Release 1.2.2

* Fix the ollama provider

.github/workflows/ci.yml

@@ -93,8 +93,11 @@ jobs:
         shell: bash
         run: |
           mkdir -p release
-          cp -rf src/r2ai.dll src/r2ai.exe src/r2ai src/r2ai.dylib src/r2ai.so release/ || true
-          cp -rf src/build/r2ai src/build/r2ai.dylib src/build/r2ai.so release/ || true
+          # 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/*

MANIFEST.in

@@ -1 +0,0 @@
-recursive-include r2ai/ui *.tcss

Makefile

@@ -1,4 +1,6 @@
-TARGETS=all clean mrproper user-install install user-uninstall uninstall
+TARGETS=all clean mrproper user-install install user-uninstall uninstall fmt indent
+
 $(TARGETS):
 	$(MAKE) -C src $@
+
 .PHONY: $(TARGETS)

README.md

@@ -19,8 +19,8 @@
 
 This repository contains two plugins for radare2:
 
-* **r2ai** - native plugin for radare2
-* **decai** - r2js plugin with special focus on decompilation
+* **r2ai** - native [AI plugin](./src/README.md) for radare2
+* **decai** - r2js plugin with special [focus on decompilation](./decai/README.md)
 
 If you are looking to use radare2 with other agents via MCP:
 
@@ -37,12 +37,21 @@ If you are looking to use radare2 with other agents via MCP:
 * RAG markdown, code or textfiles using its native vector database
 * Embed the output of an r2 command and resolve questions on the given data
 
-## Practical Examples
+## User defined Prompts
 
-* Enhanced decompilation with `r2ai -d`
-* Autoname functions with `r2ai -n`
-* Explain function with `r2ai -x`
-* Find vulnerabilities with `r2ai -V`
+```console
+[0x00000000]> r2ai -q
+explain: Explain the current function -
+devices: Find and explain devices used -
+libs: Group imports by Libraries -
+varnames: Better variable names -
+autoname: Automatically suggest a better name for this function -
+vulns: Find vulnerabilities or bugs in the current function -
+signature: Suggest an improved function signature -
+dlopen: List libraries loaded with dlopen - Some libraries are loaded
+decompile: Augmented decompilation based on LLM -
+[0x00000000]>
+```
 
 ## Installation
 
@@ -85,12 +94,15 @@ r2ai -e model=claude-3-7-sonnet-20250219
 r2ai -e max_tokens=64000
 ```
 
-## Documentation
+## Further Reading
 
 * There's [a chapter](https://book.rada.re/plugins/r2ai.html) in the official r2book
 * Cryptax on [lmstudio+gptoss](https://cryptax.medium.com/r2ai-with-lmstudio-and-gpt-oss-08efa5ea2476) blog post
-* Malware analysis [with r2ai](https://github.com/cryptax/talks/blob/master/BSidesKristiansand-2025/r2ai.pdf) by Cryptax
+* Malware analysis [with r2ai](https://arxiv.org/pdf/2504.07574) by Cryptax and Daniel Nakov
+* Analysis of [Linux/Trigona ransomware](https://cryptax.medium.com/linux-trigona-analysis-with-r2ai-3e2bd1815e52),  [Linux/Prometei botnet](https://cryptax.medium.com/reversing-a-prometei-botnet-binary-with-r2-and-ai-part-one-3cdb3dc6ffab) and [W32/SkyAI](https://cryptax.medium.com/w32-skyai-uses-ai-so-do-i-d33f04d63534with) with r2ai
 
 ## Videos
 
-- https://infosec.exchange/@radareorg/111946255058894583
+- [Solving a crackme](https://infosec.exchange/@radareorg/111946255058894583)
+- [De-obfuscation of malware Linux/Ladvix](https://asciinema.org/a/724126)
+- [Analysis of the /fast option inside Linux/Trigona ransomware](https://asciinema.org/a/pBPEaJhp6cunWSKFpBUDTgPt4)

VERSION

@@ -1 +1 @@
-1.1.1
+1.2.2

autogen.sh

@@ -1,3 +1,5 @@
 #!/bin/sh
-vim src/r2ai.c
+vim decai/decai.r2.js
+vim src/r2ai.h
+vim src/meson.build
 vim VERSION

decai/decai.r2.js

@@ -1,5 +1,6 @@
 (function () {
   // Constants and configuration
+  const VERSION = "1.2.2";
   const COMMAND = "decai";
   const DEFAULT_PROMPT =
     "Transform this pseudocode and respond ONLY with plain code (NO explanations, comments or 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:', Reduce lines of code and fit everything in a single function, Remove all dead code";
@@ -727,8 +728,9 @@ Use radare2 to resolve user requests.
         "x-api-key: " + key[0],
       ];
 
+      let res = {};
       try {
-        const res = http.post(
+        res = http.post(
           provider.defaultBaseurl + "/v1/messages",
           headers,
           JSON.stringify(payload),
@@ -765,13 +767,20 @@ Use radare2 to resolve user requests.
 
       try {
         const res = http.post(url, [], JSON.stringify(payload));
-        return utils.filterResponse(res.message.content);
-      } catch (e) {
+        if (res && res.error) {
+          throw new Error(res.error);
+        }
+        if (res.message && res.message.content) {
+          return utils.filterResponse(res.message.content);
+        }
+        throw new Error(JSON.stringify(res));
         if (res.error && res.error.indexOf("try pulling")) {
           const modelName = res.error.split(/"/g)[1];
           res.error += "\n!ollama run " + modelName;
         }
-        return "ERROR: " + (res.error || e.message);
+      } catch (e) {
+        console.error(e.stack);
+        return "ERROR: " + e.message;
       }
     },
 
@@ -866,8 +875,10 @@ Use radare2 to resolve user requests.
   // Command handlers
   const commands = {
     help: () => {
-      const msg = (m) => console.error(" " + COMMAND + " " + m);
-      console.error("Usage: " + COMMAND + " (-h) ...");
+      let helpmsg = "";
+      const msg = (m) => helpmsg += " " + COMMAND + " " + m + "\n";
+      helpmsg += "Usage: " + COMMAND + " (-h) ...\n";
+      helpmsg += "Version: " + VERSION + "\n";
       msg("-a [query] - solve query with auto mode");
       msg("-d [f1 ..] - decompile given functions");
       msg("-dd [..]   - same as above, but ignoring cache");
@@ -888,6 +899,7 @@ Use radare2 to resolve user requests.
       msg("-v         - show local variables");
       msg("-V         - find vulnerabilities");
       msg("-x[*]      - eXplain current function (-x* for r2 script)");
+      r2.log(helpmsg.trim());
     },
 
     decompile: (args, extraQuery, useCache, recursiveCalls) => {

examples/benchmark/Makefile

@@ -1,41 +1,14 @@
 SHELL=/bin/bash
 
-PROVIDER=ollama
+MODELS=gpt-oss:20b gemma3:4b gemma3:12b
+MODELS+=granite3-dense:8b mannix/jan-nano:latest yi-coder:latest qwen3:latest deepseek-coder-v2:latest hhao/qwen2.5-coder-tools:latest qwen2.5-coder:latest
 MODELS=granite3-dense:8b mannix/jan-nano:latest yi-coder:latest qwen3:latest deepseek-coder-v2:latest hhao/qwen2.5-coder-tools:latest qwen2.5-coder:latest
-MODELS=gemma3:1b
 LANGS=Dart Perl Ruby Javascript PHP C
-BINARIES=files/complex1 files/complex2 files/complex3
-MODELS=gpt-4o
-PROVIDER=openai
 
 all:
-	mkdir -p tmp
-	for b in $(BINARIES) ; do \
-		for a in $(MODELS) ; do \
-			echo "$$b $$a"; \
-			./run.sh $$b $$a ${PROVIDER} ${LANGS} > tmp/$${b##*/}_$${a//\//_}.html ; \
+	for a in ${MODELS} ; do \
+		declare -a ARGS; \
+		for l in ${LANGS} ; do \
+			ARGS+=("echo \"decai -e model=$$a;decai -e lang=$$l;s 0x100000924;af;decai -d\" | r2 -q /bin/ls"); \
 		done ; \
-	done
-
-.PHONY: run index evaluate
-
-compile: $(BINARIES)
-
-files/complex1: files/complex1.c
-	gcc -o $@ $<
-
-files/complex2: files/complex2.c
-	gcc -o $@ $<
-
-files/complex3: files/complex3.c
-	gcc -o $@ $<
-
-lmstudio:
-	for b in $(BINARIES) ; do \
-		for a in $(MODELS) ; do \
-			./run.sh $$b $$a lmstudio ${LANGS} > $${b##*/}_$${a//\//_}.html ; \
-		done ; \
-	done
-
-clean:
-	rm -rf tmp
+		./run.sh "$${ARGS[@]}" > ${a//\//_}.html ; done

examples/benchmark/evaluate.sh

@@ -1,59 +0,0 @@
-#!/bin/bash
-
-# Evaluate decompiled outputs using mai and determine best model+lang
-
-echo "Evaluating decompiled outputs..."
-
-tmpfile=$(mktemp)
-MAI="mai -q -p openai -m gpt-4o"
-# MAI="mai -q"
-
-for orig in files/complex1.c files/complex2.c files/complex3.c; do
-    base=$(basename $orig .c)
-    original_code=$(cat $orig)
-    for f in tmp/${base}_*.txt; do
-        if [ -f "$f" ]; then
-            decompiled_code=$(cat $f)
-	    #echo "Comparing $orig $f"
-	    result=`echo n | ${MAI} -q -r "/template files/CMPROMPT.txt original=@${orig} decompiled=@${f}"`
-	    score=`echo "$result" | head -n 1 | grep -oE '[0-9]+'` # |head -n 1`
-	    perfect=`echo "$result" | grep 'No sign' && echo yes`
-		    [ -z "$score" ] && score=0
-	    if [ "$perfect" = yes ]; then
-		    score=100
-	    else
-		    score="$((0+${score}))"
-	    fi
-	    echo "$score\t$f"
-            if [ -n "$score" ]; then
-                key="${f%.*}"  # remove .txt
-                # Check if key exists in tmpfile
-                existing=$(grep "^$key " $tmpfile | awk '{print $2}')
-                if [ -n "$existing" ]; then
-                    new_score=$(( existing + score ))
-                    sed -i "s/^$key .*/$key $new_score/" $tmpfile
-                else
-                    echo "$key $score" >> $tmpfile
-                fi
-            fi
-        fi
-    done
-done
-
-# Find the best (highest total score)
-best_key=""
-best_score=0
-while read key score; do
-    if [ "$score" -gt "$best_score" ]; then
-        best_score=$score
-        best_key=$key
-    fi
-done < $tmpfile
-
-rm $tmpfile
-
-if [ -n "$best_key" ]; then
-    echo "Best model+language combination: $best_key with total score $best_score"
-else
-    echo "No scores found"
-fi

examples/benchmark/files/CMPROMPT.txt

@@ -1,17 +0,0 @@
-Compare the two provided code in C with the decompiled counterpart. Ignore function or variable names, syntax differences, language-specific constructs, and formatting. Focus solely on what the code does: the operations performed, control flow, logic, and functional behavior. Determine if they are functionally equivalent in terms of their effects and computations.
-
-Output only:
-- A numeric score from 1 to 100 (100 being identical logic and functionality)
-- Followed by a single sentence describing the most relevant semantic differences (or "No significant semantic differences" if score is 95+)
-
-**Original Source Code:**
-
-```C
-{original}
-```
-
-**Decompiled Code:**
-
-```
-{decompiled}
-```

examples/benchmark/files/complex1.c

@@ -1,21 +0,0 @@
-#include <stdio.h>
-#include <stdlib.h>
-
-int main() {
-    int n = 5;
-    int *arr = (int *)malloc(n * sizeof(int));
-    if (arr == NULL) return 1;
-    for (int i = 0; i < n; i++) {
-        arr[i] = i * i;
-    }
-    int sum = 0;
-    for (int i = 0; i < n; i++) {
-        sum += arr[i];
-        for (int j = 0; j < i; j++) {
-            sum -= j;
-        }
-    }
-    printf("Sum: %d\n", sum);
-    free(arr);
-    return 0;
-}

examples/benchmark/files/complex2.c

@@ -1,39 +0,0 @@
-#include <stdio.h>
-#include <stdlib.h>
-
-struct Node {
-    int data;
-    struct Node *next;
-};
-
-int main() {
-    struct Node *head = NULL;
-    struct Node *tail = NULL;
-    for (int i = 0; i < 10; i++) {
-        struct Node *newNode = (struct Node *)malloc(sizeof(struct Node));
-        if (newNode == NULL) return 1;
-        newNode->data = i * 2;
-        newNode->next = NULL;
-        if (head == NULL) {
-            head = newNode;
-            tail = newNode;
-        } else {
-            tail->next = newNode;
-            tail = newNode;
-        }
-    }
-    struct Node *current = head;
-    int sum = 0;
-    while (current != NULL) {
-        sum += current->data;
-        current = current->next;
-    }
-    printf("Sum: %d\n", sum);
-    current = head;
-    while (current != NULL) {
-        struct Node *temp = current;
-        current = current->next;
-        free(temp);
-    }
-    return 0;
-}

examples/benchmark/files/complex3.c

@@ -1,33 +0,0 @@
-#include <stdio.h>
-
-int main() {
-    int x = 42;
-    int y = 0;
-    switch (x & 15) {
-        case 0:
-            y = x << 1;
-            break;
-        case 1:
-            y = x >> 1;
-            break;
-        case 2:
-            y = x ^ 255;
-            break;
-        case 3:
-            y = ~x;
-            break;
-        default:
-            y = x | 128;
-            break;
-    }
-    if ((y & 1) == 0) {
-        y += 10;
-    } else {
-        y -= 5;
-    }
-    for (int i = 0; i < 5; i++) {
-        y = (y << 2) | (y >> 30);
-    }
-    printf("Result: %d\n", y);
-    return 0;
-}

examples/benchmark/index.sh

@@ -1,6 +1,5 @@
 #!/bin/bash
 
-cp decai/BENCH.txt tmp/BENCH.txt
 # Find all directories containing BENCH.txt
 dirs=$(find . -name "BENCH.txt" -type f | xargs dirname | sort | uniq)
 
@@ -47,4 +46,4 @@ cat <<EOF >> benchmark_index.html
 <br />
 <br />
 </html>
-EOF
+EOF

examples/benchmark/run.sh

@@ -1,55 +1,12 @@
 #!/bin/bash
 
 # Generate HTML output for command executions with collapsible sections and width control
-# If PLAIN=1, output plain text instead
 
-if [ $# -lt 4 ]; then
-    echo "Usage: $0 <binary> <model> <provider> <lang1> [lang2 ...]"
-    echo "Or: $0 command1 [command2 ...] (legacy)"
+if [ $# -eq 0 ]; then
+    echo "Usage: $0 command1 [command2 ...]"
     exit 1
 fi
 
-# Check if first arg is a file (binary), assume new format
-if [ -f "$1" ]; then
-    binary="$1"
-    model="$2"
-    provider="$3"
-    shift 3
-    langs=("$@")
-    commands=()
-    for lang in "${langs[@]}"; do
-       # commands+=("# $lang echo \"decai -e model=$model;decai -e lang=$lang;decai -e api=$provider;af;decai -d\" | r2 -q $binary 2> /dev/null")
-       commands+=("# $lang r2 -e scr.color=0 -qcq -qc \"decai -e model=$model;decai -e lang=$lang;decai -e api=$provider;af;decai -d\" $binary")
-    done
-    set -- "${commands[@]}"
-else
-    # Legacy mode
-    commands=("$@")
-fi
-
-if [ "$PLAIN" = "1" ]; then
-    # In plain mode, execute each command and print raw output.
-    # Also, if TXTFILE is set, save combined raw output there.
-    tmp_plain=$(mktemp)
-    for cmd in "$@"; do
-        if [[ "$cmd" =~ ^#[[:space:]]*(.*) ]]; then
-            exec_cmd="${BASH_REMATCH[1]}"
-        else
-            exec_cmd="$cmd"
-        fi
-        tmpout=$(mktemp)
-        eval "$exec_cmd" > "$tmpout" 2>&1 || true
-        mv "$tmpout" "$tmp_plain"
-        rm -f "$tmpout"
-    done
-    if [ -n "$TXTFILE" ]; then
-        mv "$tmp_plain" "$TXTFILE"
-    else
-        rm -f "$tmp_plain"
-    fi
-    exit 0
-fi
-
 # Timeout in seconds for each command (configurable via TIMEOUT environment variable)
 TIMEOUT=${TIMEOUT:-30}
 
@@ -152,27 +109,24 @@ EOF
 
 # Process each command
 for cmd in "$@"; do
-    # Extract title: if command starts with #, parse title and command, else use the command
-    if [[ "$cmd" =~ ^#[[:space:]]*([^[:space:]]+)[[:space:]]*(.*) ]]; then
+    # Extract title: if command starts with #, use the rest as title, else use the command
+    if [[ "$cmd" =~ ^#[[:space:]]*(.*) ]]; then
         title="${BASH_REMATCH[1]}"
-        exec_cmd="${BASH_REMATCH[2]}"
+        exec_cmd="${cmd#*#}"
+        exec_cmd="${exec_cmd#"${exec_cmd%%[![:space:]]*}"}"  # trim leading spaces
     else
         title="$cmd"
         exec_cmd="$cmd"
     fi
 
-    echo "RUN $exec_cmd" > /dev/stderr
     # Execute command and time it
     start_time=$(date +%s.%N)
     tmpfile=$(mktemp)
-    # echo "eval $exec_cmd > $tmpfile" > /dev/stder
-    eval "$exec_cmd" > "$tmpfile" 2> /dev/null &
+    eval "$exec_cmd" > "$tmpfile" 2>&1 &
     pid=$!
     (sleep $TIMEOUT && kill $pid 2>/dev/null) &
     wait $pid 2>/dev/null
     output=$(cat "$tmpfile")
-    #echo "$output" > /dev/stderr
-
     rm "$tmpfile"
     exit_code=$?
     end_time=$(date +%s.%N)
@@ -185,16 +139,8 @@ for cmd in "$@"; do
         execution_time=$(printf "%.3f" "$execution_time")
     fi
 
-    # Save plain output to .txt file if in new format
-    if [ -n "$binary" ]; then
-        txtfile="tmp/${binary##*/}_${model//\//_}_${title}.txt"
-        mkdir -p "$(dirname "$txtfile")" 2>/dev/null || true
-        echo "$output" > "$txtfile"
-    fi
-
     # Escape HTML in output
-    #escaped_output=$(echo "$output" | sed 's/&/\&amp;/g; s/</\&lt;/g; s/>/\&gt;/g')
-    escaped_output="$output"
+    escaped_output=$(echo "$output" | sed 's/&/\&amp;/g; s/</\&lt;/g; s/>/\&gt;/g')
 
     cat << EOF
         <div class="command-section expanded">

prompts/README.md

@@ -0,0 +1,43 @@
+# R2AI Prompts
+
+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.
+
+Lines starting with `#` are comments and ignored.
+
+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
+
+## Variable Substitution
+
+In the `Prompt` field, you can use:
+
+- `${VAR}`: Replaced with the value of the environment variable `VAR`
+- `$(cmd)`: Replaced with the output of the r2 command `cmd`
+
+## Usage
+
+- `r2ai -q`: List all available prompts
+- `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
+
+## Future Features
+
+- Support for r2js scripts
+- More advanced conditionals

prompts/autoname.r2ai.txt

@@ -0,0 +1,4 @@
+Title: Automatically suggest a better name for this function
+Author: pancake
+Command: afv;r2ai -d
+Prompt: Suggest a better name for this function. Output only the suggested function name (must be one alphanumeric word) prefixed with `'afn '. Check the code below:

prompts/decompile.r2ai.txt

@@ -0,0 +1,4 @@
+Title: Augmented decompilation based on LLM
+Author: pancake
+Command: pdc
+Prompt: $(-e r2ai.prompt). Translate the code into $(-e r2ai.lang) programming language.

prompts/devices.r2ai.txt

@@ -0,0 +1,4 @@
+Title: Find and explain devices used
+Author: pancake
+Command: izqq~^/dev
+Prompt: explain in one line the purpose of each of these devices

prompts/dlopen.r2ai.txt

@@ -0,0 +1,6 @@
+Title: List libraries loaded with dlopen
+Description: Some libraries are loaded
+Requires: analysis
+Command: iiq~dlopen
+If-Empty: This binary is not using dlopen
+If-Command: e emu.str=true;pd--5@@c:axtq sym.imp.dlopen

prompts/explain.r2ai.txt

@@ -0,0 +1,4 @@
+Title: Explain the current function
+Author: pancake
+Command: r2ai -d
+Prompt: Analyze function calls, comments and strings, ignore registers and memory accesses. Considering the references and involved loops; explain the purpose of this function in one or two short sentences. Output must be only the translation of the explanation in $(-e r2ai.hlang)

prompts/libs.r2ai.txt

@@ -0,0 +1,4 @@
+Title: Group imports by Libraries
+Author: pancake
+Commands: iiq;il
+Query: Group the imports that belong to each library and describe its purpose and usage scope

prompts/signature.r2ai.txt

@@ -0,0 +1,4 @@
+Title: Suggest an improved function signature
+Author: pancake
+Command: ?e variables;afv;?e function name:;afd;?e current signature:?e afs
+Prompt: Analyze the uses of the arguments and return value to infer the signature, identify which is the correct type for the resturn. Do NOT print the function body. Output must be *ONLY* the function signature prefixed with the `'afs ` text

prompts/tophp.r2ai.md

@@ -0,0 +1,10 @@
+---
+description: 'Decompile this code into PHP'
+version: 0.1
+author: pancake
+command: pdg
+if-empty: exit
+---
+Your goal is to $(-e r2ai.prompt).
+
+Translate the code into $(-e r2ai.lang) programming language.

prompts/varnames.r2ai.txt

@@ -0,0 +1,4 @@
+Title: Better variable names
+Author: pancake
+Command: afv
+Prompt: Output only the radare2 command without markdown, guess a better name and type for each local variable and function argument taking using. output an r2 script using afvn and afvt commands

prompts/vulns.r2ai.txt

@@ -0,0 +1,4 @@
+Title: Find vulnerabilities or bugs in the current function
+Author: pancake
+Command: pdc
+Prompt: Find vulnerabilities, do not show the code, only a short summary of the resulting analysis report including suggested good practices to address the problem and a practical exploit example.

src/.clang-format

@@ -1,40 +0,0 @@
-BasedOnStyle: LLVM
-Language: Cpp
-PointerAlignment: Right
-AlwaysBreakAfterDefinitionReturnType: None
-BinPackParameters: true
-MaxEmptyLinesToKeep: 1
-SpaceInEmptyParentheses: false
-SpacesInContainerLiterals: true
-SpaceBeforeParens: Custom
-SpaceBeforeParensOptions:
-  AfterIfMacros: true
-  AfterFunctionDefinitionName: false
-  AfterFunctionDeclarationName: false
-  AfterForeachMacros: true
-  AfterControlStatements: true
-  BeforeNonEmptyParentheses: false
-SpacesInParentheses: false
-InsertBraces: true
-ContinuationIndentWidth: 8
-IndentCaseLabels: false
-IndentFunctionDeclarationAfterType: false
-IndentWidth: 8
-UseTab: ForContinuationAndIndentation
-ColumnLimit: 0
-BreakBeforeBraces: Attach
-BreakBeforeTernaryOperators: true
-SpaceBeforeSquareBrackets: false
-SpaceInEmptyBlock: false
-AllowShortIfStatementsOnASingleLine: false
-AllowShortCaseLabelsOnASingleLine: true
-AllowShortFunctionsOnASingleLine: Inline
-AllowShortLoopsOnASingleLine: false
-AlignAfterOpenBracket: DontAlign
-AlignEscapedNewlines: DontAlign
-AlignConsecutiveMacros: true
-AlignTrailingComments: false
-AlignOperands: false
-Cpp11BracedListStyle: false
-ForEachMacros: ['r_list_foreach', 'ls_foreach', 'fcn_tree_foreach_intersect', 'r_skiplist_foreach', 'graph_foreach_anode', 'r_list_foreach_safe', 'r_pvector_foreach', 'r_rbtree_foreach', 'r_interval_    tree_foreach']
-SortIncludes: false

src/AGENTS.md

@@ -0,0 +1,32 @@
+# R2AI Plugin for radare2
+
+## Locations
+
+- All the source code of the plugin is in the current directory
+- Radare2 headers use to be in /usr/local/include/libr
+
+## Formatting Style
+
+- Use the radare2 coding style rules
+  - Always add a whitespace before `(` in function calls
+  - Indent the code with tabs
+  - Indent comments with spaces
+  - Do not define variables inside the `for` parenthesis 
+  - Do not check for null before calling free
+
+## Test
+
+Run `r2ai` cli tool or radare2 oneliners running the `r2ai` command.
+
+* `r2pm -r r2ai -h` -> help message for the r2ai tool
+* `r2 -qc 'r2ai -h' /bin/ls -> help message of the r2ai plugin
+* `r2 -qc 'r2ai -a what arch is used here' -> testing auto mode
+* `r2 -e r2ai.model=gemma3:12b -e r2ai.auto.raw=true -qc 'r2ai -a what arch is used here' -> testing auto mode
+
+## Actions
+
+Run the following commands to perform the action described
+
+- FORMAT: `make fmt`
+- COMPILE: `make` (with no extra arguments)
+- INSTALL: `make user-install`

src/Makefile

@@ -1,13 +1,21 @@
 DOTEXE=
 # DOTEXE=.exe
-DOTLIB=.$(shell r2 -H R2_LIBEXT)
+DOTLIB=.$(shell r2 -NNH R2_LIBEXT)
 
-CFLAGS+=$(shell pkg-config --cflags r_core)
-LDFLAGS+=$(shell pkg-config --libs r_core)
-R2SYSPLUGDIR=$(shell r2 -H R2_LIBR_PLUGINS)
-R2USRPLUGDIR=$(shell r2 -H R2_USER_PLUGINS)
+R2SYSPLUGDIR=$(shell r2 -NNH R2_LIBR_PLUGINS)
+R2USRPLUGDIR=$(shell r2 -NNH R2_USER_PLUGINS)
 R2BINDIR=$(shell r2pm -H R2PM_BINDIR)
+R2MANDIR=$(shell r2pm -H R2PM_MANDIR)
 
+R2_CFLAGS=$(shell pkg-config --cflags r_core)
+R2_LDFLAGS=$(shell pkg-config --libs r_core)
+ifeq ($(R2_CFLAGS),)
+R2_CFLAGS=$(shell r2 -NNH R2_CFLAGS)
+R2_LDFLAGS=$(shell r2 -NNH R2_LDFLAGS)
+endif
+
+CFLAGS+=$(R2_CFLAGS)
+LDFLAGS+=$(R2_LDFLAGS)
 CFLAGS+=-I. -fPIC
 CFLAGS+=-Wall
 CFLAGS+=-g
@@ -31,6 +39,7 @@ ifeq ($(USE_LIBCURL),1)
 endif
 
 OBJS=r2ai.o
+OBJS+=llm.o
 OBJS+=auto.o
 OBJS+=vdb.o
 OBJS+=tools.o
@@ -39,6 +48,8 @@ OBJS+=anthropic.o
 OBJS+=r2ai_http.o
 OBJS+=openai.o
 OBJS+=markdown.o
+OBJS+=query.o
+OBJS+=rawtools.o
 
 all: $(OBJS) r2check
 	$(CC) -fPIC -shared -o r2ai$(DOTLIB) $(OBJS) $(CFLAGS) $(LDFLAGS)
@@ -54,24 +65,37 @@ doc-install:
 
 vdb.o: vdb_embed.inc.c
 
-indent:
-	for a in *.c ; do python indent.py $$a ; done
+fmt format indent:
+	clang-format-radare2 *.c
 
 user-install: user-uninstall
 	mkdir -p $(R2USRPLUGDIR)
-	cp -f r2ai$(DOTLIB) $(R2USRPLUGDIR)
-	cp -f r2ai$(DOTEXE) $(R2BINDIR)
+	mkdir -p $(R2BINDIR)
+	mkdir -p $(R2MANDIR)/man1
+	mkdir -p $(HOME)/.config/r2ai/prompts
+	$(MAKE) user-uninstall
+	cp -f r2ai$(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/
 
 user-uninstall:
 	rm -f $(R2USRPLUGDIR)/r2ai$(DOTLIB)
+	rm -f $(R2BINDIR)/r2ai$(DOTEXE)
+	rm -f $(R2MANDIR)/man1/r2ai.1
 
 install:
-	cp -f r2ai$(DOTLIB) $(R2SYSPLUGDIR)
+	$(MAKE) uninstall
+	mkdir -p $(R2USRPLUGDIR)
+	mkdir -p $(R2BINDIR)
+	cp -f r2ai$(DOTLIB) $(R2SYSPLUGDIR)/r2ai$(DOTLIB)
+	cp -f r2ai$(DOTEXE) $(R2BINDIR)/r2ai$(DOTEXE)
 
 uninstall:
 	rm -f $(R2SYSPLUGDIR)/r2ai$(DOTLIB)
+	rm -f $(R2BINDIR)/r2ai$(DOTEXE)
 
 r2check:
-	@r2 -qcq --
+	@r2 -NN -qcq --
 clean:
 	rm -f *.o *.d

src/README.md

@@ -1,12 +1,10 @@
-# C-R2AI
+# R2AI
 
-This directory contains the C rewrite of the r2ai plugin for radare2.
+The r2ai plugin for radare2.
 
 ## Installation
 
-Requires radare2 from git if you want to use the stream HTTP APIs.
-
-Otherwise the only requirement is a C compiler.
+Requires a C compiler and modern version of radare2 if you want to use the stream HTTP APIs. It will fallback to curl or powershell if not available and `R2_CURL` environment is set to `1`.
 
