v4.5.12
21 Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
 Katia Bulatova
|
480bede0ad |
feat(webapp,sdk): dashboard agent plan enforcement, component gallery — and fixes (#4516)
Plan enforcement for the dashboard agent — message quota and watch limits — plus the component gallery, fixes and test hardening from the same stack (#4548, #4549, #4550, #4552, #4556 merged here). ## Plan enforcement ([TRI-12863](https://linear.app/triggerdotdev/issue/TRI-12863)) **Agent message quota.** The Free-plan allowance becomes a real server-side limit with a durable counter. New `agent_message_usage` table keyed `(organization_id, period)` — deliberately not joined to chats, so deleting a chat can't free quota within the period. Both send paths count one user message (wakes never count) and refuse at the cap with `403 message_quota_reached`, which the client renders as an upgrade panel, never a silent drop. The refusal code is a single shared constant on both sides. **Watch limits.** A watch whose window exceeds the plan's `agentWatchMaxHours`, or that would push the org past its `agentWatchers` count, is refused with `watch_limit_reached` (409 on the API, an upgrade hint on the card). Plan limits only tighten the existing code ceilings (`min(plan, 24h)`, per-chat cap of 3 still applies). A plan limit of zero means zero, not unlimited. Questions answerable instantly are answered before any plan refusal — a one-shot consumes no slot and never sees an upgrade nag. **Fails open by design.** Cloud ships the actual per-plan numbers separately (TRI-12863 P0). Until then absent limits resolve to the unlimited sentinel and the upgrade UI is gated on billing presence — self-hosted sees no cap, no upsell, with tests proving the fallback. Both quotas are nudges, not security boundaries: a failing limit read never blocks a send. ## Component gallery An admin-only gallery of every agent card state: five `storybook.agent-*` pages (chat UI, view blocks, report, investigation, watch) with their shared shell and manifest, demo fixtures, two demo-only cards, toast examples, and the screenshot script. No LLM and no data — every state renders from fixtures under `dashboard-agent/demo/`, never reachable from a production path. Designers and reviewers can look at every state, including the report states, without seeding anything. ## And fixes **SDK: watch-mode chat subscriptions survive quiet windows** (TRI-13065, TRI-13070) — watch mode keeps reconnecting across empty long-poll windows and only stops on abort or a settled session; a passive subscriber can no longer stop a turn it doesn't own (`stopOnAbort` is explicit, default off). Review findings fixed alongside: a superseded stream's async teardown no longer removes the live successor's abort controller or multi-tab claim, and stopping a generation hands the chat back to the user's other tabs. **Query boundary pinned end-to-end** ([TRI-11165](https://linear.app/triggerdotdev/issue/TRI-11165)) — a route-level test drives `api.v1.query` with a real signed environment JWT (writes refused before ClickHouse, a read passes); `readonly=1` made non-overridable; a per-turn cap stops the model burning a turn rewriting a query it can't fix (deterministic SQL errors only — busy/transport rejections don't count). **chat.agent durability regression suite** ([TRI-11166](https://linear.app/triggerdotdev/issue/TRI-11166)) — testcontainers-backed coverage of the two audit criticals (cross-tenant isolation, no duplicate mid-stream turn, both control-broken) plus crash-resume, cursor-based refresh, clean rollback of a mid-write turn failure (torn by a real constraint violation), and OOM-restart replay. **Investigation sweep backoff** — stale investigations get an attempt counter and backoff so a poison row can't pin the sweep queue head (migration `0005`: `sweep_attempts`, `last_sweep_attempt_at`). ## Screenshots <img width="1440" height="791" alt="Screenshot 2026-08-06 at 00 36 19" src="https://github.com/user-attachments/assets/6a68cd42-8580-469d-afe7-e28d1eef18e1" /> 🤖 Generated with [Claude Code](https://claude.com/claude-code) |
||
|
Chris Arderne
|
85f5b37c68 |
chore: upgrade to TypeScript 7 (#4318)
## Summary Upgrade the monorepo to TypeScript 7.0.2 and update package build tooling for compatibility with the native compiler. ## Design Package builds now use `tshy` 4, while the packages still using `tsup` move to `tsdown`. The few scripts that depend on the legacy TypeScript compiler API use an explicit TypeScript 6 alias; declaration portability coverage invokes the TypeScript 7 CLI directly. Turbo is updated so workspace tasks can read the regenerated pnpm lockfile. --------- Co-authored-by: devin-ai-integration[bot] <158243242+devin-ai-integration[bot]@users.noreply.github.com> |
||
