Repository navigation
docs(goal-conductor): dissolve Known limits, fix stale approval list, never block a wake on an ask - #37618
Conversation
UX Review (Opus 5.5) — ⏭️ skippedRevision |
Design Review (Opus 5.5) — ✅ PASSDesign-level review of Design-Verdict: PASS Only the skill text and its pins change, and it all reverts cleanly; the new approval wording matches Notes
[DESIGN-REVIEWED] 8ae1ccf Description read[DESCRIPTION-READ] 604e0a619f140a1e4cd44e6a2bb5ccb4eb5f5d5780ab3afe8dbf04e2f4f9a3c5 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 W7 changes how the conductor behaves on every patrol wake, not just its text. Its only recorded support is an owner ruling quoted in the description, and it sits close to the "non-blocking ask form" that Notes
What this change shipsInventory (8 items) — 6 justifiedIntent: make the goal-conductor skill text match what the code does by spreading its "Known limits" section across the steps it belongs to. This is a FIX.
[FIRST-PRINCIPLES-REVIEWED] 8ae1ccf Description read[DESCRIPTION-READ] 7036739842f3d92ef89904dcdb358598d660d033c28994ea610a1b9c84af54a7 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: |
72bf066 to
2aefa07
Compare
fa3b7d7 to
55aef1f
Compare
… never block a wake on an ask
55aef1f to
8ae1ccf
Compare
KiroCrewCaptain
left a comment
There was a problem hiding this comment.
Reviewed head 8ae1ccf. These checks were made on that head:
- The new "What prompts." list matches the three grant tuples in
agent_parts/spec_grants.pyand the own-child send/close verbs intool_approval.py. work_dispatchis still withheld.- The three wake markers W7 names all exist in the code.
work_ask_recordaccepts adonehuman_approvalitem.- The debate budget default of 10 replaces a cap that no longer exists.
- Each sampled Known-limits bullet is still present at its new place.
- The new and changed test assertions hold against the head file.
Done well: each stale claim is fixed against the grant code, and the body has a per-bullet table showing where every removed line went.
The design is signed off under the standing maintainer sign-off for changes by team members.
Optional:
- W4 (SKILL.md:111) is circular, because
fileis already one of the listed kinds. Afilebar also passes once the report exists. Consider saying that the conductor reads the report before it closes the itemaccepted. - W7 (SKILL.md:645) lists three "turn nobody typed" markers. #37609 lists nine, including
[monitor wake]. Consider pointing at one shared list and naming the merge order with #37609. test_skill_states_the_real_approval_costcompares prose with prose. Asserting the backticked verbs equal_CONDUCTOR_CORE_GRANTSand_CONDUCTOR_OWN_CHILD_VERBSwould catch the next drift.
Problem / Motivation
Goal: The goal-conductor skill has no "Known limits of this version" section; each of its bullets sits where it applies, and the text the conductor reads matches what the code does.
The owner asked why the skill carries a "Known limits of this version" section. It is a 96-line grab bag at the end of
goal-conductor/SKILL.md: real limits, tool-loading tips, rules that belong in a step, and duplicates. Part of it is stale. It sayssession_send/session_closeneed an approval per seed, answer and close, but the code auto-approves both into a conductor's own child. The Needs-human checklist lists four risky cases that differ from the five asks in Patrol step 2.references/worker-debate.mdsizes a debate budget from a goal item cap that no longer exists. Three passages contradict another passage in the same file.Why it matters
The conductor plans around the text. A stale approval list makes it expect clicks that never come (and budget turns for them). A second, different ask list lets it ask for a case the five asks say to decide, or skip one they say to ask. A budget tied to a removed cap has no value.
