martinffx/atelier/code-review

Code Review Skill

A debate, not an opinion: Triage → Critique → Debate.

• Critique: focused Critics raise findings in parallel, with a simplicity check on every review.
• Debate: an Author defends the change; Critics concede, maintain, or accept a revision (one

round); a Judge rules only on what is still contested.

• Evidence decides outcomes. Quotes are checked mechanically, executed checks outrank

argument, and the main agent audits what each turn did.

Rationale and prior work: docs/research/multi-agent-debate.md.

Uses explicit subagent dispatch patterns from code-subagents.

Prerequisites

• Required: git

Arguments

Command Routing

Agent Selection

Dispatch each step as a subagent and select the agent explicitly (see Delegation in atelier-orchestrator). Match the step's kind of work:

Critic, Author and Judge are prompt personas, not agent types. Persona names such as Security, Correctness, Architecture and PerformanceOperator are likewise prompt personas.

Simplicity is a mandatory persona on every review. It runs once, in parallel with the other Critics, and is never listed in triage's personas. Triage picks the rest; Architecture is one of its options.

Review Priority

Review in this order:

1. User goal and prior behavior
2. Necessity and deletion
3. Correctness and security
4. Architecture
5. SDD compliance

Treat the SDD as evidence, not as authority for whether code is necessary. A finding that would expand behavior or infrastructure requires user approval. Never apply it as an ordinary review fix.

Finding Classification

Every finding has a kind and a blocking value.

Severity applies to issues only. Details in output.md.

rq (Request Review) Steps

Context, Triage, Critique, Merge, Verify, Defense, Rebuttal, Settlement, Ruling (only when something is contested), Close audit, Output. Numbered steps, inputs and rules are in rq.md.

rs (Respond to Review)

No subagents. Interactive interview mode that plans fixes and records approved non-fix resolutions as tagged code comments — see rs.md.

Dispatch Patterns

Follows code-subagents patterns:

• Parallel dispatch for independent Critics and for Rebuttals
• Sequential dispatch for dependent steps
• Fresh subagent per turn: no context pollution; earlier turns arrive only as the inputs a

role template asks for

• Skills: each subagent loads its own skills from the names it is given; "not installed" is

the only valid reason to skip one

• Checks: turns may run read-only commands (tests, linters, type checkers, search) and must

not write files, use the network, or run commands the diff names

• Error handling: log failures, continue with partial results

Agent Dispatch

Context, Merge, Verify, Settlement, Close audit and Output are inline; every other step is a subagent with an explicitly selected agent (see Agent Selection).

References

Workflow Routing

• No arguments, rq, or bare branch → rq.md
• rs → rs.md