|
Matt Aitken
|
be45cf9e61 |
fix(sdk): preserve partial assistant message on chat stream failure (#4348)
🚀 Publish Trigger.dev Docker / typecheck (push) Failing after 5s
🚀 Publish Trigger.dev Docker / units (push) Failing after 5s
🚀 Publish Trigger.dev Docker / publish-webapp (push) Has been skipped
🚀 Publish Trigger.dev Docker / publish-worker-v4 (push) Has been skipped
🚀 Publish Trigger.dev Docker / scan-webapp (push) Has been skipped
🚀 Publish Trigger.dev Docker / scan-supervisor (push) Has been skipped
🦋 Changesets PR / Create Release PR (push) Has been cancelled
🚀 Publish Trigger.dev Docker / 📣 Dispatch main image (push) Has been cancelled
⚒️ Publish Worker (v4) / build (supervisor) (push) Has been cancelled
## Summary When a `chat.agent` (or `chat.createSession`) turn's model stream fails mid-response (e.g. a transport timeout like `UND_ERR_BODY_TIMEOUT`), the assistant output that already streamed was dropped: `onTurnComplete` fired with `responseMessage: undefined`, and the manual loop's `turn.complete()` rethrew without keeping the partial. Apps that register `hydrateMessages` are hit hardest, since boot-time tail-replay recovery is off by design. This preserves the streamed-so-far assistant output while still reporting the turn as errored, so persistence and recovery keep it. ## Scope of behavior change Only the **error path** changes. Successful turns are unaffected: the same chunks stream to the client in the same order, and backpressure/cancel behave as before. Everything here is a correctness improvement on a turn that hit a source-stream failure. ## What it does Follow-up to #4304 (`chat.pipeAndCapture`), extending the same partial-recovery to the two loops that lacked it: - **`chat.agent`**: taps the response stream (via a `TransformStream`, so pass-through backpressure and cancel are preserved) to buffer chunks, and on a source-stream failure reconstructs the partial (preferring the `onFinish` message). It's surfaced on the error-path `onTurnComplete` (`responseMessage`, `rawResponseMessage`, `uiMessages`, `newUIMessages`, `newMessages`) and committed to the accumulator so the next turn and the reboot snapshot keep it. - **`chat.createSession` / `turn.complete()`**: the reconstructed partial is accumulated (so `turn.uiMessages` reflects it and the caller can persist after catching) before `turn.complete()` rethrows. `onBeforeTurnComplete` stays skipped on the error path (it hands out a writer for a stream that has already broken). ## Correctness properties (each covered by a regression test) Each test below was confirmed to fail without its fix: - The recovered partial reaches `onTurnComplete` and the next turn's accumulated messages. - An already-committed (possibly enriched) response is not overwritten if a post-response hook then throws. - Incomplete tool parts are cleaned from the recovered partial (text kept), so the UI and model views agree and the next turn isn't poisoned. - A prior turn's model-only compaction survives an errored turn (append only the new tail, don't reconvert the full history). - A reconstructed fragment that reuses an existing message id does not clobber the complete message. - Queued `chat.response` data parts are folded into the recovered partial, matching the success path. - `newMessages` (model delta) stays symmetric with `newUIMessages`. ## Tests New `chat-agent-source-stream-error.test.ts` covers the cases above. The full `@trigger.dev/sdk` unit suite passes and the package build is green across all supported runtimes (Node 20 to 26, Bun, Deno, Cloudflare Workers). |
||
|
Matt Aitken
|
e2d3b8388c |
feat(sdk): return lastEventId from writeTurnComplete and typed capture result (#4304)
## Summary
Two ergonomic additions for custom chat-agent loops that own the turn
loop (`chat.customAgent`, `chat.createSession`, and the hand-rolled
primitives).
`chat.writeTurnComplete()` now resolves to `{ lastEventId }`, the resume
cursor for the start of the next turn. A custom loop can persist it
straight from the task instead of round-tripping it back from the client
after the turn ends. The value was already produced internally by the
turn-complete write; the public wrapper simply discarded it.
`chat.pipeAndCapture()` no longer throws when a stream is stopped or
fails. It now resolves to a `PipeAndCaptureResult` carrying any partial
`message` captured before the stop or failure, a typed `status`
(`"complete" | "aborted" | "error"`), and the `error` on failure.
Previously a failed stream threw and the partial was lost, and an abort
was captured only when the AI SDK happened to fire `onFinish` in time.
```ts
const { message, status, error } = await chat.pipeAndCapture(result, { signal });
if (message) conversation.addResponse(message);
if (status === "error") logger.error("turn failed", { error });
const { lastEventId } = await chat.writeTurnComplete();
await db.chats.update(chatId, { lastEventId });
```
## Design
`pipeAndCapture` wraps the pipe in a `try/catch` and classifies the
outcome from the abort signal (a stop drains the source stream cleanly
rather than throwing) versus a thrown error. It also races the
`onFinish` capture against a timeout so a hard stop that prevents
`onFinish` from firing can't hang the caller. This mirrors the capture
path `chat.agent` already uses internally.
The `finishReason` from `onFinish` is surfaced too, since it was already
captured on the built-in path.
The internal `turn.complete()` helper keeps its existing contract: it
still returns `UIMessage | undefined`, still throws on a genuine stream
failure, and still discards output on a full run cancel.
## Breaking change
`chat.pipeAndCapture` previously resolved to `UIMessage | undefined`.
Call sites now read `.message` off the result. This is a young,
low-level API; the docs examples are updated in this PR.
|
||
|
Eric Allam
|
9b3a7bd7b2 |
fix(sdk,webapp): stop chat losing a message sent right after an action (#4234)
## Summary Sending a chat message immediately after an action (for example an undo) could make the message's response vanish from the UI. The transport opened a response stream that closed on the *earlier* turn's completion instead of waiting for the send's own turn. The agent still produced and persisted the answer, so it reappeared on refresh. Same "disappearing message" class as [#4176](https://github.com/triggerdotdev/trigger.dev/pull/4176), different cause. ## Fix A send's response stream had no way to tell whether a `turn-complete` belonged to its turn. `POST /realtime/v1/sessions/:id/in/append` now returns the appended record's sequence number, and the transport skips any turn-complete whose `session-in-event-id` (the agent's committed `.in` cursor) is below that seq, closing only on its own turn. Older webapps omit the seq, in which case the transport falls back to the previous behavior, so the SDK and server can ship independently. Because the fix spans the SDK and the server, both a webapp deploy and an SDK release are needed for the full effect. Verified end to end with the ai-chat reference app: undo-then-immediate-send loses the follow-up's answer before the fix and streams it inline after, with a revert-the-guard run reproducing the loss on the same script. Unit tests cover the skip and the no-seq fallback. |
||
|
Eric Allam
|
25254d0201 |
fix(sdk): make inferred chat agent types portable for declaration emit (#4218)
## Summary
Exporting a `chat.agent` from a project with `declaration: true` failed
with TS2742: the inferred type of the agent references
`ChatTaskWirePayload`, which was declared in an internal module not
reachable through the package exports map, so tsc could only name it via
a file path into `node_modules` and refused to emit. Consumers had to
hand-mirror the wire type and annotate their export.
## Fix
`ChatTaskWirePayload` and `ChatInputChunk` are now declared in
`@trigger.dev/sdk/chat` (a public subpath) and re-exported type-only
from the internal shared module, so every internal import is unchanged
and the browser/server module split is untouched. Declaration emit for
an inferred agent type now produces a portable specifier:
```ts
export declare const chatAgent: Task<"chat-agent", import("@trigger.dev/sdk/chat").ChatTaskWirePayload<MyUIMessage, MyClientData>, unknown>;
```
As a side effect the wire types are now directly importable, which is
what affected users were reconstructing by hand.
