only and may miss provider prompt-cache hits");
 	r_config_set_i (core->config, "r2ai.max_tokens", 4096); // max output tokens, or max total tokens
 	r_config_desc (core->config, "r2ai.max_tokens", "Maximum tokens for LLM responses (output/total depending on provider)");
 	r_config_set_i (core->config, "r2ai.thinking_tokens", 0);
 	r_config_desc (core->config, "r2ai.thinking_tokens", "Number of tokens reserved for internal thinking/context messages");
 	r_config_set (core->config, "r2ai.temperature", "0.01");
 	r_config_desc (core->config, "r2ai.temperature", "Sampling temperature for LLM output (0 = deterministic)");
+	r_config_set_b (core->config, "r2ai.deterministic", false);
+	r_config_desc (core->config, "r2ai.deterministic", "Remove randomness from responses by forcing greedy sampling parameters (overrides r2ai.temperature)");
 	r_config_set (core->config, "r2ai.cmds", "pdc");
 	r_config_desc (core->config, "r2ai.cmds", "Default command sequence used by automation (e.g. 'pdc')");
 	r_config_set (core->config, "r2ai.lang", "C");
@@ -580,22 +750,25 @@ R_IPI bool r2ai_init(RCorePluginSession *cps) {
 	r_config_set (core->config, "r2ai.hlang", "english");
 	r_config_desc (core->config, "r2ai.hlang", "Human language for prompts/messages (e.g. english)");
 	r_config_set (
-		core->config, "r2ai.system",
-		"You are a reverse engineer. The user is reversing a binary, using "
-		"radare2. The user will ask questions about the binary and you will "
-		"respond with the answer to the best of your ability.");
+		core->config, "r2ai.system", "You are a reverse engineer. The user is reversing a binary, using "
+			"radare2. The user will ask questions about the binary and you will "
+			"respond with the answer to the best of your ability. "
+			"If none of the available tools fit the request, answer directly "
+			"from your own knowledge instead of inventing tool names. Only the "
+			"tools explicitly listed in this conversation exist.");
 	r_config_set (
-		core->config, "r2ai.prompt",
-		"Rewrite this function and respond ONLY with code, replace goto/labels "
-		"with if/else/for, use NO explanations, NO markdown, Simplify as much as "
-		"possible, use better variable names, take function arguments and "
-		"strings from comments like 'string:'");
+		core->config, "r2ai.prompt", "Rewrite this function and respond ONLY with code, replace goto/labels "
+			"with if/else/for, use NO explanations, NO markdown, Simplify as much as "
+			"possible, use better variable names, take function arguments and "
+			"strings from comments like 'string:'");
 	r_config_set (core->config, "r2ai.promptdir", "~/.config/r2ai/prompts");
 	r_config_desc (core->config, "r2ai.promptdir", "Directory containing .r2ai prompt files");
 	r_config_set_b (core->config, "r2ai.clippy", false);
 	r_config_desc (core->config, "r2ai.clippy", "Responses from the llm will be displayed by clippy");
 	r_config_set_b (core->config, "r2ai.wizard", true);
 	r_config_desc (core->config, "r2ai.wizard", "Run the setup wizard automatically on the first interactive chat session");
+	r_config_set_b (core->config, "r2ai.markdown", true);
+	r_config_desc (core->config, "r2ai.markdown", "Render LLM responses using the radare2 markdown colorizer");
 	r_config_set_b (core->config, "r2ai.stream", false);
 	r_config_desc (core->config, "r2ai.stream", "Enable streaming responses from the LLM (true/false)");
 	r_config_set_i (core->config, "r2ai.auto.max_runs", 50);
@@ -606,6 +779,8 @@ R_IPI bool r2ai_init(RCorePluginSession *cps) {
 	r_config_desc (core->config, "r2ai.auto.init_commands", "Initial commands executed when auto mode starts (semicolon separated)");
 	r_config_set_b (core->config, "r2ai.auto.yolo", false);
 	r_config_desc (core->config, "r2ai.auto.yolo", "Execute potentially dangerous commands in auto mode without asking");
+	r_config_set_b (core->config, "r2ai.history", false);
+	r_config_desc (core->config, "r2ai.history", "Keep direct queries (r2ai [query], -d, -i) in the conversation context (see -L, -R)");
 	r_config_set_b (core->config, "r2ai.auto.reset_on_query", false);
 	r_config_desc (core->config, "r2ai.auto.reset_on_query", "Reset auto-mode conversation state on new user queries");
 	r_config_set_b (core->config, "r2ai.auto.think", true);
@@ -613,12 +788,16 @@ R_IPI bool r2ai_init(RCorePluginSession *cps) {
 
 	r_config_set_i (core->config, "r2ai.http.timeout", 240);
 	r_config_desc (core->config, "r2ai.http.timeout", "HTTP client timeout (seconds) for provider API calls");
-	r_config_set_i (core->config, "r2ai.http.max_retries", 5);
+	r_config_set_i (core->config, "r2ai.http.max_retries", 3);
 	r_config_desc (core->config, "r2ai.http.max_retries", "Maximum number of HTTP retries for failed requests");
 	r_config_set_i (core->config, "r2ai.http.max_backoff", 30);
 	r_config_desc (core->config, "r2ai.http.max_backoff", "Maximum backoff time (seconds) between HTTP retries");
 	r_config_set_b (core->config, "r2ai.http.use_files", false);
 	r_config_desc (core->config, "r2ai.http.use_files", "Use temporary files to pass HTTP request/response payloads (true/false)");
+	char *http_headers = r_sys_getenv ("R2AI_HEADERS");
+	r_config_set (core->config, "r2ai.http.headers", r_str_get (http_headers));
+	free (http_headers);
+	r_config_desc (core->config, "r2ai.http.headers", "Extra HTTP headers ('\\n' separated 'Name: value'; defaults to R2AI_HEADERS env)");
 	r_config_set_b (core->config, "r2ai.auto.raw", false);
 	r_config_desc (core->config, "r2ai.auto.raw", "Use prompt engineering for tool calling instead of native API support (true/false)");
 	r_config_set_b (core->config, "r2ai.auto.usejs", true);
@@ -627,23 +806,35 @@ R_IPI bool r2ai_init(RCorePluginSession *cps) {
 	r_config_desc (core->config, "r2ai.auto.slim", "Use slim mode for auto commands (temporarily disable asm.lines.fcn and scr.utf8)");
 	r_config_set_b (core->config, "r2ai.debug", false);
 	r_config_desc (core->config, "r2ai.debug", "Enable debug output including API request/response details and curl commands");
+	r_config_set_b (core->config, "r2ai.async", false);
+	r_config_desc (core->config, "r2ai.async", "Run LLM calls in background threads (see r2ai -s, -si)");
+	r_config_set_b (core->config, "r2ai.async.purge", true);
+	r_config_desc (core->config, "r2ai.async.purge", "Drop completed/errored/cancelled async tasks after status listings");
 	r_config_lock (core->config, true);
 
+	r2ai_async_init (state);
+
 	return load_r2airc (cps);
 }
 
 R_API bool r2ai_fini(RCorePluginSession *cps) {
 	RCore *core = cps->core;
+	R2AI_State *state = cps->data;
+	r2ai_async_fini (state);
 	r_config_lock (core->config, false);
 	r_config_rm (core->config, "r2ai.api");
 	r_config_rm (core->config, "r2ai.cmds");
 	r_config_rm (core->config, "r2ai.model");
+	r_config_rm (core->config, "r2ai.apitype");
+	r_config_rm (core->config, "r2ai.cacheck");
 	r_config_rm (core->config, "r2ai.prompt");
+	r_config_rm (core->config, "r2ai.markdown");
 	r_config_rm (core->config, "r2ai.stream");
 	r_config_rm (core->config, "r2ai.system");
 	r_config_rm (core->config, "r2ai.data");
 	r_config_rm (core->config, "r2ai.data.path");
 	r_config_rm (core->config, "r2ai.data.nth");
+	r_config_rm (core->config, "r2ai.history");
 	r_config_rm (core->config, "r2ai.auto.reset_on_query");
 	r_config_rm (core->config, "r2ai.http.timeout");
 	r_config_rm (core->config, "r2ai.http.max_retries");
@@ -652,14 +843,22 @@ R_API bool r2ai_fini(RCorePluginSession *cps) {
 	r_config_rm (core->config, "r2ai.http.use_files");
 	r_config_rm (core->config, "r2ai.auto.slim");
 	r_config_rm (core->config, "r2ai.debug");
+	r_config_rm (core->config, "r2ai.async");
+	r_config_rm (core->config, "r2ai.async.purge");
 	r_config_lock (core->config, true);
 
-	R2AI_State *state = cps->data;
 	r2ai_conversation_free (state);
 
 	if (state) {
+		r2ai_tools_free (state->tools);
+		state->tools = NULL;
 		r_vdb_free (state->db);
 		state->db = NULL;
+		free (state->prompt_auto);
+		state->prompt_auto = NULL;
+		free (state->vertex_token);
+		state->vertex_token = NULL;
+		R_FREE (state->cache_prefix);
 		free (state);
 		cps->data = NULL;
 	}
@@ -675,7 +874,7 @@ static bool r_cmd_r2ai_client(RCorePluginSession *cps, const char *input) {
 }
 
 // PLUGIN Definition Info
-RCorePlugin r_core_plugin_r2ai_client = {
+RCorePlugin r_core_plugin_r2ai = {
 	.meta = {
 		.name = "r2ai",
 		.desc = "r2ai plugin in plain C",
@@ -691,7 +890,7 @@ RCorePlugin r_core_plugin_r2ai_client = {
 #ifndef R2_PLUGIN_INCORE
 R_API RLibStruct radare_plugin = {
 	.type = R_LIB_TYPE_CORE,
-	.data = &r_core_plugin_r2ai_client,
+	.data = &r_core_plugin_r2ai,
 	.abiversion = R2_ABIVERSION,
 	.pkgname = "r2ai",
 	.version = R2_VERSION

src/r2ai.h

@@ -2,7 +2,6 @@
 #define R2AI_H
 
 #define _GNU_SOURCE
-#define _POSIX_C_SOURCE 200809L
 
 #include <time.h>
 #include <stdint.h>
@@ -15,18 +14,24 @@
 #include "r_vdb.h"
 #include "markdown.h"
 
-#define R2AI_VERSION "1.3.4"
-
-#if 0
-#define R2AI_DEFAULT_MODEL "gpt-5-mini"
-#define R2AI_DEFAULT_PROVIDER "openai"
-#else
-// #define R2AI_DEFAULT_MODEL "gpt-oss:20b"
-// #define R2AI_DEFAULT_MODEL "gemma3:12b"
-#define R2AI_DEFAULT_MODEL "mannix/jan-nano"
-#define R2AI_DEFAULT_PROVIDER "ollama"
+#ifndef R_LIB_ABIVERSION
+#define R_LIB_ABIVERSION R2_ABIVERSION
 #endif
 
