pr-review

name	pr-review
description	Executes a risk-tiered, multi-perspective PR review — triages the change, runs specialist subagents in parallel (code-first, testing, skeptical, big-picture) plus an external model pass scaled to risk, then posts a consolidated review to the PR.
license	LGPL-3.0

PR Reviewer

Run a comprehensive PR review covering all four Freenet review perspectives, plus an external (non-Claude) model, and post a consolidated review to the PR.

When to Use

Invoke /freenet:pr-review <PR-NUMBER> after a PR is ready for review, before merging. The PR number is passed as the skill argument ($1); if none is given, detect the PR for the current branch with gh pr view --json number -q .number.

How This Skill Works

This skill orchestrates a review — it does not do all the perspectives by hand. It checks out the PR, triages the change to a risk tier, spawns specialist subagents in parallel (scaled to that tier) plus an external model pass, reconciles their findings into one report, and posts that report to the PR.

The four subagents ship with this plugin as first-class agent types — invoke them directly with the Agent tool's subagent_type parameter. Do not paste agent definitions into a general-purpose prompt; that is obsolete.

Step 1: Check Out the PR, Gather Context, and Triage Risk

Critical: reviewers must read the PR's actual code, and the review must not disturb the user's working tree. Check the PR out into a dedicated worktree — do NOT use gh pr checkout, which switches the user's working branch and drags any uncommitted changes onto the PR branch, contaminating the review.

PR=<PR-NUMBER>          # the $1 skill argument
BASE="$(gh pr view "$PR" --json baseRefName -q .baseRefName)"   # the PR's base branch

git fetch origin "$BASE"                 # fresh base branch for the diff
git fetch origin "pull/$PR/head"         # PR head — FETCH_HEAD now points here
REVIEW_DIR="${TMPDIR:-/tmp}/pr-review-$PR"
git worktree remove --force "$REVIEW_DIR" 2>/dev/null || true   # prune a stale prior worktree
git worktree add --detach "$REVIEW_DIR" FETCH_HEAD              # PR code, isolated
cd "$REVIEW_DIR"                                                # run the review from here

Gather context (from the worktree):

gh pr view "$PR"
gh pr diff "$PR" --name-only
gh pr checks "$PR"                    # CI status
gh issue view <ISSUE_NUMBER>          # linked issue (from "Fixes #" / "Closes #")

# Existing review feedback — read it so the review ADDRESSES it, not duplicates it:
gh pr view "$PR" --json comments,reviews
gh api repos/{owner}/{repo}/pulls/"$PR"/comments   # inline review comments

The gh api .../comments call above is the reliable way to get inline comments — they are easy to miss. (If your environment provides a gh-pr-interactions skill, it documents the comment API in more depth.)

Large diffs: if the diff exceeds ~2000 changed lines or ~40 files, instruct each subagent to review by file batches rather than loading the whole diff into context.

Cleanup (mandatory — do this even if the review aborts partway): remove the worktree with git worktree remove --force "$REVIEW_DIR".

Do NOT run a code-simplifier or any other mutating step here — this skill reviews the PR as submitted; editing the checked-out code would make reviewers judge something other than the PR.

Pick the Risk Tier

Choose the tier by checking these in order — first match wins — or honor an explicit tier the user named (e.g. "full review of PR 42"):

1. Skip — the entire diff is mechanical: typo / comment-only / formatting / version bump / CHANGELOG-only. Judge this against the whole diff, not the PR title — a PR labelled "version bump" that also edits logic is not Skip. Report "trivial — CI is the only gate" and stop; do not spawn reviewers.
2. Full — the change touches a high-risk surface (below), OR the diff is large or cross-cutting, OR you are genuinely uncertain. Run the complete process.
3. Light — everything else: a low-to-moderate-risk change with no high-risk surface. Spawn a reduced reviewer set (Step 2) plus the external model pass (Step 3).

These tiers are exhaustive — Light is the catch-all. When torn between two, pick the heavier one.

