srnnkls/review

Pre-loaded Context

Git status: !git status --short 2>/dev/null

Recent commits: !git log --oneline -5 2>/dev/null

Current branch: !git branch --show-current 2>/dev/null

Active scopes: !find scopes -maxdepth 3 -name scope.md 2>/dev/null

Open PRs: !gh pr list --limit 5 --json number,title,headRefName --jq '.[] | "#\(.number) \(.title) (\(.headRefName))"' 2>/dev/null

Review Dispatcher

Routes to the appropriate review skill based on argument type.

---

Auto-Detect Rules

Pre-parse: Extract --reviewers <aliases> from $ARGUMENTS before pattern matching. Value is a comma-separated list from {opus, sonnet, gpt, gemini}. Resolve to models per the Reviewer Selection section. Unknown alias → ask user to pick from the table. Flag is inherited by all downstream routes (Skill/code, Skill/scope, Skill/gestalt).

Apply these rules to remaining $ARGUMENTS in order:

| Pattern | Route | Action | |---|---|---| | Numeric, #N, or GitHub PR URL | PR review | Read and follow operations/pr.md | | 7+ hex chars (commit SHA) | Commit review | Skill(code, review --rev $ARGUMENTS) | | --final <name> | Final scope review | Skill(code, review --final $NAME) | | Matches scopes/*/*/scope.md or scope name | Scope review | Skill(scope, review $SCOPE_NAME) | | gestalt or --structural | Structural review | Skill(gestalt, review $REST) | | --test-audit [path] or path whose first component is test or tests | Test quality audit | Read and follow operations/test-audit.md with $TARGET = path or tests | | File path that exists | Path review | Skill(code, review --path $ARGUMENTS) | | No argument | Menu fallback | See below |

---

Menu Fallback

When no argument or ambiguous, use AskUserQuestion:

Header: Review
Question: What would you like to review?
multiSelect: false
Options:
- PR: Review a GitHub pull request — inline comments and structured summary
- Commit: Review changes in a specific commit
- Branch diff: Review all changes since diverging from base branch
- Uncommitted: Review staged and unstaged modifications
- Structural: Gestalt-driven structural review — topology, blast radius, targeted questions
- Test quality: Audit tests for oracle mirroring, mock tautologies, framework tests, trivial assertions

With "Other" covering: scope review, path review, or custom target.

Routing by selection:

| Selection | Action | |---|---| | PR | Read and follow operations/pr.md — ask for PR # | | Commit | Skill(code, review --rev ...) — ask for SHA first | | Branch diff | Skill(code, review --diff <base>..HEAD) — ask for base branch first | | Uncommitted | Skill(code, review) — auto-detects staged/unstaged | | Structural | Skill(gestalt, review) — ask for base..target range first | | Other: scope | Skill(scope, review) — scope skill asks for name | | Test quality | Read and follow operations/test-audit.md — ask for path (default: tests) |

Protocol: dispatch/protocol.md

---

Reviewer Infrastructure

Canonical configuration for multi-agent review. Domain skills compose on this.

Reference: See reference/models.md for models, reference/harnesses.md for dispatch templates, reference/report.md for YAML schemas, reference/synthesis.md for merge algorithm.

Models & Harnesses

• Claude (native subagent): opus, sonnet — dispatched by the agent via Task.
• External (codex, gemini): defined and dispatched by the peer skill — run peer list for the canonical registry (id ↔ harness ↔ model ↔ alias). Don't restate model strings here.

Full details: reference/models.md, reference/harnesses.md, peer skill

Dispatch Pattern

Cartesian product: roles × harnesses, all in single message. Domain skill defines roles. This skill defines harnesses.

Reviewer Selection

Reviewers can be specified three ways, resolved in this order:

1. --reviewers flag (non-interactive) — comma-separated aliases
2. validation.yaml review_config — persisted per-scope selection
3. AskUserQuestion — interactive fallback when neither is present

--reviewers Flag

Accepts a comma-separated list of short aliases:

| Alias | Harness | Reviewer-id | |---|---|---| | opus | claude | claude-opus | | sonnet | claude | claude-sonnet | | gpt | codex | codex-gpt5.5 | | gemini | gemini | gemini-3.5-flash |

(Model per reviewer-id: peer list.)

Examples:

• /review --reviewers opus,gpt → claude-opus + codex-gpt5.5
• /review --reviewers opus,gpt,gemini → claude-opus + codex-gpt5.5 + gemini-3.5-flash
• /implement execute --reviewers opus,gpt → same two reviewers used for Phase A.5 + Phase C

Invalid alias: Report unknown alias and ask user to pick from the table.

Interactive Fallback (no flag, no review_config)

Question 1: Select reviewers (multiSelect):

• claude-opus (Recommended), claude-sonnet, codex-gpt5.5 (Recommended), gemini-3.5-flash (Recommended)

Default: claude-opus, codex-gpt5.5, gemini-3.5-flash

Question 2: Reasoning effort (if Codex selected): low, medium, high (Recommended)

Full Model Mapping

Reviewer-id ↔ harness ↔ model is the peer skill registry (peer list). claude-opus/claude-sonnet map to the Claude Task harness; codex-gpt5.5/ gemini-3.5-flash to the external harnesses peer dispatches.

Store resolved selections in validation.yaml under review_config (whether from flag, prior config, or interactive prompt).

Report Schema

Base YAML structures for reviewer and synthesized reports. Full details: reference/report.md

Synthesis

Merge, dedup, gate aggregation, severity aggregation. Full details: reference/synthesis.md