CLAUDE.md · git:20260920.5a89c9a · 2026-09-20 · sha256 4be86999e44aa302
CLAUDE.md git:20260920.5a89c9aA
Immutable. This exact content never changes and is served at /api/v1/blob/4be86999e44aa302.
@AGENTS.md # Claude Integration Entry `AGENTS.md` is auto-injected above via `@import` — governance rules, gates, and state model are already in context. This file only adds Claude-specific dispatch. **Do NOT duplicate rules from AGENTS.md here.** ## Startup (every conversation) 1. **At the start of a task, enter the governed flow in `AGENTS.md` before any file edit or completion claim — decide up front, not at the edit moment. No silent direct edits.** 2. Classify task scope from the user's message: - `tiny-fix` (< 3 files, no semantic change) → execute directly (diff + 1-line verification). - `quick-win` (1–2 modules, clear scope) → read SSoT (Step 3), skip Step 4. - `feature` / `architecture-change` / `hotfix` / uncertain → continue. 3. Read `.agentcortex/context/current_state.md` (SSoT). *(Skip for tiny-fix.)* 4. Read `.agent/rules/engineering_guardrails.md`. *(Skip for tiny-fix and quick-win.)* 5. If `.agentcortex/context/work/<worklog-key>.md` exists, read to resume. *(Skip for tiny-fix.)* Token budget rationale: conditional loading saves ~5,000 tokens for tiny-fix, ~3,500 tokens for quick-win. Full mechanics: `AGENTS.md §vNext State Model`. ## Slash Commands `/command` dispatches via `.claude/commands/<command>.md` → canonical workflow at `.agent/workflows/<command>.md`. ## Skills Skill metadata: `.agent/skills/*`. Full instructions: `.agents/skills/*/SKILL.md`. Activation and cache policy: `AGENTS.md §Skill Safety & Precedence` and `§Shared Phase Contracts`. ## Validate Run `.agentcortex/bin/validate.sh` (bash) or `.agentcortex/bin/validate.ps1` (PowerShell).