[Bug] Ghost sessions: empty '(no messages)' rows undeletable from agents view, orphaned leases never reclaimed, partial fix #1079 has a regression

[stanleytejakusuma]

Summary

The agents view shows ghost rows labelled (no messages) that cannot be deleted from the UI. Ctrl+X either does nothing or reports "Session became active; stop it before deleting" because the daemon resolves them active from orphaned session leases. These are empty session files (no messages) whose owning process died, leaving the lease unclaimed.

Environment

• Prime Agent 0.7.2 and 0.7.3, Homebrew install
• macOS on Apple Silicon
• Default background daemon + agents view

Symptoms

1. Agents view lists rows labelled (no messages).
2. Ctrl+X on them either does nothing or reports "Session became active; stop it before deleting".
3. On disk these are empty session files (~720-849 bytes) containing only session/model_change/thinking_level_change/service_tier_change/session_state JSONL entries, zero user or assistant messages.
4. Orphaned session-lease dirs exist whose owner PID is dead and never reclaimed, so the daemon treats the row as active and refuses deletion.

Root cause (verified by bisection, 2026-08-14)

Two distinct defects:

A. Orphaned leases are never reclaimed. When the owning process dies without cleanup, the lease dir survives and the daemon continues to resolve the session as active, blocking UI deletion. There is no sweep that reclaims leases whose owner PID is dead.

B. The partial upstream fix (PR #1079) is broken. Cherry-picked and bisected locally:

• Commit 153b291 (cleanup sweep: session-file-actions.ts, session-lease.ts, daemon-mode.ts) is SAFE, targeted tests 30/30, full suite 191/191.
• Commit df9959a (session-manager.ts one-liner: shouldPersistWithoutAssistant excludes session_state, keeps only session_info) is the REGRESSION: reverting just that line restores 193/193; with it, 20 daemon-mode tests fail with ~30s timeouts.
• Mechanism: readSessionInfo() (session-manager.ts:1005) reads session_state entries to discover passive/draft sessions; callers include daemon-catalog-process.ts:70/74/99/247 and daemon-mode.ts:1088/2067/2128/3723 for RLM heartbeat restore, passive-child hydration, and saved-session catalog. Excluding session_state from the persist bypass makes legitimate passive sessions undiscoverable. A ghost draft and a legitimate passive RLM child share the same on-disk shape (session_state, no assistant message), so the exclusion breaks passive-session discovery.

Suggested fix

1. Ship only the safe half of fix(coding-agent): delete empty draft sessions on shutdown and sweep ghosts at startup #1079 (153b291): the cleanup sweep that reclaims leases whose owner PID is dead, externalized as a scheduled sweep in the daemon.
2. Do NOT ship df9959a as-is. Rework it so session_state exclusion distinguishes ghost drafts from legitimate passive RLM children (for example by lease liveness or a persisted session kind field) before touching persistence semantics.
3. Add a regression test: an empty session with a dead-owner lease must become deletable from the agents view within one sweep cycle.

Workaround

Deleting the orphaned files with rm works (the ghosts on disk are safely removable once their lease PIDs are dead), and a standalone cleanup script externalizing 153b291's sweep logic survives Homebrew upgrades. But the UI-level deletion should work without manual file surgery.

Related

• Empty draft session files (ghost sessions) and orphaned leases accumulate on shutdown — never cleaned up #1078 ghost sessions
• Unnamed sessions can remain status=working forever (0 messages, 0 clients), never reconciling to idle #1201 stuck-"working" sessions (same lease-resolution family)
• PR fix(coding-agent): delete empty draft sessions on shutdown and sweep ghosts at startup #1079 candidate fix, bisected above

[artgas1]

Re-diagnosed this on 0.8.1 (current main, c718bf3). The undeletable (no messages) rows reproduce with no orphaned lease involved: the daemon legitimately holds those sessions in memory, and the view shows a second, unclassified row for the same file.

What I measured

Three undeletable rows on my machine, and the daemon list explains all three:

{"id":"814345482e92","activeSessionId":"814345482e92","lifecycle":"draft","activity":"working",
 "messageCount":0,"attachedClients":0,"isStreaming":false,"isBashRunning":false,
 "hasRunningRlmChildren":false,"workerState":"ready","workerPid":7764,
 "sessionFile":"~/.prime/agent/sessions/01a0580a-25dc-....jsonl"}

Live workers, ready, zero messages, zero attached clients. list_saved_sessions returns the same files with messageCount: 0.

