[think] Cold-start submitMessages drain executes before onStart initializes session — "Cannot read properties of undefined (reading 'appendMessage')"

[cgrdavies]

Describe the bug

Calling the submitMessages() RPC on a cold (hibernated/evicted) Think agent durably records the submission, but the post-accept drain can execute it before the agent's onStart has assigned this.session (this.session = await this.configureSession(baseSession)). The turn then crashes in _appendMessageToHistory with Cannot read properties of undefined (reading 'appendMessage'), and the submission is marked terminally failed — the durable row survives, but the turn it represents is lost (a retry under the same idempotency key is rejected as a duplicate; a retry under a fresh key hits the same cold-start window).

Warm agents — any instance where a chat connection or other code path already ran onStart in the same lifetime — process the identical submission fine, which makes this intermittent and easy to misdiagnose.

To Reproduce

1. A Think agent DO (stock or custom configureSession).
2. Ensure the DO is cold: no WebSocket/chat connection, evicted or never woken since deploy.
3. From another Worker, call stub.submitMessages([message], { idempotencyKey, metadata }) over a DO stub — a "completion callback"-style programmatic turn.
4. Observe [Think] Submission failed … "Cannot read properties of undefined (reading 'appendMessage')" and a terminally-failed submission row.

Sequence (published dist, v0.12.1 line refs): submitMessages accepts → _startSubmissionDrain() → keepAliveWhile(() => this._drainSubmissions()) (think.js ~5206) → _runSubmission → turn execution → _appendMessageToHistory (~1311) dereferences this.session, which the wrapped onStart (~934) has not assigned yet.

Expected behavior

submitMessages (or the drain it triggers) should await the agent's internal readiness — session initialization — before executing a submission. Callers of a public RPC shouldn't have to know about onStart timing. Notably, the alarm path Think schedules via this.schedule(0, "_drainThinkSubmissions") re-enters through the normal agent lifecycle, but the immediate keepAliveWhile drain does not wait for it.

Environment

• @cloudflare/think 0.12.1
• The same cross-worker callback pattern ran reliably for us on 0.8.8 under production traffic; after upgrading to 0.12.x we see daily failures whenever the target DO is cold.

Workaround

Subclass override that forces initialization before delegating:

override async submitMessages(...args: Parameters<Think['submitMessages']>) {
  await this.ensureInitialized() // awaits __unsafe_ensureInitialized() and asserts this.session
  return super.submitMessages(...args)
}

Impact

Any "durable completion hook" pattern (cf. #1752) that wakes an idle agent via submitMessages silently loses events whenever the target agent is cold — for us that was workflow-completion and delegated-sub-agent-completion notifications in production.