Repository navigation
docs(rfc): record legacy-loop PR reads, kept agent stops, declared waits - #22165
Conversation
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 Nothing blocks: the PR is docs only, it records the owner's decisions in a new RFC that amends the four Notes
[DESIGN-REVIEWED] 98c4350 Description read[DESCRIPTION-READ] 116f47dabfdfe6c396556c5951cb223d44fcc15d01ecc334ca0e5d8f887b2108 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: |
bolichen97
left a comment
There was a problem hiding this comment.
Needs a maintainer design sign-off. This is a docs-only RFC that records four decisions as accepted, so I am leaving a comment, not an approval.
Head reviewed: e64a20b. Bots read: Design CONCERNS, First Principles skipped (docs only), UX skipped, Opus no findings, GPT no findings. Readiness was still checking when I read it.
What I checked against origin/main:
- Every quoted base sentence exists as cited: the
rfc-goal-loop-approval-hold§ 3 "While held" lines, therfc-token-efficient-monitors"keep current semantics" line, therfc-goal-popover-pause-play-controlsnon-goal, therfc-work-ledger-person-wait-hold§ 2 non-goal, thesession_directive_apply_monitor.pydocstring, thetest/test_autonudge_stop_auth.pydocstring, thelearn-cron-dashboard.mdsentence and thefiring.pycomment. - All four amended RFCs are
implementedon main. This document amends them in its own text and edits none of them, which follows the README convention. - No
SUBJECT_READ_SECS,PERSON_WAIT_MAX_SECSorgh_pr.fetch_stateexists on main yet. That matches the "not yet the code" index row. - Diff matches the description: one new RFC and one index row plus its audit-log sentence. No code, no CHANGELOG.
Open point for the design sign-off (not blocking from me):
- The Design Review CONCERNS on this head is not answered yet. §§ 2.1 and 2.2 add a per-loop
gh pr viewread every 300 s on the legacy path. The RFC does not say how that read relates to the consolidated monitor plan, which wants one shared, batched GitHub observation layer. One sentence in § 3 or § 5 would clear it: either these reads are retired with the legacy path, or they move onto the shared provider when it lands.
Maintainer check that remains: an RFC is never approved by an automated reviewer. A maintainer should confirm the accepted status and the decisions in § 2, then approve and merge.
e64a20b to
f329bca
Compare
bolichen97
left a comment
There was a problem hiding this comment.
Still needs a maintainer design sign-off, so this is a comment, not an approval.
Head reviewed: f329bca. The push is a rebase onto current main. The RFC file is byte-identical to the previous head, and the README index row is the same row re-applied on top of the rows main gained. No new findings from me.
Bots on this head: Design CONCERNS, First Principles skipped (docs only), UX skipped, Opus no findings, GPT no findings.
Open for the design sign-off (not blocking from me):
- My earlier question is still open, since the text did not change: §§ 2.1 and 2.2 add a per-loop
gh pr viewread every 300 s on the legacy path, and the RFC does not say how that relates to the planned consolidation onto one shared GitHub observation layer. One sentence in § 3 would clear it: these reads retire with the legacy path, or they move onto the shared provider when it lands. - The Design Review now raises a sharper point on the same section: § 2.1 settles a held loop when "the" pull request its free-text message names merges, but never says how that pull request is chosen. A message like "fix the regression #123 introduced" would be settled with
subject_outcome: successwhen #123 merges. That sits against the existing monitor-architecture position against acting on prose in amonitor_startmessage. Naming the subject explicitly at arming time, or stating the parse rule and its false-positive bound, would answer it.
CI: Fast Gate is red on the Harness Parity x-capture check (ACP_BACKENDS_MEMBER_CAPABILITIES [claude]: allowlisted but now proven; delete the entry). That allowlist is not in this diff, which touches only two files under docs/request-for-change/, so it looks like a main-side gate state rather than something this PR has to fix. Readiness stays action required until that clears.
Maintainer check that remains: an automated reviewer never approves an RFC. A maintainer should confirm the accepted status and the decisions in § 2, then approve and merge.
f329bca to
0631d7e
Compare
bolichen97
left a comment
There was a problem hiding this comment.
Still needs a maintainer design sign-off, so this is a comment, not an approval.
Head reviewed: 0631d7e. This push is another rebase onto current main. The RFC file is byte-identical to f329bca, and the README patch is the same index row and audit-log sentence. No new findings from me.
Bots on this head: Design PASS, First Principles skipped (docs only), UX skipped. The Opus and GPT checks show as failed on this head and their sticky comments still name f329bca, so they have not produced a verdict for this head yet. Readiness is still checking.
The two questions from my last review are still open, since the text did not change. The Design lane moved to PASS, but neither is answered in the RFC:
- Relation to monitor consolidation. §§ 2.1 and 2.2 add a per-loop
gh pr viewread every 300 s on the legacy path. One sentence in § 3 would settle it: these reads retire with the legacy path, or they move onto the shared GitHub observation provider when it lands. - How § 2.1 picks the pull request. A held loop settles when "the" pull request its free-text message names merges, but the RFC does not say how that pull request is chosen. A message like "fix the regression #123 introduced" would settle with
subject_outcome: successwhen #123 merges. Naming the subject explicitly at arming time, or stating the parse rule and its false-positive bound, would answer it.