 ```console
 make
@@ -37,24 +35,53 @@ Usage: r2ai   [-args] [...]
 | r2ai -d [query]         Ask a question on the current function
 | r2ai -dr                Decompile current function (+ 1 level of recursivity)
 | r2ai -a [query]         Resolve question using auto mode
-| r2ai -e                 Same as '-e r2ai.'
+| r2ai -e (k(=v))         Same as '-e r2ai.'
 | r2ai -h                 Show this help message
 | r2ai -i [file] [query]  read file and ask the llm with the given query
 | r2ai -m                 show selected model, list suggested ones, choose one
-| r2ai -n                 suggest a better name for the current function
+| r2ai -p [provider]      set LLM provider (openai, anthropic, gemini, etc.)
+| r2ai -q                 list available query prompts
+| 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 -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)
-| r2ai -s                 function signature
-| r2ai -x                 explain current function
-| r2ai -v                 suggest better variables names and types
-| r2ai -V[r]              find vulnerabilities in the decompiled code (-Vr uses -dr)
-| r2ai [arg]              send a post request to talk to r2ai and print the output
+| r2ai [query]            query the selected model+provider with the given query
 [0x100003f58]> 
 ```
 
+## 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.
+
+### Available Prompts
+
+- `autoname`: Automatically rename functions and variables
+- `decompile`: Decompile the current function
+- `devices`: Analyze device-related code
+- `dlopen`: Analyze dynamic library loading
+- `explain`: Explain the purpose of the current function
+- `libs`: Analyze library usage
+- `signature`: Improve function signatures
+- `varnames`: Rename variables based on context
+- `vulns`: Find vulnerabilities in the current function
+
+### Usage
+
+- `r2ai -q`: List all available query prompts
+- `r2ai -q [name]`: Run the specified prompt on the current function
+- `r2ai -q [name] [instructions]`: Run the prompt with additional custom instructions
+
+### Example
+
+```console
+[0x00001303]> r2ai -q explain
+The main function implements a simple substitution cipher program that reads user input, encrypts it using a hardcoded alphabet, and then decrypts it back to demonstrate the algorithm.
+```
+
+Query prompts automate repetitive analysis tasks by executing predefined radare2 commands and sending context-aware prompts to the LLM.
+
 ## Configuration
 
 These can be set with `r2ai -e <keyname>=<value>`
@@ -76,7 +103,6 @@ These can be set with `r2ai -e <keyname>=<value>`
 | r2ai.http.use_files | If you get CURL argument too long error, set this to true and arguments will be stored in a file.|
 
 
-
 ## Example
 
 - Get the [substitute example binary](https://github.com/radareorg/r2ai/tree/master/examples)
@@ -215,8 +241,6 @@ Note: System prompt is applied automatically but not stored in history
     {"command":"pdf @ main"}
 ```
 
-
-
 ## TODO
 
 * add "undo" command to drop the last message
@@ -225,7 +249,6 @@ Note: System prompt is applied automatically but not stored in history
 * keep context between direct commands unless explicitly reset
 * modify context if tool command is user edited
 
-
 ## Messages API
 
 The r2ai project includes a new API for managing messages and tool calls in conversations with AI models. This API is implemented in `messages.c` and `messages.h`.
@@ -268,4 +291,4 @@ In r2ai, as lm-studio uses an OpenAI-like API, configure:
 
 - `r2ai -e api=openai`
 - `r2ai -e baseurl=http://LM-STUDIO-IP:PORT`
-- `r2ai -e model=?` to list available models
+- `r2ai -e model=?` to list available models

src/anthropic.c

@@ -1,10 +1,11 @@
 #include "r2ai.h"
 
 static const char *modelname(const char *model_name) {
-	return model_name ? model_name : "claude-3-7-sonnet-20250219";
+	return model_name? model_name: "claude-3-7-sonnet-20250219";
 }
 
-R_IPI R2AI_ChatResponse *r2ai_anthropic(RCore *core, R2AIArgs args) {
+R_IPI R2AI_ChatResponse *r2ai_anthropic(RCorePluginSession *cps, R2AIArgs args) {
+	RCore *core = cps->core;
 	const char *model = args.model;
 	char **error = args.error;
 	const R2AI_Tools *tools = args.tools;
@@ -40,7 +41,7 @@ R_IPI R2AI_ChatResponse *r2ai_anthropic(RCore *core, R2AIArgs args) {
 	// Create messages JSON
 	char *messages_json = NULL;
 
-	if (messages_input && messages_input->n_messages > 0) {
+	if (messages_input && r_list_length (messages_input->messages) > 0) {
 		// Convert directly to JSON without filtering
 		messages_json = r2ai_msgs_to_anthropic_json (messages_input);
 		if (!messages_json) {
@@ -98,8 +99,13 @@ R_IPI R2AI_ChatResponse *r2ai_anthropic(RCore *core, R2AIArgs args) {
 	free (messages_json);
 
 	// Save the full JSON for debugging
-	r_file_dump ("/tmp/r2ai_anthropic_request.json", (const ut8 *)data, strlen (data), 0);
-	R_LOG_DEBUG ("Full request saved to /tmp/r2ai_anthropic_request.json");
+	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);
 
 	// Make the API call
@@ -120,8 +126,13 @@ R_IPI R2AI_ChatResponse *r2ai_anthropic(RCore *core, R2AIArgs args) {
 	}
 
 	// Save the response for inspection
-	r_file_dump ("/tmp/r2ai_anthropic_response.json", (const ut8 *)res, strlen (res), 0);
-	R_LOG_DEBUG ("Anthropic API response saved to /tmp/r2ai_anthropic_response.json");
+	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);
+
 	R_LOG_DEBUG ("Anthropic API response: %s", res);
 
 	// Parse the response

src/auto.c

@@ -1,76 +1,60 @@
 #include "r2ai.h"
-#include <time.h>
 
-// Forward declaration of the r2ai_llmcall function
-extern R2AI_ChatResponse *r2ai_llmcall(RCore *core, R2AIArgs args);
-
-// Add a global structure to track timing and costs
-typedef struct {
-	double total_cost;
-	double run_cost;
-	int total_tokens;
-	int run_tokens;
-	int total_prompt_tokens;
-	int run_prompt_tokens;
-	int total_completion_tokens;
-	int run_completion_tokens;
-	time_t start_time;
-	time_t total_start_time;
-} R2AIStats;
-
-static R2AIStats stats = { 0 };
+// Stats are now stored in R2AI_State->stats, no longer need global variable
 
 // Helper function to format time duration
 static char *format_time_duration(time_t seconds) {
 	if (seconds < 60) {
-		return r_str_newf ("%llds", (long long)seconds);
+		return r_str_newf ("%" PFMT64d "s", (long long)seconds);
 	}
 	if (seconds < 3600) {
-		return r_str_newf ("%lldm%llds", (long long)(seconds / 60), (long long)(seconds % 60));
+		return r_str_newf ("%" PFMT64d "m%" PFMT64d "s", (long long) (seconds / 60), (long long) (seconds % 60));
 	}
-	return r_str_newf ("%lldh%lldm%llds",
-		(long long)(seconds / 3600),
-		(long long)((seconds % 3600) / 60),
-		(long long)(seconds % 60));
+	return r_str_newf ("%" PFMT64d "h%" PFMT64d "m%" PFMT64d "s",
+		(long long) (seconds / 3600),
+		(long long) ((seconds % 3600) / 60),
+		(long long) (seconds % 60));
 }
 
 // Initialize timing and cost tracking for a run
-static void r2ai_stats_init_run(int n_run) {
+static void r2ai_stats_init_run(R2AI_State *state, int n_run) {
 	time_t run_start = time (NULL);
 	if (n_run == 1) {
 		// First run, initialize total timing
-		stats.total_cost = 0.0;
-		stats.run_cost = 0.0;
-		stats.total_start_time = run_start;
-		stats.total_tokens = 0;
-		stats.run_tokens = 0;
-		stats.total_prompt_tokens = 0;
-		stats.run_prompt_tokens = 0;
-		stats.total_completion_tokens = 0;
-		stats.run_completion_tokens = 0;
+		state->stats.total_cost = 0.0;
+		state->stats.run_cost = 0.0;
+		state->stats.total_start_time = run_start;
+		state->stats.total_tokens = 0;
+		state->stats.run_tokens = 0;
+		state->stats.total_prompt_tokens = 0;
+		state->stats.run_prompt_tokens = 0;
+		state->stats.total_completion_tokens = 0;
+		state->stats.run_completion_tokens = 0;
 	}
-	stats.start_time = run_start;
+	state->stats.start_time = run_start;
 }
 
 // Print a simple run indicator at the start
-static void r2ai_print_run_end(RCore *core, const R2AI_Usage *usage, int n_run, int max_runs) {
+static void r2ai_print_run_end(RCorePluginSession *cps, const R2AI_Usage *usage, int n_run, int max_runs) {
+	RCore *core = cps->core;
+	R2AI_State *state = cps->data;
 	(void)n_run;
 	(void)max_runs;
-	time_t run_time = time (NULL) - stats.start_time;
-	time_t total_time = time (NULL) - stats.total_start_time;
+	time_t run_time = time (NULL) - state->stats.start_time;
+	time_t total_time = time (NULL) - state->stats.total_start_time;
 	if (usage) {
-		stats.run_tokens = usage->total_tokens;
-		stats.run_prompt_tokens = usage->prompt_tokens;
-		stats.run_completion_tokens = usage->completion_tokens;
-		stats.total_tokens += usage->total_tokens;
-		stats.total_prompt_tokens += usage->prompt_tokens;
-		stats.total_completion_tokens += usage->completion_tokens;
+		state->stats.run_tokens = usage->total_tokens;
+		state->stats.run_prompt_tokens = usage->prompt_tokens;
+		state->stats.run_completion_tokens = usage->completion_tokens;
+		state->stats.total_tokens += usage->total_tokens;
+		state->stats.total_prompt_tokens += usage->prompt_tokens;
+		state->stats.total_completion_tokens += usage->completion_tokens;
 	}
 
 	if (r_config_get_b (core->config, "r2ai.chat.show_cost") == true) {
 		// TODO: calculate cost
-		stats.run_cost = 0.0 * run_time;
-		stats.total_cost += stats.run_cost;
+		state->stats.run_cost = 0.0 * run_time;
+		state->stats.total_cost += state->stats.run_cost;
 	}
 
 	// Format times for display
@@ -78,83 +62,90 @@ static void r2ai_print_run_end(RCore *core, const R2AI_Usage *usage, int n_run,
 	char *total_time_str = format_time_duration (total_time);
 
 	// Print detailed stats
-	R2_PRINTF ("\x1b[1;34m%s | total: %d in: %d out: %d | run: %d in: %d out: %d | %s / %s\x1b[0m\n",
+	r_cons_printf (core->cons, "\x1b[1;34m%s | total: %d in: %d out: %d | run: %d in: %d out: %d | %s / %s\x1b[0m\n",
 		r_config_get (core->config, "r2ai.model"),
-		stats.total_tokens,
-		stats.total_prompt_tokens,
-		stats.total_completion_tokens,
-		stats.run_tokens,
-		stats.run_prompt_tokens,
-		stats.run_completion_tokens,
+		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);
-	R2_NEWLINE ();
-	R2_FLUSH ();
+	r_cons_newline (core->cons);
+	r_cons_flush (core->cons);
 
 	free (run_time_str);
 	free (total_time_str);
 }
 
 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"
-			   "- 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";
+			"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"
+			"- 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";
 
 // Helper function to process messages and handle tool calls recursively
-R_API void process_messages(RCore *core, R2AI_Messages *messages, const char *system_prompt, int n_run) {
+R_API void process_messages(RCorePluginSession *cps, R2AI_Messages *messages, const char *system_prompt, int n_run) {
+	RCore *core = cps->core;
+	R2AI_State *state = cps->data;
 	char *error = NULL;
 	bool interrupted = false;
 	const int max_runs = r_config_get_i (core->config, "r2ai.auto.max_runs");
 
 	if (n_run > max_runs) {
-		R2_PRINTF ("\x1b[1;31m[r2ai] Max runs reached\x1b[0m\n");
-		R2_FLUSH ();
+		r_cons_printf (core->cons, "\x1b[1;31m[r2ai] Max runs reached\x1b[0m\n");
+		r_cons_flush (core->cons);
 		return;
 	}
 
 	if (!system_prompt) {
+		system_prompt = Gprompt_auto;
 		const char *init_commands = r_config_get (core->config, "r2ai.auto.init_commands");
 		if (R_STR_ISNOTEMPTY (init_commands)) {
 			char *edited_command = NULL;
 			char *cmd_output = execute_tool (core, "r2cmd", r_str_newf ("{\"command\":\"%s\"}", init_commands), &edited_command);
-			if (cmd_output) {
-				system_prompt = r_str_newf ("%s\n\nHere is some information about the binary to get you started:\n>%s\n%s", Gprompt_auto, edited_command, cmd_output);
+			if (R_STR_ISNOTEMPTY (cmd_output)) {
+				R2AI_Message init_msg = {
+					.role = "system",
+					.content = r_str_newf ("Here is some information about the binary to get you started:\n>%s\n%s", edited_command, cmd_output)
+				};
+				r2ai_msgs_add (messages, &init_msg);
 				free (cmd_output);
 			}
 			free (edited_command);
-		} else {
-			system_prompt = Gprompt_auto;
 		}
 	}
 
