Snapshot async request configuration for HTTP sessions

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

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

.github/dependabot.yml

@@ -0,0 +1,8 @@
+version: 2
+updates:
+  - package-ecosystem: "github-actions"
+    directory: "/"
+    schedule:
+      interval: "weekly"
+    labels:
+      - buildsystem

.github/workflows/ci.yml

@@ -1,47 +1,170 @@
-# TODO: add windows
-name: ci
+name: CI and Release
 
 env:
-  R2V: 5.9.0
+  R2V: 6.2.6
 
 on:
   push:
-  pull_request:
-  workflow_dispatch:
+    branches:
+      - '**'
+    tags:
+      - '[0-9]*'
+#  pull_request:
 
 jobs:
-  build-linux:
-    runs-on: ubuntu-22.04
+  build:
+    name: Build (${{ matrix.os }} / ${{ matrix.build_system }})
+    runs-on: ${{ matrix.os }}
+    strategy:
+      fail-fast: false
+      matrix:
+        os: [ubuntu-24.04, macos-latest, windows-latest]
+        build_system: [make, meson]
+        exclude:
+          - os: windows-latest
+            build_system: make
+
     steps:
-    - uses: actions/checkout@v4
-    - name: Installing radare2
-      run: |
-        wget "https://github.com/radareorg/radare2/releases/download/${R2V}/radare2_${R2V}_amd64.deb"
-        wget "https://github.com/radareorg/radare2/releases/download/${R2V}/radare2-dev_${R2V}_amd64.deb"
-        sudo dpkg -i "radare2_${R2V}_amd64.deb"
-        sudo dpkg -i "radare2-dev_${R2V}_amd64.deb"
-    - name: Setting up the Python virtual environment
-      run: |
-        python -m venv venv
-        . venv/bin/activate
-        pip install pylint
-    - name: Install Python dependencies
-      run: |
-        . venv/bin/activate
-        pip install -r requirements.txt
-    - name: Linting
-      run: |
-        . venv/bin/activate
-        make cilint
-    - name: Running r2ai from the shell
-      run: |
-        . venv/bin/activate
-        python main.py -h
-    - name: Running r2ai from r2pipe
-      run: |
-        . venv/bin/activate
-        r2 -q -c '#!pipe python main.py -h' /bin/ls
-    - name: Running r2ai from r2
-      run: |
-        . venv/bin/activate
-        r2 -q -i main.py -c r2ai /bin/ls
+      - name: Checkout
+        uses: actions/checkout@v7
+
+      - name: Setup Meson/Ninja
+        if: matrix.build_system == 'meson'
+        run: |
+          python -m pip install --upgrade pip
+          pip install meson ninja
+
+      - name: Install radare2
+        id: r2
+        uses: radareorg/github-actions@v3
+        with:
+          version: ${{ env.R2V }}
+
+      - name: Build with Make
+        if: matrix.build_system == 'make'
+        run: make -C src
+
+      - name: Build with Meson (Unix)
+        if: matrix.build_system == 'meson' && matrix.os != 'windows-latest'
+        run: |
+          cd src
+          meson setup build || meson configure build
+          meson compile -C build
+
+      - name: Build with Meson (Windows)
+        if: matrix.build_system == 'meson' && matrix.os == 'windows-latest'
+        shell: powershell
+        run: |
+          cd src
+          $prefix = "${{ steps.r2.outputs.prefix }}"
+          meson setup build --backend ninja --prefix="$prefix" -Dr2_prefix="$prefix"
+          if ($LASTEXITCODE -ne 0) { meson configure build }
+          meson compile -C build
+
+  bindist:
+    name: Build r2pm binary (${{ matrix.os }})
+    strategy:
+      fail-fast: false
+      matrix:
+        os: [ubuntu-latest, ubuntu-24.04-arm, macos-latest]
+    runs-on: ${{ matrix.os }}
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v7
+
+      - name: Install radare2
+        uses: radareorg/github-actions@v3
+        with:
+          version: ${{ env.R2V }}
+
+      - name: Package r2pm binary
+        run: make -C src bindist
+
+      - name: Upload r2pm binary artifact
+        uses: actions/upload-artifact@v7
+        with:
+          name: bindist-${{ matrix.os }}
+          path: r2ai-*.zip
+          if-no-files-found: error
+
+  windows-bindist:
+    name: Build r2pm binary (Windows)
+    runs-on: windows-latest
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v7
+
+      - name: Setup Meson/Ninja
+        run: |
+          python -m pip install --upgrade pip
+          pip install meson ninja
+
+      - name: Install radare2
+        id: r2
+        uses: radareorg/github-actions@v3
+        with:
+          version: ${{ env.R2V }}
+
+      - name: Build with Meson
+        shell: powershell
+        run: |
+          cd src
+          $prefix = "${{ steps.r2.outputs.prefix }}"
+          meson setup build --backend ninja --prefix="$prefix" -Dr2_prefix="$prefix"
+          if ($LASTEXITCODE -ne 0) { meson configure build }
+          meson compile -C build
+
+      - name: Package r2pm binary
+        shell: pwsh
+        run: |
+          New-Item -ItemType Directory -Force -Path bindist\plugins,bindist\bin,bindist\prompts | Out-Null
+          Get-ChildItem src\build\ -Filter "*r2ai*.dll" -ErrorAction SilentlyContinue | Copy-Item -Destination bindist\plugins
+          Get-ChildItem src\..\build\ -Filter "*r2ai*.dll" -ErrorAction SilentlyContinue | Copy-Item -Destination bindist\plugins
+          Copy-Item src\build\r2ai.exe bindist\bin
+          Copy-Item prompts\*.r2ai.md bindist\prompts
+          $rv = (& radare2 -qv).Trim()
+          Compress-Archive -Path bindist\* -DestinationPath "r2ai-$rv-windows-x86-64.zip" -Force
+
+      - name: Upload r2pm binary artifact
+        uses: actions/upload-artifact@v7
+        with:
+          name: windows-bindist
+          path: r2ai-*.zip
+          if-no-files-found: error
+
+  release:
+    name: Create Release
+    needs: [bindist, windows-bindist]
+    if: startsWith(github.ref, 'refs/tags/')
+    runs-on: ubuntu-latest
+    permissions:
+      contents: write
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v7
+
+      - name: Verify tag is on master
+        run: |
+          git fetch origin master:refs/remotes/origin/master
+          if ! git branch -r --contains ${GITHUB_SHA} | grep -q "origin/master"; then
+            echo "Tag commit is not on origin/master — skipping release."
+            exit 0
+          fi
+
+      - name: Download r2pm binary artifacts
+        uses: actions/download-artifact@v8
+        with:
+          pattern: '*bindist*'
+          path: release-artifacts
+          merge-multiple: true
+
+      - name: Add decai.r2.js to release artifacts
+        run: cp decai/decai.r2.js release-artifacts/
+
+      - name: Create GitHub Release
+        uses: softprops/action-gh-release@v3
+        with:
+          files: |
+            release-artifacts/*.zip
+            release-artifacts/decai.r2.js
+          generate_release_notes: true

.gitignore

@@ -1,2 +1,17 @@
 build
-venv
+b
+src/b
+src/r2ai
+src/r2ai.dSYM
+venv
+**/__pycache__
+*.pyc
+*.pyo
+*.o
+*.so
+*.dylib
+*.a
+third_party
+
+# r2pm binary packages
+r2ai-*.zip

Attic/examples/ai.sh

@@ -0,0 +1,162 @@
+#!/bin/sh
+OPENAPI_HOST=localhost
+OPENAPI_PORT=8080
+
+if [ -z "${OLLAMA_HOST}" ]; then
+	OLLAMA_HOST=localhost
+fi
+if [ -z "${OLLAMA_PORT}" ]; then
+	OLLAMA_PORT=11434
+fi
+if [ -z "${OLLAMA_MODEL}" ]; then
+	OLLAMA_MODEL="llama3.2:1b"
+fi
+
+GEMINI_KEY=""
+GEMINI_MODEL="gemini-1.5-flash"
+if [ -f ~/.r2ai.gemini-key ]; then
+	GEMINI_KEY=$(cat ~/.r2ai.gemini-key)
+fi
+OPENAI_KEY=""
+OPENAI_MODEL="gpt-4o"
+if [ -f ~/.r2ai.openai-key ]; then
+	OPENAI_KEY=$(cat ~/.r2ai.openai-key)
+fi
+CLAUDE_KEY=""
+CLAUDE_MODEL="claude-sonnet-5-5"
+if [ -f ~/.r2ai.anthropic-key ]; then
+	CLAUDE_KEY=$(cat ~/.r2ai.anthropic-key)
+fi
+DEEPSEEK_KEY=""
+DEEPSEEK_MODEL="claude-sonnet-5-5"
+if [ -f ~/.r2ai.deepseek-key ]; then
+	DEEPSEEK_KEY=$(cat ~/.r2ai.anthropic-key)
+fi
+
+SCISSORS="------------8<------------"
+
+read_INPUT() {
+	export INPUT=`(echo "$ARG" ; echo "<INPUT>"; cat ; echo "</INPUT>" ) | jq -R -s .`
+		echo "$INPUT"
+}
+
+claude() {
+	read_INPUT
+	PAYLOAD="
+	{
+	\"model\": \"${CLAUDE_MODEL}\",
+	\"max_tokens\": 5128,
+	\"messages\": [ { \"role\": \"user\", \"content\": ${INPUT} } ]
+	}
+	"
+	echo "$SCISSORS"
+	curl -s https://api.anthropic.com/v1/messages \
+		-H "Content-Type: application/json" \
+		-H "anthropic-version: 2023-06-01" \
+		-H "x-api-key: ${CLAUDE_KEY}" \
+		-d "`printf '%s\n' \"${PAYLOAD}\"`" | jq -r '.content[0].text'
+	echo "$SCISSORS"
+}
+
+ollama() {
+	read_INPUT
+	PAYLOAD="{ \"stream\":false, \"model\":\"${OLLAMA_MODEL}\", \"messages\": [{\"role\":\"user\", \"content\": ${INPUT} }]}"
+	echo "$SCISSORS"
+	curl -s "http://${OLLAMA_HOST}:${OLLAMA_PORT}/api/chat" \
+		-H "Content-Type: application/json" \
+		-d "`printf '%s\n' \"${PAYLOAD}\"`" | jq -r .message.content
+	echo "$SCISSORS"
+}
+
+openapi() {
+	read_INPUT
+	PAYLOAD="{ \"prompt\": ${INPUT} }"
+	echo "$SCISSORS"
+	curl -s "http://${OPENAPI_HOST}:${OPENAPI_PORT}/completion" \
+		-H "Content-Type: application/json" \
+		-d "`printf '%s\n' \"${PAYLOAD}\"`" | jq -r .content
+	echo "$SCISSORS"
+}
+
+openai() {
+	read_INPUT
+	PAYLOAD="
+	{
+	\"model\": \"${OPENAI_MODEL}\",
+	\"max_completion_tokens\": 5128,
+	\"messages\": [ { \"role\": \"user\", \"content\": ${INPUT} } ]
+	}
+	"
+	echo "$SCISSORS"
+	curl -s https://api.openai.com/v1/chat/completions \
+		-H "Content-Type: application/json" \
+		-H "Authorization: Bearer ${OPENAI_KEY}" \
+		-d "`printf '%s\n' \"${PAYLOAD}\"`" | jq -r '.choices[0].message.content'
+	echo "$SCISSORS"
+}
+
+gemini() {
+	read_INPUT
+	PAYLOAD=" {
+	\"contents\":[{
+          \"parts\":[
+            {\"text\": ${INPUT}}
+          ] }] }"
+	echo "$SCISSORS"
+	curl -s -X POST "https://generativelanguage.googleapis.com/v1beta/models/gemini-1.5-flash:generateContent?key=${GEMINI_KEY}" \
+		-H "Content-Type: application/json" \
+		-d "`printf '%s\n' \"${PAYLOAD}\"`" | jq -r .candidates[0].content.parts[0].text
+	echo "$SCISSORS"
+}
+
+deepseek() {
+	read_INPUT
+	PAYLOAD=" {
+	\"model\":\"deepseek-chat\",
+	\"stream\":\"false\",
+	\"messages\":[{
+            {\"role\": \"user\", \"content\": ${INPUT}}
+          }]}"
+	echo "$SCISSORS"
+	curl -s -X POST "https://api.deepseek.com/chat/completions" \
+		-H "Authorization: Bearer ${DEEPSEEK_KEY}" \
+		-H "Content-Type: application/json" \
+		-d "`printf '%s\n' \"${PAYLOAD}\"`" | jq -r .choices[0].message.content
+	echo "$SCISSORS"
+}
+
+show_help() {
+	cat <<EOF
+$ ai [--] | [-h] | [prompt] < INPUT
+-h = show this help message
+-- = don't display the ---8<--- lines in the output
+SHAI_API = ollama | gemini | claude | openai
+OLLAMA_MODEL=hf.co/mradermacher/salamandra-7b-instruct-aina-hack-GGUF:salamandra-7b-instruct-aina-hack.Q4_K_M.gguf
+OLLAMA_HOST=localhost
+OLLAMA_PORT=11434
+GEMINI_KEY=~/.r2ai-gemini.key
+OPENAI_KEY=~/.r2ai-openai.key
+CLAUDE_KEY=~/.r2ai-anthropic.key
+DEEPSEEK_KEY=~/.r2ai-deepseek.key
+CLAUDE_MODEL=claude-sonnet-5-5
+EOF
+	exit 0
+}
+
+[ "$1" = "-h" ] && show_help
+
+if [ "$1" = "--" ]; then
+	SCISSORS=""
+	shift
+fi
+
+export ARG="$@"
+case "${SHAI_API}" in
+gemini|google) gemini ; ;;
+deepseek) deepseek ; ;;
+openapi) openapi ; ;;
+claude) claude ; ;;
+ollama) ollama ; ;;
+openai) openai ; ;;
+*) claude ; ;;
+esac

Attic/examples/benchmark/Makefile

@@ -0,0 +1,14 @@
+SHELL=/bin/bash
+
+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
+LANGS=Dart Perl Ruby Javascript PHP C
+
+all:
+	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 ; \
+		./run.sh "$${ARGS[@]}" > ${a//\//_}.html ; done

Attic/examples/benchmark/decai/BENCH.txt

@@ -0,0 +1,4 @@
+<b>
+decai -e model=$a;decai -e lang=$l;s 0x100000924;af;decai -d
+</b><br/>
+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

Attic/examples/benchmark/index.sh

@@ -0,0 +1,49 @@
+#!/bin/bash
+
+# Find all directories containing BENCH.txt
+dirs=$(find . -name "BENCH.txt" -type f | xargs dirname | sort | uniq)
+
+# Start HTML output
+cat <<EOF > benchmark_index.html
+<html>
+<table>
+EOF
+
+# Process each directory
+for dir in $dirs; do
+    cat <<EOF >> benchmark_index.html
+	<tr> <td>
+	<div>
+EOF
+
+    # Add contents of BENCH.txt
+    cat "$dir/BENCH.txt" >> benchmark_index.html
+
+    cat <<EOF >> benchmark_index.html
+	</div>
+	</td>
+	<td>
+EOF
+
+    # Find and add links to HTML files
+    for html_file in $(ls "$dir"/*.html 2>/dev/null | sort); do
+        base=$(basename "$html_file" .html)
+        cat <<EOF >> benchmark_index.html
+	<a href="$dir/$base.html">$base</a>
+	<br />
+EOF
+    done
+
+    cat <<EOF >> benchmark_index.html
+	</td>
+	</tr>
+EOF
+done
+
+# End HTML output
+cat <<EOF >> benchmark_index.html
+</table>
+<br />
+<br />
+</html>
+EOF

Attic/examples/benchmark/run.sh

@@ -0,0 +1,177 @@
+#!/bin/bash
+
+# Generate HTML output for command executions with collapsible sections and width control
+
+if [ $# -eq 0 ]; then
+    echo "Usage: $0 command1 [command2 ...]"
+    exit 1
+fi
+
+# Timeout in seconds for each command (configurable via TIMEOUT environment variable)
+TIMEOUT=${TIMEOUT:-30}
+
+cat << 'EOF'
+<!DOCTYPE html>
+<html lang="en">
+<head>
+    <meta charset="UTF-8">
+    <meta name="viewport" content="width=device-width, initial-scale=1.0">
+    <title>Command Execution Results</title>
+    <style>
+        * {
+            box-sizing: border-box;
+        }
+        :root {
+            --max-width: 400px;
+        }
+        body {
+            font-family: monospace;
+            margin: 0;
+            padding: 20px;
+            background-color: #f5f5f5;
+            overflow-x: hidden;
+        }
+        .slider-container {
+            background: white;
+            padding: 15px;
+            margin-bottom: 20px;
+            box-shadow: 0 2px 4px rgba(0,0,0,0.1);
+            width: 100%;
+        }
+        .slider-container label {
+            display: block;
+            margin-bottom: 10px;
+            font-weight: bold;
+        }
+        .slider-container input {
+            width: 100%;
+        }
+        .commands-container {
+            display: flex;
+            flex-wrap: wrap;
+            gap: 15px;
+            align-items: flex-start;
+        }
+        .command-section {
+            background: white;
+            border-radius: 8px;
+            box-shadow: 0 2px 4px rgba(0,0,0,0.1);
+            overflow: hidden;
+            width: var(--max-width);
+        }
+        .command-title {
+            background: #e9ecef;
+            padding: 12px 15px;
+            cursor: pointer;
+            border: none;
+            width: 100%;
+            text-align: left;
+            font-family: inherit;
+            font-size: inherit;
+            border-radius: 8px 8px 0 0;
+            position: relative;
+        }
+        .command-title:hover {
+            background: #dee2e6;
+        }
+        .command-time {
+            position: absolute;
+            top: 12px;
+            right: 15px;
+            color: #6c757d;
+            font-size: 0.9em;
+        }
+        .command-body {
+            padding: 15px;
+            border-top: 1px solid #dee2e6;
+            display: none;
+        }
+        .command-output {
+            background: #f8f9fa;
+            padding: 10px;
+            border-radius: 4px;
+            white-space: pre-wrap;
+            word-wrap: break-word;
+            overflow-x: auto;
+        }
+        .command-section.expanded .command-body {
+            display: block;
+        }
+    </style>
+</head>
+<body>
+    <div class="slider-container">
+        <label for="width-slider">Maximum Width: <span id="width-value">400</span>px</label>
+        <input type="range" id="width-slider" min="200" max="2000" value="400" step="50">
+    </div>
+    <div class="commands-container">
+EOF
+
+# Process each command
+for cmd in "$@"; do
+    # Extract title: if command starts with #, use the rest as title, else use the command
+    if [[ "$cmd" =~ ^#[[:space:]]*(.*) ]]; then
+        title="${BASH_REMATCH[1]}"
+        exec_cmd="${cmd#*#}"
+        exec_cmd="${exec_cmd#"${exec_cmd%%[![:space:]]*}"}"  # trim leading spaces
+    else
+        title="$cmd"
+        exec_cmd="$cmd"
+    fi
+
+    # Execute command and time it
+    start_time=$(date +%s.%N)
+    tmpfile=$(mktemp)
+    eval "$exec_cmd" > "$tmpfile" 2>&1 &
+    pid=$!
+    (sleep $TIMEOUT && kill $pid 2>/dev/null) &
+    wait $pid 2>/dev/null
+    output=$(cat "$tmpfile")
+    rm "$tmpfile"
+    exit_code=$?
+    end_time=$(date +%s.%N)
+    
+    # Calculate execution time
+    execution_time=$(echo "$end_time - $start_time" | bc -l 2>/dev/null || echo "0")
+    if [ $? -ne 0 ]; then
+        execution_time="N/A"
+    else
+        execution_time=$(printf "%.3f" "$execution_time")
+    fi
+
+    # Escape HTML in output
+    escaped_output=$(echo "$output" | sed 's/&/\&amp;/g; s/</\&lt;/g; s/>/\&gt;/g')
+
+    cat << EOF
+        <div class="command-section expanded">
+            <button class="command-title" onclick="toggleSection(this)">$title<span class="command-time">${execution_time}s</span></button>
+            <div class="command-body">
+                <div class="command-output">$escaped_output</div>
+            </div>
+        </div>
+EOF
+done
+
+cat << 'EOF'
+    </div>
+
+    <script>
+        // Slider functionality
+        const slider = document.getElementById('width-slider');
+        const widthValue = document.getElementById('width-value');
+        
+        slider.addEventListener('input', function() {
+            const value = this.value;
+            document.documentElement.style.setProperty('--max-width', value + 'px');
+            widthValue.textContent = value;
+        });
+
+        // Toggle functionality
+        function toggleSection(button) {
+            const section = button.parentElement;
+            section.classList.toggle('expanded');
+        }
+    </script>
+</body>
+</html>
+EOF

Attic/examples/gemini.py

@@ -0,0 +1,6 @@
+import google.generativeai as genai
+
+genai.configure(api_key="YOUR_API_KEY")
+model = genai.GenerativeModel("gemini-1.5-flash")
+response = model.generate_content("Explain how AI works")
+print(response.text)

Attic/examples/llama-vim.sh

@@ -0,0 +1,4 @@
+#!/bin/sh
+r2pm -r llama-server  \
+    --port 8012 -ngl 99 -fa -ub 1024 -b 1024 -dt 0.1 \
+    --ctx-size 0 --cache-reuse 256 -m ~/.r2ai.models/qwen2.5-coder-7b-instruct-q3_k_m.gguf

Attic/examples/podcast.sh

@@ -0,0 +1,15 @@
+#!/bin/sh
+#
+# Launch r2ai and use this prompt to generate the conversation from data.txt
+# Copypaste the output into podcast.txt
+#
+# -f /tmp/data.txt Create a script for a podcast between two technical people named Sam and Max, about the changelog on radare2 5.9.4, focus on important features and stuff that impacts positively to users. The length must be as long as possible. Do not show any "**" section, output must contain only the conversation, be emotional and make the two people ask questions to learn more about the details of each feature
+
+if [ ! -f podcast.txt ]; then
+ echo "Missing podcast.txt. Please read the script before running it."
+ exit 1
+fi
+
+cat podcast.txt | sed -e 's/Sam:/[[pbas 40]]/g' -e 's/Max:/[[pbas 60]]/g' > podcast.say.txt
+say -f podcast.say.txt -o podcast.aiff
+ffmpeg -i podcast.aiff podcast.mp3

Attic/examples/scrap-ddg.sh

@@ -0,0 +1,3 @@
+#!/bin/sh
+# pip install ddgr
+ddgr --json $@

Attic/examples/scrap-ddgweb.sh

@@ -0,0 +1,5 @@
+#!/bin/sh
+URLS=`./scrap-ddg.sh -n2 -- $@ |jq '.[].url'`
+for URL in ${URLS}; do
+	eval ./scrap-web.sh ${URL}
+done

Attic/examples/scrap-web.sh

@@ -7,4 +7,4 @@ fi
 # https://github.com/aaronsw/html2text =>
 # https://raw.githubusercontent.com/aaronsw/html2text/master/README.md
 
-curl -s "$1" | html2text -width 1024 -utf8 -nobs | grep -E '.{80}' | grep -v '=====' | grep -v '\*\*\*\*'|less
+curl -s "$1" | html2text -width 1024 -utf8 -nobs | grep -E '.{80}' | grep -v '=====' | grep -v '\*\*\*\*' | grep -v -e '^\d' | sed -r 's/^[[:space:]]+|[[:space:]]+$//g' | grep -v ://

Attic/examples/socialai.py

@@ -0,0 +1,68 @@
+import llama_cpp
+
+model_path="/Users/pancake/.r2ai.models/";
+# model_path+="unsloth.Q4_K_M.gguf"
+# model_path+= "llama-3-tsuki-unsloth-8b.Q5_K_M.gguf"
+model_path+= "llama-2-7b-chat-codeCherryPop.Q5_K_M.gguf"
+# model_path += "mistral-7b-instruct-v0.2.Q5_K_M.gguf"
+peers = [
+	["@kelsy", "act as a twitter user responding in one short sentence your first though on my messages, be constructive and help me discuss ideas"],
+	["@john", "act as a twitter user, be concise, respond in one short sentence be funny, help me reason my plans"],
+	["@anna", "respond in one short sentence with philosophical reasoning on my message"],
+	["@tony", "behave like a shy twitter user, respond with one or two short sentences, as a software developer, respond in short but wise sentence reasoning the best plans for implementing the topic"]
+]
+
+logs = []
+
+ai = llama_cpp.Llama(model_path=model_path, verbose=False, n_ctx=8096)
+
+def context(msg):
+	global logs
+	ats = [word for word in msg.split() if word.startswith('@')]
+	ctx = []
+	for log in logs:
+		if log == msg:
+			continue
+		if len(ats) > 0 and any(at in log for at in ats):
+			ctx.append(log)
+		if not log.startswith("@"):
+		 	ctx.append(log)
+	return ctx
+		
+def sortedpeers(msg):
+	global peers
+	ats = [word for word in msg.split() if word.startswith('@')]
+	if len(ats) == 0:
+		return peers
+	ps = []
+	for peer in peers:
+		if any(at in peer[0] for at in ats):
+			ps.insert(0, peer)
+		else:
+			ps.append(peer)
+	return ps
+
+def chat(msg):
+	logs.append(msg)
+	global ai
+	for peer in sortedpeers(msg):
+		ctx = ",".join(context(msg))
+		m = f"[INST]{peer[1]}[/INST] {msg}"
+		m = f"[INST]{peer[1]}[/INST] Consider this context: {ctx}. Respond to: {msg}```"
+		m = f"[INST]{peer[1]}[/INST] <s>{ctx}</s>. Respond in one sentence to: {msg}```"
+		mm = ai(m, max_tokens=-1)
+		r = mm["choices"][0]["text"]
+		r = "".join(r.split("\n"))
+		reply = f"{peer[0]}: {r}"
+		logs.append(reply)
+		print(f"\x1b[31m{reply}\x1b[0m")
+
+# res = ai("Hello")
+# print(res["choices"][0]["text"])
+
+
+while True:
+	msg = input()
+	if not msg:
+		break
+	chat(msg)

Attic/examples/substitute.c

@@ -0,0 +1,66 @@
+#include <stdio.h>
+#include <string.h>
+#include <ctype.h>
+
+#define ALPHABET_SIZE 26
+
+// Function to perform substitution (both encryption and decryption)
+void substitute(const char *input, const char *substitution, char *output, int reverse) {
+    char mapping[ALPHABET_SIZE];
+
+    if (reverse) {
+        // Create reverse substitution table for decryption
+        for (int i = 0; i < ALPHABET_SIZE; i++) {
+            mapping[substitution[i] - 'A'] = 'A' + i;
+        }
+    } else {
+        // Use the direct substitution table for encryption
+        for (int i = 0; i < ALPHABET_SIZE; i++) {
+            mapping[i] = substitution[i];
+        }
+    }
+
+    for (int i = 0; input[i] != '\0'; i++) {
+        char c = input[i];
+
+        // Handle uppercase letters
+        if (isupper(c)) {
+            output[i] = mapping[c - 'A'];
+        }
+        // Handle lowercase letters
+        else if (islower(c)) {
+            output[i] = tolower(mapping[c - 'a']);
+        }
+        // Leave non-alphabet characters unchanged
+        else {
+            output[i] = c;
+        }
+    }
+
+    output[strlen(input)] = '\0';  // Null-terminate the output string
+}
+
+int main() {
+    char input[256];
+    char output[256];
+
+    // Configurable substitution key (A-Z)
+    const char *substitution = "QWERTYUIOPASDFGHJKLZXCVBNM";  // Example substitution key
+
+    // Get input from the user
+    printf("Enter text to encrypt: ");
+    fgets(input, sizeof(input), stdin);
+
+    // Remove newline character from input if it exists
+    input[strcspn(input, "\n")] = 0;
+
+    // Perform encryption (forward substitution)
+    substitute(input, substitution, output, 0);
+    printf("Encrypted text: %s\n", output);
+
+    // Perform decryption (reverse substitution)
+    substitute(output, substitution, input, 1);
+    printf("Decrypted text: %s\n", input);
+
+    return 0;
+}

Attic/llmcom.py

@@ -0,0 +1,28 @@
+from transformers import AutoTokenizer
+import transformers
+import torch
+
+model = "facebook/llm-compiler-13b"
+model = "QuantFactory/llm-compiler-7b-GGUF"
+# model = "file:///tmp/llm-compiler.gguf"
+
+tokenizer = AutoTokenizer.from_pretrained(model)
+pipeline = transformers.pipeline(
+    "text-generation",
+    model=model,
+    torch_dtype=torch.float16,
+    device_map="auto",
+)
+
+sequences = pipeline(
+    '%3 = alloca i32, align 4',
+    do_sample=True,
+    top_k=10,
+    temperature=0.1,
+    top_p=0.95,
+    num_return_sequences=1,
+    eos_token_id=tokenizer.eos_token_id,
+    max_length=200,
+)
+for seq in sequences:
+    print(f"Result: {seq['generated_text']}")

