fix(session): retry resumed turns that fail against an expired Cursor agent

[justin-carper]

Problem

After a session sits idle then resumes, turns intermittently fail with:

Cursor run ended with status "error"

Timing is inconsistent because the trigger is server-side Cursor agent expiry, not any local clock. Cursor's API publishes no agent-retention TTL (only 429 backoff), and the server drops agents well before our local 7-day reuse window.

Root cause

The create-fallback guarded resume but not the send that follows it.

Path	Behavior
resumeAgent() throws	Caught → falls through to fresh createAgent (full replay). Graceful.
resumeAgent() succeeds, later send() fails	Not caught. Error propagated and failed the turn.

When a resumed agent is server-side-stale, Agent.resume(agentId) still succeeds locally; the failure only surfaces when the run completes with status === "error" inside the stream. The pooled record had already been re-pointed at the dead agentId, so the next turn resumed the same dead agent again.

Fix

agentRun now wraps the resumed-turn stream. On a resumed turn that throws before emitting any event and when not aborted, it transparently:

1. Re-creates a fresh agent (same pool key/record, no resumeAgentId).
2. Replays the full transcript (no context loss).
3. Re-pools under the same session, overwriting the dead agentId.

Guarded to a single attempt. Never retries a fresh-create turn, an already-emitting stream, or a user abort. If re-acquire itself fails, the original resume failure is chained as error.cause for diagnosability.

Tests

New test/language-model.test.ts (7 cases, drive doStream end-to-end with a mocked SDK backend):

• Resumed + error + no emit → re-create + full replay, pool re-pointed
• Resumed + error after emit → no retry (no double-emit)
• Fresh-create error → no retry
• Aborted → no retry
• Retry send also fails → propagates (single attempt)
• Non-pooled explicit-agentId resume → retries, closes both agents, pools neither
• Re-acquire throws → original failure chained as cause

Full suite 228 pass, tsc --noEmit clean, npm run build success.

Notes

• No change to session-pool.ts or agent-events.ts — retry is composed from existing primitives.
• Local TTLs (7d session, 24h/30d model cache) left unchanged; they were not the cause.