AdevX

A practical starter chatbot that can handle many tasks with tools.

It includes:

• Chat loop with memory
• Interactive "AdevX is thinking..." waiting animation
• Multi-mode assistant: chat, coding, image, research, and agent
• Optional online tool-calling mode with:
  • OpenAI (OPENAI_API_KEY)
  • OpenRouter (OPENROUTER_API_KEY, supports openrouter/free)
  • Groq (GROQ_API_KEY)
  • Together (TOGETHER_API_KEY)
  • Ollama local (ollama-local, no cloud key needed)
• Offline local minimal-reasoning mode (no API key required)
• Automatic provider failover chain (smart routing + fallback)
• Phase 2 RAG: project indexing + retrieval context injection
• Modular runtime RAG: incremental semantic indexing + hybrid retrieval (BM25 lexical + sparse semantic similarity + symbol boosts)
• Repository intelligence: AST-aware symbols, imports, call graph, and references
• Agent commands: planning, execution, and review from the terminal
• Git intelligence: repository analysis, commit summaries, and impact reports
• Memory architecture: stats, search, and consolidation over local memory
• Local benchmark + metrics commands for retrieval and command health
• Production hardening:
  • provider error secret redaction
  • private/local URL fetch blocking by default
  • safer shell command blocking
  • atomic local JSON writes
  • smarter provider retry/failover behavior
  • bounded agent execution timeout
• Built-in tools for:
  • file listing/reading/writing/appending/searching
  • safe math calculation
  • image metadata analysis (PNG/JPG/GIF/BMP/WEBP)
  • URL fetching
  • shell command execution (interactive approval)
  • text summarization

Quick Start

python taskbot.py

Modular architecture preview runtime:

python -m adevx.main

The modular runtime now uses live provider calls (OpenAI-compatible HTTP APIs) instead of scaffold echo responses, with retry + circuit-breaker routing. It also runs background incremental index refresh for workspace retrieval quality.

Autonomous reasoning docs:

• docs/ARCHITECTURE_REPORT.md
• docs/AUTONOMOUS_ENGINE.md

Online Mode Setup

Use one provider key, then run AdevX.

OpenAI:

$env:OPENAI_API_KEY="your_key"
python taskbot.py --provider openai --model gpt-4.1-mini

OpenRouter (free router):

$env:OPENROUTER_API_KEY="your_key"
python taskbot.py --provider openrouter --model openrouter/free

Groq (free plan, rate-limited):

$env:GROQ_API_KEY="your_key"
python taskbot.py --provider groq

Together AI:

$env:TOGETHER_API_KEY="your_key"
python taskbot.py --provider together

Ollama local (fully offline):

ollama pull qwen2.5:7b
python taskbot.py --provider ollama-local --model qwen2.5:7b

For demos without API credits, you can also let AdevX auto-detect Ollama:

ollama pull qwen2.5:7b
python taskbot.py
/local

Smart Routing + Fallback

• Default chain: openai -> groq -> openrouter -> together -> ollama-local
• If one provider fails, AdevX automatically tries the next provider.
• If cloud credits/rate limits fail, AdevX tries an installed Ollama model before showing raw RAG context.
• Raw RAG snippets are only shown when no cloud provider or local model is available.
• Configure chain:

$env:ADEVX_PROVIDER_CHAIN="openrouter,groq,together,ollama-local"

• Disable smart routing:

$env:ADEVX_SMART_ROUTING="0"

Offline Commands

Slash commands always work (even while online):

• /h
• /help
• /remember <note>
• /memory
• /memory stats|search <query>|consolidate
• /forget
• /about
• /models
• /use provider:model
• /autotune [max_latency_seconds]
• /speed fast|balanced|quality
• /health [timeout_seconds]
• /local
• /ollama status|models
• /modes
• /mode <name>
• /image <path>
• /rag status|rebuild|query <text>|on|off
• /repo symbols|graph|explain <symbol>|references <symbol>
• /agent plan|execute|review <text>
• /git analyze|summarize [rev]|impact [path|rev-range]
• /benchmark
• /metrics
• /phase run|status
• /status
• /online
• /offline
• /ls [path]
• /read <path>
• /write <path> <content>
• /append <path> <content>
• /search <query> [path]
• /calc <expression>
• /fetch <url>
• /shell <command>
• /summarize <text>
• /exit

Offline Natural Language (No API)

In offline mode, AdevX also supports simple plain-text tasks like:

• create a file named hello.txt with text: hi
• read hello.txt
• list files
• calculate sqrt(144) + 6
• search hello in .
• summarize <your text>
• remember my project is BChat

Model switch examples:

• /models
• /use groq:llama-3.1-8b-instant
• /use openrouter:openrouter/free
• /use ollama-local:qwen2.5:7b
• /local
• /ollama status
• /ollama models
• /autotune 15
• /speed fast
• /mode coding
• /mode image
• /image assets/logo.png
• /rag rebuild
• /rag query merge sort implementation
• /repo explain WorkspaceIndexAdapter
• /memory search provider routing
• /agent plan implement auth middleware
• /git analyze
• /phase run

Phase automation:

/phase run
/phase status

/phase run currently automates:

• RAG rebuild
• local model autotune (if available)
• speed profile optimization
• capability benchmark scoring (internal heuristic)

RAG quick start:

/rag rebuild
/rag status
/rag query auth middleware

One-shot mode

python taskbot.py --once "/calc sqrt(144) + 6"

Disable animation in interactive mode:

python taskbot.py --no-animation

For slow local models, increase local timeout:

$env:ADEVX_OLLAMA_TIMEOUT="240"

For longer local code answers:

$env:ADEVX_OLLAMA_MAX_TOKENS="1200"

For long autonomous tasks, tune the agent timeout:

$env:ADEVX_AGENT_TIMEOUT="180"

Private/local URL fetches are blocked by default for safety. Enable only when you trust the target:

$env:ADEVX_ALLOW_PRIVATE_URL_FETCH="1"

Production Reports

• docs/PRODUCTION_AUDIT.md
• docs/PRODUCTION_READINESS_REPORT.md

Tests

Run core regression tests:

python -m unittest discover -s tests -p "test_*.py" -v

Notes

• This bot is a strong foundation, but no system can literally do "any task" without limits.
• Quality depends heavily on the model/provider you choose.
• In offline mode, default behavior is rule-based (no cloud token billing).
• If Ollama is running locally, offline mode can also use a local LLM for broader reasoning.
• File operations are restricted to the current workspace for safety.
• Shell commands always require your approval before running.

Author

Aditya Pal