skills(triage): drop "run full suite" — defer to PR CI per "ship before end_turn"

[tend-agent]

Problem

tend-triage's Step 6 "If fixing" tells the bot to "Run the full test suite and lints" before commit/push. That directly conflicts with running-in-ci's "End the turn only when work is shipped" (added in #661), which forbids backgrounding pre-push work the deliverable depends on.

Today on PRQL/prql, run 27149177431 (tend-triage on issue #5988, "Compiler crash when joining a zero-row relation"):

• Bot reproduced the panic, applied the fix, added a snapshot test, ran it green — all locally.
• Following Step 6.3 ("Run the full test suite and lints"), kicked off task prqlc:pull-request (~7 min suite) with run_in_background: true.
• Used Monitor (timeout_ms: 420000) and ScheduleWakeup (delaySeconds: 180) to wait for the suite, then emitted end_turn.
• Session terminated — no commit, no push, no PR, no comment on the issue. 14 minutes of agent time on a real user bug, zero maintainer-visible output.

Last assistant text: "I've scheduled a fallback check. Waiting for the test suite to finish."

This is the second occurrence of "agent calls ScheduleWakeup/Monitor in CI expecting resumption and the deliverable never ships" — the first was tend-mention run 26347739838 (May 24, fix attempt closed as #593, gate-tightening followed in #602, and the root-cause guidance shipped in #661 — but the triage skill never picked up the alignment). PRQL is still pinned to max-sixty/tend@0.1.2, predating #661, so the conflicting Step 6.3 won. Even once consumers bump past #661, an agent reading both skills will see Step 6.3's explicit instruction and a generic principle in running-in-ci; the explicit instruction tends to win.

Solution

Rewrite triage Step 6.3 so the two skills agree: the targeted reproduction test plus a clean compile is sufficient confidence; the comprehensive suite is PR CI's job. Renumber the following two steps (4 → 3, 5 → 4) since "Confirm the test passes" merges naturally into Step 2.

Gate assessment

• Evidence level: High — 2 occurrences of "agent uses async-resume tool in CI then end_turn leaves deliverable unshipped" across tend-mention and tend-triage. Both fully diagnosed from session logs.
• Structural vs stochastic: Structural. Two skills' instructions conflict on whether to run the comprehensive suite before push; the more specific (triage Step 6.3) wins. Same conditions reproduce reliably.
• Change type: Removal / simplification — drop the conflicting "Run the full test suite and lints" line, fold "Confirm the test passes" into the same step, renumber. Per Gate 2, Low evidence bar (1 occurrence enough).
• Both gates: Pass.

Evidence gist (full per-run history): https://gist.github.com/a017dd295de05dc44af4479748ffe636

Sources and notes

• Today's triage session: PRQL/prql Run 27149177431, issue PRQL/prql#5988 (reporter kgutwin, non-maintainer).
• Session JSONL tool sequence: Bash(task prqlc:pull-request, timeout 420000) → Bash(until... done, run_in_background: true, timeout 420000) → Monitor(timeout_ms: 420000, persistent: false) → Write(/tmp/pr-body.md) → Read(.../tasks/...output) (×3) → ScheduleWakeup(delaySeconds: 180, prompt: "Check the prqlc:pull-request test output ... and continue triage of issue 5988"). Stop reason: end_turn. No git commit / git push / gh pr create / gh issue comment in the session.
• Prior tend-mention occurrence: PRQL/prql Run 26347739838, terminated at 6h cap; documented in the 2026-05 evidence gist under "Acted on (High evidence, structural): tend-mention 6h-cap idle after ScheduleWakeup in CI".
• Earlier closed PR: skills(running-in-ci): forbid ScheduleWakeup and fire-and-forget background bash in CI #593 (forbid ScheduleWakeup and fire-and-forget background bash in CI) — closed at 1 occurrence pre-gate-tightening. Comment: "no! we need to change our criteria; this is structural and it's not critical". Gates were tightened in skills(review-gates): structural classification doesn't override Gate 1 #602, then skills(running-in-ci): end the turn only when work is shipped #661 added the more nuanced framing — this PR completes the picture by removing the triage-side conflict.
• running-in-ci's rule (post-skills(running-in-ci): end the turn only when work is shipped #661): "A targeted compile plus the tests directly exercising the change is enough local confidence to ship — leave the comprehensive matrix to CI."

[tend-agent]

Same pattern elsewhere: plugins/tend-ci-runner/skills/nightly/SKILL.md line 281 still says branch, fix, run full test suite, commit, push, create PR, poll CI. By this PR's own argument — an explicit step-list instruction tends to override the generic running-in-ci rule — nightly remains vulnerable to the same mid-wait-end_turn failure mode, just one of the project's longer suites away from reproducing it. Happy to push a follow-up that drops run full test suite from that line; say the word.