@trigger.dev/sdk 4.8.5 → 4.8.7
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/dist/commonjs/v3/ai.d.ts +34 -8
- package/dist/commonjs/v3/ai.js +304 -168
- package/dist/commonjs/v3/ai.js.map +1 -1
- package/dist/commonjs/v3/auth.d.ts +4 -4
- package/dist/commonjs/v3/auth.js +21 -75
- package/dist/commonjs/v3/auth.js.map +1 -1
- package/dist/commonjs/v3/chat-client.d.ts +1 -1
- package/dist/commonjs/v3/chat.d.ts +1 -1
- package/dist/commonjs/v3/prompt.js +6 -8
- package/dist/commonjs/v3/prompt.js.map +1 -1
- package/dist/commonjs/v3/runs.d.ts +5 -1
- package/dist/commonjs/v3/runs.js.map +1 -1
- package/dist/commonjs/v3/sessions.js +10 -7
- package/dist/commonjs/v3/sessions.js.map +1 -1
- package/dist/commonjs/v3/shared.d.ts +2 -2
- package/dist/commonjs/v3/shared.js +16 -15
- package/dist/commonjs/v3/shared.js.map +1 -1
- package/dist/commonjs/v3/steeringContext.d.ts +35 -10
- package/dist/commonjs/v3/steeringContext.js +86 -14
- package/dist/commonjs/v3/steeringContext.js.map +1 -1
- package/dist/commonjs/v3/streams.js +10 -8
- package/dist/commonjs/v3/streams.js.map +1 -1
- package/dist/commonjs/v3/tasks.d.ts +2 -2
- package/dist/commonjs/v3/tasks.js.map +1 -1
- package/dist/commonjs/v3/test/mock-chat-agent.js +24 -13
- package/dist/commonjs/v3/test/mock-chat-agent.js.map +1 -1
- package/dist/commonjs/v3/toolOutputJson.d.ts +17 -0
- package/dist/commonjs/v3/toolOutputJson.js +241 -0
- package/dist/commonjs/v3/toolOutputJson.js.map +1 -0
- package/dist/commonjs/v3/webhooks.d.ts +8 -2
- package/dist/commonjs/v3/webhooks.js +20 -11
- package/dist/commonjs/v3/webhooks.js.map +1 -1
- package/dist/commonjs/version.js +1 -1
- package/dist/esm/v3/ai.d.ts +34 -8
- package/dist/esm/v3/ai.js +306 -171
- package/dist/esm/v3/ai.js.map +1 -1
- package/dist/esm/v3/auth.d.ts +4 -4
- package/dist/esm/v3/auth.js +22 -76
- package/dist/esm/v3/auth.js.map +1 -1
- package/dist/esm/v3/chat-client.d.ts +1 -1
- package/dist/esm/v3/chat.d.ts +1 -1
- package/dist/esm/v3/prompt.js +6 -8
- package/dist/esm/v3/prompt.js.map +1 -1
- package/dist/esm/v3/runs.d.ts +5 -1
- package/dist/esm/v3/runs.js.map +1 -1
- package/dist/esm/v3/sessions.js +10 -7
- package/dist/esm/v3/sessions.js.map +1 -1
- package/dist/esm/v3/shared.d.ts +2 -2
- package/dist/esm/v3/shared.js +16 -15
- package/dist/esm/v3/shared.js.map +1 -1
- package/dist/esm/v3/steeringContext.d.ts +35 -10
- package/dist/esm/v3/steeringContext.js +85 -14
- package/dist/esm/v3/steeringContext.js.map +1 -1
- package/dist/esm/v3/streams.js +10 -8
- package/dist/esm/v3/streams.js.map +1 -1
- package/dist/esm/v3/tasks.d.ts +2 -2
- package/dist/esm/v3/tasks.js.map +1 -1
- package/dist/esm/v3/test/mock-chat-agent.js +25 -14
- package/dist/esm/v3/test/mock-chat-agent.js.map +1 -1
- package/dist/esm/v3/toolOutputJson.d.ts +17 -0
- package/dist/esm/v3/toolOutputJson.js +238 -0
- package/dist/esm/v3/toolOutputJson.js.map +1 -0
- package/dist/esm/v3/webhooks.d.ts +8 -2
- package/dist/esm/v3/webhooks.js +20 -11
- package/dist/esm/v3/webhooks.js.map +1 -1
- package/dist/esm/version.js +1 -1
- package/docs/ai/observability.mdx +25 -19
- package/docs/ai/prompts.mdx +25 -19
- package/docs/ai-chat/actions.mdx +1 -1
- package/docs/ai-chat/anatomy.mdx +1 -1
- package/docs/ai-chat/backend.mdx +24 -24
- package/docs/ai-chat/background-injection.mdx +5 -5
- package/docs/ai-chat/chat-local.mdx +2 -2
- package/docs/ai-chat/compaction.mdx +47 -18
- package/docs/ai-chat/custom-agents.mdx +4 -4
- package/docs/ai-chat/error-handling.mdx +4 -4
- package/docs/ai-chat/fast-starts.mdx +8 -7
- package/docs/ai-chat/frontend.mdx +6 -2
- package/docs/ai-chat/lifecycle-hooks.mdx +11 -11
- package/docs/ai-chat/migrating-from-a-route-handler.mdx +4 -4
- package/docs/ai-chat/migrating-from-hydrate-messages.mdx +1 -1
- package/docs/ai-chat/overview.mdx +1 -1
- package/docs/ai-chat/patterns/branching-conversations.mdx +1 -1
- package/docs/ai-chat/patterns/database-persistence.mdx +2 -2
- package/docs/ai-chat/patterns/human-in-the-loop.mdx +1 -1
- package/docs/ai-chat/patterns/native-compaction.mdx +29 -9
- package/docs/ai-chat/patterns/persistence-and-replay.mdx +1 -1
- package/docs/ai-chat/patterns/redacting-telemetry.mdx +2 -2
- package/docs/ai-chat/patterns/skills.mdx +2 -2
- package/docs/ai-chat/patterns/sub-agents.mdx +1 -1
- package/docs/ai-chat/patterns/tool-result-auditing.mdx +1 -1
- package/docs/ai-chat/patterns/trusted-edge-signals.mdx +2 -2
- package/docs/ai-chat/patterns/version-upgrades.mdx +2 -2
- package/docs/ai-chat/pending-messages.mdx +5 -5
- package/docs/ai-chat/prompt-caching.mdx +7 -7
- package/docs/ai-chat/quick-start.mdx +2 -2
- package/docs/ai-chat/testing.mdx +9 -9
- package/docs/ai-chat/tools.mdx +9 -5
- package/docs/ai-chat/transcript-storage.mdx +2 -2
- package/docs/ai-chat/types.mdx +10 -3
- package/docs/apikeys.mdx +2 -1
- package/docs/cli-login-commands.mdx +12 -0
- package/docs/concurrency.mdx +40 -0
- package/docs/database-connections.mdx +1 -1
- package/docs/deploy-environment-variables.mdx +84 -2
- package/docs/deployment/trusted-publishing.mdx +4 -0
- package/docs/errors-retrying.mdx +25 -18
- package/docs/github-actions.mdx +1 -1
- package/docs/github-integration.mdx +32 -1
- package/docs/how-it-works.mdx +4 -2
- package/docs/introduction.mdx +1 -1
- package/docs/mcp-introduction.mdx +7 -1
- package/docs/mcp-remote.mdx +90 -0
- package/docs/oidc/azure.mdx +81 -0
- package/docs/oidc/databricks.mdx +99 -0
- package/docs/oidc/doppler.mdx +60 -0
- package/docs/oidc/gcp.mdx +109 -0
- package/docs/oidc/hugging-face.mdx +90 -0
- package/docs/oidc/openrouter.mdx +91 -0
- package/docs/oidc/snowflake.mdx +91 -0
- package/docs/oidc/vault.mdx +71 -0
- package/docs/oidc.mdx +2 -0
- package/docs/runs/metadata.mdx +1 -1
- package/docs/self-hosting/docker.mdx +2 -0
- package/docs/self-hosting/env/webapp.mdx +4 -1
- package/docs/self-hosting/kubernetes.mdx +2 -0
- package/docs/tasks/schemaTask.mdx +26 -13
- package/docs/tasks/streams.mdx +2 -2
- package/docs/troubleshooting.mdx +7 -0
- package/docs/webhooks/channels.mdx +6 -6
- package/docs/webhooks/connect.mdx +9 -75
- package/docs/webhooks/custom-providers.mdx +87 -0
- package/docs/webhooks/deliveries.mdx +7 -7
- package/docs/webhooks/filters.mdx +2 -2
- package/docs/webhooks/how-it-works.mdx +58 -0
- package/docs/webhooks/human-in-the-loop.mdx +3 -3
- package/docs/webhooks/linear.mdx +5 -3
- package/docs/webhooks/overview.mdx +20 -71
- package/docs/webhooks/own-handler.mdx +62 -0
- package/docs/webhooks/pricing-and-limits.mdx +13 -0
- package/docs/webhooks/providers/attio.mdx +38 -0
- package/docs/webhooks/providers/autumn.mdx +40 -0
- package/docs/webhooks/providers/cal-com.mdx +41 -0
- package/docs/webhooks/providers/calendly.mdx +41 -0
- package/docs/webhooks/providers/clerk.mdx +40 -0
- package/docs/webhooks/providers/discord.mdx +45 -0
- package/docs/webhooks/providers/docusign.mdx +41 -0
- package/docs/webhooks/providers/elevenlabs.mdx +41 -0
- package/docs/webhooks/providers/firecrawl.mdx +54 -0
