Repository navigation
fix(goal-conductor): park findings that no Done when condition needs - #37648
Conversation
UX Review (Opus 5.5) — ⏭️ skippedRevision |
First Principles Review (Opus 5.5) — ✅ PASSPremise-level review of First-Principles-Verdict: PASS Nothing blocks. The fix adds one missing rule, in the prose that let every finding become an item, which is where the loop came from. Its siblings in the charter, the team-lead skill and the RFC all change in step. Notes
What this change shipsInventory (6 items) — 5 justifiedIntent: stop a goal Bot from turning every finding into a new work item, so its goal can finish. This is a FIX.
[FIRST-PRINCIPLES-REVIEWED] 38a9cbc Description read[DESCRIPTION-READ] 1c33b79d364b42edf43d956e8c264f63735ba70eeccccc3a4fd6da816f916d5e 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, the new part sits between Plan and Changed, and both Board parsers ( Notes
[DESIGN-REVIEWED] 38a9cbc Description read[DESCRIPTION-READ] fe44ea87538d9c5fc32f874c80bed5c12cb906489a610d1e7a245437406f2ffc 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: |
c1eb7d2 to
38a9cbc
Compare
|
span=1c12c52038a7
|
Code change approver recommendation: ApproveWhy: the decider estimates a 10% chance that this change introduces a defect, within the 12% limit for approving on its own, and every check passes. Prompt- and skill-only change. It adds a rule that sends findings no "Done when" condition needs to a 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 |
|
This adds a new concept, so it needs a design sign-off from a maintainer before it can merge. A maintainer will post the decision here. The review found nothing that blocks this head. Only prompt and skill text changes: the goal-conductor and team-lead skills, the Optional:
This PR and #37652 re-pin the same conductor prompt and spec digests. Whichever merges second needs a rebase and a fresh re-pin. |
bolichen97
left a comment
There was a problem hiding this comment.
Approving 38a9cbc8df.
- The rule is consistent across all four places:
spec_prompts.py(goal text now five parts, Changed last), the goal-conductor Patrol and plan-reviewer steps, the team-lead skill, and the RFC. A finding becomes an item only when a Done when condition needs it; everything else goes onParked:until the owner moves it in. - Keeping Changed last matters because the Board reads it to the end of the text; the prompt and the skill both say so.
- Small note, not blocking: Parked shares the 2000-character goal budget, so a long Parked line could crowd out Plan. One short entry each, as written, should keep that in check.
- Ran
test_goal_conductor_parks_findings.py,test_agent_refactor_contracts.pyandtest_agent_refactor_spec_bytes.pylocally: 80 passed. CI is green on this head.
AI-assisted, sent by Bolin Review bot
Part of #37361
Problem / Motivation
Goal: A goal-conductor Bot parks a finding that its goal's "Done when" does not need, instead of turning it into a new work item, so the work can end.
In #37361 a Bot with one goal ran for six days and never finished. A review found something, the finding became a work item, the item became a merge request, and that merge request got a review that found more. Open issues went from 46 to 108 while the goal still read "beta fixed".
Why it matters
Every owner who gives a Bot a goal can hit this loop: the plan reviewer runs every round, and the skill told the conductor that each fix the reviewer finds "becomes a new item if you take it". Nothing in the skill said when not to take it, so the work grows faster than it closes.
Not a goal
accept_eval) or to the reviewer itself.What changed (motivation → approach → change)
The goal-conductor skill now tells the conductor to park any finding (from the plan reviewer, a worker's report or its own patrol) that no "Done when" condition of the goal needs. A parked finding goes on a new
Parked:line of the goal text, with where it came from, and does not become an item, a seed or adecide. Each round report lists the findings parked that round. Only the owner moves a parked finding in: the conductor then adds a "Done when" condition for it and creates its item.The goal text keeps
Changedas its last part, because the Board's Latest change line readsChangedto the end of the goal text. The conductor's charter (_LEDGER_CONDUCTOR_RULES["goal_plan"]insrc/kiro_crew/agent_parts/spec_prompts.py, which the skill points to for the goal text) and the team-lead skill now list the same five parts, so the prompt and the skill agree. The plan-review RFC's "What stays" line now says the same rule.Backwards compatibility
Compatible: this changes skill and prompt instructions only; goal texts without a
Parked:line still read the same on the Board.Removes nothing: the conductor still creates items for findings a "Done when" condition needs; the change only adds a rule for the others.
Tests
test/test_goal_conductor_parks_findings.pypins the wording in three sections of the skill: the goal text has aParkedpart beforeChanged(Round 0), the park rule and the reviewer line (Patrol), and the parked list in the round report (Close the round). A fourth test pins that the charter lists Parked before Changed and that the team-lead skill names the same five parts. Reverting any one changed hunk of the two skills or the charter on its own turns one test red.The conductor prompt text changed, so
test/fixtures/crewmate_prompts/ledger.golden.txt, the_CONDUCTOR_SYSTEM_PROMPTshape intest/test_agent_refactor_contracts.pyand thekirocrew-conductor.json/kirocrew-ledger-conductor.jsondigests intest/test_agent_refactor_spec_bytes.pyare re-pinned. Only those two spec files' digests moved; every other spec, the audit events and the state sidecar match as before.Proof: PROVEN test/test_agent_refactor_contracts.py::test_every_pre_split_name_resolves_with_its_pre_split_shape, test/test_agent_refactor_spec_bytes.py::test_every_written_spec_matches_the_pre_split_bytes[clean_over_customized], test/test_agent_refactor_spec_bytes.py::test_every_written_spec_matches_the_pre_split_bytes[customized], test/test_agent_refactor_spec_bytes.py::test_every_written_spec_matches_the_pre_split_bytes[forks], test/test_agent_refactor_spec_bytes.py::test_every_written_spec_matches_the_pre_split_bytes[fresh] (head 38a9cbc)
Determinism: the file passed 3 repeated runs and a run in reversed test order. It reads one shipped file and has no shared state.
Manual verification
N/A — the change is skill and prompt text; the tests pin the wording, and the 2767 tests that read the conductor skills and prompts pass with it.
Related Issues
Part of #37361
Sibling sweep
Swept: git grep -n -i "becomes a new item|new item if|finding.*new item" -- src/kiro_crew docs
Sibling: docs/request-for-change/rfc-conductor-plan-review-hourly-check.md:137 -- changed here
Sibling: src/kiro_crew/builtin_skills/goal-conductor/SKILL.md:678 -- changed here
Swept: git grep -n -i -E "four parts|Changed: a time and why|Done when, Plan, Changed" -- src test
Sibling: src/kiro_crew/agent_parts/spec_prompts.py:140 -- changed here
Sibling: src/kiro_crew/builtin_skills/team-lead/SKILL.md:43 -- changed here
Sibling: test/fixtures/crewmate_prompts/ledger.golden.txt:127 -- changed here
Pattern harvest
Not generalizable: the defect is one missing rule in one skill's prose about when a finding becomes work, not a code pattern a rule can catch.
Checklist
feat|fix|docs|style|refactor|perf|test|chore|ci|build|revert: ...)