菜单
Use when you need a code-review verdict on a PR or current branch, without auto-applying any fixes.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Stage B of /devflow-weekly: given the bundled context of every occurrence PR for one recurring pattern, re-derive the root cause, make the intervention edits in the working tree, and return the touched paths + PR title + PR body as JSON. Invoked as a subagent on a branch the orchestrator already created.
Run the weekly devflow self-improvement loop locally: scan freshly-merged watched-author PRs, write per-PR retrospective entries (LLM only for PRs that fail the mechanical clean-gate), derive recurring patterns, and open one human-reviewed intervention PR per actionable pattern. Use when running the weekly devflow retrospective + audit.
Use when setting up external documentation for the first time, performing a comprehensive documentation refresh, or when large portions of internal docs need corresponding external docs created.
Use when setting up internal documentation for the first time, when the docs directory is empty or poorly organized, or when a codebase has no structured developer documentation yet.
Use when a PR has customer-visible changes (new features, bug fixes, UI changes) that need a release note entry, or when finalizing a branch before merge.
Use when internal documentation has been updated and external customer-facing docs need to be aligned, or when checking for outdated, missing, or confidential content in external docs.
| name | review |
| description | Use when you need a code-review verdict on a PR or current branch, without auto-applying any fixes. |
| argument-hint | pr-number |
You are the review engine orchestrator. Run a four-phase review and present an APPROVE/REJECT verdict.
Input: Optional PR number as $ARGUMENTS. If omitted, review current branch vs main.
Engine sharing. Phases 0 through 4.3 of this skill are also executed verbatim by /devflow:review-and-fix (which wraps them in a fix loop and skips Phase 4.4 entirely — no GitHub post; its final report is emitted to chat only). When modifying engine behavior here — Phase 3 agent prompts, Phase 1 batching, Phase 0.5 classification, Phase 4 verdict criteria — verify /devflow:review-and-fix still produces the same findings; that's where divergence has historically slipped in. /devflow:review-and-fix's SKILL.md deliberately keeps no paraphrase of these phases, so changes here propagate automatically as long as the file is reachable at the path **/devflow/skills/review/SKILL.md.
/devflow:review-and-fix instead.origin/main (or the PR base) and a long-lived branch diff will swamp Phase 1 with stale items.Run:
git status --porcelain
If there is output, warn: "You have uncommitted changes that will not be included in this review."
If $ARGUMENTS is a PR number:
gh pr diff $ARGUMENTS
gh pr view $ARGUMENTS --json headRefName,baseRefOid,headRefOid --jq '.'
If either command fails (non-zero exit code), stop immediately and report: "Failed to retrieve diff. Verify the PR number exists and you have required permissions."
Use the PR diff output for Phase 1. Store the head branch name, baseRefOid as $PR_BASE_SHA, and headRefOid as $PR_HEAD_SHA — Phase 1's per-file slicing needs them (see Phase 1.1).
Note on gh pr diff path filtering. gh pr diff <N> does NOT support path arguments — gh pr diff <N> -- <file> errors with accepts at most 1 arg(s) (cli/cli#5398, unresolved). When you need per-file slicing in Phase 1.1, use git diff "$PR_BASE_SHA...$PR_HEAD_SHA" -- <paths> instead, or pipe the full gh pr diff through filterdiff -i '<pattern>' if patchutils is installed.
If no argument (review current branch):
git diff origin/main...HEAD
git diff origin/main...HEAD --name-only
If either command fails (non-zero exit code), stop immediately and report: "Failed to retrieve diff. Verify origin/main is reachable and you are on a valid branch."
Use the diff output for Phase 1. The current branch is the review target.
If the diff is empty, report: "No changes to review. Branch is identical to main." and stop.
Cache the diff to disk. Write the diff fetched above to .devflow/review/<slug>/diff.patch — fetch once, do not re-run gh pr diff / git diff. Compute <slug> as:
pr-<N> where <N> is the PR number from $ARGUMENTS./ with -, lowercase, drop any character that isn't [a-z0-9._-]. (Matches the workpad slug convention /devflow:review-and-fix already uses.)Combine the initial fetch with the cache write in one shot using tee so the diff is captured exactly once and stdout remains available for Phase 1 consumption:
mkdir -p .devflow/review/<slug>
gh pr diff $ARGUMENTS | tee .devflow/review/<slug>/diff.patch
# or, in current-branch mode:
# git diff origin/main...HEAD | tee .devflow/review/<slug>/diff.patch
This replaces the bare gh pr diff / git diff invocation at the top of Phase 0.2 — use the tee form instead. Store <slug> and the resolved diff path (e.g. .devflow/review/pr-863/diff.patch) so Phase 3 can substitute it into its agent prompts via {DIFF_PATH}. The directory creation is harmless if it already exists; the file is overwritten on every run.
.devflow/ should be gitignored (it's ephemeral working state). This skill does not add the entry itself (that's a repo-level concern); flag missing .gitignore coverage in the chat output if .devflow/ is not already ignored. When /devflow:review is invoked standalone (not from /devflow:review-and-fix), this cached diff is the only file in the directory — independent of the fix-loop's iter-<N>.json workpad files that live in the same place.
