code-review

name	code-review
description	Review Devolutions Gateway pull requests and diffs using focused contract, compression, documentation, prose, and skeptical passes. Use whenever Copilot code review is requested or a proposed change needs review.

Code review

Inspect the diff, its stated goal, and repository guidance before selecting conditional reviewers:

• Run external-contract-reviewer only when the change may affect HTTP or OpenAPI behavior, WebSocket behavior, relay or tunneling semantics, configuration schemas, or agent and platform protocols.
• Run code-compressor for every diff, limiting it to changed code.
• Run docs-compressor when documentation, rustdoc, explanatory comments, or instructional prose changes.
• Run prose-verifier when hand-maintained prose changes in a format where source line breaks are stylistic.
• Run skeptical-reviewer last for every diff.

Prefer one rubber-duck sub-agent per applicable skill. Run the contract, compression, documentation, and prose reviews in parallel. Give every agent the change goal, review scope, base and head references, relevant repository guidance, and an instruction to invoke its named skill, inspect the diff itself, make no edits, and return only evidence-backed findings. If sub-agents or skill invocation are unavailable, apply the same skills directly and keep the ordering.

Before the skeptical pass, normalize the initial results. Drop claims that lack a concrete location or mechanism, merge findings with the same root cause, and retain disagreements instead of resolving them by vote. Give the skeptical reviewer the diff plus the retained findings and ask it to independently review the change, verify or reject each supplied concern, and identify material issues the focused passes missed.

Evaluate the final evidence yourself. Report correctness and external-contract findings first, ordered by severity, then optional compression suggestions. Each retained item needs a location, concrete impact, and actionable correction or simplification. Merge duplicates across reviewers and do not inflate a maintainability preference into a defect. Briefly name skipped conditional reviewers and why; if nothing material remains, say that no findings were identified.