Review Work

Codex Harness Tool Compatibility

This skill may include examples copied from the OpenCode harness. In Codex, do not call OpenCode-only tools such as call_omo_agent(...), task(...), background_output(...), or team_*(...) literally. Translate those examples to Codex native tools:

Role-specific behavior must be described in a self-contained message. Use fork_context: false to start the child with only the initial prompt (no parent history); use fork_context: true only when full parent history is truly required. Include any required conversation context, files, diffs, constraints, and requested skill names directly in the spawned agent's message. OMO installs these selectable agent roles into ~/.codex/agents/: explorer, librarian, plan, momus, metis, lazycodex-code-reviewer, lazycodex-qa-executor, and lazycodex-gate-reviewer — pass the matching name as agent_type so the child gets that role's model and instructions. If the spawn tool exposes no agent_type parameter, omit it and describe the role inside message. If a code block below conflicts with this section, this section wins.

Codex exposes ONE of two subagent tool surfaces per session; check your own tool list and route accordingly. If multi_agent_v1.* tools exist, use the table above as written. If instead a flat spawn_agent with a required task_name exists (multi_agent_v2), rewrite every multi_agent_v1.* example: multi_agent_v1.spawn_agent({...,"fork_context":false}) becomes spawn_agent({"task_name":"<lowercase_digits_underscores>","message":...,"agent_type":...,"fork_turns":"none"}) ("all" only when full parent history is truly required); send_input becomes send_message; do not call close_agent/resume_agent (finished agents end on their own; followup_task re-tasks one, interrupt_agent stops one); wait_agent takes only timeout_ms and returns on any child mailbox activity. agent_type works the same on both surfaces. If a code block below conflicts with this section, this section wins.

For work likely to exceed one wait cycle, require the child to send WORKING: <task> - <current phase> before long passes and BLOCKED: <reason> only when progress stops. A multi_agent_v1.wait_agent timeout only means no new mailbox update arrived. Treat a running child as alive. Fallback only when the child is completed without the deliverable, ack-only after followup, explicitly BLOCKED:, or no longer running.

Codex Subagent Reliability

Every multi_agent_v1.spawn_agent message must be self-contained. Start with TASK: <imperative assignment>, then name DELIVERABLE, SCOPE, and VERIFY. State that it is an executable assignment, not a context handoff. Role or specialty instructions belong inside message. Use fork_context: false unless full history is truly required; paste only the review context that worker needs.

Review lanes are leaf agents: a lane does its own reading, running, and judging inline and never spawns sub-reviewers of its own. Reviewers are one-shot: a lane ends at its verdict; a re-review after fixes is a fresh spawn scoped to the delta plus current evidence, never a followup_task to a long-lived reviewer carrying stale context.

Plan and reviewer agents may run for a long time; spawn them in the background and keep doing independent root work. Between multi_agent_v1.wait_agent calls, back off — double the timeout up to ~5 minutes — instead of spinning short cycles.

Treat child status as a progress signal, not a timeout counter. For work likely to exceed one wait cycle, require the child to send WORKING: <task> - <current phase> before long reading, testing, or review passes, and BLOCKED: <reason> only when it cannot progress. While any child is active, keep the parent visibly alive with active subagent count, agent names, latest WORKING: phase, and whether the parent is waiting for mailbox updates. Track spawned agent names locally. Use multi_agent_v1.wait_agent for mailbox signals, not proof of completion. A timeout only means no new mailbox update arrived. Treat a running child as alive. Fallback only when the child is completed without the deliverable, ack-only after followup, explicitly BLOCKED:, or no longer running. Then mark that review lane INCONCLUSIVE, do not count it as PASS or approval, close if safe, and respawn a smaller fork_context: false reviewer with the missing deliverable. Preserve completed lane results immediately. If the retry budget is exhausted, keep the lane INCONCLUSIVE and still emit a final aggregate result.

Review Work - Gate Review Orchestrator

Review completed implementation work through exactly two lanes: your own hands-on manual QA on the real surface, and ONE gate reviewer sub-agent that audits the whole change set against the goal, the constraints, and your QA evidence. The review passes only when the QA matrix has no failing row AND the gate reviewer returns APPROVE.

One reviewer, not a panel. A single gate reviewer holding the full context (goal, diff, history, QA evidence) catches what a fan-out of narrow reviewers misses between their seams, and it costs one agent instead of five. Never add review lanes; widen the gate reviewer's checklist instead.

Phase 0: Gather Review Context

Before running anything, collect these inputs. Extract from conversation history first - the user's original request, constraints discussed, and decisions made are usually already in the thread. Only ask if truly missing.

• GOAL: The original objective. What was the user trying to achieve? Pull from the initial request in this conversation.
• CONSTRAINTS: Rules, requirements, or limitations. Tech stack restrictions, performance targets, API contracts, design patterns to follow, backward compatibility needs.
• BACKGROUND: Why this work was needed. Business context, user stories, related systems, prior decisions that informed the approach.
• CHANGED_FILES: Auto-collect via git diff --name-only HEAD~1 or against the appropriate base (branch point, specific commit).
• DIFF: Auto-collect via git diff HEAD~1 or against the appropriate base.
• FILE_CONTENTS: The full content of each changed file plus the neighboring files that show the established patterns. Required verbatim when the reviewer cannot read files (oracle); when your surface's gate reviewer can read files and run commands, pass the paths and the diff instead of pasting everything.
• RUN_COMMAND: How to start/run the application. Check package.json scripts, Makefile, docker-compose.yml, or ask the user.
• CONTEXT_MINING: What the history and the trackers say about this area (collected below).

Review PRs and branches from a dedicated review worktree only: create or attach one with git worktree add <path> <branch> before collecting changed files, diff, file contents, or running checks, then immediately lock it with git worktree lock <path> --reason "review:<pr-or-branch>". The main worktree is read-only context; never checkout, test, or edit the review branch there.

Auto-collection sequence:

# 1. Get changed files
git diff --name-only HEAD~1  # or: git diff --name-only main...HEAD

# 2. Get diff
git diff HEAD~1  # or: git diff main...HEAD

# 3. Detect run command
# Check package.json -> "scripts.dev" or "scripts.start"
# Check Makefile -> default target
# Check docker-compose.yml -> services

# 4. Mine the context the implementation may have missed (keep the output short)
git log --oneline -20 -- <each changed file>            # recent changes and their reasons
git log --all --oneline --grep="<keywords from goal>"    # related commits, reverts
gh issue list --search "<keywords>" --state all           # related issues (when gh is available)
gh pr list --search "<keywords>" --state all              # related PRs and their review comments
rg -n "TODO|FIXME|HACK" <changed files>                   # warnings left by previous authors
# plus: files that import the changed modules, tests touching the same paths,
# docs and config that reference the changed behavior

Record CONTEXT_MINING as a short list: source -> finding -> why it matters for this change. Slack, Notion, and Discord searches belong here too when those tools exist.

For GOAL, CONSTRAINTS, BACKGROUND - review the full conversation history. The user's original message almost always contains the goal. Constraints often emerge during discussion. If anything critical is ambiguous, ask ONE focused question - not a checklist.

Phase 1: Manual QA (you run it)

You are the QA lane. Do not delegate hands-on QA to a sub-agent: the orchestrator owns the real-surface proof, exactly as the ulw-loop final gate records manualQa under the main session.