- package/docs/webhooks/providers/github.mdx +46 -0
- package/docs/webhooks/providers/gitlab.mdx +41 -0
- package/docs/webhooks/providers/granola.mdx +44 -0
- package/docs/webhooks/providers/hookdeck.mdx +98 -0
- package/docs/webhooks/providers/infisical.mdx +41 -0
- package/docs/webhooks/providers/jira.mdx +41 -0
- package/docs/webhooks/providers/langfuse.mdx +40 -0
- package/docs/webhooks/providers/linear.mdx +45 -0
- package/docs/webhooks/providers/loops.mdx +40 -0
- package/docs/webhooks/providers/metronome.mdx +42 -0
- package/docs/webhooks/providers/mux.mdx +40 -0
- package/docs/webhooks/providers/openai.mdx +40 -0
- package/docs/webhooks/providers/pagerduty.mdx +40 -0
- package/docs/webhooks/providers/plain.mdx +41 -0
- package/docs/webhooks/providers/planetscale.mdx +40 -0
- package/docs/webhooks/providers/recall-ai.mdx +54 -0
- package/docs/webhooks/providers/replicate.mdx +41 -0
- package/docs/webhooks/providers/resend.mdx +44 -0
- package/docs/webhooks/providers/retell.mdx +41 -0
- package/docs/webhooks/providers/sentry.mdx +41 -0
- package/docs/webhooks/providers/shopify.mdx +49 -0
- package/docs/webhooks/providers/slack.mdx +48 -0
- package/docs/webhooks/providers/square.mdx +40 -0
- package/docs/webhooks/providers/stripe.mdx +86 -0
- package/docs/webhooks/providers/supabase.mdx +41 -0
- package/docs/webhooks/providers/tailscale.mdx +43 -0
- package/docs/webhooks/providers/typeform.mdx +42 -0
- package/docs/webhooks/providers/vapi.mdx +47 -0
- package/docs/webhooks/providers/vercel.mdx +40 -0
- package/docs/webhooks/providers/whatsapp.mdx +45 -0
- package/docs/webhooks/providers/workos.mdx +41 -0
- package/docs/webhooks/providers/zendesk.mdx +41 -0
- package/docs/webhooks/providers.mdx +53 -0
- package/docs/webhooks/quick-start.mdx +131 -0
- package/docs/webhooks/session-routing.mdx +6 -6
- package/docs/webhooks/signing-secrets.mdx +80 -0
- package/docs/webhooks/slack.mdx +11 -2
- package/docs/webhooks/verification.mdx +139 -0
- package/docs/webhooks/waiters.mdx +2 -2
- package/docs/webhooks/webhook-tasks.mdx +46 -0
- package/package.json +4 -5
- package/docs/migrating-from-v3.mdx +0 -332
- package/docs/webhooks/public-beta.mdx +0 -70
- package/docs/webhooks/sources.mdx +0 -252
package/dist/esm/v3/ai.js
CHANGED
|
@@ -1,4 +1,4 @@
|
|
|
1
|
-
import { accessoryAttributes, apiClientManager,
|
|
1
|
+
import { accessoryAttributes, apiClientManager, canFallbackToLocalPublicToken, generateJWT, controlSubtype, getSchemaParseFn, headerValue, InputStreamOncePromise, isSchemaZodEsque, logger, ManualWaitpointPromise, OutOfMemoryError, resourceCatalog, SemanticInternalAttributes, SESSION_IN_CONSUMED_ID_HEADER, SESSION_IN_EVENT_ID_HEADER, sessionStreams, SessionChannelRouter, InputStreamTimeoutError, taskContext, TRIGGER_CONTROL_SUBTYPE, SESSION_CLOSED_HEADER, SESSION_CLOSED_REASON_HEADER, tryCatch, } from "@trigger.dev/core/v3";
|
|
2
2
|
import { chatEvent, normalizeKeyString } from "./webhooks.js";
|
|
3
3
|
import { fetchPublicFile } from "./modelFileDownload.js";
|
|
4
4
|
// Runtime VALUES go through the ESM/CJS shim so the CJS build can `require`
|
|
@@ -13,7 +13,7 @@ import { withToolResultsInCallOrder } from "./toolResultOrder.js";
|
|
|
13
13
|
import { ManagedChatResponse, createOrderedChatWriter } from "./managedChatResponse.js";
|
|
14
14
|
import { reduceUIMessageChunks } from "./uiMessageChunks.js";
|
|
15
15
|
import { forwardSubAgentChunks } from "./subAgentStream.js";
|
|
16
|
-
import { convertSteeredMessages, retainStepMessages, steeringMarkers, withoutInjectionMarkers, } from "./steeringContext.js";
|
|
16
|
+
import { convertSteeredMessages, generationResponseMessages, retainStepMessages, steeringMarkers, withoutInjectionMarkers, } from "./steeringContext.js";
|
|
17
17
|
let transcriptStorageOverride;
|
|
18
18
|
/**
|
|
19
19
|
* Test-only override for the storage `chat.agent` persists through, so a
|
|
@@ -41,6 +41,7 @@ import { withResolvedExternalDeploymentId } from "./externalDeploymentId.js";
|
|
|
41
41
|
import { resolvePinToFollow } from "./chatVersionSkew.js";
|
|
42
42
|
import { sessions, } from "./sessions.js";
|
|
43
43
|
import { createTask } from "./shared.js";
|
|
44
|
+
import { withJsonToolResults } from "./toolOutputJson.js";
|
|
44
45
|
import { triggerConcurrencyBody } from "./concurrency-shared.js";
|
|
45
46
|
import { markChatAgentRunForStreamsWarning } from "./streams.js";
|
|
46
47
|
import { tracer } from "./tracer.js";
|
|
@@ -57,12 +58,21 @@ function toModelMessages(messages, context) {
|
|
|
57
58
|
// conditional spread keeps the options object byte-identical to the no-tools
|
|
58
59
|
// path when nothing was declared.
|
|
59
60
|
const tools = locals.get(chatResolvedToolsKey);
|
|
60
|
-
return convertSteeredMessages(messages, async (batch) => convertToModelMessages(batch, {
|
|
61
|
+
return convertSteeredMessages(messages, async (batch) => withJsonToolResults(await convertToModelMessages(batch, {
|
|
61
62
|
ignoreIncompleteToolCalls: true,
|
|
62
63
|
...(tools ? { tools } : {}),
|
|
63
|
-
}), locals.get(chatSteeringInjectionsKey) ?? new Map(), context ?? locals.get(chatCurrentUIMessagesKey) ?? messages, context !== undefined || locals.get(chatCurrentUIMessagesKey) !== undefined);
|
|
64
|
+
})), locals.get(chatSteeringInjectionsKey) ?? new Map(), context ?? locals.get(chatCurrentUIMessagesKey) ?? messages, context !== undefined || locals.get(chatCurrentUIMessagesKey) !== undefined);
|
|
64
65
|
}
|
|
65
66
|
const chatTurnContextKey = locals.create("chat.turnContext");
|
|
67
|
+
/**
|
|
68
|
+
* The running chat turn's `chatId`, `turn` and `clientData`: the one place every callback event
|
|
69
|
+
* the SDK builds takes them from, so a callback never lacks the `clientData` its siblings get.
|
|
70
|
+
* All three are undefined outside a turn. One locals lookup; the values are not copied.
|
|
71
|
+
*/
|
|
72
|
+
function chatEventContext() {
|
|
73
|
+
const ctx = locals.get(chatTurnContextKey);
|
|
74
|
+
return { chatId: ctx?.chatId, turn: ctx?.turn, clientData: ctx?.clientData };
|
|
75
|
+
}
|
|
66
76
|
/**
|
|
67
77
|
* Per-run slot holding the Session handle that backs this chat's `.in` /
|
|
68
78
|
* `.out` channels. Populated at the top of `chatAgent`'s run function from
|
|
@@ -2050,6 +2060,33 @@ function emptyUsage() {
|
|
|
2050
2060
|
outputTokenDetails: { textTokens: undefined, reasoningTokens: undefined },
|
|
2051
2061
|
};
|
|
2052
2062
|
}
|
|
2063
|
+
/**
|
|
2064
|
+
* The last step's usage of a `streamText` result: what the model's context window held on
|
|
2065
|
+
* its final call. It comes from `steps` because the result's `usage` is the last step's on
|
|
2066
|
+
* AI SDK 6 but the turn's sum on AI SDK 7. A custom stream source with no `steps` falls
|
|
2067
|
+
* back to its `usage`, which such a source supplies as the last step's. Raced against a 2s
|
|
2068
|
+
* timeout like the turn's `totalUsage`, since the promises can hang on a stop-abort.
|
|
2069
|
+
* Returns undefined when neither resolves in time.
