calm-guide

CALM MCP workflow — quick trigger

Read AGENTS.md at the project root in full now. It is the single source of truth for this workflow — this skill exists to make sure you actually reach for it at the right moment, not to hold a second copy of its content that would silently drift out of sync with it over time. If AGENTS.md doesn't exist at the project root, this isn't a CALM-instrumented repo (or you're in the wrong directory) — stop here, this skill doesn't apply.

If you only have room for the short version right now

1. Orient — repo_overview() always first, then hotspots()/ fitness_report() as needed.
2. Locate — locate(query) (search + file_overview + symbol_info in one call) or search(query, kind=...).
3. Inspect — source(symbol) for a symbol-precise read, or understand(symbol) for locate+source+callers together.
4. Trace — callers(symbol), callees(symbol), path(from, to), dependencies(path).
5. Pre-Edit — edit_context(symbol): mandatory before touching any real code symbol, no exceptions. (A Markdown/text heading is the one exception — it never carries a blast radius to check.)
6. Edit — edit_symbol(...)/edit_lines(...): CALM's own write path, hash-verified, risk-gated, reindexes immediately. Fall back to native Edit/Write only for a brand-new file or a path CALM doesn't index (dotdirs, target/, node_modules/, dist/, build/, etc.).
7. Verify — diff_impact(staged=true): mandatory before any commit or push, no exceptions.
8. Recover — session_context() after 10+ calls without progress; remember(topic, content)/recall(topic) for durable cross-session notes.

Steps 5 and 7 are the only hard-enforced ones under Claude Code (a hook denies the native equivalent without them) — the rest is strongly recommended, not blocking. Every signal, edge case, and the preset/toolset reference lives in AGENTS.md — go read it in full if you haven't this session.