Skip to content

fix(autonudge): pause on an unanswered approval; keep open ledgers running - #16993

Merged
bolichen97 merged 1 commit into
mainfrom
fix/autonudge-approval-pause
Oct 7, 2026
Merged

bolichen97 merged 1 commit into
mainfrom
fix/autonudge-approval-pause

Conversation

@iamwhatever

@iamwhatever iamwhatever commented Oct 5, 2026 •

Copy link
Copy Markdown
Contributor

Problem / Motivation

Goal: A goal loop waits when a tool approval goes unanswered and picks up again by itself once someone is back. A conductor's work-ledger watch keeps running while its ledger still has open work.

A long patrol loop ended for good at cycle 47 of 200. One cycle asked for a tool approval while nobody was at the dashboard. The prompt timed out, and the next wake turned the loop off (stopped_reason: approval_stalled). Only a person re-arming it could bring it back. A work-ledger watch has the same problem with its own limits: when its cycle cap or time budget runs out, it stops even though the ledger still has open items.

Why it matters

Unattended patrols are meant to run overnight. Today one missed approval prompt, or one spent cap, ends the whole patrol. Nobody hears about it until someone opens the dashboard and starts it again by hand.

Not a goal

  • Changing the conductor skill or patrol_budget.py. A sibling PR owns those.
  • Keeping more stop records or gateway.log history. Another PR owns that.
  • Structured monitors (gate=False controller records). Their approval-stall handling is unchanged.
  • A new notification when a loop starts waiting. The approval-timeout card already appears in the chat.

What changed (motivation → approach → change)

Case Before After
A cycle's approval goes unanswered the loop turns off the loop waits: it stays active, runs no cycles, and uses none of its cycle cap or time budget; one notification says so
Someone comes back they must re-arm the loop the loop resumes by itself when a person answers an approval (any channel), types a message in the dashboard, or presses Play
A work-ledger watch hits its cap or budget with open items the watch stops the spent limit is raised (by a quarter, at least 10 cycles or 1 hour), logged at WARNING
A watch older than the runtime ceiling (7 days as shipped) — nothing is raised, and the limit stops it as before (logged at WARNING)

Decision record. rfc-goal-loop-approval-hold.md records the hold, the held-time credit and the work-ledger extension, with the alternatives and risks. It landed on main as its own docs PR, #17033, and this PR is rebased onto it.

Root cause. In the timer (autonudge_service/firing.py), the approval stall was a stop, just like a spent cap. So nothing ever turned the loop back on.

Approach. The timer now treats the stall as a hold. Setting the hold is written to the store first and announced only after the write lands (_record_approval_hold). It then sends one notification on the monitor channel (_notify_nudge_held in slack/gateway.py) naming what resumes it, in place of the old "loop stopped" one. (It replaces the old stop notice, which the hold removed; it is not a new kind.) It sits after the cycle cap and before the time budget, so the wall clock running during the wait cannot end the loop. When the hold ends, the waiting time is given back to the budget clock (created_ts). The reconciler skips a held loop. One release function (release_approval_hold, in autonudge_service/timers.py) writes the cleared hold to the store, waits for that write, and only then announces it and re-arms the loop. If the write fails, the hold stays as it was. It is bound on the service in src/kiro_crew/autonudge.py, next to a best-effort release_approval_hold_for helper. Its callers are: the dashboard runner's approval closer (only when no host route decided it), the Slack and Discord approval waits (slack/gateway.py, discord/renderer.py), the shared registry in messaging/approval.py, a message a person typed in the dashboard (notify_user_input(human=True), set in chat_handlers.py), and the manual fire route, which refuses the press with approval_hold_release_failed if the release cannot be written. A Slack or Discord message on its own does not release it; answering an approval there does. An agent's or app's turn does not release it.

For the ledger watch, the timer reads open items through a new has_open_items function in probes/work_ledger.py, the module that is already the store's read seam for watches. The tick takes itself off the timer table before it raises the bound, so the update cannot cancel it before the WARNING runs. It only counts real evidence, so a ledger it cannot read leaves the limit in place. The runaway backstop is the existing monitoring.max_runtime_secs ceiling.

