Repository navigation
docs(rfc): one reroute-first ask policy for every crew kind - #37623
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 Docs-only RFC with a stated rollback; skipping the plan gate removes a chat wait, not an approval. Notes
[DESIGN-REVIEWED] d40eb73 Description read[DESCRIPTION-READ] 961fca8ac83d40d077341e12c8d40e364085c43764e6cb4d4ae06fd93fe237da 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: |
|
This RFC changes when a conductor waits for a person's go-ahead on its plan. In a turn nobody typed, it would treat the plan as approved and dispatch round 1 at once. A design that lets an agent skip a person's approval step is for the maintainers to decide, so this review does not approve it. A maintainer will post the decision here. The front matter says What the review checked on this head:
One thing done well: section 7 reports results that cut against the design, such as the rise in long-task asks and the three fixes that were not re-benched. It does not hide them. For the security review:
Optional:
|
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 adds docs: an accepted RFC and its index row, written so the implementation in #37609 can land on an RFC that is already on main. I checked the RFC's "current state on main" quotes against the code and they are there:
The gateway markers that section 5.7 relies on also exist where it cites them:
Trust check: the Round 0 plan gate is a chat wait, not an approval. A person can already skip it by typing "go ahead". Tool approvals, the sandbox and the deny gate stay unchanged, so someone who forges a marker gains nothing they couldn't already get by typing "go ahead". The trust boundary do 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 |
bolichen97
left a comment
There was a problem hiding this comment.
Design review of the RFC and its README row. I read the whole diff on this head (docs only, no code). I did not run tests locally and did not re read the code citations on main.
What it is. One ASK_CORE text replaces about nine differently worded "when to ask the owner" lists. Every crew kind reads it at agent rank and in every CRITICAL RULES variant. It says: take another allowed path first, act on a clear best option and report it, ask only for five named cases (a) to (e), and keep the safety holds. Section 5.7 also skips the Round 0 plan gate in a turn nobody typed, detected by a fixed gateway marker at the start of the turn.
Pros
- One source of truth instead of copies that drift, with a test that the list wording lives only in
output_contract.py. - The replay bench moves over asks from 18/55 to 9/55, and all 20 control moments where asking is right still ask.
- Section 5.7 changes only prompt text. The markers it relies on already exist in code, and a marker quoted inside a tool result or file does not count.
- Blocks are never re spelled: the same effect by another tool is still the blocked action, and approvals and required reviews are not bypassed.
Cons and risks
- Long task unneeded asks per run went from 0.10 to 0.24. The RFC says the sample is too small to tell, so it shows no gain there.
- Bot moments with lessons loaded only move from 10/15 to 8/15.
- The three fixes in 7.3 were not re benched (the RFC lists this as an open question).
- A goal the owner types into session A and A forwards with
session_sendnow runs without the plan gate. The RFC accepts this and says why. - The deny notice and lesson header changes reach every session, not only crew sessions. Always on guidance grows by about 1.1k to 1.5k characters per crew kind.
Tech debt. The two re bench items in section 11. The README row is a long "verified at" entry that will need a re audit when #37609 lands.
Recommendation. Worth doing, and the direction (reroute before asking, five fixed ask cases) is sound. Because this is a design RFC that changes the deny notice wording and the plan gate, I am leaving it as a comment and not approving. It needs the maintainer sign off on the direction, and the implementation PR #37609 should follow the same sign off. All 96 checks are green.
bolichen97
left a comment
There was a problem hiding this comment.
Direction signed by maintainer (bolichen97), 11-15 PST.
Approving the RFC and its README row. I read the whole diff on this head earlier (docs only, no code), and the head has not changed since. All 97 checks are green. I did not run tests locally and did not re read the code citations on main.
Two notes for the implementation PR (#37609). Neither blocks this document:
- Re measure the unneeded question rate with a bigger sample. The long task bench shows 0.10 unneeded asks per run on main and 0.24 on the new design, and the RFC itself says the sample is too small to tell them apart.
- The deny notice wording change must not refuse anything main allows. It should only change what the agent says after a block, not which calls are blocked.
Problem / Motivation
Goal: Record the owner's decision that crew agents reroute first and ask only for five named cases, and that a goal in a turn nobody typed does not wait at the plan gate, so the code in #37609 lands on an accepted RFC.
Crew agents stop and ask the owner too often, even when a fix or a clear recommendation exists. The ask cases are copied about nine times in different words, lessons say they override default behavior, the policy deny notice says to stop, and the goal-conductor Round 0 gate waits for a go-ahead even in a turn no person typed (cron, another session, a patrol wake), where nobody can answer. #37609 changes all of that, and its First Principles review asks for the RFC to be on main first, because an RFC shipped in the same PR does not count.
Why it matters
Without an accepted record, #37609 stays blocked, and the change to every crew kind's default ask and reroute behavior has no decision on file to audit or revert against.
Not a goal
What changed (motivation → approach → change)
Adds
docs/request-for-change/rfc-reroute-first-ask-policy.mdwith statusaccepted(asked for by the repository owner in the conductor chat on 2026-11-15; takes effect when this document merges) and its row indocs/request-for-change/README.md.ASK_COREtext at agent rank and in every CRITICAL RULES variant; every other ask list points at (a)-(e)Section 6 states the safety holds (no re-spelled blocks, no loosening, approvals and required reviews never bypassed, the plan gate skipped only on a gateway marker). Section 7 has the A/B bench results.
Backwards compatibility
Compatible: documentation only; nothing reads it at runtime.
Removes nothing: it adds one document and one index row.
Tests
Proof: not provable -- docs or config only
scripts/docs-lint.shpasses locally.Manual verification
N/A — documentation only.
Related Issues
Part of #37606
Implementation: #37609.
Checklist
feat|fix|docs|style|refactor|perf|test|chore|ci|build|revert: ...)