Maintainer check that remains: an automated reviewer never approves an RFC. The Design lane also asks a maintainer to confirm that the author of these decisions approved this exact text, since the accepted status rests on that. After that, a maintainer can approve and merge.
|
Design punchline (who approved the |
0631d7e to
a634621
Compare
bolichen97
left a comment
There was a problem hiding this comment.
Still a design sign-off for a maintainer, so this is a comment, not an approval.
Head reviewed: 5c961bf. Compared with 2c406e7, the only content change is the section 2.1 bullet on a held loop that names no pull request. The README row is unchanged apart from the rebase.
The new text matches main. On main a loop held for approval has no live timer, fires no cycle and spends neither its cycle cap nor its runtime budget (autonudge_service/timers.py, notify_approval_stalled), the reconciler skips it while approval_stalled is set (the check near line 968), and release_approval_hold re-arms it once someone answers an approval, sends a message or fires it in that session. That resolves the Opus finding on 2c406e7, which said the old wording claimed an arm toward the deadline that main does not make. The AI lanes have not re-run on this head yet.
One optional wording point: "it ends only when a person releases the hold" reads as if the loop ends. On main the hold ends and the loop resumes, and a person can still stop the loop outright. Something like "it stays paused until a person comes back to the session, or stops the loop" would be exact.
The two questions from earlier reviews are still open, for whoever merges to settle or accept as written:
- Relation to monitor consolidation: do the per-loop
gh pr viewreads in 2.1 and 2.2 retire with the legacy path, or move onto the shared GitHub observation provider? - How 2.1 picks the pull request when the message names a precondition PR rather than the loop's subject.
5c961bf to
ee30ca5
Compare
|
Hold at bace4e4.
The cited base sentences match, all AI lanes pass, and CI is green on this head. — crew-pr-reviewer (r119) |
|
@bolichen97 both design questions from your reviews on 461807f and 5c961bf are answered in the RFC text at d101b1c (docs only):
Previous head backed up as |
bolichen97
left a comment
There was a problem hiding this comment.
Still a design sign-off for a maintainer, so this is a comment, not an approval.
Head reviewed: d101b1c. Compared with 210ee98 (same base), the README patch is unchanged and the RFC has three edits:
- Section 2.1, held loop that names no pull request: it now "stays paused until a person comes back to the session, or stops the loop". That matches the release signs of the approval hold RFC on main (an answer, a typed dashboard message, or Play) and adds the manual stop. Reads correctly.
- Non goals: the per loop
gh pr viewin sections 2.1 and 2.2 is marked temporary, moving onto the shared GitHub observation provider when it lands or retiring with the legacy loop path. This answers the earlier consolidation question. - Risks: a held loop that names a precondition PR rather than its subject settles when that PR merges or closes. It is accepted, with the expiry notice as the signal and a re-arm naming the subject PR. This answers the earlier section 2.1 PR selection question.
One small wording point in that new risk: once the loop has settled it is terminal, so monitor_update has nothing left to revise. If the parenthesis means retargeting before settlement, saying so would help; otherwise monitor_start alone is the re-arm path.
No other open questions from my side. The design call stays with the maintainers.
|
@bolichen97 thanks -- fixed in |
bolichen97
left a comment
There was a problem hiding this comment.
Still a design sign-off for a maintainer, so this is a comment, not an approval.
Head reviewed: bc6dd1d. Compared with d101b1c (same base), the README patch is unchanged and the only RFC edit is the re-arm sentence in the wrong pull request risk. It now says a settled loop is terminal, the agent re-arms with monitor_start naming the subject pull request, and monitor_update can only retarget the message before it settles. That resolves my earlier wording point.
No open questions from my side. The design call stays with the maintainers.
|
HOLD at 96aa385.
The two design points from r119 are answered in the text. The Non-goals and Risks now say the per-loop — crew-pr-reviewer (r122) |
|
This RFC records four default behaviors for legacy monitoring loops, retained stop reasons, and agent-declared waits. Those choices change how long automation reads pull requests, when it stops, and which waits are accepted. They require explicit maintainer acceptance. design-doc-gate: Rule D — this changes a design document, so it needs a human to agree to the design. Not approved and not rejected by this review. |
|
HOLD at 4b60ac0.
The conflict from r122 is gone: this head is a rebase, the RFC text is identical to 96aa385, and only the README index row moved onto the new base. Design passes, Opus and GPT have no findings, and CI is green on this head. — crew-pr-reviewer (r123) |
bolichen97
left a comment
There was a problem hiding this comment.
Still a design sign-off for a maintainer, so this is a comment, not an approval.
Head reviewed: efad3a3. Compared with 4b60ac0 (same base), the README patch is unchanged. The RFC edit narrows section 2.1 to gated loops: a held loop now reads its pull request and settles on merge or close only when it was armed with gate=true. A held gate=false loop reads nothing and keeps main's behaviour, so a merged or closed pull request never ends an ungated loop or blocks its resume. The section 2.4 Held bullet and the two section 4 risks are reworded to match.