1. Reuse first. If this session already captured real-surface evidence for the FINAL tree (an ultrawork or ulw-loop evidence directory, a visual-qa verdict on this same build), consume it as QA rows instead of re-running. A fix committed after a capture stales that capture: re-run the rows it covered.
2. Pick the channel that faithfully exercises the surface and capture the artifact:
   • HTTP: curl -i (or an API request context) - status line, headers, body.
   • CLI / TUI: a real pty - drive the command and keep the transcript; for color or layout evidence render through a browser-based terminal, never a tmux capture-pane dump.
   • Web: the real page in a real browser - the harness's in-process surface (a Bun.WebView / playwright-core code cell, or Codex's Browser plugin) or the agent-browser CLI - action log plus screenshot.
   • Desktop / GUI: OS-level automation against the running app - action log plus screenshot.
   • Library / SDK: a script that imports and exercises the public API - transcript.
   • Data-shaped work (migrations, configs, generated files): the resulting artifact itself, diffed or dumped.
3. Cover at least: the happy path the goal names, the riskiest edge (empty, boundary, malformed, or concurrent input), and one regression on adjacent behavior the change could have broken. Add a row for every stated success criterion.
4. Build the QA matrix - one row per scenario:

A row without an artifact path is not PASS. If the application cannot even start, that is an immediate FAIL.

Any FAIL ends the review here: report REVIEW FAILED with the failing rows and skip the gate reviewer - reviewing code that does not work wastes the reviewer. Fix first, then re-enter at Phase 0 with the delta.

Phase 2: Launch the Gate Reviewer (one agent)

Launch exactly one reviewer, in the background, then keep doing independent root work (teardown prep, report scaffolding) while it runs.

oracle cannot read files or run commands: it receives everything inline (DIFF + FILE_CONTENTS + CONTEXT_MINING + the QA matrix). If your surface's gate reviewer has read and shell tools, still paste the diff and the QA matrix, and hand it file paths instead of full contents.

task(
  subagent_type="oracle",
  run_in_background=true,
  load_skills=[],
  description="Gate-review the completed work against goal, constraints, and QA evidence",
  prompt="""
<review_type>GATE REVIEW</review_type>

<original_goal>
{GOAL - paste the user's original request and any clarifications}
</original_goal>

<constraints>
{CONSTRAINTS - every rule, requirement, or limitation discussed}
</constraints>

<background>
{BACKGROUND - why this work was needed, broader context}
</background>

<changed_files>
{CHANGED_FILES - list of modified file paths}
</changed_files>

<file_contents>
{FILE_CONTENTS - full content of every changed file plus neighboring files that show existing patterns; or the paths, when the reviewer can read files}
</file_contents>

<diff>
{DIFF - the actual git diff}
</diff>

<context_mining>
{CONTEXT_MINING - git history, related issues and PRs, docs and config that reference the changed behavior, warnings from previous authors}
</context_mining>

<manual_qa_matrix>
{QA MATRIX - every row with its artifact path}
</manual_qa_matrix>

Role: final gate reviewer. You do not implement fixes. Assume every success claim is unverified until you reproduce it from the artifacts: executors can be wrong, tests can be too narrow, success prose can be misleading.

Review from the user's perspective first: what did they originally want, what result did they expect to receive, and does the shipped change actually satisfy that outcome? Then work the checklist. Be obsessively thorough - the point of this review is to catch what the implementer missed.

REVIEW CHECKLIST:

1. **Goal completeness**: break the goal into every sub-requirement (explicit AND implied). Mark each ACHIEVED / MISSED / PARTIAL with code evidence. An implied requirement a reasonable engineer would have addressed counts.
2. **Constraint compliance**: list every constraint and verify each with specific evidence. A violated constraint is a blocker.
3. **Behavioral correctness**: trace 3+ representative scenarios and 5+ edge cases (empty, boundary, malformed, concurrent, failure paths) through the code. Logic errors, off-by-one, null handling, races, leaks, unhandled rejections.
4. **Code quality**: pattern consistency with the neighboring files, naming, error handling (no swallowed errors), type safety (no `as any` or suppressions), performance on hot paths, abstraction level, tests that are meaningful rather than tautological or implementation-mirroring, API and breaking-change hygiene.
5. **Security**: input validation and injection vectors (SQL, XSS, command, path traversal, SSRF), authentication and authorization, secrets in code or logs, data exposure, new dependencies and lockfile consistency, unsafe file and network handling.
6. **Missed context**: does the change respect the reasons recorded in history, issues, PR reviews, and docs? Does anything that imports or documents the changed behavior need to move with it?
7. **QA evidence audit**: for every matrix row, does the artifact exist and prove what the row claims? Name any success criterion that has no covering row.
8. **Scope**: anything added that was not asked for - unnecessary abstraction, speculative generality, unrequested hardening - is a NOTE unless it breaks a constraint.

