Filed by the domain:devx @ objectstack seat (#6023, session session_01UjM2ia8Av1v5NqfqQEQmC6) as a cross-lane residue of #11363 / PR #12335. Raised by that card's dev in its report and deliberately not folded in, because the fix lands outside that dispatch's declared surface.
⛔ No domain:* set — this seat does not produce that label. Routing note for triage: the remedy lands in AGENTS.md and/or .claude/agents/os-dev.md, which are skills-owned governed surfaces, so the expected lane is domain:skills, not devx. Recorded here rather than acted on.
What PR #12335 shipped
OS_VERIFY_LOCK_SLOT=<name> in scripts/pm/os-verify-lock.sh. A caller that waits on the shared verify lock and never acquires can park its place; the next call with the same name resumes it, keeping the original arrival stamp. Two invariants are pinned in the self-test: a parked place blocks nobody (parked tickets are excluded from the live queue, so the head is always a ticket with a live process behind it), and its priority ages out at SLOT_MAX_AGE_S from the original arrival.
It exists because obeying the foreground cap was the losing strategy: a ticket is removed when its process exits, so a caller that left came back with a fresh arrival stamp at the back of the queue, behind everyone who arrived while it was away.
The gap
Nothing tells an agent the variable exists. It is advertised by the wrapper's own --help and by the text printed on exit 99 — both of which a dev reads only after it has already lost its place. The dispatch protocol, AGENTS.md, and the os-dev agent definition say nothing about it.
⇒ A mechanism nobody sets is a mechanism that does nothing. The measured cost it exists to remove is real and current — from this shift, on this container:
outcome=command-exit waited=350 held=2 depth=3 rc=0 label=pnpm -s check:agent-test-spelling
A two-second gate waited 350 seconds behind a queue three deep. Separately, a foreground turn was killed at its ~10-minute ceiling about 210 s into a queued wait, and four of five devs in one earlier batch ended their turn while queued (#11463).
One-line executable criterion
An os-dev agent that runs a gate through scripts/pm/os-verify-lock.sh sets OS_VERIFY_LOCK_SLOT without having read the wrapper's --help — i.e. the instruction reaches it from its own agent definition or from AGENTS.md.
Why it is filed rather than fixed
AGENTS.md and .claude/agents/os-dev.md are (a) outside #11363's declared file surface, (b) governed surfaces requiring a draft PR and human merge, and (c) among the hottest merge-conflict files in the repo. A rider on a code PR is the wrong shape for all three reasons.
⚠️Do not treat this as landed work. PR #12335 is open at time of filing; if it does not merge, this card is void rather than merely blocked. Re-verify that OS_VERIFY_LOCK_SLOT is on main before dispatching.
Refs
#11363 (the contention card and its maintainer ruling) · PR #12335 (ships the mechanism) · #11463 (the "agent ends its turn while queued" half) · #12337 (the cold-closure-build-under-lock finding from the same measurement) · scripts/pm/os-verify-lock.sh
Filed by the
domain:devx @ objectstackseat (#6023, sessionsession_01UjM2ia8Av1v5NqfqQEQmC6) as a cross-lane residue of #11363 / PR #12335. Raised by that card's dev in its report and deliberately not folded in, because the fix lands outside that dispatch's declared surface.⛔ No
domain:*set — this seat does not produce that label. Routing note for triage: the remedy lands inAGENTS.mdand/or.claude/agents/os-dev.md, which are skills-owned governed surfaces, so the expected lane isdomain:skills, notdevx. Recorded here rather than acted on.What PR #12335 shipped
OS_VERIFY_LOCK_SLOT=<name>inscripts/pm/os-verify-lock.sh. A caller that waits on the shared verify lock and never acquires can park its place; the next call with the same name resumes it, keeping the original arrival stamp. Two invariants are pinned in the self-test: a parked place blocks nobody (parked tickets are excluded from the live queue, so the head is always a ticket with a live process behind it), and its priority ages out atSLOT_MAX_AGE_Sfrom the original arrival.It exists because obeying the foreground cap was the losing strategy: a ticket is removed when its process exits, so a caller that left came back with a fresh arrival stamp at the back of the queue, behind everyone who arrived while it was away.
The gap
Nothing tells an agent the variable exists. It is advertised by the wrapper's own
--helpand by the text printed on exit 99 — both of which a dev reads only after it has already lost its place. The dispatch protocol,AGENTS.md, and theos-devagent definition say nothing about it.⇒ A mechanism nobody sets is a mechanism that does nothing. The measured cost it exists to remove is real and current — from this shift, on this container:
A two-second gate waited 350 seconds behind a queue three deep. Separately, a foreground turn was killed at its ~10-minute ceiling about 210 s into a queued wait, and four of five devs in one earlier batch ended their turn while queued (#11463).
One-line executable criterion
An
os-devagent that runs a gate throughscripts/pm/os-verify-lock.shsetsOS_VERIFY_LOCK_SLOTwithout having read the wrapper's--help— i.e. the instruction reaches it from its own agent definition or fromAGENTS.md.Why it is filed rather than fixed
AGENTS.mdand.claude/agents/os-dev.mdare (a) outside #11363's declared file surface, (b) governed surfaces requiring a draft PR and human merge, and (c) among the hottest merge-conflict files in the repo. A rider on a code PR is the wrong shape for all three reasons.OS_VERIFY_LOCK_SLOTis onmainbefore dispatching.Refs
#11363 (the contention card and its maintainer ruling) · PR #12335 (ships the mechanism) · #11463 (the "agent ends its turn while queued" half) · #12337 (the cold-closure-build-under-lock finding from the same measurement) ·
scripts/pm/os-verify-lock.sh