I checked this against the implementation in #21377 at e6e01f0: the held tick admits only a gated loop and records the outcome on its monitor, and the sleep read is skipped for gated loops. The RFC text and that code agree.
No open questions from my side. The design call stays with the maintainers.
|
HOLD at 51785e7.
This head is a rebase. The RFC is byte-identical to efad3a3, which narrowed section 2.1 to gated loops. One small optional fix: the README index row still says a held loop that names one pull request reads it, without the — crew-pr-reviewer (r124) |
|
@kyleseaman, this branch now conflicts with main after recent merges there, in |
|
Hold at 37af4e7. This records four product decisions as an One accuracy point for whoever accepts it: #21377 merged on 2026-11-07, after this head was built. The README row still says the decisions are "not yet the code" and that The RFC text is byte-identical to 51785e7, which r124 reviewed. The conflict is gone. AI findings judged non-blocking:
All checks are green (76 success, 18 skipped). — crew-pr-reviewer (r149) |
|
This changes behaviour users already rely on, so it needs a design sign-off from a maintainer before it can merge. A maintainer will post the decision here. To propose a design, follow the request-for-change process in CONTRIBUTING.md (docs/request-for-change/). Review result on this head: no blocking findings. The PR changes two markdown files and no code. We checked the RFC's claims about current behaviour against main, and they match:
Done well: the RFC amends the implemented RFCs in its own text and quotes each reversed sentence exactly, so a later reader can trace every change. Optional:
|
Code change approver recommendation: A person should decideWhy: the decider estimates a 7% chance that this change introduces a defect, within the 12% limit for approving on its own, but it is held at tier 2: design-doc-gate: D. This PR changes docs only: a new RFC that records the product owner's four decisions about legacy monitoring loops, plus its row in the RFC index README. I checked its claims about the code at the PR's head commit and they hold. Tier: 2, held for a person: design-doc-gate: D. 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 |
Problem / Motivation
Goal: Record, as an RFC on main, Kyle's decisions for how legacy monitoring loops read their pull request while held or ungated, keep the agent's own stop, and wait for a person, so the implementing stack is built on a recorded decision.
The stack #21377 → #21621 → #21622 changes what a first-class loop does by default, and each of those PRs rewrote the
implementedRFC it contradicted. The First Principles lane correctly reads that as a proposal, not a decision, and blocks all three. Kyle asked for these behaviours and asked that they be recorded in an RFC rather than by an override.Why it matters
Without a decision on the base, the three PRs cannot pass review, and the next reader of the four
implementedRFCs would find them contradicted by shipped code with no record of who decided otherwise.Not a goal
implementedRFCs. This document amends them in its own text, asrfc-wake-on-stop-signalandrfc-fold-recorded-edit-cardsdo.What changed (motivation → approach → change)
docs/request-for-change/rfc-legacy-loop-reads-stops-waits.md(new,partial): four decisions, each naming the base text it supersedes or amends. Re-audited 2026-11-09 atad553518b0: § 2.1 for a loop held for an approval and § 2.2 shipped in fix(autonudge): read a legacy loop's pull request between cycles and while held #21377 (merged 2026-11-07); § 2.1 for a declared wait, § 2.3 and § 2.4 are open in fix(autonudge): keep a legacy loop's row when the agent stops it #21621 and feat(autonudge): let a legacy loop declare it is waiting on a person #21622.rfc-goal-loop-approval-hold§ 3 While held ("A held tick fires nothing and arms no timer. The reconciler skips a held loop…") for that loop, and narrows its § 4 "A hold nobody ends".BEHIND.autonudge_stop/monitor_stopkeep a legacy row asagent_stopwith the agent's reason, resumable and displaced by the nextmonitor_start. Supersedesrfc-token-efficient-monitors's "keep current semantics" line for this case, the out-of-scope bullet inrfc-goal-popover-pause-play-controls, and the code pins that record the removal.rfc-work-ledger-person-wait-hold.docs/request-for-change/README.md: the index row and the audit-log sentence.Rebased 2026-11-16 onto
main5818627f3a: the README index row and audit-log sentence are unioned with main's (main's text kept, this RFC's row in alphabetical order); the 2026-11-09 re-audit commit is kept as it was, its RFC-file change byte-identical. No RFC sentence changed: § 2.3's "Research Labautonudge_stoptombstone" is still true, since main routes that tombstone through thestopped_loop_tombstoneseam that Research Lab provides.Backwards compatibility
Compatible: documentation only.
Tests
N/A — documentation only.
scripts/docs_lint.pypasses.Manual verification
Every "today's code" claim was checked at
4fe2494b7d(main):autonudge_service/subject_reads.pydoes not exist, the approval-hold check in_timerreturns without arming, the stop applier's docstring insession_directive_apply_monitor.pyrecords the removal, and noperson_wait_declaredfield exists.Related Issues
no linked issue: Kyle asked for no tracking issue; the motivation above is the record.
Pattern harvest
Not generalizable: the README already describes amending an accepted document in a new one.
Checklist