Repository navigation
feat(conductor): share the long-task and team rules with the Bot and root the parity test at conductor mode's agent - #37725
Conversation
…root the parity test at conductor mode's agent Per docs/decisions/2026-11-15-conductor-mode-runs-on-the-conductor-by-default.md: eight rules move into _LEDGER_CONDUCTOR_RULES, so kirocrew-conductor's charter and a Bot's conductor.patrol both render them every turn (item budget, one owner per file / one integrator per output / author never tests own work, merging two lines, plan reviewer, pr_merged bar, hourly quiet-wake check, routine incidents, report before park). The tracker rule is root-only and allowlisted. The team-lead rules worth keeping are folded into goal-conductor. test_mate_conductor_parity.py now roots at crew_mode.crew_mode_agent() and checks every shared rule is in both installed prompts.
UX Review (Opus 5.5) — ⏭️ skippedRevision |
Design Review (Opus 5.5) — ✅ PASSDesign-level review of Design-Verdict: PASS No design-level problem: the change is prompt text added through the existing shared-rule mechanism, it removes nothing, and the parity test now starts from the agent conductor mode actually runs. Notes
[DESIGN-REVIEWED] 218277c Description read[DESCRIPTION-READ] dee2ab3093a3eedc12ed2924fd56ebd240fd03cfd9a0d0a63ae97e7cb5139a34 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 one weak premise is the new Notes
What this change shipsInventory (10 items) — 7 justifiedIntent: both conductors (a Bot's and conductor mode's) keep the long-task and team rules every turn, even after compaction; ADDITION.
[FIRST-PRINCIPLES-REVIEWED] 218277c Description read[DESCRIPTION-READ] 8dea54ff740e8d00fb1290b832df6a731aeb0934304bef30e1f0199c908c3fc7 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: A person should decideWhy: the decider thinks the change is probably fine but estimates an 18% chance that this change introduces a defect, above the 12% limit for approving on its own. Prompt-only change. It adds eight rules to _LEDGER_CONDUCTOR_RULES, and each one appears in both the root charter and the Bot's conductor.patrol. The root-only tracker rule is allowlisted with a reason, and the new goal-conductor skill text matches the anchors the test checks. The parity test's root now comes from crew_mode_agent(). It adds a check that each rule renders on both installed conductors, plus a working positive control. The re-pinned digests are consistent with the charter growing. The pr_merged bar with repo is supported by accept_eval and by the ledger's check that a bar is filled in. All CI checks and AI review lanes passed. The open questions are policy points, not concrete defects: pr_merged combined with the pending budget, and no written record for the pr_merged default. 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 |
KiroCrewCaptain
left a comment
There was a problem hiding this comment.
Reviewed head 218277c. The change is prompt text plus tests. The eight new shared rules in _LEDGER_CONDUCTOR_RULES reach both the root charter and the Bot's conductor.patrol by reference. Every new anchor phrase is present in the head goal-conductor skill. The re-pinned digests move only kirocrew-conductor.json, kirocrew-ledger-conductor.json and the _CONDUCTOR_SYSTEM_PROMPT shape. No grant, scope or gate code changes. PR Readiness is green on this head, and the branch merges cleanly onto current main.
Done well: each rule is written once and referenced from both sides. The new two-sided render test also ships with a positive control, which proves the test fails when one side loses a rule.
The design is signed off under the standing maintainer sign-off for changes by team members.
Optional:
bar_pr_mergedand the 24h pending budget:is_pending_overdueexempts onlyhuman_approval, so a green PR that waits more than 48h for review can end up closedabandoned. Consider exempting items that are waiting on a merge, or moving this rule to its own PR with a decision record.- The
routine_incidentsprompt text leaves out three things the skill says: "save its unpushed work first", "regenerable build output", and the ask-case carve-out. Copying them over would keep the prompt rule as strong as the skill rule. - #37652 conflicts with this PR on the same prompt text and digest pins. Whichever PR lands second needs a rebase and a re-pin.
Problem / Motivation
Goal: A Bot's conductor and conductor mode's conductor follow the same long-task and team rules every turn, and the parity test proves it against the agent conductor mode actually runs.
