manciple-review

name	manciple-review
description	Review completed Manciple task work by loading the task packet, checking review readiness with manciple_verify --profile review, evaluating acceptance criteria against run log and git diff evidence, and recording a deterministic verdict (approve/request-changes/block). Use when the user asks Codex to review, approve, or give feedback on a completed Manciple task.

Manciple Review

Use this skill when reviewing a completed Manciple task. The reviewer evaluates whether the implementation satisfies the task spec, records evidence, and decides a verdict.

This is a review-only workflow. Do not implement changes, fix bugs, or expand scope. If the review uncovers issues, record them and let the implementor handle them.

Required Tools

Bind the review to the current repository before using tools. Start every review by resolving the repo root with:

git rev-parse --show-toplevel

The expected Manciple state is the .manciple/ directory under that repo root. Pass that root as the repo argument on every Manciple MCP call. Manciple MCP is global; it must return data scoped to the requested repo, not to the server process working directory.

Prefer deterministic MCP tool results over agent judgment. Use repo-local CLI commands only when MCP tools are unavailable.

• manciple_get_task_packet <task-id> — load compact task context
• manciple_verify --profile review — run deterministic review checks
• manciple_verify --profile worker — re-run worker verification if needed
• manciple_run_log — record the review outcome
• manciple_set_status — update task status after the review
• manciple_get_task <task-id> — load full YAML spec when the packet is insufficient

Use these CLI commands for verdict recording when MCP tools are unavailable:

• manciple approve <task-id>
• manciple request-changes <task-id>
• manciple block-review <task-id>

When using CLI fallback, run commands from the resolved repo root and prefer pnpm exec manciple ... so the local Manciple implementation is used.

Review Workflow

Step 1: Load the task

Resolve the current repo root and verify that .manciple/ exists there. Pass that absolute root as repo on every Manciple MCP call.

Identify the task to review. If the user did not name a task, list tasks in needs_review status:

pnpm exec manciple list --status needs_review

When no specific task is named, review every reviewable needs_review task by default. Sort the queue by the timestamp of each task's latest non-review-outcome run log under .manciple/runs/, oldest first, so the task completed least recently is reviewed first. If a needs_review task has no run log, treat it as not review-ready and record/request changes for missing evidence rather than silently skipping it.

Call manciple_get_task_packet with { "repo": "<repo-root>", "task_id": "<task-id>" } to get the compact task context. Review the packet fields: status, type, goal, acceptance criteria, implementation notes, verification commands, outputs required, and notes.

If the packet status is not needs_review, stop and report the mismatch. Do not review a task that has not been submitted for review.

Step 2: Check review readiness

Run the deterministic review verification:

manciple_verify --profile review

or via CLI:

pnpm exec manciple verify --profile review

Check that the task has a run log entry. If one does not exist, check with:

pnpm exec manciple list --status needs_review --domain <domain>

If review readiness checks fail (e.g., missing run log, incomplete verification), record the issues and set status to blocked or request-changes depending on severity.

Step 3: Load review evidence

Load the run log to see what the worker did:

Run logs are Markdown files under .manciple/runs/, usually named with a timestamp prefix such as .manciple/runs/<timestamp>-<task-id>.md. Use the latest non-review-outcome run log for implementation evidence, and use *-review-outcome.md files only as prior review history.

Inspect the git diff to see what changed:

git diff <base-ref>...HEAD -- <allowed_paths>

Check that only allowed_paths were modified and no forbidden_paths were touched. Verify that the changes match the task scope.

Step 4: Evaluate acceptance criteria

Walk through each acceptance criterion from the task spec. For each criterion:

1. State the criterion verbatim.
2. Check the evidence: run log entries, git diff, verification results, file contents.
3. Determine pass/fail/not-applicable.

Use manciple_get_task with the same repo argument to get the full spec if the packet is insufficient. If MCP is unavailable, use pnpm exec manciple task-packet <task-id> or read only the repo-local task YAML needed for review context.

If the implementation notes mention specific constraints, verify those were respected.

Step 5: Decide a verdict

Choose one of:

Verdict	Condition	Next status
Approve	All acceptance criteria pass. Verification passes. No issues found.	complete
Request changes	Some acceptance criteria fail or issues are found. The work is partially correct but needs fixes.	needs_review (after fixes) or stay needs_review
Block	The task cannot proceed: wrong scope, missing dependencies, fundamental approach errors, or review readiness checks failed.	blocked

When in doubt between request-changes and block, prefer request-changes unless the task is fundamentally unsalvageable.

Step 6: Record the review outcome

Record the review by calling manciple_run_log with:

• task_id: the reviewed task
• repo: the resolved repo root
• task_status: the new status after the review
• result: one of complete, partial, blocked, or failed
• acceptance_criteria_evidence: list of evidence lines showing how each criterion passed or failed
• follow_ups: any issues found that need addressing
• risks: residual concerns
• verify_receipt: the receipt from manciple_verify --profile review
• notes: review notes and verdict rationale

If MCP is unavailable, record the verdict via repo-local CLI:

pnpm exec manciple approve <task-id>
# or
pnpm exec manciple request-changes <task-id>
# or
pnpm exec manciple block-review <task-id>

Set the final task status with manciple_set_status, passing the same repo:

• complete when approved
• blocked when blocked
• Leave as needs_review when changes are requested (the implementor will re-submit)

Step 7: Report

In your final response, include:

• task id and title reviewed
• verdict (approve / request-changes / block)
• acceptance criteria results: pass/fail per criterion
• verification receipt from manciple_verify --profile review or pnpm exec manciple verify --profile review
• files changed (from git diff)
• follow-up tasks or issues identified
• risks

Keep the report concise but preserve specific evidence references.

Verdict Decision Trees

Approve

All of:

• Task status is needs_review
• manciple_verify --profile review or pnpm exec manciple verify --profile review passes
• All acceptance criteria are satisfied
• Changes are scoped to allowed_paths
• No forbidden_paths were modified
• Run log exists and is complete
• No residual risks that block shipping

Request Changes

Any of:

• One or more acceptance criteria not met
• Verification fails
• Run log missing key evidence (files_changed, tests_run, verify_receipt)
• Changes touch unexpected areas
• Implementation notes not followed

Block

Any of:

• Task is not in needs_review status
• Missing dependent tasks that must complete first
• Fundamental approach is wrong — the task needs re-specification
• Changes are destructive or outside the allowed scope
• Review readiness checks fail critically

Scope Boundaries

This skill is for reviewing Manciple task work only. Do not:

• Edit task specs directly (use manciple_set_status for status)
• Implement fixes or improvements found during review
• Review tasks not in needs_review status
• Expand the scope of the review beyond the task spec

If you encounter issues outside the task scope, report them as follow-ups.