From the diff, extract the list of changed files (use --name-only output or parse from PR diff). Store this list — it's needed for Phase 1 and Phase 3.
Attempt to find the related issue number using these methods in order:
From PR body (look for Resolves #N, Fixes #N, or Closes #N):
If a PR number was provided:
ISSUE_NUM=$(gh pr view $ARGUMENTS --json body --jq '.body' | grep -oiE '(resolves|fixes|closes)[[:space:]]+#[0-9]+' | grep -oE '[0-9]+' | head -1)
If no PR number:
ISSUE_NUM=$(gh pr view HEAD --json body --jq '.body' 2>/dev/null | grep -oiE '(resolves|fixes|closes)[[:space:]]+#[0-9]+' | grep -oE '[0-9]+' | head -1)
From branch name (fallback — matches issue-{number} pattern set by /implement):
if [ -z "$ISSUE_NUM" ]; then
# If reviewing a PR, use the stored head branch name from Phase 0.2
# If reviewing current branch, use git branch --show-current
BRANCH_NAME="${STORED_HEAD_BRANCH:-$(git branch --show-current)}"
ISSUE_NUM=$(echo "$BRANCH_NAME" | grep -oE 'issue-[0-9]+' | grep -oE '[0-9]+')
fi
If an issue number was found, fetch the issue:
gh issue view $ISSUE_NUM --json title,body
Truncation rule: Only use the first 200 lines of the issue body. This captures the summary and desired behavior while skipping excessive implementation detail.
Store the issue title and truncated body as issue_context. If no issue was found, set issue_context to empty and note: "No related issue found — skipping issue compliance check."
Before launching anything, classify the diff. The classification scales agent dispatch so that tiny / config-only PRs don't pay the full engine cost (and so type-design-analyzer is dispatched only when there are actually new types, not when "class" happens to appear as a word elsewhere in the diff).
Compute four flags:
small_diff = (total changed lines < 100) AND (changed-file count ≤ 3)config_only = every changed file has an extension in {.yml, .yaml, .json, .md, .toml, .ini, .lock, .txt}has_new_types = the added-lines slice of the diff (lines starting with + but not +++) contains, in a code file (file extension NOT in the config_only set above), a line that matches ^\+\s*(?:(?:final|abstract|readonly|export(?:\s+default)?|public|pub)\s+)*(class|interface|type|enum|struct|trait)\s+\w+. The optional leading modifiers catch language-specific qualifiers (e.g. final class, abstract class, readonly class, export class, export default class, public class) — without them, the regex would silently miss genuinely-new-type diffs in languages whose declarations begin with a visibility / modality keyword.engine_self_modifying = any changed file's path matches skills/** OR agents/** OR lib/** (the DevFlow engine's own files, which live at the repo root in the devflow-autopilot repo). These are the SKILL.md / agent-definition / helper-script files that are the review engine — a typo here silently breaks every future review. lib/** is included because helper scripts and test fixtures under lib/ are part of the engine surface. (This gate only fires when reviewing a PR against the DevFlow repo itself; on an adopter's repo these paths normally won't match the engine.)Compute counts from the diff already fetched in 0.2/0.3 — no extra gh calls.
Apply the engine profile per the table below. The first row overrides all others when its flag is set; otherwise the remaining rows apply per their combinations. Output one line announcing the chosen profile so the human reader knows the engine ran a leaner path on purpose, not by accident:
| Combination | Engine behavior |
|---|---|
engine_self_modifying (any combination of the other flags) | Override the other flags: run the full engine (no Phase 1+2 skip, no agent gating in Phase 3.1). The risk surface is "every future review breaks if this is wrong," which dwarfs the per-PR cost saving from a leaner profile. |
small_diff AND config_only | Skip Phase 1 + Phase 2 (checklist gen + verify) entirely. Set checklist_skipped = "intentional". In Phase 3.1, skip pr-test-analyzer and pr-review-toolkit:type-design-analyzer. |
config_only (but not small_diff) | Run Phase 1+2 normally. In Phase 3.1, skip pr-test-analyzer and pr-review-toolkit:type-design-analyzer. |
small_diff (but not config_only) | Run Phase 1+2 normally. In Phase 3.1, skip pr-test-analyzer if no test files (*test*, *spec*, language-specific test naming conventions, etc.) appear in the diff. |
| neither flag set | Run the full engine. In Phase 3.1, still apply the has_new_types gate for type-design-analyzer. |
Concretely: when engine_self_modifying is true, the orchestrator does NOT set checklist_skipped = "intentional" regardless of small_diff / config_only, and the Phase 3.1 engine-profile gates listed below are bypassed (every Phase 3 agent in the launch list runs). The override is not an aesthetic tag on the announcement line — it is the load-bearing rule that keeps the full engine wired through Phase 1's skip predicate AND Phase 3.1's per-agent gates for engine-self-modifying diffs.
has_new_types is the canonical predicate for the type-design-analyzer gate in Phase 3.1; the previous heuristic ("check for class in the diff") fires false-positives on YAML/markdown comments and is superseded.