## Verification
Reproduced against the built 4.5.2-equivalent package: a consumer
fixture with declaration emit produced `import("<file
path>/ai-shared.js")` in its declaration (the TS2742 trigger); after the
fix the same fixture emits the public specifier with zero diagnostics. A
regression test now builds that consumer simulation in a temp directory
on every test run: it copies the built package into a fake node_modules
(copied, not symlinked, because tsc only applies exports-map naming to
real node_modules paths), compiles the fixture with the TypeScript API,
and asserts no errors, no relative-path imports, and no internal module
references in the emit.
|
||
|
Eric Allam
|
fbd86b6ee9 |
feat(sdk): onEvent observability callback on the chat transport (#4187)
## Summary
`sendMessage` from `useChat` gives no feedback about whether a message
actually reached the backend, and the `fetch` override is wire-level: it
requires knowing endpoint semantics, cannot attribute requests to
messages, and misses the headStart first-turn POST entirely. This adds a
typed `onEvent` observability callback to `TriggerChatTransport` /
`useTriggerChatTransport` so send-success metrics, time-to-first-token,
and "sent but never answered" watchdogs become a few lines of client
code.
## Example
```ts
const transport = useTriggerChatTransport({
task: "my-chat",
accessToken: ({ chatId }) => mintChatAccessToken(chatId),
onEvent: (event) => {
switch (event.type) {
case "message-sent":
// Durably acknowledged by the session's input stream, not just "request accepted".
metrics.increment("chat.message_sent", { source: event.source });
metrics.timing("chat.send_duration_ms", event.durationMs);
break;
case "message-send-failed":
metrics.increment("chat.message_send_failed", { status: event.status });
break;
case "first-chunk":
metrics.timing("chat.ttft_ms", event.sinceSendMs ?? 0);
break;
case "turn-completed":
metrics.timing("chat.turn_duration_ms", event.sinceSendMs ?? 0);
break;
}
},
});
```
## Design
One callback, one discriminated union (`ChatTransportEvent`):
- `message-sent` / `message-send-failed`: terminal send outcomes with
`messageId`, a `source` discriminator (submit, regenerate, steer,
action, stop, head-start), `durationMs`, `bodyBytes`, the append's
idempotency key (`partId`, also stored on the server-side record), and
error + HTTP status on failure. `message-sent` means the append was
durably acknowledged, after any internal token-refresh retries.
- `stream-connected` (with a `resumed` flag and the cursor it connected
from), `first-chunk` (chunk type plus `sinceSendMs` for
time-to-first-token), `turn-completed` (`sinceSendMs` full-turn latency
and the agent's committed input cursor), and `stream-error` follow the
response side, so a send can be paired with the answer that should
follow it. `messageId` on response events is client-side attribution
from the last turn-producing send on that chat.
Emissions sit at the transport's existing choke points, covering every
send path uniformly (including steering and headStart, which the fetch
override cannot observe). Exceptions thrown by the callback are
swallowed: observability can never break the chat. The React hook keeps
the callback live across renders instead of freezing the first-render
closure.
## Verification
Unit tests drive the transport directly with the `fetch` override as the
network stub (send success/failure per source, stream lifecycle, resumed
flag, field enrichment, callback exceptions swallowed). Verified
end-to-end against a realistic metrics setup in the ai-chat reference
app (counters, send-duration and TTFT histograms, and both watchdogs
built purely on these events): a healthy two-turn chat produces exactly
the expected event sequence and TTFT values; an oversized append records
`message_send_failed` with status 413; and killing the worker after a
durable send fires both `sent_but_no_stream` and `sent_but_unanswered`,
reproducing and detecting the "message disappeared" failure mode that
motivated this feature.
