Skip to content

ci: a weekly merge-queue duration reading against a 22-minute median budget and a 32-minute p90 alarm, filing a speed-up card when over #35322

Description

@os-warren

review-split thread · os-warren · session_01JQwxbYzM6act7DgRZuPrxu

Source: maintainer — Warren, 2026-11-19, review-split discussion thread: 03:16Z 「以上都同意」 (CI speed budget for objectstack; 「变慢两成非常大了」 rejected a 20% allowance); 04:24Z 「1+2+4」, point 2 as amended by the review: budget = median 22 min, alarm = p90 32 min, measured weekly by a script that files a speed-up card when over; a PR's own before/after reading is advisory and never blocks a merge.

Filing gate: ③ a task the maintainer dispatched. Child of epic #35205.

Grade: p2 · area:fleet · Task · fleet machinery, no product surface

维护者速读

  • 改了什么:每周自动量一次合并队列耗时(中位、九成、失败率),超过 22 分钟中位或 32 分钟九成就自动开一张提速卡。
  • 为什么:你 11-19 定的 CI 速度预算;现在没人盯趋势,变慢没人知道。
  • 最坏:量错会多开或漏开提速卡;不挡任何合并,可随时撤。
  • 需要你做的:无(改 workflow 与 AGENTS.md,到时请你批)。

Baseline

merge_group runs 2026-11-17..19: 176 all-green head SHAs, first start to last end over the queue workflows (waiting for a runner included), p50 21.7 min, p90 31.9 min, max 34.5 min.

What changes

  1. scripts/ci/queue-duration.mjs (new, read-only): reads the last 7 days of merge_group runs of the workflows that run on the queue, groups them by head SHA, and for all-green SHAs measures first start to last end. Prints p50, p90, max, sample size, the queue failure rate (SHAs with any red run), and the five slowest jobs on the critical path. Takes --repo <owner/name> and --since; --json for the workflow. --self-test per AGENTS.md "Writing a --self-test" (battery names, floor, handshake), offline on fixtures.
  2. scripts/ci/queue-budget.json (new): budgetMedianMinutes: 22, alarmP90Minutes: 32, the baseline above with its window, and the list of queue workflow names. Changing a number is the maintainer's call (say so in the file and in the script header).
  3. A weekly scheduled job (a job in fleet-patrols.yml beside governed-merges-audit, or its own workflow if the patrol pins make that cleaner): runs the script for objectstack; when the 7-day median exceeds the budget or the p90 exceeds the alarm, files or refreshes ONE deduplicated card (fixed title prefix + plain-text body marker, the scheduled-full-run-card pattern; labels on create only: tooling, area:fleet, domain:devx, priority:p2, pm:queue, issue type Task). Writes nothing on pull_request. Also prints hotcrm's queue reading (--repo objectstack-ai/hotcrm, public) to the run summary, report-only, no card.
  4. AGENTS.md, Build & Test: one sentence that the merge queue has a speed budget held in scripts/ci/queue-budget.json, read weekly, over budget files a speed-up card; a PR that changes CI or adds a gate reports the affected job's before/after median as advisory.

Done when

  • node scripts/ci/queue-duration.mjs --self-test green; a live dry run on the last 7 days prints a reading close to the baseline (cite it with the commit it was taken on).
  • The scheduled job runs on workflow_dispatch without writing on a PR event; the de-dup holds (second run refreshes, never files twice).

Activity

  1. self-assigned this
    on Nov 19, 2026
  2. added
    toolingFix lands in tooling, not a product package — enters pm:queue only with Unblocks: or a named surface
    on Nov 19, 2026
  3. added theissue type on Nov 19, 2026
  4. added
    pm:dispatchedDispatched to a dev agent by /pm-dispatch
    domain:devxDomain lane — seat card indexed by label:pm:seat
    on Nov 19, 2026
  5. os-warren commented on Nov 19, 2026

    @os-warren
    CollaboratorAuthor

    Claim: review-split order 2026-11-19
    Session: session_01JQwxbYzM6act7DgRZuPrxu
    Account: os-warren
    Branch: claude/issue-35322-queue-duration-budget
    Worktree: objectstack-issue-35322
    Domain: domain:devx
    File surface: scripts/ci/queue-duration.mjs (new), scripts/ci/queue-budget.json (new), .github/workflows/fleet-patrols.yml or one new workflow, the patrol pin script if it must learn the new job, AGENTS.md (one sentence in Build & Test)
    Container & model: M, mode:subagent, effort: high
    Clause-②: no
    Metadata-protocol: none
    Responsibility: n/a — not a defect card
    Thread-read: none
    Premise: holds (baseline measured 2026-11-19 over merge_group 11-17..19)
    Serial constraints cleared: none (AGENTS.md also edited by #35321 in a different section)


    Generated by Claude Code

  6. objectstack-fleet commented on Nov 19, 2026

    @objectstack-fleet
    Contributor

    triage#1 · hotlong · session_01Q4EjjMnyLqJdzuktEmRPXH

    Release: #35322 · holder os-warren is out of tokens (seat-offline, brief 7005006297 on #32588) · back to pm:queue for an online sibling seat
    Resume-from: claude/issue-35322-queue-duration-budget@e1626c63 (PR #35343)

    2026-11-19T09:26Z · No fleet-lane seat is online; it waits in the queue. The claimant picks up PR #35343 where it stands.

  7. added
    pm:queueReady for the PM dispatch loop
    and removed
    pm:dispatchedDispatched to a dev agent by /pm-dispatch
    on Nov 19, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    area:fleetdomain:devxDomain lane — seat card indexed by label:pm:seatpm:queueReady for the PM dispatch looppriority:p2Medium: important, M3toolingFix lands in tooling, not a product package — enters pm:queue only with Unblocks: or a named surface

    Type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions