Repository navigation
docs: decide goal blocker lifecycle - #13013
Conversation
|
Intent: Record the maintainer decision that remediable blockers and human-only approvals keep autonomous goals active within finite budgets and a strict no-self-grant authority ceiling. |
GPT 6.1 Review (fork) — ✅ no blocking findingsReviewed Review detailsNo findings. |
Design Review (Opus 5.5, fork) — ✅ PASSDesign-level review of Design-Verdict: PASS Nothing blocks: this is a draft RFC with no runtime change, and it splits each costly phase (1b, 2) behind its own recorded maintainer decision. Notes
[DESIGN-REVIEWED] bcfdf4e Description read[DESCRIPTION-READ] b9a316d0cdd917f4bf723555cf64c8d6d915cfa64c10dbe0ee684d09bcf60aaf 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 (fork) — 🔴 changes requested (blocking)Reviewed The new BLOCKING — docs/request-for-change/README.md:89
|
795067b to
671cf22
Compare
671cf22 to
f63249a
Compare
|
span=ab637e60142f
|
|
span=14be823d2b60
|
|
span=14be823d2b60
|
|
Rebased onto main
|
|
Rebased onto main Checked on this head: the documentation tests that read the RFC index (1,094 passed), docs lint and the other script gates, the flake-pattern and wait-discipline ratchets, and the single-commit push guard. The body's current-head paragraph is regenerated: all 50 repository paths the RFC cites exist on this main, and the identifiers it cites beside the six cited files main changed since |
|
This is the round push for The README stays main's plus this RFC's sentence and row. Checked on this head: the documentation tests that read the RFC index (1,094 passed), docs lint and the other script gates, the flake-pattern and wait-discipline ratchets, and the single-commit push guard. The body's current-head paragraph is regenerated. |
bolichen97
left a comment
There was a problem hiding this comment.
Re-reviewed the change since 5e8f9a4 on head b24b202. The head was rebased onto main; the README edits are unchanged, and the RFC changed in two places.
- The "Fail closed" bullet no longer overclaims. It now says plainly that a missing or aliased record leaves the row uncommitted, so the cap of section 2 does not apply and the loop's own session can raise that row's bounds until an owner recommits. That matches section 3, which decides a failed leaf check as no record. Good correction.
- New open question 3 lays out the two choices for that gap. Option (b) rests on the default
agent.sandboxof"auto"; that default matches main.
One small gap: open question 3 says the Phase 1 cap "already waits on open question 1, and this choice joins it", but the Phase 0 exit criteria and the Phase 1 paragraph still name only open question 1 as what the bound-write cap waits on. Suggest adding open question 3 to both, so the acceptance record cannot close with this choice unrecorded.
Design sign-off on this RFC stays with the maintainers; this review does not approve it.
|
This is the round push for The other reds on Checked on this head: the documentation tests that read the RFC index (1,094 passed), docs lint and the other script gates, the flake-pattern and wait-discipline ratchets, and the single-commit push guard. The body's current-head paragraph is regenerated. |
bolichen97
left a comment
There was a problem hiding this comment.
Re-reviewed the change since b24b202 on head f547c9c. The head was rebased onto main; the README edits and the commit message are unchanged, and the RFC changed in two lines.
- The Phase 0 exit criteria now require answers to open questions 1 and 3, or the Phase 1 bound-write cap is split out to wait for them.
- The Phase 1 paragraph now says the bound-write cap is blocked on open questions 1 and 3.
That closes the gap raised in the previous review: the acceptance record can no longer close with the open question 3 choice unrecorded. No new findings.
Design sign-off on this RFC stays with the maintainers; this review does not approve it.
|
Round push for
Main's #23166 then added an audit sentence and an index row to Checked on |
|
HOLD at c301cd0.
The aliased-leaf item from r117 is resolved: the replacement record keeps the — crew-pr-reviewer (r122) |
|
Rebase for the README conflict: Main's #23208 added an audit sentence and an index row to crew-pr-reviewer r122 item 3 and the Design note on Of the 51 repository paths the RFC cites, main changed four more since Checked on this head: the documentation tests that read the RFC index (1,098 passed), docs lint and the other script gates, the flake-pattern and wait-discipline ratchets, and the single-commit push guard. The body's current-head paragraph and citation counts are regenerated. |
|
Rebased onto main All 51 repository paths the RFC cites still exist on |
|
Rebased onto main All 51 repository paths the RFC cites still exist on |
|
HOLD at 046c21e.
The diff is documentation only. The AI lanes report no blocking findings. Design's two notes on this head are guidance for that acceptance record and do not block. Our 2026-10-11 change request is out of date and can be dismissed. — crew-pr-reviewer (r135) |
|
Rebased onto main All 51 repository paths the RFC cites still exist on |
|
This branch conflicts with main again, most likely in |
|
@bolichen97 Rebased onto main All 51 repository paths the RFC cites still exist on |
|
Rebased onto main All 51 repository paths the RFC cites still exist on |
|
@iamwhatever could you re-review? The hole from your change request at 50ca304 looks closed at eb2a3c8. At load, an open record with no row and no pending-transition mark now becomes a retained stop that the next Open question 1 (the §2 cap against |
|
Maintainer decision: the direction is signed, including the §2 cap on a goal loop's own budget writes. For open question 1 we pick (b): the cap admits a renewal on a This is a decision on the RFC's direction, not an approval of the PR. |
|
Thanks for the decision. I'll record open question 1 = (b) in the RFC, linked to your decision, in one push after iamwhatever's re-review of Per Phase 0's exit criteria, the document still needs:
|
|
HOLD at eb2a3c8: the branch conflicts with main again, in Our r99 point is closed on this head. At load, an open record with no row and no pending-transition mark now becomes a retained stop that the next To clear:
The remaining Phase 0 items you listed (open question 3, the chat-stop record, the Phase 2 decision, the acceptance record) are exit criteria inside the RFC, not blockers for merging a draft. Design sign-off stays with the maintainers. — crew-pr-reviewer (r146) |
|
Rebased onto main The same push records bolichen97's decision on open question 1, option (b): the cap admits a renewal on a All 51 repository paths the RFC cites still exist on |
|
Checked 1708020: open question 1 now records option (b) as decided, with the same limits (three renewals, 1,000 cycles, seven days), counted in the sealed record, and the reason matches the decision. §2, §3 and the new Phase 1 exit criterion for the counted renewal say the same thing. Status stays No further change needed from me on this point. |
|
Rebased onto main All 51 repository paths the RFC cites still exist on |
…omous work Builds on rfc-goal-loop-approval-hold, whose hold (kirodotdev#16993) decides the runtime's response to an unanswered approval, and proposes what that document does not decide: - agent-facing contracts that treat a remediable blocker as work rather than a stop condition, across the base prompt, the /goal instruction, the autonudge_stop description and the bundled and repo skills; - an authority ceiling that covers the loop's own budget: a self-session bound write is capped at the pair the owner committed, save one counted renewal on a work-ledger watch, a stop the agent can trigger retains the commitment, and a re-arm or proxy at a committed slot inherits it; - a per-slot commitment record in a leaf the sandbox seals, the file-edit gate refuses and only the gateway writes; - a commitment generation and sealed provenance for scheduled automation, so a run, cron or completion launched under an ended commitment arms nothing; and - a finite 50-cycle default for the dashboard goal surfaces and the refusal of a negative cap at every boundary. It keeps rfc-permanent-patrol's decision (kirodotdev#33382, kirodotdev#33385): a loop the model arms picks no cycle cap, and a work-ledger watch takes no bound from the model. What the ceiling still bounds for such a loop is its runtime window, which a self-session stop and re-arm would otherwise restart. The bound-write cap collided with the conductor's agent-side renewal, which the approval-hold document kept. Open question 1 records the maintainer's answer, (b): the cap admits a renewal on a work-ledger watch within three renewals, 1,000 cycles and seven days, counted in the sealed record. Main has since retired that renewal (kirodotdev#33385, kirodotdev#33599), so the answer is left for the maintainer to confirm. The Phase 1 cap still waits on open question 3. Adds the document's row to the request-for-change index.
|
Rebased onto main Since the last head, main accepted rfc-permanent-patrol (#33382, implemented by #33385) and removed the One point for bolichen97: open question 1's answer, (b), admits a counted renewal on a All 52 repository paths the RFC cites exist on |
Code change approver recommendation: A person should decideWhy: the decider thinks the change is probably fine but estimates a 15% chance that this change introduces a defect, above the 12% limit for approving on its own. This PR changes only docs. It adds a draft RFC (rfc-goal-blocker-lifecycle) plus one index row and one audit sentence in the RFC README, and changes no runtime code. I checked most of the index row's runtime claims against the repo and they hold on current main. The 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 one draft RFC that builds on rfc-goal-loop-approval-hold, the goal-loop decisions that design leaves open: remediation contracts, an authority ceiling on the loop's own budget with its sealed commitment record, an owner's ending that automation cannot undo, and finite dashboard goal defaults.
rfc-goal-loop-approval-hold.md (#17033) and its implementation #16993 decided the runtime's response to an unanswered approval: the loop holds. They do not decide how agents treat remediable blockers, who may raise or renew a loop's budget, whether an owner's ending survives a workflow run or cron the loop launched, or what a fresh dashboard goal's budget is. On main the agent-facing contracts still end a goal on a generic blocker, a loop's own session can raise its bounds, stop and re-arm for a fresh budget, or replace its row through
ctx.nudge, and a fresh popover goal arms unbounded.Why it matters
The budget is the service-enforced ending left to a loop that keeps working while an approval holds nothing else back, and today the loop can renew it from its own turn or through automation it scheduled. Agents also keep retiring goals on blockers they could repair, and a dashboard goal nobody budgeted runs until a person stops it.
Not a goal
ctx.nudgeor the Issue Radar crew runtime.What changed (motivation → approach → change)
The RFC is reconciled against the merged hold: every part #17033 decided or #16993 made moot is removed, the undecided parts stand alone, and each kept runtime claim was re-read on main. The file goes from 6,120 to 904 lines.
Removed, by the heading it had:
approval_stalldisposition is unchanged and outside docs(rfc): goal loop approval hold and work-ledger bound extension #17033's scope, and the record-read of the marker depended on the removed sealed marker. Its definition seal is kept in the new §4.stopped_reason": their case was the stall stop; the owner's-ending case is one alternative now. "Spend the approval credit only on a stall attributed to a delivered cycle": decided by docs(rfc): goal loop approval hold and work-ledger bound extension #17033, which spends nothing.Kept, as the RFC's decisions:
/goalinstruction, the base prompt, theautonudge_stopdescription, babysit, kirocrew-prepare-pr, the self-nudge recipe and scaffold) and the goal-loop skill's stale approval-stall sentence are aligned to the hold.work-ledgerwatch (open question 1, answered (b)), a stop the agent can trigger retains the commitment, a re-arm or proxy at a committed slot inherits it, and an arm at a held loop is refused.autonudge-commitments.jsonunder the OS seal, the file-edit gate and a gateway-only writer, one alias-checking reader, record-over-row authority, write ordering and the owner-recommit signal./goalprovenance gate, channel provenance, and app watchdogs that honor an ended commitment.POST /api/autonudge, the popover's provenance and draft marker, and negative caps refused everywhere.New: a "Relationship to rfc-goal-loop-approval-hold" section, a "Relationship to rfc-permanent-patrol" section, and three open questions. Open question 1 records that the §2 cap refuses the
monitor_updateraise rfc-goal-popover-pause-play-controls describes, and refused the conductor'spatrol_budget.py renew, which #17033 kept until #33385 retired it; bolichen97's maintainer decision answers it with (b), so the cap admits a renewal on awork-ledgerwatch within three renewals, 1,000 cycles and seven days, counted in the sealed record. Main has since retired agent-side renewal (#33385, #33599), so the RFC asks the maintainer to confirm that (b) stands. Open question 2 asks whether the hold's fields move into the sealed record. Open question 3 asks what a self-session write may do to a slot with no readable record: be capped at the row's stored bounds, or leave the gap the Security section states; the Phase 1 cap is blocked on it. The acceptance requested here is a maintainer record that answers open question 3, or splits the Phase 1 bound-write cap out to wait for it (Phase 0).Front matter follows the directory's conventions: status
draft(no maintainer acceptance recorded),doc-pr: 13013, no implementation PRs. The README gains this RFC's index row beside rfc-goal-loop-approval-hold and one audit sentence.Backwards compatibility
Compatible: documentation only. No runtime behavior, API, stored data or user-visible surface on main changes.