+#if R_LIB_ABIVERSION >= 125
+#define r2ai_cmd_help(core,msg) r_cons_cmd_help ((core)->cons, msg)
+#else
+#define r2ai_cmd_help(core,msg) r_core_cmd_help (core, msg)
+#endif
+
+#if R2_ABIVERSION < 83
+#define r_core_call(x,y) r_core_cmd_call(x,y)
+#endif
+
+#define R2AI_VERSION "1.5.4"
+#define R2AI_DEFAULT_MODEL "gpt-oss:20b"
+#define R2AI_DEFAULT_PROVIDER "ollama"
+
 #define R2AI_DEFAULT_VECTORS 16
 
 // HTTP configuration structure
@@ -41,7 +46,7 @@ typedef struct {
 	R2AI_HttpConfig config;
 	const char *url;
 	const char *data;
-	const char * const *headers;
+	const char **headers;
 } HTTPRequest;
 
 // HTTP response structure
@@ -55,8 +60,6 @@ typedef struct {
 #error Your radare2 is too old
 #endif
 
-
-
 // Tool definition structure
 typedef struct {
 	char *name;
@@ -64,6 +67,12 @@ typedef struct {
 	char *parameters; // JSON string of parameters/input_schema
 } R2AI_Tool;
 
+typedef struct {
+	char *output;
+	char *edited_command;
+	char *comment;
+} R2AI_ToolResult;
+
 // Tools array management (now just RList<R2AI_Tool *>)
 
 typedef struct {
@@ -72,10 +81,6 @@ typedef struct {
 	const char *id;
 } R2AI_ToolCall;
 
-/**
- * Free an R2AI_ToolCall instance
- */
-R_API void r2ai_tool_call_free(R2AI_ToolCall *tc);
 
 typedef struct {
 	char *type;
@@ -95,11 +100,6 @@ typedef struct {
  */
 R_API RList *r2ai_content_blocks_new(void);
 
-/**
- * Free a content blocks RList
- */
-R_API void r2ai_content_blocks_free(RList *cb);
-
 typedef struct {
 	char *role;
 	char *content;
@@ -146,6 +146,7 @@ typedef struct {
 	int max_tokens;
 	int thinking_tokens;
 	float temperature;
+	bool deterministic; // greedy sampling to remove randomness from responses
 	bool dorag;
 	char **error;
 } R2AIArgs;
@@ -162,131 +163,50 @@ typedef struct {
 	time_t total_start_time;
 } R2AIStats;
 
+// Forward declaration for async task queue (defined in r2ai_priv.h)
+struct r2ai_task_queue_t;
+
 // Main state structure to hold all global state
 typedef struct r2ai_state_t {
 	RList *conversation; // Global conversation messages (from messages.c)
 	R2AIStats stats; // Global stats (from auto.c)
 	RList *tools; // Global tools instance (from tools.c)
-	RMarkdownTheme current_theme; // Global theme (from markdown.c)
-	bool theme_initialized; // Global theme flag (from markdown.c)
+	RMarkdown markdown; // Markdown rendering context
 	void *help_msg; // Global help message (from r2ai.c)
 	RVdb *db; // Vector database for embeddings
+	struct r2ai_task_queue_t *async; // Async task queue (from async.c)
+	char *prompt_auto; // Auto-mode system prompt
+	char *vertex_token; // Cached Vertex AI OAuth2 token
+	ut64 vertex_token_expiry; // Monotonic microseconds when the cached token expires
+	char *cache_prefix; // Last chat request prefix used by r2ai.cacheck
 } R2AI_State;
 
-/**
- * Initialize a new empty messages array
- */
-R_API RList *r2ai_msgs_new(void);
-
-/**
- * Initialize the conversation container (call during plugin init)
- */
+// conversation
 R_API void r2ai_conversation_init(R2AI_State *state);
-
-/**
- * Get the conversation instance (returns NULL if not initialized)
- */
 R_API RList *r2ai_conversation_get(R2AI_State *state);
-
-/**
- * Clear all messages in a container without freeing the container
- */
-R_API void r2ai_msgs_clear(RList *msgs);
-
-/**
- * Add a message to the array
- * All strings are duplicated, so caller can free their copies
- */
-R_API bool r2ai_msgs_add(RList *msgs, const R2AI_Message *msg);
-
-/**
- * Add a tool call to the last message in the array
- * All strings are duplicated, so caller can free their copies
- */
-R_API bool r2ai_msgs_add_tool_call(RList *msgs, const R2AI_ToolCall *tc);
-
-/**
- * Parse a JSON response string and add the messages to the array
- * Returns true on success, false on failure
- */
-R_API bool r2ai_msgs_from_response(RList *msgs, const char *json_str);
-
-/**
- * Parse a RJson object directly and add the messages to the array
- * Returns true on success, false on failure
- */
-R_API bool r2ai_msgs_from_json(RList *msgs, const RJson *json);
-
-/**
- * Convert messages array to JSON string
- * Caller must free the returned string
- */
-R_API char *r2ai_msgs_to_json(const RList *msgs);
-
-/**
- * Convert messages array to Anthropic format JSON string
- * Caller must free the returned string
- */
-R_API char *r2ai_msgs_to_anthropic_json(const RList *msgs);
-
-/**
- * Free a messages array and all associated data
- */
-R_API void r2ai_msgs_free(RList *msgs);
-
-/**
- * Free the conversation (call during plugin unload)
- */
 R_API void r2ai_conversation_free(R2AI_State *state);
 
-/**
- * Free a R2AI_Message structure
- */
+// messages
+R_API RList *r2ai_msgs_new(void);
+R_API void r2ai_msgs_clear(RList *msgs);
+R_API bool r2ai_msgs_add(RList *msgs, const R2AI_Message *msg);
+R_API char *r2ai_msgs_to_json(const RList *msgs, bool raw_tool_args);
+R_API char *r2ai_msgs_to_anthropic_json(const RList *msgs);
+R_API void r2ai_msgs_free(RList *msgs);
 R_API void r2ai_message_free(R2AI_Message *msg);
-
-/**
- * Delete the last N messages from the message array
- * If n <= 0, defaults to deleting just the last message
- */
+R_IPI void r2ai_chat_response_free(R2AI_ChatResponse *res);
+R_API void r2ai_message_fini(R2AI_Message *msg);
 R_API void r2ai_delete_last_messages(RList *messages, int n);
 
-/**
- * Get the tools instance from state
- * Returns a pointer to the tools RList
- */
+// tools
 R_API RList *r2ai_get_tools(RCore *core, R2AI_State *state);
-
-/**
- * Parse OpenAI format tools JSON into internal tools RList
- * Caller must free the result with r2ai_tools_free
- */
 R_API RList *r2ai_tools_parse(const char *tools_json);
-
-/**
- * Convert tools RList to OpenAI format JSON
- * Caller must free the returned string
- */
 R_API char *r2ai_tools_to_openai_json(const RList *tools);
-
-/**
- * Convert tools RList to Anthropic format JSON
- * Caller must free the returned string
- */
 R_API char *r2ai_tools_to_anthropic_json(const RList *tools);
-
-/**
- * Free a tools RList and all associated data
- */
 R_API void r2ai_tools_free(RList *tools);
-
-/**
- * Execute a tool and return the output 
- * The command to execute is within args. The user may interactively
- * modify this command, hence edited_command contains the real
- * command which was run
- * Caller must free edited_command
- */
-R_API char *execute_tool(RCore *core, const char *tool_name, const char *args, char **edited_command, char **comment_out);
+R_API void r2ai_tool_result_fini(R2AI_ToolResult *result);
+R_API void r2ai_tool_call_free(R2AI_ToolCall *tc);
+R_API R2AI_ToolResult execute_tool(RCorePluginSession *cps, const char *tool_name, const char *args);
 
 /**
  * Execute r_core_cmd_str with slim mode if enabled
@@ -306,7 +226,7 @@ R_API char *r2ai_cmdstr(RCore *core, const char *cmd);
  * @param use_files Use files instead of arguments
  * @return Response body as string (must be freed by caller) or NULL on error
  */
-R_API char *r2ai_http_post(RCore *core, const char *url, const char *headers[], const char *data, int *code, int *rlen);
+R_API char *r2ai_http_post(RCore *core, const char *url, const char **headers, const char *data, int *code, int *rlen);
 
 /**
  * Send an HTTP GET request
@@ -317,19 +237,20 @@ R_API char *r2ai_http_post(RCore *core, const char *url, const char *headers[],
  * @param rlen Pointer to store the response length
  * @return Response body as string (must be freed by caller) or NULL on error
  */
-R_API char *r2ai_http_get(RCore *core, const char *url, const char *headers[], int *code, int *rlen);
+R_API char *r2ai_http_get(RCore *core, const char *url, const char **headers, int *code, int *rlen);
 
 /**
  * Get the base URL for a given provider
  *
  * @param core RCore instance for configuration
  * @param provider The provider name (e.g., "openai", "anthropic", etc.)
- * @return Base URL for the provider, or NULL if unknown
+ * @return Newly allocated base URL for the provider, or NULL if unknown
  */
-R_IPI const char *r2ai_get_provider_url(RCore *core, const char *provider);
+R_IPI char *r2ai_get_provider_url(RCore *core, const char *provider);
 
 // anthropic
 R_IPI R2AI_ChatResponse *r2ai_anthropic(RCorePluginSession *cps, R2AIArgs args);
+R_IPI R2AI_ChatResponse *r2ai_anthropic_parse_response(char *json, char **error);
 
 // openai
 R_IPI R2AI_ChatResponse *r2ai_openai(RCorePluginSession *cps, R2AIArgs args);
@@ -339,6 +260,8 @@ R_IPI R2AI_ChatResponse *r2ai_gemini(RCorePluginSession *cps, R2AIArgs args);
 
 // auto mode
 R_IPI void cmd_r2ai_a(RCorePluginSession *cps, const char *user_query);
+R_IPI char *r2ai_auto_system_prompt(RCorePluginSession *cps);
+R_IPI void r2ai_print_response(RCore *core, const char *text);
 // R_API char *r2ai(RCore *core, R2AI_State *state, R2AIArgs args);
 R_API char *r2ai(RCorePluginSession *cps, R2AIArgs args);
 R_API bool r2ai_init(RCorePluginSession *cps);
@@ -346,10 +269,9 @@ R_API bool r2ai_fini(RCorePluginSession *cps);
 
 R_IPI R2AI_ChatResponse *r2ai_llmcall(RCorePluginSession *cps, R2AIArgs args);
 