Attic/pancake-vdb.swift

@@ -0,0 +1,249 @@
+import Foundation
+
+// Token representation
+struct Token {
+    var token: String
+    var count: Int
+    var df: Int
+}
+
+// KD-Tree Node
+class KDNodeData {
+    let point: [Float]
+    let text: String
+    var left: KDNodeData?
+    var right: KDNodeData?
+    let splitDim: Int
+    
+    init(point: [Float], text: String, left: KDNodeData? = nil, right: KDNodeData? = nil, splitDim: Int) {
+        self.point = point
+        self.text = text
+        self.left = left
+        self.right = right
+        self.splitDim = splitDim
+    }
+}
+
+// Result for K-Nearest Neighbors
+struct KNNResult {
+    let node: KDNodeData
+    let distanceSq: Float
+}
+
+class VectorDB {
+    // Core database structures
+    private let dimension: Int
+    private var root: KDNodeData?
+    private var tokens: [Token] = []
+    private var totalDocs: Int = 0
+    private var size: Int = 0
+    
+    init(dimension: Int) {
+        self.dimension = dimension
+    }
+    
+    // Validate token (similar to C implementation)
+    private func isValidToken(_ token: String) -> Bool {
+        let invalidTokens = ["pancake", "author", "radare2"]
+        return !invalidTokens.contains(token)
+    }
+    
+    // Find a token in the token list
+    private func findToken(_ tokens: [Token], _ token: String) -> Token? {
+        return tokens.first { $0.token == token }
+    }
+    
+    // Compute embedding using TF-IDF approach
+    private func computeEmbedding(_ text: String) -> [Float] {
+        // Zero the embedding vector
+        var embedding = [Float](repeating: 0.0, count: dimension)
+        
+        // Tokenize and normalize
+        let words = text
+            .lowercased()
+            .replacingOccurrences(of: "[^a-z0-9\\s]", with: " ", options: .regularExpression)
+            .split(separator: " ")
+            .map { String($0) }
+            .filter { !$0.isEmpty }
+        
+        // Local document token frequency
+        var docTokens: [Token] = []
+        for token in words {
+            if let existingToken = findToken(docTokens, token) {
+                let index = docTokens.firstIndex { $0.token == token }!
+                docTokens[index].count += 1
+            } else {
+                docTokens.append(Token(token: token, count: 1, df: 0))
+            }
+        }
+        
+        // Increment total documents
+        totalDocs += 1
+        
+        // Update global document frequencies
+        for docToken in docTokens {
+            if let globalToken = findToken(self.tokens, docToken.token) {
+                if isValidToken(docToken.token) {
+                    let index = self.tokens.firstIndex { $0.token == docToken.token }!
+                    self.tokens[index].count += 1
+                    self.tokens[index].df += 1
+                }
+            } else {
+                self.tokens.append(Token(token: docToken.token, count: docToken.count, df: 1))
+            }
+        }
+        
+        // Compute TF-IDF and update embedding
+        for docToken in docTokens {
+            // Compute term frequency
+            let tf = 1 + log(Float(docToken.count))
+            
+            // Find global token to get document frequency
+            let dfValue = findToken(self.tokens, docToken.token)?.df ?? 1
+            
+            // Compute inverse document frequency
+            let idf = log(Float(totalDocs + 1) / Float(dfValue + 1)) + 1
+            
+            // Compute weight
+            let weight = tf * idf
+            
+            // Hash-based index assignment
+            let hash = simpleHash(docToken.token)
+            let index = hash % dimension
+            
+            // Add weighted value to embedding
+            embedding[index] += weight
+        }
+        
+        // L2 Normalize the embedding
+        return normalizeVector(embedding)
+    }
+    
+    // Simple string hash function
+    private func simpleHash(_ str: String) -> Int {
+        var hash = 0
+        for char in str.unicodeScalars {
+            hash = ((hash << 5) - hash) + Int(char.value)
+            hash = hash & hash // Convert to 32-bit integer
+        }
+        return abs(hash)
+    }
+    
+    // Normalize vector to unit length
+    private func normalizeVector(_ vec: [Float]) -> [Float] {
+        let normSq = vec.reduce(0) { $0 + $1 * $1 }
+        if normSq > 0 {
+            let norm = sqrt(normSq)
+            return vec.map { $0 / norm }
+        }
+        return vec
+    }
+    
+    // Compute squared Euclidean distance
+    private func squaredDistance(_ a: [Float], _ b: [Float]) -> Float {
+        guard a.count == b.count else { return Float.greatestFiniteMagnitude }
+        return zip(a, b).reduce(0) { $0 + pow($1.0 - $1.1, 2) }
+    }
+    
+    // Recursive KD-Tree insertion
+    private func insertRecursive(_ node: KDNodeData?, point: [Float], text: String, depth: Int) -> KDNodeData {
+        guard let currentNode = node else {
+            return KDNodeData(point: point, text: text, splitDim: depth % dimension)
+        }
+        
+        let axis = currentNode.splitDim
+        if point[axis] < currentNode.point[axis] {
+            currentNode.left = insertRecursive(currentNode.left, point: point, text: text, depth: depth + 1)
+        } else {
+            currentNode.right = insertRecursive(currentNode.right, point: point, text: text, depth: depth + 1)
+        }
+        
+        return currentNode
+    }
+    
+    // Insert a document into the vector database
+    func insert(_ text: String) {
+        guard !text.isEmpty else { return }
+        
+        // Compute embedding
+        let embedding = computeEmbedding(text)
+        
+        // Insert into KD-Tree
+        root = insertRecursive(root, point: embedding, text: text, depth: 0)
+        size += 1
+    }
+    
+    // K-Nearest Neighbors search
+    func query(_ text: String, k: Int) -> [String] {
+        guard !text.isEmpty, root != nil else { return [] }
+        
+        // Compute query embedding
+        let queryEmbedding = computeEmbedding(text)
+        
+        // Initialize result set
+        var results: [KNNResult] = []
+        
+        // Recursive KNN search
+        func searchRecursive(_ node: KDNodeData?, query: [Float], depth: Int) {
+            guard let node = node else { return }
+            
+            // Compute distance and potentially insert into results
+            let distSq = squaredDistance(query, node.point)
+            
+            if results.count < k {
+                results.append(KNNResult(node: node, distanceSq: distSq))
+                // Maintain max-heap property
+                results.sort { $0.distanceSq > $1.distanceSq }
+            } else if distSq < results[0].distanceSq {
+                results[0] = KNNResult(node: node, distanceSq: distSq)
+                results.sort { $0.distanceSq > $1.distanceSq }
+            }
+            
+            // Decide which subtree to search first
+            let axis = node.splitDim
+            let diff = query[axis] - node.point[axis]
+            let first = diff < 0 ? node.left : node.right
+            let second = diff < 0 ? node.right : node.left
+            
+            // Recursively search
+            searchRecursive(first, query: query, depth: depth + 1)
+            
+            // Potentially search the other subtree
+            if results.count < k || diff * diff < results[0].distanceSq {
+                searchRecursive(second, query: query, depth: depth + 1)
+            }
+        }
+        
+        // Perform the search
+        searchRecursive(root, query: queryEmbedding, depth: 0)
+        
+        // Sort results and return texts
+        return results
+            .sorted { $0.distanceSq < $1.distanceSq }
+            .map { $0.node.text }
+    }
+    
+    // Get total number of documents
+    func getSize() -> Int {
+        return size
+    }
+}
+
+// Example usage
+func main() {
+    // Create a vector database with 64-dimensional embeddings
+    let db = VectorDB(dimension: 64)
+    
+    // Insert some sample documents
+    db.insert("The quick brown fox jumps over the lazy dog")
+    db.insert("A quick brown fox is very quick")
+    db.insert("The author of radare2 is name / named pancake")
+    db.insert("The lazy dog sleeps all day")
+    
+    // Query for similar documents
+    let results = db.query("who is the author of radare2?", k: 2)
+    print("Similar documents:", results)
+}
+
+// Run example
+main()

Attic/py/Makefile

@@ -0,0 +1,81 @@
+R2_USER_PLUGINS=$(shell r2 -H R2_USER_PLUGINS)
+
+PWD=$(shell pwd)
+R2PM_BINDIR=$(shell r2pm -H R2PM_BINDIR)
+PV=3.12
+ifeq ($(shell which python${PV} > /dev/null && echo ok),ok)
+PYTHON?=python${PV}
+else
+PYTHON?=python3
+endif
+PIP=$(PYTHON) -m pip
+
+ifeq ($(R2PM_BINDIR),)
+MISSING RADARE2
+endif
+
+LINTED=r2ai/code_block.py
+LINTED+=r2ai/bubble.py
+LINTED+=r2ai/const.py
+LINTED+=r2ai/backend/kobaldcpp.py
+# LINTED+=r2ai/index.py
+# LINTED+=r2ai/voice.py
+# LINTED+=r2ai/anthropic.py
+r2ai-python py python r2aipy: venv
+	@./r2ai.sh
+
+large:
+	. venv/bin/activate ; $(PYTHON) -m r2ai.cli -l
+
+all.old:
+	@test -n "${VIRTUAL_ENV}" || (echo "Run:"; echo ". venv/bin/activate" ; exit 1)
+	$(PYTHON) main.py || $(MAKE) deps
+
+venv:
+	$(PYTHON) -m venv venv
+	if [ -z "`find venv | grep llama_cpp`" ]; then . venv/bin/activate ; pip install . ; fi
+
+deps: venv
+	#test -n "${VIRTUAL_ENV}" || (echo "Run: . venv/bin/activate" ; exit 1)
+	. venv/bin/activate && export CMAKE_ARGS="-DLLAMA_METAL=on -DLLAMA_METAL_EMBED_LIBRARY=ON" && \
+		pip install --force-reinstall -U --no-cache-dir .
+clean:
+	rm -rf venv
+	rm -rf build
+	find . -name "*.egg-info" -exec rm -rf {} +
+
+user-install:
+	rm -f $(R2PM_BINDIR)/r2ai
+	ln -fs $(PWD)/r2ai.sh $(R2PM_BINDIR)/r2ai
+
+install: user-install
+	-mkdir -p /usr/local/share/man/man1/r2ai.1
+	-cp doc/usage/r2ai.1 /usr/local/share/man/man1/r2ai.1
+
+install-plugin user-install-plugin:
+	ln -fs $(PWD)/r2ai/plugin.py $(R2_USER_PLUGINS)/r2ai.py
+
+uninstall user-uninstall:
+	rm -f $(R2PM_BINDIR)/r2ai
+	-rm -f /usr/local/share/man/man1/r2ai.1
+
+user-uninstall-plugin uninstall-plugin:
+	rm -f $(R2_USER_PLUGINS)/r2ai.py
+
+pub:
+	$(PYTHON) -m build
+	twine check dist/*
+	twine upload -u __token__ --repository-url https://upload.pypi.org/legacy/ --verbose dist/*
+
+pylint lint cilint:
+	. venv/bin/activate ; pylint $(LINTED)
+
+lintall:
+	pylint *.py r2ai/*.py
+
+.PHONY: lint lintall
+
+deps-global:
+	export CMAKE_ARGS="-DLLAMA_METAL=on -DLLAMA_METAL_EMBED_LIBRARY=ON" && \
+		$(PIP) install --force-reinstall -U --break-system-packages --no-cache-dir .
+

Attic/py/README.md

@@ -0,0 +1,155 @@
+# r2ai-python
+
+This directory contains the original implementation of the AI support for radare2 written in Python.
+
+It is considered _deprecated_, but it's still useful and aims to be maintained, we recommend using decai or the C rewrite as alternatives.
+
+The Python implementation consists in a module can be used in 3 different ways:
+
+* commandline repl using r2pipe to spawn or connect to remote radare2 instances
+* radare2 core plugin that is instantiated when placing it in the radare2 plugin's directory
+* Python API to be used for writing your own scripts
+
+## API Providers
+
+The Python implementation focused on supporting the LlamaCpp Python module, which is somehow heavy to load in many environments, but it also supports litellm which provides access to many external connectors like ollama, openai, anthropic, gemini, etc.
+
+## Deprecation Reasons
+
+There are several reasons why this implementation is considered deprecated and not recommended to use:
+
+1) It's Python.
+
+ - Aka, it's slow, heavy loads in same process
+ - Your system probably have multiple versions installed
+ - The shell and radare2 python bindings are probably not the same
+ - Required venv to be created (at least an extra 1GB)
+ - Not typed, regressions and bugs appear at runtime
+
+2) Goto 1
+
+### Windows
+
+On Windows you may follow the same instructions, just ensure you have the right python environment ready and create the venv to use
+
+```cmd
+git clone https://github.com/radareorg/r2ai
+cd r2ai
+set PATH=C:\Users\YOURUSERNAME\Local\Programs\Python\Python39\;%PATH%
+python3 -m pip install .
+python3 main.py
+```
+
+### Selecting the model
+
+- List all downloaded models: `-m`
+- Get a short list of models: `-MM`
+- Help: `-h`
+
+**Example selecting a remote models:**
+
+```
+[r2ai:0x00006aa0]> -m anthropic:claude-sonnet-5-5
+[r2ai:0x00006aa0]> -m openai:gpt-4
+```
+
+**Example downloading a free local AI: Mistral 7B v0.2:**
+
+Launch r2ai, select the model and ask a question. If the model isn't downloaded yet, r2ai will ask you which precise version to download.
+
+```
+[r2ai:0x00006aa0]> -m TheBloke/Mistral-7B-Instruct-v0.2-GGUF
+```
+
+Then ask your question, and r2ai will automatically download if needed:
+
+```
+[r2ai:0x00006aa0]> give me a short algorithm to test prime numbers
+Select TheBloke/Mistral-7B-Instruct-v0.2-GGUF model. See -M and -m flags
+[?] Quality (smaller is faster): 
+ > Small | Size: 2.9 GB, Estimated RAM usage: 5.4 GB
+   Medium | Size: 3.9 GB, Estimated RAM usage: 6.4 GB
+   Large | Size: 7.2 GB, Estimated RAM usage: 9.7 GB
+   See More
+
+[?] Quality (smaller is faster): 
+ > mistral-7b-instruct-v0.2.Q2_K.gguf | Size: 2.9 GB, Estimated RAM usage: 5.4 GB
+   mistral-7b-instruct-v0.2.Q3_K_L.gguf | Size: 3.6 GB, Estimated RAM usage: 6.1 GB
+   mistral-7b-instruct-v0.2.Q3_K_M.gguf | Size: 3.3 GB, Estimated RAM usage: 5.8 GB
+   mistral-7b-instruct-v0.2.Q3_K_S.gguf | Size: 2.9 GB, Estimated RAM usage: 5.4 GB
+   mistral-7b-instruct-v0.2.Q4_0.gguf | Size: 3.8 GB, Estimated RAM usage: 6.3 GB
+   mistral-7b-instruct-v0.2.Q4_K_M.gguf | Size: 4.1 GB, Estimated RAM usage: 6.6 GB
+   mistral-7b-instruct-v0.2.Q4_K_S.gguf | Size: 3.9 GB, Estimated RAM usage: 6.4 GB
+   mistral-7b-instruct-v0.2.Q5_0.gguf | Size: 4.7 GB, Estimated RAM usage: 7.2 GB
+   mistral-7b-instruct-v0.2.Q5_K_M.gguf | Size: 4.8 GB, Estimated RAM usage: 7.3 GB
+   mistral-7b-instruct-v0.2.Q5_K_S.gguf | Size: 4.7 GB, Estimated RAM usage: 7.2 GB
+   mistral-7b-instruct-v0.2.Q6_K.gguf | Size: 5.5 GB, Estimated RAM usage: 8.0 GB
+   mistral-7b-instruct-v0.2.Q8_0.gguf | Size: 7.2 GB, Estimated RAM usage: 9.7 GB
+
+[?] Use this model by default? ~/.r2ai.model: 
+ > Yes
+   No
+
+[?] Download to ~/.local/share/r2ai/models? (Y/n): Y
+```
+
+**Example selecting a local model served by Ollama**
+
+Download a model and make it available through Ollama:
+
+```
+$ ollama ls
+NAME                  ID              SIZE      MODIFIED     
+codegeex4:latest      867b8e81d038    5.5 GB    23 hours ago  
+```
+
+Use it from r2ai by prefixing its name with `ollama/`
+
+```
+[r2ai:0x00002d30]> -m ollama/codegeex4:latest
+[r2ai:0x00002d30]> hi
+Hello! How can I assist you today?
+```
+
+### Standard/Auto mode
+
+The standard mode is invoked by directly asking the question.
+For the Auto mode, the question **must be prefixed** by `' ` (quote + space). The AI may instruct r2ai to run various commands. Those commands are run on *your host*, so you will be asked to review them before they run.
+
+Example in "standard" mode:
+
+```
+[r2ai:0x00006aa0]> compute 4+5
+4 + 5 = 9
+[r2ai:0x00006aa0]> draw me a pancake in ASCII art
+Sure, here's a simple ASCII pancake:
+
+  _____  
+ (     )
+ (     )
+  -----
+```
+
+Example in auto mode:
+
+```
+[r2ai:0x00006aa0]>' Decompile the main
+[..]
+r2ai is going to execute the following command on the host
+Want to edit? (ENTER to validate) pdf @ fcn.000015d0
+This command will execute on this host: pdf @ fcn.000015d0. Agree? (y/N) y
+```
+
+If you wish to edit the command, you can do it inline for short one line commands, or an editor will pop up.
+
+### r2ai Configuration settings
+
+List all settings with `-e`
+
+| Key         | Explanation                           |
+| ----------- | ------------------------------------- |
+| debug_level | All verbose messages for level 1. Default is 2 |
+| auto.max_runs | Maximum number of questions the AI is allowed to ask r2 in auto mode. |
+| auto.hide_tool_output | By default false, consequently output of r2cmd, run_python etc is shown. Set to `true` to hide those internal messages. |
+| chat.show_cost | Show the cost of each request to the AI if true |
+

Attic/py/pyproject.toml

@@ -1,7 +1,7 @@
 [project]
 name = "r2ai"
 dynamic = ["readme"]
-version = "0.8.0"
+version = "0.9.6"
 description = "Artificial intelligence tooling for radare2"
 
 license = {text = "MIT License"}
@@ -9,44 +9,56 @@ authors = [
   {name = "pancake", email = "[email protected]"}
 ]
 
-[project.urls]
-homepage = "https://www.radare.org/"
-repository = "https://github.com/radareorg/r2ai"
-
-[project.scripts]
-r2ai = "r2ai:main"
-
-[tool.setuptools.packages.find]
-where = [".", "r2ai"]
-include = ["main", "r2ai"]
-namespaces = true
-
-[tool.setuptools.dynamic]
-readme = {file = "README.md", content-type = "text/markdown"}
-
-[build-system]
-requires = ["setuptools"]
-build-backend = "setuptools.build_meta"
-
 dependencies = [
   "rich",
   "r2pipe",
   "inquirer",
-  "llama-cpp-python",
-  "huggingface_hub==0.22.0",
+  "llama-cpp-python==0.3.7",
+  "huggingface_hub",
   "appdirs",
   "unidecode",
   "jsonref",
   "transformers",
   "pydantic",
-  "torch"
+  "pyreadline3",
+  "tokentrim",
+  "boto3",
+  "colorama",
+  "textual",
+  "litellm>=1.60.5",
+  "numpydoc"
 ]
 
 [project.optional-dependencies]
 extras = [
+  "chromadb",
   "openai",
   "anthropic",
   "groq",
   "google-generativeai",
   "google-cloud-aiplatform"
 ]
+
+
+[project.urls]
+homepage = "https://www.radare.org/"
+repository = "https://github.com/radareorg/r2ai"
+
+[project.scripts]
+r2ai = "r2ai.main:run" 
+
+[tool.setuptools]
+include-package-data = true
+
+[tool.setuptools.packages.find]
+where = ["."]
+include = ["r2ai*"]
+namespaces = true
+
+[tool.setuptools.dynamic]
+readme = {file = "README.md", content-type = "text/markdown"}
+
+[build-system]
+requires = ["setuptools", "wheel"]
+build-backend = "setuptools.build_meta"
+

Attic/py/r2ai.sh

@@ -0,0 +1,22 @@
+#!/bin/sh
+unset DYLD_LIBRARY_PATH
+unset LD_LIBRARY_PATH
+export TRANSFORMERS_NO_ADVISORY_WARNINGS=1
+PYTHON=python3
+if [ -h "$0" ]; then
+	F=`readlink $0`
+else
+	F="$0"
+fi
+D=`dirname "$F"`
+RD=`realpath "$D"`
+[ -n "${RD}" ] && D="$RD"
+[ -n "$D" ] && cd "$D"
+if [ -d venv ]; then
+	. venv/bin/activate
+else
+	$PYTHON -m venv venv
+	. venv/bin/activate
+	pip3 install -e .
+fi
+exec $PYTHON -m r2ai.cli "$@"

Attic/py/r2ai/__init__.py

@@ -0,0 +1,24 @@
+import os
+import logging
+from rich.logging import RichHandler
+
+
+VERSION = "0.8.2"
+
+# 0 NOTSET, 1 DEBUG, 2 INFO, 3 WARNING, 4 ERROR, 5 CRITICAL; multipied by 10
+LOG_LEVEL = int(os.environ.get('R2AI_LOG', '2')) * 10
+LOG_FILE = os.environ.get('R2AI_LOGFILE', None)
+
+for tag in ["httpx", "openai", "httpcore"]:
+    _logger = logging.getLogger(tag)
+    _logger.setLevel(logging.CRITICAL)
+    _logger.propagate = False  # Disable child loggers too
+
+handlers = [RichHandler()]
+if LOG_FILE:
+    handlers.append(logging.FileHandler(LOG_FILE))
+
+LOGGER = logging.getLogger(__name__)
+logging.basicConfig(format="%(name)s - %(levelname)s - %(message)s",
+                    handlers=handlers)
+LOGGER.setLevel(LOG_LEVEL)

Attic/py/r2ai/auto.py

@@ -0,0 +1,511 @@
+import builtins
+import json
+import sys
+import re
+from . import LOGGER
+import litellm
+from litellm import _should_retry, acompletion, utils, ModelResponse
+import asyncio
+from .pipe import get_filename
+from .tools import r2cmd, run_python, execute_binary, schemas, print_tool_call
+import json
+import signal
+from .spinner import spinner
+from .completion import create_chat_completion
+import uuid
+import time
+import readline
+import subprocess
+import shlex
+import os
+import tempfile
+
+# litellm.drop_params = True
+# litellm.set_verbose=True
+
+ANSI_REGEX = re.compile(r'\x1B(?:[@-Z\\-_]|\[[0-?]*[ -/]*[@-~])')
+
+SYSTEM_PROMPT_AUTO = """
+You are a reverse engineer and you are using radare2 to analyze a binary.
+The user will ask questions about the binary and you will respond with the answer to the best of your ability.
+
+# Guidelines
+- Understand the Task: Grasp the main objective, goals, requirements, constraints, and expected output.
+- Reasoning Before Conclusions**: Encourage reasoning steps before any conclusions are reached.
+- Assume the user is always asking you about the binary, unless they're specifically asking you for radare2 help.
+- The binary has already been loaded. You can interact with the binary using the r2cmd tool.
+- `this` or `here` might refer to the current address in the binary or the binary itself.
+- If you need more information, try to use the r2cmd tool to run commands before answering.
+- 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.
+- 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.
+- Don't just regurgitate the same code, figure out what it's doing and rewrite it to be more understandable.
+- If you need to run a command in r2 before answering, you can use the r2cmd tool
+- Do not repeat commands if you already know the answer.
+- Formulate a plan. Think step by step. Analyze the binary as much as possible before answering.
+- You must keep going until you have a final answer.
+- Double check that final answer. Make sure you didn't miss anything.
+- Make sure you call tools and functions correctly.
+"""
+
+SYSTEM_PROMPT_AUTO_REASONING = f"""
+You are an expert reverse engineer.
+The user will ask questions about the binary and you will respond with the answer to the best of your ability.
+You can use the available tools to help you answer the user's question.
+The binary has already been loaded. You can interact with the binary using the r2cmd tool.
+If you need to run a command in r2 before answering, you can use the r2cmd tool
+"""
+
+class ChatAuto:
+    def __init__(self, model, max_tokens = 32000, top_p=0.95, temperature=0.0, interpreter=None, system=None, tools=None, ask_to_execute=True, messages=None, tool_choice='auto', llama_instance=None, timeout=60, stream=True, cb=None, max_runs=100):
+        self.logger = LOGGER
+        self.functions = {}
+        self.tools = []
+        self.model = model
+        self.max_tokens = max_tokens
+        self.top_p = top_p
+        self.temperature = temperature
+        self.system = system
+        self.messages = messages
+        self.interpreter = interpreter
+        self.system_message = None
+        self.timeout = timeout
+        self.stream = stream
+        self.cost = 0
+        self.is_reasoning = False
+        self.n_runs = 0
+        self.max_runs = max_runs
+        self._last_response = None
+        self._start_time = time.time()
+        self._last_run_time = time.time()
+        self.ask_to_execute = ask_to_execute
+        self.llama_instance = llama_instance or interpreter.llama_instance if interpreter else None
+        
+        init_prompt = ""
+        
+        if not self.llama_instance:
+            rmodel = re.sub(r'ft:([^:]+).*', r'\1', self.model)
+            model_info = litellm.get_model_info(rmodel)
+            self.max_tokens = min(self.max_tokens, model_info['max_tokens'])
+
+            if self.model.startswith('openai/o'):
+                self.is_reasoning = True
+                self.max_tokens = model_info['max_tokens']
+                self.timeout = 60 * 60
+            init_commands = self.interpreter.env["auto.init_commands"]
+            
+            if init_commands:
+                init_prompt = f"""
+Here is some information about the binary to get you started:
+> {init_commands}
+{r2cmd(init_commands)}
+"""
+        if self.is_reasoning:
+            if not system and not self.messages[0]['role'] == 'developer':
+                self.messages.insert(0, { "role": "developer", "content": SYSTEM_PROMPT_AUTO_REASONING + init_prompt })
+            self.stream = False
+        else:
+            if messages and messages[0]['role'] != 'system':
+                self.messages.insert(0, { "role": "system", "content": system or SYSTEM_PROMPT_AUTO + init_prompt })
+
+
+        if cb:
+            self.cb = cb
+        else:
+            self.cb = lambda *args: None
+        self.tool_choice = None
+        if tools:
+            for tool in tools:
+                if tool.__name__ in schemas:
+                    schema = schemas[tool.__name__]
+                else:
+                    schema = utils.function_to_dict(tool)
+                
+                self.tools.append({ "type": "function", "function": schema })
+                self.functions[tool.__name__] = tool
+            self.tool_choice = tool_choice
+        #self.tool_end_message = '\nNOTE: The user saw this output, do not repeat it.'
+
+    async def process_tool_calls(self, tool_calls):
+        if tool_calls:
+            for tool_call in tool_calls:
+                self.logger.debug(f"tool_call: {tool_call}")
+                tool_name = tool_call["function"]["name"]
+                if "id" not in tool_call:
+                    tool_call["id"] = str(uuid.uuid4())
+                try:
+                    tool_args = json.loads(tool_call["function"]["arguments"])
+                except Exception:
+                    if "arguments" in tool_call["function"] and type(tool_call["function"]["arguments"]) == dict:
+                        tool_args = tool_call["function"]["arguments"]
+                    else:
+                        self.logger.error(f'Error parsing JSON: {tool_call["function"]["arguments"]}')
+                        # raise Exception('Error parsing JSON')
+                        self.messages.append({"role": "tool", "name": tool_name, "content": "Error: Unable to parse JSON" , "tool_call_id": tool_call["id"]})
+                        continue
+                if tool_name not in self.functions:
+                    self.logger.error(f'Tool not found: {tool_name}')
+                    self.messages.append({"role": "tool", "name": tool_name, "content": "Error: Tool not found" , "tool_call_id": tool_call["id"]})
+                    continue
+
+                if self.ask_to_execute:
+                    # tool_call typically contains:
+                    # { 'id': 'toolu_01Hr6XTatRtp7rX54iBu1voC', 'type': 'function', 'function': {'name': 'r2cmd', 'arguments': '{"command": "pdf @ main"}'}}
+                    answer = ''
+                    will_execute = True # a command (modified or not) will be executed
+                    while answer.lower() != 'y':
+                        try:
+                            command = tool_args.get('command', '')
+
+                            if len(command.splitlines()) > 1:
+                                print(f'r2ai is going to execute the following command on the host:')
+                                print(f'{command}')
+                                want_edit = input('Want to edit? (Y/n) ')
+                                if want_edit.lower() != 'n':
+                                    # open editor when several lines
+                                    with tempfile.NamedTemporaryFile(mode='w+', delete=False) as temp_file:
+                                        temp_file.write(command) 
+                                        temp_filename = temp_file.name 
+
+                                    # Fallback to 'nano' if EDITOR is not set (I still prefer emacs)
+                                    editor = os.environ.get('EDITOR', 'nano')  
+                                    subprocess.run(shlex.split(editor + ' ' + temp_filename))
+                                    with open(temp_filename, "r") as f:
+                                        new_command = f.read().strip()
+                                    os.remove(temp_filename)
+                                else:
+                                    self.logger.debug(f'User agrees with not editing the command: {command}')
+                                    answer = 'y'
+
+                            else:
+                                # inline short edit
+                                print(f'r2ai is going to execute the following command on the host:')
+                                print(f'> {command}')
+                                
+                                new_command = input(f"Type a new command or press ENTER to use {command} ") or command
+
+                            if answer.lower() != 'y':
+                                answer = input(f"\033[91mThis command will execute on this host: {new_command}. Agree? (y/N)\033[0m ")
+                                if answer.lower() == 'y':
+                                    # we need to refresh tool_call and tool_args with edited values
+                                    tool_args['command'] = new_command
+                                    tool_call['function']['arguments'] = json.dumps(tool_args)
+
+                        except (json.JSONDecodeError, TypeError, AttributeError) as e:
+                            # in case tool_call is not formatted as expected, like missing command, arguments, function...
+                            self.logger.error(f'Unexpected format for {tool_call}')
+                            self.messages.append({"role": "tool", "name": tool_name, "content": "User refused to execute the command" , "tool_call_id": tool_call["id"]})
+                            # error case: we'll skip the tool
+                            will_execute = False
+                            break
+
+                    if not will_execute:
+                        self.logger.debug(f'Skipping tool={tool_call}')
+                        continue
+
+                self.logger.debug(f'Executing tool id={tool_call["id"]}, name={tool_name}, arguments={tool_args}')
+                self.cb('tool_call', { "id": tool_call["id"], "function": { "name": tool_name, "arguments": tool_args } })
+                if asyncio.iscoroutinefunction(self.functions[tool_name]):
+                    tool_response = await self.functions[tool_name](**tool_args)
+                else:
+                    tool_response = self.functions[tool_name](**tool_args)
+                if self.interpreter.env["auto.hide_tool_output"] == "false":
+                    self.cb('tool_response', { "id": tool_call["id"] + '_response', "content": tool_response })
+                self.messages.append({"role": "tool", "name": tool_name, "content": ANSI_REGEX.sub('', tool_response), "tool_call_id": tool_call["id"]})
+
+        return await self.get_completion()
+
+    async def process_streaming_response(self, resp):
+        tool_calls = []
+        msgs = []
+        parts = []
+        current_message = { "role": "assistant", "content": "", "tool_calls": [] }
+        first_chunk = True
+        async for chunk in resp:
+            delta = None
+            choice = chunk.choices[0]
+            delta = choice.delta
+            if delta.tool_calls:
+                delta_tool_calls = delta.tool_calls[0]
+                index = delta_tool_calls.index
+                fn_delta = delta_tool_calls.function
+                tool_call_id = delta_tool_calls.id
+                if len(current_message['tool_calls']) < index + 1:
+                    tool_call = {
+                            "id": tool_call_id or str(uuid.uuid4()),
+                            "type": "function",
+                            "function": {
+                                "name":fn_delta.name,
+                                "arguments": fn_delta.arguments
+                            }
+                        }
+                    current_message['tool_calls'].append(tool_call)
+                else:
+                    if fn_delta.name:
+                        current_message['tool_calls'][index]["function"]["name"] = fn_delta.name
+                    current_message['tool_calls'][index]["function"]["arguments"] += fn_delta.arguments
+            else:
+                m = None
+                done = False
+                if delta.content is not None:
+                    if first_chunk:
+                        self.cb('message', { 'content': "\n\x1b[1;32massistant\x1b[0m\n", 'done': True })
+                        first_chunk = False
+                    m = delta.content
+                    if m is not None:
+                        current_message['content'] += m
+                    self.cb('message', { "content": m, "id": 'message_' + chunk.id, 'done': False })
+                if 'finish_reason' in choice and choice['finish_reason'] == 'stop':
+                    done = True
+                    self.cb('message', { "content": "", "id": 'message_' + chunk.id, 'done': True })
+                self.cb('message_stream', { "content": m if m else '', "id": 'message_' + chunk.id, 'done': done })
+        self.messages.append(current_message)
+        if len(current_message['tool_calls']) == 0:
+            del current_message['tool_calls']
+            self.show_cost()
+        else:
+            self.show_cost()
+            await self.process_tool_calls(current_message['tool_calls'])
+                    
+        return current_message
+
+    async def process_response(self, resp):
+        content = resp.choices[0].message.content
+        
+        current_message = { 'role': 'assistant', 'content': content or '', 'tool_calls': [] }
+        for i, tool_call in enumerate(resp.choices[0].message.tool_calls or []):
+            current_message['tool_calls'].append({
+                "id": tool_call.id,
+                "type": "function",
+                "index": i,
+                "function": {
+                    "name": tool_call.function.name,
+                    "arguments": tool_call.function.arguments
+                },
+            })
+        if len(current_message['tool_calls']) == 0:
+            try:
+                tool_call = json.loads(content)
+                if 'name' in tool_call and tool_call['name'] in self.functions:
+                    args = None
+                    if 'arguments' in tool_call:
+                        args = tool_call['arguments']
+                    elif 'parameters' in tool_call:
+                        args = tool_call['parameters']
+                    if args:
+                        current_message['tool_calls'].append({ "id": resp.id, "function": { "name": tool_call['name'], "arguments": json.dumps(args) } })
+            except Exception:
+                pass
+        self.messages.append(current_message)
+        if current_message['content'] != "":
+            self.cb('message', { 'content': "\n\x1b[1;32massistant\x1b[0m\n", 'done': True })
+            self.cb('message', { "content": current_message['content'], "id": 'message_' + resp.id, 'done': True })
+
+        if len(current_message['tool_calls']) == 0:
+            del current_message['tool_calls']
+            self.show_cost()
+        else:
+            self.show_cost()
+            await self.process_tool_calls(current_message['tool_calls'])
+
+        return current_message
+        
+    async def async_response_generator(self, response):
+        for item in response:
+            resp = ModelResponse(stream=True, **item)
+            yield resp
+
+    async def attempt_completion(self):
+        stream = self.stream
+        if self.llama_instance:        
+            args = {
+                "temperature": self.temperature,
+                "top_p": self.top_p,
+                "max_tokens": self.max_tokens,
+                "stream": stream,
+            }
+            if self.interpreter.env["chat.rawdog"] == "true":
+                res = create_chat_completion(self.interpreter, messages=self.messages, tools=self.tools, **args)
+            else:
+                res = self.llama_instance.create_chat_completion(self.messages, tools=self.tools, **args)
+            if args['stream']:
+                return self.async_response_generator(res)
+            else:
+                return ModelResponse(**next(res))
+        additional_args = {}
+        if self.is_reasoning:
+            additional_args['reasoning_effort'] = self.interpreter.env["chat.reasoning_effort"]
+        else:
+            additional_args['temperature'] = self.temperature
+        if self.interpreter.env["chat.openai_store"] == "true":
+            additional_args['store'] = True
+        if not self.check_max():
+            additional_args['tools'] = self.tools
+            additional_args['tool_choice'] = self.tool_choice
+        compl = await acompletion(
+            model=self.model,
+            messages=self.messages,
+            timeout=self.timeout,
+            max_completion_tokens=self.max_tokens,
+            **additional_args,
+            stream=stream,
+        )
+        return compl
+
+    def check_max(self):
+        if self.max_runs > 0 and self.n_runs >= self.max_runs - 1:
+            return True
+        return False
+    
+    def show_cost(self):
+        def format_time(seconds):
+            minutes = int(seconds // 60)
+            remaining_seconds = int(seconds % 60)
+            if minutes > 0:
+                return f"{minutes}m {remaining_seconds}s"
+            return f"{remaining_seconds}s"
+        self.n_runs += 1
+        run_time = format_time(time.time() - self._last_run_time)
+        self._last_run_time = time.time()
+        total_time = format_time(time.time() - self._start_time)
+        
+        if self.interpreter.env["chat.show_cost"] == "true" and self.n_runs > 0:
+            run_cost = 0
+            if not self.llama_instance:
+                run_cost = litellm.completion_cost(completion_response=self._last_response, messages=self.messages, model=self.model)
+                self.cost += run_cost
+            cb('usage', { 
+                "model": self.model, 
+                "run_cost": run_cost, 
+                "total_cost": self.cost, 
+                "n_runs": self.n_runs, 
+                "max_runs": self.max_runs,
+                "run_time": run_time,
+                "total_time": total_time
+            })
+
+    async def get_completion(self):
+        stream = self.stream
+        if self.llama_instance:
+            response = await self.attempt_completion()
+            if stream:
+                return await self.process_streaming_response(response)
+            else:
+                return await self.process_response(response)
+        max_retries = 5
+        base_delay = 2
+        
+        for retry_count in range(max_retries):
+            try:
+                response = await self.attempt_completion()
+                self._last_response = response
+                self.logger.debug(f'chat completion {response}')
+                res = None
+                if stream:
+                    res = await self.process_streaming_response(response)
+                else:
+                    res = await self.process_response(response)
+
+                return res
+            except Exception as e:
+                self.logger.error(f'Error getting completion: {e}')
+                if not _should_retry(getattr(e, 'status_code', None)) or retry_count == max_retries - 1:
+                    raise
+                
+                delay = base_delay * (2 ** retry_count)
+                self.logger.info(f"Retrying in {delay} seconds...")
+                await asyncio.sleep(delay)
+        
+        raise Exception("Max retries reached. Unable to get completion.")
+
+    async def achat(self, messages=None) -> str:
+        if messages:
+            self.messages = messages
+        self.logger.debug(self.messages)
+        response = await self.get_completion()
+        return response
+    
+    def chat(self, **kwargs) -> str:
+        return asyncio.run(self.achat(**kwargs))
+
+def cb(type, data):
+    spinner.stop()
+    if type == 'message_stream':
+        if 'content' in data:
+            sys.stdout.write(data['content'])
+    elif type == 'tool_call':
+        builtins.print()
+        print_tool_call(data)
+    elif type == 'tool_response':
+        if 'content' in data:
+            sys.stdout.write(data['content'])
+            sys.stdout.flush()
+            builtins.print()
+    elif type == 'usage':
+        sys.stdout.flush()
+        builtins.print()
+        sys.stdout.write(f'\x1b[1;34m{data["model"]} | total: ${float(data["total_cost"]):.10f} | run: ${float(data["run_cost"]):.10f} | {data["n_runs"]} / {data["max_runs"]} | {data["run_time"]} / {data["total_time"]}\x1b[0m')
+        sys.stdout.flush()
+        builtins.print()
+    elif type == 'message' and data['done']:
+        if data['content']:
+            sys.stdout.flush()
+            sys.stdout.write(data['content'])
+            sys.stdout.flush()
+        else:
+            builtins.print()
+
+def signal_handler(signum, frame):
+    raise KeyboardInterrupt
+
+def chat(interpreter, **kwargs):
+    model = interpreter.model.replace(":", "/")
+    tools = [r2cmd, run_python, execute_binary]
+    messages = interpreter.messages
+    tool_choice = 'auto'
+
+    try:
+        loop = asyncio.get_event_loop()
+        if loop.is_closed() or loop.is_running():
+            loop = asyncio.new_event_loop()
+            asyncio.set_event_loop(loop)
+    except RuntimeError:
+        loop = asyncio.new_event_loop()
+        asyncio.set_event_loop(loop)
+
+    chat_auto = ChatAuto(
+        model, 
+        max_tokens=int(interpreter.env["llm.maxtokens"]), 
+        top_p=float(interpreter.env["llm.top_p"]), 
+        temperature=float(interpreter.env["llm.temperature"]), 
+        interpreter=interpreter, 
+        tools=tools, 
+        ask_to_execute=interpreter.env["auto.ask_to_execute"] == "true",
+        messages=messages, 
+        tool_choice=tool_choice, 
+        llama_instance=interpreter.llama_instance, 
+        stream=interpreter.env["chat.stream"] == "true",
+        max_runs=int(interpreter.env["auto.max_runs"]),
+        cb=cb
+    )
+    
+    original_handler = signal.getsignal(signal.SIGINT)
+
+    try:
+        signal.signal(signal.SIGINT, signal_handler)
+        spinner.start()
+        return loop.run_until_complete(chat_auto.achat())
+    except KeyboardInterrupt:
+        builtins.print("\033[91m\nOperation cancelled by user.\033[0m")
+        tasks = asyncio.all_tasks(loop=loop)
+        for task in tasks:
+            task.cancel()
+        try:
+            loop.run_until_complete(asyncio.gather(*tasks, return_exceptions=True))
+            loop.run_until_complete(asyncio.sleep(0.1))
+        except asyncio.CancelledError:
+            pass
+        return None
+    finally:
+        signal.signal(signal.SIGINT, original_handler)
+        spinner.stop()
+        litellm.in_memory_llm_clients_cache.flush_cache()