Announce one line, e.g.:
Diff classification: engine_self_modifying (overrides other flags) → running full engine — this diff modifies the review engine itself.Diff classification: small_diff + config_only → skipping Phase 1+2 and pr-test-analyzer + type-design-analyzer.Diff classification: config_only → skipping pr-test-analyzer + type-design-analyzer (Phase 1+2 still run).Diff classification: full engine.Output: Phase 1/4: Generating verification checklist...
Skip this entire phase (and Phase 2) when Phase 0.5 set checklist_skipped = "intentional" (small_diff AND config_only). Proceed directly to Phase 3. The verdict rule in 4.2 distinguishes this intentional skip from a checklist-gen failure.
Count the changed files. If 10 or fewer, launch one checklist-generator agent. If more than 10, split into batches of 10 and launch one agent per batch. Slice the diff to only the batch's files before passing it. To slice:
git diff "$PR_BASE_SHA...$PR_HEAD_SHA" -- <file1> <file2> .... Do NOT use gh pr diff $ARGUMENTS -- <file> — that form errors with accepts at most 1 arg(s) (cli/cli#5398, unresolved). Alternatively, pipe the cached full diff through filterdiff -i '<glob>' if patchutils is installed.git diff origin/main...HEAD -- <file1> <file2> ....^diff --git headers.Passing the full diff to every batch is wasteful and increases dup rate. Tell each batch which other files are being handled by sibling batches so it does not generate items for them.
Merge the resulting checklists by concatenating all items. If batching ran (>1 batch), proceed to Phase 1.5: Dedup before renumbering. If only one batch ran, renumber IDs sequentially (VC-1, VC-2, ...) and skip Phase 1.5.
In-batch sanity dedup still applies before Phase 1.5 hands the array off:
source_file. "Same claim" = same defect/contract under scrutiny, not identical wording (e.g., the same path/format assertion appears in both batches → keep one). When Phase 1.5 runs, this is mostly a no-op — the deduper agent does the heavy lifting via claim_signature..gitignore anchoring — should appear at most once each in the merged list, not once per batch. The category for these is "api_contract" by convention.If the merged-and-deduped checklist has more than 100 items, sort by priority and keep the top 100:
dependency_interaction items (cross-boundary contracts — highest drift risk).test_mock_alignment items (mocks-vs-real divergence is a classic PR-killer).api_contract items.data_format_assumption items.Drop items below the cap. This is a cost cap: every checklist item triggers a verifier subagent in Phase 2. Real-world runs on medium PRs have produced 150+ items when generators are exhaustive on doc-heavy diffs, but the load-bearing signal (cross-boundary contracts, mock-vs-real divergence, issue acceptance) is usually captured well within 100. Announce the cap in chat: Capped checklist at 100 of {N} items (dropped {M} items by category: dependency_interaction: K1, api_contract: K2, ...; priority kept: issue-acceptance, dependency_interaction, ...). so the human reader knows which categories took the hit, not just that coverage was truncated. (In /devflow:review-and-fix mode the same data also lands in the workpad's cap_drops block and the report's ## Coverage section; in standalone /devflow:review runs the chat announcement is the only surface.)
Record what was dropped. When the cap fires, summarize the dropped items by category so the orchestrator can surface coverage gaps in the final report (and the fix-loop wrapper can record it in the workpad — see cap_drops in /devflow:review-and-fix's workpad schema). Compute and return alongside the truncated checklist:
{
"count": M,
"by_category": {
"dependency_interaction": K1,
"api_contract": K2,
"test_mock_alignment": K3,
"data_format_assumption": K4,
"...": "..."
}
}
where M is the total dropped count (N - 100) and the per-category counts sum to M. If the cap did not fire, return {"count": 0, "by_category": {}}. The orchestrator stores this for the report's ## Coverage section in /devflow:review-and-fix and for the chat announcement in standalone /devflow:review runs.
Use the Agent tool with subagent_type: "devflow:checklist-generator".
Pass the following prompt:
Here is the git diff for this PR:
<diff>
{paste the full diff output here}
</diff>
Changed files to analyze:
{paste the file list here}
Generate the verification checklist. Return the JSON array in a ```json code fence.
If issue_context is not empty, append this to the prompt:
The following GitHub issue describes the intended behavior for this PR. In addition to code-correctness items, include checklist items that verify the PR implements the key requirements from the issue's summary and desired behavior sections. Focus on functional requirements — not stylistic suggestions or background context in the issue.
<issue>
Title: {issue_title}
Body (first 200 lines):
{truncated_issue_body}
</issue>
If the caller is /devflow:review-and-fix on iteration N≥2 (the fix-loop wrapper supplies prior_checklist from iter-<N-1>.json), append this to the prompt:
This is iteration N (N≥2) of an auto-fix loop. The previous iteration's verification checklist is supplied below. Operate in variance-recovery mode per your agent contract (Step 2b):
- Generate claims NOT already present in the prior checklist (dedup against `claim_signature`).
- Prioritize claim categories that are underrepresented in the prior iteration.
- The goal is variance recovery — surfacing what a second-look pass would catch — NOT re-litigation of items already considered.
Return an empty JSON array `[]` if a second pass surfaces nothing new.