High-risk surfaces — always Full: concurrency / async, cryptography / security / auth, state authorization, data or schema migration, wire format / protocol / serialization (freenet-stdlib enums), consensus / routing, transport / NAT traversal, contract or delegate WASM, deploy / release / CI config.

State the chosen tier and the one-line reason before proceeding.

Step 2: Spawn the Review Subagents in Parallel

Spawn the reviewers with the Agent tool in a single message so they run concurrently, each with run_in_background: true. Which reviewers run depends on the tier picked in Step 1:

• Full — spawn all four.
• Light — spawn freenet:skeptical-reviewer; also spawn freenet:big-picture-reviewer if the diff removes code or spans multiple components.

subagent_type	Perspective
freenet:code-first-reviewer	Reads the code before the description; flags gaps between stated intent and implementation
freenet:testing-reviewer	Test-coverage gaps at unit / integration / simulation / E2E levels
freenet:skeptical-reviewer	Adversarial — bugs, race conditions, edge cases, failure modes
freenet:big-picture-reviewer	Goal alignment, removed tests/fixes, scope creep, stale skills/docs

These subagent_type values are plugin-namespaced: freenet: is the plugin name and the files in agents/ carry the bare name (skeptical-reviewer, etc.). Use the freenet:-prefixed form exactly as shown.

In each subagent's prompt, include: the PR number, the repo (owner/repo), and the path to the review worktree from Step 1 ($REVIEW_DIR), so the agent can Read/Grep the PR's actual code — not just the diff — for surrounding context. Each agent already carries its own review methodology; you do not need to supply it.

Step 3: External Model Review (runs concurrently with Step 2)

Run this for both Light and Full tiers — the external model is the highest-value single pass, because its blind spots do not correlate with Claude-authored code. Spawn it concurrently with Step 2's subagents. Run a non-Claude model from the review worktree, diffing against the PR's base branch ($BASE, fetched fresh in Step 1) — never a possibly-stale local main:

codex review --base "origin/$BASE"

(If your environment provides a codex-review skill, it wraps this command.) For a Full review of a high-risk PR, add a third independent model when one is available — e.g. a gemini-cli-review skill or the gemini CLI.

When external models are unavailable — fall back, never skip

External reviewers go down (quota exhaustion, capacity limits, API outages). If codex fails, try gemini; if both are genuinely unavailable, do not skip the independent pass and do not fail the review. A codex review that returns no findings summary, errors out, or reports exhausted budget counts as unavailable — retry once, then treat it as down.

Prefer waiting when you can. If the change is not time-sensitive and the external quota reset is near (within a few hours, or by the next working session), prefer to wait for the external model and note the blockage on the PR — that preserves the strongest signal. Fall back to the Claude-lens pass only when waiting isn't practical (the reset is far off, the change is needed sooner, or the user asked you to proceed).

Substitute a diverse-Claude-lens pass (per ~/.claude/rules/multi-model-review.md). A single extra Claude reviewer is not enough — it shares the author model's blind spots, which is the whole reason the external pass exists. Instead spawn at least three independent reviewers, each with a DISTINCT adversarial lens and each blind to the others, scaling to 4–5 for a Full-tier high-risk change:

• skeptical bug-hunt — assume bugs exist; hunt races, edge cases, failure modes.
• code-first — read the code before the PR description; flag intent/impl mismatches.
• the failure mode specific to THIS change — e.g. for a wire/protocol change: compat + serialization; for auth: authz bypass; for a migration: data-loss / rollback.

These can be the existing reviewer subagents plus extra adversarial lenses, or general-purpose subagents with explicit lens prompts. Each reads the actual checked-out code and is told NOT to rubber-stamp. Synthesize them the same way as Step 5.

If you cannot spawn subagents (you are yourself a background or dispatched agent without the Agent/Task tool), do not skip and do not fail the review: run the lenses serially within your own context instead — one distinct adversarial pass after another (not a single combined read), then synthesize. Serial-within-self is the required fallback whenever spawning subagents is unavailable.