Visible. The goal popover (website/src/components/AutoNudgePopover.tsx, typed in autoNudgeLoop.ts) shows a held loop as paused, with a new paused_approval_hold line in every website/src/i18n/locales/ catalog (en-XA regenerated): "Paused · an approval in this chat timed out. Press Play or send a message to resume." Play resumes it, and Pause still stops it. The stopped-goal Clear link stays with stopped loops. monitor_inspect reports paused_for_approval, and the dashboard websocket frame carries approval_stalled, which the record normalizer (website/src/monitoring/automation.ts) and the popover bridge (SessionAutomationPopover.tsx) now pass through. The spec docs (monitor-architecture.md, learn-cron-dashboard.md, messaging.md) describe the hold and the extension.

The cycle-cap check still runs first, so a cap reached before the hold began still ends the loop.

Backwards compatibility

Compatible: stored rows load as before. The new approval_stalled_at field defaults to 0, and a row already stopped with approval_stalled keeps that reason and can still be re-armed.

Removes nothing: the stop on an unanswered approval becomes a hold, and no field, route or key is removed.

Tests

  • test/test_autonudge_approval_stall.py: an unanswered approval holds the loop (still active, no fire, no expired, nothing armed). Waiting time is given back to the budget. A person typing releases the hold and an app send does not. A release re-arms, and the next cycle fires. The reconciler leaves a held loop alone. A hold or a release whose write fails is undone and announces nothing. The release helper is best-effort. monitor_inspect shows the hold.
  • test/test_autonudge_ledger_extension.py: a tick running as the real armed timer survives its own update and logs. A spent cap or budget with open items is raised, with the right amounts and the clamp. A finished ledger, a non-ledger loop and a user stop all stop as before. The backstop refuses past the ceiling, an unreadable ledger counts as no open items, and there is a seven-day fallback.
  • test/test_messaging_approval.py: an answer releases the hold and a timeout does not.
  • test/test_dashboard_approval_window.py: the dashboard release is gated on the host attribution and is not in the timeout branch.
  • test/test_autonudge_manual_fire.py: the fire route releases the hold before firing, and refuses with 503 when it cannot.
  • website/src/test/automationNormalizer.test.ts and SessionAutomationPopover.test.tsx: approval_stalled survives the normalizer and the bridge, so the shipped popover shows the hold.
  • website/src/test/AutoNudgePopover.test.tsx: a held loop reads paused (new line, warn box, Play resumes, Pause still stops it, no Clear).
  • test/test_monitor_notice_channel.py: the hold notification lands on the monitor channel and names what resumes it.
  • The contract manifests (test_autonudge_refactor_contract.py, test_autonudge_handlers_cov80.py) are updated for the new members and field.

Manual verification

I rendered the goal popover with a held loop and with a running loop, in an isolated capture page against the Vite dev server (screenshots below). The scoped local gates passed. Four backend tests and five frontend test files fail the same way on an unmodified base checkout on this host (sandbox and symlinked node_modules), so they are environmental. CI is the authority on the full suites.

Screenshots / video

Held for approval: the title reads Paused, the warn box says what resumes it, Play resumes it and Pause still stops it.

held

Running loop, unchanged

running

Related Issues

Closes #16976

Decision: #17033 (docs/request-for-change/rfc-goal-loop-approval-hold.md), merged.

Related: #13000 takes a different path (it keeps firing through permission blockers). This PR holds the loop instead.

Pattern harvest

Not generalizable: the fix changes one terminal decision into a hold with a release. That is a product choice in one timer, not a code pattern a rule could catch.

Checklist

  • At most two commits (one is the norm), with a Conventional Commits title
  • Existing tests pass and new tests added for new functionality
  • Self-review completed; code follows project style guidelines
  • Documentation updated (if applicable)
  • No secrets, credentials, or internal references in the diff

@iamwhatever
iamwhatever requested a review from a team as a code owner October 5, 2026 09:48
@iamwhatever
iamwhatever requested a review from CrysisDeu October 5, 2026 09:48
@github-actions github-actions Bot added the readiness: checking Automated validation is still running label Oct 5, 2026
@github-actions

github-actions Bot commented Oct 5, 2026 •

Copy link
Copy Markdown
Contributor

Design Review (Opus 5.5) — ✅ PASS

Design-level review of 76f61c707ebd3b634feadf2f9f07ec6eb6431d21 — updated in place on each push. A BLOCK verdict blocks PR readiness; PASS/CONCERNS are advisory.

