Repository navigation
docs(rfc): drop the fixed goal concurrency default; host pressure wakes the conductor - #36304
Conversation
First Principles Review (Opus 5.5) — ⏭️ skippedRevision |
UX Review (Opus 5.5) — ⏭️ skippedRevision |
Design Review (Opus 5.5) — ✅ PASSDesign-level review of Design-Verdict: PASS Docs-only RFC changes: the decision is recorded with its alternatives, its compatibility story (stored 6s are kept) and its landing order (pressure first), and nothing ships yet. Notes
[DESIGN-REVIEWED] 7d35767 Description read[DESCRIPTION-READ] d7151b0db9ca5a3a421a0b0fe48ab56902303f9f6076487464bc39c66989aada 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 Review detailsNo findings. Notes:
[OPUS-REVIEWED] 7d35767 Verdict parsed from the review's SHA-scoped output markers for commit False positive or not applicable? A repository writer can comment: |
7b8b284 to
4ce398e
Compare
Code change approver recommendation: ApproveWhy: the decider estimates an 8% chance that this change introduces a defect, within the 12% limit for approving on its own, and every check passes. Docs-only PR: amends the concurrency-cap RFC (§7, default becomes 32, status partial) and adds a resource-pressure RFC. Every code claim checked holds at head: DEFAULT_GOAL_CONCURRENCY_CAP=6 and MAX_ITEMS_PER_CONDUCTOR=32; _admit_create_locked checks the store bound first, so a 32 default never fires; to_dict always stores concurrency_cap and absent or out-of-range values read as the default; the existing refusal text contradicts the 'never ask to raise it' prompt; the notifier constants and episode model match; _fire, _watched_boards, _ALL_CLOSED_ENTRY, admission_check and cached_admission_check exist with the stated fail-open behaviour. Trust: the planned 409 reads the live cached verdict, not the new file, so a stale or forged host_pressure.json can only cause an extra pause. The record expires after 90 s, the RFC adds no new authority, and it requires the pressure code to land before or with the default change. Nothing blocks. 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 |
|
Not from this PR (docs only). Two reds on head 4ce398e:
|
4ce398e to
7d35767
Compare
bolichen97
left a comment
There was a problem hiding this comment.
Approve. I read the whole diff on this head. It is docs only: the README rows, a §7 amendment to rfc-work-ledger-concurrency-cap.md, and the new rfc-work-ledger-resource-pressure.md. All checks are green.
What it decides
- The concurrency-cap amendment (§7) drops the fixed default of 6 open items per goal. The default becomes
MAX_ITEMS_PER_CONDUCTOR(32), andconcurrency_capstays as a bound the user can set. The refusal text says to wait for a seat instead of asking to raise the cap.resource_statusdecides each wave, in both the Ultra Mode and Bot conductor rules. - The pressure RFC has the gateway's existing
ResourcePressureNotifierwrite a hosthost_pressurerecord that expires after 90 s, and show it inwork_ledger_read. It wakes each watched conductor once per level change, through the existing_fireand probe budgets. Attightthe conductor pauses dispatch, which is advice only. Atcritical,createis refused with 409host_pressure_critical.
Why it is fine
- Dropping the 6 loosens a default nobody measured. A board that already stored 6 keeps it, and no migration rewrites stored values.
- The
criticalrefusal is not new strictness of its own. It reusescached_admission_check, which already refuses background work (cron, task-runner steps, helper turns) at critical on main. It admits onunknownand whenagent.admission_gateis false, and writes nothing. This applies main's own rule to work-ledgercreate, a background path that skipped it. - The pressure record is the safety net that §7 needs once the fixed 6 is gone, and §10 asks for it to land first, or in the same PR.
- No new tool, status, event kind or config key. Cost is about one conductor turn per conductor per level change.
Long-term debt: low. The known limits are written down: memory only, one host, and advice only at tight.
Direction signed by maintainer (bolichen97): both RFCs as written. The code PRs still need their own review, with the §5 tests.
Problem / Motivation
Goal: Record the owner's decision that host resources, not a fixed number, bound how many workers a goal runs, and that host pressure is written on the work ledger and wakes the conductor so it pauses.
Every goal board is capped at 6 open items (
DEFAULT_GOAL_CONCURRENCY_CAPinsrc/kiro_crew/work_ledger_records.py). The 6 was never measured and does not follow the host. When host memory gets tight, nothing tells a conductor:ResourcePressureNotifieronly notifies the person, and a work-ledgercreatedoes not consultadmission_check.Why it matters
A fixed 6 holds back a healthy host and does nothing for a starved one. A conductor that never hears about pressure keeps dispatching into a host about to freeze. Both code changes change conductor defaults, so the First Principles lane needs this decision on main first.
Not a goal
concurrency_capfield.What changed (motivation → approach → change)
The owner asked to stop micro-managing with a fixed number, and to have tight resources recorded on the ledger and wake the conductor. Two linked RFCs record that, each saying the other is needed.
rfc-work-ledger-concurrency-cap.md§7 (amended)concurrency_capbecomesMAX_ITEMS_PER_CONDUCTOR(32). The field stays for a bound the user sets. The refusal says wait for a seat.resource_statusdecides each wave, in Ultra Mode and Bot conductor rules. Reverses the earlier owner ruling that kept 6.rfc-work-ledger-resource-pressure.md(new)host_pressurerecord.work_ledger_readshows it. Each watched conductor wakes once per episode level. Tight: stop dispatching (advice). Critical:createrefused 409host_pressure_critical. Clear: resume.The concurrency-cap RFC moves to
partial: §3 shipped in #19338, §7 is not built. The RFC index (docs/request-for-change/README.md) carries both rows.Backwards compatibility
Compatible: documents only; no code, config or schema changes.
Removes nothing: the diff adds and amends RFC text; the 2026-10-12 decision text stays as the record.
Tests
Proof: not provable -- docs or config only
test/test_pr_quality_gates.py(the one related backend test file) passes locally: 97 passed. The pressure RFC lists the tests the code PR must carry.Manual verification
Code references re-read at
9b32941be6(origin/main): the cap check in_admit_create_locked, the notifier's 30 s / 600 s / 30 min constants,_fire,_watched_boards,_ALL_CLOSED_ENTRY,admission_check, and thatConductorRecord.to_dictalways storesconcurrency_cap(so boards that stored 6 keep it).Related Issues
Closes #36301
Checklist
feat|fix|docs|style|refactor|perf|test|chore|ci|build|revert: ...)