v0.18.0

[0.18.0] - 2026-05-30

Schema Compatibility

• Added a machine-readable schema policy and checker that require durable
  harness schemas to declare compatibility, migration, deprecation, and
  changelog rules before 1.0.
• Added stable schemas for permission compile output, explain block records,
  runtime parity reports, and state export manifests; runtime parity JSON now
  carries schemaVersion: 1.
• Explain block records now accept mode: "bypass" and an optional
  fingerprint field for bypass-audit diagnostics.

Documentation

• Added generated guides for policy pack authoring, runtime parity scorecards,
  and team/PR adoption.
• Documented the Codex native hook-fire probe and the
  AHK_E2E_CODEX_REQUIRE_HOOKS=1 release switch for making missing lifecycle
  hook artifacts blocking.

Runtime Parity

• Added codex-parity-probes.mjs and npm run check:codex-parity-probes
  as the release wrapper for strict Codex hook and reviewer artifact probes.
• The strict Codex parity wrapper now emits per-probe JSON outcomes so release
  logs identify whether hook-fire or reviewer-artifact parity failed, passed,
  remained planned, or was not reached before the driver failed.
• The kit repo release gate now validates root-local failure records in
  addition to generated-template failure record fixtures, so observed runtime
  gaps cannot sit outside readiness.
• Runtime parity JSON now includes per-runtime status counts plus category,
  promotion criteria, and next-step metadata for partial capability rows.
• The real Codex E2E smoke now probes generated SessionStart and SessionEnd
  lifecycle hook artifacts. The probe is non-blocking by default while Codex
  hook-fire parity remains partial, and becomes a hard failure when
  AHK_E2E_CODEX_REQUIRE_HOOKS=1 is set. The smoke now initializes a git
  workspace and captures Codex feature/JSONL diagnostics when lifecycle
  artifacts are missing.
• The same E2E smoke now probes Codex reviewer decision artifact creation and
  can make it blocking with AHK_E2E_CODEX_REQUIRE_REVIEWER_ARTIFACT=1.
• Added runtime-parity-report.mjs --fail-partial so release lanes can turn
  partial runtime capabilities into blocking failures without changing the
  default warning-only scorecard behavior.

Explain Layer

• Added agent-harness-kit explain and generated .harness/scripts/explain.mjs
  diagnostics for last-block, task, permission, evidence, and readiness modes.
• agent-harness-kit explain --bypass <fingerprint> now explains bypass audit
  rows and approved request coverage from the same central diagnostics surface.
• agent-harness-kit explain --task <id> now points at
  task-evidence-check --task=<id> instead of the Stop-hook-only
  --active-task mode.
• agent-harness-kit explain --permission <tool> --task <id> no longer reports
  .harness/permissions.json as missing when the task contract supplies the
  relevant allow/deny decision.
• agent-harness-kit explain --task <id> now reports linked evidence pass-proof
  gaps such as missing diffSummary or UI artifacts, matching
  explain --evidence diagnostics.
• agent-harness-kit explain now reports repo-escaping evidence paths as unsafe
  path fields instead of mislabeling them as ordinary missing files.
• Task-scoped permission explanations now keep sourceRule on the task contract
  even when .harness/permissions.json exists but was not part of the decision.
• agent-harness-kit explain --last-block now skips remediation telemetry such
  as block_remediated so resolved blocks do not hide the last active block.
• agent-harness-kit explain --last-block now recognizes canonical block
  telemetry events and routes task-evidence or permission-denied blocks to the
  most direct repair command.
• Generated installs now include .harness/docs/explain.md, documenting JSON
  output, repair commands, and override expectations.

Evidence Attestation

• Added check-evidence-attestation.mjs as a strict readiness gate for passing
  evidence bundles. It requires attested pass checks with command metadata,
  stdout/stderr sidecar hashes, and a replay plan produced by
  task-evidence-check --verify-hashes --replay-plan.
• Generated installs now include harness:evidence:attestation,
  evidenceAttestation config, and the evidence-attestation readiness gate.
• verify-ui summaries now include route, assertion, and DOM snapshot hash
  metadata, and passing UI evidence must carry that browser proof metadata.
• Generated installs now include .harness/docs/evidence-attestation.md.

Permission Compiler

• Added task-aware permission compilation, permissions diff, and
  permissions explain decision-chain output for task, skill, and default
  policies.
• Added check-permissions-drift.mjs and wired generated readiness plus npm
  scripts to use the dedicated drift wrapper.
• High-risk task contracts now fail permission compilation when they use
  wildcard or broad Bash permissions.
• Permission compilation now emits runtime hook expectations for Claude and
  Codex, and fails when generated hook matchers drift from the compiled runtime
  contract, including Codex apply_patch mutation coverage.
• Generated installs now write .harness/permissions.compiled.json during
  rendering and merge compiler-derived Claude permission hints into
  .claude/settings.json.
• harness-report now runs the permission compiler, surfaces compiled skill and
  task contract counts, and fails release JSON output on compiler-detected
  permission drift or high-risk task permission errors.

Bypass Governance

• Added the structured bypass request workflow: bypass request,
  bypass audit --strict, and bypass explain now share the same strict audit
  engine used by release readiness.
• Strict bypass audit now accepts only approved, unexpired request scopes, rejects
  scope mismatches, and requires failure-record links for
  converted-to-failure-record acknowledgements.

Harness Noise Reporting

• Added report-harness-noise.mjs and generated npm scripts for ranking noisy
  rules from block telemetry, bypass records, false-positive acknowledgements,
  review latency, and Stop-hook loop-guard activations.
• Statusline last-block alerts now filter telemetry for real block records, so
  idle or permission prompt notifications do not masquerade as blocking gates.
• harness-report now embeds harness-noise status so release dashboards expose
  false-positive and override pressure instead of burying it in logs.

Upgrade

• Upgrade now reconciles executable bits for unchanged managed scripts recorded
  in the install lockfile, while preserving user-modified sidecar targets.

Eval Tasks

• The kit repo check:eval-tasks release gate now uses a deterministic Node
  wrapper instead of shell chaining, with aggregate JSON output covering both
  root-local and generated-template eval task directories.
• Added a package-script regression test that blocks unquoted shell chaining in
  npm scripts, keeping multi-directory gates on deterministic Node wrappers.

Failure Learning

• The kit repo check:failure-records release gate now uses a deterministic
  Node wrapper instead of shell chaining, with aggregate JSON output covering
  both generated-template and root-local failure record directories.
• check-failure-records now reports the records directory it validated in
  both text and JSON output, making multi-directory release gates easier to
  audit.

---

Full history: CHANGELOG.md
Install: npx agent-harness-kit@0.18.0 init