Removes nothing: the diff adds one RFC and one index row to
docs/request-for-change/and deletes no file, record or documented capability on main.Tests
python scripts/docs_lint.py: all documentation checks passed.BRAND_BASE_REF=origin/main python scripts/check_brand_name.py: no misspellings in added lines.COMMENT_HISTORY_BASE_REF=origin/main python scripts/check_comment_history.py: passed.bcfdf4ee09is one commit directly on9111b0d2cc. Since2dccca3b71it changes the README audit line, where this RFC's sentence stays last on main's line, after every sentence that line carries on9111b0d2cc(main's line is kept as it is), and the aliased-leaf replacement, which now keeps thecutovermarker. No runtime code changed. The RFC is re-audited againstea45e802e9(last-audited: 2026-10-20,audited-at: ea45e802e9): every source path and code identifier it cites exists there (subject_outcomeis the field the open fix(autonudge): read a legacy loop's pull request between cycles and while held #21377 proposes), except the ones the rfc-permanent-patrol re-read below cites on9111b0d2cc. All 52 repository paths it cites exist on9111b0d2cc; main changed 48 of them since2d5bb2f86d, the base of the previous check. Every one of the RFC's 153 backticked identifiers thatsrc/,test/orwebsite/src/held on2d5bb2f86dis still there on9111b0d2cc, and each of the 31 cited on the same line as a repository path that held there still appears in that file. The six the sweep does not find on either base are names the RFC or an open PR proposes (autonudge-commitments.json,subject_outcome), three written asClass.member, and the audit shaea45e802e9. Each citation of code main moved since52a302b914names its new owner:_handle_goal_commandand the/goalinstruction (dashboard/chat_turn/slash_commands.py, refactor(dashboard): move the turn queue and eager spawn into chat_turn #21918), theautonudge_stopdescription (mcp_tools/control_parts/schemas.py, refactor: split mcp_tools control.py and spawn.py into owner packages #21575), the_monitor_updateapplier (dashboard/session_directive_apply_monitor.py, refactor(dashboard): split change_cards, chat_tags and directive apply #22138), the store's agent-writable statement (autonudge_service/store_loader.py, refactor: split history_projection, snapshot and autonudge into owners #22116), and the Slack cron turn that addsacked_itemsto the prompt (slack/gateway_runtime/cron_turn.py). The Design round ona738a0cb55is answered. A new Alternatives entry, "Hold Phase 2 until it also bounds a self-scheduled cron", weighs Phase 2's cost against the cron residual. Phase 0's exit criteria require the accepting maintainer to record that a person's chat stop request is a stop from a turn, retained under §2. Phase 1 lands as two PRs, Phase 1a first and on its own. The Design round ond4879d809dis answered too: Phase 1b, which carries that §2 rule, does not start without the record; acceptance can be recorded in one record or two, so Phase 1a can start once §1, §5 and the negative-cap refusals are accepted; and Phase 2 waits for a separate maintainer decision on its definition-digest cost against the cron residual. The index row now citesea45e802e9, the commit the audit records, and its six claims hold there and on current main. The review of16d2df0b40is answered as well: the Status line at the top now says acceptance comes in one record or two and the status flips toacceptedonce both are recorded, and it names Phase 1's two PRs; "Hold Phase 2 until it also bounds a self-scheduled cron" now calls its rejection this document's recommendation, with the decision itself the accepting maintainer's under Phase 0. The Opus finding on0ff5fd4ff7is answered too: a row written before Phase 1 now has its record sealed from its stored bounds at the first load after the upgrade, so the agent-writable row never decides a later load (Backward compatibility, the cutover exception in §3, and a new Phase 1 exit criterion). bolichen97's two points on25188f8cb1are