Best-practices review fixes: risk-signal dead logic, Standard audit path, closeout extraction, triage-log writer (0.16.0)

[stid]

From a full review of the plugin's prompt surfaces (skills, commands, agents, rules, templates, hooks) against Anthropic's current published guidance on skill authoring, subagents, hooks, and instruction framing.

Fixed — dead logic in the routing rule

The risk-signal definition ("irreversibility OR multi-session cold-pickup") contained a branch that could never fire: Deep-level irreversibility is already a hard trigger, which routes to Deep before the 4+-signals branch is evaluated. The rule now names the one operative risk signal — multi-session cold-pickup — in all three places it is stated (triage SKILL, rules/lcd.md, README), with the why in parentheses.

Changed — closeout contract extracted; Standard audit gets its own path

• The closeout contract (~40 lines: log line, opt-in evaluator, living-spec fold) lived inline in the triage SKILL — paid at triage time for actions that fire at the finish line — and was restated verbatim in /lcd:audit. It now lives once in skills/triage/references/closeout.md (the fanout.md pattern); both call-sites point there. Triage drops 237 → 207 lines, under the ~200-line instruction-file adherence budget.
• /lcd:audit no longer threads Standard-lane items through a Deep-numbered list via exceptions. Standard has its own section, including the previously unspecified failure branch: BLOCKED names each uncovered AC-N (SURFACE) and writes no closeout line. PASS (test-presence) joins the documented audit-result vocabulary (it was emitted but absent from the contract enum).

Added — bin/lcd-triage-log.sh (+ fixture test)

The triage log's read side was always a script; the write side was model-formatted — but /lcd:tidy machine-reads both line shapes. A format is only a contract if a program owns both ends: the script validates every field (lane/hard/risk enums, the audit-result vocabulary), creates the log with its header when missing, and echoes the appended line. Triage and the closeout reference now call it.

Docs drift pass

• Plan template Constitution table gains the commits.md row the command already required.
• Quick lane records decisions as D-NNN blocks (the DECISIONS template's format), not free-form lines.
• STUCK header names all three halt reasons (iteration-cap was missing).
• Hardcoded "~5 min" prompt-cache TTL dropped from rules/lcd.md and docs/why-lcd.md — TTLs vary by plan and load; the cache-miss argument stands without the dated number.
• none/EVAL combination rule stated flatly ("none combines with nothing; EVAL is mixable with any real surface") instead of the self-contradicting "opposite of none but like none".
• doctor/tidy drop the literal argument-hint: "(no args)" placeholder.
• One sample finding line each in the lcd-evaluator / lcd-reviewer digest contracts.
• lcd-recon notes its Context7 allowlist assumes the server is registered as context7 (other registrations put the tools outside the allowlist; the agent falls through to web search).

Deliberately left alone

• MultiEdit stays in the hooks matcher — back-compat with older Claude Code versions; it simply never matches on current ones.
• user-invocable frontmatter untouched — works today; default-value semantics not worth the risk in this pass.

Verification

• tests/run.sh: 11/11 (includes the new test-triage-log.sh).
• claude plugin validate .: passed.
• Prompt changes here are behavioural (triage routing wording, audit restructure) — an eval row per evals/run-eval.sh would be the usual evidence; flagging for a local run since eval runs cost API tokens.

🤖 Generated with Claude Code

https://claude.ai/code/session_0123cKLoJu9PxoCQfBCFU1yN