Design-Verdict: PASS

Check that the 7-day ledger backstop is meant to measure active time only. Releasing a hold moves created_ts forward, so a watch that holds repeatedly can run past seven wall-clock days.

[DESIGN-REVIEWED] 76f61c7

Description read

[DESCRIPTION-READ] 6f1dc03ef1a7fba88a815f4130d22a1bff78cfb3a531ef45235d99a4a0017c33

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 .github/scripts/pr-description-capture.sh over the same inputs; if it differs, the description or that evidence changed after this verdict was formed and any finding drawn from either is unproven.

@github-actions

github-actions Bot commented Oct 5, 2026 •

Copy link
Copy Markdown
Contributor

UX Review (Opus 5.5) — 🟡 CONCERNS

UX-level review of 76f61c707ebd3b634feadf2f9f07ec6eb6431d21 — updated in place on each push. A BLOCK verdict blocks PR readiness; PASS/CONCERNS are advisory.

UX-Verdict: CONCERNS

The held-loop notice sends people looking for an approval that has already timed out, and it doesn't say whether resuming discards it.

Watch

  • The notice points at something the reader can't act on. The new status line says "Paused · an approval in this chat timed out. Press Play or send a message to resume." The blind reader understood it, then said: "I have no idea which approval it means or where to find it… it does make me want to find that approval first." They also said they were "unsure whether resuming skips the approval that timed out." The result is that someone hunts through the chat for a dead prompt before pressing Play. This happens on every hold, which is the overnight-patrol path this PR targets. Smallest fix: say what happens to the action and drop the pointer, e.g. "Paused · an approval timed out, so that action was skipped. Press Play or send a message to resume."
  • Pause looks like it does nothing on a held loop. In the held state the title already reads "Paused", yet Pause stays live and actually stops the loop. The blind reader said: "it's already paused, so I don't know what it would do here." The one way to stop a held loop from this popover therefore looks like a no-op. This matters less, because only someone who wants to end a held loop hits it. Smallest fix: when heldForApproval is true, set the Pause button's label and title to "Stop goal".

Suggestions

  • In paused_approval_hold, drop the leading "Paused ·" because the title right above it already says "Paused". The line becomes "An approval timed out, so…", and the first words carry the reason.

[UX-REVIEWED] 76f61c7

Description read

[DESCRIPTION-READ] beda21e88913d3ab7a7689063529331ce0b5cf0336da163276eac1f17fbb3fab

That sha256 names everything this verdict read: the PR description, plus the 5 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 .github/scripts/pr-description-capture.sh over the same inputs; if it differs, the description or that evidence changed after this verdict was formed and any finding drawn from either is unproven.

@github-actions

github-actions Bot commented Oct 5, 2026 •

Copy link
Copy Markdown
Contributor

First Principles Review (Opus 5.5) — ✅ PASS

Premise-level review of 76f61c707ebd3b634feadf2f9f07ec6eb6431d21 — why this exists and whether the shipped surface is the smallest honest version. Updated in place on each push. A BLOCK verdict blocks PR readiness; PASS/CONCERNS are advisory.

First-Principles-Verdict: PASS

Check that a held loop bound to a Teams or iMessage session resumes: only the dashboard, Slack, Discord and the messaging/approval.py registry call the release.

What this change ships

Inventory (10 items) — 10 justified

Intent: keep an unattended goal patrol running overnight after a missed approval or a spent bound, while its ledger still has open work. FIX.

  1. An unanswered approval now pauses a goal loop instead of ending it — justified
  2. When the pause starts, one monitor-channel notice says so; it replaces the old "loop stopped" notice — justified
  3. The loop resumes by itself when someone answers an approval, types in the dashboard, or presses Play (same 4 call sites that set the stall) — justified
  4. Time spent paused is given back to the runtime budget — justified
  5. A work-ledger watch with open items gets a higher cap or budget instead of stopping — justified
  6. The extension stops at the monitoring.max_runtime_secs ceiling, logged at WARNING — justified
  7. The goal popover shows "Paused · an approval… timed out", with a new i18n line in every locale — justified
  8. monitor_inspect reports paused_for_approval, and the websocket/REST row passes approval_stalled through — justified
  9. Play refuses with 503 approval_hold_release_failed when the release write fails — justified
  10. A new persisted approval_stalled_at field on loop rows (defaults to 0) — justified

[FIRST-PRINCIPLES-REVIEWED] 76f61c7

Description read

[DESCRIPTION-READ] 761f04c786cb45777c863761038101fc88820386a1747021ba6205788e5fda13

That sha256 names the PR description this verdict read. Recompute it with .github/scripts/pr-description-capture.sh; if it differs, the description changed after this verdict was formed and any finding drawn from the description is unproven.

