qa(checklist-test §2): every QA runner dispatch states its effort by card shape — high for a run, xhigh only for a security fix or adversarial review — and a cloud card records the effort it inherits #27390

domain:devx#1 · sunhaolin · session_017xRBcVy4VXmkQjn2MGeaWk

Filing gate: ③ a maintainer-directed task.

• The maintainer, in session session_017xRBcVy4VXmkQjn2MGeaWk on 2026-10-29, asked about the six wave-3 runner cards of [sweep] v18.3.0 release test sweep #27377: 「你派的任务都需要 extra 吗」 ("do all the tasks you dispatched need extra-high effort?").
• The seat answered that they do not. A run card needs high, and xhigh belongs to a security fix dispatch and to adversarial review.
• The maintainer then said 「立卡更新 skills」.

Business reason: the wave-3 runner cards were dispatched with no effort stated. create_session takes no effort parameter. The dispatch-runbook (.claude/skills/pm-dispatch/references/dispatch-runbook.md, the trigger-flow line ②) says a trigger-flow session inherits the dispatching seat's tier. So each card ran at whatever effort the dispatching session had, chosen by no one.

Cost of that, measured on #27377's cards at about 30 minutes in:

• W3-C1 had spent $33.04, mostly on environment setup and spot-checks.
• W3-C3b had spent $14.98.

A run card's work does not need the top effort:

• per-area vitest batches;
• REST/CLI steps scored against server truth;
• browser steps scored against RUNNER's oracle ladder.

execution-duties.md already requires effort to be set by card difficulty and passed explicitly, ⛔ never left empty. checklist-test is the one dispatcher that does not apply this. Its §2 names the model tier, through dispatch-gates --tier, and says nothing about effort.

Reader: every session that dispatches checklist-test runners, whether as subagents or as cloud cards. Today that is the release-wave PM sessions.

Dedupe (open AND closed): "checklist-test runner effort level high xhigh cloud session dispatch" returns 45 hits, and none of them sets runner effort. #27383 (this session's periodic-testing card) touches the same SKILL.md, in §5. This card is §2 only, so the two should land serially or be merged at claim time.

What changes in .claude/skills/checklist-test/SKILL.md §2

1. Effort by card shape, stated in every runner dispatch:
   • high for a checklist run card: Tier 1 pin batches, manual API/CLI sets, manual browser sets, the P0 smoke, and a UX walkthrough;
   • xhigh only for the os-dev dispatched to fix a security finding, and for an adversarial re-verification of a contested verdict;
   • ⛔ never left empty.
2. The two backends:
   • A runner subagent spawned with the Agent tool gets effort passed explicitly. This skill line is the instruction that makes passing it lawful.
   • A cloud runner card (create_session) cannot set effort and inherits the dispatching session's effort. The dispatcher records the inherited level in the anchor's dispatch note. When that level is above the card's shape, the dispatcher starts the cards from a session at the intended level. Measure how a cloud session's effort is actually set and inherited before writing this, and cite the runbook line rather than restating it.
3. The spend line: the anchor's dispatch note and each runner's final report carry each card's spend at report time, so the next wave can calibrate. get_session exposes usage.cost_usd.

Out of scope: changing effort on sessions already running; the model tier rule, which stays with dispatch-gates --tier; pm-dispatch's own effort rules.

Gates to expect

• The pm-skill line ratchet, if it covers this file. qa(checklist-test): writes through the pm-dispatch doors, every defect filed in the run, auth/authz fixes dispatched at once #26567 had to touch it. A net increase cites this ruling, and the seat posts it on [sweep] v18.3.0 release test sweep #27377 as a comment id.
• skip-changeset.

Tier: .claude/** is governed Tier S. Clause-②: no. Metadata-protocol: none.

Generated by Claude Code