Repository navigation
Releases
v1.9.0
1.9.0 - Experimental GitHub Copilot agentStop hook (opt-in, contract-verified only)
- New
stop-hook --format copilotand installer (proctor setup --agents github-copilotorproctor install-copilot-hook) writing.github/hooks/proctor.json. Blocking findings emit the documented{"decision":"block","reason":...}contract with exit 0; clean checks,stop_hook_active: trueand malformed input are silent, so proctor blocks at most once per turn and never approaches Copilot's documented 8-block runaway guard. Both documented input dialects accepted;bash/powershell/commandfields so the Copilot CLI and cloud agent run the same handler. - Installer merges into an existing
proctor.jsonpreserving foreign events/entries/fields, never touches sibling files in.github/hooks/, andproctor uninstallremoves only proctor's entry. Installs only on explicit opt-in, never on detection alone. - Verification status: contract-verified against GitHub's Copilot hooks reference only, never live-verified against a real Copilot CLI or cloud agent session. Synthetic tests cover the documented payload shapes, block/silence contract, schema, merge/uninstall and setup gating - they are not evidence of real enforcement or remediation. Treat the hook as unproven until run live.
proctor doctorreports the Copilot hook as experimental/opt-in.- Fixes a README error that credited Cursor with a Stop hook integration it does not have.
No change to the live-verified tier (Claude Code, Codex CLI end-to-end; Antigravity catch+feedback). No market or coverage claims.
1.8.0: Go support in opt-in dynamic checks
Go modules with *_test.go files now run in the opt-in dynamic checks through go test -count=1 -json ./... under the existing bubblewrap write/network isolation, with GOPROXY=off and GOTOOLCHAIN=local so runs never fetch modules or toolchains.
Which checks execute for Go: DY001 (baseline test regression) and DY005 (inconsistent repeated test, now recognizing an added time.Sleep). Which do not: DY002 probes and DY004 mutation-lite remain Node/Python-only; DY003's expectation-literal recognition is JS/Python-shaped and does not fire for Go. A changed go.mod/go.sum stops a dynamic run as inconclusive, matching the existing dependency-manifest pinning.
Verification, with exact execution scope: full vitest suite green locally (batched runs; lint clean; verify:pack passes from a fresh tarball install) and PR/main CI green on macOS/Ubuntu/Windows plus the proctor self-check and tuned-dev gates. On CI runners (no bubblewrap), the 4 unit-level Go tests (JSON event parsing, adapter selection, test-location) pass on every platform/Node leg while the 5 sandbox integration cases skip — end-to-end Go execution was verified only on Linux with Go 1.26.6 and bubblewrap 0.6.1, not cross-platform. Those authored integration cases run through the real sandbox: a removed guard pinned by a baseline test flags DY001 naming example.com/age::TestRejectsOutOfRange; an honest red-to-green fix stays silent; a build failure reports skipped with a reason and no finding; a deterministic authored regression fixture (a time.Sleep-added test with visit-counter outcomes, not a real scheduling flake) flags DY005; the working tree is left untouched.
Evidence integrity: named results come only from go test -json pass/fail events carrying a Test field; package-level pass/fail lines, skips, cached results (-count=1 disables caching), empty packages, and build/package-level failures produce no named evidence — build failures and exit-without-named-failure states report skipped/inconclusive, never a clean pass or a synthetic finding. Duplicate or conflicting named outcomes make the run inconclusive.
Contract-verified through the sandboxed worker only: no live-harness, catch-rate, or all-language claims. Modules whose dependencies are not already in the local module cache fail honestly rather than fetch, and cold Go build caches can approach the default 60s dynamic budget on large modules (reports skipped, never a finding). The historical Windows 15/16 count-loss incident remains an unresolved caveat; see docs/UNRESOLVED-INTEGRITY.md.
1.7.0
The explicit result contract: check --result-json versioned envelope and --require-dynamic (exit 3, never a cheating finding, for incomplete or review-required requested dynamic evidence); verification-aware receipts, badges, Markdown and SARIF; legacy --json unchanged.
The Stop hook and unresolved --verify now pass --require-requested-dynamic: requested dynamic evidence that is incomplete or review-required is unverified at those state-clearing consumers instead of a false clean, so the block budget and unresolved record are retained until a complete check with an observed successful state update. Static-only projects and empty diffs are unaffected; the public optional-dynamic exit contract is unchanged.
The original Windows 15/16 count-loss incident remains an unresolved historical event with unknown cause; this release does not claim to explain it.