-	r2ai_stats_init_run (n_run);
+	r2ai_stats_init_run (state, n_run);
+
+	r_cons_printf (core->cons, "\x1b[35m[DEBUG] About to call r2ai_llmcall with n_run=%d\x1b[0m\n", n_run);
+	r_cons_flush (core->cons);
 
 	// Set up args for r2ai_llmcall call with tools directly
 	R2AIArgs args = {
 		.messages = messages,
 		.error = &error,
 		.dorag = true,
-		.tools = r2ai_get_tools (),
+		.tools = r2ai_get_tools (), // Always send tools in auto mode
 		.system_prompt = system_prompt
 	};
 
-	// Call r2ai_llmcall to get a response
-	R2AI_ChatResponse *response = r2ai_llmcall (core, args);
+	R2AI_ChatResponse *response = r2ai_llmcall (cps, args);
 
 	if (!response) {
 		return;
@@ -169,18 +160,34 @@ R_API void process_messages(RCore *core, R2AI_Messages *messages, const char *sy
 		return;
 	}
 
+	// Debug logging for rawtools
+	R_LOG_DEBUG ("[DEBUG] Model response - Role: %s", message->role? message->role: "null");
+	if (message->content) {
+		R_LOG_DEBUG ("[DEBUG] Content: %s", message->content);
+	}
+	if (message->reasoning_content) {
+		R_LOG_DEBUG ("[DEBUG] Reasoning: %s", message->reasoning_content);
+	}
+	if (message->tool_calls && message->n_tool_calls > 0) {
+		R_LOG_DEBUG ("[DEBUG] Tool calls: %d", message->n_tool_calls);
+		for (int i = 0; i < message->n_tool_calls; i++) {
+			const R2AI_ToolCall *tc = &message->tool_calls[i];
+			R_LOG_DEBUG ("[DEBUG] Tool %d: %s - %s", i, tc->name? tc->name: "null", tc->arguments? tc->arguments: "null");
+		}
+	}
+
 	// Process the response - we need to add it to our messages array
 	if (message->content || message->reasoning_content) {
-		R2_PRINTF ("\x1b[31m[Assistant]\x1b[0m\n\n");
+		r_cons_printf (core->cons, "\x1b[31m[Assistant]\x1b[0m\n\n");
 		if (message->reasoning_content) {
-			R2_PRINTF ("\x1b[90m<thinking>\n%s\n</thinking>\x1b[0m\n", message->reasoning_content);
-			R2_NEWLINE ();
-			R2_FLUSH ();
+			r_cons_printf (core->cons, "\x1b[90m<thinking>\n%s\n</thinking>\x1b[0m\n", message->reasoning_content);
+			r_cons_newline (core->cons);
+			r_cons_flush (core->cons);
 		}
 		if (message->content) {
-			R2_PRINTF ("%s", message->content);
-			R2_NEWLINE ();
-			R2_FLUSH ();
+			r_cons_printf (core->cons, "%s", message->content);
+			r_cons_newline (core->cons);
+			r_cons_flush (core->cons);
 		}
 	}
 
@@ -190,16 +197,17 @@ R_API void process_messages(RCore *core, R2AI_Messages *messages, const char *sy
 
 	// Check for tool calls and process them
 	if (message->tool_calls && message->n_tool_calls > 0) {
-		R_LOG_DEBUG ("Found %d tool call (s)", message->n_tool_calls);
+		R_LOG_DEBUG ("Found %d tool call(s)", message->n_tool_calls);
 
 		// Process each tool call
 		for (int i = 0; i < message->n_tool_calls; i++) {
 			const R2AI_ToolCall *tool_call = &message->tool_calls[i];
 
 			if (!tool_call->name || !tool_call->arguments || !tool_call->id) {
+				R_LOG_DEBUG ("Skipping invalid tool call %d", i);
 				continue;
 			}
-			R_LOG_DEBUG ("Tool call: %s", tool_call->name);
+			R_LOG_DEBUG ("Tool call %d: %s with args: %s", i, tool_call->name, tool_call->arguments);
 			// Don't log the full arguments which might get truncated
 			char *tool_name = strdup (tool_call->name);
 			char *tool_args = strdup (tool_call->arguments);
@@ -216,15 +224,46 @@ R_API void process_messages(RCore *core, R2AI_Messages *messages, const char *sy
 			} else {
 				char *edited_command = NULL;
 				cmd_output = execute_tool (core, tool_name, tool_args, &edited_command);
-				// TODO: need to edit the R2AI_Messages* and modify the command of the last tool_use
+				if (edited_command) {
+					// Update the last message's tool call arguments with the edited command
+					R2AI_Message *last_msg = r_list_get_n (messages->messages, r_list_length (messages->messages) - 1);
+					if (last_msg && last_msg->tool_calls && last_msg->n_tool_calls > 0) {
+						for (int j = 0; j < last_msg->n_tool_calls; j++) {
+							R2AI_ToolCall *tc = (R2AI_ToolCall *)&last_msg->tool_calls[j];
+							if (tc->id && strcmp (tc->id, tool_call->id) == 0) {
+								// For r2cmd, update the command in arguments
+								if (strcmp (tool_name, "r2cmd") == 0) {
+									char *args_dup = strdup (tc->arguments);
+									RJson *args_json = r_json_parse (args_dup);
+									if (args_json) {
+										RJson *cmd_json = (RJson *)r_json_get (args_json, "command");
+										if (cmd_json && cmd_json->str_value) {
+											// Update the command field
+											free ((char *)cmd_json->str_value);
+											cmd_json->str_value = strdup (edited_command);
+											// Serialize back to JSON
+											char *new_args = r_json_to_string (args_json);
+											if (new_args) {
+												free ((void *)tc->arguments);
+												tc->arguments = new_args;
+											}
+										}
+										r_json_free (args_json);
+									}
+								}
+								break;
+							}
+						}
+					}
+				}
 				free (edited_command);
 			}
 
 			free (tool_name);
 			free (tool_args);
 			if (strcmp (cmd_output, "R2AI_SIGINT") == 0) {
-				R2_PRINTF ("\n\n\x1b[1;31m[r2ai] Processing interrupted after tool execution\x1b[0m\n\n");
-				R2_FLUSH ();
+				r_cons_printf (core->cons, "\n\n\x1b[1;31m[r2ai] Processing interrupted after tool execution\x1b[0m\n\n");
+				r_cons_flush (core->cons);
 				free (cmd_output);
 				cmd_output = strdup ("<user interrupted>");
 				interrupted = true;
@@ -233,32 +272,41 @@ R_API void process_messages(RCore *core, R2AI_Messages *messages, const char *sy
 			// Create a tool call response message
 			R2AI_Message tool_response = {
 				.role = "tool",
-				.tool_call_id = tool_call->id,
+				.tool_call_id = strdup (tool_call->id),
 				.content = cmd_output
 			};
 
 			// Add the tool response to our messages array
 			r2ai_msgs_add (messages, &tool_response);
+			R_LOG_DEBUG ("[DEBUG] Added tool response to messages: %s", cmd_output? cmd_output: "null");
+			r_cons_printf (core->cons, "\x1b[32m[DEBUG] Tool result: %s\x1b[0m\n", cmd_output? cmd_output: "no output");
 			free (cmd_output);
 		}
 
-		r2ai_print_run_end (core, usage, n_run, max_runs);
+		r2ai_print_run_end (cps, usage, n_run, max_runs);
 
 		// Check if we should continue with recursion
 		if (!interrupted && message->tool_calls && message->n_tool_calls > 0) {
-			process_messages (core, messages, system_prompt, n_run + 1);
+			R_LOG_DEBUG ("[DEBUG] Recursing to process_messages with n_run=%d", n_run + 1);
+			process_messages (cps, messages, system_prompt, n_run + 1);
+		} else {
+			R_LOG_DEBUG ("[DEBUG] Auto mode loop ending - no more tool calls or interrupted");
 		}
 	} else {
-		r2ai_print_run_end (core, usage, n_run, max_runs);
+		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);
 }
 
-R_IPI void cmd_r2ai_a(RCore *core, const char *user_query) {
+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, "\x1b[35m[DEBUG] cmd_r2ai_a called with query: %s\x1b[0m\n", user_query);
+	r_cons_flush (core->cons);
 	// Get conversation
-	R2AI_Messages *messages = r2ai_conversation_get ();
+	R2AI_Messages *messages = r2ai_conversation_get (state);
 	if (!messages) {
 		R_LOG_ERROR ("Conversation not initialized");
 		return;
@@ -266,24 +314,24 @@ R_IPI void cmd_r2ai_a(RCore *core, const char *user_query) {
 
 	// Add user query to the conversation (no system prompt)
 	// If this is the first message in a new conversation, clear previous history
-	if (messages->n_messages == 0 || r_config_get_b (core->config, "r2ai.auto.reset_on_query")) {
+	if (r_list_length (messages->messages) == 0 || r_config_get_b (core->config, "r2ai.auto.reset_on_query")) {
 		r2ai_msgs_clear (messages);
 	}
 
 	// Add user query
 	R2AI_Message user_msg = {
 		.role = "user",
-		.content = user_query
+		.content = (char *)user_query
 	};
 	r2ai_msgs_add (messages, &user_msg);
 
-	process_messages (core, messages, NULL, 1);
+	process_messages (cps, messages, NULL, 1);
 }
 
 // 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, bool always_show_length) {
 	if (!content || *content == '\0') {
-		R2_PRINTF ("%s\n", empty_msg ? empty_msg : "<no content>");
+		r_cons_printf (core->cons, "%s\n", empty_msg? empty_msg: "<no content>");
 		return;
 	}
 
@@ -293,24 +341,26 @@ static void print_content_with_length(RCore *core, const char *content, const ch
 	if (content_len > max_display) {
 		// Truncate long content and show length
 		char *truncated = r_str_ndup (content, max_display);
-		R2_PRINTF ("%s... \x1b[1;37m(length: %zu chars)\x1b[0m\n",
+		r_cons_printf (core->cons, "%s... \x1b[1;37m(length: %zu chars)\x1b[0m\n",
 			truncated, content_len);
 		free (truncated);
 	} else if (always_show_length) {
 		// Always show length for certain types (like tool responses)
-		R2_PRINTF ("%s \x1b[1;37m(length: %zu chars)\x1b[0m\n",
+		r_cons_printf (core->cons, "%s \x1b[1;37m(length: %zu chars)\x1b[0m\n",
 			content, content_len);
 	} else {
-		R2_PRINTF ("%s\n", content);
+		r_cons_printf (core->cons, "%s\n", content);
 	}
 }
 
 // Add this function right after cmd_r2ai_a
-R_IPI void cmd_r2ai_logs(RCore *core) {
+R_IPI void cmd_r2ai_logs(RCorePluginSession *cps) {
+	RCore *core = cps->core;
+	R2AI_State *state = cps->data;
 	// Get conversation
-	R2AI_Messages *messages = r2ai_conversation_get ();
-	if (!messages || messages->n_messages == 0) {
-		R2_PRINTF ("No conversation history available\n");
+	R2AI_Messages *messages = r2ai_conversation_get (state);
+	if (!messages || r_list_length (messages->messages) == 0) {
+		r_cons_printf (core->cons, "No conversation history available\n");
 		return;
 	}
 
@@ -326,88 +376,79 @@ R_IPI void cmd_r2ai_logs(RCore *core) {
 
 		pj_a (pj);
 
-		for (int i = 0; i < messages->n_messages; i++) {
-			const R2AI_Message *msg = &messages->messages[i];
-
+		for (int i = 0; i < r_list_length (messages->messages); i++) {
+			const R2AI_Message *msg = r_list_get_n (messages->messages, i);
 			pj_o (pj);
-
-			pj_ks (pj, "role", msg->role ? msg->role : "unknown");
-			pj_ks (pj, "content", msg->content ? msg->content : "");
-
+			pj_ks (pj, "role", msg->role? msg->role: "unknown");
+			pj_ks (pj, "content", msg->content? msg->content: "");
 			if (msg->tool_calls && msg->n_tool_calls > 0) {
 				pj_ka (pj, "tool_calls");
-
 				for (int j = 0; j < msg->n_tool_calls; j++) {
 					const R2AI_ToolCall *tc = &msg->tool_calls[j];
-
 					pj_o (pj);
-
 					if (tc->name) {
 						pj_ks (pj, "name", tc->name);
 					}
 					if (tc->arguments) {
 						pj_ks (pj, "arguments", tc->arguments);
 					}
-
 					pj_end (pj);
 				}
-
 				pj_end (pj);
 			}
-
 			pj_end (pj);
 		}
 
 		pj_end (pj);
 
 		char *json_str = pj_drain (pj);
-		R2_PRINTF ("%s\n", json_str);
+		r_cons_printf (core->cons, "%s\n", json_str);
 		free (json_str);
 
 		return;
 	}
 
-	R2_PRINTF ("\x1b[1;34m[r2ai] Chat Logs (%d messages)\x1b[0m\n",
-		core->cons, messages->n_messages);
+	r_cons_printf (core->cons, "\x1b[1;34m[r2ai] Chat Logs (%d messages)\x1b[0m\n",
+		r_list_length (messages->messages));
 
-	R2_PRINTF ("\x1b[1;33mNote: System prompt is applied automatically but not stored in history\x1b[0m\n\n");
+	r_cons_printf (core->cons, "\x1b[1;33mNote: System prompt is applied automatically but not stored in history\x1b[0m\n\n");
 
 	// Display each message in the conversation
-	for (int i = 0; i < messages->n_messages; i++) {
-		const R2AI_Message *msg = &messages->messages[i];
+	for (int i = 0; i < r_list_length (messages->messages); i++) {
+		const R2AI_Message *msg = r_list_get_n (messages->messages, i);
 		const char *role = msg->role;
 
 		// Format based on role
 		if (!strcmp (role, "user")) {
-			R2_PRINTF ("\x1b[1;32m[user]:\x1b[0m ");
+			r_cons_printf (core->cons, "\x1b[1;32m[user]:\x1b[0m ");
 			print_content_with_length (core, msg->content, "<no content>", false);
 		} else if (!strcmp (role, "assistant")) {
-			R2_PRINTF ("\x1b[1;36m[assistant]:\x1b[0m ");
+			r_cons_printf (core->cons, "\x1b[1;36m[assistant]:\x1b[0m ");
 			print_content_with_length (core, msg->content, "<no content>", false);
 			// Show tool calls if present
 			if (msg->tool_calls && msg->n_tool_calls > 0) {
 				for (int j = 0; j < msg->n_tool_calls; j++) {
 					const R2AI_ToolCall *tc = &msg->tool_calls[j];
-					R2_PRINTF ("  \x1b[1;35m[tool call]:\x1b[0m %s\n",
-						tc->name ? tc->name : "<unnamed>");
+					r_cons_printf (core->cons, "  \x1b[1;35m[tool call]:\x1b[0m %s\n",
+						tc->name? tc->name: "<unnamed>");
 
 					if (tc->arguments) {
-						R2_PRINTF ("    %s\n", tc->arguments);
+						r_cons_printf (core->cons, "    %s\n", tc->arguments);
 					}
 				}
 			}
 		} else if (!strcmp (role, "tool")) {
-			R2_PRINTF ("\x1b[1;35m[tool]:\x1b[0m ");
+			r_cons_printf (core->cons, "\x1b[1;35m[tool]:\x1b[0m ");
 			print_content_with_length (core, msg->content, "<no result>", true);
 
 			// Don't show the tool call ID as requested
 		} else {
 			// Other roles (system, etc.)
-			R2_PRINTF ("\x1b[1;37m[%s]:\x1b[0m ", role);
+			r_cons_printf (core->cons, "\x1b[1;37m[%s]:\x1b[0m ", role);
 			print_content_with_length (core, msg->content, "<no content>", false);
 		}
 
-		R2_NEWLINE ();
-		R2_FLUSH ();
+		r_cons_newline (core->cons);
+		r_cons_flush (core->cons);
 	}
 }

src/indent.py

@@ -1,39 +0,0 @@
-import re
-import sys
-import subprocess
-
-
-arg = sys.argv[1]
-# Step 1: Format the file in-place using clang-format
-subprocess.run(["clang-format", "-i", arg], check=True)
-
-def is_function(s):
-    return s and not s[0].isspace()
-
-def is_control_structure(s):
-    return s in {"if", "for", "while", "switch", "catch", "return"}
-
-def fix_line(line):
-    # Skip lines that are empty or only whitespace
-    if not line.strip():
-        return line
-
-    # Match function calls like: foo(bar) => foo (bar)
-    # Avoid if/for/while/catch/return and function *definitions*
-    pattern = r'\b([a-zA-Z_]\w*)\('
-
-    def replacer(match):
-        name = match.group(1)
-        if is_control_structure(name) or is_function(line):
-            return match.group(0)  # No change
-        return f'{name} ('
-
-    return re.sub(pattern, replacer, line)
-
-# Step 2: Read the file, transform it, and write it back
-with open(arg, "r", encoding="utf-8") as f:
-    lines = f.readlines()
-
-with open(arg, "w", encoding="utf-8") as f:
-    for line in lines:
-        f.write(fix_line(line))

src/llm.c

@@ -0,0 +1,364 @@
+#define R_LOG_ORIGIN "llm"
+
+#include "r2ai.h"
+#include "r2ai_priv.h"
+#include <string.h>
+
+static const R2AIProvider r2ai_providers[] = {
+	{ "openai", "https://api.openai.com/v1", true, false, false, false },
+	{ "gemini", "https://generativelanguage.googleapis.com/v1beta/openai", true, false, false, false },
+	{ "anthropic", "https://api.anthropic.com/v1", true, true, false, false },
+	{ "ollama", "http://localhost:11434/api", false, false, true, true },
+	{ "openapi", "http://127.0.0.1:11434", false, false, false, false },
+	{ "xai", "https://api.x.ai/v1", true, false, false, false },
+	{ "openrouter", "https://openrouter.ai/api/v1", true, false, false, false },
+	{ "groq", "https://api.groq.com/openai/v1", true, false, false, false },
+	{ "mistral", "https://api.mistral.ai/v1", true, false, false, false },
+	{ NULL, NULL, false, false, false, false } // sentinel
+};
+
+R_IPI const R2AIProvider *r2ai_get_provider(const char *name) {
+	if (R_STR_ISEMPTY (name)) {
+		return NULL;
+	}
+	for (int i = 0; r2ai_providers[i].name; i++) {
+		if (!strcmp (name, r2ai_providers[i].name)) {
+			return &r2ai_providers[i];
+		}
+	}
+	return NULL;
+}
+
+// Forward declaration for rawtools
+R_IPI R2AI_ChatResponse *r2ai_llmcall(RCorePluginSession *cps, R2AIArgs args) {
+	RCore *core = cps->core;
+
+	// 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 && args.tools->n_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: "");
+	}
+	if (!args.provider) {
+		args.provider = strdup (provider);
+	}
+
+	if (!args.max_tokens) {
+		args.max_tokens = r_config_get_i (core->config, "r2ai.max_tokens");
+	}
+	if (!args.temperature) {
+		const char *configtemp = r_config_get (core->config, "r2ai.temperature");
+		args.temperature = configtemp? atof (configtemp): 0;
+	}
+
+	const R2AIProvider *prov = r2ai_get_provider (provider);
+	if (!prov) {
+		R_LOG_ERROR ("Unknown provider: %s", provider);
+		return NULL;
+	}
+
+	char *api_key = NULL;
+	if (prov->requires_api_key) {
+		api_key = r2ai_get_api_key (core, 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 %s provider. Please set one with: r2ai "
+				"%s_API_KEY=YOUR_KEY",
+				provider, Provider);
+			free (Provider);
+			return NULL;
+		}
+	}
+
+	// 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");
+	}
+	if (!args.messages) {
+		args.messages = r2ai_msgs_new ();
+	}
+	int context_pullback = -1;
+	// 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");
+		if (!state->db) {
+			state->db = r_vdb_new (R2AI_DEFAULT_VECTORS);
+			r2ai_refresh_embeddings (cps);
+		}
+		RStrBuf *sb = r_strbuf_new ("");
+		r_strbuf_appendf (sb, "\n ## Query\n\n%s\n ## Context\n", args.input);
+		RVdbResultSet *rs = r_vdb_query (state->db, args.input, K);
+		if (rs) {
+			int i;
+			for (i = 0; i < rs->size; i++) {
+				RVdbResult *r = &rs->results[i];
+				KDNode *n = r->node;
+				r_strbuf_appendf (sb, "- %s.\n", n->text);
+			}
+			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->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);
+	}
+
+	args.thinking_tokens = r_config_get_i (core->config, "r2ai.thinking_tokens");
+
+	const R2AIProvider *p = r2ai_get_provider (provider);
+	if (p && p->uses_anthropic_header) {
+		res = r2ai_anthropic (cps, args);
+	} else {
+		res = r2ai_openai (cps, args);
+	}
+	if (context_pullback != -1) {
+		R2AI_Message *msg = r_list_get_n (args.messages->messages, context_pullback);
+		free ((char *)msg->content);
+		msg->content = strdup (args.input);
+	}
+	if (*args.error) {
+		R_LOG_ERROR ("%s", *args.error);
+		free (*args.error);
+		*args.error = NULL;
+	}
+
+	return res;
+}
+
+/* Return a malloc'd API key read from the environment or from ~/.r2ai.<provider>-key
+ * Caller is responsible for freeing the returned string (or NULL if not found). */
+R_IPI char *r2ai_get_api_key(RCore *core, const char *provider) {
+	(void)core;
+	char *api_key = NULL;
+	char *api_key_env = r_str_newf ("%s_API_KEY", provider);
+	r_str_case (api_key_env, true);
+	char *s = r_sys_getenv (api_key_env);
+	free (api_key_env);
+	if (R_STR_ISNOTEMPTY (s)) {
+		api_key = s;
+	} else {
+		free (s);
+		char *api_key_filename = r_str_newf ("~/.r2ai.%s-key", provider);
+		char *absolute_apikey = r_file_abspath (api_key_filename);
+		if (r_file_exists (absolute_apikey)) {
+			api_key = r_file_slurp (absolute_apikey, NULL);
+			if (api_key) {
+				r_str_trim (api_key);
+			}
+		}
+		free (api_key_filename);
+		free (absolute_apikey);
+	}
+	return api_key;
+}
+
+R_IPI const 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 (!strcmp (provider, "openai") || !strcmp (provider, "ollama")) {
+		const char *host = r_config_get (core->config, "r2ai.baseurl");
+		if (R_STR_ISNOTEMPTY (host)) {
+			if (r_str_startswith (host, "http")) {
+				if (!strcmp (provider, "openai")) {
+					return r_str_newf ("%s/v1", host);
+				} else {
+					return r_str_newf ("%s/api", host);
+				}
+			}
+			if (!strcmp (provider, "openai")) {
+				return r_str_newf ("http://%s/v1", host);
+			} else {
+				return r_str_newf ("http://%s/api", host);
+			}
+		}
+	}
+
+	return p->url;
+}
+R_IPI RList *r2ai_fetch_available_models(RCore *core, const char *provider) {
+	const char *purl = r2ai_get_provider_url (core, provider);
+	if (!purl) {
+		return NULL;
+	}
+
+	// Create models endpoint URL
+	char *models_url = NULL;
+	const R2AIProvider *prov = r2ai_get_provider (provider);
+	if (prov && prov->uses_tags_endpoint) {
+		models_url = r_str_newf ("%s/tags", purl);
+	} else {
+		models_url = r_str_newf ("%s/models", purl);
+	}
+	if (!models_url) {
+		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_get_api_key (core, provider);
+	}
+
+	int code = 0;
+	char *response = NULL;
+	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->uses_anthropic_header) {
+			// 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);
+		free (api_key);
+	} else {
+		// We have no headers
+		R_LOG_DEBUG ("GET %s", models_url);
+		response = r2ai_http_get (core, models_url, NULL, &code, NULL);
+	}
+
+	free (models_url);
+
+	if (!response || code != 200) {
+		R_LOG_DEBUG ("Failed to fetch models from %s (code: %d)", provider, code);
+		free (response);
+		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;
+		const R2AIProvider *prov = r2ai_get_provider (provider);
+		if (prov && prov->uses_tags_endpoint) {
+			data = r_json_get (json, "models");
+		} else {
+			data = r_json_get (json, "data");
+		}
+
+		if (data && data->type == R_JSON_ARRAY) {
+			const RJson *model_item = data->children.first;
+			while (model_item) {
+				const RJson *id;
+				if (prov && prov->uses_tags_endpoint) {
+					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)) {
+					R_LOG_DEBUG ("Model: %s", id->str_value);
+					r_list_append (models, strdup (id->str_value));
+				}
+				model_item = model_item->next;
+			}
+		}
+		r_json_free (json);
+	}
+
+	free (response);
+	return models;
+}
+
+R_IPI void r2ai_list_providers(RCore *core, RStrBuf *sb) {
+	size_t i;
+	for (i = 0; r2ai_providers[i].name; i++) {
+		if (sb) {
+			if (i > 0) {
+				r_strbuf_append (sb, ", ");
+			}
+			r_strbuf_append (sb, r2ai_providers[i].name);
+		} else {
+			r_cons_println (core->cons, r2ai_providers[i].name);
+		}
+	}
+}
+
+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");
+	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) {
+			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);
+	}
+	r_list_free (files);
+}

