From 8c2bface484c486afe554bed79c4200abd39ba2a Mon Sep 17 00:00:00 2001 From: Luke Curley Date: Wed, 28 Oct 2026 13:28:34 -0700 Subject: [PATCH] quest: record the local review trial's findings Co-Authored-By: Claude Opus 5.5 --- quest/m0/local-review.md | 19 +++++++++++++++---- 1 file changed, 15 insertions(+), 4 deletions(-) diff --git a/quest/m0/local-review.md b/quest/m0/local-review.md index fe5ee614..e4d9c162 100644 --- a/quest/m0/local-review.md +++ b/quest/m0/local-review.md @@ -43,14 +43,20 @@ Decided 2026-10-28 (user, /quest-plan): overrides the built-in reviewers' defaults. The Codex plugin's `adversarial-review` prompt is a good starting point for a repository that wants one. Later rounds resume the same session with the new head - and the author's free-text responses (fixes and refutations). The reviewer - accepts a refutation unless it defends the finding with new evidence. - Without them, rejected findings would recur every round. + and the author's free-text responses (fixes and refutations). Without + them, rejected findings would recur every round. A refutation stands only + when it cites evidence (`path:line`, a command's output, or a recorded + maintainer decision) the reviewer cannot rebut. An uncited one is a claim + to verify, and its finding stays: the 2026-10-28 trial below showed a + reviewer dropping a real finding on an unverifiable refutation. - Read-only: Codex in its `read-only` sandbox; Claude restricted to read tools and read-only git commands. Never a write tool or a permission bypass. + The sandbox has no network, so `quest review` fetches the base before it + launches the reviewer. - Cache: always resume. Request Claude's 1h prompt-cache TTL and OpenAI's 24h retention explicitly. The PR shows from usage output that a resumed - round reads the cache. A cold resume is accepted: it costs one rewrite and is + round reads the cache. `codex exec review --json` reports zero usage, so + measure the resumed turns. A cold resume is accepted: it costs one rewrite and is still cheaper than a fresh review. - Instructions come from the base: `REVIEW.md` is optional and read with `git show :REVIEW.md`. For a head whose `AGENTS.md`/`CLAUDE.md` @@ -60,6 +66,11 @@ Decided 2026-10-28 (user, /quest-plan): restoring the base's files in a scratch worktree. If none is clean, ask in the PR. `quest init` does not scaffold `REVIEW.md`. Without one, a round adds no rules to the built-in review. +- Trial, 2026-10-28 (Codex reviewer, by hand): `codex exec review --base` + then `codex exec resume --output-schema` worked, with 68-83% of each + resumed turn's input read from cache. On a plan-only diff both turns found + nothing; on [moq-dev/moq#7126](https://github.com/moq-dev/moq/pull/7126) + they found one real issue. The Claude side is untested. - The authoring model decides whether a change is too trivial to review. A skip is recorded in `## Review` with its reason, and that record satisfies the guide's review rule like a review does.