Not a goal
refused/errorno longer "ask"), W8 (third empty round: the conductor decides, records and reports, no ask; the round count is its own bound, not an ask case), W10 and W11 (worker-debate.mdrules 4-5: a self-set cap or budget is the conductor's call, an owner-approved spend stays (b)).agent_parts/spec_prompts.py:438-442: covered by feat(prompts): one reroute-first ask policy for every crew kind #37609, which points both at ask cases (a)-(e). This PR leaves that text as on main.What changed (motivation → approach → change)
Audit of the section → every bullet placed by kind (real limit stays short where it applies; tool tip becomes one line; misplaced rule goes into its step; duplicate is deleted) → heading removed. Every stale claim was checked on
origin/mainfd4c0d3 first. The intro gains a "What prompts." paragraph that matches the grant code; the debate budget gets a number; three contradictions are reworded. Owner ruling 2026-11-15 added two ask fixes from the audit: an unassertable item becomes afileitem (W4), and ahuman_approvalask in a turn nobody typed is recorded withwork_ask_record/needs_youinstead of a blockingask_question(W7). Code citations, the per-bullet placement and every moved safety line are under "Evidence" below.Evidence (stale claims, placement, safety lines)
session_sendandsession_stopare deliberately NOT auto-approved … one approval per item at dispatch, per question, per close"; list omitswork_ask_record,work_reportdashboard/chat_turn/tool_approval.py:353-354,447-464own-child send/close auto-approved forkirocrew-conductor/kirocrew-team-lead;config/sections_agent.py:1452-1453defaultTrue;agent_parts/spec_grants.py:207-209work_dispatchwithheld;:268-272,:483-489,:501-521granted listsDEFAULT_GOAL_ITEM_CAP = 20);work_ledger_writes.py:506caps open items onlytask-dashboardvalidation_parts/messaging_session_schemas.py:614-633work_dispatchalways createswork_dispatch, or to re-bind (skip step 2)fileis an evaluator kind (scripts/accept_eval.py)fileitem;human_approvalonly when a person must judge ithuman_approvalalways verified by a blockingask_question(up to 30 min)[auto-nudge cycle N](autonudge_service/model.py:884) or[work-ledger wake](probes/work_ledger.py:109-334);work_ask_recordis granted (agent_parts/spec_grants.py:485)work_ask_record+needs_you/[OPTIONS:], keep patrolling;ask_questiononly in a user-typed turnKnown-limits bullets, old → new:
ask_questionholds the turnhuman_approvalparagraph ("keep patrol work out of the turn that asks", the wording #37609 keeps)tool_searchidsnot_boundfor a root conductorsession_send; prompt is not an askwork_dispatchfailuressession_control_disabledwork_dispatchfailuressession_sendreports deliverysummaryis prosereportedclauseexecute_bashfallbackSafety lines moved (old line on fd4c0d3 → new section): S2/S3 evaluator takes no command, fixed argv,
refusednever routed around (1146-1158 → intro; step 3refusedline unchanged); S11 never route an item totask_run/workflow_run/spawn_*(1111-1113 → intro "What prompts"); S32 decide from verdict, notsummary(1131-1136 → step 3); S38 out-of-bounds targets refused (1142-1145 → Which agent); S39 sends to sessions you did not create prompt, you ingest external content (1114-1121 → intro "What prompts"); S16 conductor specs cannot write (1091-1095 → kept at "Which agent");ask_questionblocks (1064 → step 3).Backwards compatibility
Compatible: skill text only; no code, schema or config changes.
Removes nothing: every bullet of the removed section is kept at a new place or already said elsewhere in the same file (tables above).
Tests
test/test_conductor_agent.py::test_skill_states_the_real_approval_costnow pins the true approval text: own-child send/close do not prompt, other sends prompt,work_dispatchprompts once per item,session_stopand other core tools prompt.test/test_conductor_agent.py::test_skill_never_blocks_a_wake_on_a_question_cardpins W7 (noask_questionin a turn nobody typed) and W4 (unassertable ->fileitem).test/test_*.pythat reads the goal-conductor skill (25 files): 860 passed.Proof: not provable -- docs or config only
Manual verification
N/A — skill text only; the pinned tests read the shipped file.
Related Issues
N/A
Checklist