-R2AI_ChatResponse *r2ai_rawtools_llmcall(RCorePluginSession *cps, R2AIArgs args);
 
 R_IPI void cmd_r2ai_c(RCorePluginSession *cps);
-R_IPI void cmd_r2ai_logs(RCorePluginSession *cps);
+R_IPI void cmd_r2ai_logs(RCorePluginSession *cps, const char *flags);
 R_IPI void cmd_r2ai_lr(RCorePluginSession *cps);
 
 /**
@@ -357,16 +279,12 @@ R_IPI void cmd_r2ai_lr(RCorePluginSession *cps);
  */
 R_API void process_messages(RCorePluginSession *cps, RList *messages, const char *system_prompt, int n_run);
 
-/**
- * Helper function to convert RJson to string
- */
+// json helpers that must be moved into r2
 R_API char *r_json_to_string(const RJson *json);
 
-/**
- * Helper function to convert RJson to PJ
- */
-R_API PJ *r_json_to_pj(const RJson *json, PJ *existing_pj);
+// commands
 R_API void r2ai_cmd_q(RCorePluginSession *cps, const char *input);
+R_API void r2ai_cmd_qj(RCorePluginSession *cps, const char *input);
 R_API char *find_prompt_file(RList *search_dirs, const char *name);
 R_API R2AIPrompt *parse_prompt_file(const char *filepath);
 R_API void r2aiprompt_free(R2AIPrompt *prompt);

src/r2ai_priv.h

@@ -2,14 +2,77 @@
 #define INCLUDE_R2AI_PRIV__H
 
 #include "r2ai.h"
+#include <r_th.h>
 
 typedef enum {
 	R2AI_API_OPENAI_COMPATIBLE,
 	R2AI_API_ANTHROPIC,
 	R2AI_API_GEMINI,
-	R2AI_API_OLLAMA
+	R2AI_API_OLLAMA,
+	R2AI_API_VERTEX_GEMINI,
+	R2AI_API_VERTEX_ANTHROPIC
 } R2AI_API_Type;
 
+/* Async task subsystem (see ASYNC.md) */
+typedef enum {
+	R2AI_TASK_PENDING = 0,
+	R2AI_TASK_RUNNING,
+	R2AI_TASK_WAIT_APPROVE,
+	R2AI_TASK_WAIT_INPUT,
+	R2AI_TASK_COMPLETE,
+	R2AI_TASK_ERROR,
+	R2AI_TASK_CANCELLED,
+} R2AITaskState;
+
+typedef enum {
+	R2AI_TASK_QUERY = 0,
+	R2AI_TASK_AUTO,
+} R2AITaskKind;
+
+typedef struct r2ai_task_t {
+	int id;
+	R2AITaskKind kind;
+	R2AITaskState state;
+	char *title;
+	char *query;
+	char *system_prompt;
+	char *model;
+	char *provider;
+	RList *messages; /* R2AI_Message * - owned conversation */
+	RList *tools; /* R2AI_Tool * - snapshot taken on the main thread, items not owned */
+	int max_runs;
+	RStrBuf *output; /* flushed by -si */
+	char *error;
+	char *pending_tool_name;
+	char *pending_tool_args;
+	char *pending_tool_call_id;
+	char *tool_result; /* main -> worker handoff */
+	bool cancel_req;
+	RThread *thread;
+	RThreadLock *lock;
+	RThreadSemaphore *gate;
+	RCorePluginSession *cps;
+	RCore *llm_core; /* shallow core view with an owned request config snapshot */
+	time_t created;
+	time_t started;
+	time_t finished;
+	int steps;
+} R2AITask;
+
+typedef struct r2ai_task_queue_t {
+	RList *tasks;
+	RThreadLock *lock;
+	int next_id;
+} R2AITaskQueue;
+
+R_IPI void r2ai_async_init(R2AI_State *state);
+R_IPI void r2ai_async_fini(R2AI_State *state);
+R_IPI int r2ai_async_query(RCorePluginSession *cps,
+	const char *title, const char *query, const char *sysp);
+R_IPI int r2ai_async_auto(RCorePluginSession *cps,
+	const char *title, const char *query, const char *sysp);
+R_IPI void r2ai_async_cmd(RCorePluginSession *cps, const char *input);
+
 typedef struct {
 	const char *name;
 	const char *url;
@@ -19,8 +82,24 @@ typedef struct {
 } R2AIProvider;
 
 R_API void cmd_r2ai(RCorePluginSession *cps, const char *input);
+static inline char *r2ai_cons_editor(RCons *cons, const char *file, const char *str) {
+#if R2_ABIVERSION >= 160
+	return r_cons_editor (cons, file, str, 0, NULL);
+#elif R2_ABIVERSION >= 120
+	return r_cons_editor (cons, file, str, NULL);
+#else
+	return r_cons_editor (cons, file, str);
+#endif
+}
 R_IPI const R2AIProvider *r2ai_get_provider(const char *name);
-R_IPI const char *r2ai_get_provider_url(RCore *core, const char *provider);
+R_IPI const char *r2ai_system_prompt(RCore *core, const char *sysp);
+R_IPI R2AI_ChatResponse *r2ai_send(RCorePluginSession *cps, const R2AIProvider *p, R2AIArgs args);
+R_IPI char *r2ai_post(RCore *core, const char *name, const char *url, const char **headers, const char *data, int *code, char **error);
+R_IPI char *r2ai_gemini_request(const R2AIArgs *args);
+R_IPI R2AI_ChatResponse *r2ai_gemini_parse(char *json, char **error);
+R_IPI char *r2ai_anthropic_request(const R2AIArgs *args, bool vertex);
+R_IPI R2AI_ChatResponse *r2ai_rawtools_llmcall(RCorePluginSession *cps, const R2AIProvider *p, R2AIArgs args);
+R_IPI char *r2ai_get_provider_url(RCore *core, const char *provider);
 R_IPI RList *r2ai_fetch_available_models(RCore *core, const char *provider);
 R_IPI void r2ai_list_providers(RCore *core, RStrBuf *sb);
 R_IPI void r2ai_refresh_embeddings(RCorePluginSession *cps);
@@ -28,4 +107,18 @@ R_API char *r2ai_apikeys_path(bool *exists);
 R_API void r2ai_apikeys_edit(RCorePluginSession *cps);
 R_API char *r2ai_apikeys_get(const char *provider);
 
+/* vertex.c */
+R_IPI const char *r2ai_vertex_get_token(R2AI_State *state);
+R_IPI R2AI_ChatResponse *r2ai_vertex_gemini(RCorePluginSession *cps, R2AIArgs args);
+R_IPI R2AI_ChatResponse *r2ai_vertex_anthropic(RCorePluginSession *cps, R2AIArgs args);
+
+/* claw.c - SOUL.md / IDENTITY.md personality files */
+#define R2AI_CLAW_HINT "use 'r2ai -ide' to edit or 'r2ai -id-' to delete"
+R_API bool r2ai_claw_exists(void);
+R_API char *r2ai_claw_system_prompt(const char *base);
+R_API void r2ai_claw_create(RCorePluginSession *cps, const char *user_extra);
+R_API void r2ai_claw_show(RCorePluginSession *cps);
+R_API void r2ai_claw_edit(RCorePluginSession *cps);
+R_API void r2ai_claw_delete(RCorePluginSession *cps);
+
 #endif

src/rawtools.c

@@ -1,4 +1,4 @@
-/* r2ai - Copyright 2023-2025 pancake */
+/* r2ai - Copyright 2023-2026 pancake */
 
 #define R_LOG_ORIGIN "rawtools"
 #include "r2ai.h"
@@ -76,17 +76,11 @@ static bool parse_raw_tool_call(const char *response, char **tool_name, char **t
 }
 
 // Function to handle rawtools mode in LLM call
-R2AI_ChatResponse *r2ai_rawtools_llmcall(RCorePluginSession *cps, R2AIArgs args) {
-	if (!cps) {
-		return NULL;
+R_IPI R2AI_ChatResponse *r2ai_rawtools_llmcall(RCorePluginSession *cps, const R2AIProvider *p, R2AIArgs args) {
+	const char *user_system = args.system_prompt;
+	if (R_STR_ISEMPTY (user_system)) {
+		user_system = "You are a reverse engineer using radare2. The binary is loaded. Use r2cmd tool for analysis.";
 	}
-	RCore *core = cps->core;
-
-	// Modify the system prompt to include rawtools instructions only if no tool messages
-	const char *original_system_prompt = args.system_prompt;
-
-	// For rawtools, use a shorter system prompt to avoid token limits
-	const char *short_system_prompt = "You are a reverse engineer using radare2. The binary is loaded. Use r2cmd tool for analysis.";
 
 	// Check for initial command output in messages
 	char *init_output = NULL;
@@ -103,33 +97,18 @@ R2AI_ChatResponse *r2ai_rawtools_llmcall(RCorePluginSession *cps, R2AIArgs args)
 		}
 	}
 
-	char *enhanced_system_prompt = NULL;
-	if (init_output) {
-		enhanced_system_prompt = r_str_newf ("%s\n\n%s\n\n%s", short_system_prompt, init_output, RAWTOOLS_PROMPT);
-	} else {
-		enhanced_system_prompt = r_str_newf ("%s\n\n%s", short_system_prompt, RAWTOOLS_PROMPT);
-	}
+	char *base_prompt = init_output
+		? r_str_newf ("%s\n\n%s\n\n%s", user_system, init_output, RAWTOOLS_PROMPT)
+		: r_str_newf ("%s\n\n%s", user_system, RAWTOOLS_PROMPT);
 	free (init_output);
-	// Temporarily modify args to use enhanced prompt and no tools (since we're using prompt engineering)
+	char *enhanced_system_prompt = r2ai_claw_system_prompt (base_prompt);
+	free (base_prompt);
+
 	R2AIArgs rawtools_args = args;
 	rawtools_args.system_prompt = enhanced_system_prompt;
-	rawtools_args.tools = NULL; // Disable native tool calling
-
-	// Make the LLM call directly to provider
-	const char *provider = rawtools_args.provider? rawtools_args.provider: r_config_get (core->config, "r2ai.api");
-	if (!provider) {
-		R_LOG_ERROR ("No provider defined");
-		return NULL;
-	}
-
-	R2AI_ChatResponse *response = NULL;
-	const R2AIProvider *p = r2ai_get_provider (provider);
-	if (p && p->api_type == R2AI_API_ANTHROPIC) {
-		response = r2ai_anthropic (cps, rawtools_args);
-	} else {
-		response = r2ai_openai (cps, rawtools_args);
-	}
+	rawtools_args.tools = NULL;
 
+	R2AI_ChatResponse *response = r2ai_send (cps, p, rawtools_args);
 	free (enhanced_system_prompt);
 
 	if (!response || !response->message || !response->message->content) {
@@ -138,14 +117,12 @@ R2AI_ChatResponse *r2ai_rawtools_llmcall(RCorePluginSession *cps, R2AIArgs args)
 
