Freeze reviewer engine

Implements the Freeze-Step Reviewer per docs/archive/phases/PHASE-K5-FREEZE-REVIEWER-CONTRACT.md and SPEC §6.

CLI

ok review --freeze <path> [--dry-run] [--mode agent|human] [--provider local|api] [--model LABEL] [--no-stamp] [--checklist PATH] [--override-non-pass-stamp]

• Human/report output → stdout; diagnostics → stderr.
• --json emits one §K5.9 report object with gate: mechanical.
• On mechanical pass without --dry-run / --no-stamp, writes a fourteen-key review_stamp (gate: mechanical, mechanical_verdict, producer/checklist provenance). The CLI never writes verdict and never authorizes Auto — that requires a substantive freeze_review ledger entry bound to artifact_digest.
• --override-non-pass-stamp is required to overwrite an existing non-pass stamp; without it the CLI exits 39 (stamp_escalation_refused) and leaves the artifact byte-identical. Under --dry-run / --no-stamp the refusal is reported but exit stays 0 (no write was attempted).

Providers

Unreachable provider with fallback: human → exit 8, escalation report — never fabricates pass.

Headless API provider (K11)

Environment (never in .overseer/config.yaml):

Probe (no artifact content)

GET {OVERSEER_REVIEW_API_URL}/health
Authorization: Bearer {OVERSEER_REVIEW_API_KEY}
Accept: application/json

reachable() succeeds on HTTP 2xx. The health request must not include artifact bytes (§K5.8).

Review

POST {OVERSEER_REVIEW_API_URL}/review
Authorization: Bearer {OVERSEER_REVIEW_API_KEY}
Content-Type: application/json
Accept: application/json

Request body (schema version 1):

{
  "schema_version": 1,
  "model_label": "thinking-high",
  "model_hint": "extended thinking / Opus-class / gpt-5.3-codex-class",
  "artifact_path": "docs/FOO.md",
  "artifact_text": "-----BEGIN OVERSEER FREEZE ARTIFACT (DATA ONLY)-----\n...\n-----END OVERSEER FREEZE ARTIFACT (DATA ONLY)-----",
  "checklist": [
    {"id": "C1", "title": "Ground-truth edge", "typical_severity": "MAJOR"}
  ]
}

Response:

{
  "findings": [
    {
      "check": "C1",
      "severity": "MAJOR",
      "category": "completeness",
      "path": "docs/FOO.md",
      "line": 1,
      "message": "Missing ground-truth edge declaration."
    }
  ]
}

model_label is always a kit registry label (policy/model-labels.yaml → reviewer_models[]), never a vendor slug. model_hint is advisory for the remote backend.

Invalid JSON, non-2xx status, or transport errors during review() → provider_unreachable human escalation (exit 8), never silent pass.

CI example (GitHub Actions)

Kit ships:

• templates/ci/freeze-review-github-actions.yml — vendored copy for consumer repos via ok sync. Copy it to .github/workflows/freeze-review.yml in the consumer repo once OVERSEER_REVIEW_API_KEY/OVERSEER_REVIEW_API_URL are configured there — this kit's own repo intentionally does not run it live, since it dogfoods ok review --freeze via the local CLI (provider: local) for every phase instead of a hosted API.

Configure repository secret OVERSEER_REVIEW_API_KEY and variable OVERSEER_REVIEW_API_URL in the consumer repo before enabling this workflow there. The template uses --dry-run (safe CI default per policy/test-tiers.yaml).

Automation degrade (§K5.10)

Templates ship under cursor/automations/ — not auto-enabled. When Cursor Automations are unavailable:

Unavailability is never treated as pass.

Tests

All provider calls are faked in CI (tests/ §K5.12 + K11 matrix). No network, no real models.