The long-task rules from #37030 (per-item budget, plan reviewer, hourly check on a quiet wake) live only in the goal-conductor skill, which a conductor may not re-read after compaction. The team lead's useful team rules (one owner per file, one integrator per output, author never tests own work, merging two lines, tracker, routine incidents, report before park) live only in the team-lead skill, which
kirocrew-conductorand Bots never read.test/test_mate_conductor_parity.pyhard-codedkirocrew-conductoras its root instead of readingcrew_mode.crew_mode_agent().Why it matters
The owner ruled that the Bot and conductor mode work the same way (2026-11-06) and that conductor mode runs on
kirocrew-conductorwith the team lead's rules folded into goal-conductor (docs/decisions/2026-11-15-conductor-mode-runs-on-the-conductor-by-default.md, landed by #37705). Rules that sit only in a skill are dropped after compaction; rules that sit in one agent's prompt drift from the other.Not a goal
kirocrew-team-leadagent or theteam-leadskill (follow-up below).goal_changessection, and the Bot'sconductor.patrolalready says to record the goal first.What changed (motivation → approach → change)
Rules every turn needs → one shared text both conductors render (
_LEDGER_CONDUCTOR_RULES, the existing mechanism) → eight new keys, each placed in_CONDUCTOR_PROMPT_SECTIONSand named as a{placeholder}increwmate_prompts.jsonconductor.patrol.conductor.patrolitem_budgetown_files(one owner / one integrator / author never tests own work)merge_linesplan_reviewerbar_pr_mergedpr_mergedfor a PR the goal wants merged)wake_check(hourly quiet-floor check)routine_incidentsreport_before_parktrackerALLOWLISTprompt:trackerwith reasonDropped, not folded (decision record): "When you keep a ledger", "your own loop is never an item", and the do-it-yourself bias. goal-conductor carries none of them.
Parity test (
test/test_mate_conductor_parity.py):crew_mode.crew_mode_agent()(_ROOT), not a hard-coded name; the installer is looked up from it.test_conductor_mode_runs_the_charter_this_file_comparesfails if the default moves to an agent whose prompt is not_CONDUCTOR_PROMPT_SECTIONS.test_every_shared_rule_renders_on_both_installed_conductors: every_LEDGER_CONDUCTOR_RULESvalue is in the installed root spec'spromptAND in the Bot's rendered ledger text. Positive controltest_the_render_check_sees_a_rule_lost_on_one_sidedropswake_checkfrom the Bot side and expects the failure.test_the_long_task_and_team_rules_are_sharedpins the eight keys on both sides.CONDUCTOR_MODE_SKILL_ANCHORSgains anchors for the eight keys in goal-conductor.docs/system-specs/modules/crew-mode.md("The two conductors stay in step") names the new shared rules and the new root.Token cost per turn (chars / 4)
kirocrew-conductorkirocrew-mate-conductor(Bot with ledger)conductor.patrolkirocrew-team-lead(opt-in only)Backwards compatibility
Compatible: prompt text only; no config key, schema or API changes.
Removes nothing: every existing rule, section, allowlist entry and anchor stays; only additions.
Tests
test/test_mate_conductor_parity.py: four new tests above plus eight anchor entries; existing tests now read_ROOT.test/test_agent_refactor_spec_bytes.py/test/test_agent_refactor_contracts.py: re-pinnedkirocrew-conductor.json/kirocrew-ledger-conductor.jsondigests and the_CONDUCTOR_SYSTEM_PROMPTshape, because the charter text grew._CONDUCTOR_SYSTEM_PROMPT,_LEDGER_CONDUCTOR_RULES,conductor.patrol,goal-conductor/SKILLorcrewmate_prompts, plustest_conductor_context_cap.py: 1115 passed.Proof: not provable -- other: prompt-text feature, not a fix; the positive-control test is the guard check.
Manual verification
N/A — the rendered prompts of both agents are asserted directly by the new tests.
Related Issues
docs/decisions/2026-11-15-conductor-mode-runs-on-the-conductor-by-default.md(fix(crew): conductor mode runs on kirocrew-conductor by default #37705 merged).Follow-up
kirocrew-team-lead(service_agents._install_team_lead_agent,_TEAM_LEAD_SYSTEM_PROMPT) and theteam-leadskill once no install still setsagent.crew_mode_agentto it; then droptest_a_team_lead_set_as_the_conductor_reaches_goal_conductor.Checklist
Contribution License Agreement
Contributions are accepted under the project's Apache-2.0 license.