|
|
2070
|
+
*/
|
|
2071
|
+
async function readLastStepUsage(result) {
|
|
2072
|
+
const stepsPromise = result != null ? result.steps : undefined;
|
|
2073
|
+
const hasSteps = typeof stepsPromise?.then === "function";
|
|
2074
|
+
const source = hasSteps ? stepsPromise : result != null ? result.usage : undefined;
|
|
2075
|
+
if (typeof source?.then !== "function")
|
|
2076
|
+
return undefined;
|
|
2077
|
+
try {
|
|
2078
|
+
const resolved = await Promise.race([
|
|
2079
|
+
source,
|
|
2080
|
+
new Promise((r) => setTimeout(() => r(undefined), 2_000)),
|
|
2081
|
+
]);
|
|
2082
|
+
return hasSteps
|
|
2083
|
+
? resolved?.at(-1)?.usage
|
|
2084
|
+
: resolved;
|
|
2085
|
+
}
|
|
2086
|
+
catch {
|
|
2087
|
+
return undefined;
|
|
2088
|
+
}
|
|
2089
|
+
}
|
|
2053
2090
|
function addUsage(a, b) {
|
|
2054
2091
|
const add = (x, y) => x != null || y != null ? (x ?? 0) + (y ?? 0) : undefined;
|
|
2055
2092
|
return {
|
|
@@ -2596,6 +2633,21 @@ function managedResponse() {
|
|
|
2596
2633
|
}
|
|
2597
2634
|
return response;
|
|
2598
2635
|
}
|
|
2636
|
+
/**
|
|
2637
|
+
* The `summarize` that `chatCompact` calls between tool steps for a compaction configured on
|
|
2638
|
+
* `chat.agent`, a session turn or a `MessageAccumulator`: the event is built here once, so no
|
|
2639
|
+
* surface hands the callback less than the others do.
|
|
2640
|
+
* @internal
|
|
2641
|
+
*/
|
|
2642
|
+
function innerSummarize(compaction, steps) {
|
|
2643
|
+
return (messages) => compaction.summarize({
|
|
2644
|
+
messages,
|
|
2645
|
+
usage: steps.at(-1)?.usage,
|
|
2646
|
+
source: "inner",
|
|
2647
|
+
stepNumber: steps.length - 1,
|
|
2648
|
+
...chatEventContext(),
|
|
2649
|
+
});
|
|
2650
|
+
}
|
|
2599
2651
|
/**
|
|
2600
2652
|
* Check that no tool calls are in-flight in a step's content.
|
|
2601
2653
|
* Used before compaction to avoid losing tool state mid-execution.
|
|
@@ -2667,15 +2719,9 @@ async function applyPrepareMessages(messages, reason) {
|
|
|
2667
2719
|
const hook = locals.get(chatPrepareMessagesKey);
|
|
2668
2720
|
if (!hook)
|
|
2669
2721
|
return messages;
|
|
2670
|
-
const
|
|
2722
|
+
const { chatId = "", turn = 0, clientData } = chatEventContext();
|
|
2671
2723
|
const prepared = await tracer.startActiveSpan("prepareMessages()", async () => {
|
|
2672
|
-
return hook({
|
|
2673
|
-
messages,
|
|
2674
|
-
reason,
|
|
2675
|
-
chatId: turnCtx?.chatId ?? "",
|
|
2676
|
-
turn: turnCtx?.turn ?? 0,
|
|
2677
|
-
clientData: turnCtx?.clientData,
|
|
2678
|
-
});
|
|
2724
|
+
return hook({ messages, reason, chatId, turn, clientData });
|
|
2679
2725
|
}, {
|
|
2680
2726
|
attributes: {
|
|
2681
2727
|
[SemanticInternalAttributes.STYLE_ICON]: "task-hook-onStart",
|
|
@@ -2746,9 +2792,10 @@ function getCompactionState() {
|
|
|
2746
2792
|
*
|
|
2747
2793
|
* @example
|
|
2748
2794
|
* ```ts
|
|
2749
|
-
* prepareStep: async ({ messages, steps }) => {
|
|
2795
|
+
* prepareStep: async ({ messages, steps, responseMessages }) => {
|
|
2750
2796
|
* // your custom logic here...
|
|
2751
2797
|
* const result = await chat.compact(messages, steps, {
|
|
2798
|
+
* responseMessages, // needed on AI SDK 7 to rebuild the steps after a compaction
|
|
2752
2799
|
* threshold: 80_000,
|
|
2753
2800
|
* summarize: async (msgs) => generateText({ model, messages: msgs }).then(r => r.text),
|
|
2754
2801
|
* });
|
|
@@ -2772,7 +2819,10 @@ async function chatCompact(messages, steps, options) {
|
|
|
2772
2819
|
type: "rebuilt",
|
|
2773
2820
|
messages: await applyPrepareMessages([
|
|
2774
2821
|
{ role: "user", content: state.summary },
|
|
2775
|
-
...
|
|
2822
|
+
...generationResponseMessages({
|
|
2823
|
+
steps,
|
|
2824
|
+
responseMessages: options.responseMessages,
|
|
2825
|
+
}).slice(state.baseResponseMessageCount),
|
|
2776
2826
|
], "compaction-rebuild"),
|
|
2777
2827
|
};
|
|
2778
2828
|
}
|
|
@@ -2783,7 +2833,7 @@ async function chatCompact(messages, steps, options) {
|
|
|
2783
2833
|
const totalTokens = currentStep.usage.totalTokens;
|
|
2784
2834
|
const inputTokens = currentStep.usage.inputTokens;
|
|
2785
2835
|
const outputTokens = currentStep.usage.outputTokens;
|
|
2786
|
-
const turnCtx =
|
|
2836
|
+
const turnCtx = chatEventContext();
|
|
2787
2837
|
const stepNumber = steps.length - 1;
|
|
2788
2838
|
const shouldTrigger = options.shouldCompact
|
|
2789
2839
|
? await options.shouldCompact({
|
|
@@ -2795,9 +2845,7 @@ async function chatCompact(messages, steps, options) {
|
|
|
2795
2845
|
source: "inner",
|
|
2796
2846
|
stepNumber,
|
|
2797
2847
|
steps,
|
|
2798
|
-
|
|
2799
|
-
turn: turnCtx?.turn,
|
|
2800
|
-
clientData: turnCtx?.clientData,
|
|
2848
|
+
...turnCtx,
|
|
2801
2849
|
})
|
|
2802
2850
|
: totalTokens != null && options.threshold != null && totalTokens > options.threshold;
|
|
2803
2851
|
if (!shouldTrigger) {
|
|
@@ -2825,7 +2873,10 @@ async function chatCompact(messages, steps, options) {
|
|
|
2825
2873
|
// Store state in locals for subsequent steps
|
|
2826
2874
|
locals.set(chatCompactionStateKey, {
|
|
2827
2875
|
summary,
|
|
2828
|
-
baseResponseMessageCount:
|
|
2876
|
+
baseResponseMessageCount: generationResponseMessages({
|
|
2877
|
+
steps,
|
|
2878
|
+
responseMessages: options.responseMessages,
|
|
2879
|
+
}).length,
|
|
2829
2880
|
baseResponseStepCount: steps.length,
|
|
2830
2881
|
});
|
|
2831
2882
|
// Set model-only override — UI messages stay intact for persistence.
|
|
@@ -2851,8 +2902,7 @@ async function chatCompact(messages, steps, options) {
|
|
|
2851
2902
|
inputTokens,
|
|
2852
2903
|
outputTokens,
|
|
2853
2904
|
stepNumber,
|
|
2854
|
-
|
|
2855
|
-
turn: turnCtx?.turn,
|
|
2905
|
+
...turnCtx,
|
|
2856
2906
|
writer: { write, merge },
|
|
2857
2907
|
});
|
|
2858
2908
|
}
|
|
@@ -2918,8 +2968,8 @@ async function chatCompact(messages, steps, options) {
|
|
|
2918
2968
|
* ```
|
|
2919
2969
|
*/
|
|
2920
2970
|
function chatCompactionStep(options) {
|
|
2921
|
-
return async ({ messages, steps }) => {
|
|
2922
|
-
const result = await chatCompact(messages, steps, options);
|
|
2971
|
+
return async ({ messages, steps, responseMessages }) => {
|
|
2972
|
+
const result = await chatCompact(messages, steps, { ...options, responseMessages });
|
|
2923
2973
|
return result.type === "skipped" ? undefined : result;
|
|
2924
2974
|
};
|
|
2925
2975
|
}
|
|
@@ -3107,6 +3157,7 @@ async function drainSteeringQueue(config, messages, steps, queueOverride) {
|
|
|
3107
3157
|
injectedModelMessages: injected,
|
|
3108
3158
|
chatId: ctx?.chatId ?? "",
|
|
3109
3159
|
turn: ctx?.turn ?? 0,
|
|
3160
|
+
clientData: ctx?.clientData,
|
|
3110
3161
|
stepNumber,
|
|
3111
3162
|
});
|
|
3112
3163
|
}
|
|
@@ -3626,25 +3677,14 @@ function toStreamTextOptions(options, retain = true) {
|
|
|
3626
3677
|
const taskCompaction = locals.get(chatAgentCompactionKey);
|
|
3627
3678
|
const taskPendingMessages = locals.get(chatPendingMessagesKey);
|
|
3628
3679
|
{
|
|
3629
|
-
result.prepareStep = async ({ messages, steps, }) => {
|
|
3680
|
+
result.prepareStep = async ({ messages, steps, responseMessages, }) => {
|
|
3630
3681
|
let resultMessages;
|
|
3631
3682
|
// 1. Compaction
|
|
3632
3683
|
if (taskCompaction) {
|
|
3633
3684
|
const compactResult = await chatCompact(messages, steps, {
|
|
3685
|
+
responseMessages,
|
|
3634
3686
|
shouldCompact: taskCompaction.shouldCompact,
|
|
3635
|
-
summarize: (
|
|
3636
|
-
const ctx = locals.get(chatTurnContextKey);
|
|
3637
|
-
const lastStep = steps.at(-1);
|
|
3638
|
-
return taskCompaction.summarize({
|
|
3639
|
-
messages: msgs,
|
|
3640
|
-
usage: lastStep?.usage,
|
|
3641
|
-
source: "inner",
|
|
3642
|
-
stepNumber: steps.length - 1,
|
|
3643
|
-
chatId: ctx?.chatId,
|
|
3644
|
-
turn: ctx?.turn,
|
|
3645
|
-
clientData: ctx?.clientData,
|
|
3646
|
-
});
|
|
3647
|
-
},
|
|
3687
|
+
summarize: innerSummarize(taskCompaction, steps),
|
|
3648
3688
|
});
|
|
3649
3689
|
if (compactResult.type !== "skipped") {
|
|
3650
3690
|
resultMessages = compactResult.messages;
|
|
@@ -3667,13 +3707,27 @@ function toStreamTextOptions(options, retain = true) {
|
|
|
3667
3707
|
// 4. Parallel tool results in call order, the order the next turn's history
|
|
3668
3708
|
// will list them in. See `withToolResultsInCallOrder`.
