Freeze-review loop skill

Use when a frozen spec (frozen: true) must reach pass before an Auto build phase may start. Invoke explicitly: /freeze-review-loop or paste the prompt below. Not automatic by default — the operator or handover must request it.

Complements cursor/skills/freeze-review/SKILL.md (single-pass review). This skill runs bounded rounds until pass, escalation, or max rounds — without re-pasting the prompt each round.

Purpose

Prevent dishonest or shallow freeze approvals. A thinking-model reviewer must cite file+line findings; a fixer applies only those citations; the reviewer re-checks until clean or stopped.

When to use

• Roadmap step {step}a (Thinking) just committed a freeze artifact
• ok review --freeze <path> returned findings or you need deep review beyond the checklist
• Before clearing {step}b (Auto) in the handover

Model

• Reviewer rounds: thinking-high label — extended thinking / Opus-class in Cursor
• Fix rounds: may use Auto for mechanical citation fixes only; reviewer always re-runs thinking

Read policy/model-labels.yaml and .overseer/config.yaml → freeze_contract.reviewer.

Hard stops (stop loop immediately)

Escalate to human — do not auto-fix — when any finding is:

• security — auth, scope, secrets, injection
• irreversible — data deletion, non-revertible migrations
• real_money — billing, live spend
• gates_tier3 — artifact gates merge, staging push, live flip

Also stop when:

• Verdict is blocked
• Round count reaches max rounds (default 7)
• Fixer would need to change scope beyond cited lines (Tier 2 — ask once)

Loop (execute in one agent session — do not stop after one round)

ROUND = 1
MAX_ROUNDS = 7   # operator may lower to 3 for small specs

while ROUND <= MAX_ROUNDS:

  1. REVIEW (thinking model, independent posture)
     - Read frozen artifact + frozen_inputs it depends on
     - Read docs/OVERSEER-KIT-SPEC.md §6
     - Emit verdict: pass | findings | blocked
     - Every finding: path:line, severity, category, concrete message
     - Run checklist mentally: ground-truth edge, completeness, consistency,
       security, irreversibility, real money, Tier-3 linkage, citation discipline
     - Do NOT trust prior round summaries — re-read files

  2. If pass:
     - Run: ok review --freeze <artifact-path>
     - If CLI pass: stamp ok; update artifact Review-record table; EXIT success

  3. If blocked OR escalating finding:
     - Record in Review-record table; STOP for human

  4. If findings (non-escalating):
     - FIXER: apply minimal edits addressing ONLY cited path:line items
     - No scope expansion; no deleting requirements to greenwash
     - Commit on feature branch (Tier 1)
     - ROUND += 1; continue loop

If ROUND > MAX_ROUNDS:
  - Verdict: blocked (max rounds)
  - STOP for human

CLI between rounds (recommended)

After fixer commits, before next review round:

./cli/ok review --freeze <artifact-path> --dry-run

Checklist gate catches structural gaps; thinking review catches semantic dishonesty.

Review-record table (required)

Append each round to the artifact's freeze Review record section:

Must cite

Uncited findings are invalid — discard and re-review. Operator verifies citations, not trust.

Tier authority

• Feature-branch fixes during loop: Tier 1
• Merge to {{vcs.git.main_branch}}: Tier 3 — never part of this loop
• Enabling Cursor Automation for this: Tier 2 confirm once

Degrade

If the skill cannot run (no thinking model, context limit): run single-pass freeze-review skill + manual rounds in fresh chats.