Repository navigation
docs(goal-conductor): worker debate Round 0 biggest-risk rows and owned claim checks - #33783
Conversation
UX Review (Opus 5.5) — ⏭️ skippedRevision |
Design Review (Opus 5.5) — ✅ PASSDesign-level review of Design-Verdict: PASS Nothing blocks: this is skill-text-only and puts into practice an accepted RFC; no test or code reads the wording it replaces. Notes
[DESIGN-REVIEWED] d5c4df9 Description read[DESCRIPTION-READ] 6ff30dcc024a6a093ceb572e70709c2bca657d8af59653817d6268a245d62ed7 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: |
First Principles Review (Opus 5.5) — ✅ PASSPremise-level review of First-Principles-Verdict: PASS Every item puts into the skill text a rule that Phase 1 of the base-accepted RFC Notes
What this change shipsInventory (7 items) — 7 justifiedIntent: make worker-debate critics raise the key risk even when it falls outside their role, and give every proposer claim a named checker who did not write it. ADDITION to the skill protocol.
[FIRST-PRINCIPLES-REVIEWED] d5c4df9 Description read[DESCRIPTION-READ] ae034ef8fb3372f8f513103a8f19865a1d6fe2715a66e267d52a1abcfd0d00b4 That sha256 names the PR description this verdict read. Recompute it with |
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: |
|
First Principles disposition on head 1db76e0, Notes item 4 (note-copy check moved from every critic to one named critic): kept, no change. The defect it fixes was found in the critic-roles debate as overlap-0-4: the protocol gave some duties to every critic at once (the smallest-thing entry and the note-copy check), so each critic repeated the same check and none owned it. Naming one critic on its card gives the check a single owner; the check itself and its blocking severity are unchanged. The debate closed overlap-0-4 as fixed in round 1. Status of the other reds on this head: the BLOCK itself (product-shape change, RFC #33408 still open) is held until #33408 merges as accepted; this PR is then rebased once. Backend 3.12 shard 3 and Windows shard 3 fail on test_slim_skills_do_not_regrow_duplicate_guidance_or_wire_model_ids[kirocrew-worktree-dev-16000], a main break (that SKILL.md is 16004 bytes on main) fixed by #33795. Build Desktop failed on HTTP 504 downloads. Coverage Gate and PR Readiness follow from those. |
1db76e0 to
17bf3b9
Compare
17bf3b9 to
d5c4df9
Compare
KiroCrewCaptain
left a comment
There was a problem hiding this comment.
Approving. This head changes one file, src/kiro_crew/builtin_skills/goal-conductor/references/worker-debate.md (+43/-6, prose only). The review read it in full at d5c4df9. Each of the 7 hunks maps to one row of the change table in the accepted RFC (rfc-worker-debate.md §2.7). No test pins the changed wording. No blocking repo rule matches this path, and no recorded decision covers this protocol. PR Readiness is green on this head. A maintainer signed off the design through the accepted RFC #33408.
Done well: the new checker rule closes the gap the RFC debate found, where a critic could confirm its own adopted recommendation. It reuses the existing objection ids, relay step and critique step, and adds no new rounds, relays or work items.
Optional:
- Line 157: "gives your recommended answer" reads as the Conductor's answer, because "you" means the Conductor everywhere else in this file. The RFC and line 200 both say the critic gives it. "gives its recommended answer" would fix it.
- Line 207: the
A-M3id relies on mechanism row numbers, but the mechanism table never defines a stableM<n>row id. If rows are numbered by position, removing a row can repoint an open proof assignment. One clause saying row ids are never renumbered or reused would fix it. - Line 131: the new
same as <id>ruling is missing from the Ruling table at lines 190 to 198.
Problem / Motivation
Goal: Worker-debate critics raise the key problem even when it sits outside their role, and every claim the debate relies on has a named checker.
In a worker debate, critic roles are fixed up front. The owner worried this can miss the main point (outside every role) and that dropping roles makes critics overlap. On the 2026-11-02 debate, a role's Round 0 risks got no table rows, and a proposer's claim that "an existing part already does it" was never checked by anyone, so a wrong claim passed review and failed live.
Why it matters
A debate that misses its key risk, or lets a false claim through, produces a confident wrong answer. The protocol is only as good as what it forces critics to check.
Not a goal
docs/request-for-change/rfc-worker-debate.md); it landed separately.What changed (motivation → approach → change)
Design record: the worker debate
critic-roles(outcomestable, 2 rounds + wrap up, three critics: focus, overlap, cost). Its note and final proposal live in the Kiro Crew research directory (research/debate-critic-roles/note.md,final-A.md). RFC #33408 (worker debate claim checks,docs/request-for-change/rfc-worker-debate.md) is merged with statusaccepted(merge commit 76855bb, 2026-11-09); this PR implements its Option B as amended by that debate.One file,
src/kiro_crew/builtin_skills/goal-conductor/references/worker-debate.md:same as <id>; the proposer answers once and both critics verify.Q:from a question bank, answered with evidence only;proofcells on "existing part does it" mechanism rows; in each step-b relay one line per checker naming the ids/rows it checks, never by their author; the checker asks "what if it is wrong on the path the proposal uses it for" and lists aproof checked:line; an unchecked assignment stays open under its original id (aproofcell's id is proposal letter + row, e.g.A-M3) and is reassigned; one-critic fallback.The
A-M3id prefix is a wording fix the overlap critic asked for in its final file (two proposals could both have anM3).Cost: +43/−6 lines of skill text, 0 new rounds, 0 new relays, 0 new work items.
Backwards compatibility
Compatible: skill text only; no code, config or stored format changes.
Removes nothing: the only narrowed duty is the note-copy check, moved from every critic to one named critic; the check itself stays.
Tests
Proof: not provable -- docs or config only
Ran locally:
scripts/docs-lint.sh(pass),scripts/check_builtin_skill_scope.py(pass),test/test_conductor_skills_steer_corrections.py,test/test_goal_conductor_skill_grants.py,test/test_builtin_agent_skill_resources.py(44 passed).Manual verification
N/A — skill text; the design was verified by the three debate critics in their final files.
Related Issues
Refs #33408
Checklist