answered as well: the cutover runs only while the sealed leaf lacks itscutovermarker, which the gateway writes in the same atomic write as the records it seals, rows or none; and the Phase 1 criterion now also covers an out-of-bandmax_runtime_secsraise and a row written after the cutover, which stays uncommitted. His follow-on on2dccca3b71, which the crew-pr-reviewer raised again on60d2aee7c1, is answered too: the fresh empty record that replaces an aliased leaf (One reader that trusts no name, and the Phase 1 test item) keeps thecutovermarker, because a writer runs only after a load that has completed the cutover, so the next load does not run the cutover again and seal whatever bounds the rows hold then. The Opus finding on5e8f9a4d7fis answered as well: the Security section's "Fail closed" bullet no longer says a self-session write cannot raise a row's bounds when the row's record is missing or aliased. It now states that gap, since such a row is uncommitted and the cap of §2 applies only against a committed pair, and a new open question 3 records the choice between capping such a write at the row's stored bounds and accepting the gap. bolichen97's re-review ofb24b20282fis answered too: Phase 0's exit criteria and the Phase 1 paragraph now name open questions 1 and 3 as what the bound-write cap waits on, so the acceptance record cannot close with that choice unrecorded. The crew-pr-reviewer's r146 HOLD oneb2a3c8539is answered as well: the README conflict is cleared, and bolichen97's decision on open question 1, (b), is recorded in the RFC: under that question, in §2's bound-write rule and its renewal paragraph, in §3's record, in Phase 0, in Phase 1 with a new exit criterion for the counted renewal, and in the Summary, Backward compatibility, Security, the schema-ceiling alternative and open question 3. Main's rfc-permanent-patrol (docs: RFC for loops the model arms picking no cycle cap #33382, implemented by feat: loops the model arms pick no cycle cap and end with one final report #33385) and chore(conductor): remove the unused patrol_budget tool and script #33599 changed the loop defaults and removed agent-side renewal after that head, so this head re-reads those statements on9111b0d2cc. That covers a new "Relationship to rfc-permanent-patrol" section, the header, the Motivation'smonitor_updateandmonitor_startbullets, and §2's dispatch patrol and renewal paragraph. It also covers the §5 rows formonitor_startand the dispatch patrol, the "Why 50" paragraph, the Phase 1 renewal criterion, Security's no-new-authority bullet, two Alternatives, and one line under open question 1's recorded answer. The design of §2 to §4 and the status are unchanged, and the rest of the audit stands atea45e802e9. Earlier rounds' changes are unchanged: theautonudge.jsonsealing alternative, the definition digest's cost, theautonudge_selfarm.pyprecedent, the dispatch patrol arm (§2, §5, Phase 1 and compatibility), the runtime budget's clock, the stop/re-arm/restore renewal fix, the self-scheduled cron residual, the time-window alternative to Phase 2, and the Phase 1a/1b split. The README index row is rebased onto main's rows on9111b0d2cc, which are kept as they are, as is main's audit line: the README differs from main only by this RFC's row and its one audit sentence, which names the audit sha. Documentation lint, comment-history, brand-name, harness-parity, diff, the flake-pattern and wait-discipline ratchets, and the single-commit push guard pass.git diff --check origin/main..HEAD: clean.push_guard.py --base main --require-single-on-base:STATUS: SAFE TO PUSH (single commit on base).Manual verification
N/A — documentation only; the factual check against main is listed under Tests.
Related Issues
Builds on #17033 and #16993.
Checklist
feat|fix|docs|style|refactor|perf|test|chore|ci|build|revert: ...)Contribution License Agreement
Contributed under the project's Apache-2.0 license; see the Licensing section of CONTRIBUTING.md.