<prior_checklist iteration="N-1">
{paste the iter-(N-1) checklist JSON — id, category, claim, source_file, claim_signature, verdict}
</prior_checklist>
Extract the JSON array from the agent's response (look for the ```json code fence).
If the agent fails or returns malformed JSON, retry once. If it fails again, log: "Verification checklist generation failed. Proceeding with existing agents only." Set a checklist_skipped flag and skip to Phase 3.
Store the parsed checklist items for Phase 1.5 (if batched) or Phase 2 (if single-batch).
Output: Generated {N} verification checklist items.
When Phase 1 ran a single generator batch, skip this phase entirely — there are no cross-batch duplicates to resolve.
When Phase 1 ran in 2+ batches, dedupe via the devflow:checklist-deduper agent instead of manually. Manual cross-batch dedup is bias-prone (real-run telemetry: orchestrator collapsing ~70 items to ~40 by hand consistently dropped 3–6 legitimate distinct items per run).
Output: Phase 1.5/4: Deduping checklist across {B} batches...
Use the Agent tool with subagent_type: "devflow:checklist-deduper".
Concatenate the raw checklist items from all batches into a single JSON array. Preserve each item's original id and tag it with its source batch so traceability survives the merge — prefix each id with batch{K}: (e.g. batch1:VC-3, batch2:VC-1) before passing to the deduper.
Pass the following prompt:
Here is the concatenated raw checklist from {B} generator batches. Merge duplicates per your dedup rules and return the deduped JSON array. Preserve `merged_from` provenance on every surviving item.
<raw_checklist>
{paste the JSON array of all items from all batches, with batch-prefixed ids}
</raw_checklist>
Extract the JSON array from the deduper's response (look for the ```json code fence). The output array uses fresh sequential IDs (VC-1, VC-2, ...) and records merged_from on each item.
If the deduper agent fails or returns malformed JSON, retry once. If it fails again, fall back to manual cross-batch dedup using the In-batch sanity dedup rules from Phase 1.1 and continue — do NOT block the engine on dedup failure.
Output: Deduped to {N_after} of {N_before} items.
Output: Phase 2/4: Verifying {N} checklist items...
Split the checklist into two groups based on each item's verification_mode field (set by the generator in Phase 1):
verification_mode: "lite") — the orchestrator runs grep -n / rg directly. No agent dispatch. See 2.1a.verification_mode: "agent", or missing/unrecognized) — dispatch the devflow:checklist-verifier agent. See 2.1b.This partition supersedes the old "one verifier agent per checklist item, no batching exceptions" rule. For pure string-presence claims, an orchestrator-direct grep -n is 5–10x cheaper than spawning a verifier subagent and produces an identical verdict. The lite path is bounded to claims that reduce mechanically to substring presence/absence — see checklist-generator.md for the eligibility rules the generator applies.
When invoked by /devflow:review-and-fix on iteration N≥2, iter-(N-1)'s workpad is available and the caller has supplied (a) the iter-(N-1) checklist and (b) the set of files modified by the iter-(N-1) fix commit (fix_files). Before partitioning into lite/agent batches, the orchestrator MAY short-circuit verification for items whose verdicts are mechanically guaranteed to be unchanged.
For each item in the current iteration's checklist, reuse the prior verdict (skip verification) iff ALL of the following hold:
claim_signature.verdict is PASS.source_file is NOT in fix_files (the fix commit did not touch it).For each reused item, copy verdict, evidence, and file_checked from the prior result and tag it reused_from_iter_<N-1>: true in the workpad. Everything else — new items the generator emitted in variance-recovery mode, items whose prior verdict was FAIL or INCONCLUSIVE, items whose source_file was touched by the fix commit — verifies fresh.
Why narrow. The framing the user established: iterations exist for two distinct reasons. Fix-induced defects (did the fix introduce new bugs?) are well-served by file-intersection — a PASS item whose file the fix didn't touch is genuinely unchanged. Variance-recovered defects (did iter-1 miss something a second look would find?) are the opposite — they're the entire purpose of running Phase 1 again, and a coarse "the fix didn't touch any prior-checklist file, so skip Phase 1+2 wholesale" gate would silently dismiss them. The narrow per-item reuse here optimizes only the first case.
Output: Reused {K} of {N} checklist verdicts from iter-(N-1) (matching claim_signature, prior verdict PASS, source_file untouched by fix commit). Verifying remaining {N-K} fresh.
For each lite item, execute the probe described in lite_probe:
kind: "string_present" — run grep -nF -- "<string>" <file> (or rg -nF "<string>" <file> if available). If a line_range is present, additionally check that at least one hit falls inside [L1, L2] (inclusive). Verdict: PASS if any in-range hit (or any hit when no range), FAIL otherwise.kind: "string_absent" — run the same grep. Verdict: PASS if no hit; FAIL if any hit.Use fixed-string mode (-F) by default — lite_probe.string is a literal, not a regex. Escape shell-special characters by quoting.
