The three agent roles, short: test then review, strict on code, prose only removed - #474
Merged
Merged
Conversation
… removal A proposal for the owner. The reviewer's first act is now to establish that the head it was given is tested: CI green, the branch's own tests green, and the hardware reading where the change targets hardware. Otherwise it stops with the new verdict NOT READY FOR REVIEW. Scout first: a branch that built before it measured is sent back to measure. Prose stops being reviewed. The PROSE section and the citation policing in SOURCES are deleted; a brittle reference is never a send-back reason and is never corrected, and prose that is false, will rot or misleads is flagged REMOVE for the implementer to delete. The CTO's recommendations, for the owner to accept or strike, are section 3's three paragraphs: the BLOCKER/NOTE ranking, convergence on the delta in later rounds, and full re-running plus open-ended mutation hunting scoped to high-risk code. FIT, GROWTH, EDGE CASES, the negative-control and oracle rules and the output format stay. The every-branch list is folded into the sections its lines belong to. wc -w: 1559 before, 1478 after. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014iqcj4jDKpaiDX8B7CMvmK
…removed The owner transferred the agent roles to the orchestrator on 2026-09-21 with one preference: keep them short. The reviewer prompt drops from 1559 words to 720, and two new files carry what every brief used to repeat. reviewer.md: untested code is not reviewed (NOT READY FOR REVIEW); findings are BLOCKER, NOTE or REMOVE and only a BLOCKER sends a branch back; a later round judges the earlier BLOCKERs and what changed since; mutation hunting without limit stays on high-risk code; prose and brittle references are never corrected, only removed when they make trouble; the report is posted on the pull request. The one-rule-per-incident list is gone. implementer.md: the brief is a fence, measure before building, test before review, the pull request body is the handoff, and the standing git and harness rules every brief used to paste. orchestrator.md: one goal, at most three branches in rounds, scout then build then test then review, the judge judges, nothing but a defect turns main red. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014iqcj4jDKpaiDX8B7CMvmK
Japabu
changed the title
Japabu
marked this pull request as ready for review
September 21, 2026 07:29
Japabu
enabled auto-merge
September 21, 2026 07:29
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
The owner transferred ownership of the agent roles to the orchestrator on 2026-09-21, with one preference: keep them short. This replaces the long review prompt with a short one and adds the two roles every brief used to restate.
.claude/agents/reviewer.md, 1559 words to 720. Untested code is not reviewed: a missing green CI run, missing tests or a missing hardware reading ends the review with NOT READY FOR REVIEW. Findings are BLOCKER, NOTE or REMOVE, and only a BLOCKER sends a branch back. A later round judges the earlier BLOCKERs and what changed since. Unlimited mutation hunting stays on high-risk code. Prose and brittle references are never corrected and never cost a round; they are removed when they make trouble. The report is posted as a comment on the pull request, so a wiped scratchpad loses nothing. The list of one rule per past incident is gone..claude/agents/implementer.md, new, 507 words. The brief is a fence; measure before building on a guess; test before review; the pull request body is the handoff; the standing git and harness rules, stated once..claude/agents/orchestrator.md, new, 440 words. One goal and at most three branches in rounds; scout, build, test, review; the judge judges; nothing but a defect turnsmainred; the bench ritual.Nothing in
src/,tests/or.github/reads these files (git grepfor the three paths is empty), so no gate moves. Supersedes #429.🤖 Generated with Claude Code
https://claude.ai/code/session_014iqcj4jDKpaiDX8B7CMvmK