src/main.c

@@ -1,31 +1,114 @@
 /* Copyright r2ai - 2023-2025 - pancake */
 
-#include <r_core.h>
-#include "r2ai.h"
+#include "r2ai_priv.h"
 
-int main(int argc, char **argv) {
-	if (argc < 2) {
-		eprintf ("Usage: r2ai <prompt>\n");
-		return 1;
+static void show_help() {
+	printf ("Usage: r2ai [-vhp:m:q:E] <prompt>\n"
+	"  -v           Show version information\n"
+	"  -h           Show this help message\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"
+	"  -E           Edit the r2ai rc file\n");
+}
+
+static void show_version() {
+	printf ("r2ai " R2AI_VERSION "\n");
+}
+
+int main(int argc, const char **argv) {
+	int c;
+	const char *provider = NULL;
+	const char *model = NULL;
+	RList *queries = r_list_newf (free);
+
+	RGetopt opt;
+	r_getopt_init (&opt, argc, argv, "vhp:m:q:E");
+	while ((c = r_getopt_next (&opt)) != -1) {
+		switch (c) {
+		case 'v':
+			show_version ();
+			r_list_free (queries);
+			return 0;
+		case 'h':
+			show_help ();
+			r_list_free (queries);
+			return 0;
+		case 'p':
+			provider = opt.arg;
+			break;
+		case 'm':
+			model = opt.arg;
+			break;
+		case 'q':
+			r_list_append (queries, strdup (opt.arg));
+			break;
+		case 'E':
+			{
+				RCore *core = r_core_new ();
+				RCorePluginSession cps = {
+					.core = core
+				};
+				r2ai_init (&cps);
+				char *rc_path = r_file_home (".config/r2ai/rc");
+				r_cons_editor (core->cons, rc_path, NULL);
+				free (rc_path);
+				r_list_free (queries);
+				r2ai_fini (&cps);
+				r_core_free (core);
+				return 0;
+			}
+		default:
+			show_help ();
+			r_list_free (queries);
+			return 1;
+		}
 	}
 
-	const char *prompt = argv[1];
-
 	RCore *core = r_core_new ();
-	char *err = NULL;
-
-	R2AIArgs args = {
-		.input = prompt,
-		.error = &err,
-		.dorag = false,
+	RCorePluginSession cps = {
+		.core = core
 	};
+	r2ai_init (&cps);
 
-	char *res = r2ai (core, args);
-	if (res) {
-		R_LOG_INFO (res);
-		free (res);
+	// Process queries if any
+	if (!r_list_empty (queries)) {
+		RListIter *iter;
+		char *query;
+		r_list_foreach (queries, iter, query) {
+			r2ai_cmd_q (&cps, query);
+		}
+	} else {
+		if (opt.ind >= argc) {
+			// List available prompts
+			r2ai_cmd_q (&cps, "");
+			r_list_free (queries);
+			r2ai_fini (&cps);
+			r_core_free (core);
+			return 0;
+		}
+
+		const char *prompt = argv[opt.ind];
+
+		char *err = NULL;
+		R2AIArgs args = {
+			.input = prompt,
+			.provider = provider,
+			.model = model,
+			.error = &err,
+			.dorag = false,
+		};
+		char *res = r2ai (&cps, args);
+		if (res) {
+			r_cons_println (core->cons, res);
+			free (res);
+		}
+		free (err);
+		r_cons_flush (core->cons);
 	}
-	free (err);
+
+	r_list_free (queries);
+	r2ai_fini (&cps);
 	r_core_free (core);
 	return 0;
 }

src/markdown.c

@@ -34,13 +34,13 @@ R_API RMarkdownTheme r2ai_markdown_theme_default(void) {
 		.heading4 = "\x1b[1;4;34m", // Bold, underlined, blue
 		.heading5 = "\x1b[1;4;35m", // Bold, underlined, magenta
 		.heading6 = "\x1b[1;4;36m", // Bold, underlined, cyan
-					    // List items
+					// List items
 		.list_bullet = "• ", // Bullet character
 		.list_number = "\x1b[1m%s.\x1b[0m ", // Bold number
-						     // Checkbox states
+						// Checkbox states
 		.checkbox_checked = "\x1b[32m[✓]\x1b[0m ", // Green checkmark
 		.checkbox_unchecked = "\x1b[90m[ ]\x1b[0m ", // Gray empty box
-							     // Reset code
+							// Reset code
 		.reset = "\x1b[0m"
 	};
 	return theme;
@@ -103,7 +103,7 @@ static void append_formatted(RStrBuf *sb, const char *text, int len, MarkdownSta
 static int count_indent(const char *str) {
 	int count = 0;
 	while (*str == ' ' || *str == '\t') {
-		count += (*str == '\t') ? 4 : 1;
+		count += (*str == '\t')? 4: 1;
 		str++;
 	}
 	return count;
@@ -257,7 +257,7 @@ R_API char *r2ai_markdown(const char *markdown) {
 					append_formatted (sb, 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);
+					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];
@@ -302,7 +302,7 @@ R_API char *r2ai_markdown(const char *markdown) {
 				if (p > start) {
 					append_formatted (sb, start, p - start, state);
 				}
-				p += double_marker ? 2 : 1;
+				p += double_marker? 2: 1;
 				start = p;
 
 				if (double_marker && marker == '*') {
@@ -317,7 +317,7 @@ 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);
-				p += double_marker ? 2 : 1;
+				p += double_marker? 2: 1;
 				start = p;
 				state = NORMAL;
 			} else {

src/meson.build

@@ -1,5 +1,5 @@
 project('r2ai', 'c',
-  version : '0.1.0',
+  version : '1.2.2',
   default_options : ['warning_level=2', 'c_std=c11']
 )
 
@@ -188,7 +188,10 @@ srcs = files(
   'r2ai.c',
   'auto.c',
   'vdb.c',
+  'llm.c',
   'tools.c',
+  'query.c',
+  'rawtools.c',
   'messages.c',
   'anthropic.c',
   'r2ai_http.c',

src/meson_options.txt

@@ -1,6 +1,6 @@
 option('r2_prefix',
   type: 'string',
-  value: '',
+  value: '/usr/local',
   description: 'Path prefix where radare2 is installed (include/ and lib/).'
 )
 

src/messages.c

@@ -2,12 +2,8 @@
 
 #include "r2ai.h"
 
-// AITODO: remove globals
-// Static global message store for session persistence
-static R2AI_Messages *conversation = NULL;
-
 #define INITIAL_CAPACITY 8
-#define GROWTH_FACTOR    1.5
+#define GROWTH_FACTOR 1.5
 
 R_API void r2ai_message_free(R2AI_Message *msg) {
 	if (!msg) {
@@ -15,7 +11,7 @@ R_API void r2ai_message_free(R2AI_Message *msg) {
 	}
 
 	free (msg->role);
-	free (msg->content);
+	free ((void *)msg->content); // maybe double free
 	free (msg->reasoning_content);
 	free (msg->tool_call_id);
 
@@ -46,24 +42,19 @@ R_API void r2ai_message_free(R2AI_Message *msg) {
 	}
 }
 
-// Initialize conversation container (call this during plugin init)
-R_API void r2ai_conversation_init(void) {
-	if (conversation) {
-		// Already initialized
+// 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
 		return;
 	}
 
-	conversation = R_NEW0 (R2AI_Messages);
-	conversation->cap_messages = INITIAL_CAPACITY;
-	conversation->messages = R_NEWS0 (R2AI_Message, conversation->cap_messages);
-	if (!conversation->messages) {
-		R_FREE (conversation);
-	}
+	state->conversation = r2ai_msgs_new ();
 }
 
-// Get the conversation instance (returns NULL if not initialized)
-R_API R2AI_Messages *r2ai_conversation_get(void) {
-	return conversation;
+R_API R2AI_Messages *r2ai_conversation_get(R2AI_State *state) {
+	return state? state->conversation: NULL;
 }
 
 // Create a new temporary messages container
@@ -72,12 +63,12 @@ R_API R2AI_Messages *r2ai_msgs_new(void) {
 	if (!msgs) {
 		return NULL;
 	}
-	msgs->cap_messages = INITIAL_CAPACITY;
-	msgs->messages = R_NEWS0 (R2AI_Message, msgs->cap_messages);
+	msgs->messages = r_list_new ();
 	if (!msgs->messages) {
 		R_FREE (msgs);
 		return NULL;
 	}
+	msgs->messages->free = (RListFree)r2ai_message_free;
 	return msgs;
 }
 
@@ -86,38 +77,19 @@ R_API void r2ai_msgs_free(R2AI_Messages *msgs) {
 		return;
 	}
 
-	// Don't actually free if it's our static instance
-	if (msgs == conversation) {
-		// Just clear the messages but keep the allocated structure
-		for (int i = 0; i < msgs->n_messages; i++) {
-			r2ai_message_free (&msgs->messages[i]);
-		}
-		msgs->n_messages = 0;
-		return;
-	}
-
 	// Normal free for non-static instances
 	if (msgs->messages) {
-		for (int i = 0; i < msgs->n_messages; i++) {
-			r2ai_message_free (&msgs->messages[i]);
-		}
-		R_FREE (msgs->messages);
+		r_list_free (msgs->messages);
 	}
 
 	R_FREE (msgs);
 }
 
 // Free the conversation when plugin is unloaded
-R_API void r2ai_conversation_free(void) {
-	if (conversation) {
-		if (conversation->messages) {
-			for (int i = 0; i < conversation->n_messages; i++) {
-				r2ai_message_free (&conversation->messages[i]);
-			}
-			R_FREE (conversation->messages);
-		}
-		R_FREE (conversation);
-		conversation = NULL;
+R_API void r2ai_conversation_free(R2AI_State *state) {
+	if (state && state->conversation) {
+		r2ai_msgs_free (state->conversation);
+		state->conversation = NULL;
 	}
 }
 
@@ -127,10 +99,7 @@ R_API void r2ai_msgs_clear(R2AI_Messages *msgs) {
 		return;
 	}
 
-	for (int i = 0; i < msgs->n_messages; i++) {
-		r2ai_message_free (&msgs->messages[i]);
-	}
-	msgs->n_messages = 0;
+	r_list_purge (msgs->messages);
 }
 
 R_API bool r2ai_msgs_add(R2AI_Messages *msgs, const R2AI_Message *msg) {
@@ -138,91 +107,79 @@ R_API bool r2ai_msgs_add(R2AI_Messages *msgs, const R2AI_Message *msg) {
 		return false;
 	}
 
-	// Check if we need to resize
-	if (msgs->n_messages >= msgs->cap_messages) {
-		int new_cap = msgs->cap_messages * GROWTH_FACTOR;
-		R2AI_Message *new_messages = realloc (msgs->messages, sizeof (R2AI_Message) * new_cap);
-		if (!new_messages) {
-			return false;
-		}
-		msgs->messages = new_messages;
-		msgs->cap_messages = new_cap;
-
-		// Zero the newly allocated portion
-		memset (&msgs->messages[msgs->n_messages], 0,
-			sizeof (R2AI_Message) * (msgs->cap_messages - msgs->n_messages));
+	R2AI_Message *new_msg = R_NEW0 (R2AI_Message);
+	if (!new_msg) {
+		return false;
 	}
 
-	// Copy the message to the array
-	R2AI_Message *dest = &msgs->messages[msgs->n_messages++];
-	dest->role = msg->role ? strdup (msg->role) : NULL;
-	dest->content = msg->content ? strdup (msg->content) : NULL;
-	dest->reasoning_content = msg->reasoning_content ? strdup (msg->reasoning_content) : NULL;
+	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;
 
 	if (msg->content_blocks) {
 		R2AI_ContentBlocks *cb = R_NEW0 (R2AI_ContentBlocks);
 		if (!cb) {
-			r2ai_message_free (dest);
-			msgs->n_messages--;
+			r2ai_message_free (new_msg);
+			free (new_msg);
 			return false;
 		}
 		cb->n_blocks = msg->content_blocks->n_blocks;
 		cb->blocks = R_NEWS0 (R2AI_ContentBlock, cb->n_blocks);
 		if (!cb->blocks) {
 			free (cb);
-			r2ai_message_free (dest);
-			msgs->n_messages--;
+			r2ai_message_free (new_msg);
+			free (new_msg);
 			return false;
 		}
 		for (int i = 0; i < cb->n_blocks; i++) {
 			R2AI_ContentBlock *src = &msg->content_blocks->blocks[i];
 			R2AI_ContentBlock *dst = &cb->blocks[i];
-			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;
+			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;
 		}
-		dest->content_blocks = cb;
+		new_msg->content_blocks = cb;
 	}
 
-	dest->tool_call_id = msg->tool_call_id ? strdup (msg->tool_call_id) : NULL;
-	dest->tool_calls = NULL;
-	dest->n_tool_calls = 0;
+	new_msg->tool_call_id = msg->tool_call_id? strdup (msg->tool_call_id): NULL;
+	new_msg->tool_calls = NULL;
+	new_msg->n_tool_calls = 0;
 
 	// Copy tool calls if any
 	if (msg->tool_calls && msg->n_tool_calls > 0) {
-		dest->tool_calls = R_NEWS0 (R2AI_ToolCall, msg->n_tool_calls);
-		if (!dest->tool_calls) {
-			// Clean up and return error
-			r2ai_message_free (dest);
-			msgs->n_messages--;
+		new_msg->tool_calls = R_NEWS0 (R2AI_ToolCall, msg->n_tool_calls);
+		if (!new_msg->tool_calls) {
+			r2ai_message_free (new_msg);
+			free (new_msg);
 			return false;
 		}
 
-		dest->n_tool_calls = msg->n_tool_calls;
+		new_msg->n_tool_calls = msg->n_tool_calls;
 		for (int i = 0; i < msg->n_tool_calls; i++) {
 			const R2AI_ToolCall *src_tc = &msg->tool_calls[i];
-			R2AI_ToolCall *dst_tc = (R2AI_ToolCall *)&dest->tool_calls[i];
+			R2AI_ToolCall *dst_tc = (R2AI_ToolCall *)&new_msg->tool_calls[i];
 
-			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;
+			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 (msgs->messages, new_msg);
 	return true;
 }
 
 R_API bool r2ai_msgs_add_tool_call(R2AI_Messages *msgs, const R2AI_ToolCall *tc) {
-	if (!msgs || !tc || msgs->n_messages == 0) {
+	if (!msgs || !tc || r_list_length (msgs->messages) == 0) {
 		return false;
 	}
 
-	R2AI_Message *msg = &msgs->messages[msgs->n_messages - 1];
+	R2AI_Message *msg = r_list_get_n (msgs->messages, r_list_length (msgs->messages) - 1);
 
 	// Allocate or resize the tool_calls array
 	if (msg->n_tool_calls == 0) {
@@ -233,7 +190,7 @@ R_API bool r2ai_msgs_add_tool_call(R2AI_Messages *msgs, const R2AI_ToolCall *tc)
 	} else {
 		R2AI_ToolCall *new_tool_calls = realloc (
 			(void *)msg->tool_calls,
-			sizeof (R2AI_ToolCall) * (msg->n_tool_calls + 1));
+			sizeof (R2AI_ToolCall) *(msg->n_tool_calls + 1));
 		if (!new_tool_calls) {
 			return false;
 		}
@@ -244,9 +201,9 @@ R_API bool r2ai_msgs_add_tool_call(R2AI_Messages *msgs, const R2AI_ToolCall *tc)
 
 	// Copy the tool call
 	R2AI_ToolCall *dst_tc = (R2AI_ToolCall *)&msg->tool_calls[msg->n_tool_calls];
-	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;
+	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;
 
 	msg->n_tool_calls++;
 	return true;
@@ -294,8 +251,8 @@ R_API bool r2ai_msgs_from_json(R2AI_Messages *msgs, const RJson *json) {
 
 	// 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.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;
 	new_msg.n_tool_calls = 0;
@@ -328,14 +285,14 @@ R_API bool r2ai_msgs_from_json(R2AI_Messages *msgs, const RJson *json) {
 			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;
+			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;
 		}
 		new_msg.content_blocks = cb;
 	}
@@ -366,21 +323,32 @@ R_API bool r2ai_msgs_from_json(R2AI_Messages *msgs, const RJson *json) {
 			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;
+			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)) {
 				break;
 			}
 		}
+
+		// Add the parsed message to the messages array
+		if (!r2ai_msgs_add (msgs, &new_msg)) {
+			r2ai_message_free (&new_msg);
+			return false;
+		}
+
+		// Free the temporary message struct (strings are duplicated in add)
+		r2ai_message_free (&new_msg);
+
+		return true;
 	}
 
 	return true;
 }
 
 R_API char *r2ai_msgs_to_json(const R2AI_Messages *msgs) {
-	if (!msgs || msgs->n_messages == 0) {
+	if (!msgs || r_list_length (msgs->messages) == 0) {
 		return NULL;
 	}
 
@@ -391,13 +359,13 @@ R_API char *r2ai_msgs_to_json(const R2AI_Messages *msgs) {
 
 	pj_a (pj); // Start array
 
-	for (int i = 0; i < msgs->n_messages; i++) {
-		const R2AI_Message *msg = &msgs->messages[i];
+	for (int i = 0; i < r_list_length (msgs->messages); i++) {
+		const R2AI_Message *msg = r_list_get_n (msgs->messages, i);
 
 		pj_o (pj); // Start message object
 
 		// Add role
-		pj_ks (pj, "role", msg->role ? msg->role : "user");
+		pj_ks (pj, "role", msg->role? msg->role: "user");
 
 		if (msg->content) {
 			pj_ks (pj, "content", msg->content);
@@ -435,7 +403,7 @@ R_API char *r2ai_msgs_to_json(const R2AI_Messages *msgs) {
 				pj_o (pj); // Start function object
 