@github-actions

github-actions Bot commented Oct 5, 2026 •

Copy link
Copy Markdown
Contributor

Opus 5.5 Review — ✅ no blocking findings

Reviewed 76f61c707ebd3b634feadf2f9f07ec6eb6431d21 — this comment is updated in place on each push.

Review details

No findings that block the merge.

FINDING — src/kiro_crew/dashboard/handlers/autonudge.py:351 — "paused_for_approval": bool(loop.approval_stalled) and not is_structured_monitor_loop(loop) doesn't check active, so two kinds of stopped loop read as paused_for_approval: true. One is a held loop the user then paused, because deactivating doesn't clear the flag. The other is an old row stopped with reason approval_stalled, which still has the flag set. In both cases monitor_inspect tells the agent the loop will resume on an answer, a message or a fire, but release_approval_hold refuses inactive loops and fire_now returns 409 → Fix: add and bool(loop.active), the same check the popover makes in heldForApproval.

[OPUS-REVIEWED] 76f61c7

Verdict parsed from the review's SHA-scoped output markers for commit 76f61c707ebd3b634feadf2f9f07ec6eb6431d21.

False positive or not applicable? A repository writer can comment:
/ai-review override fable 76f61c707ebd3b634feadf2f9f07ec6eb6431d21: <one-sentence reason>

@github-actions

github-actions Bot commented Oct 5, 2026 •

Copy link
Copy Markdown
Contributor

GPT 6.1 Review — ✅ no blocking findings

GPT 6.1 completed its review of 76f61c707ebd3b634feadf2f9f07ec6eb6431d21 and found no blocking issues.

This comment is updated in place on each push.

Review details

No findings.
[GPT-REVIEWED] 76f61c7

False positive or not applicable? A repository writer can comment:
/ai-review override gpt 76f61c707ebd3b634feadf2f9f07ec6eb6431d21: <one-sentence reason>

@github-actions github-actions Bot added readiness: action required A blocking check or review needs attention and removed readiness: checking Automated validation is still running labels Oct 5, 2026
@iamwhatever

Copy link
Copy Markdown
Contributor Author
  • Commit the approval release before publishing or re-arming span=dc6b1620be5a -- fixed

Fixed in b9ea065: release_approval_hold is now async. It clears the hold and moves created_ts under _lock, awaits _write_monitor_snapshot_locked, and restores all three fields if the write raises.
Only after the write lands does it log, emit updated, re-arm and return True. Sync callers schedule it as a supervised task; the fire route awaits it.
Test: test_a_failed_release_write_keeps_the_hold_and_publishes_nothing.

@iamwhatever

Copy link
Copy Markdown
Contributor Author
  • Scheduled bound extension's update cancels the calling tick span=848c51dca410 -- fixed

Fixed in b9ea065: the tick removes itself from _timers before update, so update's _cancel_timer finds nothing to cancel and arms its own timer; the WARNING and re-arm then run.
Test: test_an_armed_tick_survives_its_own_update_and_logs runs the tick as the registered timer task; it fails with the detach removed.

@iamwhatever

Copy link
Copy Markdown
Contributor Author
  • Typing on Slack or Discord does not release the hold span=8650472cf0e3 -- fixed

Fixed in b9ea065 by taking the documentation branch of Clears-when: the PR body, learn-cron-dashboard.md and monitor-architecture.md now say a typed message releases the hold only in the dashboard.
On Slack and Discord an answered approval releases it (both approval waits call release_approval_hold_for). Wiring every channel's inbound message path is new surface beyond this fix.

