CLAUDE.md

1@AGENTS.md
2 
3# Claude Integration Entry
4 
5`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.**
6 
7## Startup (every conversation)
8 
91. **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.**
102. Classify task scope from the user's message:
11   - `tiny-fix` (< 3 files, no semantic change) → skip to Step 5.
12   - `quick-win` (1–2 modules, clear scope) → read SSoT (Step 3), skip Step 4.
13   - `feature` / `architecture-change` / `hotfix` / uncertain → continue.
143. Read `.agentcortex/context/current_state.md` (SSoT). *(Skip for tiny-fix.)*
154. Read `.agent/rules/engineering_guardrails.md`. *(Skip for tiny-fix and quick-win.)*
165. If `.agentcortex/context/work/<worklog-key>.md` exists, read to resume. *(Skip for tiny-fix.)*
17 
18Token budget rationale: conditional loading saves ~5,000 tokens for tiny-fix, ~3,500 tokens for quick-win. Full mechanics: `AGENTS.md §vNext State Model`.
19 
20## Slash Commands
21 
22`/command` dispatches via `.claude/commands/<command>.md` → canonical workflow at `.agent/workflows/<command>.md`.
23 
24## Skills
25 
26Skill metadata: `.agent/skills/*`. Full instructions: `.agents/skills/*/SKILL.md`. Activation and cache policy: `AGENTS.md §Skill Safety & Precedence` and `§Shared Phase Contracts`.
27 
28## Validate
29 
30Run `.agentcortex/bin/validate.sh` (bash) or `.agentcortex/bin/validate.ps1` (PowerShell).
31