 	R_LOG_DEBUG ("[RAWTOOLS] Raw model response: %s", response->message->content);
 
-	// Check if the response contains a raw tool call
 	char *tool_name = NULL;
 	char *tool_args = NULL;
 
 	if (parse_raw_tool_call (response->message->content, &tool_name, &tool_args)) {
 		R_LOG_DEBUG ("Raw tool call detected: %s with args: %s", tool_name, tool_args);
 
-		// Find the tool in our tools list
 		RList *tools = r2ai_get_tools (cps->core, cps->data);
 		bool tool_found = false;
 
@@ -160,67 +137,55 @@ R2AI_ChatResponse *r2ai_rawtools_llmcall(RCorePluginSession *cps, R2AIArgs args)
 			}
 		}
 
+		R2AI_Message *modified_message = R_NEW0 (R2AI_Message);
+		if (!modified_message) {
+			free (tool_name);
+			free (tool_args);
+			return response;
+		}
+		modified_message->role = strdup ("assistant");
+
 		if (tool_found) {
-			// Modify the existing response to include tool call information
-			R2AI_Message *modified_message = R_NEW0 (R2AI_Message);
-			modified_message->role = strdup ("assistant");
-			// Remove the TOOL line from content, or set to empty if that's all there is
 			char *content = strdup (response->message->content? response->message->content: "");
-			char *tool_line = strstr (content, "TOOL: ");
-			if (tool_line) {
-				// Find the end of the line
-				char *line_end = strchr (tool_line, '\n');
-				if (line_end) {
-					// Remove the TOOL line
-					memmove (tool_line, line_end + 1, strlen (line_end + 1) + 1);
-				} else {
-					// The TOOL line is the entire content
-					*tool_line = '\0';
+			if (content) {
+				char *tool_line = strstr (content, "TOOL: ");
+				if (tool_line) {
+					char *line_end = strchr (tool_line, '\n');
+					if (line_end) {
+						memmove (tool_line, line_end + 1, strlen (line_end + 1) + 1);
+					} else {
+						*tool_line = '\0';
+					}
 				}
+				modified_message->content = content;
 			}
-			modified_message->content = content;
 
-			// Set up proper tool call structure
 			modified_message->tool_calls = r_list_new ();
 			if (modified_message->tool_calls) {
 				modified_message->tool_calls->free = (RListFree)r2ai_tool_call_free;
 				R2AI_ToolCall *tc = R_NEW0 (R2AI_ToolCall);
-				tc->name = tool_name;
-				tc->arguments = tool_args;
-				char id_buf[32];
-				snprintf (id_buf, sizeof (id_buf), "rawtool_%d", (int)time (NULL));
-				tc->id = strdup (id_buf);
-				r_list_append (modified_message->tool_calls, tc);
+				if (tc) {
+					char id_buf[32];
+					snprintf (id_buf, sizeof (id_buf), "rawtool_%d", (int)time (NULL));
+					tc->name = tool_name;
+					tc->arguments = tool_args;
+					tc->id = strdup (id_buf);
+					r_list_append (modified_message->tool_calls, tc);
+					tool_name = NULL;
+					tool_args = NULL;
+				}
 			}
-			tool_name = NULL;
-			tool_args = NULL;
-
-			// Replace the response message
-			if (response->message) {
-				r2ai_message_free ((R2AI_Message *)response->message);
-			}
-			*(R2AI_Message **)&response->message = modified_message;
 		} else {
-			// Unknown tool, modify content to indicate error
-			R2AI_Message *modified_message = R_NEW0 (R2AI_Message);
-			modified_message->role = strdup ("assistant");
-			size_t result_len = strlen ("Unknown tool: ") + strlen (tool_name) + 1;
-			modified_message->content = malloc (result_len);
-			if (modified_message->content) {
-				snprintf (modified_message->content, result_len, "Unknown tool: %s", tool_name);
-			}
-
-			// Replace the response message
-			if (response->message) {
-				r2ai_message_free ((R2AI_Message *)response->message);
-			}
-			*(R2AI_Message **)&response->message = modified_message;
-			free (tool_name);
-			free (tool_args);
-			tool_name = NULL;
-			tool_args = NULL;
+			char *errmsg = r_str_newf ("Unknown tool: %s", tool_name);
+			modified_message->content = errmsg;
 		}
 
+		if (response->message) {
+			r2ai_message_free ((R2AI_Message *)response->message);
+		}
+		*(R2AI_Message **)&response->message = modified_message;
+		free (tool_name);
+		free (tool_args);
 		return response;
 	}
 
@@ -228,54 +193,26 @@ R2AI_ChatResponse *r2ai_rawtools_llmcall(RCorePluginSession *cps, R2AIArgs args)
 	free (tool_name);
 	free (tool_args);
 
-	// If the response has no content, try again with normal mode (without rawtools prompt)
 	if (!response->message || !response->message->content || !*response->message->content) {
 		R_LOG_DEBUG ("No tool call found and no content, falling back to normal mode");
 
-		// Free the current response
-		if (response->message) {
-			r2ai_message_free ((R2AI_Message *)response->message);
-		}
-		free (response);
+		r2ai_chat_response_free (response);
 
-		// Modify args to use original system prompt without rawtools enhancement
-		R2AIArgs fallback_args = args;
-		fallback_args.system_prompt = original_system_prompt;
-		fallback_args.tools = args.tools; // Keep tools for fallback
-
-		// Call provider directly
-		const char *provider = fallback_args.provider? fallback_args.provider: r_config_get (core->config, "r2ai.api");
-		if (!provider) {
-			R_LOG_ERROR ("No provider defined");
-			return NULL;
-		}
-
-		R2AI_ChatResponse *fallback_response = NULL;
-		const R2AIProvider *p = r2ai_get_provider (provider);
-		if (p && p->api_type == R2AI_API_ANTHROPIC) {
-			fallback_response = r2ai_anthropic (cps, fallback_args);
-		} else {
-			fallback_response = r2ai_openai (cps, fallback_args);
-		}
+		R2AI_ChatResponse *fallback_response = r2ai_send (cps, p, args);
 
 		if (fallback_response && fallback_response->message && fallback_response->message->content && *fallback_response->message->content) {
 			return fallback_response;
 		}
 
-		// If still no content, create a response with warning
-		if (fallback_response) {
-			if (fallback_response->message) {
-				r2ai_message_free ((R2AI_Message *)fallback_response->message);
-			}
-			free (fallback_response);
-		}
+		r2ai_chat_response_free (fallback_response);
 
-		// Create a warning response
 		R2AI_ChatResponse *warning_response = R_NEW0 (R2AI_ChatResponse);
 		R2AI_Message *msg = R_NEW0 (R2AI_Message);
-		msg->role = strdup ("assistant");
-		msg->content = strdup ("Warning: LLM provided no response content");
-		*((R2AI_Message **)&warning_response->message) = msg;
+		if (msg) {
+			msg->role = strdup ("assistant");
+			msg->content = strdup ("Warning: LLM provided no response content");
+			*((R2AI_Message **)&warning_response->message) = msg;
+		}
 		R_LOG_WARN ("LLM provided no response content");
 		return warning_response;
 	}

src/test_vertex.sh

@@ -0,0 +1,40 @@
+#!/bin/sh
+# Offline smoke tests for the Vertex AI provider (no network required)
+PASS=0
+FAIL=0
+
+# Use empty rc to avoid user config interference
+R2ENV="R2_RCFILE=/dev/null"
+
+check() {
+	desc="$1"
+	expected="$2"
+	actual="$3"
+	if echo "$actual" | grep -q "$expected"; then
+		PASS=$((PASS+1))
+	else
+		FAIL=$((FAIL+1))
+		printf "FAIL: %s\n  expected: %s\n  got: %s\n" "$desc" "$expected" "$actual"
+	fi
+}
+
+# Provider enumeration
+providers=$(env $R2ENV r2 -qc 'r2ai -e r2ai.api=?' -- 2>/dev/null)
+check "vertex in provider list" "vertex" "$providers"
+check "vertex-anthropic in provider list" "vertex-anthropic" "$providers"
+
+# Config defaults
+region=$(env $R2ENV r2 -qc 'r2ai -e r2ai.vertex.region' -- 2>/dev/null)
+check "vertex.region default is us-central1" "us-central1" "$region"
+
+# Config roundtrip
+rt=$(env $R2ENV r2 -qc 'e r2ai.vertex.project=test-proj; e r2ai.vertex.region=europe-west4; r2ai -e r2ai.vertex.project; r2ai -e r2ai.vertex.region' -- 2>/dev/null)
+check "vertex config roundtrip project" "test-proj" "$rt"
+check "vertex config roundtrip region" "europe-west4" "$rt"
+
+# Error when project not set
+err=$(env $R2ENV r2 -qc 'e r2ai.api=vertex; e r2ai.vertex.project=; r2ai hello' -- 2>&1)
+check "error when project not set" "vertex.project" "$err"
+
+printf "\n%d passed, %d failed\n" "$PASS" "$FAIL"
+[ "$FAIL" -eq 0 ]

src/tools.c

@@ -1,6 +1,6 @@
-/* Copyright r2ai - 2023-2025 - pancake */
+/* Copyright r2ai - 2023-2026 - pancake */
 
-#include "r2ai.h"
+#include "r2ai_priv.h"
 
 R_API char *strip_command_comment(const char *input, char **comment_out) {
 	const char *hash_pos = strchr (input, '#');
@@ -67,11 +67,15 @@ R_API RList *r2ai_get_tools(RCore *core, R2AI_State *state) {
 	}
 	if (!state->tools) {
 		state->tools = r_list_newf (NULL);
-		r_list_append (state->tools, (void *)&r2cmd_tool);
-		// Only add execute_js tool if enabled in config
-		if (r_config_get_b (core->config, "r2ai.auto.usejs")) {
-			r_list_append (state->tools, (void *)&qjs_tool);
-		}
+	}
+	if (!state->tools) {
+		return NULL;
+	}
+	r_list_purge (state->tools);
+	r_list_append (state->tools, (void *)&r2cmd_tool);
+	// Reflect the current config instead of caching the first-seen value.
+	if (r_config_get_b (core->config, "r2ai.auto.usejs")) {
+		r_list_append (state->tools, (void *)&qjs_tool);
 	}
 	return state->tools;
 }
@@ -266,72 +270,75 @@ static char *compose_command_with_comment(const char *command, const char *comme
 	return r_str_newf ("%s # %s", command, comment);
 }
 
-R_API char *r2ai_r2cmd(RCore *core, RJson *args, bool verbose, char **edited_command, char **comment_out) {
+R_API void r2ai_tool_result_fini(R2AI_ToolResult *result) {
+	if (!result) {
+		return;
+	}
+	free (result->output);
+	free (result->edited_command);
+	free (result->comment);
+	memset (result, 0, sizeof (*result));
+}
+
+// let the user review or edit the text before running it, returns NULL when cancelled
+static char *confirm(RCore *core, const char *text) {
+	RCons *cons = core->cons;
+	if (strchr (text, '\n')) {
+		char *edited = r2ai_cons_editor (cons, NULL, text);
+		return edited? edited: strdup (text);
+	}
+	r_cons_newline (cons);
+	r_cons_readpush (cons, text, strlen (text));
+	r_cons_readpush (cons, "\x05", 1); // ctrl+e moves the cursor to the end
+	r_line_set_prompt (cons->line, "[r2ai]> ");
+	const char *line = r_line_readline (cons);
+	if (r_cons_is_breaked (cons) || R_STR_ISEMPTY (line)) {
+		return NULL;
+	}
+	return strdup (line);
+}
+
+static R2AI_ToolResult r2ai_r2cmd(RCore *core, RJson *args, bool verbose) {
+	R2AI_ToolResult result = { 0 };
 	if (!args) {
-		return strdup ("{ \"res\":\"Command is NULL\" }");
+		result.output = strdup ("{ \"res\":\"Command is NULL\" }");
+		return result;
 	}
 
 	const RJson *command_json = r_json_get (args, "command");
 	if (!command_json || !command_json->str_value) {
-		return strdup ("{ \"res\":\"No command in tool call arguments\" }");
-	}
-
-	if (edited_command) {
-		*edited_command = NULL;
+		result.output = strdup ("{ \"res\":\"No command in tool call arguments\" }");
+		return result;
 	}
 
 	char *comment = NULL;
 	char *command = strip_command_comment (command_json->str_value, &comment);
 	if (!command) {
 		free (comment);
-		return strdup ("{ \"res\":\"Command is NULL\" }");
+		result.output = strdup ("{ \"res\":\"Command is NULL\" }");
+		return result;
 	}
 
 	if (r_str_startswith (command, "r2 ")) {
 		free (command);
 		free (comment);
-		if (comment_out) {
-			*comment_out = NULL;
-		}
-		return strdup ("{ \"res\":\"You are already in r2!\" }");
+		result.output = strdup ("{ \"res\":\"You are already in r2!\" }");
+		return result;
 	}
 
-	bool ask_to_execute = r_config_get_b (core->config, "r2ai.auto.yolo") != true;
-	if (ask_to_execute) {
-		char *input_command = compose_command_with_comment (command, comment);
-		if (!input_command) {
+	if (!r_config_get_b (core->config, "r2ai.auto.yolo")) {
+		char *input = compose_command_with_comment (command, comment);
+		char *edited = confirm (core, input);
+		free (input);
+		if (!edited) {
 			free (command);
 			free (comment);
-			return strdup ("{ \"res\":\"Failed to prepare command\" }");
+			result.output = strdup ("R2AI_SIGINT");
+			return result;
 		}
-		bool is_multiline = strchr (input_command, '\n') != NULL;
-
-		if (is_multiline) {
-			r_cons_editor (core->cons, NULL, input_command);
-		} else {
-			r_cons_newline (core->cons);
-			r_cons_readpush (core->cons, input_command, strlen (input_command));
-			r_cons_readpush (core->cons, "\x05", 1);
-			r_line_set_prompt (core->cons->line, "[r2ai]> ");
-			const char *readline_result = r_line_readline (core->cons);
-			if (r_cons_is_breaked (core->cons) || R_STR_ISEMPTY (readline_result)) {
-				R_LOG_INFO ("Command execution cancelled %s", readline_result);
-				free (input_command);
-				free (command);
-				free (comment);
-				if (comment_out) {
-					*comment_out = NULL;
-				}
-				return strdup ("R2AI_SIGINT");
-			}
-			if (R_STR_ISNOTEMPTY (readline_result)) {
-				free (input_command);
-				input_command = strdup (readline_result);
-			}
-		}
-
 		char *new_comment = NULL;
-		char *new_command = strip_command_comment (input_command, &new_comment);
+		char *new_command = strip_command_comment (edited, &new_comment);
+		free (edited);
 		if (new_command) {
 			free (command);
 			command = new_command;
@@ -340,12 +347,9 @@ R_API char *r2ai_r2cmd(RCore *core, RJson *args, bool verbose, char **edited_com
 		} else {
 			free (new_comment);
 		}
-		free (input_command);
 	}
 
-	if (edited_command) {
-		*edited_command = strdup (command);
-	}
+	result.edited_command = strdup (command);
 	R_LOG_DEBUG ("Edited command: %s", command);
 
 	if (verbose) {
@@ -363,10 +367,9 @@ R_API char *r2ai_r2cmd(RCore *core, RJson *args, bool verbose, char **edited_com
 	if (!json_cmd) {
 		free (command);
 		free (comment);
-		if (comment_out) {
-			*comment_out = NULL;
-		}
-		return strdup ("{ \"res\":\"Failed to create JSON command\" }");
+		r2ai_tool_result_fini (&result);
+		result.output = strdup ("{ \"res\":\"Failed to create JSON command\" }");
+		return result;
 	}
 
 	char *cmd_output = r2ai_cmdstr (core, json_cmd);
@@ -375,19 +378,14 @@ R_API char *r2ai_r2cmd(RCore *core, RJson *args, bool verbose, char **edited_com
 
 	if (!cmd_output) {
 		free (comment);
-		if (comment_out) {
-			*comment_out = NULL;
-		}
-		return strdup ("{ \"res\":\"Command returned no output or failed\" }");
+		r2ai_tool_result_fini (&result);
+		result.output = strdup ("{ \"res\":\"Command returned no output or failed\" }");
+		return result;
 	}
 
-	if (comment_out) {
-		*comment_out = comment;
-	} else {
-		free (comment);
-	}
-
-	return cmd_output;
+	result.output = cmd_output;
+	result.comment = comment;
+	return result;
 }
 
 /**
@@ -416,93 +414,62 @@ R_API char *r2ai_cmdstr(RCore *core, const char *cmd) {
 }
 
 // qjs function implementation
-R_API char *r2ai_qjs(RCore *core, RJson *args, bool verbose, char **edited_script_out) {
+static R2AI_ToolResult r2ai_qjs(RCore *core, R2AI_State *state, RJson *args, bool verbose) {
+	R2AI_ToolResult result = { 0 };
 	if (!args) {
-		return strdup ("{ \"res\":\"Script is NULL\" }");
-	}
-
-	if (edited_script_out) {
-		*edited_script_out = NULL;
+		result.output = strdup ("{ \"res\":\"Script is NULL\" }");
+		return result;
 	}
 
 	const RJson *script_json = r_json_get (args, "script");
 	if (!script_json) {
-		return strdup ("{ \"res\":\"No script field found in arguments\" }");
+		result.output = strdup ("{ \"res\":\"No script field found in arguments\" }");
+		return result;
 	}
 
 	if (!script_json->str_value) {
-		return strdup ("{ \"res\":\"Script value is NULL or empty\" }");
+		result.output = strdup ("{ \"res\":\"Script value is NULL or empty\" }");
+		return result;
 	}
 
-	bool ask_to_execute = r_config_get_b (core->config, "r2ai.auto.yolo") != true;
-	const char *script = script_json->str_value;
-	char *edited_script = NULL;
-
-	if (ask_to_execute) {
-		// Check if script contains newlines to determine if it's multi-line
-		bool is_multiline = strchr (script, '\n') != NULL;
-
-		if (is_multiline) {
-			// Use editor for multi-line scripts
-			edited_script = strdup (script);
-			r_cons_editor (core->cons, NULL, edited_script);
-			script = edited_script;
-		} else {
-			// For single-line scripts, push the script to input buffer
-
-			r_cons_readpush (core->cons, script, strlen (script));
-			r_cons_readpush (core->cons, "\x05", 1); // Ctrl+E - move to end
-			r_line_set_prompt (core->cons->line, "[r2ai]> ");
-			const char *readline_result = r_line_readline (core->cons);
-
-			// Check if interrupted or ESC pressed (readline_result is NULL or empty)
-			if (r_cons_is_breaked (core->cons) || R_STR_ISEMPTY (readline_result)) {
-				free (edited_script); // Free if already allocated
-				return strdup ("R2AI_SIGINT");
-			}
-
-			// Process the result
-			if (readline_result && *readline_result) {
-				edited_script = strdup (readline_result);
-				script = edited_script;
-			} else {
-				// If user just pressed enter, keep the original script
-				edited_script = strdup (script);
-				script = edited_script;
-			}
+	char *script = strdup (script_json->str_value);
+	if (!r_config_get_b (core->config, "r2ai.auto.yolo")) {
+		char *edited = confirm (core, script);
+		free (script);
+		if (!edited) {
+			result.output = strdup ("R2AI_SIGINT");
+			return result;
 		}
+		script = edited;
 	}
 
-	if (edited_script_out) {
-		char *dup_script = strdup (script);
-		if (dup_script) {
-			*edited_script_out = dup_script;
-		}
-	}
+	result.edited_command = strdup (script);
 
 	if (verbose) {
 		char *print_script = r_str_newf ("\n```js\n%s\n```", script);
-		char *print_script_rendered = r2ai_markdown (print_script);
+		char *print_script_rendered = r2ai_markdown (state? &state->markdown: NULL, print_script);
 		r_cons_printf (core->cons, "%s\n", print_script_rendered);
 		r_cons_flush (core->cons);
 		free (print_script);
 		free (print_script_rendered);
 	}
 	char *payload = r_str_newf ("var console = { log:r2log, warn:r2log, info:r2log, error:r2log, debug:r2log };%s", script);
-
-	// Free edited_script after we're done using it
-	free (edited_script);
-
-	R_LOG_DEBUG ("Payload length: %d", (int)strlen (payload));
+	free (script);
 
 	if (!payload) {
-		return strdup ("{ \"res\":\"Failed to create script payload\" }");
+		r2ai_tool_result_fini (&result);
+		result.output = strdup ("{ \"res\":\"Failed to create script payload\" }");
+		return result;
 	}
 