|
|
3669
3709
|
const base = resultMessages ?? messages;
|
|
3670
|
-
|
|
3671
|
-
|
|
3710
|
+
// 5. Tool results as plain JSON, which AI SDK 7 before 7.0.101 checks on
|
|
3711
|
+
// every step. See `withJsonToolResults`.
|
|
3712
|
+
const prepared = await withJsonToolResults(withToolResultsInCallOrder(base));
|
|
3713
|
+
return resultMessages || prepared !== base ? { messages: prepared } : undefined;
|
|
3672
3714
|
};
|
|
3673
3715
|
}
|
|
3674
3716
|
if (retain && typeof result.prepareStep === "function") {
|
|
3675
3717
|
result.prepareStep = retainStepMessages(result.prepareStep);
|
|
3676
3718
|
}
|
|
3719
|
+
/**
|
|
3720
|
+
* A steer that `pendingMessages.prepare` turns into a `system` message lands in
|
|
3721
|
+
* the model lane in that form, and so replays in every later turn's `messages`.
|
|
3722
|
+
* AI SDK 7 rejects a system message in `messages` unless this is set (its default
|
|
3723
|
+
* is `false`), so without it the step fails with an invalid-prompt error on the
|
|
3724
|
+
* turn that was steered and on every turn after it. Routing the note to
|
|
3725
|
+
* `instructions` for one step would not help the later turns, because the lane
|
|
3726
|
+
* is what the caller's own `streamText({ messages })` reads. A caller that wants
|
|
3727
|
+
* the strict behaviour passes `allowSystemInMessages: false` after the spread.
|
|
3728
|
+
* AI SDK 5 and 6 ignore the key and accept system messages already.
|
|
3729
|
+
*/
|
|
3730
|
+
result.allowSystemInMessages = true;
|
|
3677
3731
|
const channelDownload = locals.get(chatChannelDownloadKey);
|
|
3678
3732
|
if (channelDownload) {
|
|
3679
3733
|
result.experimental_download = channelAwareDownload(channelDownload);
|
|
@@ -3727,17 +3781,35 @@ function isActionTurn(value) {
|
|
|
3727
3781
|
* tail does not match the old message's conversion, nothing is changed and
|
|
3728
3782
|
* `false` is returned so the caller can fall back to a full reconversion.
|
|
3729
3783
|
*/
|
|
3730
|
-
async function replaceModelRun(lane, oldUi, newUi, tailAfter, knownOldRun) {
|
|
3784
|
+
async function replaceModelRun(lane, oldUi, newUi, tailAfter, knownOldRun, oldIsLastUiMessage = false) {
|
|
3731
3785
|
const oldRun = knownOldRun ?? (await toModelMessages([stripProviderMetadata(oldUi)]));
|
|
3732
3786
|
const newRun = await toModelMessages([stripProviderMetadata(newUi)]);
|
|
3733
|
-
// A message that converts to nothing (a pending tool call with no output yet,
|
|
3734
|
-
// which `ignoreIncompleteToolCalls` drops) locates no run in the lane. Matching
|
|
3735
|
-
// an empty slice would splice the new run in without removing what the message
|
|
3736
|
-
// actually contributed, such as a spliced head-start partial, and the lane would
|
|
3737
|
-
// then carry the same tool call twice.
|
|
3738
|
-
if (oldRun.length === 0)
|
|
3739
|
-
return false;
|
|
3740
3787
|
const end = lane.length - tailAfter;
|
|
3788
|
+
// A message that converts to nothing (a tool call awaiting approval on AI SDK 7,
|
|
3789
|
+
// or a pending one with no output yet, which `ignoreIncompleteToolCalls` drops)
|
|
3790
|
+
// locates no run in the lane. When it is the last UI message and the lane holds
|
|
3791
|
+
// none of its tool calls, it contributed nothing, so the new run goes in at the
|
|
3792
|
+
// tail. A message that is not the last one has later turns after its place, which
|
|
3793
|
+
// the tail is not. When the lane does hold one of its calls, such as a spliced
|
|
3794
|
+
// head-start partial, matching an empty slice would leave the old form behind and
|
|
3795
|
+
// the lane would carry the same tool call twice.
|
|
3796
|
+
if (oldRun.length === 0) {
|
|
3797
|
+
if (!oldIsLastUiMessage || end < 0 || end > lane.length)
|
|
3798
|
+
return false;
|
|
3799
|
+
const oldCallIds = new Set();
|
|
3800
|
+
for (const part of oldUi.parts) {
|
|
3801
|
+
const id = part.toolCallId;
|
|
3802
|
+
if (typeof id === "string")
|
|
3803
|
+
oldCallIds.add(id);
|
|
3804
|
+
}
|
|
3805
|
+
// Tool-call, tool-result and approval parts of a lane message all carry `toolCallId`.
|
|
3806
|
+
const holdsOldCall = lane.some((message) => Array.isArray(message.content) &&
|
|
3807
|
+
message.content.some((part) => oldCallIds.has(part.toolCallId)));
|
|
3808
|
+
if (holdsOldCall)
|
|
3809
|
+
return false;
|
|
3810
|
+
lane.splice(end, 0, ...newRun);
|
|
3811
|
+
return true;
|
|
3812
|
+
}
|
|
3741
3813
|
const start = end - oldRun.length;
|
|
3742
3814
|
if (start < 0 || end > lane.length)
|
|
3743
3815
|
return false;
|
|
@@ -4031,6 +4103,8 @@ function makeChannelDownloader(connectors, handleFor) {
|
|
|
4031
4103
|
return null;
|
|
4032
4104
|
};
|
|
4033
4105
|
return (url) => {
|
|
4106
|
+
if (isInlineDataUrl(url))
|
|
4107
|
+
return Promise.resolve(null);
|
|
4034
4108
|
const key = url.href;
|
|
4035
4109
|
const cached = cache.get(key);
|
|
4036
4110
|
if (cached)
|
|
@@ -4053,17 +4127,32 @@ function makeChannelDownloader(connectors, handleFor) {
|
|
|
4053
4127
|
}
|
|
4054
4128
|
function fileUrlOf(part) {
|
|
4055
4129
|
const p = part;
|
|
4056
|
-
|
|
4057
|
-
|
|
4058
|
-
|
|
4059
|
-
|
|
4060
|
-
|
|
4061
|
-
try {
|
|
4062
|
-
return new URL(raw);
|
|
4130
|
+
let raw = p?.type === "file" ? p.data : p?.type === "image" ? p.image : undefined;
|
|
4131
|
+
// AI SDK 7 tags the data (`{ type: "url", url }`); AI SDK 6 puts a URL or string there directly.
|
|
4132
|
+
if (raw && typeof raw === "object" && !(raw instanceof URL)) {
|
|
4133
|
+
const tagged = raw;
|
|
4134
|
+
raw = tagged.type === "url" ? tagged.url : undefined;
|
|
4063
4135
|
}
|
|
4064
|
-
|
|
4136
|
+
let url;
|
|
4137
|
+
if (raw instanceof URL) {
|
|
4138
|
+
url = raw;
|
|
4139
|
+
}
|
|
4140
|
+
else if (typeof raw === "string" && /^https?:\/\//i.test(raw)) {
|
|
4141
|
+
try {
|
|
4142
|
+
url = new URL(raw);
|
|
4143
|
+
}
|
|
4144
|
+
catch {
|
|
4145
|
+
return undefined;
|
|
4146
|
+
}
|
|
4147
|
+
}
|
|
4148
|
+
else {
|
|
4065
4149
|
return undefined;
|
|
4066
4150
|
}
|
|
4151
|
+
return isInlineDataUrl(url) ? undefined : url;
|
|
4152
|
+
}
|
|
4153
|
+
/** An inline `data:` upload carries its bytes in the URL, so it is never handed to a connector, cached or logged. */
|
|
4154
|
+
function isInlineDataUrl(url) {
|
|
4155
|
+
return url.protocol === "data:";
|
|
4067
4156
|
}
|
|
4068
4157
|
/**
|
|
4069
4158
|
* Fetch the channel-owned files in the messages a turn hands to `run()`, so the model call reads
|
|
@@ -4130,7 +4219,8 @@ async function resolveChannelFiles(messages, download) {
|
|
|
4130
4219
|
/**
|
|
4131
4220
|
* The AI SDK `experimental_download` for a turn: channel-owned URLs come from the connectors, and
|
|
4132
4221
|
* every other URL goes to `fallback`, or to the AI SDK's default behaviour (left for a model that
|
|
4133
|
-
* reads the URL itself, fetched from a public host otherwise).