Attic/py/r2ai/backend/bedrock.py

@@ -0,0 +1,106 @@
+from colorama import Fore, Back, Style
+
+from ..pipe import get_r2_inst
+
+BEDROCK_TOOLS_CONFIG = {
+    "tools": [
+        {
+            "toolSpec": {
+                "name": "r2cmd",
+                "description": "runs commands in radare2. You can run it multiple times or chain commands with pipes/semicolons. You can also use r2 interpreters to run scripts using the `#`, '#!', etc. commands. The output could be long, so try to use filters if possible or limit. This is your preferred tool",
+                "inputSchema": {
+                    "json": {
+                        "type": "object",
+                        "properties": {
+                            "command": {
+                                "type": "string",
+                                "description": "command to run in radare2."
+                            }
+                        },
+                        "required": [
+                            "command"
+                        ]
+                    }
+                }
+            }
+        }
+    ]
+}
+
+def build_messages_for_bedrock(messages):
+    # Bedrock needs that conversation messages alternate between user and assistant
+    # if the user wants to send multiple messages they should all be consolidated
+    # in a single entry
+
+    bedrock_msgs = []
+    for msg in messages:
+        role = msg.get("role")
+        if msg.get("role") not in ["user", "assistant"]:
+            continue
+
+        if len(bedrock_msgs) > 0 and bedrock_msgs[-1]["role"] == role:
+            last_msg = bedrock_msgs[-1]
+            # This message should be consolidated with the previous one
+            if isinstance(msg["content"], list):
+                last_msg["content"].extend(msg["content"])
+            else:
+                last_msg["content"].append({
+                    "text": msg["content"]
+                })
+
+        else:
+            # The role changed, so create a new entry
+            if isinstance(msg["content"], list) and "role" in msg:
+                # This clause is for messages that are returned from bedrock
+                # and thus are already well formatted
+                bedrock_msgs.append(msg)
+            else:
+                bedrock_msgs.append({
+                    "role": role,
+                    "content": [{"text": msg["content"]}]
+                })
+
+    return bedrock_msgs
+
+def extract_bedrock_tool_calls(response):
+    tool_calls = []
+    content = response.get("output", {}).get("message", {}).get("content", [])
+    for msg in content:
+        if not "toolUse" in msg:
+            continue
+
+        tool_calls.append(msg["toolUse"])
+
+    return tool_calls
+
+def process_bedrock_tool_calls(calls):
+    r2 = get_r2_inst()
+    messages = []
+    if not r2:
+        print("Invalid r2 instance. Can not execute commands. Did you open a file?")
+        return messages
+
+    for call in calls:
+        if call["name"] == "r2cmd":
+            cmd = call["input"]["command"]
+            print(f"\n{Fore.GREEN}Executing r2 cmd: {cmd}{Style.RESET_ALL}")
+            res = r2.cmd(cmd)
+            # print(f"{res}")
+            messages.append({
+                "role": "user",
+                "content": [{
+                    "toolResult": {
+                        "toolUseId": call.get("toolUseId"),
+                        "content": [{ "text": res }]
+                    }
+                }]
+            })
+
+    return messages
+
+def print_bedrock_response(response, print=print, output_limit=200):
+    msg = response.get("output", {}).get("message", {})
+    
+    for m in msg.get("content", []):
+        if "text" in m:
+            print(f"\n{Fore.YELLOW}[AI]> {Style.RESET_ALL}{m['text']}")

Attic/py/r2ai/backend/openapi.py

@@ -1,28 +1,38 @@
 """Implementation for kobaldcpp http api call using openai endpoint."""
 import json
 import requests
+from .. import LOGGER
 
 # MODEL="~/.r2ai.models/Lexi-Llama-3-8B-Uncensored_Q4_K_M.gguf"
 # ./llama-server --in-prefix '### User: ' --prompt 5001 \
 #	--in-suffix "### Assistant: " -m $MODEL
 
-def chat(messages, uri='http://localhost:5001'):
+def chat(messages, uri='http://localhost:5001', model='gpt-3.5-turbo', openapiKey=''):
     """Send a message to a kobaldcpp server and return the autocompletion response
     """
     if uri.endswith("/"):
         uri = uri[0:len(uri)-1]
 #    url = f'{uri}/v1/completions'
-    url = f'{uri}/v1/chat/completions'
+    url = f'{uri}/chat/completions'
     data = {
-      "model": "gpt-3.5-turbo",
+      "model": model,
       "messages": messages
     }
-    r = requests.post(url=url, data=json.dumps(data), timeout=600)
-# print(r.text)
+    headers = {
+        "HTTP-Referer": "https://rada.re", # openrouter specific: Optional, for including your app on openrouter.ai rankings.
+        "X-Title": "radare2", # openrouter specific: Optional. Shows in rankings on openrouter.ai.
+        "Authorization": f"Bearer {openapiKey}",
+        "Content-Type": "application/json"
+    }
+
+    r = requests.post(url=url, data=json.dumps(data), timeout=600, headers=headers)
     j = json.loads(r.text)
     if "choices" in j:
         choice = j["choices"][0]
         if "text" in choice:
             return j["choices"][0]["text"]
         return choice["message"]["content"]
+    if "error" in j:
+        error = j["error"]
+        LOGGER.getChild("openapi").error("OpenAIError[%s], %s", error['code'], error['message'])
     return "No response"

Attic/py/r2ai/bubble.py

@@ -14,6 +14,15 @@ CLIPPY_BEGIN = """
 
 CLIPPY_END = """`-----------------------------------------------------------------------"""
 
+PICO_BEGIN = r"""
+                                           __   __          
+                                          (o |_| o)_______    
+                                          |  _____________)  
+                                          \              / . 
+                                           \            /  |\
+                                            \__________/   | \
+.----------------------------------------------------------'  '---------"""
+
 USER_BEGIN = """   |\\
    | \\
 .--'  '--"""
@@ -48,6 +57,11 @@ def response_begin():
     """
     print(CLIPPY_BEGIN + getpad() + ".")
 
+def pico_begin():
+    """Print the beginning of the Pico le Croco response
+    """
+    print(PICO_BEGIN + getpad() + ".")
+
 def response_end():
     """Print the end of the clippy response
     """

Attic/py/r2ai/cli.py

@@ -0,0 +1,3 @@
+from .main import run # pylint: disable=wrong-import-position
+
+run()

Attic/py/r2ai/env.py

@@ -0,0 +1,13 @@
+from typing import Iterable, Callable
+
+class R2AiEnv(dict):
+    def __init__(self):
+        self._callbacks = {}
+        
+    def add_callback(self, key: str, callback: Callable):
+        self._callbacks[key] = callback
+        
+    def __setitem__(self, __key, __value) -> None:
+        if __key in self._callbacks:
+            self._callbacks[__key](__value)
+        return super().__setitem__(__key, __value)

Attic/py/r2ai/index.py

@@ -5,13 +5,18 @@ import builtins
 import requests
 import json
 import traceback
-from utils import syscmdstr
+have_chrome = True
+try:
+    import chromadb
+except Exception:
+    have_chrome = False
 from unidecode import unidecode
 import sys
+from r2ai import LOGGER
 try:
-    from .utils import slurp
+    from .utils import slurp, syscmdstr
     from .const import R2AI_HISTFILE
-except:
+except Exception:
     from utils import slurp
     R2AI_HISTFILE = "/dev/null"
 
@@ -24,7 +29,7 @@ MASTODON_KEY = ""
 try:
     if "HOME" in os.environ:
         MASTODON_KEY = slurp(os.environ["HOME"] + "/.r2ai.mastodon-key").strip()
-except:
+except Exception:
     pass
 MASTODON_INSTANCE = "mastodont.cat"
 if "MASTODON_INSTANCE" in os.environ:
@@ -189,7 +194,7 @@ def filter_line(line):
         try:
             int(b)
             continue
-        except:
+        except Exception:
             pass
         if len(b) > 0:
             words.append(b)
@@ -216,60 +221,71 @@ def vectordb_search2(text, keywords, use_mastodon):
     if use_mastodon:
         print ("[r2ai] Searching in Mastodon", text)
         lines = mastodon_lines(text, keywords, True)
-#        print("LINES", lines)
         for line in lines:
-#            print("SAVE", line)
-            vectordb_instance.save(line, {"url":text})
+            vdb_add(f"{line} file={file}, url={url}")
+            idname=nextid()
     if have_vectordb is True and vectordb_instance is not None:
-        res = []
+        text = "trump"
+        res = {}
         try:
-            res = vectordb_instance.search(text, top_n=MAXMATCHES, unique=True, batch_results="diverse")
-        except:
-            try:
-                res = vectordb_instance.search(text, top_n=MAXMATCHES)
-            except:
-                traceback.print_exc()
-                pass
-        for r in res:
-            if "distance" in r:
-                # print("distance", r["distance"])
-                if r['distance'] < 1:
-                    result.append(r["chunk"])
-            else:
-                # when mprt is not available we cant find the distance
-                result.append(r["chunk"])
-    #print(result)
+            res = vectordb_instance.query(query_texts=text)
+        except Exception:
+            traceback.print_exc()
+            pass
+        if res["documents"] is not None:
+            for docs in res["documents"]:
+                for doc in docs:
+                    result.append(doc)
     return sorted(set(result))
 
+client = None
+
 def vectordb_init():
+    global have_vectordb, client, vectordb_instance
+    if have_vectordb is False:
+        return
+    if vectordb_instance is not None:
+        return
+    try:
+        have_vectordb = True
+    except Exception:
+        print("To better data index use: pip install chromadb")
+        return
+#    try:
+#        vectordb_instance = vectordb.Memory(embeddings="normal") # normal or fast
+#    except Exception:
+#        vectordb_instance = vectordb.Memory() # normal or fast
+    # vectordb_instance = vectordb.Memory() # normal or fast
+    client = chromadb.Client() # normal or fast
+    try:
+        vectordb_instance = client.delete_collection("r2ai")
+    except Exception:
+        pass
+    vectordb_instance = client.create_collection("r2ai")
+    if vectordb_instance is not None:
+        vdb_add("radare2 is a free reverse engineering tool written by pancake, aka Sergi Alvarez i Capilla. The project started in 2006 as a tool for domestic computer forensics in order to recover some deleted files and it continued the development adding new features like debugging, disassembler, decompiler, code analysis, advanced filesystem capabilities and integration with tons of tools like Frida, Radius, Ghidra, etc") # dummy entry
+
+def memorize(prompt, content):
     global have_vectordb, vectordb_instance
     if have_vectordb is False:
-        print("LEAVING")
+        builtins.print("no vdb found")
         return
-    if vectordb_instance is not None:
+    vectordb_init()
+    if vectordb_instance is None:
         return
-    try:
-        import vectordb
-        have_vectordb = True
-    except Exception as e:
-        os.system("python -m spacy download en_core_web_sm")
-        try:
-            import vectordb
-            have_vectordb = True
-        except:
-            have_vectordb = False
-            print("To better data index use:")
-            print("  pip install vectordb2")
-            print("On macOS you'll need to also do this:")
-            print("  python -m pip install spacy")
-            print("  python -m spacy download en_core_web_sm")
-            return
-    try:
-        vectordb_instance = vectordb.Memory(embeddings="best") # normal or fast
-    except:
-        vectordb_instance = vectordb.Memory() # normal or fast
-    if vectordb_instance is not None:
-        vectordb_instance.save("radare2 is a free reverse engineering tool written by pancake, aka Sergi Alvarez i Capilla. The project started in 2006 as a tool for domestic computer forensics in order to recover some deleted files and it continued the development adding new features like debugging, disassembler, decompiler, code analysis, advanced filesystem capabilities and integration with tons of tools like Frida, Radius, Ghidra, etc", {"url":"."}) # dummy entry
+    msg = f"{prompt} = {content}"
+    bs = 1024
+    while msg != "":
+        bob = msg[0:bs]
+#        bob = f"{prompt}={bob}"
+        vdb_add(bob)
+        msg = msg[bs:]
+
+g_id = 0
+def nextid():
+    global g_id
+    g_id = g_id + 1
+    return f"id{g_id}"
 
 def vectordb_search(text, keywords, source_files, use_mastodon, use_debug):
     global have_vectordb, vectordb_instance
@@ -290,17 +306,15 @@ def vectordb_search(text, keywords, source_files, use_mastodon, use_debug):
             continue
         lines = smart_slurp(file).splitlines()
         for line in lines:
-#            vectordb_instance.save(line)
-            vectordb_instance.save(line, {"url":file}) #, "url": file})
+            vdb_add(f"{line} file={file}")
             saved = saved + 1
     if use_mastodon:
         lines = mastodon_lines(text, None, True)
         for line in lines:
             saved = saved + 1
-            vectordb_instance.save(line, {"url":text})
+            vdb_add(f"{line} file={file}, url={url}")
     if saved == 0:
-        print("[r2ai] Nothing indexed")
-        vectordb_instance.save("", {})
+        print("[r2ai] Nothing was indexed")
     else:
         print("[r2ai] VectorDB index done")
     return vectordb_search2(text, keywords, use_mastodon)
@@ -327,8 +341,7 @@ class compute_rarity():
             else:
                 self.words[a] = 1
     def pull_realtime_lines(self, text, keywords, use_vectordb):
-        if self.env["debug"] == "true":
-            print(f"Pulling from mastodon {text}")
+        LOGGER.debug(f"Pulling from mastodon {text}")
         return mastodon_lines(text, keywords, use_vectordb)
 
     def find_matches(self, text, keywords):
@@ -395,7 +408,7 @@ def find_sources(srcdir):
     files = []
     try:
         files = os.walk(srcdir)
-    except:
+    except Exception:
         return []
     res = []
     for f in files:
@@ -421,28 +434,33 @@ def source_files(datadir, use_hist):
         files.append(R2AI_HISTFILE)
     return files
 
+def vdb_add(data):
+    global have_vectordb, vectordb_instance
+    vectordb_init()
+    if vectordb_instance is not None:
+        vectordb_instance.add(ids=[nextid()], documents=[data])
+
 def find_wikit(text, keywords):
-    print("wikit")
     global have_vectordb, vectordb_instance
     vectordb_init()
     if vectordb_instance is None:
-        print("vdb not initialized")
+        print("vdb not initialized", file=sys.stderr)
         return
     if keywords is not None:
         for kw in keywords:
             print("wikit " + kw)
             res = syscmdstr("wikit -a " + kw)
             if len(res) > 20:
-                vectordb_instance.save(res, {"url":kw})
+                vdb_add(f"{res} from url={kw}")
     words = filter_line(text)
     for kw in words:
         print("wikit " + kw)
         res = syscmdstr("wikit -a " + kw)
         if len(res) > 20:
-            vectordb_instance.save(res, {"keyword":kw})
+            vdb_add(f"{res} from url={kw}")
     res = syscmdstr("wikit -a '" + " ".join(words) + "'")
     if len(res) > 20:
-        vectordb_instance.save(res, {"keyword":kw})
+        vdb_add(f"{res} from keyword={kw}")
 
 def reset():
     global vectordb_instance

Attic/py/r2ai/interpreter.py

@@ -0,0 +1,579 @@
+import builtins
+import re
+import os
+import sys
+import traceback
+import json
+import platform
+have_local = True
+import getpass
+
+try:
+    import tokentrim
+except Exception:
+    have_local = False
+
+from rich.rule import Rule
+from signal import signal, SIGINT
+
+from .env import R2AiEnv
+from .large import Large
+from .utils import merge_deltas
+from .message_block import MessageBlock
+from .code_block import CodeBlock
+from .backend import kobaldcpp
+
+from .models import get_hf_llm, new_get_hf_llm, get_default_model
+from .voice import tts
+from .const import R2AI_HOMEDIR
+from . import LOGGER, logging
+from .web import stop_http_server, server_running
+from .progress import progress_bar
+
+file_dir = os.path.dirname(__file__)
+sys.path.append(file_dir)
+import index
+
+r2clippy = False
+have_rlang = False
+try:
+    import r2lang
+    have_rlang = True
+    print = r2lang.print
+    r2clippy = True
+except Exception:
+    pass
+
+Ginterrupted = False
+def signal_handler(sig, frame):
+    global Ginterrupted
+    if Ginterrupted:
+        stop_http_server(force=True) # kill if in background
+        sys.exit(0) # throws exception
+    Ginterrupted = True
+    print("^C 0", file=sys.stderr)
+    stop_http_server(force=False)
+signal(SIGINT, signal_handler)
+
+def exception_handler(self, sig, frame):
+    global Ginterrupted
+    traceback.print_exc()
+    if Ginterrupted:
+        sys.exit(0) # throws exception
+    Ginterrupted = True
+    print("^C 1", file=sys.stderr)
+# sys.excepthook = exception_handler
+
+def incodeblock(msg):
+  return "content" in msg and msg["content"].count("```") % 2 == 1
+
+def r2eval(m):
+    if "$(" in m and have_rlang:
+        def evaluate_expression(match):
+            expression = match.group(1)
+            try:
+                result = r2lang.cmd(expression)
+                return result
+            except Exception as e:
+                return f"Error: {e}"
+        return re.sub(r'\$\((.*?)\)', evaluate_expression, m)
+    return m
+
+from utils import syscmdstr
+
+def ddg(m):
+    print("[R2AI] Crawling the web with ddg and curl+sed")
+    m = m.replace("'", "")
+    res = syscmdstr(f"cd examples; ./scrap-ddgweb.sh '{m}'")
+    return f"Considering:\n```{res}\n```\n"
+
+def is_litellm_model(model):
+    import warnings
+    warnings.filterwarnings('ignore')
+    import litellm
+    provider = None
+    model_name = None
+    if model.startswith ("/"):
+        return False
+    provider = None
+    if "/" in model:
+        provider = model.split("/")[0]
+    elif ":" in model:
+        provider = model.split(":")[0]
+    if provider in litellm.LITELLM_CHAT_PROVIDERS:
+        return True
+    return False
+
+class Interpreter:
+    def __init__(self):
+        self.logger = LOGGER
+        self.mistral = None
+        self.messages = []
+        self.terminator = "</s>" ## taken from the model using the llama api
+        self.api_key = None # openai?
+        self.print = None
+        self.auto_run = False
+        self.model = get_default_model()
+        self.last_model = ""
+        self.env = R2AiEnv()
+        self.system_message = ""
+        self.env["llm.model"] = self.model ## TODO: dup. must get rid of self.model
+        self.env["llm.gpu"] = "true"
+        self.env["llm.layers"] = "-1"
+        self.env["llm.window"] = "32768" # "4096" # context_window
+        self.env["llm.maxtokens"] = "4096" # "1750"
+        self.env["llm.maxmsglen"] = "8096" # "1750"
+        self.env["llm.temperature"] = "0.002"
+        self.env["llm.repeat_penalty"] = "1.0"
+        self.env["llm.top_p"] = "0.95"
+        self.env["llm.top_k"] = "50"
+        self.env["user.name"] = "" # TODO auto fill?
+        self.env["user.os"] = ""
+        self.env["user.arch"] = ""
+        self.env["user.cwd"] = ""
+        self.env["user.editor"] = ""
+        self.env["user.plugins"] = f"{R2AI_HOMEDIR}/plugins"
+        self.env["voice.lang"] = "en"
+        self.env["voice.model"] = "base"
+        self.env["data.use"] = "false"
+        self.env["data.path"] = f"{R2AI_HOMEDIR}/doc/data"
+        self.env["data.local"] = "false"
+        self.env["data.wikit"] = "false"
+        self.env["data.ddg"] = "false"
+        self.env["data.mastodon"] = "false"
+        self.env["data.vectordb"] = "false"
+        self.env["data.hist"] = "false"
+        self.env["key.mastodon"] = ""
+        self.env["http.port"] = "8080"
+        self.env["http.tabby"] = "false"
+        self.env["http.path"] = ""
+        self.env["http.verbose"] = "true" # not used yet
+        self.env["http.chatctx"] = "false"
+        self.env["debug_level"] = os.getenv("R2AI_LOG", "2")
+        if have_rlang:
+            self.env["chat.live"] = "false"
+        else:
+            self.env["chat.live"] = "true"
+        # self.env["chat.model"] = "" # TODO
+        self.env["chat.trim"] = "false"
+        self.env["chat.voice"] = "false"
+        self.env["chat.bubble"] = "false"
+        self.env["chat.reply"] = "false"
+        self.env["chat.code"] = "true"
+        self.env["chat.rawdog"] = "false"
+        self.env["chat.reasoning_effort"] = "high"
+        self.env["chat.stream"] = "true"
+        self.env["chat.show_cost"] = "false"
+        self.env["chat.drop_params"] = "false"
+        self.env["chat.openai_store"] = "false"
+        self.env["auto.max_runs"] = "100"
+        self.env["auto.hide_tool_output"] = "false"
+        self.env["auto.init_commands"] = "aaa;iI;afl"
+        self.env["auto.ask_to_execute"] = "true"  # ask to run tools such as r2cmd, run_python etc
+
+        self.env.add_callback("debug_level", lambda val: LOGGER.setLevel(int(val) * 10))
+
+        # No active block to start
+        # (blocks are visual representation of messages on the terminal)
+        self.active_block = None
+
+        # Note: While Open Interpreter can use Llama, we will prioritize gpt-4.
+        # gpt-4 is faster, smarter, can call functions, and is all-around easier to use.
+        # This makes gpt-4 better aligned with Open Interpreters priority to be easy to use.
+        self.llama_instance = None
+        self.large = Large(self)
+
+    def get_info_for_system_message(self):
+        """Gets relevent information for the system message."""
+        info = ""
+        # Add user info
+        getpass.getuser()
+        os.getcwd()
+        platform.system()
+        # info += f"[User Info]\nName: {username}\nCWD: {current_working_directory}\nOS: {operating_system}"
+        return info
+
+    def reset(self):
+        """Resets the interpreter."""
+        self.messages = []
+
+    def load(self, messages):
+        self.messages = messages
+
+    def save(self, f):
+        json.dumps(self.messages, f, indent=2)
+
+    def handle_undo(self, arguments):
+        """Remove all messages after the most recent user entry (and the entry itself).
+        Therefore user can jump back to the latest point of conversation.
+        Also gives a visual representation of the messages removed."""
+        if len(self.messages) == 0:
+            return
+        # Find the index of the last 'role': 'user' entry
+        last_user_index = None
+        for i, message in enumerate(self.messages):
+            if message.get('role') == 'user':
+                last_user_index = i
+        removed_messages = []
+        # Remove all messages after the last 'role': 'user'
+        if last_user_index is not None:
+            removed_messages = self.messages[last_user_index:]
+            self.messages = self.messages[:last_user_index]
+
+        # Print out a preview of what messages were removed.
+        for message in removed_messages:
+            if message.get("content"):
+                print(f"**Removed message:** `\"{message['content'][:30]}...\"`")
+            elif 'function_call' in message:
+                print(f"**Removed codeblock**") # TODO: Could add preview of code removed here.
+
+    # reimplement using nltk or index.match
+    def keywords_ai(self, text):
+        # kws = self.keywords_ai("who is the author of radare?") => "author,radare2"
+        words = []
+        mmname = "TheBloke/Mistral-7B-Instruct-v0.2-GGUF"
+        ctxwindow = int(self.env["llm.window"])
+        mm = new_get_hf_llm(self, mmname, ctxwindow)
+        msg = f"Considering the sentence \"{text}\" as input, Take the KEYWORDS or combination of TWO words from the given text and respond ONLY a comma separated list of the most relevant words. DO NOT introduce your response, ONLY show the words"
+        msg = f"Take \"{text}\" as input, and extract the keywords and combination of keywords to make a search online, the output must be a comma separated list" #Take the KEYWORDS or combination of TWO words from the given text and respond ONLY a comma separated list of the most relevant words. DO NOT introduce your response, ONLY show the words"
+        response = mm(msg, stream=False, temperature=0.1, stop="</s>", max_tokens=1750)
+        self.logger.debug(response)
+        text0 = response["choices"][0]["text"]
+        text0 = text0.replace('"', ",")
+        if text0.startswith("."):
+            text0 = text0[1:].strip()
+        try:
+            text0 = text0.split(":")[1].strip()
+        except Exception:
+            pass
+        # print(text0)
+        mm = None
+        return [word.strip() for word in text0.split(',')]
+
+    @progress_bar("Thinking", color="yellow") 
+    def chat(self, message=None):
+        global print
+        global Ginterrupted
+        if self.print is not None:
+            print = self.print
+
+        if self.last_model != self.model:
+#            self.llama_instance = None
+            self.last_model = self.model
+
+        if not message and self.env["chat.code"] == "true":
+            self.end_active_block()
+            self.logger.error("Missing message")
+            return
+
+        if self.env["data.use"] == "true":
+            use_hist = self.env["data.hist"] == "true"
+            use_wikit = self.env["data.wikit"] == "true"
+            use_mastodon = self.env["data.mastodon"] == "true"
+            use_vectordb = self.env["data.vectordb"] == "true"
+            use_debug = LOGGER.level == logging.DEBUG
+            datadir = None
+            keywords = None
+            if use_mastodon:
+                keywords = self.keywords_ai(message)
+            if self.env["data.local"] == "true":
+                datadir = self.env["data.path"]
+            if self.env["data.ddg"] == "true":
+                results = ddg(message)
+                if use_vectordb:
+                    index.memorize(message, results)
+                else:
+                    self.messages.append({"role": "hint", "content": results})
+            matches = index.match(message, keywords, datadir, use_hist, use_mastodon, use_debug, use_wikit, use_vectordb)
+            if not matches:
+                matches = []
+
+            if len(matches) > 0:
+                for m in matches:
+                    self.logger.debug("HINT: " + m)
+                    self.messages.append({"role": "hint", "content": r2eval(m)})
+
+        self.logger.debug(message)
+
+        # print(message)
+        # Local model -- TODO: assume local models don't have : in the name?
+        if (
+            not self.model.startswith("openai:") and
+            not self.model.startswith("openapi:") and
+            not self.model.startswith("google:") and
+            not self.model.startswith("kobaldcpp") and
+            not self.model.startswith("bedrock:") and
+            not is_litellm_model(self.model)
+        ):
+            self.logger = LOGGER.getChild(f"local:{self.model}")
+            # Find or install Code-Llama
+            try:
+                ctxwindow = int(self.env["llm.window"])
+                self.llama_instance = new_get_hf_llm(self, self.model, ctxwindow)
+                if self.llama_instance is None:
+                    self.logger.error("Cannot find model " + self.model)
+                    return
+            except Exception:
+                traceback.print_exc()
+
+        # If it was, we respond non-interactively
+        self.messages.append({"role": "user", "content": message})
+        response = None
+        try:
+            response = self.respond()
+            self.clear_hints()
+        except Exception:
+            if Ginterrupted:
+                Ginterrupted = False
+            else:
+                traceback.print_exc()
+        self.end_active_block()
+        return response
+    
+    def end_active_block(self):
+        # if self.env["chat.code"] == "false":
+        # return
+        if self.active_block:
+            self.active_block.end()
+            self.active_block = None
+
+    def environment(self):
+        kvs = ""
+        if self.env["user.name"] != "":
+            kvs += "Name: " + self.env["user.name"] + "\n"
+        if self.env["user.os"] != "":
+            kvs += "OS: " + self.env["user.os"] + "\n"
+        if self.env["user.cwd"] != "":
+            kvs += "CWD: " + self.env["user.cwd"] + "\n"
+        if self.env["user.editor"] != "":
+            kvs += "EDITOR: " + self.env["user.editor"] + "\n"
+        if self.env["user.arch"] != "":
+            kvs += "ARCH: " + self.env["user.arch"] + "\n"
+        # info += f"[User Info]\nName: {username}\nCWD: {current_working_directory}\nOS: {operating_system}"
+        if kvs != "":
+            return "[User Info]\n" + kvs
+        return ""
+
+    def clear_hints(self):
+        res = []
+        for msg in self.messages:
+            if "role" in msg and msg["role"] != "hint":
+                res.append(msg)
+        self.messages = res
+
+    def respond(self):
+        global Ginterrupted
+        self.logger = LOGGER.getChild(self.model.split(":")[0])
+        maxtokens = int(self.env["llm.maxtokens"])
+        # Add relevant info to system_message
+        # (e.g. current working directory, username, os, etc.)
+        info = self.get_info_for_system_message()
+
+        # This is hacky, as we should have a different (minified) prompt for CodeLLama,
+        # but for now, to make the prompt shorter and remove "run_code" references, just get the first 2 lines:
+        self.system_message = "\n".join(self.system_message.split("\n")[:2])
+        # self.system_message += "\nOnly do what the user asks you to do, then ask what they'd like to do next."
+
+        system_message = self.system_message + "\n\n" + info
+        system_message += self.environment()
+
+        if self.env["chat.trim"] == "true":
+            ## this stupid function is slow as hell and doesn not provides much goodies
+            ## just ignore it by default
+            messages = tokentrim.trim(self.messages, max_tokens=maxtokens, system_message=system_message)
+        else:
+            messages = self.large.compress_messages(self.messages)
+
+            self.logger.debug(messages)
+
+        # builtins.print(prompt)
+        response = None
+        if self.auto_run:
+            from . import auto
+            import litellm
+            if self.env["chat.drop_params"] == "true":
+                litellm.drop_params = True
+            else:
+                litellm.drop_params = False
+
+            if(is_litellm_model(self.model)):
+                response = auto.chat(self)
+            else:
+                self.llama_instance = new_get_hf_llm(self, self.model, int(self.env["llm.window"]))
+                response = auto.chat(self)
+            return
+
+        elif self.model.startswith("kobaldcpp"):
+            if self.system_message != "":
+                message = f"Context:\n```\n{self.system_message}\n```\n"
+            else:
+                message = ""
+            #f"{Your name is r2ai, an assistant for radare2. User will ask about actions and you must respond with the radare2 command associated or the answer to the question. Be precise and concise when answering"
+            for m in messages:
+                role = m["role"]
+                content = m["content"]
+                if role == "user":
+                    message += f"User: {content}\n"
+                elif role == "assistant":
+                    message += f"AI: {content}\n"
+            response = ""
+            if ":" in self.model:
+                uri = self.model.split(":")[1:]
+                response = kobaldcpp.chat(message, ":".join(uri))
+            else:
+                response = kobaldcpp.chat(message)
+            if "content" in self.messages[-1]:
+                last_message = self.messages[-1]["content"]
+            if self.env["chat.reply"] == "true":
+                self.messages.append({"role": "assistant", "content": response})
+            print(response)
+            return
+
+        elif is_litellm_model(self.model):
+            # [
+            #  {"role": "system", "content": "You are a poetic assistant, be creative."},
+            #  {"role": "user", "content": "Compose a poem that explains the concept of recursion in programming."}
+            # ]
+            from litellm import completion as litellm_completion
+            completion = None
+            if self.model.startswith("ollama/"):
+                completion = litellm_completion(
+                    model=self.model,
+                    messages=self.messages,
+                    drop_params=True,
+                    max_completion_tokens=maxtokens,
+                    temperature=float(self.env["llm.temperature"]),
+                    top_p=float(self.env["llm.top_p"]),
+                )
+            elif self.model.startswith("mistral/"):
+                completion = litellm_completion(
+                    drop_params=True,
+                    model=self.model.replace(":", "/"),
+                    messages=self.messages,
+                    max_completion_tokens=maxtokens,
+                    temperature=float(self.env["llm.temperature"]),
+                    top_p=float(self.env["llm.top_p"]),
+                )
+            else:
+                completion = litellm_completion(
+                    model=self.model.replace(":", "/"),
+                    messages=self.messages,
+                    max_completion_tokens=maxtokens,
+                    temperature=float(self.env["llm.temperature"]),
+                    top_p=float(self.env["llm.top_p"]),
+                )
+
+            response = completion.choices[0].message.content
+            if "content" in self.messages[-1]:
+                last_message = self.messages[-1]["content"]
+                if self.env["chat.reply"] == "true":
+                    self.messages.append({"role": "assistant", "content": response})
+                print(response)
+                return response
+
+        else:
+            # non-openai aka local-llama model
+            if self.llama_instance == None:
+                self.logger.critical("Llama is not instantiated")
+                return
+            try:
+                if type(self.terminator).__name__ == "list":
+                    terminator = self.terminator
+                else:
+                    terminator = [self.terminator]
+                chat_args = {
+                    "stream": True,
+                    "temperature": float(self.env["llm.temperature"]),
+                    "repeat_penalty": float(self.env["llm.repeat_penalty"]),
+                    "top_p": float(self.env["llm.top_p"]),
+                    "top_k": int(self.env["llm.top_k"]),
+                    "stop": terminator,
+                    "max_tokens": maxtokens
+                }
+                if self.env["chat.rawdog"] == "true":
+                    from .completion import messages_to_prompt
+                    allmsg = messages.copy()
+                    allmsg.insert(0, ChatMessage(role=MessageRole.SYSTEM, content=self.system_message))
+                    prompt = messages_to_prompt(self, allmsg)
+                    response = self.llama_instance(prompt, **chat_args)
+                else:
+                    all_messages = messages.copy()
+                    all_messages.insert(0, {"role": "system", "content": self.system_message})
+                    response = self.llama_instance.create_chat_completion(all_messages, **chat_args)
+            except Exception as err:
+                traceback.print_exc()
+                print(Exception, err)
+                if Ginterrupted:
+                    Ginterrupted = False
+                    return
+
+        if response is None:
+            print("No response")
+            ctxwindow = int(self.env["llm.window"])
+            self.llama_instance = new_get_hf_llm(self, self.model, ctxwindow)
+            return
+        # Initialize message, function call trackers, and active block
+        self.messages.append({})
+        in_function_call = False
+        self.active_block = MessageBlock()
+        for chunk in response:
+            if Ginterrupted:
+                Ginterrupted = False
+                break
+            text = ''
+            delta = None
+            if self.env["chat.rawdog"] == "true":
+                if "content" not in messages[-1]:
+                    text = chunk["choices"][0]['text'].capitalize()
+                    messages[-1] = { "content": '', "role": "assistant" }
+                else:
+                    text = chunk["choices"][0]['text']
+                messages[-1] = merge_deltas(messages[-1], {"content": text})
+            else:
+                if "content" not in messages[-1]:
+                    if "text" in chunk["choices"][0]['delta']:
+                        text = chunk["choices"][0]['delta']["content"]
+                    messages[-1] = { "content": text, "role": "assistant" }
+                    delta = {"content": text}
+                elif "content" in chunk['choices'][0]['delta']:
+                    delta = {"content": chunk['choices'][0]['delta']['content']}
+                # Accumulate deltas into the last message in messages
+                if delta:
+                    self.messages[-1] = merge_deltas(self.messages[-1], delta)
+            if self.env["chat.live"] != "true":
+                continue
+            flushed = False
+            if self.env["chat.code"] == "true":
+                if incodeblock(self.messages[-1]):
+                    if in_function_call == False:
+                        in_function_call = True
+                        self.active_block.update_from_message(self.messages[-1])
+                        self.end_active_block()
+                        # flushed = True
+                        self.active_block = CodeBlock()
+                else:
+                    if in_function_call == True:
+                        in_function_call = False
+                        self.end_active_block()
+                        flushed = True
+                        self.active_block = MessageBlock()
+            if self.env["chat.live"] == "true": # and self.env["chat.code"] == "true":
+                self.active_block.update_from_message(self.messages[-1])
+                if flushed:
+                    self.messages[-1]["content"] = ""
+            continue # end of for loop
+        self.end_active_block()
+        self.active_block = None
+        output_text = ""
+        if len(self.messages) > 0 and "content" in self.messages[-1]:
+            output_text = self.messages[-1]["content"].strip()
+        if self.env["chat.reply"] == "true":
+            self.messages.append({"role": "assistant", "content": output_text})
+        if self.env["chat.voice"] == "true":
+            tts("(assistant)", output_text, self.env["voice.lang"])
+        elif self.env["chat.live"] != "true":
+            try:
+                r2lang.print(output_text)
+            except Exception:
+                print(output_text)
+            # print(str(self.messages))