+	R_LOG_DEBUG ("Payload length: %d", (int)strlen (payload));
+
 	char *payload_base64 = r_base64_encode_dyn ((const ut8 *)payload, strlen (payload));
 	if (!payload_base64) {
 		free (payload);
-		return strdup ("{ \"res\":\"Failed to encode script\" }");
+		r2ai_tool_result_fini (&result);
+		result.output = strdup ("{ \"res\":\"Failed to encode script\" }");
+		return result;
 	}
 
 	char *cmd = r_str_newf ("#!qjs -e base64:%s", payload_base64);
@@ -513,183 +480,152 @@ R_API char *r2ai_qjs(RCore *core, RJson *args, bool verbose, char **edited_scrip
 	char *json_cmd = to_cmd (cmd);
 	if (!json_cmd) {
 		free (cmd);
-		return strdup ("{ \"res\":\"Failed to execute qjs\" }");
+		r2ai_tool_result_fini (&result);
+		result.output = strdup ("{ \"res\":\"Failed to execute qjs\" }");
+		return result;
 	}
 
 	char *cmd_output = r_core_cmd_str (core, json_cmd);
 	free (json_cmd);
 	free (cmd);
 
-	return cmd_output;
+	result.output = cmd_output;
+	return result;
 }
 
-R_API char *execute_tool(RCore *core, const char *tool_name, const char *args, char **edited_command, char **comment_out) {
+static char *tool_logs_to_string(const RJson *logs_json) {
+	if (!logs_json || logs_json->type != R_JSON_ARRAY) {
+		return NULL;
+	}
+	RStrBuf *sb = r_strbuf_new ("");
+	if (!sb) {
+		return NULL;
+	}
+	for (size_t i = 0; i < logs_json->children.count; i++) {
+		const RJson *log = r_json_item (logs_json, i);
+		if (log && log->type == R_JSON_OBJECT) {
+			const RJson *type = r_json_get (log, "type");
+			const RJson *message = r_json_get (log, "message");
+			if (message && message->type == R_JSON_STRING) {
+				if (type && type->type == R_JSON_STRING && R_STR_ISNOTEMPTY (type->str_value)) {
+					r_strbuf_appendf (sb, "%s: ", type->str_value);
+				}
+				r_strbuf_appendf (sb, "%s\n", message->str_value);
+			}
+		}
+	}
+	char *s = r_strbuf_drain (sb);
+	if (R_STR_ISEMPTY (s)) {
+		free (s);
+		return NULL;
+	}
+	return s;
+}
+
+static void normalize_tool_output(RCore *core, bool verbose, const char *tool_name, const char *args, R2AI_ToolResult *result) {
+	if (R_STR_ISEMPTY (result->output)) {
+		free (result->output);
+		result->output = strdup ("<no output>");
+		return;
+	}
+	if (!strcmp (result->output, "R2AI_SIGINT")) {
+		return;
+	}
+	if (!r_str_startswith (result->output, "{") || !strchr (result->output, '}')) {
+		return;
+	}
+
+	RJson *json = r_json_parse (result->output);
+	if (!json) {
+		R_LOG_WARN ("Failed to parse JSON response from tool execution");
+		return;
+	}
+
+	char *normalized = NULL;
+	const RJson *logs_json = r_json_get (json, "logs");
+	const RJson *error_json = r_json_get (json, "error");
+	if (error_json && (error_json->type == R_JSON_BOOLEAN || error_json->type == R_JSON_STRING) &&
+		((error_json->type == R_JSON_BOOLEAN && error_json->num.u_value) ||
+			(error_json->type == R_JSON_STRING && error_json->str_value[0]))) {
+		normalized = tool_logs_to_string (logs_json);
+		if (!normalized && error_json->type == R_JSON_STRING) {
+			normalized = strdup (error_json->str_value);
+		}
+		if (!normalized) {
+			normalized = strdup ("Error occurred (no details available)");
+		}
+	} else {
+		const RJson *res_json = r_json_get (json, "res");
+		if (res_json && res_json->type == R_JSON_STRING) {
+			normalized = strdup (res_json->str_value);
+		} else {
+			normalized = strdup (result->output);
+		}
+		if (R_STR_ISEMPTY (normalized)) {
+			free (normalized);
+			normalized = tool_logs_to_string (logs_json);
+		}
+		if (!normalized && !strcmp (tool_name, "r2cmd")) {
+			normalized = strdup ("<no output>");
+		}
+		if (!normalized) {
+			normalized = strdup ("<no output>");
+		}
+	}
+	r_json_free (json);
+
+	if (!normalized) {
+		normalized = r_str_newf ("Error running tool: Unknown error\nCommand: %s", args);
+	}
+	free (result->output);
+	result->output = normalized;
+
+	if (verbose) {
+		r_cons_newline (core->cons);
+		r_cons_printf (core->cons, "%s\n", result->output);
+		r_cons_newline (core->cons);
+		r_cons_flush (core->cons);
+	}
+	r_str_ansi_strip (result->output);
+}
+
+R_API R2AI_ToolResult execute_tool(RCorePluginSession *cps, const char *tool_name, const char *args) {
+	R2AI_ToolResult result = { 0 };
+	if (!cps || !cps->core) {
+		result.output = strdup ("{ \"res\":\"Missing r2ai session\" }");
+		return result;
+	}
+	RCore *core = cps->core;
+	R2AI_State *state = cps->data;
 	if (!tool_name || !args) {
-		return strdup ("{ \"res\":\"Tool name or arguments are NULL\" }");
+		result.output = strdup ("{ \"res\":\"Tool name or arguments are NULL\" }");
+		return result;
 	}
 
 	R_LOG_DEBUG ("Args: %s", args);
 
 	// Check if args is valid JSON before parsing
 	if (!r_str_startswith (args, "{") || !strchr (args, '}')) {
-		return r_str_newf ("Invalid JSON arguments: %s", args);
+		result.output = r_str_newf ("Invalid JSON arguments: %s", args);
+		return result;
 	}
 
 	RJson *args_json = r_json_parse ((char *)args);
 	if (!args_json) {
-		return r_str_newf ("Failed to parse arguments: %s", args);
+		result.output = r_str_newf ("Failed to parse arguments: %s", args);
+		return result;
 	}
 
 	bool verbose = r_config_get_b (core->config, "r2ai.auto.verbose");
-	char *tool_result = NULL;
 
 	if (strcmp (tool_name, "r2cmd") == 0) {
-		tool_result = r2ai_r2cmd (core, args_json, verbose, edited_command, comment_out);
+		result = r2ai_r2cmd (core, args_json, verbose);
+	} else if (strcmp (tool_name, "execute_js") == 0) {
+		result = r2ai_qjs (core, state, args_json, verbose);
 	} else {
-		if (comment_out) {
-			*comment_out = NULL;
-		}
-		if (strcmp (tool_name, "execute_js") == 0) {
-			tool_result = r2ai_qjs (core, args_json, verbose, edited_command);
-		} else {
-			tool_result = strdup ("{ \"res\":\"Unknown tool\" }");
-		}
+		result.output = strdup ("{ \"res\":\"Unknown tool\" }");
 	}
-
-	// Check for interruption after executing the tool
-	if (tool_result && strcmp (tool_result, "R2AI_SIGINT") == 0) {
-		free (tool_result);
-		return strdup ("R2AI_SIGINT");
-	}
-
-	// Try to parse as JSON (equivalent to json.loads in Python)
-	char *result = NULL;
-
-	// Check for empty or invalid response that could cause a crash
-	if (R_STR_ISEMPTY (tool_result)) {
-		free (tool_result);
-		return strdup ("{ \"res\":\"Error: Empty or invalid response from QJS execution\" }");
-	}
-
-	// Validate that the tool_result looks like valid JSON before parsing
-	if (!r_str_startswith (tool_result, "{") || !strchr (tool_result, '}')) {
-		// Not a JSON object, return as plain text
-		return tool_result;
-	}
-
-	// Try to parse the JSON safely
-	RJson *json = r_json_parse (tool_result);
-
-	if (!json) {
-		// JSON parsing failed, return original content
-		R_LOG_WARN ("Failed to parse JSON response from tool execution");
-		return tool_result;
-	}
-
-	if (json) {
-		// Check for error field
-		const RJson *error_json = r_json_get (json, "error");
-		if (error_json && (error_json->type == R_JSON_BOOLEAN || error_json->type == R_JSON_STRING) &&
-			((error_json->type == R_JSON_BOOLEAN && error_json->num.u_value) ||
-				(error_json->type == R_JSON_STRING && error_json->str_value[0]))) {
-
-			// Build error message from logs if available
-			const RJson *logs_json = r_json_get (json, "logs");
-			if (logs_json && logs_json->type == R_JSON_ARRAY) {
-				RStrBuf *sb = r_strbuf_new ("");
-
-				for (size_t i = 0; i < logs_json->children.count; i++) {
-					const RJson *log = r_json_item (logs_json, i);
-					if (log && log->type == R_JSON_OBJECT) {
-						const RJson *message = r_json_get (log, "message");
-						if (message && message->type == R_JSON_STRING) {
-							r_strbuf_appendf (sb, "%s\n", message->str_value);
-						}
-					}
-				}
-
-				result = r_strbuf_drain (sb);
-			} else {
-				// If no logs, return the error message directly
-				if (error_json->type == R_JSON_STRING) {
-					result = strdup (error_json->str_value);
-				} else {
-					result = strdup ("Error occurred (no details available)");
-				}
-			}
-		} else {
-			// Extract res field
-			const RJson *res_json = r_json_get (json, "res");
-			if (res_json && res_json->type == R_JSON_STRING) {
-				result = strdup (res_json->str_value);
-			} else {
-				// Just return the entire command output if res field not found
-				result = strdup (tool_result);
-			}
-		}
-
-		r_json_free (json);
-	} else {
-		// JSON parse failed, handle like JSONDecodeError in Python
-		char *trimmed = r_str_trim_dup (tool_result);
-		if (trimmed) {
-			// Manual line splitting to avoid r_str_split issues
-			RList *lines = r_str_split_list (trimmed, "\n", 0);
-			if (lines) {
-				int line_count = r_list_length (lines);
-				if (line_count > 0) {
-					char *last_line = r_list_get_n (lines, line_count - 1);
-					if (last_line && r_str_startswith (last_line, "{\"res\":\"\"")) {
-						// Remove the last line by joining all lines except the last
-						RStrBuf *sb = r_strbuf_new ("");
-						int i = 0;
-
-						// Use manual iteration instead of r_list_foreach_n
-						RListIter *iter;
-						char *line;
-						r_list_foreach (lines, iter, line) {
-							// Skip the last line
-							if (line != last_line) {
-								if (i > 0) {
-									r_strbuf_append (sb, "\n");
-								}
-								r_strbuf_append (sb, line);
-								i++;
-							}
-						}
-
-						result = r_strbuf_drain (sb);
-					} else {
-						// Keep the output as is
-						result = strdup (tool_result);
-					}
-				} else {
-					result = strdup (tool_result);
-				}
-				r_list_free (lines);
-			} else {
-				result = strdup (tool_result);
-			}
-			free (trimmed);
-		} else {
-			result = strdup (tool_result);
-		}
-	}
-
-	free (tool_result);
-
-	if (!result) {
-		return r_str_newf ("Error running tool: Unknown error\nCommand: %s", args);
-	}
-	if (verbose) {
-		r_cons_newline (core->cons);
-		r_cons_printf (core->cons, "%s\n", result);
-		r_cons_newline (core->cons);
-		r_cons_flush (core->cons);
-	}
-
-	r_str_ansi_strip (result);
-
 	r_json_free (args_json);
+	normalize_tool_output (core, verbose, tool_name, args, &result);
 	return result;
 }

src/vdb.c

@@ -1,4 +1,4 @@
-/* r2ai - MIT - Copyright 2024-2025 pancake */
+/* r2ai - MIT - Copyright 2024-2026 pancake */
 
