Repository navigation
docs(rfc): conductor reads each open worker's new transcript lines on patrol and steers drift - #36360
Conversation
First Principles Review (Opus 5.5) — ⏭️ skippedRevision |
UX Review (Opus 5.5) — ⏭️ skippedRevision |
Design Review (Opus 5.5) — ✅ PASSDesign-level review of Design-Verdict: PASS Nothing blocks: the RFC is docs-only and records an owner-requested patrol change with a bounded cost. The code references cited in it check out. Notes
[DESIGN-REVIEWED] b792440 Description read[DESCRIPTION-READ] 26255c94045ffe17b942b4d94fef5942e7fd3642fe78a1e9803accf1f970fb44 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 |
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: ApproveWhy: the decider estimates a 7% chance that this change introduces a defect, within the 12% limit for approving on its own, and every check passes. This PR only changes docs: one new accepted RFC plus its index row. The RFC records the owner's decision that the goal conductor reads each open local worker's new transcript lines on every patrol turn, bounded by a 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 |
… patrol and steers drift
740ece9 to
b792440
Compare
|
bolichen97
left a comment
There was a problem hiding this comment.
Reviewed the whole diff at b792440. It adds one RFC document and its README row, and changes no code. This is a design brief, not an approval.
What it decides: on every patrol turn of the goal conductor, floor turn included, the conductor reads each open local worker's new transcript lines (session_read_message, since the stored cursor, limit 10), compares them with the item's title, acceptance and goal, and sends a steer correction through session_send if the worker has drifted. It records the correction with work_ledger_record and keeps the cursor in its session ledger's next. A remote-dispatched item is skipped on a patrol wake and checked on the next owner turn.
Pros:
- It closes a real gap: today the conductor sees only each worker's own
work_reportclaims, so a worker that widens scope looks on track until itsdonefails acceptance, after the item's run is spent. - It adds no turns, tools, fields or stored formats. It rides on wakes that already happen and on reads the conductor is already granted.
- The cost is bounded and written down: at most two reads per item per turn, 20 rows per item, and at the default cap of 6 open items 12 calls and 120 rows, each row capped at 4,000 characters.
- It keeps the earlier decisions intact: liveness stays with the
staleandorphanedflags, acceptance stays withaccept_eval, and a transcript never passes or fails an item. - The read goes through the existing session read path, which authorizes the target and redacts each row, so the conductor reads only what it could read by hand and nothing leaves its own turn.
- The monitor judge is considered and rejected for sound reasons (it screens quiet ticks only, its criteria are capped at 500 characters, it cannot steer, and its targets move with every dispatch).
Cons and risks:
- Drift is a model judgement. A false steer sends a worker a correction it did not need, and a missed drift still waits for acceptance. The document says on-track means silence, which keeps the false-steer cost low, but it has no measure of either error rate.
- Remote-dispatched items get no drift check on a patrol wake, so the gap stays open for them until an owner turn.
- Each patrol turn now carries up to 120 more rows of context for a full crew. That is bounded, but it is context spent on every wake, including the quiet floor turn.
Long-term debt: low. It is a skill and prompt change with no new schema. The one thing to watch is the per-item cursor kept in the session ledger's next, which is free-form text the skill must keep writing in the same shape, so a later change to that field should name it.
My stance: the direction is right and small, and I have no objection to it. The sign-off on this RFC stays with the maintainers. The implementing change is a separate pull request, and I would want its parity test and agent-spec goldens to move as this document says.
KiroCrewCaptain
left a comment
There was a problem hiding this comment.
Approving. The author is the repository owner, and this RFC records the owner's own decision, so the standing sign-off for design docs from team members applies.
Review result on this head: nothing blocks. The PR adds docs only: one new RFC and one index row. The cited code checks out at the PR head and at 27e9359419:
MAX_READ_CONTENT_CHARS,MAX_JUDGE_TARGETSandMAX_JUDGE_CRITERION_CHARShave the values the RFC states.DEFAULT_GOAL_CONCURRENCY_CAPmatches, and so do thecursor_unavailableandnext_sincebehaviour of the read and the QUIET-only branch of_monitor_tick_is_quiet.