Attic/py/r2ai/large.py

@@ -1,4 +1,5 @@
 from .models import get_hf_llm, new_get_hf_llm, get_default_model
+from . import LOGGER
 import re
 import json
 
@@ -10,6 +11,7 @@ class Large:
         self.maxtokens = 5000
         # self.model = "TheBloke/Mistral-7B-Instruct-v0.2-GGUF"
         self.model = "FaradayDotDev/llama-3-8b-Instruct-GGUF"
+        self.logger = LOGGER.getChild(f"large[{self.model}]")
         if ai is not None:
             self.ai = ai
         else:
@@ -29,7 +31,7 @@ class Large:
 
     def compress_text(self, msg):
         if self.mistral == None:
-            self.mistral = new_get_hf_llm(self.ai, self.model, False, self.window)
+            self.mistral = new_get_hf_llm(self.ai, self.model, self.window)
         # q = f"Rewrite this code into shorter pseudocode (less than 500 tokens). keep the comments and essential logic:\n```\n{msg}\n```\n"
         #q = f"Rewrite this code into shorter pseudocode (less than 200 tokens). keep the relevant comments and essential logic:\n```\n{msg}\n```\n"
         q = f"Resumen y responde SOLO la información relevante del siguiente texto:\n{msg}"
@@ -58,19 +60,18 @@ class Large:
         # kws = self.keywords_ai("who is the author of radare?") => "author,radare2"
         words = []
         ctxwindow = int(self.ai.env["llm.window"])
-        mm = new_get_hf_llm(self.ai, self.model, False, ctxwindow)
+        mm = new_get_hf_llm(self.ai, self.model, ctxwindow)
         msg = f"Considering the sentence \"{text}\" as input, Take the KEYWORDS or combination of TWO words from the given text and respond ONLY a comma separated list of the most relevant words. DO NOT introduce your response, ONLY show the words"
         msg = f"Take \"{text}\" as input, and extract the keywords and combination of keywords to make a search online, the output must be a comma separated list" #Take the KEYWORDS or combination of TWO words from the given text and respond ONLY a comma separated list of the most relevant words. DO NOT introduce your response, ONLY show the words"
         response = mm(msg, stream=False, temperature=0.001, stop="</s>", max_tokens=1750)
-        if self.ai.env["debug"] == "true":
-            print("KWSPLITRESPONSE", response)
+        self.logger.debug(response)
         text0 = response["choices"][0]["text"]
         text0 = text0.replace('"', ",")
         if text0.startswith("."):
             text0 = text0[1:].strip()
         try:
             text0 = text0.split(":")[1].strip()
-        except:
+        except Exception:
             pass
         # print(text0)
         mm = None
@@ -111,7 +112,7 @@ class Large:
         words = []
         if self.mistral == None:
             ctxwindow = int(self.ai.env["llm.window"])
-            self.mistral = new_get_hf_llm(self.ai, self.model, False, ctxwindow)
+            self.mistral = new_get_hf_llm(self.ai, self.model, ctxwindow)
         # q = f"Rewrite this code into shorter pseudocode (less than 500 tokens). keep the comments and essential logic:\n```\n{msg}\n```\n"
         q = f"Rewrite this code into shorter pseudocode (less than 200 tokens). keep the relevant comments and essential logic:\n```\n{msg}\n```\n"
         response = self.mistral(q, stream=False, temperature=0.1, stop="</s>", max_tokens=4096)

Attic/py/r2ai/main.py

@@ -0,0 +1,174 @@
+#!/usr/bin/env python3
+
+import os
+os.environ["TRANSFORMERS_NO_ADVISORY_WARNINGS"] = "1"
+import sys
+import builtins
+import traceback
+import appdirs
+import argparse
+import litellm
+from r2ai.repl import r2ai_singleton
+from r2ai.utils import slurp
+from r2ai.repl import runline, r2ai_repl, help_message
+
+from r2ai.pipe import open_r2, get_r2_inst
+from r2ai.const import R2AI_RCFILE
+
+OPENAI_KEY = ""
+HAVE_RLANG = False
+HAVE_R2PIPE = False
+RCFILE_LOADED = False
+
+def r2ai_rlang_plugin(unused_but_required_argument):
+    ai = r2ai_singleton()
+    def _call(s):
+        if not s.startswith("r2ai"):
+            return False
+        try:
+            run_rcfile_once(ai)
+            if len(s) == 4:
+                builtins.print(help_message)
+            else:
+                usertext = s[4:].strip()
+                runline(ai, usertext)
+        except Exception as e:
+            builtins.print(e)
+            traceback.print_exc()
+        return True
+
+    return {
+        "name": "r2ai",
+        "license": "MIT",
+        "desc": "run llama language models inside r2",
+        "call": _call,
+    }
+
+# TODO: see repl.run_script as replacement
+def run_rcfile(ai):
+    try:
+        lines = slurp(R2AI_RCFILE)
+        
+        for line in lines.split("\n"):
+            if line.strip() != "":
+                if ai is None:
+                    ai = r2ai_singleton()
+                runline(ai, line)
+    except Exception:
+        pass
+    if ai is None:
+        ai = r2ai_singleton()
+
+def run_rcfile_once(ai):
+    global RCFILE_LOADED
+    if not RCFILE_LOADED:
+        run_rcfile(ai)
+        RCFILE_LOADED = True
+
+
+def main(args, commands, dorepl=True):
+
+    os.environ["TOKENIZERS_PARALLELISM"] = "false"
+    os.environ["TRANSFORMERS_NO_ADVISORY_WARNINGS"] = "1"
+
+    try:
+        r2aihome = os.path.dirname(os.path.realpath(__file__))
+        sys.path.append(r2aihome)
+        # if available
+        sys.path.append(
+            os.path.join(r2aihome, "..", "vectordb")
+        )
+    except Exception:
+        traceback.print_exc()
+
+    home_dir = os.path.expanduser("~")
+    # create symlink if it doesnt exist
+    try:
+        dst = os.path.join(home_dir, ".r2ai.models")
+        udd = appdirs.user_data_dir("r2ai")
+        src = os.path.join(udd, "models")
+        if not os.path.exists(dst):
+            os.symlink(src, dst)
+    except Exception:
+        traceback.print_exc()
+
+    r2_openai_file = os.path.join(home_dir, ".r2ai.openai-key")
+    if os.path.isfile(r2_openai_file):
+        apikey = slurp(r2_openai_file).strip()
+        os.environ["OPENAI_API_KEY"] = apikey
+        print("[R2AI] OpenAI API key loaded from ~/.r2ai.openai-key", file=sys.stderr)
+
+    r2_mistral_file = os.path.join(home_dir, ".r2ai.mistral-key")
+    if os.path.isfile(r2_mistral_file):
+        apikey = slurp(r2_mistral_file).strip()
+        os.environ["MISTRAL_API_KEY"] = apikey
+        print("[R2AI] Mistral API key loaded from ~/.r2ai.mistral-key", file=sys.stderr)
+
+    r2_gemini_file = os.path.join(home_dir, ".r2ai.gemini-key")
+    if os.path.isfile(r2_gemini_file):
+        apikey = slurp(r2_gemini_file).strip()
+        os.environ["GEMINI_API_KEY"] = apikey
+        print("[R2AI] Gemini API key loaded from ~/.r2ai.gemini-key", file=sys.stderr)
+
+    r2_anthropic_file = os.path.join(home_dir, ".r2ai.anthropic-key")
+    if os.path.isfile(r2_anthropic_file):
+        apikey = slurp(r2_anthropic_file).strip()
+        os.environ["ANTHROPIC_API_KEY"] = apikey
+        print("[R2AI] Anthropic API key loaded from ~/.r2ai.anthropic-key", file=sys.stderr)
+    
+    ai = r2ai_singleton()
+    if "R2PIPE_IN" in os.environ:
+        pass
+    elif args.bin:
+        open_r2(vars(args)["bin"], flags=["-2"])
+
+    if commands is not None:
+        for c in commands:
+            if c.startswith("_"):
+                runline(ai, "-" + c[1:])
+            else:
+                runline(ai, c)
+    if dorepl:
+        r2ai_repl(ai)
+
+def massage_args(args):
+    runrepl = True
+    if args.command is None:
+        args.command = []
+    if args.webserver:
+        args.command.append("-w")
+    if args.eval:
+        if args.eval == "default":
+            args.command.append("-e")
+            runrepl = False
+        else:
+            args.command.append(f"-e {args.eval}")
+    if args.port:
+        if args.port == "default":
+            runrepl = False
+            args.command.append("-e http.port")
+        else:
+            args.command.append(f"-e http.port={args.port}")
+    if args.model:
+        if args.model == "default":
+            args.command.append("-mm")
+            runrepl = False
+        else:
+            args.command.append(f"-m {args.model}")
+    return runrepl, args
+
+def run():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("bin", nargs="?", type=str)
+    parser.add_argument("-w", "--webserver", action="store_true",
+        help="Start the r2ai webserver. Same as r2ai -c=-w")
+    parser.add_argument("-p", "--port", type=str, nargs="?", const="default",
+        help="Change listen port number")
+    parser.add_argument("-e", "--eval", type=str, nargs="?", const="default",
+        help="Change configuration variable")
+    parser.add_argument("-m", "--model", type=str, nargs="?", const="default",
+        help="Select model name")
+    parser.add_argument("-c", "--command", action="append",
+        help="Command to be executed. Can be passed multiple times.")
+    runrepl, args = massage_args(parser.parse_args())
+    main(args, args.command, runrepl)

Attic/py/r2ai/models.py

@@ -1,18 +1,14 @@
 from .utils import slurp, dump
-from huggingface_hub import hf_hub_download, login
-from huggingface_hub import HfApi, list_repo_tree, get_paths_info
 from typing import Dict, List, Union
 import appdirs
 import builtins
 import inquirer
 import json
-import llama_cpp
 import os
 import shutil
 import subprocess
 import sys
 import traceback
-
 # DEFAULT_MODEL = "TheBloke/CodeLlama-34B-Instruct-GGUF"
 # DEFAULT_MODEL = "TheBloke/llama2-7b-chat-codeCherryPop-qLoRA-GGUF"
 # DEFAULT_MODEL = "-m TheBloke/dolphin-2_6-phi-2-GGUF"
@@ -28,28 +24,47 @@ def get_default_model():
         fd.close()
         if "default" in usermodels:
             return usermodels["default"]
-    except:
+    except Exception:
         pass
     return DEFAULT_MODEL
 
 def Markdown(x):
-  return x
+    return x
 
 def mainmodels():
     return """
-Local:
+Decai:
+-m hugging-quants/Llama-3.2-3B-Instruct-Q8_0-GGUF
+-m QuantFactory/granite-8b-code-instruct-4k-GGUF
+-m cognitivecomputations/dolphin-2.9.3-mistral-nemo-12b-gguf
+-m bartowski/Gemma-2-9B-It-SPPO-Iter3-GGUF
+-m Qwen/Qwen2.5-Coder-7B-Instruct-GGUF
+-m trufae/granite3-instruct-8b-gguf
+Code:
+-m unsloth/Llama-3.2-1B-Instruct-GGUF
+-m mradermacher/starcoder-i1-GGUF
+-m cognitivecomputations/dolphin-2.9.4-llama3.1-8b-gguf
 -m FaradayDotDev/llama-3-8b-Instruct-GGUF
--m TheBloke/Mistral-7B-Instruct-v0.2-GGUF
+-m second-state/Mistral-Nemo-Instruct-2407-GGUF
 -m bartowski/gemma-2-9b-it-GGUF
--m dagbs/dolphin-2.8-mistral-7b-v02-GGUF
--m meetkai/functionary-small-v2.2-GGUF
+-m mradermacher/salamandra-7b-instruct-aina-hack-GGUF
+Functionary:
+-m meetkai/functionary-small-v3.2-GGUF
+-m mradermacher/granite-20b-functioncalling-i1-GGUF
+-m mradermacher/LocalAI-Llama3.1-8b-Function-Call-v0.3-i1-GGUF
 Uncensored:
--m Orenguteng/Llama-3-8B-Lexi-Uncensored-GGUF
--m TheBloke/Unholy-v2-13B-GGUF
--m Undi95/Utopia-13B-GGUF
+-m Ttimofeyka/MistralRP-Noromaid-NSFW-Mistral-7B-GGUF
+-m bartowski/Phi-3.5-mini-instruct_Uncensored-GGUF
+-m Orenguteng/Llama-3.1-8B-Lexi-Uncensored-V2-GGUF
+-m LWDCLS/DarkIdol-Llama-3.1-8B-Instruct-1.2-Uncensored-GGUF-IQ-Imatrix-Request
+-m Undi95/Llama-3-Unholy-8B-GGUF
+-m mradermacher/Llama-3.2-1B-Instruct-Uncensored-GGUF
 Remote:
 -m openapi:http://localhost:5001
--m anthropic:claude-3-5-sonnet-20240620
+-m openai:gpt-4
+-m anthropic:claude-sonnet-5-5
+-m bedrock:anthropic.claude-3-5-sonnet-20241022-v1
+-m kobaldcpp:http://localhost:5001
 """
 
 def models():
@@ -59,22 +74,22 @@ KobaldCpp:
 -m kobaldcpp:http://localhost:5001
 OpenAPI:
 -m openapi:http://localhost:5001
+-m openai:http://localhost:5001:gpt-4
 OpenAI:
 -m openai:gpt-3.5-turbo
 -m openai:gpt-4
 -m openai:gpt-4-1106-preview
 Anthropic:
--m anthropic:claude-2.1
--m anthropic:claude-3-haiku-20240307
--m anthropic:claude-3-opus-20240229
--m anthropic:claude-3-sonnet-20240229
--m anthropic:claude-3-5-sonnet-20240620
+-m anthropic:claude-opus-4-8
+-m anthropic:claude-sonnet-5-5
+-m anthropic:claude-haiku-4-5-20251001
 groq:
 -m groq:gemma-7b-it
 -m groq:llama2-70b-4096
 -m groq:mixtral-8x7b-32768
 Google:
 -m google:gemini-1.0-pro
+-m google:gemini-1.5-flash
 -m google:gemini-1.5-pro-latest
 Functionary:
 -m meetkai/functionary-7b-v1.4-GGUF
@@ -131,9 +146,7 @@ Other:
 def gpulayers(ai):
     if "llm.gpu" in ai.env:
         if ai.env["llm.gpu"] == "true":
-            print("[r2ai] Using GPU")
-            return -1
-    print("[r2ai] Using CPU")
+            return int(ai.env["llm.layers"])
     return 0
 
 def get_hf_llm(ai, repo_id, debug_mode, context_window):
@@ -143,9 +156,9 @@ def get_hf_llm(ai, repo_id, debug_mode, context_window):
             fd = open(r2ai_model_json)
             usermodels = json.load(fd)
             fd.close()
-        except:
+        except Exception:
             usermodels = {}
-            pass
+
         model_path = "" # slurp(r2ai_model_json)
         if not repo_id:
             repo_id = get_default_model()
@@ -154,9 +167,10 @@ def get_hf_llm(ai, repo_id, debug_mode, context_window):
         if usermodels is not None and repo_id in usermodels:
             model_path = usermodels[repo_id]
 #            print(f"[r2ai] Using {r2ai_model_json} {model_path}")
-            return llama_cpp.Llama(model_path=model_path, n_gpu_layers=gpulayers(ai), verbose=debug_mode, n_ctx=context_window)
-    except:
+            return get_llama_inst(repo_id, model_path=model_path, n_gpu_layers=gpulayers(ai), verbose=debug_mode, n_ctx=context_window)
+    except Exception:
         traceback.print_exc()
+
     print(f"Select {repo_id} model. See -M and -m flags", file=sys.stderr)
     raw_models = list_gguf_files(repo_id)
     if not raw_models:
@@ -242,6 +256,7 @@ def get_hf_llm(ai, repo_id, debug_mode, context_window):
 
             # Check if model was originally split
             split_files = [model["filename"] for model in raw_models if selected_model in model["filename"]]
+            from huggingface_hub import hf_hub_download
             if len(split_files) > 1:
                 # Download splits
                 for split_file in split_files:
@@ -253,18 +268,14 @@ def get_hf_llm(ai, repo_id, debug_mode, context_window):
                     hf_hub_download(
                         repo_id=repo_id,
                         filename=split_file,
-                        local_dir=default_path,
-                        local_dir_use_symlinks=False,
-                        resume_download=True)
+                        local_dir=default_path)
                 # Combine and delete splits
                 actually_combine_files(default_path, selected_model, split_files)
             else:
                 hf_hub_download(
                     repo_id=repo_id,
                     filename=selected_model,
-                    local_dir=default_path,
-                    local_dir_use_symlinks=False,
-                    resume_download=True)
+                    local_dir=default_path)
 
             model_path = download_path
         else:
@@ -272,7 +283,7 @@ def get_hf_llm(ai, repo_id, debug_mode, context_window):
             return None
     try:
         from llama_cpp import Llama
-    except:
+    except Exception:
         if debug_mode:
             traceback.print_exc()
         # Ask for confirmation to install the required pip package
@@ -347,7 +358,73 @@ def get_hf_llm(ai, repo_id, debug_mode, context_window):
         json.dump(usermodels, fd)
         fd.close()
         print("Saved")
-    return llama_cpp.Llama(model_path=model_path, n_gpu_layers=gpulayers(ai), verbose=debug_mode, n_ctx=context_window)
+    return get_llama_inst(repo_id, model_path=model_path, n_gpu_layers=gpulayers(ai), verbose=debug_mode, n_ctx=context_window)
+
+def list_downloaded_models():
+    try:
+        fd = open(r2ai_model_json)
+        usermodels = json.load(fd)
+        fd.close()
+        for m in usermodels:
+            if m.find("/") != -1:
+                M = m.ljust(40)
+                try:
+                    size = round(os.path.getsize(usermodels[m]) / 1024 / 1024 / 1024, 2)
+                    size = f"{size}".rjust(6)
+                    print(f" {size}G {M}")
+                except Exception as e:
+                    print(f"     ??? {M}")
+        user_data_dir = appdirs.user_data_dir("r2ai")
+        models_path = os.path.join(user_data_dir, "models")
+        home_dir = os.path.expanduser("~")
+        sympath = os.path.join(home_dir, ".r2ai.models")
+        files = os.listdir(models_path)
+        for file in files:
+            if file.endswith(".gguf") == False:
+                continue
+            found = False
+            for m in usermodels:
+                last_part = usermodels[m].split("/")[-1]
+                if last_part == file:
+                    found = True
+            if not found:
+                M = f"{sympath}/{file}"
+                size = round(os.path.getsize(M) / 1024 / 1024 / 1024, 2)
+                size = f"{size}".rjust(6)
+                print(f" {size}G {M}")
+    except Exception as e:
+        print(e)
+    return None
+
+def delete_downloaded_model(name):
+    fd = None
+    try:
+        name = name.strip()
+        if not os.path.exists(name):
+            print("Cannot find " + name)
+            return None
+        if name.startswith("/"):
+            if name.endswith(".gguf"):
+                os.remove(name)
+                return None
+            print("Cant delete non-gguf files")
+            return None
+        fd = open(r2ai_model_json)
+        usermodels = json.load(fd)
+        fd.close()
+        fd = None
+        if name in usermodels:
+            modelpath = usermodels[name]
+            os.remove(modelpath)
+            del usermodels[name]
+            fd = open(r2ai_model_json, "w")
+            json.dump(usermodels, fd)
+            fd.close()
+    except Exception as e:
+        print(e)
+    if fd is not None:
+        fd.close()
+    return None
 
 def set_default_model(repo_id):
     usermodels = {"default": repo_id}
@@ -355,7 +432,7 @@ def set_default_model(repo_id):
         fd = open(r2ai_model_json)
         usermodels = json.load(fd)
         fd.close()
-    except:
+    except Exception:
         pass
     usermodels["default"] = repo_id
 #    usermodels[repo_id] = model_path
@@ -383,6 +460,7 @@ def list_gguf_files(repo_id: str) -> List[Dict[str, Union[str, float]]]:
     """
 
     try:
+      from huggingface_hub import HfApi
       api = HfApi()
       tree = list(api.list_repo_tree(repo_id))
       files_info = [file for file in tree if file.path.endswith('.gguf')]
@@ -475,7 +553,8 @@ def enough_disk_space(size, path) -> bool:
 
     return False
 
-def new_get_hf_llm(ai, repo_id, debug_mode, context_window):
+def new_get_hf_llm(ai, repo_id, context_window):
+    debug_mode = True if ai.env["debug_level"] == "1" else False
     if repo_id.startswith("openai:") or repo_id.startswith("anthropic:") or repo_id.startswith("groq:") or repo_id.startswith("google:"):
         return repo_id
     if not os.path.exists(repo_id):
@@ -489,7 +568,7 @@ def new_get_hf_llm(ai, repo_id, debug_mode, context_window):
     model_path = repo_id
     try:
         from llama_cpp import Llama
-    except:
+    except Exception:
         if debug_mode:
             traceback.print_exc()
         # Ask for confirmation to install the required pip package
@@ -546,4 +625,15 @@ def new_get_hf_llm(ai, repo_id, debug_mode, context_window):
     # Initialize and return Code-Llama
     if not os.path.isfile(model_path):
         print("Model is not a file")
-    return llama_cpp.Llama(model_path=model_path, n_gpu_layers=gpulayers(ai), verbose=debug_mode, n_ctx=context_window)
+    return get_llama_inst(repo_id, model_path=model_path, n_gpu_layers=gpulayers(ai), verbose=debug_mode, n_ctx=context_window)
+
+
+def get_llama_inst(repo_id, **kwargs):
+    import llama_cpp
+    if 'functionary' in repo_id:
+        from llama_cpp.llama_tokenizer import LlamaHFTokenizer
+        kwargs['tokenizer'] = LlamaHFTokenizer.from_pretrained(repo_id)
+        filename = os.path.basename(kwargs.pop('model_path'))
+        kwargs['echo'] = True
+        return llama_cpp.Llama.from_pretrained(**kwargs, repo_id=repo_id, filename=filename)
+    return llama_cpp.Llama(**kwargs)