Edge cases:
evidence: "file not found".lite_probe field missing despite verification_mode: "lite" (malformed item) → promote the item to the agent path; do not silently PASS.grep exit code 2 (real error, not just no-match) → INCONCLUSIVE with the stderr text in evidence.Record the result in the same JSON shape as agent verdicts:
{"id": "VC-N", "verdict": "PASS|FAIL|INCONCLUSIVE", "evidence": "lite probe: 2 hits in lines 113, 117", "file_checked": "path/to/file.py"}
Examples:
claim: "License header <expected literal> appears in path/to/new_source_file". lite_probe: {kind: "string_present", string: "<expected literal>", file: "path/to/new_source_file"}. The orchestrator greps; no agent needed.claim: "Mock return value of <symbol> in path/to/test_file matches the real signature in path/to/impl_file". Two files, semantic shape comparison — must dispatch the verifier.Split the agent items into batches of up to 8. For each batch, launch all agents in parallel using multiple Agent tool calls in a single message.
Use the Agent tool with subagent_type: "devflow:checklist-verifier" for each item.
Pass the following prompt for each:
Verify this claim against the actual source code. Read the referenced files, compare the claim to reality, and report PASS, FAIL, or INCONCLUSIVE.
Checklist item:
{paste the JSON checklist item here}
The `source_line` field (if present) is best-effort from the generator and may be approximate. Treat it as a starting hint; if the symbol/claim isn't at that line, grep the file for the relevant identifier rather than reporting INCONCLUSIVE. Report INCONCLUSIVE only when the source of truth is genuinely unreachable (file missing, claim too vague to locate, external API not consultable).
When a claim's wording is technically inaccurate but the underlying code is correct (e.g., the claim oversimplifies a branch the code handles correctly), prefer **PASS** with an evidence note explaining the wording-vs-code distinction. Reserve FAIL for cases where the code itself is wrong or contradicts the claim's intent.
Report your verdict as JSON in a ```json code fence: {"id": "VC-N", "verdict": "PASS|FAIL|INCONCLUSIVE", "evidence": "...", "file_checked": "..."}
Collect verdicts from BOTH paths — lite probes (2.1a) and agent batches (2.1b). Parse the JSON verdict from each agent response.
If an agent times out or fails, record that item as:
{"id": "VC-N", "verdict": "INCONCLUSIVE", "evidence": "Verifier agent failed or timed out.", "file_checked": "N/A"}
Store all verification results in a single combined array (lite + agent), keyed by id.
Output: Verified: {pass_count} passed, {fail_count} failed, {inconclusive_count} inconclusive ({lite_count} via lite probe, {agent_count} via agent).
Output: Phase 3/4: Running review agents...
Launch all agents in a single message using multiple Agent tool calls. For each agent, pass a prompt telling it to review the changes.
Phase 3 always re-runs on every iteration of the fix loop. Unlike Phase 1+2 (where individual items can be narrow-reused via claim_signature + untouched-file checks — see Phase 2.0.5), Phase 3's review agents are the main lever for variance recovery: an LLM reviewer asked the same question twice in different sessions will not always surface the same findings, and that variance is the whole point of iterating. Skipping Phase 3 on a later iteration because "the fix didn't touch any flagged file" silently throws away the second-look signal — exactly the false-pass mode this engine is designed to avoid.
Prior-findings context (fix-loop callers only). When invoked by /devflow:review-and-fix on iteration N≥2, prepend the following block to every Phase 3 agent's prompt (between the standard task description and the defect_signature paragraph). The caller supplies iter-(N-1)'s phase3_findings from the workpad:
The following findings were raised by a prior review pass on this same code and have already been considered (some fixed, some pushed back as false positives, some deferred). Treat them as PRIOR ART, not as a checklist to re-derive:
- Do NOT re-raise a finding identical to one in the prior set unless you have new evidence the prior decision was wrong.
- DO look for *new* defects the prior pass missed — your value on this iteration is variance recovery, not corroboration.
- If you would have raised an identical finding, you may skip it; the orchestrator already has it.
<prior_findings iteration="N-1">
{paste the iter-(N-1) phase3_findings JSON — agent, severity, description, defect_signature, fix_decision}
</prior_findings>
Diff path: Substitute the cached diff path computed in Phase 0.2 (.devflow/review/<slug>/diff.patch) into {DIFF_PATH} in the prompts below. Phase 3 agents Read this file directly via their Read tool — no shell command, no gh API call, no redundant re-fetches across the 4–5 parallel agents. The previous {DIFF_CMD} substitution (which had every agent re-run gh pr diff $ARGUMENTS or git diff origin/main...HEAD) is superseded.
Required defect_signature block. Every Phase-3 finding from every Phase-3 review-agent — both the ones listed below AND any added by future maintainers — MUST carry a defect_signature object so corroboration (Phase 3.2) is mechanical, not interpretive. Append this paragraph verbatim to every Phase-3 review-agent prompt — it's the only way to instruct external pr-review-toolkit agents we cannot edit:
For every finding you report, include a `defect_signature` field with the following shape:
defect_signature:
file: "<path/to/file>" # required; the primary file the defect lives in
line_range: [<start>, <end>] # required when locatable; null only when the defect spans an unbounded region (e.g. "missing test file")
kind: "<one of: null_deref | unhandled_exception | leak | race | logic_error | api_misuse | type_design | comment_drift | test_gap | security | style | other>"
Place this field on each finding alongside severity and description. If your normal output format is a markdown bullet list, append the signature as a fenced JSON block right under the bullet. Without `defect_signature`, the orchestrator cannot corroborate your finding against other agents and may downweight it.