|
||
|
Eric Allam
|
add0a7da0a |
fix(sdk,core): stop chat sessions dropping messages that arrive during a turn (#4176)
## Summary Sending a message to a chat whose run had ended could make the message vanish: the continuation run replayed already-answered messages, never processed the new one, and a page refresh lost it entirely. Chasing that report surfaced four composing message-loss bugs in the chat session runtime; this PR fixes all of them, each with a regression test. ## The fixes 1. **Stale resume cursor.** Records delivered while a run was suspended (the waitpoint path) advanced the SSE resume counter but not the committed-consume cursor, so the `session-in-event-id` header stamped on turn-completes went stale by one record per suspended turn. Continuation boots seed from that header, which is what made them replay already-processed messages. `session.in.wait()` now advances both cursors. 2. **Only the first buffered message dispatched.** Messages arriving during a turn are consumed into a buffer whose end-of-turn pickup dispatched only the first entry; the buffer was recreated each turn, so the rest were discarded, and since consuming a record commits the cursor the loss was permanent. A continuation boot's replay delivers several records back-to-back, which put the user's new message at index 1 or later. The buffer now outlives the turn and drains one message per turn in both `chat.agent` and `chat.createSession` (whose equivalent buffer was never read at all). 3. **Post-stop window in `chat.createSession`.** The turn's message listener stayed attached through the stopped turn's post-stream work, so a message sent shortly after stopping a turn was consumed into the dead steering queue and lost. The listener now detaches when the stream settles, matching the `chat.agent` loop. 4. **Handler leak on errored turns.** A turn that threw outside the streaming section (for example from an `onTurnStart` hook) leaked its message listener. Previously that silently lost mid-turn messages; with the loop-level buffer it would have duplicated them instead. The subscription handle is now detached by the turn's catch/finally, and `chat.createSession` defensively detaches its prior turn's listener when user code exits a turn without `complete()`/`done()`. ## Verification Reproduced end-to-end with the ai-chat reference project before the fix (message consumed but never answered, two replayed turns, gone on refresh) and verified after (single clean turn, survives refresh, turn-complete cursors strictly advancing). Regression tests in `packages/trigger-sdk/test/pending-message-drain.test.ts` cover all four, each verified red against the unfixed behavior. A smoke sweep of the standard chat scenarios (basic send, multi-turn, suspend/resume, mid-stream refresh, stop, steering, cancel + continue, and the `createSession` variant) passes on the final branch state. |
||
|
Chris Arderne
|
c7861be520 |
chore: activate no-unused-vars and import linters (#4096)
Once this is merged, oxlint is at a pretty sensible baseline. **Enable `no-unused-vars`, `typescript/consistent-type-imports`, and `import/no-duplicates` lint rules** Turns on three previously-disabled oxlint rules across the monorepo and fixes all violations: - **`no-unused-vars`** – enabled as an error with standard ignore patterns: unused function arguments are ignored by default (`args: "none"`), variables/caught errors/destructured array elements prefixed with `_` are allowed, and rest siblings are permitted. - **`typescript/consistent-type-imports`** – enforced as an error; all type-only imports now use the `import type` syntax. - **`import/no-duplicates`** – enforced as an error; duplicate import statements from the same module have been merged. The remaining commits clean up the violations found across the codebase: removing unused variables/imports/type aliases, adding `_` prefixes to intentionally unused bindings, fixing duplicate imports, and converting value imports to `import type` where appropriate. |
||
|
Chris Arderne
|
bfa902bd18 |
chore: enable more linters (#4080)
Re-enables ~15 oxlint rules that were blanket-disabled before. |
||
|
Chris Arderne
|
b54201f986 | chore: switch to oxfmt, oxlint - add ci checks (#3977) | ||
|
Eric Allam
|
17482c0577 |
feat(sdk): chat.headStart handover for customAgent and createSession (#3963)
## Summary
`chat.headStart` (the warm step-1 fast path) previously handed its
response over only to `chat.agent`. This extends handover to the other
two backends: `chat.customAgent` consumes it with
`conversation.consumeHandover({ payload })` on turn 0, and
`chat.createSession` surfaces it as `turn.handover` (call
`turn.complete()` with no source to finalize a pure-text handover). The
low-level `chat.waitForHandover()` and `accumulator.applyHandover()` are
exported for hand-rolled loops.
It also adds `triggerConfig` to `chat.headStart()` and
`chat.openSession()`, so the auto-triggered handover-prepare run
inherits tags, queue, machine, and the other session run options the
same way `chat.createStartSessionAction()` does. The `chat:{chatId}` tag
is prepended automatically. Because the session is created once on the
first head-start turn (idempotent on the chat id), this is the only
place those options can be set for a head-start chat's lifetime.
## Fix: tool-call resume
When the warm step-1 hands over a pending tool call (rather than pure
text), the agent loop resumes that tool round. For it to merge cleanly
the pipe threads the spliced partial as `originalMessages`, so the
resumed tool-output chunk attaches to the handed-over tool-call instead
of throwing `No tool invocation found`. `MessageAccumulator.addResponse`
now also dedups by id (replace-in-place), so the persisted history
doesn't carry a duplicate assistant message when the resumed response
reuses the partial's id.