|
|
4222
|
+
* reads the URL itself, fetched from a public host otherwise). An inline `data:` URL is neither: it
|
|
4223
|
+
* is left to the AI SDK to decode, and reaches no connector, fallback or fetch.
|
|
4134
4224
|
*/
|
|
4135
4225
|
function channelAwareDownload(download, fallback) {
|
|
4136
4226
|
return async (requests) => {
|
|
@@ -4138,7 +4228,9 @@ function channelAwareDownload(download, fallback) {
|
|
|
4138
4228
|
const file = await download(request.url);
|
|
4139
4229
|
return file ? { data: file.data, mediaType: file.mediaType } : null;
|
|
4140
4230
|
}));
|
|
4141
|
-
|
|
4231
|
+
// An inline `data:` file is left to the AI SDK, which decodes it itself: it is no connector's
|
|
4232
|
+
// file, and `fetchPublicFile` would refuse it with the whole upload in the error message.
|
|
4233
|
+
const misses = requests.flatMap((request, i) => results[i] || request.url.protocol === "data:" ? [] : [i]);
|
|
4142
4234
|
if (misses.length === 0)
|
|
4143
4235
|
return results;
|
|
4144
4236
|
const rest = fallback
|
|
@@ -4766,28 +4858,46 @@ function makeChannelStreamTap(editor, options = {}) {
|
|
|
4766
4858
|
}
|
|
4767
4859
|
export { makeChannelStreamEditor as __makeChannelStreamEditorForTests, makeChannelStreamTap as __makeChannelStreamTapForTests, makeChannelProgressObserver as __makeChannelProgressObserverForTests, makeChannelLifecycleReporter as __makeChannelLifecycleReporterForTests, };
|
|
4768
4860
|
/**
|
|
4769
|
-
* Per-chat
|
|
4770
|
-
*
|
|
4771
|
-
*
|
|
4772
|
-
*
|
|
4773
|
-
|
|
4774
|
-
|
|
4775
|
-
const
|
|
4776
|
-
|
|
4777
|
-
|
|
4778
|
-
|
|
4779
|
-
|
|
4780
|
-
|
|
4781
|
-
|
|
4782
|
-
|
|
4783
|
-
|
|
4784
|
-
|
|
4785
|
-
|
|
4786
|
-
|
|
4861
|
+
* Per-chat latches for work the run loop does AFTER it writes `trigger:turn-complete`. The loop
|
|
4862
|
+
* opens a latch before writing the chunk and settles it once that work has finished, so a caller
|
|
4863
|
+
* that saw the chunk can wait for it instead of guessing with timers. One latch per chat at a
|
|
4864
|
+
* time (opening an open one keeps it); a chat with no open latch is quiet.
|
|
4865
|
+
*/
|
|
4866
|
+
function createChatLatches() {
|
|
4867
|
+
const latches = new Map();
|
|
4868
|
+
return {
|
|
4869
|
+
open(chatId) {
|
|
4870
|
+
if (latches.has(chatId))
|
|
4871
|
+
return;
|
|
4872
|
+
let settle;
|
|
4873
|
+
const promise = new Promise((resolve) => (settle = resolve));
|
|
4874
|
+
latches.set(chatId, { promise, settle });
|
|
4875
|
+
},
|
|
4876
|
+
/** `"run-ended"` marks a settle made as the run exits, so a waiter knows the run itself is over. */
|
|
4877
|
+
settle(chatId, reason = "settled") {
|
|
4878
|
+
const latch = latches.get(chatId);
|
|
4879
|
+
if (!latch)
|
|
4880
|
+
return;
|
|
4881
|
+
latches.delete(chatId);
|
|
4882
|
+
latch.settle(reason);
|
|
4883
|
+
},
|
|
4884
|
+
/** Why the latch settled; `undefined` when nothing is open. */
|
|
4885
|
+
async wait(chatId) {
|
|
4886
|
+
return latches.get(chatId)?.promise;
|
|
4887
|
+
},
|
|
4888
|
+
};
|
|
4787
4889
|
}
|
|
4890
|
+
/** The channel egress after the chunk: the final `send()` and the done/error reactions. */
|
|
4891
|
+
const channelEgressLatches = createChatLatches();
|
|
4892
|
+
/** The post-turn hooks after the chunk: `onTurnComplete` and the turn-complete transcript save. */
|
|
4893
|
+
const postTurnHookLatches = createChatLatches();
|
|
4788
4894
|
/** Resolves once the post-turn channel egress for `chatId` has finished (immediately when none is running). */
|
|
4789
4895
|
export async function __waitForChannelEgressForTests(chatId) {
|
|
4790
|
-
|
|
4896
|
+
return channelEgressLatches.wait(chatId);
|
|
4897
|
+
}
|
|
4898
|
+
/** Resolves once the post-turn hooks for `chatId` have finished (immediately when none are running). */
|
|
4899
|
+
export async function __waitForPostTurnHooksForTests(chatId) {
|
|
4900
|
+
return postTurnHookLatches.wait(chatId);
|
|
4791
4901
|
}
|
|
4792
4902
|
function chatCustomAgent(options) {
|
|
4793
4903
|
const { clientDataSchema, onClientDataValidationError, clientDataReportErrorAt, run: userRun, ...restOptions } = options;
|
|
@@ -6690,6 +6800,7 @@ function chatAgent(options) {
|
|
|
6690
6800
|
message: lastUIMessage,
|
|
6691
6801
|
chatId: currentWirePayload.chatId,
|
|
6692
6802
|
turn,
|
|
6803
|
+
clientData: chatEventContext().clientData,
|
|
6693
6804
|
});
|
|
6694
6805
|
}
|
|
6695
6806
|
catch {
|
|
@@ -6905,6 +7016,7 @@ function chatAgent(options) {
|
|
|
6905
7016
|
messages: cleanedUIMessages,
|
|
6906
7017
|
chatId: currentWirePayload.chatId,
|
|
6907
7018
|
turn,
|
|
7019
|
+
clientData: chatEventContext().clientData,
|
|
6908
7020
|
trigger: currentWirePayload.trigger,
|
|
6909
7021
|
});
|
|
6910
7022
|
}, {
|
|
@@ -7044,7 +7156,11 @@ function chatAgent(options) {
|
|
|
7044
7156
|
if (idx !== -1) {
|
|
7045
7157
|
const previous = accumulatedUIMessages[idx];
|
|
7046
7158
|
accumulatedUIMessages[idx] = mergeIncomingIntoHydrated(previous, incoming);
|
|
7047
|
-
replacedPairs.push({
|
|
7159
|
+
replacedPairs.push({
|
|
7160
|
+
previous,
|
|
7161
|
+
merged: accumulatedUIMessages[idx],
|
|
7162
|
+
isLast: idx === accumulatedUIMessages.length - 1,
|
|
7163
|
+
});
|
|
7048
7164
|
replaced = true;
|
|
7049
7165
|
}
|
|
7050
7166
|
else {
|
|
@@ -7054,8 +7170,8 @@ function chatAgent(options) {
|
|
|
7054
7170
|
}
|
|
7055
7171
|
if (replaced) {
|
|
7056
7172
|
let inPlace = true;
|
|
7057
|
-
for (const { previous, merged } of replacedPairs) {
|
|
7058
|
-
if (!(await replaceModelRun(accumulatedMessages, previous, merged, 0))) {
|
|
7173
|
+
for (const { previous, merged, isLast } of replacedPairs) {
|
|
7174
|
+
if (!(await replaceModelRun(accumulatedMessages, previous, merged, 0, undefined, isLast))) {
|
|
7059
7175
|
inPlace = false;
|
|
7060
7176
|
break;
|
|
7061
7177
|
}
|
|
@@ -7589,6 +7705,21 @@ function chatAgent(options) {
|
|
|
7589
7705
|
new Promise((r) => setTimeout(r, 2_000)),
|
|
7590
7706
|
]);
|
|
7591
7707
|
}
|
|
7708
|
+
/**
|
|
7709
|
+
* A stream-level failure (an `error` chunk) is the turn's finish
|
|
7710
|
+
* reason. A step rejected before the model runs (an invalid prompt,
|
|
7711
|
+
* a throwing `prepareStep`) ends the stream with one but no finish
|
|
7712
|
+
* reason of its own, or the previous step's. Tool errors do not
|
|
7713
|
+
* count: they are `tool-input-error` / `tool-output-error` parts
|
|
7714
|
+
* that go back to the model, and the turn finishes normally.
