Repository navigation
docs(goal-conductor): size dispatch waves by resource_status, not two or three per round - #36298
Conversation
8410a87 to
b4e5f2d
Compare
UX Review (Opus 5.5) — ⏭️ skippedRevision |
Design Review (Opus 5.5) — ✅ PASSDesign-level review of Design-Verdict: PASS Prompt-only change that swaps a fixed brake for the existing Notes
[DESIGN-REVIEWED] 3c1f41e Description read[DESCRIPTION-READ] 5cf2e3745948afdd4dda31a74eff8627d659e5f37a8ff9eaec9967cdbd55401d 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 base RFC Notes
What this change shipsInventory (7 items) — 6 justifiedIntent: let a goal conductor run independent items in parallel when the host has room, instead of a fixed two or three per round. This is a FIX.
[FIRST-PRINCIPLES-REVIEWED] 3c1f41e Description read[DESCRIPTION-READ] 964c92e9287ae0ad623bd4b9b53c5712e0eea401e443a63717534b41beaca0b0 That sha256 names the PR description this verdict read. 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: |
Code change approver recommendation: A person should decideWhy: the decider thinks the change is probably fine but estimates a 20% chance that this change introduces a defect, above the 12% limit for approving on its own. Prompt-only change. It replaces goal-conductor's fixed "two or three items per round" limit with a rule that sizes each wave from 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 |
b4e5f2d to
55d5142
Compare
|
55d5142 to
b6363a3
Compare
b6363a3 to
d99573f
Compare
|
d99573f to
606df60
Compare
KiroCrewCaptain
left a comment
There was a problem hiding this comment.
Thanks for dropping the host_pressure dependency. The new sentence now names only things that exist on main. One case from the earlier review is still open on this head.
src/kiro_crew/builtin_skills/goal-conductor/SKILL.md:196: whenresource_statusreadscriticalat the first wave, the rule at :191-193 dispatches nothing, so nosession_createorbindruns. On main, only those two paths callconductor_patrol.ensure_patrol(dashboard/handlers/session_control.py:554anddashboard/handlers/work_ledger_conductor.py:480). The Patrol section (:324-330) tells the conductor to armmonitor_startonly after it reads a create or bind reply, and a wave that sent nothing gets no reply. Line 196 then says "Keep the patrol loop armed while anything is queued; the patrol tick resumes you", but no loop exists. Outcome: the conductor says the queue will resume, and nothing resumes it. The goal stays at zero items until a person messages it. Smallest fix: add one sentence, for example "If a wave dispatched nothing and no patrol is armed yet, armmonitor_start(watch=\"work-ledger\")yourself." Or, oncriticalwith nothing in flight, say in the round report that dispatch is paused until the user resumes it.
Earlier items:
- Item 1,
host_pressurewake: fixed. Line 196 no longer names that wake. - Item 1, first-wave case: still stands (item 1 above).
- Optional items from the first review (the reset bullet in "What changed", and "dispatch one" on
tight/unknown): answered.
Optional:
- The "and nothing is queued" stop clause (:884) still has no named owner.
spec_prompts.pypatrol_endsandpatrol_final_report, theconductor_patrol.pytick texts andsrc/kiro_crew/docs/work-ledger.md:318still say to stop when every ledger item is terminal. Please name an issue that will sync them. team-lead/SKILL.md:153says to queue the rest of the wave ontight/unknown, and the same sentence hands the count to goal-conductor, which says to dispatch one. #36312 adds a third count. Please give the per-posture count one owner.- The checklist ticks "new tests added", but the diff adds no test. Either untick it or add a pin in
test/test_conductor_team_management_rules.py.
298a011 to
fa1cf88
Compare
|
@KiroCrewCaptain thanks, all addressed in fa1cf88:
Optional:
Could you re-review? |
… or three per round
fa1cf88 to
3c1f41e
Compare
|
|
Thanks for closing the first-wave gap and for replacing team-lead's second copy of the capacity rule with a pointer. Nothing in the diff blocks a merge, but this head cannot be approved yet. Earlier change request, item by item:
What still blocks: PR Readiness is red on this head because the CodeQL Optional:
|
Withdrawn: every item in this review is fixed or answered on the current head (see the follow-up comment). Other holds on this PR are listed there.
buluoray
left a comment
There was a problem hiding this comment.
🤖 Kiro Crew Auto-Pipeline [operator: buluoray]:
This PR is still needed, and I found no defect in it.
Current main still says "Keep concurrency small and constant — two or three items per round" in goal-conductor/SKILL.md. That line contradicts the rest of main. The shared conductor rule resource_wave in spec_prompts.py already says: on ample dispatch every ready item, on tight or unknown dispatch one and queue the rest, on critical dispatch none. The concurrency-cap RFC §7 (accepted 2026-11-13) also says resource_status decides each wave. This PR swaps the old line for the same counts that resource_wave uses, so the skill and the prompt now agree.
Earlier review points, checked on this head:
host_pressurewake named before it existed on main: fixed. The sentence now names only the patrol tick.- First wave on
critical, with nothing armed to resume the conductor: fixed. The rule now says to armmonitor_start(watch="work-ledger")when a wave dispatched nothing. - Who owns the "and nothing is queued" stop clause: answered with #37192.