Agents to launch:
pr-review-toolkit:code-reviewer — prompt:
Review the code changes in this PR. Read the cached diff at `{DIFF_PATH}`. Read CLAUDE.md for project conventions. Focus on CLAUDE.md compliance, bugs, and code quality. Only report issues with confidence >= 80.
{paste the defect_signature paragraph above}
pr-review-toolkit:silent-failure-hunter — prompt:
Review the error handling in the code changes. Read the cached diff at `{DIFF_PATH}`. Read the full changed files. Check for silent failures, inadequate error handling, and inappropriate fallback behavior.
{paste the defect_signature paragraph above}
pr-review-toolkit:comment-analyzer — prompt:
Analyze the code comments in the changes. Read the cached diff at `{DIFF_PATH}`. Check that docstrings and comments are accurate, helpful, and not misleading.
{paste the defect_signature paragraph above}
pr-review-toolkit:pr-test-analyzer — prompt:
Analyze test coverage for the changes. Read the cached diff at `{DIFF_PATH}`. Check if tests adequately cover new functionality and edge cases.
{paste the defect_signature paragraph above}
pr-review-toolkit:type-design-analyzer — launched only when the has_new_types gate is true (see Phase 3.1 gates below), and always when engine_self_modifying is set; skipped otherwise — prompt:
Analyze the type design in the code changes. Read the cached diff at `{DIFF_PATH}`. Evaluate the types actually introduced or modified in this diff for encapsulation, invariant expression, usefulness, and enforcement. Do not report on pre-existing types the diff does not touch.
{paste the defect_signature paragraph above}
General-purpose final-pass reviewer — dispatch a Task with subagent_type: general-purpose and instruct it to invoke the /superpowers:requesting-code-review skill (that skill renders its own reviewer prompt; we do not inline it). This dispatch assumes the superpowers plugin is installed in the executing environment; if /superpowers:requesting-code-review is not available, the subagent will surface that and the orchestrator should fall back to relying on the other Phase-3 reviewer agents above.
Prompt:
Invoke the `/superpowers:requesting-code-review` skill to perform a final-pass code review. Pass the following context into the skill:
- Description: {one-line summary — "PR #<N>: <title>" or "Current branch <name> vs main"}
- Plan / Requirements: {the PR body if available, else the originating issue body from Phase 0.4, else "No spec available — review against general project standards from CLAUDE.md"}
- Base SHA: {PR_BASE_SHA or origin/main HEAD}
- Head SHA: {PR_HEAD_SHA or current HEAD}
- Diff path: `{DIFF_PATH}` (the full diff, cached to disk by Phase 0.2 — Read it directly rather than re-fetching)
- Prior-iteration findings (already considered, look for new): {iter-(N-1) phase3_findings JSON if fix-loop iteration N≥2, else "none"}
Return your findings in the standard Phase-3 output format: ### Strengths / ### Issues (grouped by Critical / Important / Suggestion) / ### Recommendations / ### Assessment. Every issue MUST carry a `defect_signature` block per the contract below.
{paste the defect_signature paragraph above}
Phase 0.5 engine-profile gates (apply to this launch list):
These gates are BYPASSED entirely when engine_self_modifying is set in Phase 0.5 — every Phase 3 agent in the launch list runs regardless of config_only / small_diff / has_new_types for engine-self-modifying diffs (see Phase 0.5's override row). Apply the gates below only when engine_self_modifying is false.
pr-review-toolkit:pr-test-analyzer when config_only is set, OR when small_diff is set AND no test files appear in the diff.pr-review-toolkit:type-design-analyzer when has_new_types is false. (This replaces the older "check for class in the diff" predicate, which over-fired on the literal word class appearing in YAML / markdown / comments.)In all other cases, pr-review-toolkit:type-design-analyzer is launched when has_new_types is true.
Collect all agent responses. Extract findings, their severity labels (Critical, Important/Major, Suggestion/Minor), and their defect_signature blocks.
For each finding, compute a corroboration count — the number of Phase 3 agents that raised the same defect. Corroboration is now mechanical, not interpretive:
Two findings corroborate iff they have the same
defect_signature.file, overlappingdefect_signature.line_range(treatnullas overlapping any range in the same file whenkindmatches), AND identicaldefect_signature.kind.
A finding without a defect_signature block falls back to a one-line text-based agreement heuristic (same described file + same described defect kind in prose), but flag it in the report so the human knows the agent skipped the signature contract. Agents that systematically omit defect_signature should be re-prompted with the contract reminder.
Corroboration count is a stronger calibrator than the individual agent's verbalized confidence: a finding raised by 3 of 5 agents is much more likely to be a true positive than a 95%-confidence finding raised by only one. Single-source findings are not automatically wrong — they're flagged so a human reader can apply extra scrutiny.
If an agent fails, note: "[agent-name] did not return results." in the report. Track the count of failed agents. Failed agents do not reduce the denominator for the corroboration count of findings other agents raised.
Output: Phase 4/4: Aggregating findings...