|
|
7715
|
+
*/
|
|
7716
|
+
// Looked up once per turn; the buffer is only scanned again from
|
|
7717
|
+
// `errorScanLength` on, in case a chunk lands after this point.
|
|
7718
|
+
const errorScanLength = turnBufferedChunks.length;
|
|
7719
|
+
const earlyErrorChunk = turnBufferedChunks.find(isErrorUIMessageChunk);
|
|
7720
|
+
if (earlyErrorChunk) {
|
|
7721
|
+
capturedFinishReason = "error";
|
|
7722
|
+
}
|
|
7592
7723
|
capturedResponseMessage =
|
|
7593
7724
|
(await locals.get(chatManagedResponseKey)?.snapshot()) ?? capturedResponseMessage;
|
|
7594
7725
|
if (capturedResponseMessage)
|
|
@@ -7613,18 +7744,7 @@ function chatAgent(options) {
|
|
|
7613
7744
|
/* non-fatal — usage capture failed */
|
|
7614
7745
|
}
|
|
7615
7746
|
}
|
|
7616
|
-
|
|
7617
|
-
if (typeof lastStepUsagePromise?.then === "function") {
|
|
7618
|
-
try {
|
|
7619
|
-
lastStepUsage = (await Promise.race([
|
|
7620
|
-
lastStepUsagePromise,
|
|
7621
|
-
new Promise((r) => setTimeout(() => r(undefined), 2_000)),
|
|
7622
|
-
]));
|
|
7623
|
-
}
|
|
7624
|
-
catch {
|
|
7625
|
-
/* non-fatal — usage capture failed */
|
|
7626
|
-
}
|
|
7627
|
-
}
|
|
7747
|
+
lastStepUsage = await readLastStepUsage(runResult);
|
|
7628
7748
|
if (turnUsage) {
|
|
7629
7749
|
cumulativeUsage = addUsage(cumulativeUsage, turnUsage);
|
|
7630
7750
|
previousTurnUsage = turnUsage;
|
|
@@ -7815,7 +7935,8 @@ function chatAgent(options) {
|
|
|
7815
7935
|
if (capturedResponseMessage) {
|
|
7816
7936
|
responseCommitted = true;
|
|
7817
7937
|
capturedPartialResponse = capturedResponseMessage;
|
|
7818
|
-
const errorChunk =
|
|
7938
|
+
const errorChunk = (earlyErrorChunk ??
|
|
7939
|
+
turnBufferedChunks.slice(errorScanLength).find(isErrorUIMessageChunk));
|
|
7819
7940
|
channelStreamErrorText = errorChunk?.errorText;
|
|
7820
7941
|
channelTurnText = turnBufferedChunks
|
|
7821
7942
|
.map((chunk) => {
|
|
@@ -7854,9 +7975,7 @@ function chatAgent(options) {
|
|
|
7854
7975
|
usage: contextUsage,
|
|
7855
7976
|
turnUsage,
|
|
7856
7977
|
totalUsage: cumulativeUsage,
|
|
7857
|
-
|
|
7858
|
-
turn,
|
|
7859
|
-
clientData,
|
|
7978
|
+
...chatEventContext(),
|
|
7860
7979
|
source: "outer",
|
|
7861
7980
|
});
|
|
7862
7981
|
if (shouldTrigger) {
|
|
@@ -7876,9 +7995,7 @@ function chatAgent(options) {
|
|
|
7876
7995
|
messages: accumulatedMessages,
|
|
7877
7996
|
usage: turnUsage,
|
|
7878
7997
|
totalUsage: cumulativeUsage,
|
|
7879
|
-
|
|
7880
|
-
turn,
|
|
7881
|
-
clientData,
|
|
7998
|
+
...chatEventContext(),
|
|
7882
7999
|
source: "outer",
|
|
7883
8000
|
});
|
|
7884
8001
|
// Apply compactModelMessages/compactUIMessages callbacks, or defaults.
|
|
@@ -7924,8 +8041,7 @@ function chatAgent(options) {
|
|
|
7924
8041
|
inputTokens: turnUsage.inputTokens,
|
|
7925
8042
|
outputTokens: turnUsage.outputTokens,
|
|
7926
8043
|
stepNumber: -1, // outer loop, not a step
|
|
7927
|
-
|
|
7928
|
-
turn,
|
|
8044
|
+
...chatEventContext(),
|
|
7929
8045
|
writer: { write, merge },
|
|
7930
8046
|
});
|
|
7931
8047
|
}
|
|
@@ -8044,7 +8160,8 @@ function chatAgent(options) {
|
|
|
8044
8160
|
}
|
|
8045
8161
|
}
|
|
8046
8162
|
settleRecoveredTurn(currentWirePayload);
|
|
8047
|
-
|
|
8163
|
+
channelEgressLatches.open(currentWirePayload.chatId);
|
|
8164
|
+
postTurnHookLatches.open(currentWirePayload.chatId);
|
|
8048
8165
|
// Write turn-complete control chunk — closes the frontend stream.
|
|
8049
8166
|
const turnCompleteResult = await writeTurnCompleteChunk(currentWirePayload.chatId, turnAccessToken);
|
|
8050
8167
|
// Tool calls that have not had controls posted yet; a resume with the rest still
|
|
@@ -8243,7 +8360,7 @@ function chatAgent(options) {
|
|
|
8243
8360
|
});
|
|
8244
8361
|
}
|
|
8245
8362
|
}
|
|
8246
|
-
|
|
8363
|
+
channelEgressLatches.settle(currentWirePayload.chatId);
|
|
8247
8364
|
// Fire onTurnComplete — stream is closed, use for persistence.
|
|
8248
8365
|
if (onTurnComplete) {
|
|
8249
8366
|
await tracer.startActiveSpan("onTurnComplete()", async () => {
|
|
@@ -8342,6 +8459,7 @@ function chatAgent(options) {
|
|
|
8342
8459
|
});
|
|
8343
8460
|
}
|
|
8344
8461
|
}
|
|
8462
|
+
postTurnHookLatches.settle(currentWirePayload.chatId);
|
|
8345
8463
|
} // end if (!isAction)
|
|
8346
8464
|
// NOTE: We intentionally do NOT await deferred work from onTurnComplete here.
|
|
8347
8465
|
// Promises deferred in onTurnComplete (e.g. background self-review via
|
|
@@ -8455,12 +8573,13 @@ function chatAgent(options) {
|
|
|
8455
8573
|
if (turnError instanceof Error &&
|
|
8456
8574
|
turnError.name === "AbortError" &&
|
|
8457
8575
|
runSignal.aborted) {
|
|
8458
|
-
|
|
8576
|
+
channelEgressLatches.settle(currentWirePayload.chatId);
|
|
8577
|
+
postTurnHookLatches.settle(currentWirePayload.chatId);
|
|
8459
8578
|
channelLifecycle.dispose();
|
|
8460
8579
|
// Full run cancellation — exit immediately
|
|
8461
8580
|
throw turnError;
|
|
8462
8581
|
}
|
|
8463
|
-
|
|
8582
|
+
channelEgressLatches.settle(currentWirePayload.chatId);
|
|
8464
8583
|
// OOM errors must escape the turn loop so the task runtime can
|
|
8465
8584
|
// honor `retry.outOfMemory.machine` (set on chat.agent via
|
|
8466
8585
|
// `oomMachine`). Catching them here would keep the dead worker
|
|
@@ -8581,6 +8700,7 @@ function chatAgent(options) {
|
|
|
8581
8700
|
writer.write({ type: "error", errorText });
|
|
8582
8701
|
});
|
|
8583
8702
|
// Signal turn complete so the client knows this turn is done
|
|
8703
|
+
postTurnHookLatches.open(currentWirePayload.chatId);
|
|
8584
8704
|
errorTurnCompleteResult = await writeTurnCompleteChunk(currentWirePayload.chatId);
|
|
8585
8705
|
// A later action's snapshot reuses this cursor, so it has to move
|
|
8586
8706
|
// here too or that snapshot resumes from before the failed turn.
|
|
@@ -8781,6 +8901,7 @@ function chatAgent(options) {
|
|
|
8781
8901
|
});
|
|
8782
8902
|
}
|
|
8783
8903
|
}
|
|
8904
|
+
postTurnHookLatches.settle(currentWirePayload.chatId);
|
|
8784
8905
|
if (locals.get(chatCloseRequestedKey)) {
|
|
8785
8906
|
await performChatClose();
|
|
8786
8907
|
return;
|
|
@@ -8836,6 +8957,12 @@ function chatAgent(options) {
|
|
|
8836
8957
|
// reaching one of the loop's close checks (a turn timeout, an OOM
|
|
8837
8958
|
// re-throw). `performChatClose` is idempotent, so the ordinary path
|
|
8838
8959
|
// having already run it costs nothing here.
|
|
8960
|
+
// Any exit that skips a turn's own settle (an OOM re-throw, a throw from the turn's
|
|
8961
|
+
// error handler) leaves its post-turn latches open, and the module-level entries would
|
|
8962
|
+
// outlive this run and block the next one for the same chat. Settling is a no-op when
|
|
8963
|
+
// nothing is open, and runs first so a throwing close cannot skip it.
