[finding] Cloud dev sessions get auto-subscribed to their own PR with a drive-to-green posture that contradicts os-dev L2 (report-at-draft; PM owns CI convergence) #7512

Observation-class finding, filed by the domain:engine-core PM seat (session session_01MhAQCKYvakd4fx78uZXUB2) from a dev's stop-and-flag during #7477. Unassigned, deliberately not queued — triage grades it. PM-skill/protocol cards mandate claude-fable-5 per the model-tiering ruling.

What was observed

When the #7477 dev session (session_018vok76Sp1NgLNbWHYSoLtj, cloud card) created draft PR #7502, the platform auto-subscribed that session to its own PR and injected instructions carrying a drive-to-green posture — stay resident, react to CI events, do not end a CI-failure wake without a push or a blocker reply, arm hourly self check-ins until MERGED.

That posture directly contradicts the protocol the dev was dispatched under:

• os-dev L2 ([skill] pm-dispatch: 5 lessons from the domain:metadata seat's first term (8 cards MERGED / 0 rework / 4 dev-side falsifications) #6644, in .claude/agents/os-dev.md and the dispatch brief): report at draft-PR time, record gate status honestly (in_progress included), ⛔ do not idle-poll CI — the PM owns CI convergence, the ready-flip and landing.
• The dispatch brief's turn-end constraint: legitimate endpoints are branch + draft PR + report, or a hard blocker.

The dev resolved it correctly (followed L2, delivered the report, flagged the conflict in open_questions instead of silently picking a side) — this card exists so the conflict gets ruled once instead of being re-litigated inside every future cloud dispatch.

Why it matters

• The two postures cannot both be followed; whichever the dev picks looks like disobedience against the other instruction stream. An instruction conflict that every cloud card hits is a standing coin-flip.
• The drive-to-green residency is exactly the idle-poll shape [finding] os-dev agents idle-poll CI after their work is finished, burning the session limit and holding a seat's water level behind a task that is functionally done #7156 measured as a session-limit burn (closed into the [batch] PM-skill optimization: maintainer-accepted 9-point plan, consolidated with #5925 item 7 + #6644 L2 + #7055 + #7156 #7341 batch, which is also where L2 came from) — the platform message re-introduces it from outside the repo's control.
• Double ownership of one PR's CI (dev subscription + PM landing window subscription) is the two-drivers problem the L2 split exists to prevent — e.g. both sides re-queuing or both pushing fixes.

Non-goals / boundaries

• Not a defect in docs(spec,objectql): declare that after* hooks fire inside the unit of work (#7477) #7502 (all green, in queue) nor in the dev's conduct — the worked example went right.
• The subscription message is platform-injected; the repo cannot edit its text. The fix direction is deliberately NOT pre-ruled. Candidate shapes, for whoever takes this up: (a) an explicit os-dev.md / dispatch-brief clause ranking the dispatch contract above platform-injected PR-subscription postures for dispatched dev sessions (status quo made textual — the dev this time had to derive it); (b) instruct devs to unsubscribe from their own PR after the report; (c) accept the dev-side subscription and re-draw the L2 ownership line instead. Each has different failure modes on the "who owns a red gate" boundary.

Refs

• afterDelete fires inside a rollback-able transaction, so a rolled-back cascade leaves a hook that fired for a row that still exists #7477 / PR docs(spec,objectql): declare that after* hooks fire inside the unit of work (#7477) #7502 — the observation site (dev report open_questions[2], 2026-08-11 04:01Z).
• [skill] pm-dispatch: 5 lessons from the domain:metadata seat's first term (8 cards MERGED / 0 rework / 4 dev-side falsifications) #6644 (L2 ruling), [finding] os-dev agents idle-poll CI after their work is finished, burning the session limit and holding a seat's water level behind a task that is functionally done #7156 (idle-poll finding), [finding] os-dev agents systematically fail to terminate cleanly after opening their PR — 4/4 today: two silent deaths, two passive-wait stalls on nonexistent wakers #6586 (termination failures), [batch] PM-skill optimization: maintainer-accepted 9-point plan, consolidated with #5925 item 7 + #6644 L2 + #7055 + #7156 #7341 (batch that consolidated them).

Filed unassigned — recording the finding, not claiming it.