Attic/py/r2ai/partial_json_parser.py

@@ -0,0 +1,223 @@
+class IncompleteJSONParser:
+    def __init__(self, s):
+        self.s = s
+        self.len = len(s)
+        self.pos = 0
+
+    def parse(self):
+        self.skip_whitespace()
+        if self.pos >= self.len:
+            return None
+        char = self.s[self.pos]
+        if char == '{':
+            return self.parse_object()
+        elif char == '[':
+            return self.parse_array()
+        else:
+            return None  # Top-level must be object or array
+
+    def parse_object(self):
+        if self.s[self.pos] != '{':
+            return None
+        self.pos += 1  # Skip '{'
+        self.skip_whitespace()
+        obj = {}
+        parsed_any = False  # Flag to check if any key-value pair was parsed
+
+        while self.pos < self.len:
+            self.skip_whitespace()
+            if self.pos >= self.len:
+                break
+            if self.s[self.pos] == '}':
+                self.pos += 1
+                return obj if parsed_any else None
+            key = self.parse_string()
+            if key is None:
+                # Incomplete key, skip the rest
+                break
+            self.skip_whitespace()
+            if self.pos >= self.len or self.s[self.pos] != ':':
+                # Missing colon, skip this key
+                break
+            self.pos += 1  # Skip ':'
+            self.skip_whitespace()
+            value = self.parse_value()
+            if value is None:
+                # Incomplete value, skip this key-value pair
+                break
+            obj[key] = value
+            parsed_any = True
+            self.skip_whitespace()
+            if self.pos >= self.len:
+                break
+            if self.s[self.pos] == ',':
+                self.pos += 1
+                continue
+            elif self.s[self.pos] == '}':
+                self.pos += 1
+                return obj if parsed_any else None
+            else:
+                # Unexpected character, skip
+                break
+        # Auto-close the object
+        return obj if parsed_any else None
+
+    def parse_array(self):
+        if self.s[self.pos] != '[':
+            return None
+        self.pos += 1  # Skip '['
+        self.skip_whitespace()
+        array = []
+        parsed_any = False  # Flag to check if any element was parsed
+
+        while self.pos < self.len:
+            self.skip_whitespace()
+            if self.pos >= self.len:
+                break
+            if self.s[self.pos] == ']':
+                self.pos += 1
+                return array if parsed_any else None
+            value = self.parse_value()
+            if value is None:
+                # Incomplete value, skip this element
+                break
+            array.append(value)
+            parsed_any = True
+            self.skip_whitespace()
+            if self.pos >= self.len:
+                break
+            if self.s[self.pos] == ',':
+                self.pos += 1
+                continue
+            elif self.s[self.pos] == ']':
+                self.pos += 1
+                return array if parsed_any else None
+            else:
+                # Unexpected character, skip
+                break
+        # Auto-close the array
+        return array if parsed_any else None
+
+    def parse_value(self):
+        self.skip_whitespace()
+        if self.pos >= self.len:
+            return None
+        char = self.s[self.pos]
+        if char == '"':
+            return self.parse_string()
+        elif char == '{':
+            return self.parse_object()
+        elif char == '[':
+            return self.parse_array()
+        elif char in '-0123456789':
+            return self.parse_number()
+        elif self.s.startswith('true', self.pos):
+            self.pos += 4
+            return True
+        elif self.s.startswith('false', self.pos):
+            self.pos += 5
+            return False
+        elif self.s.startswith('null', self.pos):
+            self.pos += 4
+            return None
+        else:
+            return None  # Invalid value
+
+    def parse_string(self):
+        if self.s[self.pos] != '"':
+            return None
+        self.pos += 1  # Skip opening quote
+        result = ""
+        while self.pos < self.len:
+            char = self.s[self.pos]
+            if char == '\\':
+                if self.pos + 1 >= self.len:
+                    result += '\\'
+                    self.pos += 1
+                    break  # Incomplete escape, return what we have
+                self.pos += 1
+                escape_char = self.s[self.pos]
+                if escape_char == '"':
+                    result += '"'
+                elif escape_char == '\\':
+                    result += '\\'
+                elif escape_char == '/':
+                    result += '/'
+                elif escape_char == 'b':
+                    result += '\b'
+                elif escape_char == 'f':
+                    result += '\f'
+                elif escape_char == 'n':
+                    result += '\n'
+                elif escape_char == 'r':
+                    result += '\r'
+                elif escape_char == 't':
+                    result += '\t'
+                elif escape_char == 'u':
+                    # Unicode escape
+                    if self.pos + 4 >= self.len:
+                        # Incomplete unicode escape
+                        break
+                    hex_digits = self.s[self.pos+1:self.pos+5]
+                    try:
+                        code_point = int(hex_digits, 16)
+                        result += chr(code_point)
+                        self.pos += 4
+                    except ValueError:
+                        break  # Invalid unicode escape
+                else:
+                    # Invalid escape character
+                    result += '\\' + escape_char  # Keep it as is
+            elif char == '"':
+                self.pos += 1  # Skip closing quote
+                return result
+            else:
+                result += char
+            self.pos += 1
+        # Return the partial string if incomplete
+        return result
+        
+    def parse_number(self):
+        start = self.pos
+        if self.s[self.pos] == '-':
+            self.pos += 1
+        if self.pos >= self.len:
+            return None
+        if self.s[self.pos] == '0':
+            self.pos += 1
+        elif '1' <= self.s[self.pos] <= '9':
+            while self.pos < self.len and self.s[self.pos].isdigit():
+                self.pos += 1
+        else:
+            return None  # Invalid number
+        if self.pos < self.len and self.s[self.pos] == '.':
+            self.pos += 1
+            if self.pos >= self.len or not self.s[self.pos].isdigit():
+                return None  # Incomplete fraction
+            while self.pos < self.len and self.s[self.pos].isdigit():
+                self.pos += 1
+        if self.pos < self.len and self.s[self.pos] in 'eE':
+            self.pos += 1
+            if self.pos < self.len and self.s[self.pos] in '+-':
+                self.pos += 1
+            if self.pos >= self.len or not self.s[self.pos].isdigit():
+                return None  # Incomplete exponent
+            while self.pos < self.len and self.s[self.pos].isdigit():
+                self.pos += 1
+        num_str = self.s[start:self.pos]
+        try:
+            if '.' in num_str or 'e' in num_str or 'E' in num_str:
+                return float(num_str)
+            else:
+                return int(num_str)
+        except ValueError:
+            return None  # Invalid number
+
+    def skip_whitespace(self):
+        while self.pos < self.len and self.s[self.pos] in ' \t\n\r':
+            self.pos += 1
+
+def parse_incomplete_json(s):
+    parser = IncompleteJSONParser(s)
+    result = parser.parse()
+    return result

Attic/py/r2ai/pipe.py

@@ -1,12 +1,17 @@
 import os
 import traceback
+import r2pipe
+
 
 have_rlang = False
 r2lang = None
 
+r2 = None
+filename = None
 class FakeLang:
     def __init__(self, r2 = None):
         self.r2 = r2
+
     def ai(self, x):
         try:
             from r2ai.repl import r2ai_singleton, runline2
@@ -15,9 +20,10 @@ class FakeLang:
                 print("No global r2ai instance found")
                 return ""
             return runline2(ai, x)
-        except:
+        except Exception:
             traceback.print_exc()
             return None
+
     def cmd(self, x):
         if self.r2 is None:
             return ""
@@ -28,28 +34,42 @@ class FakeLang:
 try:
     import r2lang
     have_rlang = True
-except:
+except Exception:
     import r2pipe
     try:
         if r2pipe.in_r2():
             r2lang = FakeLang(r2pipe.open())
             r2lang.cmd("?V") # r2pipe throws only here
         else:
-            raise Error("must spawn")
-    except:
+            raise Exception("must spawn")
+    except Exception:
         try:
             have_rlang = False
             if os.environ.get('R2AI') is None:
                 ppid = os.getppid()
                 os.environ["R2AI"] = "1"
-                r2lang = FakeLang(r2pipe.open("/bin/ls"))
+                filename = "/bin/ls"
+                r2lang = FakeLang(r2pipe.open(filename))
             else:
                 r2lang = FakeLang(None)
-        except:
+        except Exception:
             print("Cannot instantiate this FakeLang class with r2pipe")
             r2lang = FakeLang(None)
-            pass
 
 def r2singleton():
-    global r2lang
+    global r2lang, r2
+    return r2lang or r2
+
+def get_r2_inst():
+    return r2singleton()
+
+def open_r2(file, flags=[]):
+    global r2, filename, r2lang
+    r2 = r2pipe.open(file, flags=flags)
+    r2lang = FakeLang(r2)
+    filename = file
     return r2lang
+
+def get_filename():
+    global filename
+    return filename

Attic/py/r2ai/plugin.py

@@ -0,0 +1,47 @@
+"""Entrypoint for the r2ai plugin and repl."""
+import sys
+import os
+import builtins
+import traceback
+
+current_dir = os.path.dirname(os.path.realpath(__file__))
+parent_dir = os.path.dirname(current_dir)
+sys.path.insert(0, parent_dir)
+os.environ["TRANSFORMERS_NO_ADVISORY_WARNINGS"] = "1"
+try:
+    venv_dir = os.path.join(parent_dir, 'venv')
+    if os.path.exists(venv_dir):
+        site_packages = os.path.join(venv_dir, 'lib', 'python{}.{}'.format(*sys.version_info[:2]), 'site-packages')
+        if os.path.exists(site_packages):
+            sys.path.insert(0, site_packages)
+except Exception:
+    pass
+
+import r2lang
+from r2ai.main import r2ai_singleton, run_rcfile_once, runline, help_message
+
+def r2ai_rlang_plugin(unused_but_required_argument):
+    ai = r2ai_singleton()
+    def _call(s):
+        if not s.startswith("r2ai"):
+            return False
+        try:
+            run_rcfile_once(ai)
+            if len(s) == 4:
+                builtins.print(help_message)
+            else:
+                usertext = s[4:].strip()
+                runline(ai, usertext)
+        except Exception as e:
+            builtins.print(e)
+            traceback.print_exc()
+        return True
+
+    return {
+        "name": "r2ai",
+        "license": "MIT",
+        "desc": "run llama language models inside r2",
+        "call": _call,
+    }
+
+r2lang.plugin("core", r2ai_rlang_plugin)

Attic/py/r2ai/progress.py

@@ -0,0 +1,100 @@
+from rich.progress import Progress, SpinnerColumn, TimeElapsedColumn, Task
+from rich.console import Console
+from .web import server_running, server_in_background
+from inspect import signature
+from functools import wraps
+
+
+def _support_total(sig, *args, **kwargs):
+    bound = sig.bind(*args, **kwargs)
+    bound.apply_defaults()
+    return "__progress_cls" in bound.arguments and "__progress_task" in bound.arguments
+
+
+def progress_bar(text, color=None, total=None, infinite=False):
+    def decorator(func):
+        @wraps(func)
+        def wrapper(*args, **kwargs):
+            sig = signature(func)
+            has_total = total is not None and _support_total(
+                sig, *args, **kwargs)
+            is_infinite = infinite or not has_total
+
+#            if server_running() and not server_in_background():
+            return func(*args, **kwargs)
+
+            with Progress(SpinnerColumn(), *Progress.get_default_columns(), console=Console(no_color=not bool(color)), transient=False) as p:
+                task_text = f"[{color}]{text}" if color else text
+                task = p.add_task(
+                    task_text, total=None if is_infinite else total)
+
+                if has_total:
+                    result = func(
+                        *args,
+                        **kwargs,
+                        __progress_cls=p,
+                        __progress_task=task)
+                else:
+                    result = func(*args, **kwargs)
+
+            return result
+        return wrapper
+    return decorator
+
+# For consistency with the above
+class ProgressBar:
+    def __init__(self, text, color=None, total=None, infinite=False) -> None:
+        self.text = text
+        self.color = color
+        self.total = total
+        self.infinite = infinite
+        self.progress: Progress = None
+        self.task: Task = None
+
+    def __enter__(self):
+        self.progress = Progress(
+            SpinnerColumn(),
+            *Progress.get_default_columns(),
+            console=Console(
+                no_color=not bool(
+                    self.color)),
+            transient=False)
+        if self.color:
+            self.task = self.progress.add_task(
+                f"[{self.color}]{self.text}", total=None if self.infinite else self.total)
+        else:
+            self.task = self.progress.add_task(
+                f"{self.text}", total=None if self.infinite else self.total)
+        self.progress.start()
+        return self
+
+    def __exit__(self, exc_type, exc_val, exc_tb):
+        if self.progress:
+            self.progress.stop()
+
+# from .progress import ProgressBar, progress_bar
+# progress_bar is used as an decorator
+# @progress_bar(Title, color="yellow")
+# def foo():
+#    bar
+#
+#
+# unlike in the class, the decorated functin can only use progressive progress only if
+#   __progress_cls and __progress_task are used as positional arguments. else it defaults to infinite
+# @progress_bar("Title", color="yellow", total=100)
+# def foo(a,b, __progress_cls=None, __progress_task=None):
+#   i = 1
+#   while True:
+        # progress_cls.update(p.task, advance=i)
+        # i+=1
+        # time.sleep(1)
+#
+#
+# ProgressBar is made for consistency with the decorator
+# import time
+# with ProgressBar("Title", color="Yellow", total=50) as p:
+#   i = 0
+#   while True:
+        # p.progress.update(p.task, advance=i)
+        # i+=1
+        # time.sleep(1)

Attic/py/r2ai/repl.py

@@ -1,6 +1,7 @@
 import builtins
-from .models import set_default_model, models, mainmodels
+from .models import set_default_model, list_downloaded_models, delete_downloaded_model, models, mainmodels
 from .utils import slurp
+from .large import Large
 import traceback
 from .const import R2AI_HISTFILE, R2AI_HOMEDIR, R2AI_RCFILE, R2AI_USERDIR
 from .web import start_http_server
@@ -9,36 +10,31 @@ from datetime import datetime
 import sys
 import os
 
-from .tab import tab_init, tab_hist, tab_write, tab_evals
-
+from .tab import tab_init, tab_hist, tab_write, tab_evals, tab_list
+from .interpreter import Interpreter
+from .pipe import have_rlang, r2lang, r2singleton
+from r2ai import bubble, LOGGER
+from .test import run_test
 tab_init()
 
 print_buffer = ""
 ais = []
 autoai = None
-from .pipe import have_rlang, r2lang, r2singleton
 r2 = r2singleton()
 
 def r2ai_singleton():
     global ais
     if len(ais) == 0:
-        from .interpreter import Interpreter
         ai = Interpreter()
         ais.append(R2AI(ai))
     return ais[0].ai
 
 def r2_cmd(x):
-    global have_rlang, ai, r2, r2_file
-    have_rlang = True
-    res = x
-    if have_rlang:
-        oc = r2lang.cmd('e scr.color').strip()
-        r2lang.cmd('e scr.color=0')
-        res = r2lang.cmd(x)
-        r2lang.cmd('e scr.color=' + oc)
-    elif r2 is not None:
+    r2 = r2singleton()
+    res = None
+    if r2 is not None:
         oc = r2.cmd('e scr.color').strip()
-        r2.cmd('e scr.color=0')
+        r2.cmd('e scr.color=3')
         res = r2.cmd(x)
         r2.cmd('e scr.color=' + oc)
     return res
@@ -51,6 +47,7 @@ help_message = """Usage: r2ai [-option] ([query] | [script.py])
  r2ai ?e [msg]          echo a message
  r2ai ?t [query]        run an query and show it's timing
  r2ai !ls               run a system command
+ r2ai !                 show prompt history
  r2ai -a                query with audio voice
  r2ai -A                enter the voice chat loop
  r2ai -k                clear the screen
@@ -61,21 +58,26 @@ help_message = """Usage: r2ai [-option] ([query] | [script.py])
  r2ai -h                show this help (same as ?)
  r2ai -H ([var])        show path variables like it's done in r2 -H
  r2ai -i [file] ([q])   load the file contents and prompt it with the given optional query
+ r2ai -j [query]        convert the user prompt into a meaningful json
  r2ai -m [file/repo]    select model from huggingface repository or local file
+ r2ai -m-[repo/model]   delete a local downloaded model (see -mm for listing them)
+ r2ai -mm               list all downloaded models
  r2ai -M                shorter list of models
  r2ai -MM               list supported and most common models from hf
  r2ai -n [num]          select the nth language model
  r2ai -q                quit/exit/^C
  r2ai -L                show chat logs (See -Lj for json)
  r2ai -L-[N]            delete the last (or N last messages from the chat history)
+ r2ai -p1c0             enter the Pico le Croco assistant mode
  r2ai -repl             enter the repl (only when running via r2pipe)
  r2ai -r [sysprompt]    define the role of the conversation
  r2ai -r2               enter the r2clippy assistant mode
  r2ai -rf [doc/role/.f] load contents of a file to define the role
  r2ai -R                reset the chat conversation context
- r2ai -t [temp]         from 0.0001 to 10 your scale to randomness in my replies
  r2ai -v                show r2ai version (same as ?V)
  r2ai -w ([port])       start webserver (curl -D hello http://localhost:8000)
+ r2ai -W ([port])       start webserver in background
+ r2ai -VV               visual mode
  r2ai -W ([port])       start webserver in background"""
 
 def myprint(msg, file=None):
@@ -113,11 +115,10 @@ def runplugin(ai, arg):
           for file in files:
               if file.endswith(".py"):
                   print(file.replace(".py", ""))
-    except:
+    except Exception:
       pass
 
 def r2ai_version():
-    import sys
     import llama_cpp
     print("python: " + sys.version)
     print("llama: " + llama_cpp.__version__)
@@ -147,14 +148,41 @@ def run_script(ai, script):
         for line in lines.split("\n"):
             if line.strip() != "":
                 runline(ai, line)
-    except:
+    except Exception:
         pass
 
 class R2AI:
     def __init__(self,ai):
         self.ai = ai
-    def cmd(x):
-        return runline2(self.ai, cmd)
+
+    def cmd(self, x):
+        return runline2(self.ai, x)
+
+def slurp_until(endword):
+    text = ""
+    while True:
+      line = sys.stdin.readline()
+      if line.strip() == endword:
+          break
+      text += line
+    return text
+
+def list_models():
+    print("ollama/llama3:latest")
+    print("ollama/granite-code:8b")
+    print("anthropic/claude-sonnet-5-5")
+    print("openai/gpt-4o")
+    print("mistral/mistral-tiny")
+
+def set_model(model):
+    ai = ais[0].ai
+    if ai.model != model:
+      ai.llama_instance = None
+    ai.model = model
+    if autoai:
+        autoai.model = model
+    ai.env["llm.model"] = ai.model
+    set_default_model(ai.model)
 
 def runline(ai, usertext):
     global print
@@ -177,10 +205,17 @@ def runline(ai, usertext):
     if usertext.startswith("-H"):
         try:
             return r2ai_vars(ai, usertext[2:].strip())
-        except:
+        except Exception:
             traceback.print_exc()
+    if usertext.startswith("-VV"):
+        from ui.app import R2AIApp # pylint: disable=import-error
+        R2AIApp(ai=ai).run()
+        return
     if usertext.startswith("?V") or usertext.startswith("-v"):
         r2ai_version()
+    elif usertext.startswith("<<"):
+        newline = slurp_until(usertext[2:])
+        return runline(ai, newline)
     elif usertext.startswith("?e"):
         print(usertext[2:].strip())
     elif usertext.startswith("?t"):
@@ -188,29 +223,40 @@ def runline(ai, usertext):
         runline(ai, usertext[2:].strip())
         tend = datetime.now()
         print(tend - tstart)
-    elif usertext.startswith("?") or usertext.startswith("-h"):
+    elif usertext.startswith("?") or usertext.startswith("-h") or usertext.startswith("-?"):
         print(help_message)
     elif usertext.startswith("clear") or usertext.startswith("-k"):
         print("\x1b[2J\x1b[0;0H\r")
+        if ai.messages:
+            ai.messages = []
+        if autoai and autoai.messages:
+            autoai.messages = []
     elif usertext.startswith("-MM"):
         print(models().strip())
     elif usertext.startswith("-M"):
         print(mainmodels().strip())
+    elif usertext.startswith("-mm"):
+        list_downloaded_models()
+    elif usertext.startswith("-m-"):
+        delete_downloaded_model(usertext[3:])
+    elif usertext.startswith("r2ai"):
+        print("ERROR: Cant run r2ai from r2ai, you must install the C plugin and run this command from radare2")
+    elif usertext == "-m?":
+        list_models()
     elif usertext.startswith("-m"):
         words = usertext.split(" ")
         if len(words) > 1:
-            ai.model = words[1]
-            ai.env["llm.model"] = ai.model
-            set_default_model(ai.model)
+            set_model(words[1])
         else:
             print(ai.model)
     elif usertext == "reset" or usertext.startswith("-R"):
         ai.reset()
-    elif usertext.startswith("-t"):
-        if usertext == "-t":
-            print(ai.env["llm.temperature"])
-        else:
-            ai.env["llm.temperature"] = usertext[2:].strip()
+        if autoai:
+            autoai.reset()
+    elif usertext.startswith("-j"):
+        q = usertext[2:].strip()
+        query = f"Please respond using ONLY in JSON with the following fields if relevant: topic, array of key words, location, url, name, description, target, amounts and other details if necessary without providing any response. Question: \"{q}\""
+        ai.chat(query)
     elif usertext == "-A":
         from .voice import stt
         ai.env["chat.voice"] = "true"
@@ -246,6 +292,13 @@ def runline(ai, usertext):
         ai.env["data.path"] = f"{R2AI_HOMEDIR}/doc/"
         ai.env["chat.bubble"] = "true"
         runline(ai, f"-rf {R2AI_HOMEDIR}/doc/role/r2clippy.txt")
+    elif usertext == "-p1c0":
+        ai.env["data.use"] = "true"
+        ai.env["data.hist"] = "true"
+        ai.env["data.path"] = f"{R2AI_HOMEDIR}/doc/"
+        ai.env["chat.bubble"] = "true"
+        ai.env["chat.picolecroco"] = "true"
+        runline(ai, f"-rf {R2AI_HOMEDIR}/doc/role/picolecroco.txt")
     elif usertext.startswith("-ed"):
         editor = "vim" # Defaults to the only real editor
         if ai.env["user.editor"] != "":
@@ -283,28 +336,36 @@ def runline(ai, usertext):
             else:
                 try:
                     print(ai.env[k])
-                except:
+                except Exception:
                     print("Invalid config key", file=sys.stderr)
                     pass
     elif usertext.startswith("-l"):
         try:
-            l = Large()
+            l = Large(ai)
             t = slurp("doc/samples/qcw.txt")
             print(l.summarize_text(t))
-        except:
+        except Exception:
             traceback.print_exc()
         sys.exit(0)
     elif usertext.startswith("-W"):
         if len(usertext) > 2:
-            port = int(usertext[2:])
-            if port > 0:
-                runline2(ai, f"-e http.port={port}")
+            try:
+                port = int(usertext[2:])
+                if port > 0:
+                    runline2(ai, f"-e http.port={port}")
+            except:
+                print("Use -w [port-number]")
+                return
         start_http_server(ai, runline2, True)
     elif usertext.startswith("-w"):
         if len(usertext) > 2:
-            port = int(usertext[2:])
-            if port > 0:
-                runline2(ai, f"-e http.port={port}")
+            try:
+                port = int(usertext[2:])
+                if port > 0:
+                    runline2(ai, f"-e http.port={port}")
+            except:
+                print("Use -w [port-number]")
+                return
         start_http_server(ai, runline2, False)
     elif usertext.startswith("-s"):
         ai.runline2 = runline2
@@ -314,7 +375,7 @@ def runline(ai, usertext):
             fname = usertext[3:].strip()
             try:
                 ai.system_message = slurp(fname)
-            except:
+            except Exception:
                 print(f"Cannot open file {fname}", file=sys.stderr)
         else:
             print(ai.system_message)
@@ -328,13 +389,13 @@ def runline(ai, usertext):
             amount = int(usertext[3:])
             if amount < 1:
                 amount = 1
-        except:
+        except Exception:
             amount = 1
         try:
             for i in range(amount):
                 ai.messages.pop() # delete user message
                 ai.messages.pop() # delete assistant message
-        except:
+        except Exception:
             pass
     elif usertext.startswith("-Lj"):
         print(ai.messages)
@@ -347,7 +408,7 @@ def runline(ai, usertext):
         try:
             res = slurp(text)
             ai.chat(res)
-        except:
+        except Exception:
             print("Cannot load file", file=sys.stderr)
     elif usertext.startswith("-i"):
         text = usertext[2:].strip()
@@ -378,7 +439,6 @@ def runline(ai, usertext):
             if index < len(ais):
                 ai = ais[index].ai
             else:
-                from .interpreter import Interpreter
                 ai0 = Interpreter()
                 ai0.model = ai.model
                 ais.append(R2AI(ai0))
@@ -391,13 +451,33 @@ def runline(ai, usertext):
                 que = words[1]
             else:
                 que = input("[Query]> ")
-        except:
+        except Exception:
             print("")
             return
+
         tag = "```\n" # TEXT, INPUT ..
         ai.chat(f"{que}:\n{tag}\n{res}\n{tag}\n")
     elif usertext[0] == "!":
-        os.system(usertext[1:])
+        import shutil
+        columns, rows = shutil.get_terminal_size()
+        if len(usertext) == 1:
+            count = 0
+            for item in tab_list():
+                print(item, file=sys.stderr)
+                count = count + 1
+                if count == rows - 1:
+                    print("-- More -- (q)")
+                    count = 0
+                    if input() == 'q':
+                        break
+        else:
+            safe_int = lambda x: int(x) if str(x).isdigit() else 0
+            n = safe_int(usertext[1:])
+            if n > 0:
+                items = tab_list()
+                runline(ai, items[n - 1])
+            else:
+                os.system(usertext[1:])
     elif usertext[0] == ".":
         #if len(usertext) > 1 and usertext[1] == ".": # ".." - run user plugins
         #    runplugin(ai, usertext[2:].strip())
@@ -410,29 +490,32 @@ def runline(ai, usertext):
             else:
                 for line in file.split("\n"):
                     runline(ai, line)
-        except Exception as e:
+        except Exception:
             traceback.print_exc()
-            pass
+
     elif usertext.startswith("' "):
         if not autoai:
-            from .interpreter import Interpreter
             autoai = Interpreter()
-            autoai.auto_run = True
+            autoai.env = ai.env
+            autoai.auto_run = True 
+
         autoai.chat(usertext[2:])
+
     elif usertext[0] == ":":
         if r2 is None:
             print("r2 is not available", file=sys.stderr)
         else:
             builtins.print(r2_cmd(usertext[1:]))
+    elif usertext.startswith("--test"):
+        run_test(usertext[7:])
     elif usertext.startswith("-"):
-        print("Unknown flag. See 'r2ai -h' for help", file=sys.stderr)
+        print(f"Unknown flag '{usertext}'. See 'r2ai -h' for help", file=sys.stderr)
     else:
         ai.chat(usertext)
 
 def r2ai_repl(ai):
-    from r2ai import bubble
     tab_evals(ai.env.keys())
-    oldoff = r2_cmd("s").strip()
+    oldoff = r2_cmd("?vx $$").strip()
     if oldoff == "0x0" or oldoff == "":
         oldoff = "0x00000000"
     olivemode = ai.env["chat.live"]
@@ -442,9 +525,9 @@ def r2ai_repl(ai):
         if os.name != "nt":
             prompt = f"\001\x1b[33m\002{prompt}" # \001\x1b[0m\002"
         if r2 is not None:
-            off = r2_cmd("s").strip()
+            off = r2_cmd("?vx $$").strip()
             if off == "":
-                off = r2_cmd("s").strip()
+                off = r2_cmd("?vx $$").strip()
             if len(off) > 5 and len(off) < 20:
                 oldoff = off
         if ai.active_block is not None:
@@ -456,7 +539,7 @@ def r2ai_repl(ai):
                 builtins.print("\001\x1b[0m\002", end="")
         except EOFError:
             break
-        except:
+        except Exception:
             traceback.print_exc()
             break
         try:
@@ -466,14 +549,17 @@ def r2ai_repl(ai):
                         break
                 else:
                     bubble.query(usertext)
-                    bubble.response_begin()
+                    if ai.env["chat.picolecroco"] == "true":
+                        bubble.pico_begin()
+                    else:
+                        bubble.response_begin()
                     if runline(ai, usertext) == "q":
                         break
                     bubble.response_end()
             else:
                 if runline(ai, usertext) == "q":
                     break
-        except:
+        except Exception:
             traceback.print_exc()
             continue
         tab_write()

Attic/py/r2ai/spinner.py

@@ -0,0 +1,42 @@
+import itertools
+import threading
+import time
+import sys
+
+class Spinner:
+    def __init__(self, message="Loading...", delay=0.1):
+        self.spinner = itertools.cycle([
+            "\033[1;31m⠋\033[0m", "\033[1;32m⠙\033[0m", "\033[1;33m⠹\033[0m", "\033[1;34m⠸\033[0m",
+            "\033[1;35m⠼\033[0m", "\033[1;36m⠴\033[0m", "\033[1;37m⠦\033[0m", "\033[1;31m⠧\033[0m",
+            "\033[1;32m⠇\033[0m", "\033[1;33m⠏\033[0m"
+        ])
+        self.message = message
+        self.delay = delay
+        self.running = False
+        self.thread = None
+        self.start_time = None
+
+    def start(self):
+        """Start the spinner in a separate thread."""
+        self.running = True
+        self.start_time = time.time()
+        self.thread = threading.Thread(target=self._spin)
+        self.thread.start()
+
+    def _spin(self):
+        """Spin the spinner while running is True."""
+        while self.running:
+            elapsed_time = time.time() - self.start_time
+            sys.stdout.write(f"\r{self.message} {next(self.spinner)} {elapsed_time:.1f}s")
+            sys.stdout.flush()
+            time.sleep(self.delay)
+        sys.stdout.write('\r' + ' ' * (len(self.message) + 20) + '\r')  # Clear the line
+        sys.stdout.flush()
+
+    def stop(self):
+        """Stop the spinner."""
+        self.running = False
+        if self.thread is not None:
+            self.thread.join()
+
+spinner = Spinner("")

Attic/py/r2ai/tab.py

@@ -1,4 +1,5 @@
 import os
+import sys
 from .const import R2AI_HISTFILE, R2AI_HOMEDIR, R2AI_RCFILE, R2AI_USERDIR
 from .models import models
 
@@ -67,7 +68,7 @@ class MyCompleter(object):  # Custom completer
 
     def display_matches(self, substitution, matches, longest_match_length):
         line_buffer = readline.get_line_buffer()
-        columns = environ.get("COLUMNS", 80)
+        columns = os.environ.get("COLUMNS", 80)
         print()
         tpl = "{:<" + str(int(max(map(len, matches)) * 1.2)) + "}"
         buffer = ""
@@ -99,11 +100,11 @@ try:
     import readline
     import rlcompleter
     have_readline = True
-except:
+except Exception:
     have_readline = True
     pass # readline not available
 
-def tab_hist(text):
+def tab_hist():
     if not have_readline:
         print("Cannot find readline", file=sys.stderr)
         return False
@@ -118,6 +119,17 @@ def tab_write():
         return False
     readline.write_history_file(R2AI_HISTFILE)
 
+def tab_list():
+    global readline
+    if not have_readline:
+        return []
+    amount = readline.get_current_history_length()
+    res = []
+    for i in range(1, amount):
+        item = readline.get_history_item(i)
+        res.append(f"{i}  {item}")
+    return res
+
 def tab_init():
     if not have_readline:
         print("Cannot find readline", file=sys.stderr)
@@ -127,9 +139,15 @@ def tab_init():
         readline.read_history_file(R2AI_HISTFILE)
     except FileNotFoundError:
         pass
-    readline.set_completer(completer.complete)
-    readline.set_completer_delims('\t\n;')
-    readline.set_completion_display_matches_hook(completer.display_matches)
-    #readline.parse_and_bind("bind -e")
-    readline.parse_and_bind('tab: complete') # linux tab (gnu readline)
-    readline.parse_and_bind("bind ^I rl_complete") ## macos tab (libedit)
+    except Exception:
+        pass
+    try:
+        readline.set_completer(completer.complete)
+        readline.set_completer_delims('\t\n;')
+        readline.set_completion_display_matches_hook(completer.display_matches)
+        if readline.__doc__.find("GNU") != -1:
+            readline.parse_and_bind('tab: complete')
+        else:
+            readline.parse_and_bind("bind ^I rl_complete")
+    except Exception as e:
+        pass

Attic/py/r2ai/test.py

@@ -0,0 +1,37 @@
+import builtins
+from .tools import run_python, execute_binary, r2cmd
+import subprocess
+from .pipe import get_filename
+import time
+py_code = """
+print('hello test')
+"""
+
+def run_test(args):
+    if not args or len(args) == 0:
+        res = run_python(py_code).strip()
+        print(f"run_python: {res}", len(res))
+        assert res == "hello test"
+        print("run_python: test passed")
+        r2cmd("o--;o /bin/ls")
+        res = execute_binary(args=["-d", "/etc"]).strip()
+        subp = subprocess.run(["/bin/ls", "-d", "/etc"], capture_output=True, text=True)
+        print("exec result", res)
+        print("subp result", subp.stdout)
+        assert ''.join(res).strip() == subp.stdout.strip()
+        print("execute_binary with args: test passed")
+    else:
+        cmd, *args = args.split(" ", 1)
+        if cmd == "get_filename":
+            builtins.print(get_filename())
+        elif cmd == "run_python":
+            builtins.print(f"--- args ---")
+            builtins.print(args)
+            builtins.print(f"--- end args ---")
+            builtins.print(f"--- result ---")
+            builtins.print(run_python(args[0]))
+            builtins.print(f"--- end result ---")
+        elif cmd == "r2cmd":
+            builtins.print(f"--- {args} ---")
+            builtins.print(r2cmd(args))
+            builtins.print("--- end ---")