Skip this step entirely in current-branch mode (no PR → no body to read). On standalone branch reviews, there is no Scope-Acknowledged Findings block; jump straight to 4.1.
When $ARGUMENTS is a PR number, the engine consults the Scope-Acknowledged Findings block in the PR body (delimited by <!-- DEVFLOW_DEFERRED_FINDINGS_START --> / <!-- DEVFLOW_DEFERRED_FINDINGS_END -->) and demotes any current finding that matches a validated deferral entry to Informational. This is the consumer side of the contract /implement Phase 4.0.5 produces; without it, /devflow:review re-raises findings that /implement already filed follow-up issues for, creating the policy mismatch the contract is meant to prevent. (See ${CLAUDE_SKILL_DIR}/../../scripts/match-deferrals.py for the matcher's exact guard order and matching rule.)
Serialize the Phase 3 findings collected in 3.2 to a JSON array with one object per finding:
[
{"file": "...", "line_range": [N, M], "kind": "...", "description": "...",
"severity": "Critical|Important|Suggestion", "agent": "..."}
]
The order matters — index N in this array becomes the matcher's finding_index reference.
Pipe the JSON to the matcher via stdin (the review allowed-tools profile in claude-runner.yml is read-only and does not grant the Write tool, so the orchestrator cannot write a findings.json file; stdin is the load-bearing alternative):
printf '%s' "$FINDINGS_JSON" | ${CLAUDE_SKILL_DIR}/../../scripts/match-deferrals.py \
--pr $ARGUMENTS \
--diff ".devflow/review/<slug>/diff.patch" \
--findings -
Capture the matcher's stdout (the JSON report described below). When invoked from /implement Phase 3.3 via /devflow:review-and-fix (which DOES have the Write tool), the file form --findings .devflow/review/<slug>/findings.json is equally supported — pick whichever the surrounding profile permits.
The matcher always exits 0 when it ran (any result, including no block found). Read the output JSON:
block_present: false → PR has no Scope-Acknowledged Findings block; proceed to 4.1 with all findings intact.pr_author_trusted: false → PR author is not in claude.allowed_bots; every deferral is rejected with reason untrusted-filer. All findings flow through unchanged. Include the rejection list in 4.1's ## Deferrals section so the human reader sees the contract was claimed but not honorable.honored[]: the finding at findings[finding_index] is demoted to Informational for the rest of Phase 4. Record the deferral_id + follow_up_issue so the 4.1 line annotation can cite them.rejected_deferrals[]: the deferral did not apply (issue closed, missing cross-link, widens-surface re-check failed, or no matching current finding). The corresponding current finding (if any) is not demoted — flag it explicitly in 4.1's ## Deferrals section with the reason.If the matcher itself errors out (exit code 2), log the failure (Deferral matcher failed: {stderr}; proceeding without demotions.) and continue to 4.1 with all findings intact. Never block the review on a matcher failure — the safe default is to surface findings, not hide them.
Caching note. The matcher hits the GitHub API once for the PR body + author and once per follow_up.issue for the cross-link guard. For a PR with N deferrals, this is N+1 API calls. Tolerable; if it ever becomes a bottleneck, batch the issue reads via gh api graphql.
Construct the report in this format:
# Review Report
## Verdict: {APPROVE | APPROVE with notes | APPROVE WITH CAVEAT | APPROVE WITH ADVISORY NOTES | REJECT} ({summary})
## Issue Compliance
{If issue found: "Reviewed against issue #{number}: {title}. Requirement-based checklist items are included in the verification results below."}
{If no issue found: "No related issue found — requirement compliance not checked."}
## Verification Checklist Results
- ({total} checked, {pass} passed, {fail} failed, {inconclusive} inconclusive)
{for each FAIL or INCONCLUSIVE item: "- VC-N: VERDICT — claim [source_file:source_line]"}
PASS items are summarized in the count line above; do not list them individually. Callers that render the report in environments supporting collapsible Markdown (e.g. GitHub PR comments) MAY wrap a per-item PASS list in a `<details>` block, but the skill itself does not emit one.
## Code Review Findings
{for each finding: "- [agent-name] severity: description (raised by N/{total Phase 3 agents that returned results} agents)"}
{for findings whose index appears in the matcher's honored[] list, append " [Deferred → #{follow_up_issue}]" to the line and render the finding under a separate sub-heading "### Informational — Deferred" rather than under its original severity bucket.}
{group Critical findings first, then Important/Major, then Suggestion/Minor, then Informational — Deferred. Within each severity, list corroborated findings (N≥2) before single-source ones (N=1) so the highest-confidence items lead.}
## Deferrals
{Omit this section entirely when 4.0 was skipped (current-branch mode) or block_present was false. Otherwise render:}
- Honored: {stats.honored}
{for each honored entry: " - {deferral_id} → #{follow_up_issue} ({category})"}
- Rejected: {len(rejected_deferrals)}
{for each rejected entry: " - {deferral_id} — rejected: {reason}"}
{If pr_author_trusted is false, prepend a single line: "**Block claimed but not honored — PR author is not in `claude.allowed_bots`. All deferrals rejected.**"}
## Verdict Criteria
- Any FAIL in verification checklist → REJECT
- Any INCONCLUSIVE in verification checklist → REJECT (manual check needed)
- Any Critical finding from review agents → REJECT (excluding findings demoted to Informational via Phase 4.0's deferral match)
- Checklist generation failed → max APPROVE WITH CAVEAT
- 2+ review agents failed → partial review coverage
- Only Important/Suggestion findings → APPROVE with notes
- No findings → APPROVE
Apply these rules in order (first match wins). For every rule that counts findings by severity, exclude findings demoted to Informational by Phase 4.0's deferral match — they appear in the report under the "Informational — Deferred" sub-heading but do not contribute to verdict computation. (Rejected-deferral entries do not demote their corresponding finding; those flow through at their original severity.)
