bug(core): stuck session with exited runtime never auto-promotes to terminated — lingers in sidebar indefinitely #1933

Repro

Spawn a session, let the agent crash or exit without acknowledging (reportWatcherActiveTrigger: no_acknowledge). The lifecycle manager probes, sees signal_disagreement between the native activity signal (activity=exited) and the initial runtime probe (runtime=alive), runs through 50 detecting attempts, escalates, and parks the session at state: stuck, reason: probe_failure. The runtime probe eventually updates to exited, but the session-level state never advances.

Observed state (real session ao-165)

"lifecycle": {
  "session":  { "state": "stuck",  "reason": "probe_failure",  "terminatedAt": null,
                "lastTransitionAt": "2026-05-19T13:14:54.490Z" },
  "runtime":  { "state": "exited", "reason": "process_missing",
                "lastObservedAt": "2026-05-19T13:14:54.490Z" },
  "pr":       { "state": "none",   "reason": "not_created" }
},
"status": "stuck",
"detectingAttempts": "50",
"detectingEscalatedAt": "2026-05-19T12:27:51.146Z",
"lifecycleEvidence": "signal_disagreement runtime=alive process=dead activity_signal=valid via_native activity=exited",
"reportWatcherActiveTrigger": "no_acknowledge",
"reportWatcherTriggerCount": "41"

The agent process is gone, the runtime is gone, no PR was ever created — but the session is not terminated.

Why this is a bug

The sidebar visibility gate in packages/web/src/components/ProjectSidebar.tsx:454-457 only hides sessions whose getAttentionLevel(session) === \"done\". A stuck lifecycle state routes to respond in packages/web/src/lib/types.ts:504-512:

if (session.lifecycle?.sessionState === \"stuck\" || ...) return \"respond\";

So the session sits on the sidebar with an orange dot forever. Neither the Show killed nor the Show done toggle helps — they only gate done-level sessions. The only ways to clear it today:

• ao kill <id> (manual)
• Click Restore (spawns a new runtime, leaves the record)
• rm the JSON by hand

That's not great UX for a session the lifecycle manager has already concluded is unrecoverable.

Proposed fix (core, not surface)

When resolveProbeDecision (in packages/core/src/lifecycle-manager.ts) determines:

• runtime.state === \"exited\" or \"missing\", and
• native activity signal is exited, and
• no live PR depends on the runtime (pr.state ∈ {none, closed_unmerged}),

…it should promote session.state from stuck → terminated with reason: probe_failure (or a new agent_process_exited reason). deriveLegacyStatus already maps error_in_process / probe_failure → errored, which is in the done bucket, so the sidebar would then hide it by default once showDone is off.

Today the lifecycle treats probe_failure as "I can't tell, escalate to a human". But when every signal agrees the agent is gone, there's nothing to escalate — just terminate it.

Out of scope (separate concerns)

• Restore button correctness: covered by Restore: requiresNativeRestore() refuses fallback when no native session ever existed; UI shows Restore button anyway #1908.
• Activity-detection redesign: covered by refactor(core): redesign activity detection — split inbox vs liveness, retire per-plugin cascade #1899.
• This is specifically about lifecycle promotion when runtime + native signal agree the agent has exited.

File:line index

• Visibility gate: packages/web/src/components/ProjectSidebar.tsx:454-457
• \"stuck\" → \"respond\" mapping: packages/web/src/lib/types.ts:504-512
• isDashboardSessionDone: packages/web/src/lib/types.ts:392-410
• deriveLegacyStatus terminal-reason mapping: packages/core/src/lifecycle-state.ts:444-460
• Probe decision pipeline: packages/core/src/lifecycle-manager.ts (resolveProbeDecision)