Skip to content

Latest commit

 

History

History
62 lines (47 loc) · 3.2 KB

File metadata and controls

62 lines (47 loc) · 3.2 KB

Codex review process and instruction delivery

Status: accepted, September 4, 2026.

Decision

Adapt Codex's detached review-agent inspection process into Diffwarden's fresh, ephemeral, schema-constrained app-server turn. Deliver the stable review contract as developer instructions and the resolved target, focus, provenance, and patch as user input. Preserve Codex's model-specific base instructions.

This is an adaptation of the detached review process, not a native detached request. Existing native-inline review remains an explicit option. No extra model call or prose translation stage is introduced.

Evidence

The original rubric port used Codex 462deb0426bf on May 28. This refresh inspected stable rust-v0.153.4, including:

  • codex-rs/skills/src/assets/samples/review-agent/SKILL.md, introduced July 14 by 83a4187837: whole-diff inspection, surrounding code, tests and callers, and no recursive delegation.
  • codex-rs/prompts/templates/review/rubric.md, updated July 21 by 81de4f251c: applicable repository rules and verified attribution in findings.
  • codex-rs/app-server/src/request_processors/turn_processor.rs: detached review clones the server configuration and invokes an agent with a history fork. It does not reliably inherit the parent thread's requested model, instructions, and restrictions.
  • codex-rs/ext/agent/src/lib.rs: the detached invocation does not supply a final output schema. The bundled skill requests prose.

Schema generation using the published 0.153.4 binary confirmed that review/start accepts only threadId, target, and delivery. turn/start exposes outputSchema, and thread/start exposes developerInstructions and ephemeral.

A separate read-only source review confirmed these blockers before implementation. An ephemeral parent's suitability for the native persisted-history fork was not proven; the chosen integration does not depend on that behavior.

Product constraints

The supplied patch remains authoritative. Project rules cannot override read-only behavior, review scope, or the output schema. Reviews inspect test source but do not execute tests, builds, health checks, services, or other agents. They complete without clarification questions and omit unsubstantiated findings.

Completed review output takes precedence over streamed progress. Async questions and commentary do not replace the final answer. Cancellation interrupts the review's own thread before disconnecting, including when using a shared daemon. Monitoring failure explanations are preserved without automatically continuing the turn.

Tradeoff and validation

The adaptation retains one-call structured output, bounded scope, and existing configuration. It does not automatically load future revisions of the upstream skill. Review-source revision, tested runtime, and selected model are separate maintenance facts; an installed CLI version is not the rubric's version.

Credential-free tests cover instruction delivery, output schema, async/final ordering, shared-server cancellation, and failure explanations. Source compatibility and these tests do not establish equivalent review quality to native detached review. Live model acceptance remains opt-in and must be reported separately.