@iamwhatever

Copy link
Copy Markdown
Contributor Author
  • _extend_for_open_ledger overrides limits the user set span=ff1bfcdc6a7a -- needs-a-decision

Raising the cap of a work-ledger watch that still has open items is the requested behaviour: a conductor patrol should end only when its ledger is finished or someone stops it, with a hard runaway ceiling.
It applies only to watch: work-ledger loops, each raise is logged at WARNING and visible as the new max_cycles/max_runtime_secs on the record, and the configured runtime ceiling bounds it.
Whether an explicitly typed cap should be exempt is the same product question First Principles raises; it is put to the maintainer with that lane.

@iamwhatever

Copy link
Copy Markdown
Contributor Author
  • heldForApproval changes the title but controls follow loop.active span=4e764e4367f2 -- fixed

Fixed in b9ea065: the controls now follow runsNow = active && !heldForApproval, so a held loop gets the paused control set.

@iamwhatever

Copy link
Copy Markdown
Contributor Author
  • Pause stays enabled and the glyph stays Zap span=4d8250d09e2c -- fixed

Fixed in b9ea065: Pause is disabled, the glyph is Play and the name is resume_loop while held; Clear is offered as for any paused loop.

@iamwhatever

Copy link
Copy Markdown
Contributor Author
  • No resume or play button on a held loop span=602397f6e115 -- fixed

Fixed in b9ea065: a held loop shows the Play button named Resume loop and nudge now; pressing it fires, and the fire route releases the hold first. New screenshot in the body.

@iamwhatever

Copy link
Copy Markdown
Contributor Author
  • The user cannot recover from inside the panel span=a4542bcd8a88 -- fixed

Fixed in b9ea065: Resume is live on a held loop (its overdue deadline no longer disables it), so the panel itself resumes the loop.

@iamwhatever

Copy link
Copy Markdown
Contributor Author
  • Render the paused control set when held span=b4ce0f411563 -- fixed

Fixed in b9ea065 as suggested; pinned by the popover test for a held loop (Pause off, Play resumes, Clear offered).

@iamwhatever

Copy link
Copy Markdown
Contributor Author
  • The status line names an approval but not where it is span=7a8c7881b53a -- rebutted

Kept the existing translated string on purpose: the timed-out prompt is already gone, so there is no approval left to point to, and the chat already shows the approval-timeout card naming the tool and the window.
With the paused control set the next step is now on screen (Resume). New wording would need new keys in all 13 catalogs; that copy change is better made with the maintainer's ruling on the hold, which may reshape this surface.

@iamwhatever

Copy link
Copy Markdown
Contributor Author
  • Item 1: approval hold is a product-shape change without a recorded decision span=f98ccd61aab1 -- needs-a-decision

Correct that no RFC records this. The behaviour was requested by the operator after a 200-cycle conductor patrol stopped at cycle 47 on one unanswered prompt.
Maintainer question: accept the hold (pause, auto-resume on a person's return) as the new default, or should it ship behind an RFC first?

@iamwhatever

Copy link
Copy Markdown
Contributor Author
  • Item 2: held time is handed back to the runtime budget span=8e1e2e304ce5 -- needs-a-decision

This is the bookkeeping the in-progress RFC rules out. Without it a hold longer than the budget would end the loop on release, which defeats the hold.
Maintainer question: keep the hand-back, or keep the budget a pure wall-clock bound and accept that long holds expire?

@iamwhatever

Copy link
Copy Markdown
Contributor Author
  • Item 3: max_cycles raised automatically for an open ledger span=1540e292cb1e -- needs-a-decision

Scoped to work-ledger watches with a non-terminal item, logged at WARNING, bounded by the runtime ceiling.
Maintainer question: is raising a typed cap acceptable for this one loop kind, or should only an untyped default be raised?

@iamwhatever

Copy link
Copy Markdown
Contributor Author
  • Item 7: has_open_items described as an existing seam span=2ebc6c18d9f9 -- fixed

Fixed in the body: it now says has_open_items is a new function in probes/work_ledger.py, the module that is already the store's read seam for watches.

@iamwhatever

Copy link
Copy Markdown
Contributor Author
  • Product-shape change without accepted RFC span=1712f4955dba -- needs-a-decision

Items 1 to 3 are one product decision and need a maintainer: either an RFC lands first, or a maintainer records the decision with /ai-review override first-principles <head>: <decision>.
The code and tests are otherwise ready; no code change can clear this item.

