Repository navigation
docs(rfc): record the worker-debate design and propose claim checks for critics - #33408
Conversation
First Principles Review (Opus 5.5) — ⏭️ skippedRevision |
UX Review (Opus 5.5) — ⏭️ skippedRevision |
Design Review (Opus 5.5) — ✅ PASSDesign-level review of Design-Verdict: PASS Docs-only RFC with no take-away; the corrected [DESIGN-REVIEWED] 5f0fed3 Description read[DESCRIPTION-READ] 44a9f5f021648b61364efbef5c470add8f4a5ad1b8998b3d0587535d5d401d1d That sha256 names everything this verdict read: the PR description, plus the 1 evidence file(s) the workflow collected from it and handed this review. It covers that evidence because the media strip replaces every attachment URL with the same placeholder, so an edit swapping one attachment for another would leave the description's captured text identical while what this review judged changed. Recompute it with |
GPT 6.1 Review — ✅ no blocking findingsGPT 6.1 completed its review of This comment is updated in place on each push. Review detailsNo findings. False positive or not applicable? A repository writer can comment: |
Opus 5.5 Review — ✅ no blocking findingsReviewed Verdict parsed from the review's SHA-scoped output markers for commit False positive or not applicable? A repository writer can comment: |
1ebc6bd to
c3ab8d0
Compare
c3ab8d0 to
589eb2d
Compare
|
RFC to record what already exists |
|
This head (589eb2d) conflicts with main. GitHub reports it as not mergeable. All 97 checks are green. Could you rebase onto main? The conflict is likely in the long |
|
Option B checks the proposer's claims. I'd like it to check the critic's own work too. I run mixed-model review panels often, the kind §3.3 points to
I ended up checking every critic before reading its answer:
Anything that fails gets rerun before synthesis. If it still fails, the synthesis says how many critics were valid, for example "7 of 8 valid". For the debate protocol, a cheap version of this would fit next to Option B. A critic's objection counts only if it quotes text it read during the debate. And the Conductor checks a critic's run before ruling on its objections, the same way the raising critic checks evidence before |
589eb2d to
430d853
Compare
430d853 to
2552279
Compare
2552279 to
3966d8c
Compare
3966d8c to
cd0c123
Compare
|
Thanks for the rebase. cd0c123 merges cleanly with main and all 97 checks are green. This is a comment with a design brief, not an approval, as with every RFC. Design brief:
Doc only; it changes no code. |
KiroCrewCaptain
left a comment
There was a problem hiding this comment.
Approving. This is a docs-only RFC with no code, skill text, config or test changes, and PR Readiness is green on this head. A maintainer signed off the design.
The review checked the RFC's file and line references into worker-debate.md, goal-conductor/SKILL.md, the grill skill, crew-mode.md and pod/README.md at the audited commit, and they hold. Every relative link resolves at this head. The two edits to rfc-worker-messaging.md now match the skill text: hard_rounds replaces max_rounds, and open question 4 is marked answered by #26932. The RFC files are byte-identical to the commit the tests ran on. Done well: the RFC says plainly where its evidence stops (no saved grill transcript, no saved mid-turn model failure) and does not overclaim.
Optional:
- rfc-worker-debate.md:105 gives the anchor budget as "half the goal's item cap", but #33601 removed that cap. A §1.3 drift row, plus a concrete default in #33783 or a note that it is deferred, would keep §1.2 accurate.
- rfc-worker-debate.md:256 says "Three small rules", but Option B lists four (:259, :266, :270, :281).
- Q5 (:532-535) puts a wording change into the Phase 1 PR, but the Migration plan limits Phase 1 to the §2.7 table. Marking Q5 as deferred, or adding it to the Phase 1 scope, would remove the ambiguity.
cd0c123 to
5f0fed3
Compare
|
rfc change, merge first and code will follow |
Problem / Motivation
Goal: The team can read the whole worker-debate design, its history and the improvement the owner adopted in a single RFC, so the skill change in #33783 traces to an accepted decision.
The worker-debate protocol (a Conductor has several workers argue one question in files) is skill text only. It grew through four docs PRs (#26316, #26932, #27969, #29475) and never had an RFC. The one RFC that describes it,
rfc-worker-messaging.md, predates three of them and still namesmax_rounds. The owner also observed that a debate's result is weaker than a "grill me" session's, and that the move from mixed models to one model is recorded only in a PR body.Why it matters
The owner wants to share the design with the team. Without one record, a reader has to rebuild it from skill text and four PR bodies, and the reasons for the model change and the stop rules are easy to lose. Shallow critic objections cost debate rounds and let untested assumptions through to later work.
Not a goal
rfc-worker-messaging.md.What changed (motivation → approach → change)
New
docs/request-for-change/rfc-worker-debate.md(statusaccepted), indexed indocs/request-for-change/README.md. The status records the repository owner's decision of 2026-11-09 in the conductor chat, verbatim: "直接改 skill 实现这个方案" (adopt Option B as amended by the §2.7 debate). It takes effect when this document merges, the same wayrfc-drop-goal-item-cap.mdwas recorded.implementation-prs: [33783].file:lineon mainb94f8cb3ac; a table of where other docs disagree with the skill textgrillskill, Research Lab grill tree, an external grill-me interviewer described in our own words); what the critic role card, Round 0 list and objection table lack; evidence from a real 2026-11-02 debate note (50 objections, 40 closed by adding text, plan grew 2.4x, an unchecked proposer claim aboutpod up --seed, an assumption about the fake backend that failed in use); four options compared; recommends Option B, claim checks: question-form objections on a quoted claim, answered with evidence, aproofcolumn, a named non-author checker,fixed_verifiedonly after a critic confirms the evidence holdsllm-councilpass on the final)focus,overlap,cost; 46 rows; open blocking 2 → 2 → 0; 2 rounds + wrap up; about 80 min; 4 work items; 0 resends;OUTCOME: stable). Keeps roles set up front: no roles and roles after a roleless Round 0 are rejected; the one real miss (2026-11-02 fake backend) was a depth miss inside a role. Adds a Round 0 biggest-risk entry with a row, one Conductor side-by-side read in the existing round-1 relay, and a same-as rule. Amends Option B (§2.4): a named non-author checker for everyproofcell and adopted recommendation, asking "what if it is wrong on its path", in the existing critique step;Q:marks question-form rows. Skill change about 34 lines, 0 new relays, rounds or work items. Known limits listedrfc-worker-messaging.md: two stale lines fixed in place, since the code wins and the doc is the bug: trigger T-b now nameshard_rounds, and open question 4 (round check-ins) is marked answered by #26932 with a pointer to the new RFC.Backwards compatibility
Compatible: docs only; no tool, schema, skill text or stored record changes.
Removes nothing: the two edited lines in
rfc-worker-messaging.mdcorrect names to match the current skill text; no path, key or capability is taken away.Tests
Proof: not provable -- docs or config only
Ran locally on head
5f0fed3deb(docs-lint; the 238 tests ran on2552279648, same RFC files):scripts/docs-lint.sh(pass),test/test_pr_quality_gates.pyandtest/test_pipeline_conductor_skill_contract.py(238 passed).Manual verification
N/A — docs only. Every
file:linein the RFC was re-read on mainb94f8cb3ac(thegoal-conductor/SKILL.mdandcrew-mode.mdline numbers moved and were updated).Related Issues
Closes #33405
Companion skill PR (Phase 1, implements §2.7 in
goal-conductor/references/worker-debate.md): #33783. Its First Principles lane traces to this document, so merge this PR first, then rebase #33783.Checklist
Contribution License Agreement
Contributed under the project's Apache-2.0 license.