 				// Add name
-				pj_ks (pj, "name", tc->name ? tc->name : "");
+				pj_ks (pj, "name", tc->name? tc->name: "");
 
 				// Add arguments if present
 				if (tc->arguments) {
@@ -459,7 +427,7 @@ R_API char *r2ai_msgs_to_json(const R2AI_Messages *msgs) {
 }
 
 R_API char *r2ai_msgs_to_anthropic_json(const R2AI_Messages *msgs) {
-	if (!msgs || msgs->n_messages == 0) {
+	if (!msgs || r_list_length (msgs->messages) == 0) {
 		return NULL;
 	}
 
@@ -470,14 +438,14 @@ R_API char *r2ai_msgs_to_anthropic_json(const R2AI_Messages *msgs) {
 
 	pj_a (pj); // Start array
 
-	for (int i = 0; i < msgs->n_messages; i++) {
-		const R2AI_Message *msg = &msgs->messages[i];
+	for (int i = 0; i < r_list_length (msgs->messages); i++) {
+		const R2AI_Message *msg = r_list_get_n (msgs->messages, i);
 
 		pj_o (pj); // Start message object
 
 		// Add role
-		const char *role = msg->role ? msg->role : "user";
-		pj_ks (pj, "role", strcmp (role, "tool") == 0 ? "user" : role);
+		const char *role = msg->role? msg->role: "user";
+		pj_ks (pj, "role", strcmp (role, "tool") == 0? "user": role);
 
 		if (msg->content_blocks) {
 			pj_ka (pj, "content"); // Start content array
@@ -550,12 +518,12 @@ R_API char *r2ai_msgs_to_anthropic_json(const R2AI_Messages *msgs) {
 					const R2AI_ToolCall *tc = &msg->tool_calls[j];
 					pj_o (pj); // Start tool_calls object
 					pj_ks (pj, "type", "tool_use");
-					pj_ks (pj, "id", tc->id ? tc->id : "");
-					pj_ks (pj, "name", tc->name ? tc->name : "");
+					pj_ks (pj, "id", tc->id? tc->id: "");
+					pj_ks (pj, "name", tc->name? tc->name: "");
 
 					// Create a non-const copy for r_json_parse
-					char *arguments_copy = tc->arguments ? strdup (tc->arguments) : NULL;
-					RJson *arguments = arguments_copy ? r_json_parse (arguments_copy) : NULL;
+					char *arguments_copy = tc->arguments? strdup (tc->arguments): NULL;
+					RJson *arguments = arguments_copy? r_json_parse (arguments_copy): NULL;
 
 					pj_ko (pj, "input"); // Start input object
 					if (arguments) {
@@ -588,7 +556,7 @@ R_API char *r2ai_msgs_to_anthropic_json(const R2AI_Messages *msgs) {
 
 // Function to delete the last N messages from conversation history
 R_API void r2ai_delete_last_messages(R2AI_Messages *messages, int n) {
-	if (!messages || messages->n_messages == 0) {
+	if (!messages || r_list_length (messages->messages) == 0) {
 		return;
 	}
 
@@ -598,17 +566,15 @@ R_API void r2ai_delete_last_messages(R2AI_Messages *messages, int n) {
 	}
 
 	// Make sure we don't try to delete more messages than exist
-	if (n > messages->n_messages) {
-		n = messages->n_messages;
+	int len = r_list_length (messages->messages);
+	if (n > len) {
+		n = len;
 	}
 
-	// Free the last n messages
-	for (int i = messages->n_messages - n; i < messages->n_messages; i++) {
-		r2ai_message_free (&messages->messages[i]);
+	// Pop the last n messages
+	for (int i = 0; i < n; i++) {
+		r_list_pop (messages->messages);
 	}
-
-	// Update the message count
-	messages->n_messages -= n;
 }
 
 // Helper function to convert RJson to PJ without draining
@@ -617,7 +583,7 @@ R_API PJ *r_json_to_pj(const RJson *json, PJ *existing_pj) {
 		return existing_pj;
 	}
 
-	PJ *pj = existing_pj ? existing_pj : pj_new ();
+	PJ *pj = existing_pj? existing_pj: pj_new ();
 	if (!pj) {
 		return NULL;
 	}
@@ -759,23 +725,3 @@ R_API char *r_json_to_string(const RJson *json) {
 	PJ *pj = r_json_to_pj (json, NULL);
 	return pj? pj_drain (pj): NULL;
 }
-
-// Create a conversation with optional initial user message
-R_API R2AI_Messages *create_conversation(const char *user_message) {
-	// Create a temporary message container (not using static storage)
-	R2AI_Messages *msgs = r2ai_msgs_new ();
-	if (!msgs) {
-		return NULL;
-	}
-
-	// Add user message if provided (no system message - that's added during processing)
-	if (R_STR_ISNOTEMPTY (user_message)) {
-		R2AI_Message user_msg = {
-			.role = "user",
-			.content = user_message
-		};
-		r2ai_msgs_add (msgs, &user_msg);
-	}
-
-	return msgs;
-}

src/openai.c

@@ -20,37 +20,19 @@ typedef struct {
 	int error_flags; // Bitfield of ModelErrorFlags
 } ModelCompat;
 
-// Hash table to store model compatibility info
-static HtPP *model_compat_db = NULL;
-
-#if 0
-// Function to check if a model has a specific error flag
-static bool model_has_error(const char *provider, const char *model, ModelErrorFlags flag) {
-	if (!model_compat_db) {
-		return false;
-	}
-
-	char *key = r_str_newf ("%s:%s", provider, model ? model : "default");
-	bool found_flag = false;
-	ModelCompat *compat = ht_pp_find (model_compat_db, key, &found_flag);
-	free (key);
-
-	if (found_flag && compat) {
-		return (compat->error_flags & flag) != 0;
-	}
-	return false;
-}
-#endif
-
 // Function to add an error flag to a model
-static void model_add_error(const char *provider, const char *model, ModelErrorFlags flag) {
-	if (!model_compat_db) {
-		model_compat_db = ht_pp_new0 ();
+static void model_add_error(R2AI_State *state, const char *provider, const char *model, ModelErrorFlags flag) {
+	if (!state) {
+		return;
 	}
 
-	char *key = r_str_newf ("%s:%s", provider, model ? model : "default");
+	if (!state->model_compat_db) {
+		state->model_compat_db = ht_pp_new0 ();
+	}
+
+	char *key = r_str_newf ("%s:%s", provider, model? model: "default");
 	bool found_flag = false;
-	ModelCompat *compat = ht_pp_find (model_compat_db, key, &found_flag);
+	ModelCompat *compat = ht_pp_find (state->model_compat_db, key, &found_flag);
 
 	if (found_flag && compat) {
 		// Update existing entry
@@ -60,7 +42,7 @@ static void model_add_error(const char *provider, const char *model, ModelErrorF
 		compat = R_NEW0 (ModelCompat);
 		compat->model_id = strdup (key);
 		compat->error_flags = flag;
-		ht_pp_insert (model_compat_db, key, compat);
+		ht_pp_insert (state->model_compat_db, key, compat);
 	}
 	free (key);
 }
@@ -78,23 +60,27 @@ static bool model_compat_free_cb(void *user, const void *k, const void *v) {
 }
 
 // Function to free the model_compat_db hash table
-R_IPI void r2ai_openai_fini(void) {
-	if (model_compat_db) {
-		ht_pp_foreach (model_compat_db, model_compat_free_cb, NULL);
-		ht_pp_free (model_compat_db);
-		model_compat_db = NULL;
+R_IPI void r2ai_openai_fini(R2AI_State *state) {
+	if (state && state->model_compat_db) {
+		ht_pp_foreach (state->model_compat_db, model_compat_free_cb, NULL);
+		ht_pp_free (state->model_compat_db);
+		state->model_compat_db = NULL;
 	}
 }
 
-R_IPI R2AI_ChatResponse *r2ai_openai(RCore *core, R2AIArgs args) {
+R_IPI R2AI_ChatResponse *r2ai_openai(RCorePluginSession *cps, R2AIArgs args) {
+	RCore *core = cps->core;
+	R2AI_State *state = cps->data;
 	// Initialize compatibility database if needed
-	if (!model_compat_db) {
-		model_compat_db = ht_pp_new0 ();
+	if (!state->model_compat_db) {
+		state->model_compat_db = ht_pp_new0 ();
 	}
+	args.provider = r_config_get (core->config, "r2ai.api");
+	args.model = r_config_get (core->config, "r2ai.model");
 
 	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";
+	const char *model_name = args.model? args.model: "gpt-4o-mini";
 	char **error = args.error;
 	const R2AI_Tools *tools = args.tools;
 	// create a temp conversation to include the system prompt and the rest of the messages
@@ -107,7 +93,7 @@ R_IPI R2AI_ChatResponse *r2ai_openai(RCore *core, R2AIArgs args) {
 	}
 	R2AI_Message system_msg = {
 		.role = "system",
-		.content = args.system_prompt
+		.content = (char *)args.system_prompt
 	};
 	// Add system message if available from args.system_prompt
 	if (R_STR_ISNOTEMPTY (args.system_prompt)) {
@@ -115,10 +101,10 @@ R_IPI R2AI_ChatResponse *r2ai_openai(RCore *core, R2AIArgs args) {
 		// 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 = args.system_prompt;
+			system_msg.content = (char *)args.system_prompt;
 		} else {
 			system_msg.role = "system";
-			system_msg.content = args.system_prompt;
+			system_msg.content = (char *)args.system_prompt;
 		}
 		r2ai_msgs_add (temp_msgs, &system_msg);
 	} else {
@@ -131,20 +117,20 @@ R_IPI R2AI_ChatResponse *r2ai_openai(RCore *core, R2AIArgs args) {
 			} else {
 				system_msg.role = "system";
 			}
-			system_msg.content = sysprompt;
+			system_msg.content = (char *)sysprompt;
 			r2ai_msgs_add (temp_msgs, &system_msg);
 		}
 	}
 	if (args.messages) {
-		for (int i = 0; i < args.messages->n_messages; i++) {
-			r2ai_msgs_add (temp_msgs, &args.messages->messages[i]);
+		for (int i = 0; i < r_list_length (args.messages->messages); i++) {
+			r2ai_msgs_add (temp_msgs, r_list_get_n (args.messages->messages, i));
 		}
 	} else {
 		R_LOG_WARN ("No messages");
 	}
 	// Safely print debug info about first message
-	if (temp_msgs && temp_msgs->n_messages > 0 && temp_msgs->messages && temp_msgs->messages[0].role) {
-		R_LOG_DEBUG ("First message role: %s", temp_msgs->messages[0].role);
+	if (temp_msgs && r_list_length (temp_msgs->messages) > 0 && ((R2AI_Message *)r_list_get_n (temp_msgs->messages, 0))->role) {
+		R_LOG_DEBUG ("First message role: %s", ((R2AI_Message *)r_list_get_n (temp_msgs->messages, 0))->role);
 	}
 	if (error) {
 		*error = NULL;
@@ -160,20 +146,16 @@ R_IPI R2AI_ChatResponse *r2ai_openai(RCore *core, R2AIArgs args) {
 		headers[0] = "Content-Type: application/json";
 		headers[1] = auth_header;
 	}
-
-	const char *openai_url;
-	if (strcmp (args.provider, "ollama") == 0) {
-		openai_url = r_str_newf ("%s/chat", base_url);
-	} else {
-		openai_url = r_str_newf ("%s/chat/completions", base_url);
-	}
-	R_LOG_DEBUG ("OpenAI URL: %s", openai_url);
+	const char *urlfmt = strcmp (args.provider, "ollama")
+		? "%s/chat/completions"
+		: "%s/chat";
+	char *openai_url = r_str_newf (urlfmt, base_url);
 
 	// Create a messages JSON object, either from input messages or from content
 	char *messages_json = NULL;
 
-	if (temp_msgs && temp_msgs->n_messages > 0) {
-		R_LOG_DEBUG ("Using input messages: %d messages", temp_msgs->n_messages);
+	if (temp_msgs && r_list_length (temp_msgs->messages) > 0) {
+		R_LOG_DEBUG ("Using input messages: %d messages", r_list_length (temp_msgs->messages));
 		messages_json = r2ai_msgs_to_json (temp_msgs);
 		if (!messages_json) {
 			if (error) {
@@ -202,19 +184,31 @@ R_IPI R2AI_ChatResponse *r2ai_openai(RCore *core, R2AIArgs args) {
 	pj_ks (pj, "model", model_name);
 	pj_kb (pj, "stream", false);
 
+	if (strcmp (args.provider, "ollama") == 0) {
+		// 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);
-	}
+		// 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);
-	} else {
-		pj_kn (pj, "max_completion_tokens", args.max_tokens ? args.max_tokens : 5128);
+		if (strcmp (args.provider, "mistral") == 0) {
+			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);
+		}
 	}
 
 	pj_end (pj);
@@ -265,8 +259,13 @@ R_IPI R2AI_ChatResponse *r2ai_openai(RCore *core, R2AIArgs args) {
 
 	// Save the full JSON to a file for inspection
 	// XXX: only create request/response files when r2ai.debug is set
-	r_file_dump ("/tmp/r2ai_openai_request.json", (const ut8 *)complete_json, strlen (complete_json), 0);
-	R_LOG_DEBUG ("Full request saved to /tmp/r2ai_openai_request.json");
+	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);
 
 	// Make the API call
@@ -274,6 +273,7 @@ R_IPI R2AI_ChatResponse *r2ai_openai(RCore *core, R2AIArgs args) {
 	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);
@@ -281,30 +281,25 @@ R_IPI R2AI_ChatResponse *r2ai_openai(RCore *core, R2AIArgs args) {
 			R_LOG_ERROR ("OpenAI API error response: %s", res);
 			// Check for specific error types in the response
 			ModelErrorFlags error_flag = MODEL_ERROR_NONE;
-			const char *model_name = args.model ? args.model : "gpt-5-mini";
 
 			// Check for temperature errors
 			if (strstr (res, "temperature")) {
-				R_LOG_DEBUG ("Detected temperature error for %s model %s", args.provider, model_name);
+				R_LOG_DEBUG ("Detected temperature error for %s model %s", args.provider, args.model);
 				error_flag |= MODEL_ERROR_TEMPERATURE;
 			}
 
-			// Add more error type checks as needed
-			// if (strstr (res, "top_p")) {
-			//     error_flag |= MODEL_ERROR_TOP_P;
-			// }
-
 			if (error_flag != MODEL_ERROR_NONE) {
 				// Record the error flags for this provider/model
-				model_add_error (args.provider, model_name, error_flag);
+				model_add_error (state, args.provider, args.model, error_flag);
 
 				// Clean up
 				free (auth_header);
 				free (res);
 
 				// Retry the call (it will skip problematic parameters this time)
-				R_LOG_INFO ("Retrying request with adjusted parameters for %s/%s", args.provider, model_name);
-				return r2ai_openai (core, args);
+				R_LOG_INFO ("Retrying request with adjusted parameters for %s/%s", args.provider, args.model);
+				RCorePluginSession retry_cps = { .core = core, .data = state };
+				return r2ai_openai (&retry_cps, args);
 			}
 		}
 		free (auth_header);
@@ -313,8 +308,13 @@ R_IPI R2AI_ChatResponse *r2ai_openai(RCore *core, R2AIArgs args) {
 	}
 
 	// Save the response for inspection
-	r_file_dump ("/tmp/r2ai_openai_response.json", (const ut8 *)res, strlen (res), 0);
-	R_LOG_DEBUG ("OpenAI API response saved to /tmp/r2ai_openai_response.json");
+	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);
+	R_LOG_DEBUG ("OpenAI API response saved to %s", res_path);
+	free (res_path);
+	free (tmpdir);
+
 	R_LOG_DEBUG ("OpenAI API response: %s", res);
 
 	// Parse the response into our messages structure
@@ -327,84 +327,75 @@ R_IPI R2AI_ChatResponse *r2ai_openai(RCore *core, R2AIArgs args) {
 		R2AI_Usage *usage = R_NEW0 (R2AI_Usage);
 		if (message) {
 			// Process the response using our r2ai_msgs_from_json logic
-			const RJson *usage_json = r_json_get (jres, "usage");
-			if (usage_json && usage_json->type == R_JSON_OBJECT) {
-				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 *usage_json = NULL;
+			if (strcmp (args.provider, "ollama") == 0) {
+				// Ollama has usage info at top level
+				usage_json = jres;
+			} else {
+				usage_json = r_json_get (jres, "usage");
 			}
-			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) {
-					const RJson *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");
 
-						// 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) {
-							message->reasoning_content = strdup (reasoning_content->str_value);
-						}
-
-						// Handle tool calls if present
-						const RJson *tool_calls = r_json_get (message_json, "tool_calls");
-						if (tool_calls && tool_calls->type == R_JSON_ARRAY && tool_calls->children.count > 0) {
-							int n_tool_calls = tool_calls->children.count;
-							R2AI_ToolCall *tool_calls_array = R_NEWS0 (R2AI_ToolCall, n_tool_calls);
-
-							if (tool_calls_array) {
-								for (int i = 0; i < n_tool_calls; 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");
-
-									if (id && id->type == R_JSON_STRING) {
-										tool_calls_array[i].id = strdup (id->str_value);
-									}
-
-									if (name && name->type == R_JSON_STRING) {
-										tool_calls_array[i].name = strdup (name->str_value);
-									}
-
-									if (arguments && arguments->type == R_JSON_STRING) {
-										tool_calls_array[i].arguments = strdup (arguments->str_value);
-									}
-								}
-
-								message->tool_calls = tool_calls_array;
-								message->n_tool_calls = n_tool_calls;
-							}
-						}
+			if (usage_json && usage_json->type == R_JSON_OBJECT) {
+				if (strcmp (args.provider, "ollama") == 0) {
+					// 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 (strcmp (args.provider, "ollama") == 0) {
+				// 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");
+
+				// 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) {
+					message->reasoning_content = strdup (reasoning_content->str_value);
+				}
+				// TODO: Handle tool calls if present?
+			}
 		}
 		r_json_free (jres);
 		R2AI_ChatResponse *result = R_NEW0 (R2AI_ChatResponse);