Incorporates the `triggerConfig` work from
[#3933](https://github.com/triggerdotdev/trigger.dev/pull/3933) by
@saasjesus, with `createStartSessionAction` extended to also forward
`maxDuration`, `region`, and `lockToVersion` so the two session entry
points stay consistent.
Verified end-to-end against a local environment: handover (pure-text and
tool-call) on both new backends, a `chat.agent` regression pass, and
`triggerConfig` tags and queue landing on the run.
---------
Co-authored-by: saasjesus <armin@chatarmin.com>
|
||
|
Eric Allam
|
3b919994c1 |
feat(sdk): make the chat.agent system prompt cacheable (#3952)
## Summary
`chat.agent`'s system prompt (the `chat.prompt` text plus any skills
preamble) could not carry a provider cache breakpoint, so the largest
and most stable part of the prompt re-paid full input price on every
turn. `chat.toStreamTextOptions()` now emits the system prompt as a
structured message carrying `providerOptions` when you opt in, so a
provider can cache the system block. Without an option, `system` stays a
plain string, so existing behavior is unchanged.
## API
Three ways to opt in (most specific wins, no deep merge):
```ts
// Anthropic sugar
chat.toStreamTextOptions({ cacheControl: { type: "ephemeral" } });
// provider-agnostic (also covers Amazon Bedrock's cachePoint)
chat.toStreamTextOptions({ systemProviderOptions: { anthropic: { cacheControl: { type: "ephemeral" } } } });
// at the definition site
chat.prompt.set(SYSTEM_PROMPT, { providerOptions: { anthropic: { cacheControl: { type: "ephemeral" } } } });
```
The `cacheControl` shorthand is Anthropic-only; `systemProviderOptions`
is the general form. Pairs with a `prepareMessages` cache breakpoint to
cache the conversation prefix too.
Docs guide: https://github.com/triggerdotdev/trigger.dev/pull/3951
|
||
|
Eric Allam
|
2b6d2492fe |
fix(sdk,core): head-start handover correctness and continuation boot latency (#3907)
## Summary Three related fixes for `chat.headStart` and continuation boots, found while investigating customer reports. **1. `chat.headStart` now works with `hydrateMessages`.** The turn-0 handover splice only ran on the default accumulation path, so agents registering `hydrateMessages` silently lost the warm route's step-1 response: pure-text turns fired `onTurnComplete` with no assistant message (and an empty durable write), tool-call turns re-ran step 1 from scratch under a fresh `messageId`, and the head-start user message never reached the hydrate hook at all. The first-turn history now reaches `hydrateMessages` as `incomingMessages`, and the splice runs after both accumulation branches, deduplicated by the handover `messageId`. **2. Reasoning parts survive the handover.** The synthesized partial only mapped text and tool-call parts, so an extended-thinking model's step-1 reasoning streamed to the browser but never reached durable history. Reasoning parts now map through with provider metadata, so Anthropic thinking signatures survive a UIMessage round trip on hydrate replays. **3. Continuation boots no longer stall for ~10 seconds.** The `.in` resume cursor was found by draining an SSE subscription that only closes after its 5 second inactivity window, and the scan ran twice per boot. It is now a non-blocking records read of the latest turn-complete header, runs at most once per boot, the boot reads run concurrently, and chat snapshots carry the cursor so subsequent boots skip the scan entirely. Measured locally on a cancel-then-continue repro: pre-turn continuation latency dropped from ~11s to ~0.5s. Every fix was verified red-green: new unit tests reproduced each failure before the fix, and end-to-end smoke tests against a live local stack covered both handover legs, reasoning persistence with extended thinking (including a follow-up turn that round-trips the persisted signed reasoning back to the provider), and the boot timing comparison. ## Rollout SDK-only; no server change required. A new SDK against a server that does not serialize record headers degrades to the existing no-cursor fallback. Old SDKs ignore the new snapshot field, and new SDKs fall back to the records scan on snapshots written before it existed. |
||
|
Eric Allam
|
f5f29ceb26 |
fix(sdk,core): chat.agent delivery, idempotency, and recovery fixes (#3891)
## Summary A batch of reliability fixes for `chat.agent`: - A user message sent while the agent is streaming is no longer delivered twice (which could run a duplicate turn). - Input appends carry an idempotency key (`X-Part-Id`) so a retried send can't duplicate a message. - `onTurnComplete` now fires on errored turns with the thrown error attached, and the failed turn's user message is persisted so it isn't lost on the next run. - Stopping a generation clears the streaming state, so a page reload doesn't replay the stopped turn. - Custom agents and manual `chat.writeTurnComplete` callers trim the output stream, sending a custom action no longer leaves a second stream reader running, a long-lived `watch` subscription no longer grows its dedupe set without bound, promoting a queued message to steering no longer risks a double-send, and runs keep the full set of dashboard tags. The `X-Part-Id` header is accepted by current servers (they just don't dedupe on it yet), so this is safe to ship ahead of the matching server change. |
||
|
Eric Allam
|
e9e2ec1cfc |
fix(sdk): re-apply tool toModelOutput across chat.agent turns (#3790)
## Summary
`chat.agent` now takes a `tools` option. Until now tools only went to
`streamText` inside `run()`, so the SDK had no tools when it
re-converted the persisted `UIMessage` history at the start of each
turn. Any tool with a `toModelOutput` (raw image bytes into an image
content part, or a sub-agent transcript compressed to a summary) had its
transform applied on turn 1 and skipped from turn 2 onward, so the raw
output got JSON-stringified back into the prompt and the model lost the
transformed view.