@iamwhatever
iamwhatever force-pushed the fix/autonudge-approval-pause branch from 53a3255 to b9ea065 Compare October 5, 2026 11:33
@github-actions github-actions Bot added readiness: checking Automated validation is still running readiness: action required A blocking check or review needs attention and removed readiness: action required A blocking check or review needs attention readiness: checking Automated validation is still running labels Oct 5, 2026
rubencu added a commit to rubencu/KiroCrew that referenced this pull request Oct 10, 2026
…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, 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.

The bound-write cap collides with the conductor's agent-side renewal,
which the approval-hold document keeps; that is open question 1 and the
Phase 1 cap is blocked on it.

Adds the document's row to the request-for-change index.
@rubencu rubencu mentioned this pull request Oct 10, 2026
5 tasks done
rubencu added a commit to rubencu/KiroCrew that referenced this pull request Oct 10, 2026
…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, 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.

The bound-write cap collides with the conductor's agent-side renewal,
which the approval-hold document keeps; that is open question 1 and the
Phase 1 cap is blocked on it.

Adds the document's row to the request-for-change index.
rubencu added a commit to rubencu/KiroCrew that referenced this pull request Oct 11, 2026
…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, 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.

The bound-write cap collides with the conductor's agent-side renewal,
which the approval-hold document keeps; that is open question 1 and the
Phase 1 cap is blocked on it.

Adds the document's row to the request-for-change index.
rubencu added a commit to rubencu/KiroCrew that referenced this pull request Oct 11, 2026
…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, 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.

The bound-write cap collides with the conductor's agent-side renewal,
which the approval-hold document keeps; that is open question 1 and the
Phase 1 cap is blocked on it.

Adds the document's row to the request-for-change index.
rubencu added a commit to rubencu/KiroCrew that referenced this pull request Oct 11, 2026
…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, 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.

The bound-write cap collides with the conductor's agent-side renewal,
which the approval-hold document keeps; that is open question 1 and the
Phase 1 cap is blocked on it.

Adds the document's row to the request-for-change index.
rubencu added a commit to rubencu/KiroCrew that referenced this pull request Oct 11, 2026
…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, 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.

The bound-write cap collides with the conductor's agent-side renewal,
which the approval-hold document keeps; that is open question 1 and the
Phase 1 cap is blocked on it.

Adds the document's row to the request-for-change index.
rubencu added a commit to rubencu/KiroCrew that referenced this pull request Oct 11, 2026
…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, 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.

The bound-write cap collides with the conductor's agent-side renewal,
which the approval-hold document keeps; that is open question 1 and the
Phase 1 cap is blocked on it.

Adds the document's row to the request-for-change index.
rubencu added a commit to rubencu/KiroCrew that referenced this pull request Oct 11, 2026
…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, 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.

The bound-write cap collides with the conductor's agent-side renewal,
which the approval-hold document keeps; that is open question 1 and the
Phase 1 cap is blocked on it.

Adds the document's row to the request-for-change index.
rubencu added a commit to rubencu/KiroCrew that referenced this pull request Oct 12, 2026
…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, 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.

The bound-write cap collides with the conductor's agent-side renewal,
which the approval-hold document keeps; that is open question 1 and the
Phase 1 cap is blocked on it.

Adds the document's row to the request-for-change index.
rubencu added a commit to rubencu/KiroCrew that referenced this pull request Oct 12, 2026
…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, 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.

The bound-write cap collides with the conductor's agent-side renewal,
which the approval-hold document keeps; that is open question 1 and the
Phase 1 cap is blocked on it.

Adds the document's row to the request-for-change index.
rubencu added a commit to rubencu/KiroCrew that referenced this pull request Oct 12, 2026
…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, 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.

The bound-write cap collides with the conductor's agent-side renewal,
which the approval-hold document keeps; that is open question 1 and the
Phase 1 cap is blocked on it.

Adds the document's row to the request-for-change index.
rubencu added a commit to rubencu/KiroCrew that referenced this pull request Oct 13, 2026
…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, 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.

The bound-write cap collides with the conductor's agent-side renewal,
which the approval-hold document keeps; that is open question 1 and the
Phase 1 cap is blocked on it.