src/query.c

@@ -0,0 +1,346 @@
+/* r2ai - Copyright 2023-2025 pancake */
+
+#include "r2ai.h"
+
+R_API void r2aiprompt_free(R2AIPrompt *prompt) {
+	if (!prompt) {
+		return;
+	}
+	free (prompt->title);
+	free (prompt->author);
+	free (prompt->desc);
+	free (prompt->command);
+	free (prompt->prompt);
+	free (prompt->requires);
+	free (prompt->if_empty);
+	free (prompt->if_command);
+	free (prompt->model);
+	free (prompt->provider);
+	free (prompt);
+}
+
+static R2AIPrompt *parse_prompt_file(const char *filepath) {
+	R2AIPrompt *prompt = R_NEW0 (R2AIPrompt);
+	char *content = r_file_slurp (filepath, NULL);
+	if (!content) {
+		free (prompt);
+		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 {
+		// Original parsing for .txt files
+		RList *lines = r_str_split_list (content, "\n", -1);
+		RListIter *iter;
+		char *line;
+		r_list_foreach (lines, iter, line) {
+			r_str_trim (line);
+			if (*line == '#' || *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->title = strdup (value);
+			} else if (!strcmp (key, "Author")) {
+				prompt->author = strdup (value);
+			} else if (!strcmp (key, "Description")) {
+				prompt->desc = strdup (value);
+			} else if (!strcmp (key, "Command") || !strcmp (key, "Commands")) {
+				prompt->command = strdup (value);
+			} else if (!strcmp (key, "Prompt") || !strcmp (key, "Query")) {
+				prompt->prompt = strdup (value);
+			} else if (!strcmp (key, "Depends")) {
+				prompt->
+					requires
+				= strdup (value);
+			} else if (!strcmp (key, "If-Empty")) {
+				prompt->if_empty = strdup (value);
+			} else if (!strcmp (key, "If-Command")) {
+				prompt->if_command = strdup (value);
+			}
+		}
+		r_list_free (lines);
+	}
+
+	free (content);
+	return prompt;
+}
+
+static char *run_commands(RCore *core, const char *commands) {
+	if (!commands) {
+		return NULL;
+	}
+	RStrBuf *sb = r_strbuf_new ("");
+	char *cmds_str = strdup (commands);
+	RList *cmds = r_str_split_list (cmds_str, ";", -1);
+	RListIter *iter;
+	char *cmd;
+	r_list_foreach (cmds, iter, cmd) {
+		r_str_trim (cmd);
+		if (*cmd) {
+			char *out = r_core_cmd_str (core, cmd);
+			r_strbuf_append (sb, out);
+			free (out);
+		}
+	}
+	r_list_free (cmds);
+	free (cmds_str);
+	return r_strbuf_drain (sb);
+}
+
+static char *replace_vars(RCore *core, const char *text) {
+	if (!text) {
+		return NULL;
+	}
+	RStrBuf *sb = r_strbuf_new ("");
+	const char *p = text;
+	while (*p) {
+		if (*p == '$' && *(p + 1) == '{') {
+			p += 2;
+			const char *end = strchr (p, '}');
+			if (end) {
+				char *var = r_str_ndup (p, end - p);
+				char *val = getenv (var);
+				if (val) {
+					r_strbuf_append (sb, val);
+				}
+				free (var);
+				p = end + 1;
+			} else {
+				r_strbuf_append (sb, "${");
+				p += 2;
+			}
+		} else if (*p == '$' && *(p + 1) == '(') {
+			p += 2;
+			const char *end = strchr (p, ')');
+			if (end) {
+				char *cmd = r_str_ndup (p, end - p);
+				char *val = r_core_cmd_str (core, cmd);
+				r_str_trim (val);
+				r_strbuf_append (sb, val);
+				free (val);
+				free (cmd);
+				p = end + 1;
+			} else {
+				r_strbuf_append (sb, "$(");
+				p += 2;
+			}
+		} else {
+			r_strbuf_append_n (sb, p, 1);
+			p++;
+		}
+	}
+	return r_strbuf_drain (sb);
+}
+
+R_API void r2ai_cmd_q(RCorePluginSession *cps, const char *input) {
+	RCore *core = cps->core;
+	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;
+	}
+	if (r_list_empty (search_dirs)) {
+		R_LOG_ERROR ("No prompt directories found");
+		r_list_free (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);
+		}
+	} else {
+		// run prompt
+		char *name = strdup (input);
+		char *extra = strchr (name, ' ');
+		if (extra) {
+			*extra++ = 0;
+			r_str_trim (extra);
+		}
+		r_str_trim (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);
+				break;
+			}
+			free (filepath_txt);
+			free (filepath_md);
+		}
+		if (!filepath) {
+			R_LOG_ERROR ("Cannot find prompt file: %s.r2ai.txt or %s.r2ai.md", name, name);
+			r_list_free (search_dirs);
+			free (name);
+			return;
+		}
+		R2AIPrompt *prompt = parse_prompt_file (filepath);
+		if (!prompt) {
+			R_LOG_ERROR ("Cannot read prompt file: %s", filepath);
+			free (filepath);
+			r_list_free (search_dirs);
+			free (name);
+			return;
+		}
+		if (!prompt->title || !prompt->command) {
+			R_LOG_WARN ("Prompt %s is missing required Title or Command directive", name);
+		}
+		free (filepath);
+		// run commands
+		char *cmd_output = run_commands (core, prompt->command);
+		// run if_command if output not empty
+		if (cmd_output && *cmd_output && prompt->if_command) {
+			char *extra_out = run_commands (core, prompt->if_command);
+			if (extra_out) {
+				char *new_output = r_str_newf ("%s\n%s", cmd_output, extra_out);
+				free (cmd_output);
+				cmd_output = new_output;
+				free (extra_out);
+			}
+		}
+		// determine the prompt to use
+		char *use_prompt = prompt->prompt;
+		if ((!cmd_output || !*cmd_output) && prompt->if_empty) {
+			if (!strcmp (prompt->if_empty, "exit")) {
+				R_LOG_ERROR ("Command output is empty, exiting prompt");
+				r2aiprompt_free (prompt);
+				free (cmd_output);
+				free (name);
+				r_list_free (search_dirs);
+				return;
+			} else {
+				use_prompt = prompt->if_empty;
+			}
+		}
+		// replace vars in prompt
+		char *replaced_prompt = replace_vars (core, use_prompt);
+		RStrBuf *sb = r_strbuf_new (replaced_prompt);
+		free (replaced_prompt);
+		if (cmd_output) {
+			r_strbuf_appendf (sb, "\n%s", cmd_output);
+			free (cmd_output);
+		}
+		if (extra) {
+			r_strbuf_appendf (sb, "\n%s", extra);
+		}
+		char *final_prompt = r_strbuf_drain (sb);
+		// now send to LLM
+		char *error = NULL;
+		char *res = r2ai (cps, (R2AIArgs){ .input = final_prompt, .model = prompt->model, .provider = prompt->provider, .error = &error, .dorag = true });
+		if (error) {
+			R_LOG_ERROR ("%s", error);
+			free (error);
+		} else if (res) {
+			r_cons_printf (core->cons, "%s\n", res);
+			free (res);
+		}
+		r2aiprompt_free (prompt);
+		free (final_prompt);
+		free (name);
+	}
+	r_list_free (search_dirs);
+}

src/r2ai.1

@@ -0,0 +1,158 @@
+.TH R2AI 1 "October 2025" "radare2" "User Commands"
+.SH NAME
+r2ai \- AI assistant for radare2
+.SH SYNOPSIS
+.B r2ai
+[\fIoptions\fR] [\fIquery\fR]
+.SH DESCRIPTION
+r2ai is an AI-powered plugin and command-line tool for radare2 that provides assistance with reverse engineering tasks, including decompilation, function analysis, vulnerability detection, and more. It integrates with various LLM providers like OpenAI, Anthropic, and Ollama.
+.SH OPTIONS
+.TP
+.B \-d
+Decompile the current function.
+.TP
+.B \-d \fIquery\fR
+Ask a question about the current function.
+.TP
+.B \-dr
+Decompile the current function with one level of recursivity.
+.TP
+.B \-a \fIquery\fR
+Resolve a question using auto mode, which allows the AI to use tools.
+.TP
+.B \-e \fIk(=v)\fR
+Set configuration option (same as 'r2ai.' prefix).
+.TP
+.B \-h
+Show help message.
+.TP
+.B \-i \fIfile\fR \fIquery\fR
+Read a file and ask the LLM with the given query.
+.TP
+.B \-m
+Show selected model, list suggested ones, and choose one.
+.TP
+.B \-p \fIprovider\fR
+Set LLM provider (openai, anthropic, gemini, etc.).
+.TP
+.B \-q
+List available query prompts.
+.TP
+.B \-q \fIname\fR [\fIinstructions\fR]
+Run predefined prompt with optional custom instructions.
+.TP
+.B \-r
+Enter the chat REPL.
+.TP
+.B \-L
+Show chat logs (use -Lj for JSON). Only for auto mode.
+.TP
+.B \-L-\fIN\fR
+Delete the last N messages from the chat history.
+.TP
+.B \-R
+Reset the chat conversation context.
+.TP
+.B \-Rq [\fItext\fR]
+Refresh and query embeddings (see r2ai.data).
+.TP
+.B \fIquery\fR
+Query the selected model and provider with the given query.
+.SH CONFIGURATION
+Configuration options can be set with \fBr2ai -e \fIkeyname\fR=\fIvalue\fR.
+.TP
+.B r2ai.api
+Name of the provider (e.g., openai). List possibilities with \fBr2ai -e r2ai.api=?\fR.
+.TP
+.B r2ai.model
+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).
+.TP
+.B r2ai.max_tokens
+Maximum output tokens or total tokens.
+.TP
+.B r2ai.temperature
+Creativity level (0=not creative, 1=very creative).
+.TP
+.B r2ai.cmds
+R2 command to issue and send output in context to model.
+.TP
+.B r2ai.lang
+Programming language for decompilation (only for \fBr2ai -d\fR).
+.TP
+.B r2ai.hlang
+Language for AI responses.
+.TP
+.B r2ai.prompt
+User prompt for \fBr2ai -d\fR.
+.TP
+.B r2ai.auto.max_runs
+Maximum loops in auto mode (\fBr2ai -a\fR).
+.TP
+.B r2ai.auto.hide_tool_output
+Hide tool output in auto mode.
+.TP
+.B r2ai.auto.yolo
+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.
+.SH QUERY PROMPTS
+Query prompts are predefined templates for common tasks. Available prompts:
+.TP
+.B autoname
+Automatically rename functions and variables.
+.TP
+.B decompile
+Decompile the current function.
+.TP
+.B devices
+Analyze device-related code.
+.TP
+.B dlopen
+Analyze dynamic library loading.
+.TP
+.B explain
+Explain the purpose of the current function.
+.TP
+.B libs
+Analyze library usage.
+.TP
+.B signature
+Improve function signatures.
+.TP
+.B varnames
+Rename variables based on context.
+.TP
+.B vulns
+Find vulnerabilities in the current function.
+.SH EXAMPLES
+.TP
+Decompile current function:
+.B r2ai -d
+.TP
+Ask about current function:
+.B r2ai -d "what is the input buffer size?"
+.TP
+Use auto mode:
+.B r2ai -a "find vulnerabilities in this function"
+.TP
+List prompts:
+.B r2ai -q
+.TP
+Run a prompt:
+.B r2ai -q explain
+.TP
+Set provider:
+.B r2ai -p openai
+.TP
+Enter REPL:
+.B r2ai -r
+.SH SEE ALSO
+.BR radare2 (1),
+.BR r2pm (1),
+.BR r2agent (1)
+.SH AUTHOR
+radare2 community

src/r2ai.h

@@ -1,25 +1,36 @@
 #ifndef R2AI_H
 #define R2AI_H
 
+#include <time.h>
+#include <stdint.h>
+#include <stdbool.h>
+#include <stddef.h>
+#include <stdlib.h>
 #include <r_core.h>
 #include <r_util/r_json.h>
 #include "r_vdb.h"
 #include "markdown.h"
 
-#if R2_VERSION_NUMBER >= 50909
-#define R2_PRINTF(...) r_cons_printf (core->cons, __VA_ARGS__)
-#define R2_FLUSH() r_cons_flush (core->cons)
-#define R2_NEWLINE() r_cons_newline (core->cons)
-#define R2_PRINTLN(x) r_cons_println (core->cons, x)
-#define R2_INTERRUPTED() r_cons_is_breaked (core->cons)
+#define R2AI_VERSION "1.2.2"
+
+#if 0
+#define R2AI_DEFAULT_MODEL "gpt-5-mini"
+#define R2AI_DEFAULT_PROVIDER "openai"
 #else
-#define R2_PRINTF(...) r_cons_printf (__VA_ARGS__)
-#define R2_FLUSH() r_cons_flush ()
-#define R2_NEWLINE() r_cons_newline()
-#define R2_PRINTLN(x) r_cons_println(x)
-#define R2_INTERRUPTED() r_cons_is_breaked ()
+// #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"
 #endif
 
+#define R2AI_DEFAULT_VECTORS 16
+
+#if R2_VERSION_NUMBER < 60000
+#error Your radare2 is too old
+#endif
+
+
+
 // Tool definition structure
 typedef struct {
 	char *name;
@@ -40,14 +51,14 @@ typedef struct {
 } R2AI_ToolCall;
 
 typedef struct {
-	const char *type;
-	const char *id;
-	const char *name;
-	const char *input;
-	const char *data;
-	const char *thinking;
-	const char *signature;
-	const char *text;
+	char *type;
+	char *id;
+	char *name;
+	char *input;
+	char *data;
+	char *thinking;
+	char *signature;
+	char *text;
 } R2AI_ContentBlock;
 
 typedef struct {
@@ -56,15 +67,28 @@ typedef struct {
 } R2AI_ContentBlocks;
 
 typedef struct {
-	const char *role;
-	const char *content;
-	const char *reasoning_content;
-	const R2AI_ContentBlocks *content_blocks;
-	const char *tool_call_id;
-	const R2AI_ToolCall *tool_calls;
+	char *role;
+	char *content;
+	char *reasoning_content;
+	R2AI_ContentBlocks *content_blocks;
+	char *tool_call_id;
+	R2AI_ToolCall *tool_calls;
 	int n_tool_calls;
 } R2AI_Message;
 
+typedef struct {
+	char *title;
+	char *author;
+	char *desc;
+	char *command;
+	char *prompt;
+	char *requires;
+	char *if_empty;
+	char *if_command;
+	char *model;
+	char *provider;
+} R2AIPrompt;
+
 typedef struct {
 	uint64_t prompt_tokens;
 	uint64_t completion_tokens;
@@ -78,9 +102,7 @@ typedef struct {
 
 // Messages array management
 typedef struct {
-	R2AI_Message *messages;
-	int n_messages;
-	int cap_messages;
+	RList *messages; // RList<R2AI_Message *>
 } R2AI_Messages;
 
 typedef struct {
@@ -98,6 +120,32 @@ typedef struct {
 	char **error;
 } R2AIArgs;
 
+// Stats structure from auto.c
+typedef struct {
+	double total_cost;
+	double run_cost;
+	int total_tokens;
+	int run_tokens;
+	int total_prompt_tokens;
+	int run_prompt_tokens;
+	int total_completion_tokens;
+	int run_completion_tokens;
+	time_t start_time;
+	time_t total_start_time;
+} R2AIStats;
+
+// Main state structure to hold all global state
+typedef struct r2ai_state_t {
+	R2AI_Messages *conversation; // Global conversation messages (from messages.c)
+	R2AIStats stats; // Global stats (from auto.c)
+	R2AI_Tools *tools; // Global tools instance (from tools.c)
+	RMarkdownTheme current_theme; // Global theme (from markdown.c)
+	bool theme_initialized; // Global theme flag (from markdown.c)
+	void *help_msg; // Global help message (from r2ai.c)
+	RVdb *db; // Vector database for embeddings
+	HtPP *model_compat_db; // Model compatibility database (from openai.c)
+} R2AI_State;
+
 /**
  * Initialize a new empty messages array
  */
@@ -106,12 +154,12 @@ R_API R2AI_Messages *r2ai_msgs_new(void);
 /**
  * Initialize the conversation container (call during plugin init)
  */
-R_API void r2ai_conversation_init(void);
+R_API void r2ai_conversation_init(R2AI_State *state);
 
 /**
  * Get the conversation instance (returns NULL if not initialized)
  */
-R_API R2AI_Messages *r2ai_conversation_get(void);
+R_API R2AI_Messages *r2ai_conversation_get(R2AI_State *state);
 
 /**
  * Clear all messages in a container without freeing the container
@@ -162,13 +210,18 @@ R_API void r2ai_msgs_free(R2AI_Messages *msgs);
 /**
  * Free the conversation (call during plugin unload)
  */
-R_API void r2ai_conversation_free(void);
+R_API void r2ai_conversation_free(R2AI_State *state);
 
 /**
  * Free a R2AI_Message structure
  */
 R_API void r2ai_message_free(R2AI_Message *msg);
 
+/**
+ * Free a R2AIPrompt structure
+ */
+R_API void r2aiprompt_free(R2AIPrompt *prompt);
+
 /**
  * Delete the last N messages from the message array
  * If n <= 0, defaults to deleting just the last message
@@ -236,7 +289,7 @@ 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(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
@@ -248,29 +301,29 @@ R_API char *r2ai_http_get(const char *url, const char *headers[], int *code, int
 R_IPI const char *r2ai_get_provider_url(RCore *core, const char *provider);
 
 // anthropic
-R_IPI R2AI_ChatResponse *r2ai_anthropic(RCore *core, R2AIArgs args);
+R_IPI R2AI_ChatResponse *r2ai_anthropic(RCorePluginSession *cps, R2AIArgs args);
 
 // openai
-R_IPI R2AI_ChatResponse *r2ai_openai(RCore *core, R2AIArgs args);
-R_IPI void r2ai_openai_fini(void);
+R_IPI R2AI_ChatResponse *r2ai_openai(RCorePluginSession *cps, R2AIArgs args);
+R_IPI void r2ai_openai_fini(R2AI_State *state);
 
 // auto mode
-R_IPI void cmd_r2ai_a(RCore *core, const char *user_query);
-R_API char *r2ai(RCore *core, R2AIArgs args);
+R_IPI void cmd_r2ai_a(RCorePluginSession *cps, const char *user_query);
+// 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);
+R_API bool r2ai_fini(RCorePluginSession *cps);
 
-R_IPI R2AI_ChatResponse *r2ai_llmcall(RCore *core, R2AIArgs args);
+R_IPI R2AI_ChatResponse *r2ai_llmcall(RCorePluginSession *cps, R2AIArgs args);
 
-R_IPI void cmd_r2ai_logs(RCore *core);
+R2AI_ChatResponse *r2ai_rawtools_llmcall(RCorePluginSession *cps, R2AIArgs args);
 
-/**
- * Create a conversation with system prompt and optional user message
- */
-R_API R2AI_Messages *create_conversation(const char *user_message);
+R_IPI void cmd_r2ai_logs(RCorePluginSession *cps);
 
 /**
  * Process messages through LLM and handle tool calls recursively
  */
-R_API void process_messages(RCore *core, R2AI_Messages *messages, const char *system_prompt, int n_run);
+R_API void process_messages(RCorePluginSession *cps, R2AI_Messages *messages, const char *system_prompt, int n_run);
 
 /**
  * Helper function to convert RJson to string
@@ -281,5 +334,6 @@ 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);
+R_API void r2ai_cmd_q(RCorePluginSession *cps, const char *input);
 
 #endif

src/r2ai_http.c

@@ -18,60 +18,80 @@
 // Global flag for tracking interrupt status
 static volatile sig_atomic_t r2ai_http_interrupted = 0;
 
+// Helper function to get HTTP configuration with defaults
+static void get_http_config(RCore *core, int *timeout, int *max_retries, int *max_backoff) {
+	*timeout = 120;
+	*max_retries = 10;
+	*max_backoff = 30;
+	if (core) {
+		int t = r_config_get_i (core->config, "r2ai.http.timeout");
+		if (t > 0) {
+			*timeout = t;
+		}
+		int mr = r_config_get_i (core->config, "r2ai.http.max_retries");
+		if (mr >= 0) {
+			*max_retries = mr;
+		}
+		int mb = r_config_get_i (core->config, "r2ai.http.max_backoff");
+		if (mb > 0) {
+			*max_backoff = mb;
+		}
+	}
+}
+
+#ifdef _WIN32
+// Helper to append headers to PowerShell command
+static void append_headers_to_cmd(RStrBuf *cmd, const char *headers[]) {
+	r_strbuf_appendf (cmd, "$ProgressPreference='SilentlyContinue';$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);
+				r_strbuf_appendf (cmd, "'%s'='%s';", key, value);
+				free (key);
+				free (value);
+			}
+			free (header);
+		}
+	}
+	r_strbuf_appendf (cmd, "};");
+}
+#endif
+
 // Signal handler for SIGINT
 static void r2ai_http_sigint_handler(int sig) {
-    (void)sig;
-    r2ai_http_interrupted = 1;
+	(void)sig;
+	r2ai_http_interrupted = 1;
 }
 
 /*
  * Portable install/restore helpers for SIGINT handler.
- * Some platforms may not have struct sigaction available; use signal()
+ * Some platforms may not have struct sigaction available; use signal ()
  * as a fallback. We allocate storage for the old handler when
  * sigaction is available so the restore function can restore it later.
  */
 static void install_sigint_handler_local(void **out_old, int *out_old_is_sigaction) {
-    void (*old)(int) = signal (SIGINT, r2ai_http_sigint_handler);
-    *out_old = (void *)old;
-    *out_old_is_sigaction = 0;
+	void (*old) (int) = signal (SIGINT, r2ai_http_sigint_handler);
+	*out_old = (void *)old;
+	*out_old_is_sigaction = 0;
 }
 
 static void restore_sigint_handler_local(void *old, int old_is_sigaction) {
-    (void)old_is_sigaction;
-    signal (SIGINT, (void (*)(int))old);
+	(void)old_is_sigaction;
+	signal (SIGINT, (void (*) (int))old);
 }
 
-// Helper function to implement exponential backoff sleep
-static void r2ai_sleep_with_backoff(int retry_count, int max_sleep_seconds) {
-	// Calculate sleep time with exponential backoff: 2^retry * base_time with jitter
-	int base_time_ms = 1000; // 500ms base time
-	int max_sleep_ms = max_sleep_seconds * 1000;
-
-	// Calculate exponential delay time with upper bound
-	int delay_ms = (1 << retry_count) * base_time_ms;
-	if (delay_ms > max_sleep_ms) {
-		delay_ms = max_sleep_ms;
+// Simple retry sleep
+static void sleep_retry(int retry_count, int max_sleep_seconds) {
+	int delay_seconds = retry_count + 1; // sleep 1, 2, 3, ... seconds
+	if (delay_seconds > max_sleep_seconds) {
+		delay_seconds = max_sleep_seconds;
 	}
-
-	// Add jitter (±20%)
-	int jitter = delay_ms / 5;
-	if (jitter > 0) {
-		srand (time (NULL) + retry_count);
-		delay_ms += (rand () % (jitter * 2)) - jitter;
-	}
-
-	// Ensure delay stays positive and within max limits
-	if (delay_ms <= 0) {
-		delay_ms = base_time_ms;
-	} else if (delay_ms > max_sleep_ms) {
-		delay_ms = max_sleep_ms;
-	}
-
-	// Sleep for the calculated time
-	struct timespec ts;
-	ts.tv_sec = delay_ms / 1000;
-	ts.tv_nsec = (delay_ms % 1000) * 1000000;
-	nanosleep (&ts, NULL);
+	r_sys_sleep (delay_seconds);
 }
 
 #if HAVE_LIBCURL
@@ -121,33 +141,13 @@ static char *curl_http_post(const char *url, const char *headers[], const char *
 		return NULL;
 	}
 
-	// Get timeout and retry configuration from config if available
-	int timeout = 120; // Default timeout in seconds
-	int max_retries = 10; // Default max retries
-	int max_backoff = 30; // Default max backoff in seconds
-
-	RCore *core = r_cons_singleton ()->user;
-	if (core) {
-		timeout = r_config_get_i (core->config, "r2ai.http.timeout");
-		if (timeout <= 0) {
-			timeout = 120; // Use default if invalid
-		}
-
-		max_retries = r_config_get_i (core->config, "r2ai.http.max_retries");
-		if (max_retries < 0) {
-			max_retries = 10; // Use default if invalid
-		}
-
-		max_backoff = r_config_get_i (core->config, "r2ai.http.max_backoff");
-		if (max_backoff <= 0) {
-			max_backoff = 30; // Use default if invalid
-		}
-	}
+	int timeout, max_retries, max_backoff;
+	get_http_config (NULL, &timeout, &max_retries, &max_backoff);
 
