v4.4.0

A check the gate cannot run reports SKIPPED instead of WARN, so a clean change can return PASS, and solumbe calibrate grades the gate's verdict against what happened next. No command or field was removed. A check's status gains the value SKIPPED and the verdict schemaVersion moves to 3, so a client that reads checks[].status must accept it.

Added

• solumbe calibrate --gate grades the gate's verdict and the convergence band against history.
  • It replays the real local gate and the real convergence score on each commit, checked out into a temporary worktree, and joins the result to the same repaired outcome solumbe calibrate uses for risk flags.
  • It reports the repair rate for PASS, WARN and FAIL, the alarm rate, the share of alarms on changes nothing repaired, a row for each check that warned, and the rate for each convergence band. Rates carry a Wilson 95% interval and are withheld under 30 commits.
  • On this repository: 111 commits, PASS repaired 16.0%, WARN 36.7%, with overlapping intervals, and an alarm on 27.0% of changes.
  • It shows whether a verdict separates repaired changes from the rest, not whether the gate prevents a repair. The optional analyzers are not run against historical trees.
• solumbe calibrate --follow-through reports how often a warning had cleared on the next run.
  • It reads the local usage log, pairs each WARN or FAIL with the next gate run on the same repository inside 60 minutes, and reports per check how often that check then passed. A convergence score below aligned is paired with the next score of the same task.
  • Gate and convergence events in the usage log now record each check's status under its name, whether the gate was local or a pull request, and hashes of the repository root, the changed-file set and the request. No path or request text is logged, and none of it is shared.
• SOLUMBE_CHECK_MODE=trial runs a trial with a control arm on the checks an agent makes before it commits.
  • Off by default. Each change falls in a shown or a withheld arm by a hash of the repository root and its head commit; SOLUMBE_CHECK_SHOWN_SHARE sets the split.
  • In the withheld arm the MCP convergence_score and local review_gate tools compute and log the result and tell the agent only that it was recorded. A blocking FAIL is shown in both arms. The command line, the pre-commit hook, CI and the pull request gate are unaffected.
  • solumbe calibrate --trial grades the arms on the first commit made on each head. scripts/hooks/check-outcomes.mjs grades them on corrections and rework in the same Claude Code session.

Changed

• A check that cannot run in the current mode is SKIPPED, not WARN, and never moves the verdict.
  • Local review_gate reported Review state as a WARN on every run, and Compliance controls as a WARN whenever bouncer runs only in CI. A clean local change could never return PASS.
  • Both now report status: "SKIPPED" with a "Not checked locally" summary. verdict is still PASS, WARN or FAIL; only a check's status can be SKIPPED, so a client that reads checks[].status must accept the new value.
  • A configured bouncer whose binary cannot be found stays a WARN: that install is actionable. Staged mode still reports Review state as PASS.
  • The terminal, markdown, review_verdict, solumbe pass-pr and herdr trust-pane reports show a skipped check on its own line (info, not warning). A clean local run now ends "ready for a PR; skipped checks run there".
  • solumbe eval --gate-effectiveness accepts SKIPPED in expectedChecks.
  • The pull request gate reports it too, for evidence GitHub did not supply: Review decision when GitHub returns none, CODEOWNERS when the file cannot be read or an owner cannot be matched to a reviewer, and Branch protection when the base branch or the protection API is unavailable. A clean pull request ends "ready to merge once review decision is confirmed".
  • The company and high-risk profiles are unchanged: they require those checks to be PASS, so a SKIPPED one still fails them. An unprotected base branch, a missing CODEOWNERS file, missing status checks and a missing approval under solo governance still WARN.
  • The verdict schemaVersion is now 3, for the new check status.

Fixed

• Release discipline fires only when a version field changes.
  • It read the version from the working tree and compared it with the base, so under --head, --staged or a PR whose checkout had moved on, any package.json edit looked like a version bump. 20 of the 22 FAILs in a 495-PR backtest were dependency edits.
  • It now reads the version from the exact subject: the head commit tree, the staged index tree, or the PR head commit. It falls back to the working tree only when that copy is unavailable.
  • A package.json with no version, and a version mismatch or non-SemVer value the change did not introduce, no longer trip it.
  • In a Release Please repository (a release-please-config.json or .release-please-manifest.json at the base), a version change outside a release PR is reported as "Version bumped by hand in a Release Please repository", not as a missing changelog entry. A release PR is recognised by its release-please--* branch or by a change to the manifest.
• Secret safety stops flagging placeholders.
  • .env.production.local.example and every other *.example, *.sample, *.template or *.dist file are templates, not secret files. Only .env.example and its three siblings were exempt before.
  • The heuristic hardcoded credential rule no longer runs on test files (*.spec.*, *.test.*, __tests__/), documentation, fixture directories or templates, where accessToken: 'valid-access-token' is a made-up value.
  • Known secret formats (AWS, Stripe live, GitHub, Slack and the rest) still fail in every file, test files included.