refactor(coding-agent): harden session action races

[snimu]

Summary

Final PR of the session-action stack: deletes machinery the new lifecycle has made provably redundant - most notably the entire broad turn-admission lease - hardens the remaining races with a deterministic test matrix, and removes the last dead legacy-protocol code. Net -59 production LOC on top of #4.

Stack: #572 contracts <- #573 kernel <- #574 admission <- #575 protocol <- #576 cleanup (this)

Motivation

The broad FIFO admission lease existed to stop direct prompts from overtaking queued work. After #3 there are no direct prompts: store arrival order is the FIFO, so the lease - promise tail, owner token, AsyncLocalStorage reentrancy, pause coordination - protected nothing. Deleting safety machinery is only safe with proof, so this PR pairs each deletion with an audit of its call sites and adds race tests at the await points that remain.

Code changes

• Deleted the broad admission machinery: _acquireTurnAdmission, _acquireSessionActionAdmission, _waitForQueuedWorkAdmission, _wakeQueuedWorkAdmissionWaiters, _queuedWorkAdmissionWaiters, _turnAdmissionTail, _turnAdmissionOwner, _turnAdmissionContext, and the admission leases around prompts, injected turns, pump preparation, and branch navigation.
• Replaced by a narrow commit fence: held only across dispatch commit edges and session-command transcript-row commits - never across extension hooks or LLM awaits (an extension calling ctx.navigateTree() from a session_before_compact hook must not deadlock; regression-tested).
• Kept, with the race each protects now named: pump epoch (stale preparation after branch pause), pump-requested flag (overlapping wakeups), pump suspension (abort/restart), queued-work pauses (overlapping branch-mutation leases), checkpoint waiters (lifecycle notifications), arrival epoch (stale-continuation invalidation).
• Deleted dead legacy code: DaemonClient.requestLegacy, requestLegacyData, the protocol-3 heartbeat branch, the unused "handoff" busy-check variant.
• Final rename: reportExecutionError -> queueVisible (the field's actual meaning).
• New test/suite/agent-session-action-races.test.ts: deterministic pause/clear/restart/dispose injection at commit-fence boundaries for turns and session commands; asserts every ticket settles exactly once and every action is restored xor delivered.
• Review-driven fixes included: suspended-queue admission preserved for steer/followUp/agent-message queueing after abort (cron and a2a paths); no deferral-observer leak on a2a direct delivery; a2a acceptance still awaits actual delivery.

LOC: +390/-399 production (net -9), +765/-165 tests (includes review-fix commits: fence revalidation and rescheduling, dispose linkage, settlement/ordering/idle guards, injected-admission serialization; and the compensating-state deletion commit).

Stack result (vs main)

Feature code +3058/-2659 (net +399). Tests +2633/-1030 (net +1603). One typed owner for every turn-producing input; exactly-once delivery via two-edge records; queue display, /compact successor detection, and headless completion correct by construction.

Note

Harden session action scheduling with commit-fence serialization to fix race conditions

• Replaces the previous turn-admission FIFO and waiter set with a new FIFO commit-fence mechanism (_acquireSessionActionCommitFence, _acquireDirectTurnAdmissionFence) that serializes session action commits using AsyncLocalStorage scoping and a disposal abort controller.
• Adds hasPendingSessionWork getter to AgentSession and field to HeartbeatCronSessionActivity to track non-durable committing turns; updates heartbeat deferral logic in shouldDeferHeartbeatCronJob so steer heartbeats proceed during streaming while follow-up heartbeats still defer.
• Moves durable message append and delivery settlement for session commands into _executeSelectedSessionCommand under the commit fence, so commands that lose a handoff race are rolled back rather than executed.
• Fixes waitForSessionInputCheckpoint and waitForIdle to drain the agent event queue before reporting idle, reducing post-idle races.
• Removes shouldYieldLowerRun, mandatoryCheckpointsComplete, and requestLegacy as part of the cleanup; manageHeartbeat now unconditionally requires the heartbeat_management server capability.

^{Macroscope summarized 07c9f4f.}

---

Note

High Risk
Core session scheduling, prompt admission, compaction continuation, and heartbeat timing changed; incorrect fence or pending-work semantics could cause duplicate delivery, early headless exit, or stuck queues.

Overview
Session input scheduling is consolidated on the action store lifecycle: the broad turn-admission lease (_acquireTurnAdmission, pause waiters, shouldYieldLowerRun, mandatoryCheckpointsComplete) is removed in favor of a FIFO commit fence (_sessionActionCommitFence) held only across dispatch commits and session-command transcript writes—not across extension hooks or LLM work.

The input pump no longer wraps preparation in admission; turns batch all delivery records (prefix + primary + next-turn) at commit; session commands record the slash message before running /compact, /refine, etc.; checkpoints and waitForIdle wait on the commit fence and agent event queue; goal context always queues through the store instead of bypassing during pump/streaming.

hasPendingSessionWork distinguishes queued/preparing/committing work from in-flight agent turns, driving heartbeat deferral (shouldDeferHeartbeatCronJob), headless completion, and post-compaction continue logic.

Daemon clients drop requestLegacy and protocol-3 heartbeat fallbacks; heartbeat management requires server capabilities.

Tests add agent-session-action-races.test.ts and expand queue/compaction/heartbeat coverage for commit-fence pause, clear, restart, dispose, navigation reentrancy, and ordering.

^{Reviewed by Cursor Bugbot for commit 07c9f4f. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 84bb107. Configure here.}