Sunil Pai

fix(think): messenger delivery, prompt truncation and media eviction (#2387) * fix(think): messenger delivery, prompt truncation and media eviction review fixes - keep Chat SDK thread locks alive (agents/chat-sdk lockHeartbeat) and give messenger queue entries a 30 minute TTL - drain queued messages after an interrupted reply is recovered - refresh the typing indicator until first text, tolerate its failure, and never post a blank reply before the apology - detect mentions on skipped burst messages, including unsubscribed threads - strip raw payloads/bytes from persisted metadata; map inline attachment data - keep start metadata when a recovery continuation extends a message - truncate inline media/text file items in older tool results and match results by message position - align media eviction with the stepped truncation cutoff; truncationStep Infinity disables truncation - resolve the default model lazily; reject async getGateway clearly Co-authored-by: Cursor <cursoragent@cursor.com> * fix(think,agents): keep beforeTurn model override for extensions; stop lock heartbeats on disconnect - Extension beforeTurn snapshots use the subclass model override instead of reading ctx.model, which lazily resolved the default model (and its AI binding/gateway) even when unused. - ChatSdkStateAdapter.disconnect() clears lock heartbeat timers, and heartbeats never start or tick while disconnected. - Test that the installed chat package still has the private getLockKey/drainQueue methods the recovery drain relies on. Co-authored-by: Cursor <cursoragent@cursor.com> * fix(think): drain queued messenger messages only after recovery settles A failed messenger reply recovery that the Tasks run will replay no longer starts the queue drain, so follow-ups queued behind it are not answered ahead of the original reply. Success and terminal failures still drain. Co-authored-by: Cursor <cursoragent@cursor.com> * test(agents): pin repeated tool id matching past provider-executed results convertToModelMessages emits provider-executed results in the assistant message, so they never occupy a position among tool-message results. Clarify producesToolResult and cover the case with the real converter. Co-authored-by: Cursor <cursoragent@cursor.com> --------- Co-authored-by: Sunil Pai <18808+threepointone@users.noreply.github.com> Co-authored-by: Cursor <cursoragent@cursor.com>

fix(agents): agent-tool replay sequencing, re-attach, and fiber recovery (#2384) * fix(agents): agent-tool replay sequencing, re-attach, and fiber recovery review fixes Co-authored-by: Cursor <cursoragent@cursor.com> * fix(agents): keep managed fiber outcomes, bound replay child resolution, stop progress frames shifting child tail numbering - Record the fiber body's outcome (completed/error/aborted + message) on the run row with its settlement stamp, and settle a managed ledger whose own settle write failed with that outcome instead of always 'completed'. Schema v14 adds cf_agents_runs.outcome / error_message. - Connect-time replay now bounds child resolution too, sharing one per-run budget across resolve, chunk read, and milestone inspection. - The shared broadcast interceptor no longer advances the child's live counter for progress/milestone frames, and ai-chat/Think tails emit those frames outside the stored-position high-water dedupe, so a chunk stored and broadcast during a tail's drain is not duplicated and progress is not dropped. Co-authored-by: Cursor <cursoragent@cursor.com> * fix(agents): keep agent-tool numbering across unstored chunks, cold re-attach, and repeated progress - A chunk too large to store is broadcast live but never replayed. The child's broadcast snoop now gives it the next stored position without consuming it plus a unique `unstoredId`; child tails forward it outside the stored-position dedupe, the parent does not advance its sequence for it, and `agentToolEventDedupeKey` keys it on the id. Live and replayed numbering stay aligned after a skipped chunk. - ai-chat and Think tails seed a cold live counter from the stored backlog before registering the forwarder, instead of realigning after the drain / inspection, so a chunk broadcast during those awaits after a restart is forwarded rather than dropped. - Each `reportProgress` frame carries a unique `id` that the dedupe key uses, so identical progress payloads are no longer dropped by `useAgentToolEvents`. Co-authored-by: Cursor <cursoragent@cursor.com> --------- Co-authored-by: Sunil Pai <18808+threepointone@users.noreply.github.com> Co-authored-by: Cursor <cursoragent@cursor.com>

