Skip to content

Agentic Harness H9 Hardening Audit

Status

Audit updated on 2026-08-06 against main at ab7bbe1, including PR #233 (04b93e2), PR #234 (bb67e1d), and merged PR #235. This document records existing evidence and the smallest remaining hardening gap; it does not authorize provider execution, network access, or mutation.

Existing evidence

H9 concernRepository evidenceResult
Cross-platform compatibility.github/workflows/compatibility.yml runs typecheck, lint, formatting, build, and tests on Ubuntu, macOS, and Windows with Node 22 and 24.Covered in hosted CI.
Governance and source safety.github/workflows/governance.yml validates commit range, attribution, whitespace, and production-file budgets; CodeQL analyzes JavaScript/TypeScript and Actions.Covered for pull requests and protected main.
Dependency and release readinessDependency Review checks dependency-changing PRs; .github/workflows/release.yml is manual-dispatch only, restricts refs to main/v*, uses npm trusted publishing, tests and builds before publication, and checks the tree remains clean.Covered by existing release controls.
Adoption gatetests/harness-adoption-gate.test.ts covers passing scorecards, missing/failing/stale scorecards, and missing approvals.Fail-closed contract covered.
Rollback and migration recoverytests/transaction-consistency.test.ts, tests/cli-adopt-apply.test.ts, and tests/cli-pack-update.test.ts cover transactional rollback, stale plans, and migration journals.Existing application recovery evidence.
Retention and purgetests/harness-state.test.ts covers checkpoint purge; tests/evidence-cache-revocation-cross-surface.test.ts covers provider-cache purge and revocation across surfaces.Existing bounded purge evidence.
Composed H9 evidencetests/harness-h9-evidence-contract.test.ts uses versioned deterministic fixtures to compose adoption-gate failure, replay, cross-project resume rejection, checkpoint purge, and transactional rollback recovery.Covered without wall-clock thresholds.
Hosted H9 compatibilityPR #233 adds a Windows Node 22-only timeout budget for the packed CLI process test; PR #235 adds the macOS Node 22 evidence-analysis timeout correction and reconciles the Windows packed-adoption hook timeout. Compatibility, Governance, and CodeQL passed.Correctness evidence is green; timing remains observational.

Remaining gap

There is no stable harness performance budget suitable for a hosted matrix, and the deterministic contract is not a certification claim. Adding wall-clock thresholds to the test suite would be flaky across the existing six compatibility jobs. PR #233 fixed a Windows Node 22 test timeout without turning that observed runner duration into a product budget.

Bounded next slice

The merged H9 performance benchmark specification defines a local, provider-neutral measurement contract that consumes the existing H9 fixtures without becoming a release gate until its variance is characterized. Any future benchmark runner requires its own implementation review and must keep fixture version h9-evidence-drill@1 and its expected terminal states (passed gate, failed gate, deterministic replay, purged checkpoint, and completed rollback) stable.

Explicitly out of scope

  • provider or model adapters;
  • remote or hosted state;
  • generic shell, network, or subprocess capability;
  • automatic activation, commits, merges, releases, or publication;
  • certification or universal safety claims.