Attic/py/r2ai/tools.py

@@ -0,0 +1,176 @@
+from r2ai.pipe import get_r2_inst
+import json
+import builtins
+import base64
+from .pipe import get_filename
+from . import LOGGER
+import time
+import sys
+from io import StringIO
+import subprocess
+import os
+is_plugin = False
+try:
+    import r2lang
+    is_plugin = True
+except Exception:
+    is_plugin = False
+    pass
+
+def r2cmd(command: str):
+    """
+    Run a r2 command and return the output
+
+    Parameters
+    ----------
+    command: str
+        The r2 command to run
+
+    Returns
+    -------
+    dict
+        The output of the r2 command
+    """
+    r2 = get_r2_inst()
+    r2.cmd('e scr.color=3')
+    if command.startswith('r2 '):
+        return "You are already in r2!"
+    cmd = '{"cmd":' + json.dumps(command) + '}'
+    res = r2.cmd(cmd)
+
+    try:
+        res = json.loads(res)
+        if 'error' in res and res['error'] is True:
+            error_message = res['error']
+            log_messages = '\n'.join(log['message'] for log in res.get('logs', []))
+            # return { 'type': 'error', 'output': log_messages }
+            return log_messages
+        
+        return res['res']
+    except json.JSONDecodeError:
+        if type(res) == str:
+            spl = res.strip().split('\n')
+            if spl[-1].startswith('{"res":""'):
+                res = '\n'.join(spl[:-1])
+        return res
+    except Exception as e:
+        # return { 'type': 'error', 'output': f"Error running r2cmd: {e}\nCommand: {command}\nResponse: {res}" }
+        return f"Error running r2cmd: {e}\nCommand: {command}\nResponse: {res}"
+    
+def run_python(command: str):
+    """
+    Run a python script and return the output
+
+    Parameters
+    ----------
+    command: str
+        The python script to run
+
+    Returns
+    -------
+    str
+        The output of the python script
+    """
+    r2 = get_r2_inst()
+    res = ""
+    is_plugin = False
+    python_path = sys.executable
+    try:
+        proc = subprocess.run([python_path, '-c', command], 
+                            capture_output=True,
+                            text=True)
+        res = proc.stdout
+        if proc.stderr:
+            res += proc.stderr
+    except Exception as e:
+        res = str(e)
+    
+    # if is_plugin:
+    #     base64cmd = base64.b64encode(command.encode('utf-8')).decode('utf-8')
+    #     res += r2cmd(f'#!python -e base64:{base64cmd} > .r2ai_tmp.log')
+    #     res += r2cmd('cat .r2ai_tmp.log')
+    #     r2cmd('rm .r2ai_tmp.log')
+    # else:
+    #     with open('r2ai_tmp.py', 'w') as f:
+    #         f.write(command)
+    #     r2 = get_r2_inst()
+    #     res += r2cmd('#!python r2ai_tmp.py > .r2ai_tmp.log')
+    #     time.sleep(0.1)
+    #     res += r2cmd('!cat .r2ai_tmp.log')
+    #     LOGGER.debug(f'run_python: {res}')
+    #     # r2cmd('rm r2ai_tmp.py')
+    #     # r2cmd('rm .r2ai_tmp.log')
+    return res        
+    
+
+schemas = {
+    "execute_binary": {
+        "name": "execute_binary",
+        "description": "Execute a binary with the given arguments and stdin",
+        "parameters": {
+        "type": "object",
+        "properties": {
+            "args": {
+                "description": "The arguments to pass to the binary. Do not include the file name.",
+                "type": "array",
+                "items": {
+                    "type": "string"
+                }
+            },
+            "stdin": {
+                "type": "string"
+            }
+            }
+        }
+    }
+}
+
+def execute_binary(args: list[str] = [], stdin: str = ""):
+    filename = get_filename()
+    if filename:
+        if os.path.isabs(filename):
+            abs_path = os.path.abspath(filename)
+            if os.path.exists(abs_path):
+                filename = abs_path
+        else:
+            cwd_path = os.path.join(os.getcwd(), filename) 
+            if os.path.exists(cwd_path):
+                filename = cwd_path
+        try:
+            cmd = [filename] + args
+            proc = subprocess.run(cmd, input=stdin, capture_output=True, text=True)
+            res = proc.stdout
+            if proc.stderr:
+                res += proc.stderr
+            return res
+        except Exception as e:
+            return str(e)
+    return ""
+    # r2 = get_r2_inst()
+    # if stdin:
+    #     r2.cmd(f'dor stdin={json.dumps(stdin)}')
+    # if len(args) > 0:
+    #     r2.cmd(f"ood {' '.join(args)}")
+    # else:
+    #     r2.cmd("ood")
+    # res = r2cmd("dc")
+    # return res
+
+
+
+def print_tool_call(msg):
+    if msg['function']['name'] == 'r2cmd':
+        builtins.print('\x1b[1;32m> \x1b[4m' + msg['function']['arguments']['command'] + '\x1b[0m')
+    elif msg['function']['name'] == 'run_python':
+        builtins.print('\x1b[1;32m> \x1b[4m' + "#!python" + '\x1b[0m')
+        builtins.print(msg['function']['arguments']['command'])
+    elif msg['function']['name'] == 'execute_binary':
+        filename = get_filename() or 'bin'
+        stdin = msg['function']['arguments']['stdin'] if 'stdin' in msg['function']['arguments'] else None
+        args = msg['function']['arguments']['args'] if 'args' in msg['function']['arguments'] else []
+        cmd = filename
+        if args and len(args) > 0:
+            cmd += ' ' + ' '.join(args)
+        if stdin and len(stdin) > 0:
+            cmd += f' stdin={stdin}'
+        builtins.print('\x1b[1;32m> \x1b[4m' + cmd + '\x1b[0m')

Attic/py/r2ai/ui/app.py

@@ -0,0 +1,350 @@
+from textual.app import App, ComposeResult, SystemCommand
+from textual.containers import ScrollableContainer, Container, Horizontal, VerticalScroll, Grid, Vertical  # Add Vertical to imports
+from textual.widgets import Header, Footer, Input, Button, Static, DirectoryTree, Label, Tree, Markdown
+from textual.command import CommandPalette, Command, Provider, Hits, Hit
+from textual.screen import Screen, ModalScreen, SystemModalScreen
+from textual.message import Message
+from textual.reactive import reactive
+from .model_select import ModelSelect
+from r2ai.pipe import open_r2, get_filename, r2
+from r2ai.interpreter import is_litellm_model
+from typing import Iterable
+import os
+from pathlib import Path 
+from textual import work
+from textual.widget import Widget
+from textual.css.query import NoMatches
+from textual import log
+from litellm import validate_environment
+from markdown_it import MarkdownIt
+# from ..repl import set_model, r2ai_singleton
+# ai = r2ai_singleton()
+from .chat import chat
+import asyncio
+import json
+import re
+ANSI_REGEX = re.compile(r'\x1B(?:[@-Z\\-_]|\[[0-?]*[ -/]*[@-~])')
+
+
+class ModelConfigDialog(SystemModalScreen):
+    def __init__(self, keys: list[str]) -> None:
+        super().__init__()
+        self.keys = keys
+
+    def compose(self) -> ComposeResult:
+        for key in self.keys:
+            yield Input(placeholder=key, id=f"{key}-input")
+        yield Button("Save", variant="primary", id="save-button")
+
+    def on_button_pressed(self, event: Button.Pressed) -> None:
+        if event.button.id == "save-button":
+            for key in self.keys:
+                os.environ[key] = self.query_one(f"#{key}-input", Input).value
+            self.dismiss()
+class ChatMessage(Widget):
+    markdown: reactive[str] = reactive("")
+    sender = "User"
+    is_done = reactive(False, layout=True, recompose=True)
+
+    def __init__(self, id: str, sender: str, content: str, is_done: bool = False, **kwargs) -> None:
+        super().__init__(id=id, classes='chat-message-container', **kwargs)
+        # self.markdown = f"*{sender}*: {content}"
+        self.markdown = content
+        self.sender = sender
+        self.is_done = is_done
+    async def watch_markdown(self, markdown: str) -> None:
+        try:
+            mkd = self.query_exactly_one(".text1")
+            text = self.markdown
+            if hasattr(mkd, 'update') and callable(mkd.update):
+                update_method = mkd.update
+                if asyncio.iscoroutinefunction(update_method):
+                    await update_method(text)
+                else:
+                    update_method(text)
+        except NoMatches:
+            pass
+
+    def add_text(self, markdown: str) -> None:
+        self.markdown += markdown
+
+    def compose(self) -> ComposeResult:
+
+        if self.sender == "User":
+            yield Static(f"[bold green]{self.sender}", classes='label_sender', markup=True)
+            yield Static(self.markdown, classes='text1', markup=True)
+        elif self.sender == 'AI':
+            yield Static(f"[bold magenta]{self.sender}", markup=True)
+            if self.is_done:
+                yield Markdown(self.markdown, classes='text1', parser_factory=lambda: MarkdownIt('gfm-like'))
+            else:
+                yield Static(self.markdown, classes='text1', markup=True)
+        elif self.sender == 'Tool Call':
+            yield Static(f"[bold blue]{self.sender}", markup=True)
+            yield Static(self.markdown, classes='text1', markup=True)
+        elif self.sender == 'Tool Response':
+            yield Static(f"[bold blue]{self.sender}", markup=True)
+            yield VerticalScroll(Static(self.markdown, classes='text1', markup=True), classes='tool_response_container')
+        
+        
+        
+            
+
+class R2AIApp(App):
+    CSS_PATH = "app.tcss"
+    BINDINGS = [
+        ("ctrl+p", "show_command_palette", "Command Palette"),
+        ("ctrl+m", "select_model", "Select Model"),
+        ("ctrl+o", "load_binary", "Load Binary")
+    ]
+    TITLE = "r2ai"
+    SUB_TITLE = None
+
+    def __init__(self, *args, **kwargs) -> None:
+        super().__init__()
+        if 'ai' in kwargs:
+            self.ai = kwargs['ai']
+        else:
+            class FakeAI:
+                model = 'gpt-4o'
+                messages = []
+            self.ai = FakeAI()
+        self.update_sub_title(get_filename())
+
+    def update_sub_title(self, binary: str = None) -> str:
+        sub_title = None
+        model = self.ai.model
+        if binary and model:
+            binary = Path(binary).name
+            sub_title = f"{model} | {binary}"
+        elif binary:
+            sub_title = binary
+        else:
+            sub_title = model
+        self.sub_title = sub_title
+
+    def compose(self) -> ComposeResult:
+        yield Header()
+        yield Container(
+            VerticalScroll(
+
+                    id="chat-container",
+            ),
+            ScrollableContainer(
+                    Horizontal(
+                        Input(placeholder="Type your message here...", id="chat-input"),
+                        Button("Send", variant="primary", id="send-button"),
+                        id="input-container",
+                    ),
+                    id="input-area",
+            ),
+            id="content",
+        )
+        yield Footer()
+
+    def on_mount(self) -> None:
+        self.install_screen(CommandPalette(), name="command_palette")
+        self.query_one("#chat-input", Input).focus()
+        # self.install_screen(BinarySelectDialog(), name="binary_select_dialog")
+
+    def action_show_command_palette(self) -> None:
+        self.push_screen("command_palette")
+    
+    
+    async def select_model(self) -> None:
+        model = await self.push_screen_wait(ModelSelect())
+        if model:
+            await self.validate_model()
+            self.ai.model = model
+            self.notify(f"Selected model: {self.ai.model}")
+        self.update_sub_title()
+
+    @work
+    async def action_select_model(self) -> None:
+        await self.select_model()
+
+    @work
+    async def action_load_binary(self) -> None:
+        binary = await self.push_screen_wait(BinarySelectDialog())
+        if binary:
+            self.notify(f"Selected binary: {binary}")
+        self.update_sub_title(binary)
+        
+    def get_system_commands(self, screen: Screen) -> Iterable[SystemCommand]:
+        yield from super().get_system_commands(screen)
+        yield SystemCommand("Models", "Select Model", self.action_select_model)
+        yield SystemCommand("Load Binary", "Load Binary", self.action_load_binary)  # Add this command
+
+    def on_button_pressed(self, event: Button.Pressed) -> None:
+        if event.button.id == "send-button":
+            self.send_message()
+
+    async def on_input_submitted(self, event: Input.Submitted) -> None:
+        if event.input.id == "chat-input":
+            await self.send_message()
+
+    def on_message(self, type: str, message: any) -> None:
+        if type == 'message':
+            existing = None
+            try:
+                existing = self.query_one(f"#{message['id']}")
+                if message["content"] is not None:
+                    existing.add_text(message["content"])
+                if message['done']:
+                    existing.is_done = True
+                self.scroll_to_bottom()
+            except NoMatches:
+                existing = self.add_message(message["id"], "AI", message["content"])
+        elif type == 'tool_call':
+            if 'command' in message['function']['arguments']:
+                self.add_message(message["id"], "Tool Call", f"{message['function']['name']} > {message['function']['arguments']['command']}")
+            elif message['function']['name'] == 'execute_binary':
+                args = message['function']['arguments']
+                output = get_filename() or "bin"
+                if 'args' in args and len(args['args']) > 0:
+                    output += f" {' '.join(args['args'])}\n"
+                if 'stdin' in args and len(args['stdin']) > 0:
+                    output += f" stdin={args['stdin']}\n"
+                
+                self.add_message(message["id"], "Tool Call", f"{message['function']['name']} > {output}")
+        elif type == 'tool_response':
+            self.add_message(message["id"], "Tool Response", ANSI_REGEX.sub('', message['content']))
+        
+    async def send_message(self) -> None:
+        input_widget = self.query_one("#chat-input", Input)
+        message = input_widget.value.strip()
+        if message:
+            self.add_message(None, "User", message)
+            input_widget.value = ""
+            try:
+                await self.validate_model()
+                self.chat(message)
+            except Exception as e:
+                self.notify(str(e), severity="error")
+    
+    @work(thread=True)
+    async def chat(self, message: str) -> None:
+        await chat(self.ai, message, lambda type, message: self.call_from_thread(self.on_message, type, message))
+
+    async def validate_model(self) -> None:
+        model = self.ai.model
+        if not model:
+            await self.select_model()
+        if is_litellm_model(model):
+            model = self.ai.model.replace(':', '/')
+            keys = validate_environment(model)
+            if keys['keys_in_environment'] is False:
+                await self.push_screen_wait(ModelConfigDialog(keys['missing_keys']))
+            
+        return True
+
+    def add_message(self, id: str, sender: str, content: str) -> None:
+        chat_container = self.query_one("#chat-container", VerticalScroll)
+        try:
+            msg = ChatMessage(id, sender, content)
+            chat_container.mount(msg)
+            return msg
+        except Exception as e:
+            pass
+        finally:
+            self.scroll_to_bottom()
+        
+
+    def scroll_to_bottom(self) -> None:
+        chat_scroll = self.query_one("#chat-container", VerticalScroll)
+        chat_scroll.scroll_end(animate=False)
+class Message(Widget):
+    def __init__(self, message: str) -> None:
+        super().__init__()
+        self.content = f'[bold]{message.role}[/] {message.content}'
+
+    def render(self) -> str:
+        return Markdown(self.content)
+
+class Messages(Container):
+    def __init__(self, messages) -> None:
+        self.messages = messages
+    def compose(self) -> ComposeResult:
+        for message in self.messages:
+            yield Message(message)
+
+class FilteredDirectoryTree(DirectoryTree):
+    input_path: reactive[Path] = reactive(Path.home())
+
+    def filter_paths(self, paths: Iterable[Path]) -> Iterable[Path]:
+        ps = [path for path in paths if str(path).startswith(str(self.input_path))]
+        return ps
+
+class BinarySelectDialog(SystemModalScreen):
+    BINDINGS = [
+        ("up", "cursor_up", "Move cursor up"),
+        ("down", "cursor_down", "Move cursor down"),
+        ("enter", "select_cursor", "Select item"),
+        ("escape", "app.pop_screen", "Close"),
+        ("backspace", "go_up", "Go up one level"),  # Add this binding
+    ]
+
+    def compose(self) -> ComposeResult:
+        with Vertical():
+            yield Input(placeholder="Enter path here...", id="path-input")
+            yield FilteredDirectoryTree(Path.home(), id="file-browser")
+
+    def on_mount(self) -> None:
+        self.path_input = self.query_one("#path-input", Input)
+        self.file_browser = self.query_one("#file-browser", DirectoryTree)
+        self.path_input.value = str(get_filename() or Path.home())
+        self.path_input.focus()
+        self.watch(self.path_input, "value", self.update_tree)
+
+    @work(thread=True)
+    def update_tree(self) -> None:
+        path = Path(self.path_input.value)
+        
+        if path.exists():
+            self.file_browser.path = str(path)
+        elif path.parent.exists():
+            self.file_browser.path = str(path.parent)
+        self.file_browser.input_path = str(path)
+
+    def on_button_pressed(self, event: Button.Pressed) -> None:
+        if event.button.id == "up-button":
+            self.go_up()
+
+    def action_go_up(self) -> None:
+        current_path = Path(self.file_browser.path)
+        parent_path = current_path.parent
+        if parent_path != current_path:
+            self.file_browser.path = str(parent_path)
+            self.path_input.value = str(parent_path)
+
+    def on_directory_tree_file_selected(self, event: DirectoryTree.FileSelected) -> None:  
+        self.path_input.value = str(event.path)
+        self.open_and_analyze_binary(str(event.path))
+        self.dismiss(str(event.path))
+    
+    @work(thread=True)
+    def open_and_analyze_binary(self, path: str) -> None:
+        global r2
+        r2 = open_r2(path)
+        res = json.loads(r2.cmd('{ "cmd":"aaa" }'))
+        if 'error' in res and res['error'] is True:
+            self.notify(res['error'], severity="error")
+
+    def on_directory_tree_directory_selected(self, event: DirectoryTree.DirectorySelected) -> None:
+        self.path_input.value = str(event.path)
+
+    def action_cursor_up(self) -> None:
+        self.file_browser.action_cursor_up()
+
+    def action_cursor_down(self) -> None:
+        self.file_browser.focus()
+        self.file_browser.action_cursor_down()
+
+    def action_select(self) -> None:
+        node = self.file_browser.cursor_node
+        if hasattr(node.data, 'is_file') and node.data.is_file:
+            self.open_and_analyze_binary(str(node.data.path))
+            self.dismiss(str(node.data.path))
+
+# app = R2AIApp()
+# app.run()

Attic/py/r2ai/ui/app.tcss

@@ -0,0 +1,129 @@
+Placeholder {
+    height: 100%;
+}
+
+#sidemenu {
+  dock: left;
+  width: 15;
+  height: 100;
+}
+Screen {
+    background: $surface;
+    layers: base overlay;
+}
+
+#content {
+    height: 1fr;
+    margin: 1 2;
+}
+
+#chat-scroll {
+    height: 1fr;
+    border: solid $primary;
+    background: $surface-darken-1;
+}
+
+#chat-container {
+    padding: 1;
+    overflow-y: scroll;
+    layout: vertical;
+    layer: base;
+}
+
+#input-area {
+    layer: base;
+    height: auto;
+    margin-top: 1;
+}
+
+#input-container {
+    height: 3;
+    background: $boost;
+    border: solid $primary;
+}
+
+#chat-input {
+    dock: left;
+    width: 1fr;
+    background: $boost;
+    border: none;
+}
+
+#send-button {
+    width: 10;
+    margin-left: 1;
+}
+
+/* ... existing styles ... */
+
+
+#binary-select-dialog Label {
+    padding: 1 2;
+}
+
+#binary-select-dialog #path-input {
+    margin: 0 2 1 2;
+}
+
+#binary-select-dialog #path-header {
+    height: 3;
+    margin-bottom: 1;
+}
+
+#binary-select-dialog #up-button {
+    width: 3;
+    min-width: 3;
+    margin-left: 1;
+}
+
+
+.chat-message-container {
+  height: auto;
+  width: 100%;
+}
+MarkdownFence {
+    height: auto;
+    overflow: auto;
+    max-height: 9999;
+}
+Static.text1 {
+    padding-left: 2;
+    padding-bottom: 1;
+}
+Static.label_sender {
+    width: auto;
+}
+
+.tool_response_container {
+    max-height: 50;
+    height: auto;
+    overflow: auto;
+    padding-bottom: 1;
+    border: solid white; 
+}
+
+ModelSelect {
+    display: none;
+    align: center middle;
+    Container {
+        margin: 0;
+
+        width: 60%;
+        height: 50%;
+        overflow: hidden;
+        display: block;
+        OptionList {
+            margin-bottom: 3;
+        }
+    }
+}
+
+BinarySelectDialog {
+    display: none;
+    align: center middle;
+    Vertical {
+        background: transparent;
+        height: 50%;
+        width: 50%;
+    }
+}

Attic/py/r2ai/ui/chat.py

@@ -0,0 +1,25 @@
+from litellm import acompletion, ChatCompletionAssistantToolCall, ChatCompletionToolCallFunctionChunk
+import asyncio
+import json
+import signal
+from r2ai.pipe import get_r2_inst
+from r2ai.tools import run_python, r2cmd, execute_binary
+from r2ai.repl import r2ai_singleton
+from r2ai.auto import ChatAuto, SYSTEM_PROMPT_AUTO
+from r2ai.interpreter import is_litellm_model
+from r2ai.models import new_get_hf_llm
+
+def signal_handler(signum, frame):
+    raise KeyboardInterrupt
+
+async def chat(ai, message, cb):
+    model = ai.model.replace(":", "/")
+    tools = [r2cmd, run_python, execute_binary]
+    ai.messages.append({"role": "user", "content": message})
+    tool_choice = 'auto'
+    if not is_litellm_model(model) and ai and not ai.llama_instance:
+        ai.llama_instance = new_get_hf_llm(ai, model, int(ai.env["llm.window"]))
+    
+    chat_auto = ChatAuto(model, interpreter=ai, system=SYSTEM_PROMPT_AUTO, tools=tools, messages=ai.messages, tool_choice=tool_choice, cb=cb)
+
+    return await chat_auto.achat()

Attic/py/r2ai/ui/model_select.py

@@ -0,0 +1,86 @@
+from textual.app import ComposeResult
+from textual.widgets import Input, OptionList
+from textual.widget import Widget
+from textual.widgets.option_list import Option
+from textual.containers import Container
+from textual.message import Message
+from textual.binding import Binding
+from textual.screen import ModalScreen, SystemModalScreen
+from textual import log
+from r2ai.models import models
+# from ..repl import set_model, r2ai_singleton
+# ai = r2ai_singleton()
+
+# MODELS = models().split("\n")
+from litellm import models_by_provider
+MODELS = []
+for model in models().split("\n"):
+    if model.startswith("-m "):
+        MODELS.append(model[3:])
+for provider in models_by_provider:
+    for model in models_by_provider[provider]:
+        MODELS.append(f"{provider}/{model}")
+class ModalInput(Input):
+    BINDINGS = [
+        Binding("down", "cursor_down", "Move down"),
+    ]
+
+
+class ModelSelect(SystemModalScreen):
+    BINDINGS = [
+        Binding("up", "cursor_up", "Move up"),
+        Binding("down", "cursor_down", "Move down"),
+        Binding("enter", "select", "Select model"),
+        Binding("escape", "app.pop_screen", "Close"),
+    ]
+
+    class ModelSelected(Message):
+        """Event emitted when a model is selected."""
+        def __init__(self, model: str) -> None:
+            self.model = model
+            super().__init__()
+
+    def compose(self) -> ComposeResult:
+        self.input = ModalInput(placeholder="Type to filter...")
+        self.option_list = OptionList()
+        with Container():
+            yield self.input
+            yield self.option_list
+
+
+    def on_mount(self) -> None:
+        self.options = []
+        for t in MODELS:
+            if t.startswith("-m "):
+                self.options.append(Option(t[3:], id=t[3:]))
+            elif len(t) > 0:
+                self.options.append(Option(t, id=t))
+        self.option_list.add_options(self.options)
+        self.filtered_options = self.options.copy()
+        self.input.focus()
+    
+    def update_options(self, options):
+        self.option_list.clear_options()
+        self.option_list.add_options(options)
+        self.filtered_options = options
+    
+    def on_input_changed(self, event: Input.Changed) -> None:
+        filter_text = event.value
+        filtered_options = [option for option in self.options if filter_text.lower() in option.id.lower()]
+        self.update_options(filtered_options)
+
+    def action_cursor_up(self) -> None:
+        self.option_list.action_cursor_up()
+
+    def action_cursor_down(self) -> None:
+        if self.option_list.has_focus:
+            self.option_list.action_cursor_down()
+        else:
+            self.option_list.focus()
+
+    def on_option_list_option_selected(self, index) -> None:
+        selected_index = index.option_index
+        if 0 <= selected_index < len(self.filtered_options):
+            selected_option = self.filtered_options[selected_index]
+            if not selected_option.disabled:
+                self.dismiss(selected_option.id)

Attic/py/r2ai/utils.py

@@ -26,10 +26,8 @@ def merge_deltas(original, delta):
     return original
 
 def slurp(f):
-    fd = open(f, errors="ignore")
-    data = fd.read()
-    fd.close()
-    return str(data)
+    with open(f, errors="ignore") as fd:
+        return str(fd.read())
 
 def dump(f, x):
     fd = open(f, "w")
@@ -40,3 +38,30 @@ def syscmdstr(cmd):
     process = subprocess.Popen(cmd, stdout=subprocess.PIPE, stderr=subprocess.PIPE, shell=True)
     output, error = process.communicate()
     return output.decode().strip()
+
+
+def filter_print(*args, **kwargs):
+    _args = []
+    filter = None
+    if "filter" in kwargs:
+        filter = kwargs["filter"]
+        del kwargs["filter"]
+    for a in args:
+        new = ""
+        lines = str(a).splitlines()
+        if len(lines) > 1:
+            for line in lines:
+                if filter is not None:
+                    if filter in line:
+                        new += line + "\n"
+                else:
+                    new += line + "\n"
+        else:
+            if filter is not None:
+                if filter in str(a):
+                    new += str(a)
+            else:
+                new += str(a)
+        _args.append(new)
+    
+    print(*_args, **kwargs)

Attic/py/r2ai/voice.py

@@ -13,7 +13,7 @@ DEVICE = None
 try:
     import whisper
     HAVE_WHISPER = True
-except:
+except Exception:
     pass
 
 have_festival = os.path.isfile("/usr/bin/festival")
@@ -50,8 +50,16 @@ def stt(seconds, lang):
         return
     tts("(r2ai) listening for 5s... ", "digues?", lang)
     print(f"DEVICE IS {device}")
-    os.system("rm -f .audiomsg.wav")
-    rc = os.system(f"ffmpeg -f avfoundation -t 5 -i '{device}' .audiomsg.wav > /dev/null 2>&1")
+    try:
+        os.remove(".audiomsg.wav")
+    except OSError:
+        pass
+    result = subprocess.run(
+        ["ffmpeg", "-f", "avfoundation", "-t", "5", "-i", device, ".audiomsg.wav"],
+        stdout=subprocess.DEVNULL,
+        stderr=subprocess.DEVNULL
+    )
+    rc = result.returncode
     if rc != 0:
         tts("(r2ai)", "cannot record from microphone. missing permissions in terminal?", lang)
         return
@@ -60,7 +68,10 @@ def stt(seconds, lang):
         result = model.transcribe(".audiomsg.wav")
     else:
         result = model.transcribe(".audiomsg.wav", language=lang)
-    os.system("rm -f .audiomsg.wav")
+    try:
+        os.remove(".audiomsg.wav")
+    except OSError:
+        pass
     tts("(r2ai)", "ok", lang)
     text = result["text"].strip()
     if text == "you":

Attic/py/r2ai/web.py

@@ -1,11 +1,18 @@
-import _thread as thread
 import platform
 import json
 import re
+import time
+from threading import Thread, Event
 from .utils import get_timez
+from . import LOGGER
 
 ores = ""
 
+kill_event = None
+server = None
+server_thread = None
+server_loading = False
+
 # OpenAI API endpoint here
 def handle_v1_completions_default(self, ai, obj, runline2, method):
     if "prompt" in obj:
@@ -294,6 +301,7 @@ def handle_custom_request(self, ai, msg, runline2, method):
     return True
 
 def start_http_server_now(ai, runline2):
+    global server_loading
     import http.server
     import socketserver
     WANTCTX = ai.env["http.chatctx"] == "true"
@@ -325,15 +333,66 @@ def start_http_server_now(ai, runline2):
                 self.send_response(404)
                 self.end_headers()
                 self.wfile.write(bytes(f'Invalid request. Use {BASEPATH}'))
-    print("[R2AI] Serving at port", PORT)
+    LOGGER.getChild("server").info("[R2AI] Serving at port %s", PORT)
     Handler.protocol_version = "HTTP/1.0"
-    server = socketserver.TCPServer(("", PORT), SimpleHTTPRequestHandler)
-    server.allow_reuse_address = True
-    server.allow_reuse_port = True
-    server.serve_forever()
+    global server
+    if server is not None:
+        server_loading = False
+        print("Already running")
+        return
+    server = None
+    try:
+        server = socketserver.TCPServer(("", PORT), SimpleHTTPRequestHandler)
+        server.allow_reuse_address = True
+        server.allow_reuse_port = True
+        server.serve_forever()
+    except Exception as e:
+        print(e)
+    server_loading = False
 
 def start_http_server(ai, runline2, background):
-    if background:
-        thread.start_new_thread(start_http_server_now, (ai, runline2))
-    else:
-        start_http_server_now(ai, runline2)
+    global server_thread, server_loading
+    server_loading = True
+    server_thread = Thread(target=start_http_server_now, args=(ai, runline2))
+    server_thread.start()
+    while server_loading:
+        time.sleep(0.1)
+    global server
+    if server is None:
+        server_thread.join()
+        return
+    if not background:
+        global kill_event
+        kill_event = Event()
+        while not kill_event.is_set():
+            time.sleep(0.2)
+        kill_event = None
+
+def stop_http_server(force=False):
+    if not force and server_in_background():
+        return
+    global server
+    if server:
+        LOGGER.getChild("server").info("Stopping server")
+        _kill_server()
+        server = None
+    global server_thread
+    if server_thread:
+        server_thread.join()
+        server_thread = None
+    if not server_in_background():
+        kill_event.set()
+
+def server_running():
+    global server_thread
+    global server
+    return bool(server) and bool(server_thread)
+
+def server_in_background():
+    global kill_event
+    return server_running and not bool(kill_event)
+
+def _kill_server():
+    global server
+    server.shutdown()
+    Thread(target=server.server_close).start()

Attic/server/Makefile

@@ -0,0 +1,21 @@
+PWD=$(shell pwd)
+R2PM_BINDIR=$(shell r2pm -H R2PM_BINDIR)
+R2_BINDIR=$(shell r2 -H R2_BINDIR)
+
+install:
+#	rm -f $(R2PM_BINDIR)/r2ai-server
+#	ln -fs $(PWD)/r2ai-server/r2ai-server $(R2PM_BINDIR)/r2ai-server
+	mkdir -p $(R2_BINDIR)
+	cp -f r2ai-server $(R2_BINDIR)/r2ai-server || echo Maybe use user-install instead?
+
+uninstall:
+	rm -f $(R2_BINDIR)/r2ai-server
+
+user-install:
+	mkdir -p $(R2PM_BINDIR)
+	rm -f $(R2PM_BINDIR)/r2ai-server
+	ln -fs $(PWD)/r2ai-server $(R2PM_BINDIR)/r2ai-server
+
+user-uninstall:
+	rm -f $(R2PM_BINDIR)/r2ai-server
+

Attic/server/README.md

@@ -0,0 +1,25 @@
+# r2ai-server
+
+Start ollama, llamacpp, llamafile, r2ai, kobaldcpp and other language model webservers using a the same syntax for all of them to simplify its launching and setup.
+
+* Use -l to list the implementations available
+
+You can install more via r2pm
+
+* Use -m to list or select the model. Those models can be absolute paths or model names when downloaded via r2ai, which uses the hugging face api
+
+## Running r2ai-server
+
+- Get usage: `r2pm -r r2ai-server`
+- List available servers: `r2pm -r r2ai-server -l`
+- List available models: `r2pm -r r2ai-server -m`
+
+On Linux, models are stored in `~/.r2ai.models/`. File `~/.r2ai.model` lists the default model and other models.
+
+**Example launching a local Mistral AI server:**
+
+```
+$ r2pm -r r2ai-server -l r2ai -m mistral-7b-instruct-v0.2.Q2_K
+[12/13/24 10:35:22] INFO     r2ai.server - INFO - [R2AI] Serving at port 8080                               web.py:336
+```
+

