Project Hermit - Autonomous Self-Evolving Multi-Agent System & Dynamic MCP Tool Host

Project Hermit is a Mobile-native, closed-loop autonomous self-optimization engine and dynamic Model Context Protocol (MCP) tool host. Operating as a background daemon (hermit_daemon.py), Hermit continuously discovers, profiles, refactors, and tests Python payload routines inside isolated RAM-disk sandbox environments (sandbox.py) using LLM-driven code synthesis combined with deterministic mathematical and AST mutators (math_mutator.py, qubo_mutator.py).

Project Hermit resolves five critical operational failures common to long-horizon autonomous AI systems:

1. Queue Starvation (Phase 1 vs. Phase 2 Dual-Phase Scheduling): Prevents high-latency functions from monopolizing mutation cycles by alternating between exploration of zero-merge functions sorted by optimization headroom ($H = \text{latency} \times \text{complexity}$) and exploitation of bottleneck functions with concurrency caps (MAX_CONSECUTIVE = 3).
2. Strategy Oscillation (6-Merge Sequence Pattern Banning): Replaces simplistic count-based strategy banning with sequence pattern detection (A-B-A-B or A-B-C-A-B-C over a 6-merge window), eliminating false-positive bans on newly merged successful code.
3. RAM Measurement Distortion (Fork-Based Grandchild Isolation): Captures exact, non-cumulative peak RSS per candidate run, eliminating historical RUSAGE_CHILDREN cumulative noise.
4. Thermal & Quota Resilience: Predictively throttles execution based on CPU temperature trend slopes (/sys/class/thermal/) and enforces pacing delay guards between outbound LLM API requests.
5. Host Protection & Introspection Rollback (IMMUTABLE_FUNCTIONS): Protects core host looper functions (get_next_target, sandbox_run, detect_oscillation) from self-mutation and automatically rolls back failed introspection patches via versioned database snapshots (.hermit_backup_v*).

Project Hermit was engineered through Human-AI Collaboration between the developer and the Google Antigravity CLI:

• Development Environment & Host Hardware: Built and benchmarked entirely on a Samsung Galaxy A36 running Termux inside a PRoot virtualized Linux environment.
• Human-AI Collaboration Model:
  • Developer: Directed project goals, identified mobile hardware constraints (thermal throttling on ARM/mobile devices, queue starvation, memory reporting issues), defined system requirements, and guided architectural iterations.
  • Google Antigravity CLI: Served as the AI pair programmer in Termux/PRoot to implement AST mutator engines, write verification test suites, refactor core engines, and generate full repository documentation artifacts.

graph TD
    A[HermitDaemon - Dual-Phase Scheduler] -->|1. Select Target Skill| B[StaticMCPHost / Orchestrator]
    B -->|2. Check Quota & Pacing| C[API Key & Pacing Guard]
    B -->|3. Synthesize Mutation| D[Mutator Engines: Math / QUBO / LLM]
    D -->|4. Generate Candidate Code| E[Sandbox Execution Harness]
    E -->|5. RAM Disk Isolation tmpfs| F[AST Code Safety Analyzer]
    F -->|6. Child Fork Execution| G[Verification & Benchmark Harness]
    G -->|7. Multi-Objective Score| H[Pareto Evaluation & Baseline Hash Check]
    H -->|Pass: Merge Candidate| I[SQLite hermit_memory.db & Hot Reload]
    H -->|Fail: Log Anti-Pattern| J[Anti-Pattern Store & Banned Strategies]

All detailed documentation, architectural specs, threat models, and research ledgers generated under the Repository Reconstruction Roadmap (v1.0) standard are located in the docs directory.

cd project-hermit-git/v0.1.9
python3 -m venv venv
source venv/bin/activate
pip install google-genai psutil

# Export your Gemini API key
export GEMINI_API_KEY="your-actual-api-key"

# Run structural host verification test
python3 orchestrator.py --mode shadow_test

# Run main integration test suite (17/17 passing)
python3 test_hermit.py

python3 hermit_daemon.py

python3 monitor_evolution.py

This project is licensed under the Apache License 2.0 - see the LICENSE file for details.