CLAUDE.md · git:20260406.66d82b5 · 2026-04-06 · sha256 fee4140f69f5fcb0
CLAUDE.md git:20260406.66d82b5A
Immutable. This exact content never changes and is served at /api/v1/blob/fee4140f69f5fcb0.
<!-- code-review-graph MCP tools --> ## MCP Tools: code-review-graph **IMPORTANT: This project has a knowledge graph. ALWAYS use the code-review-graph MCP tools BEFORE using Grep/Glob/Read to explore the codebase.** The graph is faster, cheaper (fewer tokens), and gives you structural context (callers, dependents, test coverage) that file scanning cannot. ### When to use graph tools FIRST - **Exploring code**: `semantic_search_nodes` or `query_graph` instead of Grep - **Understanding impact**: `get_impact_radius` instead of manually tracing imports - **Code review**: `detect_changes` + `get_review_context` instead of reading entire files - **Finding relationships**: `query_graph` with callers_of/callees_of/imports_of/tests_for - **Architecture questions**: `get_architecture_overview` + `list_communities` Fall back to Grep/Glob/Read **only** when the graph doesn't cover what you need. ### Key Tools | Tool | Use when | |------|----------| | `detect_changes` | Reviewing code changes — gives risk-scored analysis | | `get_review_context` | Need source snippets for review — token-efficient | | `get_impact_radius` | Understanding blast radius of a change | | `get_affected_flows` | Finding which execution paths are impacted | | `query_graph` | Tracing callers, callees, imports, tests, dependencies | | `semantic_search_nodes` | Finding functions/classes by name or keyword | | `get_architecture_overview` | Understanding high-level codebase structure | | `refactor_tool` | Planning renames, finding dead code | ### Workflow 1. The graph auto-updates on file changes (via hooks). 2. Use `detect_changes` for code review. 3. Use `get_affected_flows` to understand impact. 4. Use `query_graph` pattern="tests_for" to check coverage. <!-- SHREY-JUNIOR:START claude --> # Claude Repo Routing Use `.agent/project.yaml` and `.agent/context/*` as the repo-local source of truth generated by Kiwi Control. Profile: `product-build` Execution mode: `assisted` Prefer this lightweight guidance file over invasive repo-local runtime config in v2. Authority order: 1. existing trusted repo authority files already in the repo 2. promoted canonical docs explicitly referenced by those files 3. generated Kiwi Control overlays 4. `.agent/context/*` 5. fallback templates Before medium/high complexity work or resuming a prior phase: 1. read the existing repo authority files and any promoted canonical docs they reference 2. read `.agent/state/active-role-hints.json` 3. read `.agent/state/current-phase.json` 4. read `.agent/memory/current-focus.json` and `.agent/state/checkpoints/latest.json` 5. read the latest task packet, latest handoff, latest reconcile, and latest dispatch manifest when present 6. read `.agent/context/commands.md`, `.agent/context/specialists.md`, `.agent/context/tool-capabilities.md`, and `.agent/context/mcp-capabilities.md` 7. read the relevant `.github/instructions/*.instructions.md` 8. read relevant `.github/agents/*.md`, `.agent/roles/*.md`, and `.agent/templates/role-result.md` 9. read `.agent/checks.yaml` and `.agent/scripts/verify-contract.sh` 10. read `.agent/project.yaml` 11. read `.agent/context/architecture.md` 12. read `.agent/context/conventions.md` 13. read `.agent/tasks/` packets when present Do not rely on machine-local `.claude/*` files as portable authority unless the repo explicitly routes to them. Cloud-hosted runtimes may not see machine-global files, so treat repo-local artifacts as the portable contract. Workflow rules: - trivial work may stay direct when it is local, low-risk, and does not affect contracts, auth, security, release behavior, or multiple files - use `run` for non-trivial single-owner work that needs repo-aware framing - use `fanout` for guarded or complex work that benefits from planner / implementer / reviewer / tester separation - use `dispatch`, `collect`, and `reconcile` when simultaneous small-role work is useful and outputs must be compared explicitly - require `checkpoint` and `handoff` when Claude is taking over after another tool or handing work back - consult `push-check` before recommending push on non-trivial or guarded work - consult `release-check` or `phase-close` when closing a phase or approaching release readiness - treat `.agent/state/active-role-hints.json` as the quickest repo-local pointer for current role focus, next file to open, next command to consider, latest continuity artifacts, next specialist, and suggested MCP pack - treat `.agent/memory/current-focus.json` as the smallest portable continuity summary when resuming Gate behavior: - if policy or reconcile state is blocked, treat it as a mandatory gate rather than advisory noise - if authority files disagree, stop and warn instead of proceeding confidently - if promoted canonical docs exist, surface them early instead of restarting context from scratch Specialists and MCP: - prefer the curated specialist registry in `.agent/context/specialists.md` when the task clearly maps to implementation, QA, security, refactor, docs, planning, release, or MCP-policy work - emit canonical specialist IDs in any repo-local artifact you write; treat older aliases as compatibility input only - MCP usage follows policy, capability, trust, profile, specialist, and approval rules; do not call MCP tools opportunistically <!-- SHREY-JUNIOR:END claude -->