fix(chat): settle turns correctly across replay, resume, and observers (#2383) * fix(chat): settle turns correctly across replay, resume, and observers Wire protocol (agents, ai-chat): - Record a stream's close outcome in its metadata and echo it on replayed and bare-done terminals; orphaned streams report "aborted" and a recovery-scheduled close reports "recovering". Outcomes and origin ids survive cutover/reclaim deleting the row. - A resume ACK that lands while ai-chat holds the terminal frames replays the finished stream without an early bare done. - The "No response" done reaches every connection, including the origin, with an outcome. useAgentChat: - Observers settle on a done for a request other than the observed one. - A recovering done keeps isRecovering and holds onToolCall until the successor turn. - onTurnEnd for the tab's own request fires after status settles, so sendMessage works inside it. - With resume:false, probe on reconnect so a held onToolCall is released. - Live chunks skip the merge once divergence is detected. - Cap remembered turn errors and clear them in clearHistory. Co-authored-by: Cursor <cursoragent@cursor.com> * fix(think): record recovering/aborted outcomes on chat streams for replay Co-authored-by: Cursor <cursoragent@cursor.com> * fix(chat): replay recovery/error closes to held ACKs and tighten onToolCall gates - ResumableStream replays the request's last closed stream (finished, recovering, or errored) to a resume ACK that lands while ai-chat holds the terminal frames, instead of only the pending-cutover stream. - useAgentChat tracks the recovering request id: replays and the recovering request's own frames no longer release the hold; live frames for a later request, recovering:false, or an idle reconnect probe with no recovery reported on connect do. - With resume:false the reconnect probe also runs for a recovering hold, and an offer for another request releases a stale held turn; a proactive offer holds tool calls until that stream's terminal frame. - Think clears pending resume connections whenever a chat stream ends, so a connection that never ACKs still gets the terminal frame on chat() RPC turns. Co-authored-by: Cursor <cursoragent@cursor.com> --------- Co-authored-by: Sunil Pai <18808+threepointone@users.noreply.github.com> Co-authored-by: Cursor <cursoragent@cursor.com>

fix(think): submission and turn lifecycle edge cases (#2386) * fix(think): submission and turn lifecycle edge cases - waitForSubmission throws in-turn / from onSubmissionStatus instead of deadlocking - cancelled running submissions emit terminal status once; controller registered before running hook - reset-skipped submissions visible to waiters only after their hook - activeChannel gated to the admitted turn - _lastTurnChannel set at inference start, cleared on reset - autoContinue:false mid-stream reject flushes deferred drop at finalize - dropGenerationAfterToolCall keeps later steps' text - regression tests, including ai-chat #2352 port and #2353 gaps Co-authored-by: Cursor <cursoragent@cursor.com> * fix(think): keep continuation channel when a turn fails before streaming Record _lastTurnChannel only after inference preparation (beforeTurn) succeeds and the stream starts, so a channel-less continuation does not route to a turn that never ran. Co-authored-by: Cursor <cursoragent@cursor.com> --------- Co-authored-by: Sunil Pai <18808+threepointone@users.noreply.github.com> Co-authored-by: Cursor <cursoragent@cursor.com>

fix(chat): skip replayed continuation chunks the client already applied (#1951) (#2348) * fix(chat): skip replayed continuation chunks the client already applied (#1951) Co-authored-by: Cursor <cursoragent@cursor.com> * fix(chat): mark Think continuation streams for replay; skip onData for applied replays Co-authored-by: Cursor <cursoragent@cursor.com> * fix(chat): keep chunk seq counting across streams restarted under one request Co-authored-by: Cursor <cursoragent@cursor.com> * fix(chat): bound the applied-chunk ledger to recently applied requests Co-authored-by: Cursor <cursoragent@cursor.com> --------- Co-authored-by: Sunil Pai <18808+threepointone@users.noreply.github.com> Co-authored-by: Cursor <cursoragent@cursor.com>

feat(chat): originating user message ids on terminal response frames (#2365) * feat(chat): originating user message ids on terminal response frames (#2280) Co-authored-by: Cursor <cursoragent@cursor.com> * fix(chat): carry originating message ids onto recovered turns (#2280) Co-authored-by: Cursor <cursoragent@cursor.com> * Keep originating message ids for turns recovered before their stream started Co-authored-by: Cursor <cursoragent@cursor.com> * Drop the duplicated sentence from the terminal message ids changeset Co-authored-by: Cursor <cursoragent@cursor.com> * Carry originating message ids onto a live re-run of a turn that failed before any part Co-authored-by: Cursor <cursoragent@cursor.com> * Keep origin message ids on wake exhaustion and after cutover Co-authored-by: Cursor <cursoragent@cursor.com> * Scope recovery origin ids to the recovery callback's async context Co-authored-by: Cursor <cursoragent@cursor.com> --------- Co-authored-by: Sunil Pai <18808+threepointone@users.noreply.github.com> Co-authored-by: Cursor <cursoragent@cursor.com> Co-authored-by: x <x@x>