 #include "r2ai.h"
 
@@ -35,10 +35,16 @@ static float squared_distance(const Vector *a, const Vector *b) {
 /*-------------------------------
 KDNode Utility Functions
 -------------------------------*/
-static KDNode *create_kdnode(const Vector *v, const char *text, int split_dim) {
-	KDNode *node = (KDNode *)malloc (sizeof (KDNode));
+static KDNode *create_kdnode(const Vector *v, const char *text, int split_dim) {	KDNode *node = (KDNode *)calloc (1, sizeof (KDNode));
+	if (!node) {
+		return NULL;
+	}
 	node->point.dim = v->dim;
 	node->point.data = (float *)malloc (sizeof (float) * v->dim);
+	if (!node->point.data) {
+		free (node);
+		return NULL;
+	}
 	memcpy (node->point.data, v->data, sizeof (float) * v->dim);
 	node->text = text? strdup (text): NULL;
 	node->split_dim = split_dim;
@@ -77,16 +83,21 @@ void r_vdb_free(RVdb *db) {
 }
 
 /* Recursive KD-tree insertion */
-static KDNode *kd_insert_recursive(KDNode *node, const Vector *v, const char *text, int depth, int dimension) {
-	if (node == NULL) {
+static KDNode *kd_insert_recursive(KDNode *node, const Vector *v, const char *text, int depth, int dimension) {	if (node == NULL) {
 		int split_dim = depth % dimension;
 		return create_kdnode (v, text, split_dim);
 	}
 	int axis = node->split_dim;
+	KDNode *child = kd_insert_recursive (
+		(v->data[axis] < node->point.data[axis]) ? node->left : node->right,
+		v, text, depth + 1, dimension);
+	if (!child) {
+		return node; // allocation failed, keep existing tree
+	}
 	if (v->data[axis] < node->point.data[axis]) {
-		node->left = kd_insert_recursive (node->left, v, text, depth + 1, dimension);
+		node->left = child;
 	} else {
-		node->right = kd_insert_recursive (node->right, v, text, depth + 1, dimension);
+		node->right = child;
 	}
 	return node;
 }
@@ -98,8 +109,7 @@ void r_vdb_insert(RVdb *db, const char *text) {
 		return;
 	}
 	float *embedding = (float *)calloc (db->dimension, sizeof (float));
-	// New call: pass the db pointer so TF-IDF stats are updated.
-	compute_embedding (db, text, embedding, db->dimension);
+	compute_embedding (db, text, embedding, db->dimension, true);
 	Vector v;
 	v.dim = db->dimension;
 	v.data = embedding;
@@ -109,9 +119,15 @@ void r_vdb_insert(RVdb *db, const char *text) {
 }
 
 // K-NN Search Data Structures and Helpers
-static RVdbResultSet *create_knn_result_set(int capacity) {
-	RVdbResultSet *rs = (RVdbResultSet *)malloc (sizeof (RVdbResultSet));
-	rs->results = (RVdbResult *)malloc (sizeof (RVdbResult) * capacity);
+static RVdbResultSet *create_knn_result_set(int capacity) {	RVdbResultSet *rs = (RVdbResultSet *)calloc (1, sizeof (RVdbResultSet));
+	if (!rs) {
+		return NULL;
+	}
+	rs->results = (RVdbResult *)calloc (capacity, sizeof (RVdbResult));
+	if (!rs->results) {
+		free (rs);
+		return NULL;
+	}
 	rs->capacity = capacity;
 	rs->size = 0;
 	return rs;
@@ -227,16 +243,22 @@ void kd_search_knn_recursive(KDNode *node, const Vector *query, RVdbResultSet *r
  * Find the k nearest neighbors to the embedding computed from `query_data`.
  * Returns a RVdbResultSet that must be freed by the caller.
  */
-RVdbResultSet *r_vdb_query_embedding(RVdb *db, const float *query_data, int k) {
-	if (!db || db->size == 0 || k <= 0) {
+RVdbResultSet *r_vdb_query_embedding(RVdb *db, const float *query_data, int k) {	if (!db || db->size == 0 || k <= 0) {
 		return NULL;
 	}
 	Vector query_vec = vector_new (db->dimension);
+	if (!query_vec.data) {
+		return NULL;
+	}
 	for (int i = 0; i < db->dimension; i++) {
 		query_vec.data[i] = query_data[i];
 	}
 	// query_vec is already normalized by compute_embedding ()
 	RVdbResultSet *rs = create_knn_result_set (k);
+	if (!rs) {
+		vector_free (&query_vec);
+		return NULL;
+	}
 	kd_search_knn_recursive (db->root, &query_vec, rs, 0, db->dimension);
 	/* Optional: sort the result set in ascending order of distance */
 	qsort (rs->results, rs->size, sizeof (RVdbResult), compare_knn_result);
@@ -246,7 +268,7 @@ RVdbResultSet *r_vdb_query_embedding(RVdb *db, const float *query_data, int k) {
 
 RVdbResultSet *r_vdb_query(RVdb *db, const char *text, int k) {
 	float *query_embedding = (float *)calloc (db->dimension, sizeof (float));
-	compute_embedding (db, text, query_embedding, db->dimension);
+	compute_embedding (db, text, query_embedding, db->dimension, false);
 	// No extra normalization is needed.
 	RVdbResultSet *res = r_vdb_query_embedding (db, query_embedding, k);
 	free (query_embedding);

src/vdb_embed.inc.c

@@ -1,4 +1,4 @@
-/* r2ai - MIT - Copyright 2024-2025 pancake */
+/* r2ai - MIT - Copyright 2024-2026 pancake */
 
 #include <r_util.h>
 
@@ -6,9 +6,10 @@
 
 // Create a new global DF entry.
 static void gtfidf_add(RList *db_tokens, const char *token) {
-	RVdbToken *t = R_NEW (RVdbToken);
+	RVdbToken *t = R_NEW0 (RVdbToken);
 	t->token = r_str_trim_dup (token);
 	t->count = 1;
+	t->df = 1.0f;
 	r_list_append (db_tokens, t);
 }
 
@@ -48,7 +49,7 @@ static bool valid_token(const char *a) {
 
 #if USE_OLLAMA_EMBED
 // experimental ollama
-static void compute_embedding(RVdb *db, const char *text, float *embedding, unsigned int dim) {
+static void compute_embedding(RVdb *db, const char *text, float *embedding, unsigned int dim, bool update_stats) {
 	// curl http://localhost:11434/api/embed -d '{ "model": "llama3:latest", "input": "text" }' |jq -r '.embeddings[0]'
 	char *json_text = r_str_escape_utf8_for_json (text, -1);
 	const char *model = "llama3:latest";
@@ -74,7 +75,7 @@ static void compute_embedding(RVdb *db, const char *text, float *embedding, unsi
 	free (s);
 }
 #else
-static void compute_embedding(RVdb *db, const char *text, float *embedding, unsigned int dim) {
+static void compute_embedding(RVdb *db, const char *text, float *embedding, unsigned int dim, bool update_stats) {
 	// gtfidf_list (db);
 
 	// Zero the embedding vector.
@@ -111,24 +112,24 @@ static void compute_embedding(RVdb *db, const char *text, float *embedding, unsi
 		token = strtok_r (NULL, " \t\r\n", &saveptr);
 	}
 	free (buffer);
-	db->total_docs++;
 
-	/* --- Step 2. Update Global Document Frequencies --- */
-	// Here we use the global definition of token_df (do not re-declare it locally).
+	/* --- Step 2. Update Global Document Frequencies (only when indexing, never for queries) --- */
 	RListIter *iter;
 	RVdbToken *dt_token;
-	r_list_foreach (doc_tokens, iter, dt_token) {
-		RVdbToken *t = gtfidf_find (db->tokens, dt_token->token);
-		if (t) {
-			if (valid_token (dt_token->token)) {
-				t->count++;
-				t->df += 1.0f;
+	if (update_stats) {
+		db->total_docs++;
+		r_list_foreach (doc_tokens, iter, dt_token) {
+			RVdbToken *t = gtfidf_find (db->tokens, dt_token->token);
+			if (t) {
+				if (valid_token (dt_token->token)) {
+					t->count++;
+					t->df += 1.0f;
+				}
+			} else {
+				gtfidf_add (db->tokens, dt_token->token);
 			}
-		} else {
-			gtfidf_add (db->tokens, dt_token->token);
 		}
 	}
-	// Increment the total number of documents.
 
 	/* --- Step 3. Compute TF-IDF for Each Token and Update the Embedding --- */
 	RVdbToken *dt;