|
|
8964
|
+
channelEgressLatches.settle(currentWirePayload.chatId, "run-ended");
|
|
8965
|
+
postTurnHookLatches.settle(currentWirePayload.chatId, "run-ended");
|
|
8839
8966
|
await performChatClose();
|
|
8840
8967
|
// `stopSub` is registered post-preload so the close-during-preload
|
|
8841
8968
|
// early-return path may exit before it ever attached. Guard the
|
|
@@ -9740,6 +9867,10 @@ async function flushDeferredChatCustomAgentClientDataErrors() {
|
|
|
9740
9867
|
}
|
|
9741
9868
|
}
|
|
9742
9869
|
}
|
|
9870
|
+
/** Whether a buffered chunk is a stream-level `error` chunk. */
|
|
9871
|
+
function isErrorUIMessageChunk(chunk) {
|
|
9872
|
+
return chunk.type === "error";
|
|
9873
|
+
}
|
|
9743
9874
|
/**
|
|
9744
9875
|
* Pass every chunk through untouched while recording it in `buffer`. Handles
|
|
9745
9876
|
* both the `AsyncIterable` and `ReadableStream` shapes `toUIMessageStream()`
|
|
@@ -9921,6 +10052,7 @@ class ChatMessageAccumulator {
|
|
|
9921
10052
|
const ui = replace ? [] : [...this.uiMessages];
|
|
9922
10053
|
const lane = replace ? [] : [...this.modelMessages];
|
|
9923
10054
|
let appendedModelCount = 0;
|
|
10055
|
+
let appendedUiCount = 0;
|
|
9924
10056
|
let rebuild = replace;
|
|
9925
10057
|
for (const raw of messages) {
|
|
9926
10058
|
const message = raw.role === "assistant" ? cleanupAbortedParts(raw) : raw;
|
|
@@ -9933,13 +10065,18 @@ class ChatMessageAccumulator {
|
|
|
9933
10065
|
const idx = ui.findIndex((m) => m.id === resolved.id);
|
|
9934
10066
|
const previous = ui[idx];
|
|
9935
10067
|
ui[idx] = mergeIncomingIntoHydrated(previous, resolved);
|
|
9936
|
-
|
|
10068
|
+
// The lane's tail is only this message's place when nothing came after it
|
|
10069
|
+
// before this call's own appends.
|
|
10070
|
+
const isLast = idx === ui.length - 1 - appendedUiCount;
|
|
10071
|
+
if (!rebuild &&
|
|
10072
|
+
!(await replaceModelRun(lane, previous, ui[idx], appendedModelCount, undefined, isLast))) {
|
|
9937
10073
|
rebuild = true;
|
|
9938
10074
|
}
|
|
9939
10075
|
continue;
|
|
9940
10076
|
}
|
|
9941
10077
|
ui.push(message);
|
|
9942
10078
|
if (!rebuild) {
|
|
10079
|
+
appendedUiCount++;
|
|
9943
10080
|
const converted = await toModelMessages([message]);
|
|
9944
10081
|
lane.push(...converted);
|
|
9945
10082
|
appendedModelCount += converted.length;
|
|
@@ -10112,13 +10249,14 @@ class ChatMessageAccumulator {
|
|
|
10112
10249
|
const comp = this._compaction;
|
|
10113
10250
|
const pm = this._pendingMessages;
|
|
10114
10251
|
const queue = this._steeringQueue;
|
|
10115
|
-
return retainStepMessages(async ({ messages, steps }) => {
|
|
10252
|
+
return retainStepMessages(async ({ messages, steps, responseMessages, }) => {
|
|
10116
10253
|
let resultMessages;
|
|
10117
10254
|
// 1. Compaction
|
|
10118
10255
|
if (comp) {
|
|
10119
10256
|
const result = await chatCompact(messages, steps, {
|
|
10257
|
+
responseMessages,
|
|
10120
10258
|
shouldCompact: comp.shouldCompact,
|
|
10121
|
-
summarize: (
|
|
10259
|
+
summarize: innerSummarize(comp, steps),
|
|
10122
10260
|
});
|
|
10123
10261
|
if (result.type !== "skipped") {
|
|
10124
10262
|
resultMessages = result.messages;
|
|
@@ -10137,9 +10275,11 @@ class ChatMessageAccumulator {
|
|
|
10137
10275
|
}
|
|
10138
10276
|
/**
|
|
10139
10277
|
* Run outer-loop compaction if needed. Call after adding the response
|
|
10140
|
-
* and capturing usage. Pass the LAST step's usage
|
|
10141
|
-
* the context the model held on
|
|
10142
|
-
*
|
|
10278
|
+
* and capturing usage. Pass the LAST step's usage
|
|
10279
|
+
* (`(await result.steps).at(-1)?.usage`), which is the context the model held on
|
|
10280
|
+
* its final call. `result.usage` is that on AI SDK 6 but the turn's sum on AI SDK 7,
|
|
10281
|
+
* so read `steps`. The turn's sum (`result.totalUsage` on AI SDK 6, `result.usage`
|
|
10282
|
+
* on AI SDK 7) belongs in `context.turnUsage`. Applies
|
|
10143
10283
|
* `compactModelMessages` and `compactUIMessages` callbacks if configured.
|
|
10144
10284
|
*
|
|
10145
10285
|
* @returns `true` if compaction was performed, `false` otherwise.
|
|
@@ -10147,6 +10287,13 @@ class ChatMessageAccumulator {
|
|
|
10147
10287
|
async compactIfNeeded(usage, context) {
|
|
10148
10288
|
if (!this._compaction || !usage)
|
|
10149
10289
|
return false;
|
|
10290
|
+
// An explicit context wins; inside a chat turn the turn's own fills in what it leaves out.
|
|
10291
|
+
const turnContext = chatEventContext();
|
|
10292
|
+
const eventContext = {
|
|
10293
|
+
chatId: context?.chatId ?? turnContext.chatId,
|
|
10294
|
+
turn: context?.turn ?? turnContext.turn,
|
|
10295
|
+
clientData: context?.clientData ?? turnContext.clientData,
|
|
10296
|
+
};
|
|
10150
10297
|
const shouldTrigger = await this._compaction.shouldCompact({
|
|
10151
10298
|
messages: this.modelMessages,
|
|
10152
10299
|
totalTokens: usage.totalTokens,
|
|
@@ -10155,9 +10302,7 @@ class ChatMessageAccumulator {
|
|
|
10155
10302
|
usage,
|
|
10156
10303
|
turnUsage: context?.turnUsage,
|
|
10157
10304
|
totalUsage: context?.totalUsage,
|
|
10158
|
-
|
|
10159
|
-
turn: context?.turn,
|
|
10160
|
-
clientData: context?.clientData,
|
|
10305
|
+
...eventContext,
|
|
10161
10306
|
source: "outer",
|
|
10162
10307
|
});
|
|
10163
10308
|
if (!shouldTrigger)
|
|
@@ -10166,18 +10311,16 @@ class ChatMessageAccumulator {
|
|
|
10166
10311
|
messages: this.modelMessages,
|
|
10167
10312
|
usage,
|
|
10168
10313
|
totalUsage: context?.totalUsage,
|
|
10169
|
-
|
|
10170
|
-
turn: context?.turn,
|
|
10171
|
-
clientData: context?.clientData,
|
|
10314
|
+
...eventContext,
|
|
10172
10315
|
source: "outer",
|
|
10173
10316
|
});
|
|
10174
10317
|
const compactEvent = {
|
|
10175
10318
|
summary,
|
|
10176
10319
|
uiMessages: this.uiMessages,
|
|
10177
10320
|
modelMessages: this.modelMessages,
|
|
10178
|
-
chatId:
|
|
10179
|
-
turn:
|
|
10180
|
-
clientData:
|
|
10321
|
+
chatId: eventContext.chatId ?? "",
|
|
10322
|
+
turn: eventContext.turn ?? 0,
|
|
10323
|
+
clientData: eventContext.clientData,
|
|
10181
10324
|
source: "outer",
|
|
10182
10325
|
};
|
|
10183
10326
|
this.modelMessages = this._compaction.compactModelMessages
|
|
@@ -10475,6 +10618,7 @@ function createChatSession(payload, options) {
|
|
|
10475
10618
|
message: lastUIMessage,
|
|
10476
10619
|
chatId: currentPayload.chatId,
|
|
10477
10620
|
turn,
|
|
10621
|
+
clientData: chatEventContext().clientData,
|
|
10478
10622
|
});
|
|
10479
10623
|
}
|
|
10480
10624
|
catch {
|
|
@@ -10631,18 +10775,7 @@ function createChatSession(payload, options) {
|
|
|
10631
10775
|
/* non-fatal */
|
|
10632
10776
|
}
|
|
10633
10777
|
}
|
|
10634
|
-
|
|
10635
|
-
if (typeof lastStepUsagePromise?.then === "function") {
|
|
10636
|
-
try {
|
|
10637
|
-
lastStepUsage = (await Promise.race([
|
|
10638
|
-
lastStepUsagePromise,
|
|
10639
|
-
new Promise((r) => setTimeout(() => r(undefined), 2_000)),
|
|
10640
|
-
]));
|
|
10641
|
-
}
|
|
10642
|
-
catch {
|
|
10643
|
-
/* non-fatal */
|
|
10644
|
-
}
|
|
10645
|
-
}
|
|
10778
|
+
lastStepUsage = await readLastStepUsage(source);