fix(chat): harden live stream-error recovery in Think and ai-chat (#2385) * fix(chat): harden live stream-error recovery (review follow-ups) Co-authored-by: Cursor <cursoragent@cursor.com> * fix(chat): address stream-recovery review findings - Think RPC chat(): a failure after the stream drained reports status error and settles the pending stream as errored instead of leaving it awaiting a cutover. - Think: a failure in the bookkeeping before onChatResponse keeps the pending hook marker so the next start replays the hook. - ai-chat: persist: false on live recovery also drops a new turn's message persisted early for a tool approval request, so the turn is retried. - Think messengers: recovered replies are at-most-once per post; a failed post is not re-sent and the retry resumes with the next post. Co-authored-by: Cursor <cursoragent@cursor.com> * fix(think): serialize recovered messenger reply delivery per key Startup replay, live settlement and scheduled retries could interleave across awaits and post the same chunk twice. Chain deliveries per recovery key and re-read the stored cursor before each post. Co-authored-by: Cursor <cursoragent@cursor.com> --------- Co-authored-by: Sunil Pai <18808+threepointone@users.noreply.github.com> Co-authored-by: Cursor <cursoragent@cursor.com>

feat(agents): opt-in terminal-only event delivery for runAgentTool (#2364) * feat(agents): opt-in terminal-only event delivery for runAgentTool (#2298) Co-authored-by: Cursor <cursoragent@cursor.com> * fix(agents): keep terminal-only agent tool runs suppressed across child restarts (#2298) Co-authored-by: Cursor <cursoragent@cursor.com> * Declare the recovered-chunk helper on the Think agent-tool test stub Co-authored-by: Cursor <cursoragent@cursor.com> * Replay persisted agent-tool milestones and dedupe events by identity Co-authored-by: Cursor <cursoragent@cursor.com> * Key agent-tool lifecycle events on content and replay milestones when chunk reads fail Co-authored-by: Cursor <cursoragent@cursor.com> --------- Co-authored-by: Sunil Pai <18808+threepointone@users.noreply.github.com> Co-authored-by: Cursor <cursoragent@cursor.com> Co-authored-by: x <x@x>

fix(think): deliver the recovered answer to the messenger thread (#2106) (#2347) * fix(think): deliver the recovered answer to the messenger thread (#2106) Co-authored-by: Cursor <cursoragent@cursor.com> * fix(think): keep recovered messenger delivery across repeated interruptions and the visible limit Co-authored-by: Cursor <cursoragent@cursor.com> * fix(think): repost a rejected stream's partial; post the empty-response text for a blank recovered reply Co-authored-by: Cursor <cursoragent@cursor.com> * fix(think): keep a messenger reply pending while the platform re-runs recovery Co-authored-by: Cursor <cursoragent@cursor.com> * fix(think): pin the recovered messenger reply before settlement detaches Co-authored-by: Cursor <cursoragent@cursor.com> * fix(think): store a messenger reply's outcome before its recovery incident settles Co-authored-by: Cursor <cursoragent@cursor.com> * fix(think): look up a pending messenger reply's incident by its encoded key Co-authored-by: Cursor <cursoragent@cursor.com> --------- Co-authored-by: Sunil Pai <18808+threepointone@users.noreply.github.com> Co-authored-by: Cursor <cursoragent@cursor.com>

fix(think): structured output on wait TurnResult; submission messageId (#2263, #2264) (#2341) * fix(think): structured output on wait TurnResult; submission messageId (#2263, #2264) Co-authored-by: Cursor <cursoragent@cursor.com> * fix(think): capture structured output on continuations; record orphan-persisted message ids Co-authored-by: Cursor <cursoragent@cursor.com> * fix(think): keep continueLastTurn overrides on runTurn; link partials of cancelled submissions Co-authored-by: Cursor <cursoragent@cursor.com> * fix(think): capture output through delegating continueLastTurn overrides; skip unstarted rows when linking aborted partials Co-authored-by: Cursor <cursoragent@cursor.com> * fix(think): scope continuation output capture per runTurn call; keep the override's result fields Co-authored-by: Cursor <cursoragent@cursor.com> * test(think): give cancellation-link tests a stream long enough to cancel under load Co-authored-by: Cursor <cursoragent@cursor.com> --------- Co-authored-by: Sunil Pai <18808+threepointone@users.noreply.github.com> Co-authored-by: Cursor <cursoragent@cursor.com>