Attic/server/r2ai-server

@@ -4,21 +4,31 @@ MODEL=""
 PORT=8080
 MODELDIR=~/.r2ai.models
 LLAMA=llamafile
+[ -z "${R2AI}" ] && R2AI=r2ai
 
 main() {
 	if [ -z "${MODEL}" ]; then
-		echo "Use -l and -m to select a model"
+		echo "Use -l and -m to select a model" >&2
 		exit 1
 	fi
+	echo "${MODEL}" | grep -e ^/
+	if [ $? = 0 ]; then
+		MODELPATH="${MODEL}"
+	else
+		MODELPATH="${MODELDIR}/${MODEL}.gguf"
+	fi
 	case ${LLAMA} in
 	llamafile)
-		llamafile --server --nobrowser -c 0 --port ${PORT} -m "${MODELDIR}/${MODEL}.gguf"
+		llamafile --server --nobrowser -c 0 --port ${PORT} -m "${MODELPATH}"
 		;;
 	llamacpp)
-		llama-server -c 0 --port ${PORT} -m "${MODELDIR}/${MODEL}.gguf"
+		llama-server -c 0 --port "${PORT}" -m "${MODELPATH}"
 		;;
 	koboldcpp)
-		koboldcpp -c 0 --port ${PORT} -m "${MODELDIR}/${MODEL}.gguf"
+		koboldcpp -c 0 --port "${PORT}" -m "${MODELPATH}"
+		;;
+	r2ai)
+		${R2AI} --port "${PORT}" -m "${MODEL}" -w
 		;;
 	*)
 		echo "Invalid llama server selected."
@@ -29,11 +39,12 @@ main() {
 help() {
 	cat <<EOF
 Usage: r2ai-server [-flags]
- -m         list models
- -m [name]  select given model
- -l         list llama servers available
- -h         show this help message
- -p [port]  select port        
+-m         list models
+-m [name]  select given model
+-l         list llama servers available
+-l [name]  select llama software
+-h         show this help message
+-p [port]  select port
 EOF
 }
 
@@ -44,6 +55,8 @@ models() {
 }
 
 llamas() {
+	# TODO check for those bins in path before suggesting them
+	echo r2ai
 	echo llamafile
 	echo llamacpp
 	echo koboldcpp
@@ -64,11 +77,11 @@ while : ; do
 			models
 			break
 		fi
-		MODEL=$2
+		MODEL="$2"
 		shift
 		;;
 	-p)
-		PORT=$2
+		PORT="$2"
 		shift
 		;;
 	-h)
@@ -76,6 +89,7 @@ while : ; do
 		;;
 	"")
 		main
+		break
 		;;
 	esac
 	shift

LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2024 pancake
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

Makefile

@@ -1,94 +1,9 @@
-R2_USER_PLUGINS=$(shell r2 -H R2_USER_PLUGINS)
-PWD=$(shell pwd)
-R2PM_BINDIR=$(shell r2pm -H R2PM_BINDIR)
-# Note that a bunch of packages are not available for 3.12 yet
-ifeq ($(shell which python3.11 > /dev/null && echo ok),ok)
-PYTHON?=python3.11
-else
-PYTHON?=python3
-endif
-PIP=$(PYTHON) -m pip
+TARGETS=all clean mrproper user-install install user-uninstall uninstall fmt indent
 
-LINTED=r2ai/code_block.py
-LINTED+=r2ai/bubble.py
-LINTED+=r2ai/const.py
-LINTED+=main.py
-LINTED+=r2ai/backend/kobaldcpp.py
-# LINTED+=r2ai/index.py
-# LINTED+=r2ai/voice.py
-# LINTED+=r2ai/anthropic.py
+$(TARGETS):
+	$(MAKE) -C src $@
 
-ifeq ($(R2PM_BINDIR),)
-FATAL ERROR
-endif
+test: all
+	$(MAKE) -C test
 
-.PHONY: all all.old deps clean deps-global pub lint cilint
-.PHONY: install uninstall user-install user-uninstall
-
-all: venv
-	@./r2ai.sh
-
-large:
-	. venv/bin/activate ; $(PYTHON) main.py -l
-
-all.old:
-	@test -n "${VIRTUAL_ENV}" || (echo "Run:"; echo ". venv/bin/activate" ; exit 1)
-	$(PYTHON) main.py || $(MAKE) deps
-
-venv:
-	$(PYTHON) -m venv venv
-	if [ -z "`find venv | grep llama_cpp`" ]; then . venv/bin/activate ; pip install -r requirements.txt ; fi
-
-deps: venv
-	#test -n "${VIRTUAL_ENV}" || (echo "Run: . venv/bin/activate" ; exit 1)
-	. venv/bin/activate && export CMAKE_ARGS="-DLLAMA_METAL=on -DLLAMA_METAL_EMBED_LIBRARY=ON" && \
-		pip install --force-reinstall -U -r requirements.txt --no-cache-dir
-	$(MAKE) vectordb
-
-clean:
-	rm -rf venv vectordb vdb
-
-mrproper:
-	$(MAKE) clean
-
-deps-global:
-	export CMAKE_ARGS="-DLLAMA_METAL=on -DLLAMA_METAL_EMBED_LIBRARY=ON" && \
-		$(PIP) install --force-reinstall -U -r requirements.txt --break-system-packages --no-cache-dir
-
-vdb vectordb:
-	git clone https://github.com/kagisearch/vectordb vdb
-	cat vdb/setup.py | grep -v tensorflow_text > .x && mv .x vdb/setup.py
-	. venv/bin/activate \
-		&& cd vdb \
-		&& $(PIP) install setuptools tensorflow_hub \
-		&& $(PYTHON) setup.py build \
-		&& $(PYTHON) setup.py install
-
-user-install:
-	ln -fs $(PWD)/r2ai-server $(R2PM_BINDIR)/r2ai-server
-	ln -fs $(PWD)/r2ai.sh $(R2PM_BINDIR)/r2ai
-
-install: user-install
-	-mkdir -p /usr/local/share/man/man1/r2ai.1
-	-cp doc/usage/r2ai.1 /usr/local/share/man/man1/r2ai.1
-
-install-plugin user-install-plugin:
-	ln -fs $(PWD)/main.py $(R2_USER_PLUGINS)/r2ai.py
-
-uninstall user-uninstall:
-	rm -f $(R2PM_BINDIR)/r2ai
-	-rm -f /usr/local/share/man/man1/r2ai.1
-
-user-uninstall-plugin uninstall-plugin:
-	rm -f $(R2_USER_PLUGINS)/r2ai.py
-
-pub:
-	$(PYTHON) setup.py build sdist
-	twine check dist/*
-	twine upload -u __token__ --repository-url https://upload.pypi.org/legacy/ --verbose dist/*
-
-cilint:
-	pylint $(LINTED)
-
-lint:
-	pylint *.py r2ai/*.py
+.PHONY: $(TARGETS) test

README.md

@@ -1,110 +1,115 @@
-```
-,______  .______ .______  ,___
-: __   \ \____  |:      \ : __|
-|  \____|/  ____||  _,_  || : |
-|   :  \ \   .  ||   :   ||   |
-|   |___\ \__:__||___|   ||   |
-|___|        :       |___||___|
-             *       --pancake
+# R2AI - Augmented reversing with LLM for radare2
+
+```console
+         ╭─────────────────────────────────╮
+         │ ,______  .______ .______  ,___  │
+ ╭──╮    │ : __   \ \____  |:      \ : __| │
+ │ _│_   │ |  \____|/  ____||  _,_  || : | │
+ │ O O  <  |   :  \ \   .  ||   :   ||   | │
+ │  │╷   │ |   |___\ \__:__||___|   ||   | │
+ │  ││   │ |___|        :       |___||___| │
+ │ ─╯│   ╰─────────────────────────────────╯
+ ╰───╯
 ```
 
 [![ci](https://github.com/radareorg/r2ai/actions/workflows/ci.yml/badge.svg)](https://github.com/radareorg/r2ai/actions/workflows/ci.yml)
+[![radare2](https://img.shields.io/badge/radare2-6.0.4-green)](https://github.com/radareorg/radare2)
 
-Run a language model in local, without internet, to entertain you or help answering questions about radare2 or reverse engineering in general. Note that models used by r2ai are pulled from external sources which may behave different or respond unrealible information. That's why there's an ongoing effort into improving the post-finetuning using memgpt-like techniques which can't get better without your help!
+## Components
 
-<p align="center">
-  <img src="doc/r2clippy.jpg">
-</p>
+This repository contains two plugins for radare2:
+
+* **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:
+
+* **r2mcp** - the [official radare2 mcp](https://github.com/radare2/radare2-mcp)
+* **r2copilot** - the mcp with focus on CTF [r2copilot](https://github.com/darallium/r2-copilot)
+
+If you are looking for a radare2 focused autonomous agent:
+* **r2agent** - automate radare2 workflows through autonomous agents. [r2agent](https://github.com/nitanmarcel/r2agent)
 
 ## Features
 
-* Prompt the language model without internet requirements
-* Use local GGUF or remote language models (via http)
-* Index large codebases or markdown books using a vector database
-* Slurp file and perform actions on that
-* Embed the output of an r2 command and resolve questions on the given data
-* Define different system-level assistant role
-* Set environment variables to provide context to the language model
+* Configure different roles and customize prompts
+* Scriptable via r2pipe via the r2ai command
 * Live with repl and batch mode from cli or r2 prompt
-* Accessible as an r2lang-python plugin, keeps session state inside radare2
-* Scriptable from python, bash, r2pipe, and javascript (r2papi)
-* Use different models, dynamically adjust query template
-  * Load multiple models and make them talk between them
+* Support Automatic (ReAct) mode to solve tasks using function calling
+* Use local and remote language models (ollama, openai, grok, anthropic, ..)
+* RAG markdown, code or textfiles using its native vector database
+* Embed the output of an r2 command and resolve questions on the given data
+
+## User defined Prompts
+
+```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
 
-To use `r2ai`, you need to setup a venv and install all the python dependencies inside. Just running `make` or `r2pm -ci r2ai` should be enough.
+The recommended way to install any of the r2ai components is via r2pm:
 
-You can optionally install vectordb to index local data by typing: `make vdb`. If the process fails  `rm -rf vdb` and try again
-
-## Running
-
-When running installed via r2pm you can execute it like this:
-
-```bash
-r2pm -r r2ai
+```console
+$ r2pm -Uci r2ai
+$ r2pm -Uci decai
 ```
 
-Additionally you can get the `r2ai` command inside r2 to run as an rlang plugin by installing the bindings:
+## Using r2ai
 
-```bash
-r2pm -i rlang-python
-make user-install
+- Adds the **r2ai** command to the radare2 shell: `r2 -qc r2ai`
+- You can also run the wrapper in $PATH: `r2pm -r r2ai`
+
+Drop your API keys in environment variables or use the configuration file:
+
+```console
+$ export ANTHROPIC_API_KEY=sk-ant-api03-CENSORED
+$ export OPENAI_API_KEY=sk-proj-6rlSPS-zN1v...
 ```
 
-After this you should get the `r2ai` command inside the radare2 shell. Set the `R2_DEBUG=1` environment to see the reasons why the plugin is not loaded if it's not there.
+Or edit the api keys file `~/.config/r2ai/apikeys.txt` run:
 
-## Windows
-
-On Windows you may follow the same instructions, just ensure you have the right python environment ready and create the venv to use
-
-```cmd
-git clone https://github.com/radareorg/r2ai
-cd r2ai
-set PATH=C:\Users\YOURUSERNAME\Local\Programs\Python\Python39\;%PATH%
-python3 -m pip -r requirements.txt
-python3 main.py
+```console
+$ r2ai -K
 ```
 
-## Usage
+## Saving settings
 
-There are 4 different ways to run `r2ai`:
+You may customize and save your configuration settings using your OS's default settings file (e.g `~/.radare2rc` on Linux).
+For example, the following configuration sets Claude 3.7 by default, with max output tokens to 64000.
 
-* Standalone and interactive: `r2pm -r r2ai` or `python main.py`
-* Batch mode: `r2ai '-r act as a calculator' '3+3=?'`
-* As an r2 plugin: `r2 -i main.py /bin/ls`
-* From radare2 (requires `r2pm -ci rlang-python`): `r2 -c 'r2ai -h'`
-* Using r2pipe: `#!pipe python main.py`
-  * Define a macro command: `'$r2ai=#!pipe python main.py`
 
