Repository navigation
fix(crew): close the two conductor-mode rule gaps - #37617
Conversation
UX Review (Opus 5.5) — ⏭️ skippedRevision |
Design Review (Opus 5.5) — ✅ PASSDesign-level review of Design-Verdict: PASS Nothing blocks: the tracking sentence now uses the same check ( Notes
[DESIGN-REVIEWED] 619362c Description read[DESCRIPTION-READ] b4ca12d8cc5002e4d908e2988339b589346bfe59a17de2ecc090748a5af101bf 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 |
First Principles Review (Opus 5.5) — ✅ PASSPremise-level review of First-Principles-Verdict: PASS Nothing blocks. The two skill rules close gaps the base parity test already listed with "Remove it then and add the key above". The tracking gate reuses the existing What this change shipsInventory (5 items) — 5 justifiedIntent: the conductor-mode team lead follows the same shared rules as the other conductors, and a Bot turn with no worker dispatch is no longer told to track hand-offs. FIX.
[FIRST-PRINCIPLES-REVIEWED] 619362c Description read[DESCRIPTION-READ] 40386b20747478e2a870630f581bf5954829bd8cb32240b11a0fffc11d1bafa6 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: |
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 an 8% chance that this change introduces a defect, within the 12% limit for approving on its own, and every check passes. Small, well-scoped fix. 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 |
2f48232 to
2f5b8c0
Compare
2f5b8c0 to
619362c
Compare
bolichen97
left a comment
There was a problem hiding this comment.
Approving at 619362c. All 97 checks pass, no red, and no one holds auto-merge.
What I read, on the whole diff:
- Two sentences are added to the goal-conductor skill: dependent calls go in a later step, and a "requested" monitor reply is receipt only, not a cue to retry. Both are prompt text and add no gate, limit or refusal. The two matching parity-test gaps are closed by moving them from the gap list to the anchor list, and the test still fails if an anchor goes missing.
- The tracking sentence now names the ledger tools only when the operating-mode block is also taught.
dispatch_mountedis threaded through the member section builders and the five call sites, and it defaults toNone, which keeps the old behavior for any caller that does not pass it. On main the session start already asks_member_dispatch_heldthe same question for the same block (session_start.py:329), and a known mount answer wins over the configured backend, as that function documents. - This changes only which of two prompt sentences a turn gets. It does not refuse or return less to any request. No regex in product code.
- The new test covers both ways the block is withheld (mounted false with a capable backend, and unknown with an incapable backend) and the held case that keeps the ledger sentence.
- The doc edits match the code. The prompt-registry notes for the two tracking keys name the new condition.
Non-blocking: in the spec edit, the sentence that follows the new one still starts with "There", which now reads as if it points at the withheld case. Moving the new sentence after that one would read better.
One difference to know, not a problem: this builder passes the crew key from the member's config alias, while session start uses crewmate_for_session. They agree for the config names I traced, and the None fallback only matters when no session evidence reaches the build.
Problem / Motivation
Goal: Conductor mode's team lead follows the same two shared rules the other conductors already follow, and a Bot turn that cannot hand work to a worker is no longer told to track its hand-offs.
kirocrew-team-lead(conductor mode) reaches the shared conductor rules only throughgoal-conductor/SKILL.md. Two of them were missing there and listed as known gaps intest/test_mate_conductor_parity.py(CONDUCTOR_MODE_GAPS):dependent_calls(a call that depends on another goes in a later step) andpatrol_requested(arequestedmonitor reply is not a retry cue). The root conductor and the Bot conductor both carry them.docs/system-specs/modules/crew-mode.mdsays "kirocrew-conductor(conductor mode chats)". Conductor mode runskirocrew-team-lead.session_*, or this session withheld the dashboard mount) still gotconductor.tracking_ledger: "Every task you hand to a worker ... record each item ... with work_ledger_record". The turn has no worker to hand anything to.Why it matters
The team lead batches a report with the read that proves it, and retries a
requestedmonitor reply, which the other two conductors are told not to do. A Bot on a backend without dispatch is told to track hand-offs it cannot make.Not a goal
worker.principles, a worker parity test): a sibling PR.ask_only_when_neededallowlist): the v2 guidance PR feat(prompts): one reroute-first ask policy for every crew kind #37609 covers it.What changed (motivation → approach → change)
dependent_callsandpatrol_requestedgoal-conductor/SKILL.mdcarries both rules (top of "The loop", and the Patrolmonitor_startparagraph). The parity test moves both keys fromCONDUCTOR_MODE_GAPStoCONDUCTOR_MODE_SKILL_ANCHORS, so the existing anchor test now holds them.crew-mode.mdnameskirocrew-conductoras the plain and child conductor, says how the parity test holdskirocrew-team-leadthrough the skill, and the test table row says the same.build_member_sectionpicksconductor.tracking_ledgeronly when the ledger is mounted AND_member_dispatch_held(the same verdict that gates the operating-mode block insession_start) holds. It takes the live session'smember_dispatch_mountedasdispatch_mounted; every member-section caller insession_start.py,turn_message.pyand the V2 essentials path passes it through.crewmate_prompts.jsoncomments for the two tracking pieces name the new condition.No RFC:
crew-mode.md"The two conductors stay in step" already requires a conductor rule to land on every conductor or carry a reason, and the tracking change is a bug fix (the text named work the turn cannot do).Backwards compatibility
Compatible:
dispatch_mountedis a new optional keyword defaulting toNone, which falls back to the configured backend's capability, so every existing caller keeps working.Removes nothing: the tracking sentence still renders on every turn it rendered on; only its variant changes when the operating-mode block is withheld.
Tests
test_a_turn_without_the_dispatch_tools_is_not_told_to_track_its_hand_offs(both halves: session reports no mount; no session evidence and the backend cannot dispatch) gets the plain sentence; a session that holds the mount still gets the ledger sentence. Replacing the new_member_dispatch_heldcondition withTruefails both cases.test_goal_conductor_skill_carries_each_mapped_rulenow holds the two new anchors;test_every_conductor_mode_gap_is_still_a_gapruns on an empty gap list.Proof: PROVEN test/test_mate_conductor_parity.py::test_goal_conductor_skill_carries_each_mapped_rule (head 619362c)
Local: related tests (crewmate prompts, conductor parity, member prompt/essentials/turn context, skill and team-lead specs, spec bytes) passed on the rebased head (556 passed after #37164 and #37618); flake8, black, isort, mypy on touched modules, docs_lint clean.
Manual verification
N/A — unit coverage sufficient: the change is prompt text selection and skill wording, both asserted on rendered text.
Related Issues
Refs the Bot / conductor-mode sync audit (items C1, C4, N7).
Sibling sweep
Swept:
git grep -n "binding_has_ledger_tools" -- src/Sibling: src/kiro_crew/context_assembly/session_start.py -- not affected -- already gates the patrol block on _member_dispatch_held
Sibling: src/kiro_crew/agent_materialization/service_agents.py -- not affected -- install-time skill mapping, no per-turn text
Pattern harvest
Rule candidate: review-prompt
Pattern: prompt text that names tools must be gated by the same predicate that mounts them, not a narrower one
Checklist
feat|fix|docs|style|refactor|perf|test|chore|ci|build|revert: ...)Follow-up (not in this diff)
dashboard/handlers/memory.py's V2 essentials validator passes no session evidence and uses the configured-capability fallback; it builds a size check, not a turn.