Skip to content

ARCHITECTURE.md

Latest commit

 

History

History
36 lines (28 loc) · 2.96 KB

ARCHITECTURE.md

File metadata and controls

36 lines (28 loc) · 2.96 KB

CodeMechanic-Bot Architecture

CodeMechanic-Bot uses an event-driven, multi-agent architecture. The system is designed around a central EventBus in orchestrator.py that fully decouples the agents, allowing them to scale and fail independently.

The Event Pipeline

When you run python orchestrator.py, the orchestrator enters an infinite loop. Every 30 minutes, it triggers the BountyRadar to start a scan cycle.

  1. BountyRadar: Queries GitHub for issues with specific labels (bounty, algora, polar). It filters out saturated issues and emits a BOUNTY_FOUND event.
  2. ScamDetector: Listens for BOUNTY_FOUND. It queries the repository's statistics. If the repo has < 5 stars, 0 historically merged PRs, or > 50 open issues, it kills the pipeline for that issue. Otherwise, it emits BOUNTY_VERIFIED.
  3. StaticAnalyzer: Runs continuously in the background to proactively audit popular repositories (via docker). Runs Semgrep (logic bugs), Trivy (IaC and CVEs), and Gitleaks (secrets). Emits BOUNTY_VERIFIED for zero-days found.
  4. PREngineer: Listens for BOUNTY_VERIFIED.
    • Uses the GitHub API to post a "Comment First" message proposing to work on the issue.
    • Clones the repository locally.
    • Harvests context (tests, CONTRIBUTING.md, recent commits).
    • Queries the local AI (gemma4:e4b, with fallbacks to llama3 and mistral) to generate a targeted patch.
    • Tests the patch in a sandboxed Docker container (with daemon reconnection resilience).
    • Emits PR_READY.
  5. CodeReviewer: Listens for PR_READY.
    • Has the AI self-review the generated patch for security and style.
    • If approved, it uses the GitHub CLI (gh) and robust git branching (checkout -B) to fork the repository, commit the patch, and open a Pull Request.
    • Emits PR_SUBMITTED.
  6. ContentEngine / DevOpsMonitor / EarningsTracker: All listen for PR_SUBMITTED in parallel.
    • ContentEngine: Writes a markdown blog post detailing the automated fix.
    • DevOpsMonitor: Simulates tracking the CI/CD pipeline of the PR.
    • EarningsTracker: Maintains a local ledger of estimated ROIs.
  7. ReviewTracker: Runs continuously alongside the radar. It polls open PRs authored by the bot. If a maintainer leaves a review requesting changes, it emits PR_REVIEWED, which routes back to the PREngineer to iteratively fix the code.

Premium Web Dashboard

The backend communicates natively with a lightweight FastAPI server (api/main.py). The frontend (ui/) uses a beautiful Catppuccin Mocha theme with glassmorphism to control the bot. It embeds CodeMirror (with Vim mode enabled) so humans can intercept and review PRs manually if manual_approval is required in the config.

Config & Model Fallbacks

The system behavior is defined in config.yaml. Crucially, the bot implements a "Model Fallback" system. If the primary model fails or times out, it automatically tries the next models in the fallback array to ensure 24/7 uptime.