Done well: §3 rejects the monitor judge with reasons you can check in the code, and the cost bound (12 calls and 120 rows per turn at the default cap) is spelled out.
For the security review:
- No boundary moves in this PR. Every read still goes through
authorize_target, and each row goes throughredact_and_truncate. session_read_messageis already in_CONDUCTOR_DASHBOARD_GRANTS.- The drift steer is a
session_send. It skips the prompt only for the conductor's own child sessions, and only whileagent.conductor_own_child_verbsis on (the default). Otherwise it prompts.
Optional:
- §1 and §2.2 say a floor patrol comes after
_MAX_QUIET_STREAK(10) quiet ticks. When no worker has reported since the last delivered turn, which is the silent-drift case this RFC targets,_quiet_floorreturns_IDLE_LEDGER_QUIET_FLOOR(40 ticks). Consider stating the real cadence. - §2.1 quotes
rfc-conductor-work-ledger.mdline 722 with an ellipsis that drops "that path keeps its human approval prompt", and §6 says the opposite. One sentence saying this RFC amends that clause would help. §6 could also say the step adds steers as well as reads, and name the approval path above. - #36302 at its current head does not yet carry the remote-item skip or the at-most-two-reads wording that §5 lists as exit criteria.
bolichen97
left a comment
There was a problem hiding this comment.
Reviewed the whole diff at b792440. It adds one RFC document and its README row, and changes no code. The direction matches the repository owner's ask, and it is small and low in long-term debt: on every patrol turn the conductor reads each open local worker's new transcript lines with a stored cursor, a limit of 10 and at most two reads per item, and steers only when the worker has drifted. Liveness and acceptance stay where they are.
The one thing to watch in the implementation pull request is the per-item cursor kept in the session ledger's free-text next. The skill must keep writing it in the same shape, so the change should name that field and its format.
The earlier failure of the Sensitive Change Review helper on this head was followed by a passing run of the same check, and everything else is green. Auto-merge is already set by someone else, so I leave it as it is.
Direction signed by maintainer (bolichen97).
Problem / Motivation
Goal: Record the owner's decision that the goal conductor checks each open worker for drift on every patrol turn, so the skill change in #36302 traces to an accepted RFC.
#36302 changes the goal conductor's default patrol: on every patrol turn it reads each open worker's new transcript lines from a stored cursor and steers drift. First Principles blocked it (span
717730afd96c) because that is a product-shape change andrfc-conductor-work-ledger.mdreplaced the transcript-reading patrol. This document records the decision on the base branch.Why it matters
Without it the conductor only sees each worker's own claims, so a worker building the wrong thing is caught at acceptance, after the whole item's run is spent. The owner asked for the check.
Not a goal
stale/orphaned) or to acceptance (accept_eval).What changed (motivation → approach → change)
Product-shape change needs a record on the base → a short RFC, docs only →
docs/request-for-change/rfc-conductor-drift-check.md(statusaccepted, takes effect on merge) and its row indocs/request-for-change/README.md.The RFC states three things:
rfc-conductor-work-ledger.mdkept (line 722).sincecursor (an unmoved transcript returns no rows),limit: 10, at most two reads (20 rows) per open item per patrol turn, on turns the patrol already takes. Remote-dispatched items are skipped on a patrol wake, since their reads and sends need an owner turn.judgecannot replace it: it screens QUIET ticks only (_monitor_tick_is_quietinautonudge_service/gate.py), sees no item title or acceptance (two 500-char criteria, 8 targets), cannot steer, and its targets change on every dispatch.It amends
rfc-conductor-work-ledger.md. Every code reference was read at27e9359419.Backwards compatibility
Compatible: docs only.
Removes nothing: adds one RFC and one index row.
Tests
Proof: not provable -- docs or config only
Manual verification
N/A — docs only; every cited constant and line was re-read at
27e9359419.Related Issues
Part of #36299
Checklist