Repository navigation
docs(rfc): a conductor drops its chat after a turn that closes an item or fills a free seat - #22088
Conversation
…m or fills a free seat
UX Review (Opus 5.5) — ⏭️ skippedRevision |
First Principles Review (Opus 5.5) — ⏭️ skippedRevision |
Design Review (Opus 5.5) — ✅ PASSDesign-level review of Design-Verdict: PASS Before merge, check that the RFC's reset trigger ("closed an item or filled a free seat") can't fire on a turn that only re-patrolled. If it can, quiet cycles start resetting too. [DESIGN-REVIEWED] 7dc4308 Description read[DESCRIPTION-READ] b451bba2560e4630bc3b644228346e4869ffa68fb1f225ec8652d4b73b37f246 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 |
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: |
bolichen97
left a comment
There was a problem hiding this comment.
Looks good as a draft RFC. This needs a design sign-off from a maintainer, so I am leaving a comment and not an approval.
Head reviewed: 7dc4308. Bots read: Design PASS, First Principles skipped (docs only), UX skipped, Opus no blocking findings, GPT no blocking findings. Readiness was still checking when I read it.
What I verified against current main:
- The three skill quotes in section 1 are verbatim in the goal-conductor
SKILL.md: "dispatch the next item when one lands" (Concurrency cap), "When every item in the round has landed" (Close the round), and "the round is still running" (Never reset when). _CONDUCTOR_SYSTEM_PROMPTinagent.pystill says to reset once the next round is dispatched and never mid-round, so the contradiction the RFC describes is real on main._refuse_unvouched_wake_resetexists insession_directive_apply.py,test_conductor_agent.pyexists for the text pin, andwork_ledger_recordaction=decideaftersession_sendis already the skill's documented answer path.rfc-conductor-round-reset.mdisimplementedon main and is not edited here. The amendment lives in its own document, which matches how other amendments in the index are recorded.- The README change adds one index row and appends one audit sentence to the long audit paragraph. Nothing else in that paragraph changed.
On the Design lane's note: the trigger in section 3.1 requires the turn to close an item or dispatch one into a free seat, so a quiet re-patrol cannot fire it. I see no gap there.
One question for whoever accepts this: the Status line names the author as the owner of the decision. The linked issue is tiered as a change to existing behaviour, so please record the acceptance from a maintainer in section 8 or the Status line before the implementation PR opens, and say whose decision it is.
Other open RFC PRs (#21646, #21613, #19437) also edit the RFC index. They do not overlap in content. Whichever merges later needs a rebase of that one paragraph.
|
Problem / Motivation
Goal: Put the proposed conductor reset rule on main as a draft RFC, so its owner can accept it before the prompt and skill change that follows.
A goal conductor drops its chat only at a round close. Its own skill refills a seat as soon as an item lands, so under a full queue a round never closes and the reset never comes. On one operator's conductor DM, 10 patrol turns closed an item or filled a free seat in the 23 hours to 2026-10-19 05:00 UTC, and none of them reset.
Why it matters
Every patrol wake re-sends the whole chat. In that DM a request carried 450K to 550K tokens. Measured on the same host, a request after a reset was about 2 to 5 s faster and cost about 75% fewer credits. The wakes come 900 s apart, longer than the prompt cache lives, so each one pays for the whole chat.
The fix changes what the
kirocrew-conductoragent and the goal-conductor skill do by default. The First Principles lane reads that decision from an RFC on the base branch, so the RFC lands first.Not a goal
draftuntil its owner decides.rfc-conductor-round-reset.md. An implemented RFC is amended in its own document.session.autocompact_pct(KAS compaction: decide Crew's threshold and drop the dead refusal arm #18931 keeps the engine's thresholds).What changed (motivation → approach → change)
The goal conductor (
kirocrew-conductor) resets its chat only when "every item in the round has landed", and the goal-conductor skill's concurrency cap makes sure that never happens. The new RFC moves the reset point from the round to the turn.docs/request-for-change/rfc-conductor-rolling-reset.mddraftRFC. A patrol turn that closed an item, or dispatched one into a free seat, ends withsession_ledger_recordand thenreset_conversation, once every item it touched is at rest and no question is unanswered.docs/request-for-change/README.md"Mid-round" then means partway through the turn's own steps, not "other items are open". The RFC keeps the old non-goals: no timer, no reset forced by the gateway, and no change to
_refuse_unvouched_wake_reset. Its section 4 lists what survives ofrfc-conductor-round-reset.md.Backwards compatibility
Compatible: documentation only; nothing reads an RFC at runtime.
Removes nothing: the diff adds one document and one index row.
Tests
N/A — documentation only.
python3 scripts/docs_lint.pypasses on this branch.Manual verification
N/A — documentation only. The figures in the RFC were read from the conductor DM's transcript and the round-1 latency research on the same host, as section 1 says.
Related Issues
Part of #22082
Checklist
feat|fix|docs|style|refactor|perf|test|chore|ci|build|revert: ...)