 	// Install signal handler for interruption (portable)
 	void *r2ai_old_sig = NULL;
 	int r2ai_old_is_sigaction = 0;
-	install_sigint_handler_local(&r2ai_old_sig, &r2ai_old_is_sigaction);
+	install_sigint_handler_local (&r2ai_old_sig, &r2ai_old_is_sigaction);
 
 	// Reset interrupt flag
 	r2ai_http_interrupted = 0;
@@ -168,10 +168,10 @@ static char *curl_http_post(const char *url, const char *headers[], const char *
 		if (!response.data) {
 			if (retry_count < max_retries) {
 				retry_count++;
-				r2ai_sleep_with_backoff (retry_count, max_backoff);
+				sleep_retry (retry_count, max_backoff);
 				continue;
 			}
-			restore_sigint_handler_local(r2ai_old_sig, r2ai_old_is_sigaction); // Restore signal handler
+			restore_sigint_handler_local (r2ai_old_sig, r2ai_old_is_sigaction); // Restore signal handler
 			return NULL;
 		}
 		response.data[0] = '\0';
@@ -182,10 +182,10 @@ static char *curl_http_post(const char *url, const char *headers[], const char *
 			free (response.data);
 			if (retry_count < max_retries) {
 				retry_count++;
-				r2ai_sleep_with_backoff (retry_count, max_backoff);
+				sleep_retry (retry_count, max_backoff);
 				continue;
 			}
-			restore_sigint_handler_local(r2ai_old_sig, r2ai_old_is_sigaction); // Restore signal handler
+			restore_sigint_handler_local (r2ai_old_sig, r2ai_old_is_sigaction); // Restore signal handler
 			return NULL;
 		}
 
@@ -245,7 +245,7 @@ static char *curl_http_post(const char *url, const char *headers[], const char *
 			if (retry_count < max_retries) {
 				retry_count++;
 				R_LOG_INFO ("Retrying request (%d/%d) after failure...", retry_count, max_retries);
-				r2ai_sleep_with_backoff (retry_count, max_backoff);
+				sleep_retry (retry_count, max_backoff);
 				continue;
 			}
 			break; // Exit the retry loop after max retries
@@ -261,7 +261,7 @@ static char *curl_http_post(const char *url, const char *headers[], const char *
 			if (retry_count < max_retries) {
 				retry_count++;
 				R_LOG_INFO ("Retrying request (%d/%d) after rate limiting...", retry_count, max_retries);
-				r2ai_sleep_with_backoff (retry_count, max_backoff);
+				sleep_retry (retry_count, max_backoff);
 				continue;
 			}
 			break; // Exit the retry loop after max retries
@@ -280,14 +280,86 @@ static char *curl_http_post(const char *url, const char *headers[], const char *
 	}
 
 	// Restore the original signal handler
-	restore_sigint_handler_local(r2ai_old_sig, r2ai_old_is_sigaction);
+	restore_sigint_handler_local (r2ai_old_sig, r2ai_old_is_sigaction);
 
 	return result;
 }
 #endif // HAVE_LIBCURL
 
+#ifdef _WIN32
+/**
+ * Windows-specific HTTP POST using PowerShell
+ */
+static char *windows_http_post(const char *url, const char *headers[], const char *data, int *code, int *rlen, int timeout) {
+	RStrBuf *cmd = r_strbuf_new ("powershell -Command \"");
+	append_headers_to_cmd (cmd, headers);
+	r_strbuf_appendf (cmd, "$body='%s';", data);
+	r_strbuf_appendf (cmd, "try{$r=Invoke-WebRequest -Method Post -Uri '%s' -Headers $headers -Body $body -TimeoutSec %d;", url, timeout);
+	r_strbuf_appendf (cmd, "Write-Host $r.StatusCode;$r.Content}catch{Write-Host 0;$_.Exception.Message}\"");
+	char *cmd_str = r_strbuf_drain (cmd);
+	char *full_response = r_sys_cmd_str (cmd_str, NULL, NULL);
+	free (cmd_str);
+	if (full_response) {
+		char *newline = strchr (full_response, '\n');
+		if (newline) {
+			*newline = '\0';
+			*code = atoi (full_response);
+			char *body = newline + 1;
+			if (rlen) {
+				*rlen = strlen (body);
+			}
+			char *result = strdup (body);
+			free (full_response);
+			return result;
+		}
+		*code = 200;
+		if (rlen) {
+			*rlen = strlen (full_response);
+		}
+		return full_response;
+	}
+	*code = 0;
+	return NULL;
+}
+
+/**
+ * Windows-specific HTTP GET using PowerShell
+ */
+static char *windows_http_get(const char *url, const char *headers[], int *code, int *rlen, int timeout) {
+	RStrBuf *cmd = r_strbuf_new ("powershell -Command \"");
+	append_headers_to_cmd (cmd, headers);
+	r_strbuf_appendf (cmd, "try{$r=Invoke-WebRequest -Method Get -Uri '%s' -Headers $headers -TimeoutSec %d;", url, timeout);
+	r_strbuf_appendf (cmd, "Write-Host $r.StatusCode;$r.Content}catch{Write-Host 0;$_.Exception.Message}\"");
+	char *cmd_str = r_strbuf_drain (cmd);
+	char *full_response = r_sys_cmd_str (cmd_str, NULL, NULL);
+	free (cmd_str);
+	if (full_response) {
+		char *newline = strchr (full_response, '\n');
+		if (newline) {
+			*newline = '\0';
+			*code = atoi (full_response);
+			char *body = newline + 1;
+			if (rlen) {
+				*rlen = strlen (body);
+			}
+			char *result = strdup (body);
+			free (full_response);
+			return result;
+		}
+		*code = 200;
+		if (rlen) {
+			*rlen = strlen (full_response);
+		}
+		return full_response;
+	}
+	*code = 0;
+	return NULL;
+}
+#endif
+
 /**
  * Execute curl as a system command, sending data via temporary file
+ * On Windows, use PowerShell as fallback when curl is not available
  *
  * @param url The URL to send the request to
  * @param headers Array of headers, NULL terminated
@@ -301,25 +373,13 @@ static char *system_curl_post_file(RCore *core, const char *url, const char *hea
 		return NULL;
 	}
 
-	int timeout = r_config_get_i (core->config, "r2ai.http.timeout");
-	if (timeout <= 0) {
-		timeout = 120; // Use default if invalid
-	}
-
-	int max_retries = r_config_get_i (core->config, "r2ai.http.max_retries");
-	if (max_retries < 0) {
-		max_retries = 10; // Use default if invalid
-	}
-
-	int max_backoff = r_config_get_i (core->config, "r2ai.http.max_backoff");
-	if (max_backoff <= 0) {
-		max_backoff = 30; // Use default if invalid
-	}
+	int timeout, max_retries, max_backoff;
+	get_http_config (core, &timeout, &max_retries, &max_backoff);
 
 	// Install signal handler for interruption (portable)
 	void *r2ai_old_sig = NULL;
 	int r2ai_old_is_sigaction = 0;
-	install_sigint_handler_local(&r2ai_old_sig, &r2ai_old_is_sigaction);
+	install_sigint_handler_local (&r2ai_old_sig, &r2ai_old_is_sigaction);
 
 	// Reset interrupt flag
 	r2ai_http_interrupted = 0;
@@ -336,7 +396,7 @@ static char *system_curl_post_file(RCore *core, const char *url, const char *hea
 			R_LOG_ERROR ("Failed to create temporary file for curl data");
 			if (retry_count < max_retries) {
 				retry_count++;
-				r2ai_sleep_with_backoff (retry_count, max_backoff);
+				sleep_retry (retry_count, max_backoff);
 				continue;
 			}
 			break;
@@ -349,13 +409,88 @@ static char *system_curl_post_file(RCore *core, const char *url, const char *hea
 				free (temp_file);
 				if (retry_count < max_retries) {
 					retry_count++;
-					r2ai_sleep_with_backoff (retry_count, max_backoff);
+					sleep_retry (retry_count, max_backoff);
 					continue;
 				}
 				break;
 			}
 		}
 
+#ifdef _WIN32
+		// On Windows, use PowerShell for HTTP requests
+		RStrBuf *cmd = r_strbuf_new ("powershell -Command \"");
+		r_strbuf_appendf (cmd, "$ProgressPreference = 'SilentlyContinue'; ");
+		r_strbuf_appendf (cmd, "$headers = @{");
+
+		// Add headers as PowerShell hashtable
+		for (int i = 0; headers[i] != NULL; 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);
+				r_strbuf_appendf (cmd, "'%s'='%s';", key, value);
+				free (key);
+				free (value);
+			}
+			free (header);
+		}
+		r_strbuf_appendf (cmd, "}; ");
+
+		// Add body
+		if (use_files) {
+			r_strbuf_appendf (cmd, "$body = Get-Content '%s' -Raw; ", temp_file);
+		} else {
+			char *escaped_data = r_str_replace_all (strdup (data), "'", "''");
+			r_strbuf_appendf (cmd, "$body = '%s'; ", escaped_data);
+			free (escaped_data);
+		}
+
+		r_strbuf_appendf (cmd, "try { $response = Invoke-WebRequest -Method Post -Uri '%s' -Headers $headers -Body $body -TimeoutSec %d; ", url, timeout);
+		r_strbuf_appendf (cmd, "Write-Host $response.StatusCode; $response.Content } catch { Write-Host $_.Exception.Response.StatusCode.Value__; $_.Exception.Response.GetResponseStream().ReadToEnd() }\"");
+
+		// Execute the PowerShell command
+		char *cmd_str = r_strbuf_drain (cmd);
+
+		R_LOG_DEBUG ("Running PowerShell: %s", cmd_str);
+		char *full_response = r_sys_cmd_str (cmd_str, NULL, NULL);
+
+		free (cmd_str);
+		r_file_rm (temp_file);
+		free (temp_file);
+
+		// Parse PowerShell response: first line is status code, rest is body
+		if (full_response) {
+			char *newline = strchr (full_response, '\n');
+			if (newline) {
+				*newline = '\0';
+				*code = atoi (full_response);
+				char *body = newline + 1;
+				if (rlen) {
+					*rlen = strlen (body);
+				}
+				result = strdup (body);
+			} else {
+				*code = 200; // Assume success if no newline
+				if (rlen) {
+					*rlen = strlen (full_response);
+				}
+				result = full_response;
+			}
+			free (full_response);
+			success = true;
+		} else {
+			if (retry_count < max_retries) {
+				retry_count++;
+				R_LOG_INFO ("Retrying request (%d/%d) after failure...", retry_count, max_retries);
+				sleep_retry (retry_count, max_backoff);
+				continue;
+			}
+			*code = 0;
+			break;
+		}
+#else
 		// Compose curl command
 		RStrBuf *cmd = r_strbuf_new ("curl -s");
 
@@ -406,7 +541,7 @@ static char *system_curl_post_file(RCore *core, const char *url, const char *hea
 			R_LOG_DEBUG ("HTTP request was interrupted by user");
 			free (response);
 			response = NULL;
-			restore_sigint_handler_local(r2ai_old_sig, r2ai_old_is_sigaction);
+			restore_sigint_handler_local (r2ai_old_sig, r2ai_old_is_sigaction);
 			return NULL;
 		}
 
@@ -423,16 +558,17 @@ static char *system_curl_post_file(RCore *core, const char *url, const char *hea
 			if (retry_count < max_retries) {
 				retry_count++;
 				R_LOG_INFO ("Retrying request (%d/%d) after failure...", retry_count, max_retries);
-				r2ai_sleep_with_backoff (retry_count, max_backoff);
+				sleep_retry (retry_count, max_backoff);
 				continue;
 			}
 			*code = 0;
 			break;
 		}
+#endif
 	}
 
 	// Restore the original signal handler
-		restore_sigint_handler_local(r2ai_old_sig, r2ai_old_is_sigaction);
+	restore_sigint_handler_local (r2ai_old_sig, r2ai_old_is_sigaction);
 
 	return result;
 }
@@ -452,32 +588,13 @@ static char *system_curl_get(const char *url, const char *headers[], int *code,
 	}
 
 	// Get timeout and retry configuration from config if available
-	int timeout = 120; // Default timeout in seconds
-	int max_retries = 10; // Default max retries
-	int max_backoff = 30; // Default max backoff in seconds
-
-	RCore *core = r_cons_singleton ()->user;
-	if (core) {
-		timeout = r_config_get_i (core->config, "r2ai.http.timeout");
-		if (timeout <= 0) {
-			timeout = 120; // Use default if invalid
-		}
-
-		max_retries = r_config_get_i (core->config, "r2ai.http.max_retries");
-		if (max_retries < 0) {
-			max_retries = 10; // Use default if invalid
-		}
-
-		max_backoff = r_config_get_i (core->config, "r2ai.http.max_backoff");
-		if (max_backoff <= 0) {
-			max_backoff = 30; // Use default if invalid
-		}
-	}
+	int timeout, max_retries, max_backoff;
+	get_http_config (NULL, &timeout, &max_retries, &max_backoff);
 
 	// Install signal handler for interruption (portable)
 	void *r2ai_old_sig = NULL;
 	int r2ai_old_is_sigaction = 0;
-	install_sigint_handler_local(&r2ai_old_sig, &r2ai_old_is_sigaction);
+	install_sigint_handler_local (&r2ai_old_sig, &r2ai_old_is_sigaction);
 
 	// Reset interrupt flag
 	r2ai_http_interrupted = 0;
@@ -488,6 +605,20 @@ static char *system_curl_get(const char *url, const char *headers[], int *code,
 	bool success = false;
 
 	while (!success && retry_count <= max_retries && !r2ai_http_interrupted) {
+#ifdef _WIN32
+		result = windows_http_get (url, headers, code, rlen, timeout);
+		if (result) {
+			success = true;
+		} else if (retry_count < max_retries) {
+			retry_count++;
+			R_LOG_INFO ("Retrying request (%d/%d) after failure...", retry_count, max_retries);
+			sleep_retry (retry_count, max_backoff);
+			continue;
+		} else {
+			*code = 0;
+			break;
+		}
+#else
 		// Compose curl command
 		RStrBuf *cmd = r_strbuf_new ("curl -s");
 
@@ -528,7 +659,7 @@ static char *system_curl_get(const char *url, const char *headers[], int *code,
 		if (r2ai_http_interrupted) {
 			R_LOG_DEBUG ("HTTP request was interrupted by user");
 			free (response);
-			restore_sigint_handler_local(r2ai_old_sig, r2ai_old_is_sigaction);
+			restore_sigint_handler_local (r2ai_old_sig, r2ai_old_is_sigaction);
 			return NULL;
 		}
 
@@ -545,16 +676,17 @@ static char *system_curl_get(const char *url, const char *headers[], int *code,
 			if (retry_count < max_retries) {
 				retry_count++;
 				R_LOG_INFO ("Retrying request (%d/%d) after failure...", retry_count, max_retries);
-				r2ai_sleep_with_backoff (retry_count, max_backoff);
+				sleep_retry (retry_count, max_backoff);
 				continue;
 			}
 			*code = 0;
 			break;
 		}
+#endif
 	}
 
 	// Restore the original signal handler
-		restore_sigint_handler_local(r2ai_old_sig, r2ai_old_is_sigaction);
+	restore_sigint_handler_local (r2ai_old_sig, r2ai_old_is_sigaction);
 
 	return result;
 }
@@ -562,32 +694,13 @@ static char *system_curl_get(const char *url, const char *headers[], int *code,
 // Socket implementation with interrupt handling and retry logic
 static char *socket_http_post_with_interrupt(const char *url, const char *headers[], const char *data, int *code, int *rlen) {
 	// Get timeout and retry configuration from config if available
-	int timeout = 120; // Default timeout in seconds
-	int max_retries = 10; // Default max retries
-	int max_backoff = 30; // Default max backoff in seconds
-
-	RCore *core = r_cons_singleton ()->user;
-	if (core) {
-		timeout = r_config_get_i (core->config, "r2ai.http.timeout");
-		if (timeout <= 0) {
-			timeout = 120; // Use default if invalid
-		}
-
-		max_retries = r_config_get_i (core->config, "r2ai.http.max_retries");
-		if (max_retries < 0) {
-			max_retries = 10; // Use default if invalid
-		}
-
-		max_backoff = r_config_get_i (core->config, "r2ai.http.max_backoff");
-		if (max_backoff <= 0) {
-			max_backoff = 30; // Use default if invalid
-		}
-	}
+	int timeout, max_retries, max_backoff;
+	get_http_config (NULL, &timeout, &max_retries, &max_backoff);
 
 	// Install signal handler for interruption (portable)
 	void *r2ai_old_sig = NULL;
 	int r2ai_old_is_sigaction = 0;
-	install_sigint_handler_local(&r2ai_old_sig, &r2ai_old_is_sigaction);
+	install_sigint_handler_local (&r2ai_old_sig, &r2ai_old_is_sigaction);
 
 	// Reset interrupt flag
 	r2ai_http_interrupted = 0;
@@ -634,7 +747,7 @@ static char *socket_http_post_with_interrupt(const char *url, const char *header
 				if (retry_count < max_retries) {
 					retry_count++;
 					R_LOG_INFO ("Retrying request (%d/%d) after error...", retry_count, max_retries);
-					r2ai_sleep_with_backoff (retry_count, max_backoff);
+					sleep_retry (retry_count, max_backoff);
 					continue;
 				}
 				break; // Exit the retry loop after max retries
@@ -646,7 +759,7 @@ static char *socket_http_post_with_interrupt(const char *url, const char *header
 			if (retry_count < max_retries) {
 				retry_count++;
 				R_LOG_INFO ("Retrying request (%d/%d) after failure...", retry_count, max_retries);
-				r2ai_sleep_with_backoff (retry_count, max_backoff);
+				sleep_retry (retry_count, max_backoff);
 				continue;
 			}
 			break; // Exit the retry loop after max retries
@@ -657,7 +770,7 @@ static char *socket_http_post_with_interrupt(const char *url, const char *header
 	}
 
 	// Restore the original signal handler
-		restore_sigint_handler_local(r2ai_old_sig, r2ai_old_is_sigaction);
+	restore_sigint_handler_local (r2ai_old_sig, r2ai_old_is_sigaction);
 
 	return result;
 }
@@ -669,32 +782,13 @@ static char *curl_http_get(const char *url, const char *headers[], int *code, in
 	}
 
 	// Get timeout and retry configuration from config if available
-	int timeout = 120; // Default timeout in seconds
-	int max_retries = 10; // Default max retries
-	int max_backoff = 30; // Default max backoff in seconds
-
-	RCore *core = r_cons_singleton ()->user;
-	if (core) {
-		timeout = r_config_get_i (core->config, "r2ai.http.timeout");
-		if (timeout <= 0) {
-			timeout = 120; // Use default if invalid
-		}
-
-		max_retries = r_config_get_i (core->config, "r2ai.http.max_retries");
-		if (max_retries < 0) {
-			max_retries = 10; // Use default if invalid
-		}
-
-		max_backoff = r_config_get_i (core->config, "r2ai.http.max_backoff");
-		if (max_backoff <= 0) {
-			max_backoff = 30; // Use default if invalid
-		}
-	}
+	int timeout, max_retries, max_backoff;
+	get_http_config (NULL, &timeout, &max_retries, &max_backoff);
 
 	// Install signal handler for interruption (portable)
 	void *r2ai_old_sig = NULL;
 	int r2ai_old_is_sigaction = 0;
-	install_sigint_handler_local(&r2ai_old_sig, &r2ai_old_is_sigaction);
+	install_sigint_handler_local (&r2ai_old_sig, &r2ai_old_is_sigaction);
 
 	// Reset interrupt flag
 	r2ai_http_interrupted = 0;
@@ -715,10 +809,10 @@ static char *curl_http_get(const char *url, const char *headers[], int *code, in
 		if (!response.data) {
 			if (retry_count < max_retries) {
 				retry_count++;
-				r2ai_sleep_with_backoff (retry_count, max_backoff);
+				sleep_retry (retry_count, max_backoff);
 				continue;
 			}
-			restore_sigint_handler_local(r2ai_old_sig, r2ai_old_is_sigaction); // Restore signal handler
+			restore_sigint_handler_local (r2ai_old_sig, r2ai_old_is_sigaction); // Restore signal handler
 			return NULL;
 		}
 		response.data[0] = '\0';
@@ -729,10 +823,10 @@ static char *curl_http_get(const char *url, const char *headers[], int *code, in
 			free (response.data);
 			if (retry_count < max_retries) {
 				retry_count++;
-				r2ai_sleep_with_backoff (retry_count, max_backoff);
+				sleep_retry (retry_count, max_backoff);
 				continue;
 			}
-			restore_sigint_handler_local(r2ai_old_sig, r2ai_old_is_sigaction); // Restore signal handler
+			restore_sigint_handler_local (r2ai_old_sig, r2ai_old_is_sigaction); // Restore signal handler
 			return NULL;
 		}
 
@@ -794,7 +888,7 @@ static char *curl_http_get(const char *url, const char *headers[], int *code, in
 			if (retry_count < max_retries) {
 				retry_count++;
 				R_LOG_INFO ("Retrying request (%d/%d) after failure...", retry_count, max_retries);
-				r2ai_sleep_with_backoff (retry_count, max_backoff);
+				sleep_retry (retry_count, max_backoff);
 				continue;
 			}
 			break; // Exit the retry loop after max retries
@@ -810,7 +904,7 @@ static char *curl_http_get(const char *url, const char *headers[], int *code, in
 			if (retry_count < max_retries) {
 				retry_count++;
 				R_LOG_INFO ("Retrying request (%d/%d) after rate limiting...", retry_count, max_retries);
-				r2ai_sleep_with_backoff (retry_count, max_backoff);
+				sleep_retry (retry_count, max_backoff);
 				continue;
 			}
 			break; // Exit the retry loop after max retries
@@ -829,7 +923,7 @@ static char *curl_http_get(const char *url, const char *headers[], int *code, in
 	}
 