- One owner for the per-posture count: fixed. team-lead §5 now points to goal-conductor's "Capacity decides the wave" and no longer keeps its own copy.
- Test checklist box: fixed, it is unticked.
- CodeQL red on PR Readiness: cleared. All checks on this head now pass (80 pass, 18 skipped, 0 failing).
Merge with current main (66dc266de3): clean, no conflicts. The merged skill reads as one rule. The only other wave text is main's host-pressure section, which says a present host_pressure level wins and resource_status sizes waves again once it clears. That order is stated, so the two do not disagree. The "Rounds are not gated, but spend is bounded" section already says "resource_status decides the wave", which matches.
Tests on the merged tree:
test/test_conductor_agent.pyplus the parity, skill-grant, skill-resource, frontmatter, line-cap, team-lead and ledger-conductor tests: 624 passed.- All
test_conductor_*,test_builtin_skill*andtest_skill*files: 3158 passed, 881 skipped, 1 failed. The failure wastest_skill_search_index.py::TestTransientLockDoesNotLatch::test_a_busy_database_is_retried, a lock-retry test. It passed when run alone, and it does not read either skill file. test_back_to_back_rounds_keep_a_spend_bound: passed.
Optional, not blocking:
- Round 0 still says "name the concurrency" and lists "how many items per round" as an assumption to pick. Both pull against "Carry no number of your own". Saying that
resource_statussizes each wave there would keep one rule. rfc-work-ledger-concurrency-cap.md§7.1 quotes "Carry no number of your own" as living in team-lead §5. After this PR that phrase lives in goal-conductor. The RFC pins its reads to an older commit, so this is only a stale pointer.- No test pins the new paragraph. One anchor phrase in
CONDUCTOR_MODE_SKILL_ANCHORS["resource_wave"]intest/test_mate_conductor_parity.pywould catch a later edit that drops it or breaks team-lead's pointer.
Verdict: APPROVE
Problem / Motivation
Goal: A goal conductor dispatches every ready, independent item at once when the host has room, instead of one or two at a time.
The goal-conductor skill says "Keep concurrency small and constant — two or three items per round. More rounds beats more parallelism". Conductors follow it literally and dispatch one or two workers per round even when
resource_statusreadsample, so goals with many independent items run in many slow rounds.Why it matters
Every goal run through a conductor (Ultra Mode, mate conductor) is slowed by a fixed brake that ignores real host capacity. The owner noticed conductors are "especially careful, one or two at a time" and asked that the conductor dispatch freely, checking resources first.
Not a goal
DEFAULT_GOAL_CONCURRENCY_CAP, "at most 6 open") or its prose. That stays the hard ceiling.What changed (motivation → approach → change)
Fixed brake in prose → reuse the rule the team-lead skill already ships (section 5, "Capacity decides the wave, not your plan") →
src/kiro_crew/builtin_skills/goal-conductor/SKILL.mdreplaces the "two or three per round" paragraph with that rule: dispatch every ready, independent item up to the ledger cap; callresource_statusbefore every wave and reseed wave;ampledispatches the wave,tight/unknowndispatch one item and queue the rest,criticaldispatches none; each landing re-readsresource_statusand dispatches from the queue; the patrol loop stays armed while anything is queued (the patrol tick resumes the conductor; a wave that dispatched nothing arms its ownmonitor_start(watch="work-ledger"), since no create or bind reply arms one), and patrol ends only when every item is terminal and nothing is queued; the chat is never reset while anything is queued, since a queued item has no ledger row and the chat is its only copy; carry no number of your own.team-lead/SKILL.md§5 drops its own copy of the capacity rule and keeps only a pointer to goal-conductor's "Capacity decides the wave" plus the "give the capacity back" step, so the team-lead agent (which loads both skills) reads one rule, not two that disagree. Behaviour change for a team lead: ontightorunknownit now dispatches one item and queues the rest, where team-lead §5 used to queue the whole wave.Swept for other copies of the brake (
two or three,More rounds,small and constant) acrosssrc/,docs/andtest/: this was the only one.agent_parts/spec_prompts.pyalready says "Capacity, before dispatching —resource_status".Backwards compatibility
Compatible: prompt text only; the ledger cap still refuses past its limit.
Removes nothing: the paragraph is replaced by a capacity rule; no key, path or function changes.
Tests
Proof: not provable -- docs or config only
Ran the covering skill/conductor tests locally (
test_builtin_agent_skill_resources,test_builtin_skill_index_line,test_conductor_*,test_mate_conductor_parity,test_goal_conductor_skill_grants,test_crew_mode_steer_ledger,test_ledger_conductor_agent,test_team_lead_agent,test_work_ledger_doc): 258 passed. No test pinned the old prose.Manual verification
N/A — prompt-only change; behavior shows in the next goal run.
Related Issues
Closes #36294
Follow-up: #37192 syncs the "and nothing is queued" stop clause into
patrol_ends, the tick texts andwork-ledger.md.Stands alone. No dependency on another PR: the rule uses only
resource_statusand the patrol tick, both on main.Checklist