src/vertex.c

@@ -0,0 +1,92 @@
+/* Copyright r2ai - 2023-2026 - pancake */
+
+#include "r2ai.h"
+#include "r2ai_priv.h"
+
+#define R2AI_VERTEX_TOKEN_TTL_US (3300ULL * R_USEC_PER_SEC) /* 55 minutes, tokens are valid for ~60 */
+
+static char *vertex_fetch_token(void) {
+	char *version = r_sys_cmd_str ("gcloud version", NULL, NULL);
+	if (R_STR_ISEMPTY (version)) {
+		free (version);
+		R_LOG_ERROR ("gcloud CLI not found. Install it from https://cloud.google.com/sdk/docs/install");
+		return NULL;
+	}
+	free (version);
+	char *token = r_sys_cmd_str ("gcloud auth application-default print-access-token", NULL, NULL);
+	if (R_STR_ISEMPTY (token)) {
+		free (token);
+		R_LOG_ERROR ("gcloud is not authenticated. Run 'gcloud auth application-default login'");
+		return NULL;
+	}
+	r_str_trim (token);
+	if (R_STR_ISEMPTY (token)) {
+		free (token);
+		return NULL;
+	}
+	return token;
+}
+
+/* The returned pointer is owned by state and must NOT be freed by the caller */
+R_IPI const char *r2ai_vertex_get_token(R2AI_State *state) {
+	ut64 now = r_time_now_mono ();
+	if (state->vertex_token && now < state->vertex_token_expiry) {
+		return state->vertex_token;
+	}
+	free (state->vertex_token);
+	state->vertex_token = vertex_fetch_token ();
+	state->vertex_token_expiry = state->vertex_token? now + R2AI_VERTEX_TOKEN_TTL_US: 0;
+	return state->vertex_token;
+}
+
+// returns the publisher endpoint url or NULL when the cloud project or region are not set
+static char *vertex_url(const char *publisher, const char *model, const char *method, char **error) {
+	char *project = r_sys_getenv ("GOOGLE_CLOUD_PROJECT");
+	char *region = r_sys_getenv ("GOOGLE_CLOUD_REGION");
+	char *url = NULL;
+	if (R_STR_ISNOTEMPTY (project) && R_STR_ISNOTEMPTY (region)) {
+		url = r_str_newf ("https://%s-aiplatform.googleapis.com/v1/projects/%s/locations/%s/publishers/%s/models/%s:%s",
+			region, project, region, publisher, model, method);
+	} else {
+		*error = strdup ("Set GOOGLE_CLOUD_PROJECT and GOOGLE_CLOUD_REGION env vars first");
+	}
+	free (project);
+	free (region);
+	return url;
+}
+
+static char *vertex_post(RCorePluginSession *cps, R2AIArgs *args, const char *name, const char *publisher, const char *method, const char *data) {
+	char *url = vertex_url (publisher, args->model, method, args->error);
+	if (!url) {
+		return NULL;
+	}
+	char *auth = r_str_newf ("Authorization: Bearer %s", args->api_key);
+	const char *headers[] = { "Content-Type: application/json", auth, NULL };
+	int code = 0;
+	char *res = r2ai_post (cps->core, name, url, headers, data, &code, args->error);
+	free (auth);
+	free (url);
+	return res;
+}
+
+R_IPI R2AI_ChatResponse *r2ai_vertex_gemini(RCorePluginSession *cps, R2AIArgs args) {
+	char *data = r2ai_gemini_request (&args);
+	char *res = vertex_post (cps, &args, "vertex_gemini", "google", "generateContent", data);
+	free (data);
+	R2AI_ChatResponse *result = res? r2ai_gemini_parse (res, args.error): NULL;
+	free (res);
+	return result;
+}
+
+R_IPI R2AI_ChatResponse *r2ai_vertex_anthropic(RCorePluginSession *cps, R2AIArgs args) {
+	char *data = r2ai_anthropic_request (&args, true);
+	if (!data) {
+		*args.error = strdup ("No input or messages provided");
+		return NULL;
+	}
+	char *res = vertex_post (cps, &args, "vertex_anthropic", "anthropic", "rawPredict", data);
+	free (data);
+	R2AI_ChatResponse *result = res? r2ai_anthropic_parse_response (res, args.error): NULL;
+	free (res);
+	return result;
+}

src/wizard.c

@@ -1,4 +1,4 @@
-/* r2ai wizard - Copyright 2025 pancake */
+/* r2ai wizard - Copyright 2025-2026 pancake */
 
 #define R_LOG_ORIGIN "r2ai_wizard"
 
@@ -65,7 +65,7 @@ static bool wizard_mark_ready(void) {
 static void show_clippy_message(RCore *core, const char *message) {
 	r_cons_clear00 (core->cons);
 	char *cmd = r_str_newf ("?E %s", message);
-	r_core_cmd_call (core, cmd);
+	r_core_call (core, cmd);
 	free (cmd);
 	r_cons_flush (core->cons);
 }
@@ -130,11 +130,40 @@ static bool wizard_step_setup(RCore *core) {
 
 	show_clippy_message (core, "🔑 Time to configure those API keys!");
 	r_cons_printf (core->cons, "\nOpening API key configuration...\n");
-	r_core_cmd_call (core, "r2ai -K");
+	r_core_call (core, "r2ai -K");
 
 	return true;
 }
 
+static void wizard_claw_invoke(RCore *core, const char *extra) {
+	char *cmd = R_STR_ISNOTEMPTY (extra)
+		? r_str_newf ("r2ai -id %s", extra)
+		: strdup ("r2ai -id");
+	r_core_call (core, cmd);
+	free (cmd);
+}
+
+static bool wizard_step_personality(RCore *core) {
+	if (r2ai_claw_exists ()) {
+		return true;
+	}
+	show_clippy_message (core, "🎭 Want to give r2ai a custom personality?");
+	r_cons_printf (core->cons,
+		"\nr2ai defaults to 'r2clippy'. Later you can %s.\n\n",
+		R2AI_CLAW_HINT);
+	if (!r_cons_yesno (core->cons, 'n', "Generate a custom personality now? (y/N)")) {
+		return true;
+	}
+	r_cons_printf (core->cons,
+		"\nDescribe it in a few words (e.g. 'grumpy pirate hacker'),\n"
+		"or just press Enter to let r2ai pick something at random:\n");
+	const char *desc = r_line_readline (core->cons);
+	char *trimmed = desc? r_str_trim_dup (desc): NULL;
+	wizard_claw_invoke (core, trimmed);
+	free (trimmed);
+	return true;
+}
+
 static bool wizard_step_basic_usage(RCore *core) {
 	show_clippy_message (core, "🧩 Let's learn the basics of AI-powered reversing!");
 
@@ -166,7 +195,7 @@ static bool wizard_step_basic_usage(RCore *core) {
 
 			if (r_str_startswith (line, "shell")) {
 				r_cons_printf (core->cons, "Entering r2ai shell mode (type 'exit' to return)...\n");
-				r_core_cmd_call (core, "r2ai -r");
+				r_core_call (core, "r2ai -r");
 			} else if (r_str_startswith (line, "back")) {
 				break;
 			} else if (r_str_startswith (line, "skip")) {
@@ -185,6 +214,7 @@ static bool wizard_step_decompilation(RCore *core) {
 		"\n"
 		"🔍 Decompilation with AI:\n"
 		"   r2ai -d                    # Decompile current function\n"
+		"   r2ai -do                   # Decompile with source offsets\n"
 		"   r2ai -d explain this       # Ask about current function\n"
 		"   r2ai -dr                   # Recursive decompilation\n\n"
 		"🎛️ Advanced decompilation:\n"
@@ -326,6 +356,10 @@ R_API bool r2ai_wizard(RCore *core) {
 		return false;
 	}
 
+	if (!wizard_step_personality (core)) {
+		return false;
+	}
+
 	if (!wizard_step_basic_usage (core)) {
 		return false;
 	}

test/Makefile

@@ -0,0 +1,4 @@
+all:
+	r2r -u db/extras
+
+.PHONY: all

test/db/extras/openai-compat

@@ -0,0 +1,33 @@
+NAME=OpenAI-compatible base URL
+FILE=-
+TIMEOUT=10
+CMDS=<<EOF
+!HOME=/tmp/r2ai-r2r-no-home R2_RCFILE=/dev/null ../src/r2ai -p ollama -m fixture-model -b http://127.0.0.1:1/v1/ -e debug=true -e http.backend=socket -e http.max_retries=0 -e http.timeout=1 'fixture prompt' 2>&1 | sed -n '/^Curl command:/{h;s/.*\(Content-Type: application\/json\).*/\1/p;g;s/.*"\(http[^\"]*\)".*/\1/p;}'
+EOF
+EXPECT=<<EOF
+Content-Type: application/json
+http://127.0.0.1:1/v1/chat/completions
+EOF
+RUN
+
+NAME=Native Ollama API root
+FILE=-
+TIMEOUT=10
+CMDS=<<EOF
+!HOME=/tmp/r2ai-r2r-no-home R2_RCFILE=/dev/null ../src/r2ai -p ollama -m fixture-model -b http://127.0.0.1:1/api -e debug=true -e http.backend=socket -e http.max_retries=0 -e http.timeout=1 'fixture prompt' 2>&1 | sed -n '/^Curl command:/s/.*"\(http[^\"]*\)".*/\1/p'
+EOF
+EXPECT=<<EOF
+http://127.0.0.1:1/api/chat
+EOF
+RUN
+
+NAME=Native Ollama generate API
+FILE=-
+TIMEOUT=10
+CMDS=<<EOF
+!HOME=/tmp/r2ai-r2r-no-home R2_RCFILE=/dev/null ../src/r2ai -p ollama -m fixture-model -b http://127.0.0.1:1/v1 -e apitype=generate -e debug=true -e http.backend=socket -e http.max_retries=0 -e http.timeout=1 'fixture prompt' 2>&1 | sed -n '/^Curl command:/s/.*"\(http[^\"]*\)".*/\1/p'
+EOF
+EXPECT=<<EOF
+http://127.0.0.1:1/api/generate
+EOF
+RUN

version.sh

@@ -0,0 +1,54 @@
+#!/bin/sh
+# version.sh - print current version or bump to a new one
+# usage: ./version.sh [new-version]
+# Updates VERSION, decai/src/constants.ts, src/r2ai.h, and src/meson.build
+
+VERSION_FILE="VERSION"
+TS_FILE="decai/src/constants.ts"
+R2AI_H="src/r2ai.h"
+MESON_FILE="src/meson.build"
+
+NEW_VERSION="${1:-}"
+
+get_version() {
+	if [ -f "$VERSION_FILE" ]; then
+		tr -d '[:space:]' < "$VERSION_FILE"
+	else
+		cat <<EOF >&2
+error: $VERSION_FILE not found
+EOF
+		exit 1
+	fi
+}
+
+CURRENT=$(get_version)
+
+if [ -z "$NEW_VERSION" ]; then
+	printf '%s\n' "$CURRENT"
+	exit 0
+fi
+
+# basic semver-ish validation: digits.digits.digits with optional pre-release/build
+if ! printf '%s' "$NEW_VERSION" | grep -Eq '^[0-9]+\.[0-9]+\.[0-9]+([-+.]?[A-Za-z0-9.-]*)?$'; then
+	cat <<EOF >&2
+error: invalid version '$NEW_VERSION'
+expected format: MAJOR.MINOR.PATCH[-prerelease]
+EOF
+	exit 1
+fi
+
+printf '%s\n' "$NEW_VERSION" > "$VERSION_FILE"
+
+if [ -f "$TS_FILE" ]; then
+	sed -i -E "s/export const VERSION = \"[^\"]+\";/export const VERSION = \"$NEW_VERSION\";/" "$TS_FILE"
+fi
+
+if [ -f "$R2AI_H" ]; then
+	sed -i -E "s/#define R2AI_VERSION \"[^\"]+\"/#define R2AI_VERSION \"$NEW_VERSION\"/" "$R2AI_H"
+fi
+
+if [ -f "$MESON_FILE" ]; then
+	sed -i -E "s/version : '[^']+'/version : '$NEW_VERSION'/" "$MESON_FILE"
+fi
+
+printf '%s -> %s\n' "$CURRENT" "$NEW_VERSION"