Record the substitution in the posted review: which external models you tried, the exact failure (quota / capacity / outage), that you used the Claude-lens fallback, and which lenses ran — so a reader can see that independent review happened and why it took this form. When the external models come back before merge and the change is high-risk, prefer running the real external pass too rather than relying on the fallback alone.

Step 4: Freenet Bug-Pattern Check

Do this for Full reviews, and for Light reviews whose change has non-trivial logic.

When reviewing freenet-core, the canonical and continuously-updated bug-pattern list lives at .claude/rules/bug-prevention-patterns.md in that repo. Read it and check the PR against every pattern listed there — it supersedes any snapshot in this skill or in the subagent definitions.

Recurring patterns (non-exhaustive — the in-repo file is authoritative): biased; select starvation, fire-and-forget spawns, incomplete state cleanup on failure, backoff without jitter, .send().await on bounded channels inside event/recv loops, protocol-enum / wire-format breaks for older consumers, paired Option fields that must co-occur, and manually-mirrored telemetry counters that rot after op migrations.

Step 5: Synthesize — Reconcile and Verify

When all spawned subagents and the external model pass have returned, do not just concatenate their reports. Synthesize:

1. Deduplicate — the same finding will surface from multiple reviewers; merge it into one entry.
2. Reconcile contradictions — if two reviewers disagree, investigate the code yourself and decide; note the disagreement in the report if it is genuinely open.
3. Verify before reporting — subagents and Codex can cite wrong line numbers or hallucinate. Open every cited file:line and confirm the finding is real before it goes in the report. Drop false positives; downgrade speculative ones to questions.
4. Classify severity — map every surviving finding to Must Fix / Should Fix / Consider (see Output Format).
5. Check documentation — new public items have doc comments? CLI/config changes documented? Any existing doc now contradicted by the change?

Step 6: Post the Consolidated Review to the PR

Post the synthesized report to the PR as a review comment:

gh pr review "$PR" --comment --body-file <report-file>

Use --comment — not --approve or --request-changes. This skill produces a review; it does not gate merge or speak for a human approver. Also print the report in the conversation. End the posted body with [AI-assisted - Claude].

Output Format

Produce a consolidated review report:

## Comprehensive PR Review: #<NUMBER>

### Summary
- **PR Title:** <title>
- **Type:** <feat/fix/refactor/etc>
- **CI Status:** <passing/failing/pending>
- **Linked Issues:** <issue numbers or "none">
- **Review tier:** <Light / Full>
- **Reviewers run:** <which reviewers and external models actually ran>

---

### Code-First Analysis
**Independent Understanding:** <what the code appears to do>
**Stated Intent:** <from PR description>
**Alignment:** <matches/partially matches/misaligned>
**Gaps:** <any discrepancies>

---

### Testing Assessment
**Coverage Level:** <adequate/insufficient/excessive>

| Test Type | Status | Notes |
|-----------|--------|-------|
| Unit | ✅/❌/⚠️ | <details> |
| Integration | ✅/❌/⚠️ | <details> |
| Simulation | ✅/❌/N/A | <details> |
| E2E | ✅/❌/N/A | <details> |

**Regression Test:** <present and valid / missing / insufficient>
**Missing Tests:** <list specific gaps>

---

### Skeptical Findings
**Risk Level:** <low/medium/high>

| Concern | Severity | Location | Details |
|---------|----------|----------|---------|
| <issue> | <high/med/low> | <file:line — verified> | <explanation> |

---

### Big Picture Assessment
**Goal Alignment:** <yes/partial/no>
**Anti-Patterns Detected:** <list or "none">
**Removed Code Concerns:** <list or "none">
**Scope Assessment:** <focused/some creep/significant creep>

---

### Documentation
- Code docs: <complete/incomplete/missing>
- Architecture docs: <up-to-date/needs-update/n/a>
- User docs: <up-to-date/needs-update/n/a>