checklist_skipped = "failure") → maximum verdict is APPROVE WITH CAVEAT — verification checklist not generated (never a clean APPROVE)
4b. If Phase 1+2 were skipped intentionally by Phase 0.5 (checklist_skipped = "intentional", i.e. small_diff AND config_only) → no caveat; the verdict follows the remaining rules normally. The skip was a deliberate engine-profile choice for a low-risk diff, not a failure.Output the full report to the user.
If — and only if — $ARGUMENTS is a PR number (you are reviewing an actual PR, not the current branch), you MUST also submit the verdict as a formal GitHub Pull Request review so it becomes a visible merge signal. A REJECT verdict that lives only in a comment or in chat output is routinely missed — the PR gets marked ready and merged with the rejection still outstanding. A --request-changes review blocks the merge button (or, at minimum, forces an explicit dismissal), which is the behavior we want.
Map the verdict to a gh pr review action. The --body is a short verdict-only stub, not the full report — the auto-trigger workflow's progress comment carries the full Phase 4.1 report verbatim, and posting it in both places forces reviewers to scroll past two copies. Construct $STUB as:
## Verdict: {VERDICT} — full report in PR comment
> The complete review report (checklist results, findings, details) is in the
> Devflow Review progress comment on this PR.
where {VERDICT} is the actual verdict line (e.g. APPROVE, APPROVE with notes, APPROVE WITH CAVEAT, REJECT) — reflect what Phase 4.2 decided, do not template-fill literally. The ## Verdict: REJECT prefix is load-bearing: finalize_check greps for it on the gh pr comment fallback path.
| Verdict | Command |
|---|---|
| REJECT (any form) | gh pr review $ARGUMENTS --request-changes --body "$STUB" |
| APPROVE WITH CAVEAT / APPROVE with notes | gh pr review $ARGUMENTS --comment --body "$STUB" |
| APPROVE (clean, no findings) | gh pr review $ARGUMENTS --approve --body "$STUB" |
If gh pr review fails (e.g. you cannot review your own PR as the same GitHub identity, or the token lacks permission), fall back to gh pr comment $ARGUMENTS --body "$REPORT" — use the full $REPORT here (not $STUB), since this fallback comment is the only artifact in that path. Note in your chat output that the formal review could not be posted. Never silently skip this step on a REJECT — the whole point is that the rejection must be impossible to miss.
Then, on any APPROVE form only (APPROVE / APPROVE with notes / APPROVE WITH CAVEAT), clear a stale REJECT. A prior REJECT's --request-changes review stays the PR's effective reviewDecision until dismissed; the APPROVE-with-notes --comment review never supersedes it, and the REJECT may be a different bot identity (auto path posts as github-actions[bot], manual @claude as another), so no later review clears it either. Without this the PR is wedged at reviewDecision: CHANGES_REQUESTED forever, contradicting the green check and this APPROVE. The script dismisses only Devflow Review's own reports (body marker), never a human reviewer's --request-changes. On REJECT, skip this — the changes-request must stand. Run (re-run safe):
${CLAUDE_SKILL_DIR}/../../scripts/dismiss-stale-rejections.sh "$ARGUMENTS"
If it exits non-zero (token scope), say so in chat output and that the PR stays blocked until dismissed manually. A dismissal failure never downgrades the verdict — the verdict stands; only merge-gate housekeeping failed.
small_diff + config_only — Phase 0.5 already gates this; trust the classification rather than second-guessing it.pr-review-toolkit:type-design-analyzer on a diff where has_new_types is false — the gate exists because that analyzer over-fires when the word class appears in YAML, markdown, or comments. Honor the gate.gh pr review --request-changes — Phase 4.4 exists because chat-only rejections get missed and the PR ships anyway.CHANGES_REQUESTED review (Phase 4.4 final step) — "the required check is green so it'll merge" is the trap: a sticky changes-request keeps reviewDecision: CHANGES_REQUESTED and wedges the PR despite the green check and APPROVE verdict.engine_self_modifying override — the first row of the table overrides all others; the full engine must run on engine-self-modifying diffs because typos in SKILL.md or agent files silently break every future review./devflow:review-and-fix's Step 2.5 web-verification gate for single-source Critical findings — auto-applied fixes from confidently-stated-but-wrong external-tool claims are a known false-positive vector. (This skill itself doesn't run Step 2.5; flag it as a mistake when reviewing changes to /devflow:review-and-fix.)