Repository navigation
feat(conductor): re-plan one wrong item mid-round on a drift finding - #36618
Conversation
UX Review (Opus 5.5) — ⏭️ skippedRevision |
First Principles Review (Opus 5.5) — ✅ PASSPremise-level review of First-Principles-Verdict: PASS Nothing blocks. This is prompt text only, and its shape is covered by the accepted amendment ("Amendment 2026-11-13") to What this change shipsInventory (4 items) — 4 justifiedIntent: when a conductor's drift read shows a worker's item is itself wrong, the conductor fixes that one item in the same patrol turn instead of letting the worker spend its whole run on it. ADDITION.
[FIRST-PRINCIPLES-REVIEWED] 3f8e4a3 Description read[DESCRIPTION-READ] 3d660cef4e62804cfdb4916ff3eda339e16e2930fd357cebc3492ea8174adb46 That sha256 names the PR description this verdict read. Recompute it with |
Design Review (Opus 5.5) — ✅ PASSDesign-level review of Design-Verdict: PASS Nothing blocks: this is prompt text only and follows the accepted RFC amendment. The narrowed rule keeps whole-round re-plans at the round boundary, and both conductor surfaces change together. Notes
[DESIGN-REVIEWED] 3f8e4a3 Description read[DESCRIPTION-READ] 147584891bc8ea838300b7107cd6198083ba9bf76517dd85813874de1189ffa3 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 |
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: |
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: |
Code change approver recommendation: ApproveWhy: the decider estimates a 12% chance that this change introduces a defect, within the 12% limit for approving on its own, and every check passes. This PR changes prompt text only and carries out the accepted RFC amendment (rfc-conductor-drift-check.md, A.1–A.5). I checked each move it names against the ledger code:
An automated recommendation, not a review: a maintainer still decides. It approves on its own only when the estimated chance of a defect is at most 12% and the change has at most 400 changed lines. Decider v3c (217fe607) · commit |
b367791 to
cd1e424
Compare
|
a0bcf20 to
e12ff22
Compare
e12ff22 to
949f9a8
Compare
949f9a8 to
76409b8
Compare
76409b8 to
cc98961
Compare
cc98961 to
3f8e4a3
Compare
|
|
bolichen97
left a comment
There was a problem hiding this comment.
Approve. I read the whole diff on this head. All checks are green.
This is text only: spec_prompts.py, goal-conductor patrol step 5, and the re-plan paragraph. The rest is regenerated goldens and hashes. There is no new tool, route or storage, no regex, and nothing is refused.
Checked against the two conditions on RFC #36507:
- A bar may only move toward the goal. Both texts allow
action=acceptwith "a corrected bar never weaker than the goal asks". The other path isclose abandonedplus a replacementcreate, so an item's bar cannot be loosened in place. - The owner report stays mandatory. The skill says to "name every closed item and bar change, old and new, in your next report to the owner", and the prompt says to "name each change in your next report".
The scope also stays narrow: one item at a time, only when its own drift lines show it is wrong, and a round-level re-plan still waits for the round to end.
One note, not a blocker: the prompt version drops "old and new". Adding those words there too would keep the two texts saying the same thing.
Direction signed by maintainer (bolichen97) as RFC #36507, with its two conditions.
Problem / Motivation
Goal: When a conductor's drift read shows a work item itself is wrong, it fixes that item in the same patrol turn instead of waiting for the round to end.
The owner asked: "看一看我们是不是做了drfit检查,能够中途换方案" (do we have a drift check that can change the plan midway?). Today the conductor can only steer a worker back toward its item as written, and
goal-conductor/SKILL.mdsays "Re-planning mid-round is not: let the round finish." A worker on an item that is already done on main, mis-scoped, or has a bar that cannot be met keeps spending its run until the round closes.Why it matters
Both conductors (Ultra Mode and the Bot) waste a whole worker run on an item they already know is wrong.
Not a goal
read_message_detail, Patrol step 5's first sentence); that is feat(conductor): check each worker's new transcript for drift on every patrol turn #36302.What changed (motivation → approach → change)
This implements the accepted amendment to
docs/request-for-change/rfc-conductor-drift-check.md("Amendment 2026-11-13", merged in #36507).goal-conductor/SKILL.md, Patrol step 5: one new paragraph. When the lines show the item itself is wrong, the conductor closes itabandonedwith the reason in the close call'sdecisionfield andcreates the replacement, or corrects the bar withaction=accept, never weaker than the goal asks. It rewrites the goal text's Plan and Changed lines (the feat(crew): plan before work: a State check and a five-line first report for workers, a goal-text plan for conductors #36540 rule),session_sends the worker, recordsdecideon the item still open, and names each change in its next report.goal-conductor/SKILL.md, Close the round: the "let the round finish" line now allows one-item re-plans mid-round and keeps round-level re-plans at the boundary.spec_prompts.py, sharedgoal_planrule: one sentence with the same moves, includingdecideonly on the item still open, so the Bot conductor gets it too. Its crewmate golden fixture moves with it.kirocrew-conductor.jsonandkirocrew-ledger-conductor.json, as the RFC's exit criteria say, plus the_CONDUCTOR_SYSTEM_PROMPTshape pin.Backwards compatibility
Compatible: prompt text only; a conductor on the old text keeps working.
Removes nothing: the "let the round finish" rule is narrowed to round-level re-plans, not dropped.
Tests
Pins updated for the changed prompt:
test/test_agent_refactor_spec_bytes.py(two conductor specs in all 8 scenarios),test/test_agent_refactor_contracts.py(prompt shape),test/fixtures/crewmate_prompts/ledger.golden.txt. Locally green: those,test/test_mate_conductor_parity.py,test/test_mate_conductor_agent.py,test/test_crewmate_prompts.py,test/test_skill_budget.pyand the goal-conductor skill tests (400 passed).Manual verification
N/A — prompt text; the parity test and byte pins are the check.
Related Issues
Part of #36299
Checklist
feat|fix|docs|style|refactor|perf|test|chore|ci|build|revert: ...)