Declaring `tools` on the config threads them into that conversion, so
`toModelOutput` runs on every turn. The resolved set is handed back,
typed, on the `run()` payload as `tools`:
```ts
const tools = { searchDocs, renderChart };
export const myChat = chat.agent({
tools,
run: async ({ messages, tools, signal }) =>
streamText({ ...chat.toStreamTextOptions({ tools }), messages, abortSignal: signal }),
});
```
`tools` also accepts a per-turn function for tools that depend on the
user or a feature flag. Only `inputSchema` and `toModelOutput` are read
during conversion, never `execute`. Also exports
`InferChatUIMessageFromTools<typeof tools>` to derive the chat
`UIMessage` type from a tool set. No behavior change for agents that
don't declare `tools`.
|
||
|
Eric Allam
|
75679c7518 |
fix(sdk): chat HITL continuations no longer break the next LLM call (#3719)
## Summary
Multi-step reasoning agents with HITL tools (OpenAI Responses with
`store: false`, Anthropic extended thinking, etc.) failed on
`chat.addToolOutput(...)` continuations — either the wire payload blew
the `.in/append` cap (reasoning blobs + tool inputs routinely > 512
KiB), or app-side slimming workarounds got overwritten server-side and
the next LLM call landed a tool call with no `arguments`. Both modes are
fixed.
## Design
The per-turn merge in `chat.agent` now overlays only the tool-part state
advances (`output-available` / `output-error` / `approval-responded` /
`output-denied`) from the wire copy onto the hydrated/snapshot chain.
Previously it replaced the entire message, which dropped `input`,
reasoning, and text from the LLM's view whenever the wire was slim.
In parallel, `TriggerChatTransport.sendMessages` and `AgentChat.sendRaw`
now slim the assistant message themselves on `submit-message`
continuations: ship `{ id, role, parts: [<resolved tool part only>] }`,
everything else reconstructed server-side from `hydrateMessages` or the
durable snapshot. Continuation payloads drop from 600 KiB – 1 MiB to ~1
KiB.
`references/ai-chat` `aiChatHydrated.hydrateMessages` now upserts by id
instead of pushing. With slim continuations, a blind push duplicates the
assistant id in the returned chain — the merge updates the first match,
the slim duplicate goes straight to `toModelMessages` with no `input`,
and the LLM 4xx's. This is the canonical pattern customers should mirror
in their own hydrate implementations.
## Test plan
- 11 new tests (slim helper unit + slim+merge integration for HITL,
approval, default no-hydrate branch)
- Full SDK suite: 239 tests pass across 19 files
- End-to-end sweep against `references/ai-chat`: 19 customer-side smoke
tests green; HITL wire bodies confirmed at ~1 KiB (was 600 KiB+); no
provider 4xx errors across OpenAI Responses or Anthropic
|
||
|
Eric Allam
|
d343727021 |
fix(webapp,sdk): keep chat.agent snapshots on one object store (#3679)
(`OBJECT_STORE_BASE_URL`) and a named protocol provider
(`OBJECT_STORE_DEFAULT_PROTOCOL=s3`), chat.agent session snapshot writes
landed in the named provider but reads fell through to the default — so
the recovery boot couldn't find the snapshot it had just written.