Mechanism

1. activeLifecycleForSession classifies a message-less resident session as draft, and shouldShowAgentsViewSession keeps only lifecycle === "live", so the daemon row is hidden. That part is intended.
2. AgentsViewMode.reconcileCatalogs filters lastListedSummaries with that rule but passes this.savedSessions to reconcileUnifiedSessions unfiltered. A saved-only record is never classified, so the same file comes back as an inactive row titled (no messages), with no activeSessionId.
3. Ctrl+X twice takes the saved-session delete path. It re-lists daemon sessions and calls resolveAgentsViewActiveSummaryForPath against the unfiltered list, finds the hidden draft, and stops with Session became active; stop it before deleting. The daemon would refuse anyway: delete_saved_session throws Cannot delete the currently active session for any file bound to an active session.
4. The session cannot be stopped from the UI either, because its only row is the hidden one. Dead end, and it survives restarts of the view.

Why the file exists at all: hasUserContent() counts model_change, thinking_level_change, and service_tier_change as user intent (CONTENT_ENTRY_TYPES), and every session writes those at startup. So isEmptyDraftContent is false for every abandoned new session, the discard guard never fires, and each abandoned session leaves one permanent ghost row.

Patch

Branch: artgas1/prime-agent@main...artgas1:prime-agent:fix/agents-view-hidden-draft-rows

Saved-catalog rows get the same live-only rule as daemon rows, reusing inactiveLifecycleForSession so on-disk and resident classification cannot drift:

export function shouldShowAgentsViewSavedSession(saved: AgentConnectionSavedSessionInfo): boolean {
	return inactiveLifecycleForSession(saved) === "live";
}

applied to the saved-only branch of reconcileUnifiedSessions (daemon records are still enriched by their saved twin). Archived and crashed files are covered by the same rule, which is what deactivatePendingAgent already assumes when it writes the archived marker "so sessions do not resurface on the next scan" - today they do resurface.

Validation: test/agents-view-state.test.ts (67 passed, 2 new; the new row test fails without the guard), plus agents-view-mode, agents-view-inactive-reply, agents-view-missing-cwd, session-view-search, daemon-session-list, session-manager/session-state, and test/suite/regressions/502-unified-session-view.test.ts. npm run check clean.

What this does not fix

The ghost files still accumulate on disk, now invisibly. Two follow-ups are worth separating from the view fix:

• A sweep for message-less draft files with no resident session, which is the safe half of fix(coding-agent): delete empty draft sessions on shutdown and sweep ghosts at startup #1079 that this thread already isolated.
• hasUserContent() treating the startup model/thinking/service-tier snapshot as user intent, which is why the drafts are persisted in the first place.

Immediate workaround for anyone stuck: over the daemon socket, {"type":"kill","activeSessionId":...} for each draft with messageCount: 0 and attachedClients: 0, then {"type":"delete_saved_session","sessionPath":...}; the file goes to the trash. rm alone is not enough while the worker is resident, since the row returns until the daemon drops it.

---

Filed the 0.8.1 mechanism as its own report with a clean-state reproduction (three daemon commands, no dead process involved): #1921. Keeping this thread for the lease half, which is a different failure mode.

[artgas1]

@stanleytejakusuma if you still see the ghosts on 0.8.1, this one-liner separates your mechanism from the one I hit, no scripts needed:

prime-agent list --all --json | jq -c '.sessions[]
  | select(.messageCount == 0)
  | {id, lifecycle, activity, attachedClients, workerPid, sessionFile}'

• Rows come back with "lifecycle":"draft" and a live workerPid: the daemon is holding the session, so the file is refused by delete_saved_session and the row is drawn only by the saved catalog. That is the path I measured in [Bug] Undeletable "(no messages)" rows: the saved-session catalog skips the draft lifecycle filter #1921, and no dead process is involved.
• The ghost files exist on disk but nothing prints: the daemon does not own them, so the refusal has to come from somewhere else, which is the lease story in this thread.

For the first case the removal is two daemon commands per ghost, {"type":"kill","activeSessionId":...} then {"type":"delete_saved_session","sessionPath":...}; the file goes to the trash. Plain rm is not enough while the worker is resident: the daemon keeps the session, and the row returns.

Worth knowing which one you have, because the fixes differ and only one of them is a lease sweep.