fix(coding-agent): separate active commands from queued input

[snimu]

Summary

• stop exposing executing /compact, /refine, /goal, and /autonomous commands as queued steering or follow-up input
• separate pending queue state from active session-input work for admission, cron, heartbeats, agent messaging, daemon activity, and Agents View
• prevent queued /compact from treating itself as successor work and unnecessarily deferring compaction-triggered refinement
• make headless completion wait for the full session scheduler, not only model idleness
• expose optional active-input state to connection and RPC clients while preserving legacy daemon busy counts for mixed-version shutdown and update safety

Validation

• 736 tests passed across 22 headless, session scheduling, print, JSON, RPC, connection, daemon, restart, and Agents View suites
• 19 real daemon-worker client-mode process tests passed, covering print, JSON, RPC EOF draining, and worker ownership
• 14 credential-gated live RPC tests were skipped; equivalent faux-provider and real local daemon-worker paths passed
• npm run check

---

Note

Medium Risk
Touches session input scheduling, headless exit timing, and daemon peer summaries; wrong busy/queue classification could affect admission, heartbeats, and multi-client views, but changes are narrowly scoped to separating active commands from the durable queue.

Overview
Executing slash commands (e.g. /compact) no longer show up in user-visible queued steering or follow-up previews or empty session_action_update queue projections while they run; tests assert queuedActionCount stays 0 and unfinishedActionCount tracks the in-flight command until it finishes.

The daemon supervisor now triggers refreshWorkerSummaries on session_action_update session events (alongside turn start/end), so remote Agents View peers pick up active vs queued input changes without waiting for a turn boundary. Headless completion is covered by tests to block until an executing session command completes, not only model idleness.

Changelog records the queued-input display fix.

^{Reviewed by Cursor Bugbot for commit 7ec7438. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Separate executing slash commands from queued input in coding agent

• Ensures executing slash commands (e.g. /compact) are tracked via unfinishedActionCount rather than appearing as queued steering or follow-up messages (queuedActionCount).
• Extends DaemonSupervisor.handleWorkerFrame to call refreshWorkerSummaries on session_action_update events, so remote agent peers stay in sync with input activity.
• waitForHeadlessCompletion now waits for an executing session command to finish before settling.
• Adds tests in agent-session-queue.test.ts and daemon-supervisor-monitor.test.ts covering the new counts and headless-completion behavior.

^{Macroscope summarized 7ec7438.}

[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 7ec7438. Configure here.}

[cursor]

Overlapping refreshes can stale peer state

Medium Severity

session_action_update now kicks off refreshWorkerSummaries on every queue snapshot change, but those list RPCs are not serialized or coalesced. A slower earlier refresh can overwrite a later one, so remote peer summaries can keep a finished command as unfinished work.

 

^{Reviewed by Cursor Bugbot for commit 7ec7438. Configure here.}