|
|
10646
10779
|
// Outer-loop compaction (same logic as chat.agent)
|
|
10647
10780
|
if (sessionCompaction && turnUsage && !turnObj.stopped) {
|
|
10648
10781
|
const contextUsage = lastStepUsage ?? turnUsage;
|
|
@@ -10654,9 +10787,7 @@ function createChatSession(payload, options) {
|
|
|
10654
10787
|
usage: contextUsage,
|
|
10655
10788
|
turnUsage,
|
|
10656
10789
|
totalUsage: cumulativeUsage,
|
|
10657
|
-
|
|
10658
|
-
turn,
|
|
10659
|
-
clientData: currentPayload.metadata,
|
|
10790
|
+
...chatEventContext(),
|
|
10660
10791
|
source: "outer",
|
|
10661
10792
|
});
|
|
10662
10793
|
if (shouldTrigger) {
|
|
@@ -10664,9 +10795,7 @@ function createChatSession(payload, options) {
|
|
|
10664
10795
|
messages: accumulator.modelMessages,
|
|
10665
10796
|
usage: turnUsage,
|
|
10666
10797
|
totalUsage: cumulativeUsage,
|
|
10667
|
-
|
|
10668
|
-
turn,
|
|
10669
|
-
clientData: currentPayload.metadata,
|
|
10798
|
+
...chatEventContext(),
|
|
10670
10799
|
source: "outer",
|
|
10671
10800
|
});
|
|
10672
10801
|
const compactEvent = {
|
|
@@ -10710,12 +10839,13 @@ function createChatSession(payload, options) {
|
|
|
10710
10839
|
const hasPending = !!sessionPendingMessages;
|
|
10711
10840
|
if (!hasCompaction && !hasPending)
|
|
10712
10841
|
return undefined;
|
|
10713
|
-
return retainStepMessages(async ({ messages: stepMsgs, steps, }) => {
|
|
10842
|
+
return retainStepMessages(async ({ messages: stepMsgs, steps, responseMessages, }) => {
|
|
10714
10843
|
let resultMessages;
|
|
10715
10844
|
if (sessionCompaction) {
|
|
10716
10845
|
const compactResult = await chatCompact(stepMsgs, steps, {
|
|
10846
|
+
responseMessages,
|
|
10717
10847
|
shouldCompact: sessionCompaction.shouldCompact,
|
|
10718
|
-
summarize: (
|
|
10848
|
+
summarize: innerSummarize(sessionCompaction, steps),
|
|
10719
10849
|
});
|
|
10720
10850
|
if (compactResult.type !== "skipped") {
|
|
10721
10851
|
resultMessages = compactResult.messages;
|
|
@@ -11111,7 +11241,7 @@ function overrideRequestHeaders(authHeaders) {
|
|
|
11111
11241
|
async function callSessionsCreateWithOverride(args) {
|
|
11112
11242
|
const apiClient = apiClientManager.client;
|
|
11113
11243
|
if (!apiClient) {
|
|
11114
|
-
throw new Error("chat.createStartSessionAction: no API access token configured.
|
|
11244
|
+
throw new Error("chat.createStartSessionAction: no API access token configured. Call configure({ accessToken }) or pass apiClient: { accessToken } to createStartSessionAction before invoking the action.");
|
|
11115
11245
|
}
|
|
11116
11246
|
const authHeaders = apiClient.getAuthenticationHeaders("createSession");
|
|
11117
11247
|
const ctx = { endpoint: "sessions", chatId: args.chatId };
|
|
@@ -11135,55 +11265,60 @@ async function callSessionsCreateWithOverride(args) {
|
|
|
11135
11265
|
async function mintPublicTokenWithOverride(args) {
|
|
11136
11266
|
const apiClient = apiClientManager.client;
|
|
11137
11267
|
if (!apiClient) {
|
|
11138
|
-
throw new Error("chat.createStartSessionAction: no API access token configured for
|
|
11268
|
+
throw new Error("chat.createStartSessionAction: no API access token configured for public-token minting. Call configure({ accessToken }) or pass apiClient: { accessToken } to createStartSessionAction.");
|
|
11139
11269
|
}
|
|
11140
11270
|
const authHeaders = apiClient.getAuthenticationHeaders("createPublicToken");
|
|
11141
|
-
const accessToken = authHeaders.Authorization.slice("Bearer ".length);
|
|
11142
11271
|
const ctx = { endpoint: "auth", chatId: args.chatId };
|
|
11143
11272
|
const scopes = [`read:sessions:${args.chatId}`, `write:sessions:${args.chatId}`];
|
|
11144
|
-
const
|
|
11145
|
-
const
|
|
11146
|
-
const url = `${resolveChatStartBaseURL("auth", args.chatId, args.baseURLOption)}${endpoint}`;
|
|
11273
|
+
const baseURL = resolveChatStartBaseURL("auth", args.chatId, args.baseURLOption);
|
|
11274
|
+
const url = `${baseURL}/api/v1/auth/public-tokens`;
|
|
11147
11275
|
const init = {
|
|
11148
11276
|
method: "POST",
|
|
11149
11277
|
headers: overrideRequestHeaders(authHeaders),
|
|
11150
|
-
|
|
11151
|
-
|
|
11152
|
-
|
|
11153
|
-
|
|
11154
|
-
|
|
11155
|
-
|
|
11156
|
-
: args.expirationTime,
|
|
11157
|
-
}),
|
|
11158
|
-
}
|
|
11159
|
-
: {}),
|
|
11278
|
+
body: JSON.stringify({
|
|
11279
|
+
scopes,
|
|
11280
|
+
expirationTime: args.expirationTime instanceof Date
|
|
11281
|
+
? Math.floor(args.expirationTime.getTime() / 1000)
|
|
11282
|
+
: args.expirationTime,
|
|
11283
|
+
}),
|
|
11160
11284
|
};
|
|
11161
11285
|
const response = args.fetchOverride
|
|
11162
11286
|
? await args.fetchOverride(url, init, ctx)
|
|
11163
11287
|
: await fetch(url, init);
|
|
11164
11288
|
if (!response.ok) {
|
|
11165
|
-
|
|
11166
|
-
|
|
11167
|
-
|
|
11289
|
+
const accessToken = authHeaders.Authorization.slice("Bearer ".length);
|
|
11290
|
+
if (canFallbackToLocalPublicToken(accessToken, scopes, response.status)) {
|
|
11291
|
+
await response.arrayBuffer();
|
|
11292
|
+
const claimsURL = `${baseURL}/api/v1/auth/jwt/claims`;
|
|
11293
|
+
const claimsInit = {
|
|
11294
|
+
method: "POST",
|
|
11295
|
+
headers: overrideRequestHeaders(authHeaders),
|
|
11296
|
+
};
|
|
11297
|
+
const claimsResponse = args.fetchOverride
|
|
11298
|
+
? await args.fetchOverride(claimsURL, claimsInit, ctx)
|
|
11299
|
+
: await fetch(claimsURL, claimsInit);
|
|
11300
|
+
if (!claimsResponse.ok) {
|
|
11301
|
+
const text = await claimsResponse.text().catch(() => "");
|
|
11302
|
+
throw new Error(`auth.generateJWTClaims failed: ${claimsResponse.status} ${text}`);
|
|
11303
|
+
}
|
|
11304
|
+
const claims = (await claimsResponse.json());
|
|
11305
|
+
return generateJWT({
|
|
11306
|
+
secretKey: accessToken,
|
|
11307
|
+
payload: { ...claims, scopes },
|
|
11308
|
+
expirationTime: args.expirationTime,
|
|
11309
|
+
});
|
|
11310
|
+
}
|
|
11311
|
+
if (response.status === 404) {
|
|
11312
|
+
throw new Error("The server does not support public-token minting for this credential. Upgrade your self-hosted server.");
|
|
11168
11313
|
}
|
|
11169
11314
|
const text = await response.text().catch(() => "");
|
|
11170
11315
|
throw new Error(`auth.createPublicToken failed: ${response.status} ${text}`);
|
|
11171
11316
|
}
|
|
11172
11317
|
const responseBody = (await response.json());
|
|
11173
|
-
if (
|
|
11174
|
-
|
|
11175
|
-
throw new Error("auth.createPublicToken failed: server response did not include a token");
|
|
11176
|
-
}
|
|
11177
|
-
return responseBody.token;
|
|
11318
|
+
if (typeof responseBody.token !== "string") {
|
|
11319
|
+
throw new Error("auth.createPublicToken failed: server response did not include a token");
|
|
11178
11320
|
}
|
|
11179
|
-
return
|
|
11180
|
-
secretKey: accessToken,
|
|
11181
|
-
payload: {
|
|
11182
|
-
...responseBody,
|
|
11183
|
-
scopes,
|
|
11184
|
-
},
|
|
11185
|
-
expirationTime: args.expirationTime,
|
|
11186
|
-
});
|
|
11321
|
+
return responseBody.token;
|
|
11187
11322
|
}
|
|
11188
11323
|
/**
|
|
11189
11324
|
* Creates a server-side helper that reads a conversation from a transcript
|