Pick the smallest correct verification gate for a change in this repo, and prove it actually ran before calling it green. Use before running any `npm run verify:*` or `check:*` command, before reporting a gate as passing, when a gate finishes suspiciously fast or clean, and before trusting any agent or bot claim that a fix landed.
86%
Does it follow best practices?
Run evals on this skill
Adds up to 20 points to the overall score
Low
Low-risk findings worth noting
A green exit code is not proof. Contended gates can wait a long time before failing, early stops leave later checks unrun, and a stale worktree makes healthy-looking runs meaningless. This skill exists because those failures have cost real time more than once.
Rule: never report a gate as passing without quoting the decisive line that proves it ran. Pick one smallest sufficient gate first; do not stack broad gates unless each covers a distinct plausible failure path.
Check these before believing any result.
verify:ui under heavy-lock contention waits, then fails — it does not soft-skip green. When
another worktree holds the exclusive lease, acquireHeavyRunLock queues Playwright admission for
up to 15 minutes and throws on timeout; scripts/run-playwright.mjs catches that error and exits
75 (DATABASE_HEAVY_RUN_ADMISSION_BUSY), not 1. Do not treat a long wait or a contention
timeout (75) as a false-green soft-skip or a product test failure. When the gate does run, grep
the output for the N passed line — exit 0 alone is never proof.check:installed-lock-parity fails closed
for exactly this reason — if installed packages do not match package-lock.json, treat any test,
lint, or typecheck result as void until npm ci has run. Its own failure message says as much.verify:full stops at the first failing check. Everything after that point never ran. Do not
describe the change as broadly verified when the gate died at its second check.verify:cheap can still fail CI on formatting. During iteration, format only task-owned files.
Before a push, follow AGENTS.md: from an isolated or otherwise fully owned worktree run
npm run format, review the complete formatter diff, and commit it. Never sweep a shared dirty
checkout that contains another task's work.tail or head masks its exit code. In Bash, capture ${PIPESTATUS[0]},
or check the exit status before piping.Match the gate to what actually changed. Running a broader gate is not more rigorous if it cannot observe the change; running a narrower one is not sloppy if it can. Add a second gate only when it covers a distinct plausible regression and the incremental confidence justifies its cost.
| Change | Gate that can actually fail |
|---|---|
| Markdown / docs only | prettier --check, docs:check-links, docs:check-index |
| Localised source behavior | test:focused -- --files <paths> |
| Ordinary pre-PR check | verify:cheap — check:installed-lock-parity + check:diff-integrity + lint + typecheck + test |
| Cross-module/unknown scope | verify:full — static gates (check:runtime through check:offline-pack terminal static gate) then lint + typecheck + unit test |
| Before PR handoff | verify:pr-local (risk-routed; inspect with --dry-run) |
| UI, styling, routing, a11y | npm run ensure, affected journey, broad UI only when shared |
| Phone chrome | verify:phone-chrome (narrower than verify:ui; run it first) |
| Explicit release confidence | verify:release (includes build + full Playwright + governance/eval; provider approval required) |
lint, typecheck, and test cannot observe a markdown-only change. Say so rather than running
them for appearance.
The repository run coordinator serialises heavy work across worktrees. Leases live at
<os.tmpdir()>/clinical-kb-heavy-locks/<repoId>.lock/leases/; each holds an owner.json with pid,
mode, and command.
npm ci or npm install while any repository test, build, lint, typecheck, or server
command is active — including in another worktree.Never run provider-backed gates without explicit user confirmation: eval:rag, eval:quality,
eval:retrieval:quality, verify:release, check:supabase-project, test:live, and any GitHub,
Supabase, OpenAI, or hosted-CI call. Report the command and ask instead of running it.
Use the canonical API and provider confirmation boundary
for standing authorization, including the distinct Run PR and Clear PRs shortcut scopes.
Preserve each shortcut's duration, actions and exclusions in
the PR workflow; do not ask again for an already
authorized action or extend that authorization to another action.
A bot or agent saying it fixed something is a claim, not a result.
git log, git show); fetching from the remote is a network action subject to the
provider boundary above.git merge-base --is-ancestor returning false does not mean
the work is missing — compare file content on origin/main instead..github/workflows for auto-merge
returns nothing even though GitHub's per-PR auto-merge is in active use here, because that setting
is not a workflow file.State what ran, what passed, and what never ran.
42 passed, docs link check passed: 1287 …, not "gate green".BigSimmo/PsychSift
0d5a97a
If you maintain this skill, you can claim it as your own. Once claimed, you can manage eval scenarios, bundle related skills, attach documentation or rules, and ensure cross-agent compatibility.