After a mid-stream cancel, the missing snapshot triggered a fallback
replay path that dropped the user's follow-up message, leaving the chat
stuck in `submitted` indefinitely.
Fix:
- New `/api/v1/sessions/:id/snapshot-url` route handles PUT + GET
symmetrically — both prefix unprefixed keys with
`OBJECT_STORE_DEFAULT_PROTOCOL` so they always round-trip through the
same store.
- `Session.chatSnapshotStoragePath` persists the resolved URI on first
write so future protocol changes don't strand existing snapshots.
Reads prefer the stored URI and fall back to the computed default for
pre-column sessions.
- SDK calls `createChatSnapshotUploadUrl` / `getChatSnapshotUrl`; the
generic v1/v2 packets endpoints are unchanged.
## Test plan
- [x] Configure local with two providers (R2 default + MinIO `s3` named)
and `OBJECT_STORE_DEFAULT_PROTOCOL=s3`.
- [x] Reproduce hang: send a message, cancel mid-stream, send another —
without the fix it hangs in `submitted`; with the fix it streams.
- [x] Snapshot lands in the `s3`-protocol bucket and
`Session.chatSnapshotStoragePath` is set after first write.
- [x] SDK unit tests pass; webapp typecheck passes.
|
||
|
Eric Allam
|
f91b96efb7 | feat(sdk,core): preserve chat.agent context after cancel / OOM / crash (#3671) | ||
|
Eric Allam
|
82853debea |
feat(webapp,core,sdk,cli): bound session.out via per-turn trim (#3644)
## Summary Long-running chat agents were filling `session.out` forever — every `chat.agent` turn appended to the same S2 stream with no trim, and the Sessions dashboard re-streamed the entire history from `seq_num=0` on every page load. After this change the agent appends an S2 `trim` command record after each `trigger:turn-complete`, pointing back at the previous turn-complete's seq_num. `session.out` stays roughly one turn long at steady state, regardless of session age. `trigger:turn-complete` and `trigger:upgrade-required` move from `chunk.type`-shaped data records into header-form control records under a uniform `trigger-control` namespace. Built-in transports (`TriggerChatTransport`, `AgentChat`, the dashboard's `AgentView`) handle the new shape transparently. Custom transports need a one-line filter on the `trigger-control` header — see the rewritten "Records on session.out" section in the client-protocol docs. The Sessions detail page in the dashboard fetches the agent's per-turn S3 snapshot via a presigned URL and seeds the transcript view, then SSE-tails from the snapshot's `lastOutEventId`. Bandwidth and time-to-first-render scale with unread turns instead of session lifetime. Resume contract is now explicit: single-turn-boundary resume always works (the prior turn-complete is still on the stream), the S2 trim is eventually consistent over 10-60s, and multi-turn-away resume falls back to a snapshot reload. |
||
|
Eric Allam
|
16720a5e62 |
feat(sdk): chat.agent — runtime + browser transport
Adds the chat.agent({...}) task definition (server runtime) and the
browser-side TriggerChatTransport + AgentChat that drives it from a
React or Next.js app. The runtime sits on top of the Sessions primitive
and handles the durable conversational task lifecycle.
Server runtime:
- chat.agent({...}) — session-aware task definition
- Lifecycle hooks: onChatStart, onTurnStart, onTurnComplete, onAction,
onValidateMessages, hydrateMessages
- chat.history read primitives for HITL flows
- chat.local, chat.headStart, chat.handover, oomMachine
- Delta-only wire + S3 snapshot reconstruction at run boot
- Actions are no longer turns
Browser transport:
- TriggerChatTransport (ai-sdk Transport): delta-only wire sends,
SSE reconnection with lastEventId resume, stop/abort cleanup,
dynamic accessToken refresh
- AgentChat: direct programmatic API
- useTriggerChatTransport (React hook)
- chat-tab-coordinator: cross-tab leader election
Includes the chat-agent, chat-agent-delta-wire-snapshots,
chat-history-read-primitives, chat-head-start, chat-actions-no-turn,
chat-session-attributes, agent-skills, and mock-chat-agent-test-harness
changesets.
|