kit 6.6.2

Added

• kit browser doctor — verification-readiness diagnostics. Reports which browser
  strategy a repo actually gets (Playwright, system Chrome, CDP) or the blocker standing
  in the way, plus kit browser status, kit browser cdp-url, and
  kit browser playwright-env for wiring test runners. --json on each.

• Secret findings you have reviewed can now be accepted by name — .kit-secretsignore.
  Git history is immutable, so a test fixture or doc placeholder committed once is a
  finding forever, and secrets scan sat at warn permanently. A warning with no
  reachable green state is one nobody reads, which is how a real finding gets missed.
  Entries are <commit>:<file>:<detector>, one per line. Two properties keep it safe:
  an entry names a single commit, so it can never wave through a future occurrence of
  the same string, and a verified-live finding is never ignorable no matter what is
  listed. See docs/COMMANDS.md → "Secret findings".

Fixed

• The secrets scan now classifies example credentials instead of asking you to
  re-review them every run. Findings whose own value proves they name nothing real —
  an unreachable host (loopback, a bare compose/k8s service name, .internal,
  example.com) or a placeholder secret (pass, password, changeme, a vendor's
  published doc sample) — are counted in their own bucket rather than as reviewable
  findings. Keyed strictly on the value, never the file path: a real credential pasted
  into a test file is still flagged, and a verified-live finding is never waved through.
• npm test refuses to run when build output holds cloud-sync duplicates. In a
  synced checkout (iCloud Drive), a conflicted file gains a 2 twin; under dist/
  that twin is a stale compiled test running old assertions against new code — it can
  fail a correct change, or pass a broken one. The runner now names the files and stops
  instead of reporting a verdict it cannot trust.
• A skill installer's actual payload is now triaged, not just its wrapper.
  npx skills@latest add <owner>/<repo> triaged only the npm wrapper, passed, then
  cloned an untriaged third-party repo into ~/.agents/skills globally — the same
  fetch-and-execute blind spot create-* already closes, one level deeper. When a
  fetched package is a known repo fetcher, its owner/repo argument is now triaged too.
• The gate no longer triages a registry package npx would never fetch.
  npx tsc --noEmit was 23% of kit guard's would-block observations, gating npm's
  unrelated abandoned tsc package — but npx resolves the plain positional from local
  node_modules/.bin first, so nothing is fetched. A plain npx <name> / bunx <name>
  whose binary exists locally is now dropped; a -p/--package/--from target is
  still always gated, because npx fetches that one unconditionally.
• Concurrent agent startups no longer produce phantom TRIAGE FAILED verdicts.
  The bundled triage skill was refreshed with an in-place truncate-then-write, so
  several agents starting at once could read a half-written triage.py and crash with
  no real verdict behind it. The refresh is now atomic (write to a temp file, rename).
• The memory store opens read-only when its WAL sidecars cannot be created. The
  fallback was keyed on one SQLite error string, so a build that words the identical
  precondition differently ("attempt to write a readonly database") never took it, and
  checkMemoryInjection returned warn instead of pass. It now keys on the actual
  precondition — sidecars missing — not on the wording.
• kit statusline help distinguishes setup gaps from open action items, and states
  that Codex receives the same line via SessionStart context rather than a status bar.

---

Full changelog: https://github.com/sandstream/kit/blob/v6.6.2/CHANGELOG.md

Verify this release:

git tag -v v6.6.2
npm audit signatures