feat(coding-agent): whole-tree idle eviction

[snimu]

Worker processes used to stay alive forever, even when every session in them had been idle for hours. This PR makes the supervisor notice and stop them.

A background sweep checks each worker. If every session in it has been idle past a threshold, and nothing is attached, and no timers (heartbeats/cron jobs) are pinned to it, the worker process is stopped. Nothing is lost: the sessions become ordinary inactive entries, and attaching to one — or sending it a message — starts a fresh worker and picks up where it left off.

The threshold is a new idleEvictionMinutes setting (default 90; off disables it), which you can change in /settings.

The check errs on the side of caution: if a worker's state is at all unclear (busy session, running subagent, missing timestamps, mid-startup or mid-shutdown), it's left alone.

---

This is part 1–7 of a stack (each PR builds on the one before it):

1. feat(coding-agent): lazy child session hydration #583 — don't load finished subagents until someone needs them
2. feat(coding-agent): persist parent edges and derived depth for all sessions #584 — record each session's parent and depth on disk
3. feat(coding-agent): /rlm-max-depth session command #585 — new /rlm-max-depth command to control how deep agents can nest
4. feat(coding-agent): recursive scoped agents view #586 — browse the agent tree in the agents view
5. feat(coding-agent): delete child-agent inspector; one subagent summary line #587 — replace the child-agent inspector with one summary line
6. feat(coding-agent): whole-tree idle eviction #588 — shut down worker processes whose sessions have all gone idle
7. feat(coding-agent): intra-tree child session passivation #589 — quietly unload idle child sessions inside a running worker

Tests: the full suite at the top of the stack fails only on tests that already fail on the base branch. Each branch in the stack also passes checks and its own tests independently.

Note

Add whole-tree idle eviction to the coding agent daemon supervisor

• Adds canEvictWorker in session-action-store.ts to evaluate eviction eligibility: a worker is evictable only when ready, connected, unowned, not updating, and all sessions have been idle beyond the configured threshold with no attachments, heartbeats, or cron jobs.
• Adds idleEvictionMinutes to SettingsManager (default 90 minutes; 'off' disables eviction) and exposes it in the interactive settings UI as an 'Idle worker eviction' option.
• The supervisor periodically runs runIdleEvictionSweep, which reloads settings, evaluates candidates via canEvictWorker, and stops eligible workers. Sweep interval is bounded between 1 and 5 minutes.
• Adds mutation fencing in the supervisor so mutating commands wait for any in-flight eviction sweep to complete before proceeding.
• Extends SessionSummary with hasRegisteredHeartbeat, hasRegisteredCronJob, and lastActivityAt fields (protocol schema bumped to revision 12).
• Agent messages targeting a selector not locally resident are now forwarded to the supervisor, which resolves the saved session via the catalog, wakes the worker if needed, and delivers the message.
• Risk: existing clients consuming session list responses must handle the new schema revision 12 fields.

^{Macroscope summarized b184c50.}

---

Note

Medium Risk
Changes supervisor worker lifecycle, session protocol schema, and cross-worker messaging wake paths; mistakes could stop workers prematurely or mishandle a2a delivery, though policy is conservative and heavily tested.

Overview
Introduces whole-tree idle worker eviction: a global idleEvictionMinutes setting (default 90, off to disable) in settings and /settings, plus pure canEvictWorker policy that only stops a worker when every session is past the threshold with no attachments, active heartbeats, or registered cron jobs.

The daemon supervisor runs periodic sweeps (interval derived from the threshold), reloads settings each pass, fences mutating commands during eviction, and stops eligible workers while logging evictions. Session list rows gain hasRegisteredHeartbeat, hasRegisteredCronJob, and lastActivityAt (from message timestamps); daemon schema revision 12 documents these fields. Paused heartbeats no longer pin residency.

Agent-to-agent messaging improves for non-resident targets: the supervisor can catalog-resolve a selector (including exact names, with ambiguity errors when a name collides with an id prefix), wake a saved session through the existing create path, and deliver the message; workers validate messages before remote fallback and do not retry supervisor rejections.

Also refreshes generated model catalog entries (pricing/limits and a few provider models).

^{Reviewed by Cursor Bugbot for commit b184c50. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes and found 2 potential issues.

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

^{Reviewed by Cursor Bugbot for commit 19a2be1. Configure here.}