---

### Recommendations

#### Must Fix (Blocking)
1. <critical issues that must be addressed>

#### Should Fix (Important)
1. <significant issues that should be addressed>

#### Consider (Suggestions)
1. <minor improvements or style suggestions>

---

### Verdict
**State:** <Ready to Merge / Needs Changes — Re-review Required After Fix / Needs Changes — Light Re-Check Sufficient / Needs Discussion>
**HEAD SHA reviewed:** <sha>

<If not ready, summarize what's needed. If re-review is required, list which findings triggered it so the next pass can confirm they're addressed.>

[AI-assisted - Claude]

Re-Review After Fixes (Mandatory When Findings Were Significant)

A review is per-CODE-CONTENT, not per-PR. If the review surfaced significant problems and the author then pushed fixes, the previously-completed review is stale and does not authorize merge — the new code has not been reviewed. Run the full review process again on the updated HEAD.

When re-review is required

Re-run the full review (all of Steps 1–6) if any of the following is true after fixes are pushed:

• One or more "Must Fix" (blocking) findings were addressed — even a single blocking fix can introduce its own bugs.
• Three or more "Should Fix" findings were addressed in aggregate — many small changes compound into substantial new surface area.
• Any finding in a high-risk area was addressed: security, cryptography, state authorization, contract/delegate migration, concurrency, persistence, wire format, deployment.
• The diff changed by more than ~30 lines in response to review, regardless of severity classification.
• Conflict resolution during rebase, git commit --amend, or any force-push altered the diff vs. main — this counts as new code even if no review finding prompted it. (See ~/.claude/rules/multi-model-review.md.)

When in doubt, re-review. The cost of a second pass is small; the cost of merging unreviewed code is large.

When a follow-up pass can be lighter

Only skip re-review if all of these hold:

• All findings were trivially mechanical (typo, comment wording, rename of a private symbol).
• No "Must Fix" findings existed.
• The diff vs. the previously-reviewed HEAD is purely additive and obviously safe (e.g. a CHANGELOG line, a version bump, a single // TODO comment).
• No rebase or amend occurred.

In that case, do a focused diff-of-the-diff check (just the new commits since the prior review) and note in the verdict that a light re-check was performed rather than a full re-review.

How to re-review

1. Confirm CI is green on the current HEAD SHA, not the SHA you reviewed last time:

   gh pr view <NUMBER> --json statusCheckRollup,headRefOid
   

2. Re-run all of Steps 1–6 against the new HEAD — re-create the review worktree, re-triage, re-spawn the tier's reviewers, and re-run the external model pass. Read the full diff again, not just the fix commits — context shifts when code moves. Re-triage against the whole PR diff vs. base, not just the fix commits: the tier may rise but never falls below the original review's tier (a PR that needed a Full review does not become Skip because the last fix was one line).
3. Produce a fresh review report. In the Verdict section, explicitly note this is a re-review and reference the prior review's findings: "Re-review after fixes for findings #1, #3, #5 from prior pass — all resolved; one new concern at X."
4. If the new review surfaces its own significant findings, repeat the cycle. There is no cap on rounds; merge only when a review pass on the current HEAD comes back clean (or with only trivial findings the user accepts).

What the re-review verdict states mean

The Verdict states (defined in Output Format) are:

• Ready to Merge — review pass on current HEAD is clean or only trivial findings remain.
• Needs Changes — Re-review Required After Fix — significant findings exist; author must address them AND a fresh review must run on the updated HEAD before merge.
• Needs Changes — Light Re-Check Sufficient — only trivial mechanical findings; a diff-of-the-diff pass on the fix commits is enough.
• Needs Discussion — design-level disagreement that fixes alone won't resolve.

Never use "Ready to Merge" on the basis of a review whose HEAD SHA differs from the current PR HEAD.

Quality Standards

• Be thorough but constructive
• Cite specific files and line numbers — and verify each citation before reporting it
• Explain WHY something is a problem, not just WHAT
• Suggest fixes, not just criticisms
• Acknowledge what's done well
• Don't nitpick style if it matches codebase conventions