CODEY-V2

  ██████╗ ██████╗ ██████╗ ███████╗██╗   ██╗
 ██╔════╝██╔═══██╗██╔══██╗██╔════╝╚██╗ ██╔╝
 ██║     ██║   ██║██║  ██║█████╗   ╚████╔╝
 ██║     ██║   ██║██║  ██║██╔══╝    ╚██╔╝
 ╚██████╗╚██████╔╝██████╔╝███████╗   ██║
  ╚═════╝ ╚═════╝ ╚═════╝ ╚══════╝   ╚═╝  ─ V2
  v2.0.0 · Local AI Coding Assistant · Termux

A persistent, daemon-based AI coding agent that runs entirely on your Android device. Codey-v2 maintains state across sessions, manages a background task queue, and uses three purpose-built models — a 7B primary agent, a 0.5B planner and summarizer, and a dedicated embedding encoder — all served locally via llama.cpp.

Security notice: Codey executes shell commands and writes files based on model output. Read the security guide before use.

---

Quick Start

./install.sh          # Install everything (models, llama.cpp, PATH)
codeyd2 start         # Start all three model servers and the daemon
codey2 "your task"    # Send a task
codeyd2 status        # Check daemon health at any time

See docs/installation.md for manual setup and model download links.

---

What's New in v2.7.1

Peer CLI Delegation — Fully Repaired

The end-to-end path for delegating work to Claude, Gemini, or Qwen is now reliable:

• "Ask Claude to X" no longer gets intercepted by plannd — a peer directive regex gate in main.py bypasses the planner so the original request reaches the agent intact.
• Claude returns code Codey can apply — every delegation prompt now includes the current project files and explicit output format instructions (**\filename.py`**` + fenced code blocks). Codey extracts and writes these files automatically.
• No more permission prompts from Claude — the prompt now states upfront that Claude is responding to an automated system and must act immediately without asking clarifying questions.

Shell Safety — Consent Model Replaces Blocklist

The hard block on shell metacharacters (&&, |, ;, 2>&1, etc.) has been removed. All commands now flow through a user confirmation prompt instead. Dangerous commands (rm, curl, wget, etc.) get an explicit warning before the prompt. YOLO mode (--yolo) skips all confirmations.

Other Fixes

• Malformed JSON tool calls now trigger an explicit retry instead of being silently dropped.
• max_steps raised from 6 → 10 to handle multi-file tasks without hitting the step cap.
• Retry context now includes the failed filename and previous result so the agent doesn't repeat the wrong step.
• Planner can no longer invent function arguments or test values not mentioned in the user's request.

---

What's New in v2.7.0

Smarter Context Management

Long sessions no longer degrade. Context compression has been completely reworked:

	Before	After
Trigger threshold	75% of context window	55% — fires before things get tight
After compression	Left wherever it landed	Drops to 40% — real headroom restored
Message truncation	content[:300] before summarizing	Removed — full content passed to summarizer
Summarizer model	7B (same model doing your work)	0.5B on port 8081 — fast, independent
What gets summarized	Everything old, one flat pass	Only dropped turns — pinned messages survive
Re-summarization	Could summarize a summary	Blocked — existing summaries are pinned

Pinned messages (never dropped): file writes, patches, errors, shell results, and existing summaries. The 0.5B call is best-effort — if port 8081 is unreachable, the drop still happens and the agent keeps working.

Planner Timeout Increase

The planning call timeout has been raised from 45 s → 180 s, with the HTTP timeout set to 165 s so the network call always resolves cleanly before the outer timeout fires.

---

Capabilities

Three-Model Architecture

Model	Port	Role
Qwen2.5-Coder-7B Q4_K_M	8080	Primary agent — coding, reasoning, tool use
Qwen2.5-0.5B Q8_0	8081	Task planning and conversation summarization
nomic-embed-text-v1.5 Q4	8082	RAG retrieval encoder

All three run as independent llama-server processes, managed and watchdog-monitored by codeyd2.

Agent Features

• Persistent daemon — runs continuously in the background; state survives restarts
• Task queue — complex requests broken into steps and executed sequentially
• RAG retrieval — local knowledge base searched on every inference call; relevant docs injected automatically
• Recursive self-refinement — draft → critique → refine cycle catches bugs before they hit your files
• Error recovery — adaptive strategy switching when tools fail (write → patch, import error → install, etc.)
• Peer CLI escalation — delegates work to Claude Code, Gemini CLI, or Qwen CLI either on-demand ("ask Claude to X") or automatically when Codey exhausts its retry budget. The peer receives current project file contents and returns complete, ready-to-apply code blocks that Codey writes to disk. Requires explicit user consent before any files are shared (external services — see Security)
• Git integration — branch management, AI commit messages, conflict detection and resolution
• Voice interface — TTS output and STT input via Termux:API
• Static analysis — auto-lint on every Python write; /review command for on-demand scans
• Thermal management — monitors CPU load and battery; reduces threads automatically under stress
• Fine-tuning — export your interaction history and train a personalized adapter on Google Colab

---

Documentation

Guide	Contents
Installation	Requirements, one-line install, manual step-by-step
Commands	Full reference: codeyd2, codey2, slash commands, flags, env vars
Configuration	Config JSON, model tuning, context management, thermal settings
Architecture	System diagram, memory tiers, project structure, Python API
Knowledge Base	Setting up RAG, indexing docs, skill repos
Fine-tuning	Export data, Colab training, import adapter, rollback
Pipeline	Training data pipeline — build fine-tuning datasets from HuggingFace + synthetic data
Security	Risks, mitigations, hardening summary, reporting vulnerabilities
Troubleshooting	Common issues, performance reference, known limitations
Version History	Full changelog from v1.0.0 through v2.7.0

---

Requirements

Platform	Termux on Android, or any Linux system
RAM	6 GB+ available
Storage	~6 GB base (7B model ~4.2 GB, 0.5B ~500 MB, embed ~80 MB, toolchain ~1 GB); ~8 GB with training pipeline
Python	3.12+

---

Contributing

1. Fork the repository
2. Create a feature branch
3. Make your changes and run the tests (pytest tests/ -v)
4. Submit a pull request

Bug reports, security disclosures, and hardening contributions are especially welcome.

---

Acknowledgments

• llama.cpp — efficient on-device LLM inference
• Qwen — Qwen2.5-Coder models
• nomic-ai — nomic-embed-text embedding model
• Codey v1 — the original session-based agent this builds on

---

MIT License