GITPILOT-MCP---AI-POWERED-GITHUB-TOOL-CONTEXT-LAYER: Instructions file for Claude Code

GitPilot — Instructions for Claude Code

GitPilot is a stateful AI engineering control plane for GitHub. Full architecture lives in docs/ARCHITECTURE.md and the connection/brain/ guided-fix pipeline in docs/BRAIN.md — read those before making structural changes. Don't restate their content here; this file is about how to work in this repo, not what it does.

Source of truth

• Code correctness: the code itself and python -m pytest -q, never a doc.
• Architecture: docs/ARCHITECTURE.md, docs/BRAIN.md.
• Extension boundary: services/contracts.py — any new tracker, code host, memory backend, or AI provider implements these Protocols; don't bypass them.
• Fix-run state machine: services/memory_service.py's RUN_TRANSITIONS. Don't invent new run states outside it.

Where solved issues and patterns go

This repo has a real, durable memory system — don't create parallel markdown "brain files" for bug patterns, API quirks, or RCA notes. Use the existing tools:

• memory_remember(repository, kind, title, content, tags) — store a solved pattern, gotcha, or root cause once it's confirmed (kind="pattern" or kind="gotcha").
• memory_recall(repository, query) — check before re-diagnosing something that may already be solved.

This keeps knowledge scoped per-repository, queryable, and auditable in run_events instead of scattered across static files that go stale.

RCA discipline before proposing a fix

1. memory_recall first — has this repository seen this before?
2. code_search_repository for the actual current field/route/symbol name — never assume a name from a doc or prior session; docs and memory can be stale.
3. Confirm the failure is reproducible (test or explicit repro steps), not a one-off hallucinated symptom.
4. Only then propose a patch, and keep it to the smallest verifiable change.
5. After verification succeeds, memory_remember the root cause + fix — not the whole diff, just what a future session needs to avoid re-diagnosing it.