feat(session): warm resume for multi-message prompt interjections

[justin-carper]

Problem

When a user sends two or more messages while the Cursor agent is still processing (opencode queues them and replays the grown transcript next turn), the provider classified the turn as divergence → fresh agent + full transcript replay. Warm agent context was discarded; every multi-message interjection paid a cold-start.

Change

New continuation-multi turn kind: when the prior user-message sequence is a strict prefix and the new messages form a contiguous trailing run of user turns, resume the pooled agent and replay just the new messages as sequential turns — leading ones sent silently (no deltas), only the final run streams back to opencode. This mirrors opencode's own coalesced-loop model, where interjected messages fold into one visible assistant turn.

File	Change
transcript-fingerprint.ts	continuation-multi kind + newUserCount + contiguous-tail guard
message-map.ts	trailingUserMessages(prompt, count) + shared userTurnToCursorMessage
agent-events.ts	sendAgentTurnSilently + shared sendWithBusyRetry (busy local.force retry)
language-model.ts	sequential-send wiring, abort-aware, record rollback on partial delivery
session-pool.ts	dropSessionRecord for mid-sequence failure recovery

Safety properties

• Interleaved assistant turns disqualify (abort-then-interject shape): tail-only replay would silently drop the earlier queued message, so that shape stays divergence → full replay.
• Partial delivery rolls back: the pool record is written optimistically before sends; an error/abort mid-sequence drops the record so the next turn replays fresh instead of resuming on top of messages the agent never received.
• Edit/revert unchanged: non-prefix transcripts stay divergence.
• Abort mid-sequence cancels the in-flight run and stops further sends.
• Backend-agnostic: works through the shared AgentLike contract (in-process SDK + sidecar).

Known trade-off

Silent turns' token usage is not reported (no onDelta observer on runs 1..N-1), so multi-message turns slightly undercount usage. Documented in code.

Review

Code-reviewed (agent): 1 critical (non-contiguous tail could drop messages) + 1 important (stale record after mid-sequence failure) — both fixed with dedicated regression tests.

Verification

• tsc --noEmit clean
• 236 tests pass (20 files; +9 new/updated cases incl. integration tests driving doStream end-to-end with a mocked Cursor runtime)
• tsup build success