Adds the document's row to the request-for-change index.
rubencu added a commit to rubencu/KiroCrew that referenced this pull request Oct 13, 2026
…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, 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.

The bound-write cap collides with the conductor's agent-side renewal,
which the approval-hold document keeps; that is open question 1 and the
Phase 1 cap is blocked on it.

Adds the document's row to the request-for-change index.
rubencu added a commit to rubencu/KiroCrew that referenced this pull request Oct 13, 2026
…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, 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.

The bound-write cap collides with the conductor's agent-side renewal,
which the approval-hold document keeps; that is open question 1 and the
Phase 1 cap is blocked on it.

Adds the document's row to the request-for-change index.
rubencu added a commit to rubencu/KiroCrew that referenced this pull request Oct 14, 2026
…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, 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.

The bound-write cap collides with the conductor's agent-side renewal,
which the approval-hold document keeps; that is open question 1 and the
Phase 1 cap is blocked on it.

Adds the document's row to the request-for-change index.
rubencu added a commit to rubencu/KiroCrew that referenced this pull request Oct 14, 2026
…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, 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.

The bound-write cap collides with the conductor's agent-side renewal,
which the approval-hold document keeps; that is open question 1 and the
Phase 1 cap is blocked on it.

Adds the document's row to the request-for-change index.
rubencu added a commit to rubencu/KiroCrew that referenced this pull request Oct 16, 2026
…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, 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.

The bound-write cap collides with the conductor's agent-side renewal,
which the approval-hold document keeps; that is open question 1 and the
Phase 1 cap is blocked on it.

Adds the document's row to the request-for-change index.
rubencu added a commit to rubencu/KiroCrew that referenced this pull request Oct 17, 2026
…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, 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.

The bound-write cap collides with the conductor's agent-side renewal,
which the approval-hold document keeps; that is open question 1 and the
Phase 1 cap is blocked on it.

Adds the document's row to the request-for-change index.
rubencu added a commit to rubencu/KiroCrew that referenced this pull request Oct 17, 2026
…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, 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.

The bound-write cap collides with the conductor's agent-side renewal,
which the approval-hold document keeps; that is open question 1 and the
Phase 1 cap is blocked on it.

Adds the document's row to the request-for-change index.
rubencu added a commit to rubencu/KiroCrew that referenced this pull request Oct 17, 2026
…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, 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.

The bound-write cap collides with the conductor's agent-side renewal,
which the approval-hold document keeps; that is open question 1 and the
Phase 1 cap is blocked on it.

Adds the document's row to the request-for-change index.
rubencu added a commit to rubencu/KiroCrew that referenced this pull request Oct 19, 2026
…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, 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.

The bound-write cap collides with the conductor's agent-side renewal,
which the approval-hold document keeps; that is open question 1 and the
Phase 1 cap is blocked on it.

Adds the document's row to the request-for-change index.
rubencu added a commit to rubencu/KiroCrew that referenced this pull request Oct 19, 2026
…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, 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.

The bound-write cap collides with the conductor's agent-side renewal,
which the approval-hold document keeps; that is open question 1 and the
Phase 1 cap is blocked on it.

Adds the document's row to the request-for-change index.
rubencu added a commit to rubencu/KiroCrew that referenced this pull request Oct 20, 2026
…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, 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.

The bound-write cap collides with the conductor's agent-side renewal,
which the approval-hold document keeps; that is open question 1 and the
Phase 1 cap is blocked on it.

Adds the document's row to the request-for-change index.
rubencu added a commit to rubencu/KiroCrew that referenced this pull request Oct 20, 2026
…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, 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.

The bound-write cap collides with the conductor's agent-side renewal,
which the approval-hold document keeps; that is open question 1 and the
Phase 1 cap is blocked on it.

Adds the document's row to the request-for-change index.
rubencu added a commit to rubencu/KiroCrew that referenced this pull request Oct 20, 2026
…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, 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.

The bound-write cap collides with the conductor's agent-side renewal,
which the approval-hold document keeps; that is open question 1 and the
Phase 1 cap is blocked on it.

Adds the document's row to the request-for-change index.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

Goal loop stops for good when one tool approval goes unanswered

2 participants