APPROVE unless you can cite a specific goal item, constraint, or QA row that the artifact fails, with the evidence that proves it (including an artifact a criterion requires but that is missing). A gap you cannot tie to the stated goal - style preference, alternative design, a scenario the goal never named - is a NOTE, not a blocker. You do not judge approach optimality or hypothetical future requirements.

OUTPUT FORMAT:
<verdict>APPROVE or REJECT</verdict>
<confidence>HIGH / MEDIUM / LOW</confidence>
<summary>1-3 sentence overall assessment from the user's perspective</summary>
<goal_breakdown>
  - [ACHIEVED/MISSED/PARTIAL] Requirement - evidence (file:line or artifact)
</goal_breakdown>
<constraint_compliance>
  - [ACHIEVED/MISSED] Constraint - evidence
</constraint_compliance>
<qa_audit>
  - [VERIFIED/UNPROVEN] Row # - what the artifact shows, or what is missing
</qa_audit>
<blocking_issues>
  - Violated goal item / constraint / QA row -> observation -> file:line or artifact pointer -> exact fix. Empty if APPROVE.
</blocking_issues>
<notes>Non-blocking findings, grouped by theme (quality, security, context). Keep them short.</notes>
""")

Phase 3: Wait & Collect

Wait for the reviewer in bounded cycles while doing independent root work. Do not treat a timeout, an ack-only reply, or an empty result as APPROVE.

Store the lane verdicts independently:

If the reviewer stays silent after the reliability followup, record the lane INCONCLUSIVE and respawn one smaller reviewer scoped to the same inputs. If that retry also ends without a verdict, close the still-running agent if safe, keep the lane INCONCLUSIVE, and emit the final result naming the incomplete lane. Do not spin in repeated wait/followup cycles, and do not use a queued followup as a cancellation.

After the lane reaches a terminal state and before delivering the verdict, tear down the review worktree: run git worktree unlock <path> followed by git worktree remove <path>. The reviewer runs inside that worktree, so removing it earlier destroys its working directory; a crashed review leaves the locked tree as a recoverable marker for manual cleanup.

A re-review after fixes is a fresh Phase 0 -> 3 pass scoped to the delta plus the current evidence: re-run the affected QA rows and spawn a NEW reviewer with the delta diff and the blockers it must re-check. Never send fixes as a followup to the previous reviewer.

Phase 4: Deliver Verdict

QA matrix has no FAIL row AND the reviewer returned APPROVE → REVIEW PASSED Any QA row FAIL OR the reviewer returned REJECT → REVIEW FAILED - criteria not met Reviewer INCONCLUSIVE and nothing failed → REVIEW INCONCLUSIVE - not approved

Compile the final report in this format:

# Review Work - Final Report

## Overall Verdict: PASSED / FAILED / INCONCLUSIVE

| Lane | Verdict | Confidence |
|------|---------|------------|
| Manual QA (N rows, M artifacts) | PASS/FAIL | - |
| Gate review | APPROVE/REJECT/INCONCLUSIVE | HIGH/MED/LOW |

## Blocking Issues
[Failing QA rows first, then the reviewer's blockers - deduplicated, in fix order, each with its pointer]

## Key Findings
[Top findings across QA and review, grouped by theme]

## Recommendations
[If FAILED: exactly what to fix, in priority order]
[If PASSED: non-blocking notes worth considering]

If FAILED - be specific. The user should know exactly what to fix and in what order: the problem, the file or artifact, and the fix. No vague "consider improving X".

If PASSED - keep it short. Highlight the non-blocking notes worth considering, but don't turn a passing review into a lecture.