-## Auto mode
-
-When using OpenAI, Claude or any of the Functionary local models you can use the auto mode which permits the language model to execute r2 commands, analyze the output in loop and in a loop until it is resolved. Here's a sample session to achieve that:
-
-* Video https://infosec.exchange/@radareorg/111946255058894583
-
-```bash
-$ . env/bin/activate
-(env)$ r2 /bin/ls
-[0x00000000]> '$r2ai=#!pipe python main.py
-[0x00000000]> $r2ai '-m openai:gpt-4'
-[0x00000000]> $r2ai "' list the imports for this program"
-[0x00000000]> $r2ai "' draw me a donut"
-[0x00000000]> $r2ai "' decompile current function and explain it"
+```console
+$ r2ai -E
 ```
 
-## Examples
+then you can type the commands you want to run when the r2ai plugin is loaded:
 
-You can interact with r2ai from standalone python, from r2pipe via r2 keeping a global state or using the javascript interpreter embedded inside `radare2`.
+```console
+r2ai -e api=anthropic
+r2ai -e model=claude-sonnet-5-5
+r2ai -e max_tokens=64000
+```
 
-* [conversation.r2.js](examples/conversation.r2.js) - load two models and make them talk to each other
+## Further Reading
 
-### Development/Testing
+* 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://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
 
-Just run `make` .. or well `python3 main.py`
+## Videos
 
-### TODO
-
-* add "undo" command to drop the last message
-* dump / restore conversational states (see -L command)
-* Implement `~`, `|` and `>` and other r2shell features
+- [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

@@ -0,0 +1 @@
+1.5.4

autogen.sh

@@ -0,0 +1,6 @@
+#!/bin/sh
+vim decai/src/constants.ts
+make -C decai/src
+vim src/r2ai.h
+vim src/meson.build
+vim VERSION

decai/AGENTS.md

@@ -0,0 +1,13 @@
+# Decompiler based AI
+
+## Actions
+
+* Build: `make -C src`
+* Install: `make user-install`
+* Test: `r2 -A -qc "decai -e api=ollamacloud;decai -e model=glm-5;decai -d @ 0x100000924" /bin/ls`
+
+## Hints
+
+* Never read the minified, autogenerated `decai.r2.js`
+* Focus and work only on the TypeScript source
+* Never run `tsc` or `npx`. Always use `make` targets to compile

decai/Makefile

@@ -0,0 +1,29 @@
+PD=$(shell r2 -H R2_USER_PLUGINS)
+SPD=$(shell r2 -H R2_LIBR_PLUGINS)
+BF=/bin/dd
+
+all:
+	@echo "Run: 'make user-install' to install it in your home"
+	@echo "Run: 'make -C src' to recompile the r2.js"
+
+test:
+	r2 -i decai.r2.js -c 'decai -e deterministic=true' -caf $(BF)
+
+fmt indent:
+	$(MAKE) -C src fmt
+
+user-install:
+	mkdir -p "$(PD)"
+	cp decai.r2.js $(PD)/decai.r2.js
+
+user-uninstall:
+	rm -f $(PD)/decai.r2.js
+
+install:
+	make user-install PD=$(SPD)
+
+uninstall:
+	make user-uninstall PD=$(SPD)
+
+vs open:
+	open -a "Visual Studio Code" src

decai/README.md

@@ -0,0 +1,238 @@
+# Decai
+
+The AI based Decompiler plugin for radare2
+
+- Written in plain Javascript
+- No dependencies than radare2 and curl in PATH
+- Uses local ollama by default, but also r2ai-server
+- Use services like Anthropic, OpenAI, HF, XAI, DeepSeek, ..
+
+Features
+
+- Auto mode with function calling with ANY model
+- Uses the r2 pseudo decompiler by default, supports any other
+- Explain purpose and auto-document functions
+- Recursive decompilation to inline stubs
+- Autoname functions and perform type propagation
+- Find vulnerabilities, guide you and write exploits
+- Choose any natural language (not just English)
+- Choose output programming language (not just C)
+- Chain queries to manually fine tune the results
+- Customize decompilation prompt at any time
+
+## Installation
+
+Using r2pm: `r2pm -ci decai`
+
+From source: `make user-install`
+
+## Setup
+
+```console
+[0x00000000]> decai
+Usage: decai (-h) ...
+ decai -H         - help setting up r2ai
+ decai -d [f1 ..] - decompile given functions
+ decai -dr        - decompile function and its called ones (recursive)
+ decai -dd [..]   - same as above, but ignoring cache
+ decai -dD [query]- decompile current function with given extra query
+ decai -e         - display and change eval config vars
+ decai -E         - edit decai.txt
+ decai -h         - show this help
+ decai -i [f] [q] - include given file and query
+ decai -n         - suggest better function name
+ decai -q [text]  - query language model with given text
+ decai -Q [text]  - query on top of the last output
+ decai -r         - change role prompt (same as: decai -e prompt)
+ decai -R         - reset role prompt to default prompt
+ decai -s         - function signature
+ decai -v         - show local variables
+ decai -V         - find vulnerabilities
+ decai -x         - eXplain current function
+```
+
+Configuration options:
+
+```console
+[0x00000000]> decai -e
+decai -e api=ollama
+decai -e apitype=chat
+decai -e host=http://localhost
+decai -e port=11434
+decai -e prompt=Rewrite this function and respond ONLY with code, NO explanations, NO markdown, Change 'goto' into if/else/for/while, Simplify as much as possible, use better variable names, take function arguments and strings from comments like 'string:'
+decai -e ctxfile=
+decai -e headers=
+decai -e cmds=pdc
+decai -e cache=false
+decai -e lang=C
+decai -e hlang=English
+decai -e debug=false
+decai -e model=
+decai -e maxinputtokens=-1
+```
+
+## Running decai
+
+Decai is used from `r2` (e.g `r2 ./mybinary`). Get help with `decai -h`:
+
+```
+[0x00406cac]> decai -h
+Usage: decai (-h) ...
+ decai -H         - help setting up r2ai
+ decai -a [query] - solve query with auto mode
+ decai -d [f1 ..] - decompile given functions
+ decai -dr        - decompile function and its called ones (recursive)
+ decai -dd [..]   - same as above, but ignoring cache
+ decai -dD [query]- decompile current function with given extra query
+ ...
+```
+
+List configuration variables with `decai -e`:
+
+```
+[0x00406cac]> decai -e
+decai -e api=ollama
+decai -e host=http://localhost
+decai -e port=11434
+decai -e prompt=Rewrite this function and respond ONLY with code, NO explanations, NO markdown, Change 'goto' into if/else/for/while, Simplify as much as possible, use better variable names, take function arguments and strings from comments like 'string:'
+decai -e ctxfile=
+...
+```
+
+Persist user defaults in `~/.config/r2ai/decai.txt` with one `key=value` entry
+per line, then edit it with `decai -E`. That file is loaded on the first
+`decai` command in a session and reloaded after `decai -E` exits.
+
+Use `decai -e http=curl` to force system curl, `http=r2` to require `r2.http()`,
+or `http=auto` (the default) to use `r2.http()` when available and curl otherwise.
+Failed `r2.http()` calls return an error without falling back to curl. The
+`timeout` setting applies only to system curl. Save `http=curl` in
+`~/.config/r2ai/decai.txt` to persist the override.
+
+List possible APIs to discuss with AI: `decai -e api=?`:
+
+```
+[0x00406cac]> decai -e api=?
+r2ai
+claude
+openapi
+...
+```
+
+Extra headers can be injected globally with `decai -e headers=...` or from the
+environment with `DECAI_HEADERS` / `R2AI_HEADERS`. Use one header per line, for
+example:
+
+```console
+[0x00406cac]> decai -e headers=Authorization: Bearer sk-123\nUser-Agent: curl/8.7.1
+```
+
+### Example using a local model and ollama
+
+For example, if Ollama serves model codegeex4:latest (`ollama ls`), set decai
+API as `ollama` and model `codegeex4:latest`.
+
+```
+[0x00002d30]> decai -e api=ollama
+[0x00002d30]> decai -e model=codegeex4:latest
+[0x00002d30]> decai -q Explain what forkpty does in 2 lines
+The `forkpty` function creates a new process with a pseudo-terminal, allowing the parent to interact with the child via standard input/output/err and controlling its terminal.
+```
+
+## Examples
+
+See
+[https://github.com/radareorg/r2ai-examples](https://github.com/radareorg/r2ai-examples)
+
+```c
+$ cat stack-overflow/bug.c 
+
+#include <string.h>
+
+int main(int argc, char **argv) {
+	char buf[32];
+	strcpy (buf, argv[1]);
+	return 0;
+}
+```
+
+```c
+$ r2 buffer-overflow/a.out
+[0x100003f58]> decai -d
+int main(int argc, char **argv, char **envp) {
+    char buffer[32];
+    int result = 0;
+    
+    if (argc > 1 && argv[1] != NULL) {
+        strcpy(buffer, argv[1]);
+    }
+    
+    return result;
+}
+```
+
+### Example using a local Mistral model and r2ai-server
+
+For example, assuming we have a _local_ Mistral AI server running on port 8080
+with `r2ai-server`, we can decompile a given function with `decai -d`. The
+server shows it received the question:
+
+```
+GET
+CUSTOM
+RUNLINE: -R
+127.0.0.1 - - [13/Dec/2024 10:40:49] "GET /cmd/-R HTTP/1.1" 200 -
+GET
+CUSTOM
+RUNLINE: -i /tmp/.pdc.txt Rewrite this function and respond ONLY with code, NO explanations, NO markdown, Change goto into if/else/for/while, Simplify as much as possible, use better variable names, take function arguments and and strings from comments like string:. Transform this pseudocode into C
+```
+
+### Example using a Mistral API key
+
+Put the API key in `~/.r2ai.mistral-key`.
+
+````
+[0x000010d0]> decai -e api=mistral
+[0x000010d0]> decai -d main
+```c
+#include <stdio.h>
+#include <string.h>
+#include <stdlib.h>
+
+int main(int argc, char **argv, char **envp) {
+    char password[40];
+    char input[40];
+...
+````
+
+### Example with ChatGPT 4
+
+```
+[0x00406cac]> decai -e api=openai
+[0x00406cac]> decai -d
+#include <stdio.h>
+#include <unistd.h>
+
+void daemonize() {
+    daemon(1, 0);
+}
+...
+```
+
+## WASM HTTP
+
+With radare2's WASI HTTP build, decai defaults to `r2.http()` to reach the existing
+`r_socket_http_get/post` C APIs. Headers and JSON payloads go directly to the
+host HTTP adapter; no shell curl, temporary payload file or WebSocket proxy is
+needed. A failed/denied HTTP call is returned as an error, without a shell
+fallback. In `http=auto` mode, runtimes without `r2.http()` use system curl.
+
+Enable HTTP in the embedding application (`enabled: true` in JavaScript or
+`network=True` in Python). The radare2 sandbox must permit both `socket` and
+`network`; process execution is not needed. Browser requests remain subject
+to CORS. The host adapter controls the HTTP deadline in WASM; decai's `timeout`
+setting still controls native curl only. Responses are buffered, not streamed.
+
+This requires radareorg/radare2#27037 and the rebuilt decai script. For the
+C r2ai plugin compiled into WASM, the `auto` HTTP backend likewise selects
+radare2's C HTTP implementation; native backend selection is unchanged.

decai/decai.r2.js

@@ -0,0 +1,171 @@
+📦
+27680 /main.js
+✄
+var te="1.5.4",b="decai",w="~/.config/r2ai",N=w+"/apikeys.txt",U=w+"/decai.txt",T="Rewrite this pseudocode into concise and clean code. Output only the provided function. Do not add wrappers, helper examples, test code, or main-like functions. Replace goto with structured control flow, simplify as much as possible, infer types and use better names for variables and parameters, some strings may be appearing as comments, preserve only what is implied by the input, and remove dead code.",L={decai:`# Using Decai
+
+Decai is the radare2 plugin for decompiling functions with the help of language models.
+
+By default uses a local ollama server, but can you can pick any other service by using 'decai -e api=?'.
+
+[0x00000000]> decai -e api=?
+r2ai deepseek anthropic claude gemini groq hf lmstudio mistral ollama ollamacloud openapi openai openrouter vllm xai zen
+
+## Using Ollama
+
+* Visit https://ollama.com to install it.
+* Download the model of choice: 'ollama run llama3.3'
+* Configure decai to use the given model with: 'decai -e model=?'
+
+These are the most recommended models for decompiling in local:
+
+* hhao/qwen2.5-coder-tools:latest (18GB of ram)
+* hhao/qwen2.5-coder-tools:32b (24GB of ram required)
+
+## Common Options
+* 'decai -e baseurl=<url>' override default host and port for API endpoint (e.g., 'http://localhost:11434')
+* 'decai -e apitype=chat|generate' choose the Ollama API endpoint (default: chat)
+* 'decai -e headers=Authorization: Bearer ...\\nUser-Agent: curl/8.7.1' add or override HTTP headers
+* "export DECAI_HEADERS='Authorization: Bearer ...\\nUser-Agent: curl/8.7.1'" set extra headers from the environment
+* 'decai -e http=auto|r2|curl' select the HTTP transport (default: auto, prefers r2.http)
+* 'decai -e timeout=0' disable curl timeouts, or set a positive number of seconds
+
+* 'decai -e deterministic=true' to remove randomness from decompilation responses
+* 'decai -e lang=Python' to output the decompilation in Python instead of C
+* 'decai -e hlang=Catalan' to add comments or explanations in that language (instead of English)
+* 'decai -e cmds=pdd,pdg' use r2dec and r2ghidra instead of r2's pdc as input for decompiling
+* 'decai -e prompt=..' default prompt must be fine for most models and binaries, feel free to tweak it
+
+## API Keys
+
+Remove services like OpenAI, Mistral, Anthropic, Grok, Gemini, .. require API keys to work.
+
+See 'decai -k' to list the status of available APIkeys
+Use 'decai -K' to edit apikeys.txt and 'decai -E' to edit decai.txt.
+
+Decai will pick them from the environment or the config files in your home:
+
+* echo KEY > ~/.r2ai.openai-key
+* export OPENAI_API_KEY=...
+* put one 'key=value' pair per line in ~/.config/r2ai/decai.txt
+
+## Using the R2AI Server:
+
+Install r2ai or r2ai-server with r2pm:
+
+[0x0000000]> decai -e api=r2ai
+[0x0000000]> r2pm -ci r2ai
+
+Choose one of the recommended models (after r2pm -r r2ai):
+
+* -m ibm-granite/granite-20b-code-instruct-8k-GGUF
+* -m QuantFactory/granite-8b-code-instruct-4k-GGUF
+* -m TheBloke/Mistral-7B-Instruct-v0.2-GGUF
+
+Start the webserver:
+
+$ r2pm -r r2ai-server -l r2ai -m granite-8...
+`,auto:`# Radare2 Auto Mode
+
+Use function calling to execute radare2 commands in order to resolve the user request defined in the "User Prompt" section, analyze the responses attached in the "Command Results" section.
+
+## Function Calling
+
+Respond ONLY using plain JSON. Process user query and decide which function calls are necessary to solve the task.
+
+1. Analyze the user request to determine if we need to run commands to extend the knowledge and context of the problem.
+2. If function call is needed, construct the JSON like this:
+ - Fill the "action" key with the "r2cmd" value.
+ - Specify the "command" as a string.
+ - Optionally, provide a "reason" and "description"
+3. If the answer can be provided and no more function calls are required:
+ - Use the key "action": "reply".
+ - Include "response" with the direct answer to the user query.
+
+Return the result as a JSON object.
+
+### Sample Function Calling Communication
+
+Command Results: already performed actions with their responses
+User Prompt: "Count how many functions we have here."
+Response:
+{
+    "action": "r2cmd",
+    "command": "aflc",
+    "description": "Count functions"
+    "reason": "Evaluate if the program is analyzed before running aaa"
+}
+
+## Rules
+
+Use radare2 to resolve user requests.
+
+* Explain each step in the "reason" field of the JSON.
+* Follow the initial analysis instructions.
+* Output only valid JSON as specified.
+* Decompile and inspect functions, starting from main.
+* Run only the needed commands to gather info.
+* Use "@ (address|symbol)" to seek temporarily.
+* Output should be a verbose markdown report.
+* Use "sym." or "fcn." prefixes if "pdc" is empty.
+* If a seek fails, use "f~name" to find the symbol's address.
+
+### Initial Analysis
+
+1. Run "aflc" to count the number of functions
+2. If the output of "aflc" is "0" run "aaa" once, then "aflc" again
+3. Run only one command at a time (do not use ";")
+
+### Special cases
+
+* On Swift binaries run...
+`};var E={decopipe:{use:!1},host:"http://localhost",port:"11434",baseurl:"",extraHeaders:[],api:"ollama",pipeline:"",commands:"pdc",yolo:!1,tts:!1,language:"C",humanLanguage:"English",deterministic:!0,debug:!1,timeout:180,http:"auto",think:"",apitype:"chat",useFiles:!1,contextFile:"",model:"",cache:!1,maxInputTokens:-1,prompt:T,lastOutput:""},r={...E,decopipe:{...E.decopipe},extraHeaders:[...E.extraHeaders]};function re(e){return(r2.cmd("-e dir.tmp").trim()??".")+"/"+e}function ne(e){if(!e.startsWith("~/"))return e;let t=r2.cmd("'%HOME").trim()||r2.cmd("'%USERPROFILE").trim();return t?t+e.slice(1):e}function H(e,t){r2.cmd("'mkdir -p "+ne(e)),r2.cmd("'touch "+ne(t))}function _(e){return(r2.cmd2("test -h").logs?.[0]?.message??"").includes("-fdx")?!0:r2.cmd("'test -vf "+e).startsWith("found")}function G(e,t){return e+" ".repeat(Math.max(0,t-e.length))}function h(e,t){return e.replace(/\/+$/,"")+"/"+t.replace(/^\/+/,"")}function oe(e){return e.replace(/\x1b\[[0-9;]*m/g,"")}function ie(e){let t=e.match(/```json\s*([\s\S]*?)```/);return t&&t[1]?t[1].trim():e}function ae(e){let t=e,n=t.indexOf("{");n!==-1&&(t=t.slice(n));let o=t.lastIndexOf("}");return o!==-1&&(t=t.slice(0,o+1)),t}function $(e){return btoa(e)}function se(e,t){let n=$(t);r2.cmd("p6ds "+n+" > "+e)}function k(e){let t=e;return t=t.replace(/<think>[\s\S]*?<\/think>/gi,""),t.split(`
+`).filter(n=>!n.trim().startsWith("```")).join(`
+`)}function J(e){r.debug&&console.log(e)}function le(e){let t={};for(let n of e.split(/\r?\n/)){let o=n.trim();if(!o||o.startsWith("#"))continue;let[i,...a]=o.split("=");if(!i||a.length===0)continue;let s=a.join("=").trim(),l=i.toLowerCase().replace(/_api_key$/i,"");t[l]=s}return t}var Ue=["DECAI_HEADERS","R2AI_HEADERS"];function _e(e){let t=r2.cmd("'%"+e).trim();return t!==""&&!t.includes("=")?t:""}function Me(e){return e.trim().toLowerCase()}function Ne(e){let t=e.indexOf(":"),n=e.indexOf("="),o=t!==-1&&(n===-1||t<n)?t:n;if(o===-1)return null;let i=e.slice(0,o).trim();return i===""?null:{name:i,value:e.slice(o+1).trim()}}function Te(e){return e.value===""?`${e.name}:`:`${e.name}: ${e.value}`}function ce(e,t){for(let n of e){let o=Ne(n.trim());o&&t.set(Me(o.name),Te(o))}}function W(e){let t=new Map,n=e.replace(/\\n/g,`
+`).split(/\r?\n/g).filter(o=>{let i=o.trim();return i&&!i.startsWith("#")});return ce(n,t),Array.from(t.values())}function ue(e){return e.join("\\n")}function A(...e){let t=new Map;for(let n of e)ce(n,t);return Array.from(t.values())}function He(){for(let e of Ue){let t=_e(e);if(t)return W(t)}return[]}function de(){return A(He(),r.extraHeaders)}function M(e){return`'${e.replace(/'/g,`'"'"'`)}'`}function Ke(e){let{method:t,url:n,headers:o,payload:i}=e,a=A(["Content-Type: application/json"],o);if(r.http!=="curl"){if(r2.http)return r2.http(t,n,a.join(`
+`),i??"").body;if(r.http==="r2")throw new Error("r2.http is not available in this runtime")}let s=["curl","-s"];if(r.timeout>0&&s.push("--max-time",String(r.timeout)),a.forEach(l=>s.push("-H",M(l))),t==="POST"){if(!i)throw new Error("Payload required for POST requests");let l=r2.fdump(i);s.push("--data-binary","@"+M(l),M(n));let u=s.join(" ")+"; rm "+M(l);return J(u),r2.syscmds(u)}else{s.push(M(n));let l=s.join(" ");return J(l),r2.syscmds(l)}}function qe(e){let t=e.split(`
+`).filter(i=>i.trim()!=="");if(t.length<=1)return null;let n="",o=null;for(let i of t)try{let a=JSON.parse(i);if(a.error)return a;a.message?.content!==void 0&&(n+=a.message.content),o=a}catch{return null}return o?(o.message={role:"assistant",content:n},o):null}function pe(e){try{let t=Ke(e).trim();if(t==="")return{error:"empty response"};try{return JSON.parse(t)}catch{let n=qe(t);return n||{error:t,rawOutput:t}}}catch(t){return{error:t.message}}}function O(e,t){return pe({method:"GET",url:e,headers:t})}function fe(e,t,n){return pe({method:"POST",url:e,headers:t,payload:n})}function z(e){return e.split("_")[0].toLowerCase()}function me(e){if(!e.authKey)return;let t=Array.from(new Set([e.keyName,z(e.authKey)].filter(Boolean))),n;for(let o of t){let i=K(o,e.authKey);if(i[0])return i;!n&&i[2]!=="nope"&&(n=i)}return n||K(z(e.authKey),e.authKey)}function S(e){if(e)return typeof e=="string"?e:e.message||JSON.stringify(e)}function he(e){return r.baseurl||e.defaultBaseUrl}function ge(e,t=de()){return A(e,t)}function De(e){return e?["Authorization: Bearer "+e]:[]}function je(e){let t=["anthropic-version: 2023-06-01"];return e?A(t,["x-api-key: "+e]):t}function ye(e,t){switch(e.authStyle||"none"){case"bearer":return De(t);case"anthropic":return je(t);case"none":default:return[]}}function Be(e,t,n){let o=O(e,t),i=S(o.error);return i?(console.error(i),"error invalid response"):o.data?.map(n).join(`
+`)||""}function Fe(e,t){let n=new Set;return e.filter(o=>{let i=t(o);return n.has(i)?!1:(n.add(i),!0)})}function V(e){let t=me(e)?.[0]||null;return{baseUrl:he(e),headers:ge(ye(e,t))}}function Le(e){let{baseUrl:t,headers:n}=V(e),o=h(t,"v1/models");return Be(o,n,i=>i.id)}function Ge(e){let{baseUrl:t,headers:n}=V(e),o=O(h(t,"api/tags"),n),i=S(o.error);return i?(console.error(i),"error invalid response"):o.models?.map(a=>a.name).join(`
+`)||""}function $e(e){let{baseUrl:t,headers:n}=V(e),o=O(h(t,"v1/models"),n);return o.data?Fe(o.data,i=>i.name||i.id).map(i=>[G(i.name||i.id,30),G(""+(i.max_context_length||""),10),i.description||""].join(" ")).join(`
+`):""}var Je={openai:{buildPayload:(e,t)=>({stream:!1,model:e,messages:[{role:"user",content:t}]}),parseResponse:e=>{if(e.error)throw new Error(S(e.error)||"Unknown error");if(e.choices&&e.choices[0]?.message?.content)return k(e.choices[0].message.content);throw new Error("Invalid response format")},buildUrl:e=>h(e,"v1/chat/completions")},anthropic:{buildPayload:(e,t)=>{let n={model:e,max_tokens:5128,messages:[{role:"user",content:t}]};return Y()&&(n.thinking={type:"enabled",budget_tokens:4096},n.max_tokens=16e3),r.deterministic&&!Y()&&Object.assign(n,{temperature:0,top_p:0,top_k:1}),n},parseResponse:e=>{if(e.content&&Array.isArray(e.content)){let t=[];for(let n of e.content)n.text&&t.push(n.text);if(t.length>0)return k(t.join(`
+`))}throw e.error?new Error(S(e.error)||"Unknown error"):new Error("Invalid response format")},buildUrl:e=>h(e,"v1/messages")},ollama:{buildPayload:(e,t)=>{let n={stream:!1,model:e};return r.apitype==="generate"?n.prompt=t:n.messages=[{role:"user",content:t}],Y()?r.think==="true"||r.think==="1"?n.think=!0:n.think=r.think:n.think=!1,r.deterministic&&(n.options={repeat_last_n:0,top_p:1,top_k:1,temperature:0,repeat_penalty:1,seed:123}),n},parseResponse:e=>{if(e.error)throw new Error(S(e.error));if(typeof e.response=="string")return k(e.response);if(e.message?.content)return k(e.message.content);throw new Error(JSON.stringify(e))},buildUrl:e=>h(e,"api/"+r.apitype)},gemini:{buildPayload:(e,t)=>{let n={contents:[{parts:[{text:t}]}]},o={};return r.deterministic&&Object.assign(o,{temperature:0,topP:1,topK:1}),r.think!==""&&(X()?o.thinkingConfig={thinkingBudget:0}:o.thinkingConfig={thinkingBudget:8192}),Object.keys(o).length>0&&(n.generationConfig=o),n},parseResponse:e=>{if(e.candidates&&e.candidates[0]?.content?.parts){let n=e.candidates[0].content.parts.filter(o=>!o.thought&&o.text).map(o=>o.text);if(n.length>0)return k(n.join(`
+`))}throw e.error?new Error(S(e.error)):(console.log(JSON.stringify(e)),new Error("Invalid response format"))},buildUrl:(e,t,n)=>h(e,`v1beta/models/${t}:generateContent?key=${n}`),requiresUrlApiKey:!0}},q={anthropic:{defaultModel:"claude-sonnet-5-5",defaultBaseUrl:"https://api.anthropic.com",authKey:"ANTHROPIC_API_KEY",keyName:"anthropic",authStyle:"anthropic",apiStyle:"anthropic"},claude:{defaultModel:"claude-sonnet-5-5",defaultBaseUrl:"https://api.anthropic.com",authKey:"ANTHROPIC_API_KEY",keyName:"claude",authStyle:"anthropic",apiStyle:"anthropic"},openai:{defaultModel:"gpt-4o-mini",defaultBaseUrl:"https://api.openai.com",authKey:"OPENAI_API_KEY",authStyle:"bearer",apiStyle:"openai"},ollama:{defaultModel:"qwen2.5-coder:latest",defaultBaseUrl:"http://localhost:11434",authStyle:"none",apiStyle:"ollama"},ollamacloud:{defaultModel:"gpt-oss:120b",defaultBaseUrl:"https://ollama.com",authKey:"OLLAMA_API_KEY",keyName:"ollamacloud",authStyle:"bearer",apiStyle:"ollama"},zen:{defaultModel:"big-pickle",defaultBaseUrl:"https://opencode.ai/zen",authKey:"OPENCODE_API_KEY",keyName:"zen",authStyle:"bearer",apiStyle:"openai",hardcodedModels:["big-pickle","glm-5","kimi-k2.5"]},gemini:{defaultModel:"gemini-2.5-flash",defaultBaseUrl:"https://generativelanguage.googleapis.com",authKey:"GEMINI_API_KEY",authStyle:"none",apiStyle:"gemini",hardcodedModels:["gemini-2.0-flash","gemini-2.0-flash-lite","gemini-2.5-pro","gemini-2.5-flash","gemini-2.5-flash-lite"]},mistral:{defaultModel:"codestral-latest",defaultBaseUrl:"https://api.mistral.ai",authKey:"MISTRAL_API_KEY",authStyle:"bearer",apiStyle:"openai"},xai:{defaultModel:"grok-beta",defaultBaseUrl:"https://api.x.ai",authKey:"XAI_API_KEY",authStyle:"bearer",apiStyle:"openai",hardcodedModels:["grok-2","grok-beta"]},lmstudio:{defaultModel:"local-model",defaultBaseUrl:"http://127.0.0.1:1234",authStyle:"none",apiStyle:"openai"},deepseek:{defaultModel:"deepseek-coder",defaultBaseUrl:"https://api.deepseek.com",authKey:"DEEPSEEK_API_KEY",authStyle:"bearer",apiStyle:"openai",hardcodedModels:["deepseek-coder","deepseek-chat"]},openrouter:{defaultModel:"openai/gpt-4o-mini",defaultBaseUrl:"https://openrouter.ai/api",authKey:"OPENROUTER_API_KEY",authStyle:"bearer",apiStyle:"openai"},groq:{defaultModel:"llama-3.3-70b-versatile",defaultBaseUrl:"https://api.groq.com/openai",authKey:"GROQ_API_KEY",authStyle:"bearer",apiStyle:"openai"}};function D(e){return q[e]}function Q(){return Object.keys(q)}function be(e){let t=D(e);if(!t)throw new Error(`Unknown provider: ${e}`);if(e==="mistral")return $e(t);switch(t.apiStyle){case"openai":case"anthropic":return Le(t);case"ollama":return Ge(t);case"gemini":return"";default:return""}}function X(){return r.think==="false"||r.think==="0"}function Y(){return r.think!==""&&!X()}function We(e,t,n){let o=e;return r.think!==""&&n==="openai"&&(X()?(o+=' Answers directly and concisely, without showing any thinking steps or internal reasoning. Never include phrases like "Let me think".',o+=" /no_think"):o="Think step by step and explain the reasoning process, When answering, first output your reasoning inside <think> and </think> tags, then give the final answer."+o),t?o:r.prompt+Ye()+o}function Ye(){return`
+.Translate the code into `+r.language+` programming language
+`}function ze(e,t,n,o){let i=r.model||e.defaultModel,a=We(n,o,e.apiStyle),s=me(e);if(t.requiresUrlApiKey&&s?.[1]&&e.authKey)return`Cannot read ~/.r2ai.${z(e.authKey)}-key`;let l=t.buildPayload(i,a),u=ye(e,s?.[0]||null);e.apiStyle==="ollama"&&u.push("Accept: application/x-ndjson");let v=ge(u),c=t.buildUrl(he(e),i,s?.[0]||void 0);try{return t.parseResponse(fe(c,v,JSON.stringify(l)))}catch(d){return"ERROR: "+d.message}}function ke(e,t){let n=D(r.api);return n?ze(n,Je[n.apiStyle],e,t):`Unknown value for 'decai -e api'. Available: ${Q().join(", ")}`}function Ve(){return Object.fromEntries(Object.entries(q).filter(([,e])=>e.authKey).map(([e,t])=>[e,t.authKey]))}function K(e,t){let n=r2.cmd("'%"+t).trim();if(!n.includes("=")&&n!=="")return[n.trim(),null,"env"];let o=e.toLowerCase(),i=N;if(_(i)){let s=r2.cmd("'cat "+i),l=le(s);if(o in l)return[l[o],null,"txt"]}let a="~/.r2ai."+o+"-key";if(_(a)){let s=r2.cmd("'cat "+a);return s===""?[null,"Cannot read "+a,"no"]:[s.trim(),null,"file"]}return[null,"Not available","nope"]}function xe(){H(w,N),r2.cmd("'ed "+N)}function Z(){Object.entries(Ve()).forEach(([e,t])=>{let n=K(e,t)[2];console.log(n,"	",e)})}function ve(e){let t=D(e);if(!t){console.error(`Unknown provider: ${e}`);return}try{let n=be(e);n&&console.log(n),t.hardcodedModels&&t.hardcodedModels.forEach(o=>console.log(o)),e==="mistral"&&console.log("ministral-8b-latest"),!n&&!t.hardcodedModels&&console.log(t.defaultModel)}catch(n){let o=n;console.error(`Error listing models for ${e}:`,o.message),console.log(t.defaultModel)}}var ee=!1;function Qe(e){return e==="true"||e==="1"}function P(e){let t=r;return{get:()=>t[e],set:n=>{t[e]=Qe(n)}}}function R(e){let t=r;return{get:()=>t[e],set:n=>{t[e]=n}}}function Xe(e){let t=e.trim();return t==="chat"?"chat":t==="generate"?"generate":null}var Ee={pipeline:{get:()=>r.pipeline,set:e=>{r.pipeline=e;try{r.decopipe=JSON.parse(r2.cmd("cat "+e))}catch(t){console.error(t)}}},model:{get:()=>r.model,set:e=>{e==="?"?ve(r.api):r.model=e.trim()}},deterministic:P("deterministic"),files:P("useFiles"),think:{get:()=>r.think||"false",set:e=>{r.think=e}},apitype:{get:()=>r.apitype,set:e=>{let t=Xe(e);t?r.apitype=t:console.error("Invalid apitype. Use chat or generate.")}},debug:P("debug"),http:{get:()=>r.http,set:e=>{let t=e.trim();t==="auto"||t==="r2"||t==="curl"?r.http=t:console.error("Invalid http backend. Use auto, r2 or curl.")}},timeout:{get:()=>r.timeout,set:e=>{r.timeout=Math.max(0,parseInt(e,10)||0)}},api:{get:()=>r.api,set:e=>{e==="?"?console.error(Q().join(`
+`)):r.api=e}},lang:R("language"),hlang:R("humanLanguage"),cache:P("cache"),cmds:R("commands"),tts:P("tts"),yolo:P("yolo"),prompt:R("prompt"),ctxfile:R("contextFile"),baseurl:R("baseurl"),headers:{get:()=>ue(r.extraHeaders),set:e=>{r.extraHeaders=W(e)}},maxtokens:{get:()=>r.maxInputTokens,set:e=>{r.maxInputTokens=parseInt(e,10)||-1}}};function Ze(){let e={host:r.host,port:r.port,lastOutput:r.lastOutput};Object.assign(r,E,e),r.decopipe={...E.decopipe},r.extraHeaders=[...E.extraHeaders]}function et(e){let t=e.trim();return!t||t.startsWith("#")?null:t.startsWith("decai -e ")?t.slice(9).trim():t}function x(e){let t=e.indexOf("="),n=t===-1?e:e.slice(0,t),o=t===-1?void 0:e.slice(t+1),i=Ee[n];if(!i){console.error("Unknown config key");return}typeof o<"u"?i.set(o):console.log(i.get())}function we(){Object.entries(Ee).forEach(([e,t])=>{let n=t.get();console.log("decai -e "+e+"="+n)})}function Ae(){if(Ze(),!_(U))return;let e=r2.call("cat "+U);for(let t of e.split(/\r?\n/)){let n=et(t);n&&n[0]!="#"&&x(n)}}function Oe(){ee||(Ae(),ee=!0)}function Se(){H(w,U),r2.cmd("'ed "+U),Ae(),ee=!0}function p(e,t,n=!1){let o=(t||"").replace(/`/g,""),i=e.replace(/'/g,"");if(r.api==="r2"||r.api==="r2ai"){let s=re(".pdc.txt");se(s,o);let l=i.startsWith("-")?i:["-i",s,i].join(" "),u=r.baseurl||r.host+":"+r.port,v=h(u,"cmd/"+encodeURIComponent(l)),c=O(v,[]);return c.error?`Error: ${c.error}`:c.result||JSON.stringify(c)||"Cannot curl, use r2ai-server or r2ai -w"}if(i.startsWith("-"))return"";let a=i+`:
+`+o;return r.maxInputTokens>0&&a.length>r.maxInputTokens&&(a=a.slice(0,r.maxInputTokens)),ke(a,n)}function Re(e){let t=r2.cmd("e scr.color");r2.cmd("e scr.color=0");try{return e()}finally{r2.cmd("e scr.color="+t)}}function Pe(){let e="",t=n=>e+=" "+b+" "+n+`
+`;e+="Usage: "+b+` (-h) ...
+`,e+="Version: "+te+`
+`,t("-a [query]    - solve query with auto mode"),t("-b [url]      - set base URL (alias for decai -e baseurl)"),t("-d [f1 ..]    - decompile given functions"),t("-dd [..]      - same as above, but ignoring cache"),t("-dD [query]   - decompile current function with given extra query"),t("-dr           - decompile function and its called ones (recursive)"),t("-e            - display and change eval config vars"),t("-E            - edit and run decai.txt"),t("-h            - show this help"),t("-H            - help setting up r2ai"),t("-i [f] [q]    - include given file and query"),t("-k            - list API key status"),t("-K            - edit apikeys.txt"),t("-m [model]    - use -m? or -e model=? to list the available models"),t("-n            - suggest better function name"),t("-p [provider] - same as decai -e api (will be provider)"),t("-q [text]     - query language model with given text"),t("-Q [text]     - query on top of the last output"),t("-r [prompt]   - change role prompt (same as: decai -e prompt)"),t("-R            - reset role prompt to default prompt"),t("-s            - function signature"),t("-v            - show local variables"),t("-V            - find vulnerabilities"),t("-x[*]         - eXplain current function (-x* for r2 script)"),r2.log(e.trim())}function j(e,t,n,o){if(n){let s=r2.cmd("anos").trim();if(s.length>0)return s}let i="";if(o){let s=r2.cmd("s");i+=`## Context functions:
+`;let l=r2.cmdAt("axff~^C[2]~$$",s);for(let u of l.split(/\n/g))i+=r2.cmd("pdc@"+u);r2.cmd("s "+s)}let a=t?" "+e:"";try{let s=e.slice(2).trim(),l="";r.contextFile!==""&&r2.cmd2("test -f "+r.contextFile).value===0&&(l+=`## Context:
+[START]
+`+r2.cmd("cat "+r.contextFile)+`
+[END]
+`);let{body:u,count:v}=Re(()=>{let d=`## Before:
+`,y=0;for(let g of r.commands.split(",")){if(g.trim()==="")continue;let C=t||s.trim().length===0?g:g+"@@= "+s,I=r2.cmd(C);I.length>5&&(d+="Output of "+g+`:
+[START]
+`+I+`
+[END]
+`,y++)}return d+=`## After:
+`,{body:d,count:y}});if(v===0){console.error("Nothing to do.");return}let c="";if(r.decopipe.use){let d=r.model;try{let y=r.decopipe[r.decopipe.default],g=l+u;for(let C of y.pipeline){C.model&&(r.model=C.model);let I=C.query+". "+y.globalQuery;c=p(I,g,!0),r.debug&&console.log(`QUERY
+`,I,`
+INPUT
+`,g,`
+OUTPUT
+`,c),g=c}c=g}finally{r.model=d}}else{let d=a;l+=u+i,c=p(d,l,!1),r.lastOutput=c}return n&&c.length>1&&r2.call("ano=base64:"+$(c)),c.startsWith("```")&&(c=c.replace(/```.*\n/,"").replace(/```$/,"")),c.trim()}catch(s){let l=s;console.error(l,l.stack);return}}function tt(){let e=Re(()=>"[START]"+r.commands.split(",").map(o=>r2.cmd(o)).join(`
+`)+"[END]"),n=p("Analyze function calls, references, comments and strings, loops and ignore registers and memory accesses. Explain the purpose of this function in a single short sentence. /no_think Do not introduce or argue the response, translation of the explanation in "+r.humanLanguage,e,!0).trim().split(/\n/g);return n[n.length-1].trim()}function nt(){let e=r.language,t=r2.cmd("afv;pdc");r.language="C";let n="'afs "+p("analyze the uses of the arguments and return value to infer the signature, identify which is the correct type for the return. Do NOT print the function body, ONLY output the function signature, like if it was going to be used in a C header",t,!1),o=n.indexOf("{");return o!==-1&&(n=n.substring(0,o)),r.language=e,n}function rt(e,t){let n=[];for(;;){let o=L.auto;if(n.length>0){o+=`## Command Results
+
+`;for(let a of n){let s=JSON.parse(a);o+="### "+s.command+"\n\n```\n"+s.response+"\n```\n"}}o+=`
+
+## User Prompt
+
+`+e,r.debug&&console.log(`#### input
+`,o,`
+#### /input`),console.log("Thinking...");let i=p("",o,!0);r.debug&&console.log(`#### output
+`,i,`
+#### /output`);try{let a=JSON.parse(ae(ie(k(i))));if(a.action==="r2cmd"||a.action==="response"||a.action===a.command){let s=a.command||"";a.reason&&(console.log("[r2cmd] Reasoning: "+a.reason),r.tts&&(r2.syscmd("pkill say"),r2.syscmd("say -v Alex -r 250 '"+a.reason.replace(/'/g,"")+"' &"))),console.log("[r2cmd] Action: "+a.description),console.log("[r2cmd] Command: "+s);let l=s;r.yolo||(l=ot(s,t)),console.log("[r2cmd] Running: "+l);let u=r2.cmd2(l),v=u.logs?u.logs.map(y=>y.type+": "+y.message).join(`
+`):"",c=(u.res+v).trim();console.log(c);let d=oe(c);r.debug&&console.log(`<r2output>
+`,d,`
+<(r2output>`),n.push(JSON.stringify({action:"response",command:l,description:a.description,response:d}))}else if(a.action==="reply"){console.log(`Done
+`,a.response);break}else{console.log(`Unknown response
+`,JSON.stringify(i));break}}catch(a){let s=i.indexOf('response": "');if(s!==-1){let l=i.slice(s+12).replace(/\\n/g,`
+`).replace(/\\/g,"");console.log(l)}else console.log(i),console.error(a);break}}}function ot(e,t){for(;;){let n=r2.cmd("'?ie Tweak command? ('?' for help)").trim();if(n==="q!"){console.error("Break!");break}if(n==="?"){it();continue}else if(n.startsWith(":")){console.log(r2.cmd(n.slice(1)));continue}else if(n.startsWith("-e")){t(n);continue}else{if(n==="!")return"?e do NOT execute '"+e+"' again, continue without it";if(n.startsWith("!")){console.log(r2.syscmd(n.slice(1)));continue}else{if(n==="q")return"?e All data collected!. Do not call more commands, reply the solutions";if(n){let o=n.indexOf("#");return o!==-1?n.slice(0,o).trim():n}else return e}}}return e}function it(){console.log(" '!'     do not run this command"),console.log(" '!c'    run system command"),console.log(" 'q'     to quit auto and try to solve"),console.log(" 'q!'    quit auto without solving"),console.log(" 'c # C' use given command with comment"),console.log(" ':c'    run r2 command without feeding auto"),console.log(" '-e'    set decai configuration variables")}function at(e){return e.slice(2).trim()}function B(){return r2.cmd("afi.").trim().length>0}var F="Cannot find function at current offset";function Ce(e,t){if(e===""||!e.startsWith("-")){Pe();return}let n="",o=e[1],i=at(e);switch(o){case"H":console.log(L.decai);break;case"a":rt(i,t);break;case"m":{i==="="?x("model="):i?x("model="+i):x("model");break}case"n":case"f":{if(!B()){n=F;break}n=r2.cmd("axff~$[3]");let a=r2.cmd("fd.").trim().split(/\n/).filter(s=>!s.startsWith("secti")).join(",");n=p("give me a better name for this function. the output must be: 'afn NEWNAME'. do not include the function code, only the afn line. consider: "+a,n,!1).trim(),n+=" @ "+r2.cmd("?v $FB").trim();break}case"v":{if(!B()){n=F;break}n=r2.cmd("afv;pdc"),n=p("guess a better name and type for each local variable and function argument taking using. output an r2 script using afvn and afvt commands",n,!1);break}case"i":{let a=i.split(/\s+/).filter(Boolean);if(a.length>=2){let s=a[0],l=a.slice(1).join(" "),u=r2.cmd("cat "+s);console.log(p(l,u,!0))}else console.log("Use: decai -i [file] [query ...]");break}case"p":{i?x("api="+i):Z();break}case"r":{i?r.prompt=i:console.log(r.prompt);break}case"R":r.prompt=T;break;case"s":if(!B()){n=F;break}n=nt();break;case"V":n=p("find vulnerabilities, dont show the code, only show the response, provide a sample exploit",r.lastOutput,!1);break;case"K":xe();break;case"E":Se();break;case"k":Z();break;case"b":{i?x("baseurl="+i):console.log(r.baseurl);break}case"e":{i?x(i):we();break}case"q":try{n=p(i,null,!0)}catch(a){let s=a;console.error(s,s.stack)}break;case"Q":n=p(i,r.lastOutput,!1);break;case"x":if(!B()){n=F;break}n=tt(),(e[2]==="*"||e[2]==="r")&&(n="'CC "+n);break;case"d":e[2]==="r"?n=j(e.slice(2),!0,r.cache,!0)||"":e[2]==="d"?n=j(e,!1,!1,!1)||"":e[2]==="D"?n=j(e,!0,!1,!1)||"":n=j(e,!1,r.cache,!1)||"";break;default:Pe();break}return n||void 0}function Ie(e){Oe();let t=Ce(e,Ie);return t&&r2.log(t),!0}function st(){r2.unload("core",b),r2.plugin("core",function(){function e(t){if(t.startsWith(b)){let n=t.slice(b.length).trim();return Ie(n)}return!1}return{name:b,license:"MIT",desc:"r2 decompiler based on r2ai",call:e}})}st();

decai/dist/debian/CONFIG

@@ -0,0 +1,13 @@
+PACKAGE=decai
+DEPENDS=radare2
+SECTION=user/shell
+PRIORITY=optional
+MAINTAINER=pancake <[email protected]>
+VERSION=$(shell git tag | tail -n 1)
+# arch
+UNAMEM=$(shell uname -m)
+ifeq ($(UNAMEM),x86_64)
+ARCH=amd64
+else
+ARCH=arm64
+endif

decai/dist/debian/DESCR

@@ -0,0 +1,2 @@
+AI based decompiler plugin for radare2
+r2js plugin for radare2 that provides several features for reverse engineers, from autonaming functions, improve decompiling outputs, find bugs or explain functions to automated crackme solving using local and remote models if API keys are provided..

decai/dist/debian/Makefile

@@ -0,0 +1,34 @@
+include ./CONFIG
+
+UNAME=$(shell uname)
+SUDO?=sudo
+DEPENDS=
+CROSSARCH=x64
+R2CFG_FLAGS?=
+PWD=$(shell pwd)
+PACKAGE_DIR?=${PWD}
+
+R2_VERSION=$(shell r2 -qv)
+
+DOCKCROSS=$(PWD)/../dockcross
+R2PLUGDIR=/usr/lib/radare2/$(R2_VERSION)
+
+all: root
+	$(SUDO) rm -rf control data
+	$(MAKE) clean
+	mkdir -p data
+	cp -rf root/* data
+	$(MAKE) control
+	$(MAKE) deb
+
+root:
+	mkdir -p root/$(R2PLUGDIR)
+	cp -f ../../decai.r2.js root/$(R2PLUGDIR)
+
+purge: clean
+	rm -rf root
+
+summary:
+	echo $(VERSION)
+
+include deb.mk

decai/dist/debian/deb.mk

@@ -0,0 +1,130 @@
+# Create .deb without using dpkg tools.
+#
+# Author: Tim Wegener <[email protected]>
+#
+# Use 'include deb_hand.mak' after defining the user variables in a local
+# makefile.
+#
+# The 'data' rule must be customised in the local make file.
+# This rule should make a 'data' directory containing the full file
+# layout of the installed package.
+#
+# This makefile will create a debian-binary file a control directory and a
+# a build directory in the current directory.
+# Do 'make clobber' to remove these generated files.
+#
+# Destination:
+# PACKAGE_DIR - directory where package (and support files) will be built
+#               defaults to the current directory
+#
+# Sources:
+# SOURCE_DIR - directory containing files to be packaged
+# ICON_SOURCE - 26x26 icon file for maemo
+# DESCR - description with summary on first line
+# preinst, postinst, prerm, postrm - optional control shell scripts
+
+# These fields are used to build the control file:
+# PACKAGE = 
+# VERSION = 
+# ARCH = 
+# SECTION = 
+# PRIORITY = 
+# MAINTAINER = 
+# DEPENDS = 
+#
+# SOURCE_DIR =
+# ICON_SOURCE = 
+# (ICON_SOURCE is optional)
+
+# *** NO USER CHANGES REQUIRED BEYOND THIS POINT ***
+ifeq ($(shell uname),Darwin)
+MD5SUM=md5
+else
+MD5SUM=md5sum
+endif
+
+GAWK=awk
+PACKAGE_DIR=$(shell pwd)
+CONTROL_EXTRAS ?= ${wildcard preinst postinst prerm postrm}
+
+${PACKAGE_DIR}/control: ${PACKAGE_DIR}/data ${CONTROL_EXTRAS} DESCR \
+			${ICON_SOURCE}
+	#rm -rf $@
+	mkdir -p $@
+ifneq (${CONTROL_EXTRAS},)
+	cp ${CONTROL_EXTRAS} $@
+endif
+#       Make control file.
+	echo "Package: ${PACKAGE}" > $@/control
+	echo "Version: ${VERSION}" >> $@/control
+	echo "Section: ${SECTION}" >> $@/control
+	echo "Priority: ${PRIORITY}" >> $@/control
+	echo "Architecture: ${ARCH}" >> $@/control
+ifneq (${DEPENDS},)
+	echo "Depends: ${DEPENDS}" >> $@/control
+endif
+	echo "Installed-Size: ${shell du -s ${PACKAGE_DIR}/data|cut -f1}" \
+		>> $@/control
+	echo "Maintainer: ${MAINTAINER}" >> $@/control
+	printf "Description:" >> $@/control
+	cat DESCR | ${GAWK} '{print " "$$0;}' >> $@/control
+#ifneq (${ICON_SOURCE},)
+#	echo "Maemo-Icon-26:" >> $@/control
+#	base64 ${ICON_SOURCE} | ${GAWK} '{print " "$$0;}' >> $@/control
+#endif
+#       Make md5sums.
+	cd ${PACKAGE_DIR}/data && find . -type f -exec ${MD5SUM} {} \; \
+		| sed -e 's| \./||' \
+		> $@/md5sums
+
+${PACKAGE_DIR}/debian-binary:
+	echo "2.0" > $@
+
+${PACKAGE_DIR}/clean:
+	rm -rf ${PACKAGE_DIR}/data ${PACKAGE_DIR}/control ${PACKAGE_DIR}/build *.deb
+
+${PACKAGE_DIR}/build: ${PACKAGE_DIR}/debian-binary ${PACKAGE_DIR}/control \
+			${PACKAGE_DIR}/data
+	rm -rf $@
+	mkdir $@
+	cp ${PACKAGE_DIR}/debian-binary $@/
+	cd ${PACKAGE_DIR}/control && tar czvf $@/control.tar.gz *
+	cd ${PACKAGE_DIR}/data && \
+		COPY_EXTENDED_ATTRIBUTES_DISABLE=true \
+		COPYFILE_DISABLE=true \
+		tar cpzvf $@/data.tar.gz *
+
+# Convert GNU ar to BSD ar that debian requires.
+# Note: Order of files within ar archive is important!
+${PACKAGE_DIR}/${PACKAGE}_${VERSION}_${ARCH}.deb: ${PACKAGE_DIR}/build
+	ar -rc $@ $</debian-binary $</control.tar.gz $</data.tar.gz
+	#sed -e 's|^\([^/]\+\)/ \(.*\)|\1  \2|g' $@tmp > $@fail
+	#rm -f $@tmp
+	#mv $@fail $@
+
+.PHONY: data
+data: ${PACKAGE_DIR}/data
+
+.PHONY: control
+control: ${PACKAGE_DIR}/control
+
+.PHONY: build
+build: ${PACKAGE_DIR}/build
+
+.PHONY: clean
+clean: ${PACKAGE_DIR}/clean $(EXTRA_CLEAN)
+	rm -f debian-binary
+
+.PHONY: deb
+deb: ${PACKAGE_DIR}/${PACKAGE}_${VERSION}_${ARCH}.deb
+
+
+clobber::
+	rm -rf ${PACKAGE_DIR}/debian_binary ${PACKAGE_DIR}/control \
+		${PACKAGE_DIR}/data ${PACKAGE_DIR}/build
+
+push:
+	scp *.deb radare.org:/srv/http/radareorg/cydia/debs
+
+mrproper: clean
+	rm -rf root

decai/pipeline.json

@@ -0,0 +1,37 @@
+{
+  "use": true,
+  "default": "pipeTest",
+  "_default": "asciiStack",
+  "asciiStack": {
+    "globalQuery": "output in C, respond only with code, no explanations, no markdown",
+    "pipeline": [
+      {
+        "model": "qwen2.5-coder:latest",
+        "query": "stack frame layout for this function, output a table in ascii art"
+      }
+    ]
+  },
+  "pipeTest": {
+    "globalQuery": "output in C, respond only with code, no explanations, no markdown",
+    "pipeline": [
+      {
+        "model": "qwen2.5-coder:latest",
+        "query": "resolve stack variables and arrays size without using malloc"
+      },
+      {
+        "model": "qwen2.5-coder:14b",
+        "query": "remove boilerplate and unnecessary code, replace goto with if/else blocks"
+      }
+    ],
+    "commentedOut": [
+      {
+        "model": "qwen2.5-coder:14b",
+        "query": "use better types and names for variable and arguments"
+      },
+      {
+        "model": "qwen2.5-coder:14b",
+        "query": "inline all the functions into a single one and make the code more readable"
+      }
+    ]
+  }
+}