Registry verification: npm latest is 1.7.0 and the downloaded registry tarball matches SHA-1 d22de441092fb6ac9c34f2704b3e25a9a3f73f58 (same bytes as the local pack). A fresh npm install @kavishdua/proctor@1.7.0 passed version, clean check, setup, and the unverified-path smoke scenarios: committed dynamic: true config with skipped dynamic evidence leaves the Stop hook fail-open with bounded UNVERIFIED stderr, retains the block budget and writes no record; unresolved --verify exits 3 and keeps the record; a static-only project clears normally with no UNVERIFIED.
v1.6.4 - Visible unverified Stop state
Visible unverified Stop state
Unknown Stop checks and budget/record persistence now produce bounded UNVERIFIED diagnostics without turning infrastructure errors into cheating findings or silently clearing prior blocked continuity. Atomic updates preserve prior bytes and original history/provenance on replacement failure, and clean only temporary artifacts they successfully created.
A missing optional execution sequence uses known persisted counts, not an assumed fresh turn. Unknown budgets fail open. Enforce mode attempts to record only the established finding, explicitly not verified exhaustion/count; advisory mode still records nothing. Failed persistence stays UNVERIFIED, not a durable gate. Intentional skips remain silent and preserve continuity; clean count clearing requires successful atomic persistence.
Uncertainty has empty stdout and bounded stderr: Claude/Antigravity preserve exit0; Codex/Gemini use documented nonblocking exit1. Exit status alone does not establish clean checking or recording.
Validation
- Twenty deterministic fault/ownership regressions; local97-file suite2269 passed,1 skipped.
- Build/typecheck/lint/self-check and fresh tarball install passed.
- FixPR98, releasePR99 and postmerge1effdfc all passed five supported CI legs without reruns. Postmerge CI: https://github.com/catfish-1234/proctor/actions/runs/42735724131
- Security: https://github.com/catfish-1234/proctor/actions/runs/42735724435 ; dogfood: https://github.com/catfish-1234/proctor/actions/runs/42735724685
- Published and verified as @kavishdua/proctor@1.6.4 (latest). Downloaded registry tarball SHA-1: 9ab4328d87a3a3aa45ee62586222e525f1836b29. Fresh registry install verified version, clean check, setup and deterministic smoke checks: budget path is a directory, persistent record rename failure/owned-temp cleanup, check infrastructure failure retaining a known count across all four protocol formats, and missing optional attempt sequence retaining the known count. Expected unresolved-gate statuses also passed. Offline doctor ran and honestly reported missing setup/CI in an empty test directory and unverified remote protection, not an independent-enforcement claim.
Limits
This closes the deterministic paths described in docs/UNRESOLVED-INTEGRITY.md, not every possible state-loss path. The original Windows15/16cause remains unresolved; result-contract PR94 remains held and unshipped. No new live model/judge or semantic-coverage claims. Local hooks/records remain agent-writable feedback, not independent enforcement. No catch-rate or prevention claim.
v1.6.3
Preserve unresolved finding history across three demonstrated reset modes: a failed redundant Git-directory lookup, an EBUSY record-read failure, and malformed existing JSON. Read the already resolved path under the writer lock and refuse an unverified overwrite rather than resetting the count, original finding and baseline while claiming success. The existing caller reports UNVERIFIED when it cannot persist safely.
Three deterministic regressions fail against the prior source and pass with this fix. Per-process concurrent Stop diagnostics and the original count assertion are retained. This does not establish the cause of the observed Windows15-of-16occurrence count, fix every record-loss path, or change best-effort budget state and documented Stop fail-open behavior. Diagnosis is retained in docs/UNRESOLVED-INTEGRITY.md. Result-contract PR94 remains held; diagnostic PR95 is unmerged. No semantic detection, live-harness coverage or prevention claim.
All PR and postmerge checks passed, including the five-leg Windows-inclusive CI matrix. Fresh isolated merged-main tarball installation passed. Published and verified as @kavishdua/proctor@1.6.3 (latest). Registry tarball SHA-1: 3acd3702e773a7496dd1d3a3e8f8654c792fd570. A fresh registry install verified version, clean check, setup, unresolved gate and offline doctor. The published CLI refused to replace malformed existing history, reported UNVERIFIED with exit1 at exhausted Stop, and the unresolved gate returned3 while preserving the original bytes. Manual publication has no GitHub Actions provenance attestation.
v1.6.2
Baseline-controlled dynamic cross-runs restore a bounded set of Jest/Vitest and Python test controls, including deleted or newly added controls. Config-only changes trigger a separate run. Config-change failures remain diagnostic evidence, not cheating findings; probes/mutations do not attribute findings across changed controls.
Resolved Git hashes identify local inputs, not independently trusted revisions or runner/dependency identity. Custom setup imports and other runner controls remain outside this bounded coverage. Dynamic remains opt-in/advisory. No new live-verified harness coverage or authenticated execution claim.