 	// Restore the original signal handler
-		restore_sigint_handler_local(r2ai_old_sig, r2ai_old_is_sigaction);
+	restore_sigint_handler_local (r2ai_old_sig, r2ai_old_is_sigaction);
 
 	return result;
 }
@@ -838,32 +932,13 @@ static char *curl_http_get(const char *url, const char *headers[], int *code, in
 // Socket implementation for GET requests
 static char *socket_http_get_with_interrupt(const char *url, const char *headers[], int *code, int *rlen) {
 	// Get timeout and retry configuration from config if available
-	int timeout = 120; // Default timeout in seconds
-	int max_retries = 10; // Default max retries
-	int max_backoff = 30; // Default max backoff in seconds
-
-	RCore *core = r_cons_singleton ()->user;
-	if (core) {
-		timeout = r_config_get_i (core->config, "r2ai.http.timeout");
-		if (timeout <= 0) {
-			timeout = 120; // Use default if invalid
-		}
-
-		max_retries = r_config_get_i (core->config, "r2ai.http.max_retries");
-		if (max_retries < 0) {
-			max_retries = 10; // Use default if invalid
-		}
-
-		max_backoff = r_config_get_i (core->config, "r2ai.http.max_backoff");
-		if (max_backoff <= 0) {
-			max_backoff = 30; // Use default if invalid
-		}
-	}
+	int timeout, max_retries, max_backoff;
+	get_http_config (NULL, &timeout, &max_retries, &max_backoff);
 
 	// Install signal handler for interruption (portable)
 	void *r2ai_old_sig = NULL;
 	int r2ai_old_is_sigaction = 0;
-	install_sigint_handler_local(&r2ai_old_sig, &r2ai_old_is_sigaction);
+	install_sigint_handler_local (&r2ai_old_sig, &r2ai_old_is_sigaction);
 
 	// Reset interrupt flag
 	r2ai_http_interrupted = 0;
@@ -913,7 +988,7 @@ static char *socket_http_get_with_interrupt(const char *url, const char *headers
 				if (retry_count < max_retries) {
 					retry_count++;
 					R_LOG_INFO ("Retrying request (%d/%d) after error...", retry_count, max_retries);
-					r2ai_sleep_with_backoff (retry_count, max_backoff);
+					sleep_retry (retry_count, max_backoff);
 					continue;
 				}
 				break; // Exit the retry loop after max retries
@@ -925,7 +1000,7 @@ static char *socket_http_get_with_interrupt(const char *url, const char *headers
 			if (retry_count < max_retries) {
 				retry_count++;
 				R_LOG_INFO ("Retrying request (%d/%d) after failure...", retry_count, max_retries);
-				r2ai_sleep_with_backoff (retry_count, max_backoff);
+				sleep_retry (retry_count, max_backoff);
 				continue;
 			}
 			break; // Exit the retry loop after max retries
@@ -936,7 +1011,7 @@ static char *socket_http_get_with_interrupt(const char *url, const char *headers
 	}
 
 	// Restore the original signal handler
-		restore_sigint_handler_local(r2ai_old_sig, r2ai_old_is_sigaction);
+	restore_sigint_handler_local (r2ai_old_sig, r2ai_old_is_sigaction);
 
 	return result;
 }
@@ -974,17 +1049,37 @@ R_API char *r2ai_http_post(RCore *core, const char *url, const char *headers[],
 #if USE_LIBCURL && HAVE_LIBCURL
 			return curl_http_post (url, headers, data, code, rlen);
 #else
+#ifdef _WIN32
+			int timeout, max_retries, max_backoff;
+			get_http_config (core, &timeout, &max_retries, &max_backoff);
+			char *result = NULL;
+			int retry_count = 0;
+			bool success = false;
+			while (!success && retry_count <= max_retries) {
+				result = windows_http_post (url, headers, data, code, rlen, timeout);
+				if (result) {
+					success = true;
+				} else if (retry_count < max_retries) {
+					retry_count++;
+					R_LOG_INFO ("Retrying request (%d/%d) after failure...", retry_count, max_retries);
+					sleep_retry (retry_count, max_backoff);
+				} else {
+					*code = 0;
+				}
+			}
+			return result;
+#else
 #if USE_R2_CURL
 			r_sys_setenv ("R2_CURL", "1");
 #endif
 			return socket_http_post_with_interrupt (url, headers, data, code, rlen);
+#endif
 #endif
 		}
 	}
 }
 
-R_API char *r2ai_http_get(const char *url, const char *headers[], int *code, int *rlen) {
-	RCore *core = r_cons_singleton ()->user;
+R_API char *r2ai_http_get(RCore *core, const char *url, const char *headers[], int *code, int *rlen) {
 	const char *backend = "auto";
 
 	if (core) {
@@ -1009,10 +1104,31 @@ R_API char *r2ai_http_get(const char *url, const char *headers[], int *code, int
 #if USE_LIBCURL && HAVE_LIBCURL
 		return curl_http_get (url, headers, code, rlen);
 #else
+#ifdef _WIN32
+		int timeout, max_retries, max_backoff;
+		get_http_config (core, &timeout, &max_retries, &max_backoff);
+		char *result = NULL;
+		int retry_count = 0;
+		bool success = false;
+		while (!success && retry_count <= max_retries) {
+			result = windows_http_get (url, headers, code, rlen, timeout);
+			if (result) {
+				success = true;
+			} else if (retry_count < max_retries) {
+				retry_count++;
+				R_LOG_INFO ("Retrying request (%d/%d) after failure...", retry_count, max_retries);
+				sleep_retry (retry_count, max_backoff);
+			} else {
+				*code = 0;
+			}
+		}
+		return result;
+#else
 #if USE_R2_CURL
 		r_sys_setenv ("R2_CURL", "1");
 #endif
 		return socket_http_get_with_interrupt (url, headers, code, rlen);
+#endif
 #endif
 	}
 }

src/r2ai_priv.h

@@ -0,0 +1,22 @@
+#ifndef INCLUDE_R2AI_PRIV__H
+#define INCLUDE_R2AI_PRIV__H
+
+#include "r2ai.h"
+
+typedef struct {
+	const char *name;
+	const char *url;
+	bool requires_api_key;
+	bool uses_anthropic_header;
+	bool uses_tags_endpoint;
+	bool uses_system_ls;
+} R2AIProvider;
+
+R_IPI char *r2ai_get_api_key(RCore *core, const char *provider);
+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 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);
+
+#endif

src/rawtools.c

@@ -0,0 +1,297 @@
+/* r2ai - Copyright 2023-2025 pancake */
+
+#define R_LOG_ORIGIN "rawtools"
+#include "r2ai.h"
+#include "r2ai_priv.h"
+
+// Function to build the rawtools prompt with available tools
+static const char *RAWTOOLS_PROMPT =
+	"Common radare2 commands you can use with r2cmd:\n"
+	"- i: Show binary information\n"
+	"- iI: Show detailed binary info\n"
+	"- iz: Show strings\n"
+	"- afl: List functions\n"
+	"- pdf @ function: Disassemble function\n"
+	"- px @ address: Show hexdump\n"
+	"- aaa: Analyze all\n"
+	"- s address: Seek to address\n"
+	"- ? command: Get help\n\n"
+	"To run radare2 commands, use the r2cmd tool with JSON arguments.\n\n"
+	"TOOL: r2cmd\n"
+	"USAGE: TOOL: r2cmd {\"command\": \"your_command_here\"}\n\n"
+	"Example: TOOL: r2cmd {\"command\": \"i\"}\n\n"
+	"If you don't need to run commands, just answer directly.\n\n";
+
+// Function to parse raw tool call from response text
+static bool parse_raw_tool_call(const char *response, char **tool_name, char **tool_args) {
+	if (!response || !tool_name || !tool_args) {
+		return false;
+	}
+
+	*tool_name = NULL;
+	*tool_args = NULL;
+
+	// First try to parse "TOOL: " format
+	const char *tool_prefix = "TOOL: ";
+	const char *line_start = response;
+
+	while ((line_start = strstr (line_start, tool_prefix))) {
+		// Check if it's at the beginning of a line
+		if (line_start == response || *(line_start - 1) == '\n') {
+			// Found a tool call
+			const char *tool_start = line_start + strlen (tool_prefix);
+			const char *space_pos = strchr (tool_start, ' ');
+
+			if (space_pos) {
+				// Extract tool name
+				size_t name_len = space_pos - tool_start;
+				*tool_name = r_str_ndup (tool_start, name_len);
+
+				// Extract arguments (rest of the line)
+				*tool_args = strdup (space_pos + 1);
+
+				// Trim whitespace from arguments (simple implementation)
+				if (*tool_args) {
+					char *start = *tool_args;
+					while (*start && (*start == ' ' || *start == '\t')) {
+						start++;
+					}
+					char *end = start + strlen (start) - 1;
+					while (end > start && (*end == ' ' || *end == '\t' || *end == '\n' || *end == '\r')) {
+						end--;
+					}
+					*(end + 1) = '\0';
+					if (start != *tool_args) {
+						memmove (*tool_args, start, strlen (start) + 1);
+					}
+				}
+
+				return true;
+			}
+		}
+		line_start += strlen (tool_prefix);
+	}
+
+	// If no TOOL: format found, try to parse JSON response
+	// Look for JSON that looks like tool arguments
+	const char *json_start = strstr (response, "{\"command\"");
+	if (json_start) {
+		*tool_name = strdup ("r2cmd");
+		*tool_args = strdup (json_start);
+		// Trim trailing whitespace and newlines
+		char *end = *tool_args + strlen (*tool_args) - 1;
+		while (end > *tool_args && (*end == ' ' || *end == '\t' || *end == '\n' || *end == '\r')) {
+			*end = '\0';
+			end--;
+		}
+		return true;
+	}
+
+	return false;
+}
+
+// Function to handle rawtools mode in LLM call
+R2AI_ChatResponse *r2ai_rawtools_llmcall(RCorePluginSession *cps, R2AIArgs args) {
+	if (!cps) {
+		return NULL;
+	}
+	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;
+	if (args.messages && args.messages->messages) {
+		RList *msgs = args.messages->messages;
+		int n_msgs = r_list_length (msgs);
+		for (int i = 0; i < n_msgs; i++) {
+			R2AI_Message *msg = r_list_get_n (msgs, i);
+			if (msg && msg->role && !strcmp (msg->role, "system") && msg->content &&
+				strstr (msg->content, "Here is some information about the binary")) {
+				init_output = strdup (msg->content);
+				break;
+			}
+		}
+	}
+
+	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);
+	}
+	free (init_output);
+	// Temporarily modify args to use enhanced prompt and no tools (since we're using prompt engineering)
+	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->uses_anthropic_header) {
+		response = r2ai_anthropic (cps, rawtools_args);
+	} else {
+		response = r2ai_openai (cps, rawtools_args);
+	}
+
+	free (enhanced_system_prompt);
+
+	if (!response || !response->message || !response->message->content) {
+		return response;
+	}
+
+	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
+		const R2AI_Tools *tools = r2ai_get_tools ();
+		bool tool_found = false;
+
+		for (int i = 0; i < tools->n_tools; i++) {
+			if (tools->tools[i].name && !strcmp (tools->tools[i].name, tool_name)) {
+				tool_found = true;
+				break;
+			}
+		}
+
+		if (tool_found) {
+			// Modify the existing response to include tool call information
+			R2AI_Message *modified_message = R_NEW0 (R2AI_Message);
+			if (modified_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';
+					}
+				}
+				modified_message->content = content;
+
+				// Set up proper tool call structure
+				modified_message->tool_calls = R_NEWS0 (R2AI_ToolCall, 1);
+				if (modified_message->tool_calls) {
+					modified_message->tool_calls[0].name = tool_name;
+					modified_message->tool_calls[0].arguments = tool_args;
+					char id_buf[32];
+					snprintf (id_buf, sizeof (id_buf), "rawtool_%d", (int)time (NULL));
+					modified_message->tool_calls[0].id = strdup (id_buf);
+					modified_message->n_tool_calls = 1;
+				}
+				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);
+			if (modified_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;
+		}
+
+		return response;
+	}
+
+	// No tool call found
+	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);
+
+		// 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->uses_anthropic_header) {
+			fallback_response = r2ai_anthropic (cps, fallback_args);
+		} else {
+			fallback_response = r2ai_openai (cps, fallback_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);
+		}
+
+		// 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;
+		R_LOG_WARN ("LLM provided no response content");
+		return warning_response;
+	}
+
+	return response;
+}

src/tools.c

@@ -19,7 +19,7 @@ static R2AI_Tool r2cmd_tool = {
 
 static R2AI_Tool qjs_tool = {
 	.name = "execute_js",
-	.description = "Execute a JavaScript script in a quickjs environment. Only what you console.log will be returned.",
+	.description = "Execute a JavaScript script inside the radare2 environment. Use `var RESULT=r2.cmd(COMMAND)` to execute radare2 commands and `r2.log(MESSAGE)` instead of console.log`",
 	.parameters = "{\
 		\"type\": \"object\",\
 		\"properties\": {\
@@ -284,31 +284,19 @@ R_API char *r2ai_r2cmd(RCore *core, RJson *args, bool hide_tool_output, char **e
 		if (is_multiline) {
 			// Use editor for multi-line commands
 			*edited_command = strdup (command);
-#if R2_VERSION_NUMBER >= 50909
 			r_cons_editor (core->cons, NULL, *edited_command);
-#else
-			r_cons_editor (NULL, *edited_command);
-#endif
 			command = *edited_command;
 		} else {
 			// For single-line commands, push the command to input buffer
-			R2_NEWLINE ();
-			R2_PRINTF ("\x1b[31m[edit cmd]>\x1b[0m ");
-			R2_FLUSH ();
+			r_cons_newline (core->cons);
 			// Push the command to the input buffer
 
 			// Get user input with command pre-filled
-#if R2_VERSION_NUMBER >= 50909
 			r_cons_readpush (core->cons, command, strlen (command));
 			r_cons_readpush (core->cons, "\x05", 1); // Ctrl+E - move to end
 			const char *readline_result = r_line_readline (core->cons);
-#else
-			r_cons_readpush (command, strlen (command));
-			r_cons_readpush ("\x05", 1); // Ctrl+E - move to end
-			const char *readline_result = r_line_readline ();
-#endif
 			// Check if interrupted or ESC pressed (readline_result is NULL or empty)
-			if (R2_INTERRUPTED () || R_STR_ISEMPTY (readline_result)) {
+			if (r_cons_is_breaked (core->cons) || R_STR_ISEMPTY (readline_result)) {
 				R_LOG_INFO ("Command execution cancelled %s", readline_result);
 				return strdup ("R2AI_SIGINT");
 			}
@@ -330,8 +318,8 @@ R_API char *r2ai_r2cmd(RCore *core, RJson *args, bool hide_tool_output, char **e
 
 	if (!hide_tool_output) {
 		char *red_command = r_str_newf ("\x1b[31m%s\x1b[0m\n", *edited_command);
-		R2_PRINTF ("%s", red_command);
-		R2_FLUSH ();
+		r_cons_printf (core->cons, "%s", red_command);
+		r_cons_flush (core->cons);
 		free (red_command);
 	}
 
@@ -377,29 +365,17 @@ R_API char *r2ai_qjs(RCore *core, RJson *args, bool hide_tool_output) {
 		if (is_multiline) {
 			// Use editor for multi-line scripts
 			edited_script = strdup (script);
-#if R2_VERSION_NUMBER >= 50909
 			r_cons_editor (core->cons, NULL, edited_script);
-#else
-			r_cons_editor (NULL, edited_script);
-#endif
 			script = edited_script;
 		} else {
 			// For single-line scripts, push the script to input buffer
-			R2_PRINTF ("\x1b[31m[edit js]>\x1b[0m ");
-			R2_FLUSH ();
 
-#if R2_VERSION_NUMBER >= 50909
 			r_cons_readpush (core->cons, script, strlen (script));
 			r_cons_readpush (core->cons, "\x05", 1); // Ctrl+E - move to end
 			const char *readline_result = r_line_readline (core->cons);
-#else
-			r_cons_readpush (script, strlen (script));
-			r_cons_readpush ("\x05", 1); // Ctrl+E - move to end
-			const char *readline_result = r_line_readline ();
-#endif
 
 			// Check if interrupted or ESC pressed (readline_result is NULL or empty)
-			if (R2_INTERRUPTED () || R_STR_ISEMPTY (readline_result)) {
+			if (r_cons_is_breaked (core->cons) || R_STR_ISEMPTY (readline_result)) {
 				free (edited_script); // Free if already allocated
 				return strdup ("R2AI_SIGINT");
 			}
@@ -419,8 +395,8 @@ R_API char *r2ai_qjs(RCore *core, RJson *args, bool hide_tool_output) {
 	if (!hide_tool_output) {
 		char *print_script = r_str_newf ("\n```js\n%s\n```", script);
 		char *print_script_rendered = r2ai_markdown (print_script);
-		R2_PRINTF ("%s\n", print_script_rendered);
-		R2_FLUSH ();
+		r_cons_printf (core->cons, "%s\n", print_script_rendered);
+		r_cons_flush (core->cons);
 		free (print_script);
 		free (print_script_rendered);
 	}
@@ -435,11 +411,7 @@ R_API char *r2ai_qjs(RCore *core, RJson *args, bool hide_tool_output) {
 		return strdup ("{ \"res\":\"Failed to create script payload\" }");
 	}
 
-#if R2_VERSION_NUMBER >= 50909
 	char *payload_base64 = r_base64_encode_dyn ((const ut8 *)payload, strlen (payload));
-#else
-	char *payload_base64 = r_base64_encode_dyn ((const char *)payload, strlen (payload));
-#endif
 	if (!payload_base64) {
 		free (payload);
 		return strdup ("{ \"res\":\"Failed to encode script\" }");
@@ -481,12 +453,6 @@ R_API char *execute_tool(RCore *core, const char *tool_name, const char *args, c
 	}
 
 	bool hide_tool_output = r_config_get_b (core->config, "r2ai.auto.hide_tool_output");
-
-	char *print_name = r_str_newf ("\x1b[1;32m\x1b[4m[%s]>\x1b[0m ", tool_name);
-	R2_PRINTF ("%s\n", print_name);
-	R2_FLUSH ();
-
-	free (print_name);
 	char *tool_result = NULL;
 
 	if (strcmp (tool_name, "r2cmd") == 0) {
@@ -627,10 +593,10 @@ R_API char *execute_tool(RCore *core, const char *tool_name, const char *args, c
 		return r_str_newf ("Error running tool: Unknown error\nCommand: %s", args);
 	}
 	if (!hide_tool_output) {
-		R2_NEWLINE ();
-		R2_PRINTF ("%s\n", result);
-		R2_NEWLINE ();
-		R2_FLUSH ();
+		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);

src/vdb.c

@@ -19,7 +19,7 @@ static void vector_free(Vector *v) {
 
 /*-------------------------------
  * Distance Function Using Squared Euclidean Distance
- * For two unit vectors: dist^2 = 2 - 2 * (dot product)
+ * For two unit vectors: dist^2 = 2 - 2 *(dot product)
  *-------------------------------*/
 static float squared_distance(const Vector *a, const Vector *b) {
 	if (a->dim != b->dim) {
@@ -34,14 +34,14 @@ static float squared_distance(const Vector *a, const Vector *b) {
 }
 
 /*-------------------------------
-  KDNode Utility Functions
-  -------------------------------*/
+KDNode Utility Functions
+-------------------------------*/
 static KDNode *create_kdnode(const Vector *v, const char *text, int split_dim) {
 	KDNode *node = (KDNode *)malloc (sizeof (KDNode));
 	node->point.dim = v->dim;
 	node->point.data = (float *)malloc (sizeof (float) * v->dim);
 	memcpy (node->point.data, v->data, sizeof (float) * v->dim);
-	node->text = text ? strdup (text) : NULL;
+	node->text = text? strdup (text): NULL;
 	node->split_dim = split_dim;
 	node->left = NULL;
 	node->right = NULL;
@@ -212,8 +212,8 @@ void kd_search_knn_recursive(KDNode *node, const Vector *query, RVdbResultSet *r
 
 	int axis = node->split_dim;
 	float diff = query->data[axis] - node->point.data[axis];
-	KDNode *first = (diff < 0) ? node->left : node->right;
-	KDNode *second = (diff < 0) ? node->right : node->left;
+	KDNode *first = (diff < 0)? node->left: node->right;
+	KDNode *second = (diff < 0)? node->right: node->left;
 
 	kd_search_knn_recursive (first, query, rs, depth + 1, dim);
 

src/vdb_embed.inc.c

@@ -136,7 +136,7 @@ static void compute_embedding(RVdb *db, const char *text, float *embedding, unsi
 		// Compute term frequency: tf = 1 + log (token_count)
 		float tf = 1.0f + log ((float)dt->count);
 		RVdbToken *t = gtfidf_find (db->tokens, dt->token);
-		float df_value = t ? t->df : 1.0f;
+		float df_value = t? t->df: 1.0f;
 		// Compute inverse document frequency;
 		float idf = log (((float)db->total_docs + 1.0f) / ((float)df_value + 1.0f)) + 1.0f;
 		float weight = tf * idf;