review

name	review
description	Structured code review using tiered persona subagents, confidence-gated findings, and a merge/dedup pipeline. Use when reviewing code changes before creating a PR, as a build gate, or for standalone review.
argument-hint	[mode:autofix|mode:report-only] [branch name or file paths]

Code Review

Reviews code changes using dynamically selected reviewer personas. Spawns persona subagents that return structured findings, then merges and deduplicates into a unified report.

Agent-Intercom Communication (NON-NEGOTIABLE)

Call ping at session start. If agent-intercom is reachable, broadcast at every step. If unreachable, warn the user that operator visibility is degraded.

When the strict-safety capability pack is installed, also follow .github/instructions/strict-safety.instructions.md: for high-risk diffs, call out the ProposedAction, ActionRisk, approval, and rollback gaps that should be visible before merge or deployment.

Event	Level	Message prefix
Review start	info	[REVIEW] Starting {mode} review of {scope}
Diff analyzed	info	[REVIEW] Analyzed diff: {file_count} files, {line_count} lines changed
Persona routing	info	[REVIEW] Routing: {always_on_count} always-on + {conditional_count} conditional personas
Persona spawned	info	[SPAWN] {persona_name} for code review
Persona returned	info	[RETURN] {persona_name}: {finding_count} findings
Merge complete	info	[REVIEW] Merged: {total} findings ({p0} P0, {p1} P1, {p2} P2, {p3} P3)
Autofix applied	info	[REVIEW] Applied safe_auto fix: {finding_summary}
Review written	success	[REVIEW] Review artifact: {file_path}
Waiting for input	warning	[WAIT] Blocked on user decision
Review complete	success	[REVIEW] Complete: {summary}

Subagent Depth Constraint

This skill spawns reviewer subagents. Those subagents are leaf executors and MUST NOT spawn their own subagents. Maximum depth: review skill → persona subagent (1 hop).

Mode Detection

Check arguments for mode:autofix or mode:report-only. Strip the mode token before interpreting remaining arguments.

Mode	When	Behavior
Interactive (default)	No mode token	Review, present findings, ask for decisions
Autofix	mode:autofix	No user interaction. Apply safe_auto fixes only, write artifact, emit residual work
Report-only	mode:report-only	Read-only. Report findings with no edits, no artifacts, no follow-up item creation

Autofix mode rules

• Skip all user questions
• Apply only safe_auto findings
• Leave gated_auto, manual, and advisory findings unresolved
• Write a review artifact to docs/closure/
• Create backlog follow-up items for unresolved actionable findings
• Never commit, push, or create a PR

Report-only mode rules

• Skip all user questions
• Never edit files
• Return structured findings to caller
• Do not write a review artifact
• Do not create backlog follow-up items
• Safe for the ship agent to invoke during the build loop

Severity Scale

Level	Meaning	Build gate action
P0	Critical breakage, exploitable vulnerability, data corruption	Block commit
P1	High-impact defect in normal usage, breaking contract	Block commit
P2	Moderate issue (edge case, perf, maintainability)	Record as backlog follow-up item
P3	Low-impact, minor improvement	User's discretion

Action Routing

Class	Default owner	Meaning
safe_auto	Review skill (autofix mode)	Deterministic local fix
gated_auto	agent-intercom approval	Fix exists but changes behavior/contracts
manual	Backlog follow-up item	Actionable work requiring human judgment
advisory	Informational	Learnings, rollout notes, residual risk

Routing rules:

• Choose the more conservative route on disagreement between personas
• Only safe_auto findings enter the autofix queue
• requires_verification: true means a fix needs tests or re-review

Reviewer Personas

Always-On (every review)

Persona Subagent	Focus
Constitution Reviewer	Constitutional compliance
Rust Reviewer	Language-specific safety and correctness
Learnings Researcher	Search compound library for related past issues

Conditional (based on changed files)

Use a different model from the caller when available to force genuine diversity of critique. Cross-model is preferred but not blocking.

Persona Subagent	Select when diff touches	Suggested Model
Architecture Strategist	Module boundaries, new abstractions, dependency changes	Different from caller
Concurrency Reviewer	Concurrent/async patterns	Different from caller
Scope Boundary Auditor	Changes spanning multiple domains or exceeding expected scope	Different from caller
Agent-Native Parity Reviewer	MCP SDKs, tool handlers, agent-exposed actions, or user/agent parity-critical flows	Different from caller
Security Reviewer	Auth middleware, public endpoints, input handling, permission checks, secret management	Different from caller

Workflow

Step 1: Determine Review Scope

1. Identify changed files from git diff, explicit file list, or caller-provided scope
2. Categorize each file by type and domain
3. Identify which instruction files apply (via applyTo patterns)
4. Broadcast the diff analysis

Step 2: Route Personas

1. Always-on: spawn Constitution Reviewer, Rust Reviewer, Learnings Researcher
2. Conditional: analyze changed file paths, content patterns, and workspace agent-native signals to select additional personas:
   • Select Security Reviewer (security-reviewer.agent.md) when the diff touches: authentication or authorization code, public endpoint handlers, user input processing, permission or role checks, secret or credential management, or files matching `* src/io_utils.rs — file I/O and path handling

• src/cli.rs — user input parsing
• src/expr.rs — expression evaluation
• src/filter.rs — filter parsing
• Pattern keywords: path, file, open, read, write, execute, command, eval, parse, input, user`

3. Broadcast the routing decision with persona count

Step 3: Spawn Persona Subagents

Spawn all selected personas. Each receives:

• The list of changed files with line ranges
• The diff content relevant to their domain
• Instructions to return structured findings
• Codebase search directive (use grep/glob for context)

Broadcast each spawn.

Step 4: Collect and Merge Findings

As each persona returns:

1. Broadcast the return with finding count
2. Collect all findings
3. Deduplicate: merge findings that identify the same issue
4. Assign final severity (more conservative on disagreement)
5. Assign final action routing

Step 5: Apply Actions (mode-dependent)

Interactive mode:

1. Present findings grouped by severity (P0 first)
2. For each P0/P1, ask the user to accept, modify, or reject the recommendation
3. Apply approved fixes

Autofix mode:

1. Apply all safe_auto findings automatically
2. Create backlog follow-up items for unresolved actionable findings
3. Write review artifact to docs/closure/

Report-only mode:

1. Return structured findings to caller
2. No side effects: no edits, no review artifact, no follow-up items

When the diff changes runtime surfaces, include an explicit recommendation for whether follow-up runtime verification is required and which mode (manual, api, browser) is appropriate.

When the diff includes destructive potential, contract changes, migrations, security-sensitive edits, or other high-blast-radius work, include an explicit recommendation for whether strict-safety action classification or approval follow-up is required before merge or deployment.

When the adversarial-review capability pack is installed and this review surfaces 3 or more P0/P1 findings, recommend escalation to the adversarial-review agent for multi-model consensus validation before blocking the build.

Quality Criteria

• Every changed file is reviewed by at least the always-on personas
• All P0 findings are addressed before the review is marked complete
• P1 findings require explicit acknowledgment (fix or defer with rationale)
• The review report accurately reflects all findings and their resolution status

Model Routing

This skill operates at Tier 2 (Standard) — review coordination and finding assembly.

Generated by autoharness | Template: review/SKILL.md.tmpl