All PR and postmerge CI passed, including the full Windows-inclusive matrix. Fresh isolated tarball installation passed. Published and verified as @kavishdua/proctor@1.6.2 (latest). Registry tarball SHA-1: 748d0e65399f71a323f263e1c87a172997490114. A fresh registry install verified version, clean check, setup, unresolved gate and offline doctor. An end-to-end config-only differential ran three suites, observed failure under restored baseline controls and returned diagnostic evidence with zero cheating findings. Offline doctor correctly reports missing CI/CODEOWNERS in the scratch repository and unverified branch protection. Manual publication has no GitHub Actions provenance attestation.
v1.6.1
Conservative dynamic report consistency checks. Duplicate/conflicting test names, contradictory JSON counts/streams, TAP bailouts or incomplete plans, and incomplete/count-inconsistent JUnit now make runtime evidence inconclusive. Unsafe/missing reporter artifacts are rejected. Incomplete runs produce no cheating finding.
Dynamic stays opt-in and advisory. Project-controlled coherent fabricated reports can still pass; report consistency is not proof tests really executed. No new live-verified agent/harness coverage.
All PR and postmerge CI passed, including the Windows-inclusive matrix. Isolated tarball installation passed. Published and verified as @kavishdua/proctor@1.6.1 (latest). Registry tarball SHA-1: 727a705ddaf00d8cc1f400f190f0ec0be124ea15. A fresh registry install verified version, clean check, setup, unresolved, offline doctor, and an end-to-end duplicate-report case returning inconclusive with zero findings. Manual publication has no GitHub Actions provenance attestation.
v1.6.0
Adds proctor doctor for local enforcement-chain audits and explicit doctor --github opt-in for read-only server configuration checks. Default doctor stays offline, and --offline overrides opt-in. Incomplete protection coverage is unverified rather than absent. Local files and job dependencies are advisory evidence, not proof that an agent cannot change or bypass enforcement.
Verified: all PR and postmerge CI, including Windows; consistent manifests; isolated tarball installation; doctor default-offline behavior. Gemini CLI remains experimental/opt-in. Antigravity plugin remains repository-distributed. No all-harness or independent-prevention claim.
Published and verified on npm as @kavishdua/proctor@1.6.0 (latest). Registry tarball SHA-1: 41a0642424ed67e0f10714d72259193082b13b68. Fresh registry install verified doctor, setup, unresolved and a clean check. Publication was manual using the existing token; this release has no GitHub Actions provenance attestation.
v1.5.0
- Add Antigravity CLI Stop-hook enforcement via
proctor stop-hook --format antigravityand an
installableantigravity-plugin/. The Stop payload has nostop_hook_active, so every Stop
re-checks the diff and re-blocks, up to a per-conversation budget (default 5, set with
PROCTOR_STOP_HOOK_BLOCK_BUDGET). When the budget runs out the finding is recorded as
unresolved in the worktree's git directory;proctor unresolvedshows it and
proctor unresolved --verifyclears it only after a fresh complete check (working tree plus
commits since the finding). The pre-commit hook gates on it.PROCTOR_STOP_HOOK_MODE=advisory
is an explicit opt-in that skips the record. This is bounded local feedback, not tamper-proof:
keepproctor check --basein CI, run outside the agent's control. Verified live on agy 1.5.3. - Add Gemini CLI AfterAgent enforcement via
proctor stop-hook --format gemini, reusing the
Codex check pipeline. Blocking findings return exit-0 deny JSON with correction feedback;
clean checks, loop-guard skips, malformed input, and internal failures silently allow. - Gemini CLI support is experimental and opt-in:
proctor setupmerges a 60-second AfterAgent
hook into.gemini/settings.jsononly when Gemini CLI is explicitly selected
(--agents gemini-clior--all), never on detection alone, and says so;proctor agents
marks it experimental. Consumer Gemini CLI auth is retired by Google and the hook is
contract-verified only, never live-verified; Antigravity CLI is the supported successor. The
ruleset still deploys to GEMINI.md on detection. Setup explains hook fingerprints, one-time
trust, and/hooks;uninstallremoves only proctor's command. The Gemini extension bundles
the hook in the documentedhooks/hooks.jsonlayout and runs the shipped CLI directly. - Bump vitest to 4.1.12 (dev-only), clearing two moderate
@vitest/mockeradvisories.
v1.4.0
Codex CLI enforcement: proctor now runs as a Codex Stop hook, so a cheating turn is blocked with feedback instead of only reported after the fact. Claude Code and Cursor behavior is unchanged.
- Add Codex CLI Stop hook enforcement via
proctor stop-hook --format codex: blocking findings return an exit-0 JSON decision with feedback; clean checks, loop-guard skips and internal failures stay silent and allow the turn. proctor setupmerges a 60-second Codex Stop hook into.codex/hooks.jsonwhen Codex is a resolved target, preserving existing hooks and settings. Setup explains both project trust and hook review via/hooks;uninstallremoves only proctor's Codex command.