OrodruinLabs/nazgul:review

Nazgul Review

Examples

• /nazgul:review — Review the current IN_REVIEW task
• /nazgul:review TASK-003 — Review a specific task by ID
• /nazgul:review --materialize TASK-003 — Repair a task stuck without per-reviewer evidence: re-run the full board, write genuine reviewer files
• /nazgul:review --materialize --all — Repair every task whose evidence is missing or incomplete

Current State

• Config: !cat nazgul/config.json 2>/dev/null || echo "No config"
• Active reviewers: !jq -r '.agents.reviewers // [] | join(", ")' nazgul/config.json 2>/dev/null || echo "none"
• Plan: !head -20 nazgul/plan.md 2>/dev/null || echo "No plan"

Arguments

$ARGUMENTS

Instructions

Format all output per references/ui-brand.md — use stage banners, review verdicts, spawning indicators, and display patterns defined there.

If a task ID is provided in arguments:

1. Read the task manifest at nazgul/tasks/[TASK-ID].md
2. Verify the task is in IMPLEMENTED or IN_REVIEW status (unless --materialize is present — see repair mode below)
3. Delegate to the Review Gate agent for that task

If no task ID provided:

1. Scan nazgul/tasks/ for any task with status IMPLEMENTED or IN_REVIEW
2. If found, delegate to the Review Gate for that task
3. If none found, report that no tasks are ready for review

If --materialize is in arguments (repair mode):

Repair mode exists for tasks deadlocked by missing per-reviewer evidence (e.g., a run that wrote a consolidated summary.md instead of nazgul/reviews/[TASK-ID]/<reviewer>.md files). It re-earns evidence by running genuine reviews — it NEVER converts summary.md claims into reviewer files.

1. Select targets:
   • With a TASK-ID: that task only.
   • With --all: every task in nazgul/tasks/ whose status is DONE, APPROVED, IMPLEMENTED, IN_REVIEW, or BLOCKED with a review-evidence Blocked reason, and where any configured reviewer (jq -r '.agents.reviewers[]' nazgul/config.json) lacks an APPROVED nazgul/reviews/[TASK-ID]/<reviewer>.md. Including IMPLEMENTED is by design: the deadlock parks tasks there; a task awaiting first review simply gets its first review. Including evidence-BLOCKED tasks matters most: those are exactly what the stop hook's escalation creates. Display the matched list before dispatch.
2. Accept any post-implementation status — IMPLEMENTED, IN_REVIEW, DONE, APPROVED, or BLOCKED (when the Blocked reason mentions review evidence). If BLOCKED for another reason, skip and report.
3. Reconstruct the diff if missing: if nazgul/reviews/[TASK-ID]/diff.patch does not exist, rebuild it from the commit SHA(s) in the task manifest's ## Commits section: git show <sha> > nazgul/reviews/[TASK-ID]/diff.patch (append multiple SHAs with >>). If the manifest has no commits, set the task BLOCKED with reason cannot materialize reviews — no commit SHA in manifest and report.
4. Dispatch the FULL configured reviewer board over the diff via the Review Gate agent — each reviewer writes its own nazgul/reviews/[TASK-ID]/<reviewer>.md.
5. On all APPROVED: walk the legal state path for the current mode: first reach IN_REVIEW (IMPLEMENTED → IN_REVIEW, or BLOCKED → IN_REVIEW — legal because the reviewer files now exist), then if afk.yolo is true in nazgul/config.json: IN_REVIEW → APPROVED; otherwise IN_REVIEW → DONE. (Tasks already DONE that pass validation need no walk.) Then clear the escalation bookkeeping: remove the task's key from .safety._review_reset_counts in nazgul/config.json (jq + temp-file write), and remove any review evidence missing Blocked reason line from the manifest.
6. On rejection: normal CHANGES_REQUESTED flow — repair runs get no lower bar. A genuine rejection is signal, not friction.

Review Process

1. The Review Gate runs pre-checks (tests, lint)
2. Each reviewer evaluates the changed files
3. Reviews are written to nazgul/reviews/[TASK-ID]/
4. Consolidated feedback (if rejection) written to nazgul/reviews/[TASK-ID]/consolidated-feedback.md
5. Display the verdict and any blocking issues