@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.
Files changed (193) hide show
  1. package/dist/commonjs/v3/ai.d.ts +34 -8
  2. package/dist/commonjs/v3/ai.js +304 -168
  3. package/dist/commonjs/v3/ai.js.map +1 -1
  4. package/dist/commonjs/v3/auth.d.ts +4 -4
  5. package/dist/commonjs/v3/auth.js +21 -75
  6. package/dist/commonjs/v3/auth.js.map +1 -1
  7. package/dist/commonjs/v3/chat-client.d.ts +1 -1
  8. package/dist/commonjs/v3/chat.d.ts +1 -1
  9. package/dist/commonjs/v3/prompt.js +6 -8
  10. package/dist/commonjs/v3/prompt.js.map +1 -1
  11. package/dist/commonjs/v3/runs.d.ts +5 -1
  12. package/dist/commonjs/v3/runs.js.map +1 -1
  13. package/dist/commonjs/v3/sessions.js +10 -7
  14. package/dist/commonjs/v3/sessions.js.map +1 -1
  15. package/dist/commonjs/v3/shared.d.ts +2 -2
  16. package/dist/commonjs/v3/shared.js +16 -15
  17. package/dist/commonjs/v3/shared.js.map +1 -1
  18. package/dist/commonjs/v3/steeringContext.d.ts +35 -10
  19. package/dist/commonjs/v3/steeringContext.js +86 -14
  20. package/dist/commonjs/v3/steeringContext.js.map +1 -1
  21. package/dist/commonjs/v3/streams.js +10 -8
  22. package/dist/commonjs/v3/streams.js.map +1 -1
  23. package/dist/commonjs/v3/tasks.d.ts +2 -2
  24. package/dist/commonjs/v3/tasks.js.map +1 -1
  25. package/dist/commonjs/v3/test/mock-chat-agent.js +24 -13
  26. package/dist/commonjs/v3/test/mock-chat-agent.js.map +1 -1
  27. package/dist/commonjs/v3/toolOutputJson.d.ts +17 -0
  28. package/dist/commonjs/v3/toolOutputJson.js +241 -0
  29. package/dist/commonjs/v3/toolOutputJson.js.map +1 -0
  30. package/dist/commonjs/v3/webhooks.d.ts +8 -2
  31. package/dist/commonjs/v3/webhooks.js +20 -11
  32. package/dist/commonjs/v3/webhooks.js.map +1 -1
  33. package/dist/commonjs/version.js +1 -1
  34. package/dist/esm/v3/ai.d.ts +34 -8
  35. package/dist/esm/v3/ai.js +306 -171
  36. package/dist/esm/v3/ai.js.map +1 -1
  37. package/dist/esm/v3/auth.d.ts +4 -4
  38. package/dist/esm/v3/auth.js +22 -76
  39. package/dist/esm/v3/auth.js.map +1 -1
  40. package/dist/esm/v3/chat-client.d.ts +1 -1
  41. package/dist/esm/v3/chat.d.ts +1 -1
  42. package/dist/esm/v3/prompt.js +6 -8
  43. package/dist/esm/v3/prompt.js.map +1 -1
  44. package/dist/esm/v3/runs.d.ts +5 -1
  45. package/dist/esm/v3/runs.js.map +1 -1
  46. package/dist/esm/v3/sessions.js +10 -7
  47. package/dist/esm/v3/sessions.js.map +1 -1
  48. package/dist/esm/v3/shared.d.ts +2 -2
  49. package/dist/esm/v3/shared.js +16 -15
  50. package/dist/esm/v3/shared.js.map +1 -1
  51. package/dist/esm/v3/steeringContext.d.ts +35 -10
  52. package/dist/esm/v3/steeringContext.js +85 -14
  53. package/dist/esm/v3/steeringContext.js.map +1 -1
  54. package/dist/esm/v3/streams.js +10 -8
  55. package/dist/esm/v3/streams.js.map +1 -1
  56. package/dist/esm/v3/tasks.d.ts +2 -2
  57. package/dist/esm/v3/tasks.js.map +1 -1
  58. package/dist/esm/v3/test/mock-chat-agent.js +25 -14
  59. package/dist/esm/v3/test/mock-chat-agent.js.map +1 -1
  60. package/dist/esm/v3/toolOutputJson.d.ts +17 -0
  61. package/dist/esm/v3/toolOutputJson.js +238 -0
  62. package/dist/esm/v3/toolOutputJson.js.map +1 -0
  63. package/dist/esm/v3/webhooks.d.ts +8 -2
  64. package/dist/esm/v3/webhooks.js +20 -11
  65. package/dist/esm/v3/webhooks.js.map +1 -1
  66. package/dist/esm/version.js +1 -1
  67. package/docs/ai/observability.mdx +25 -19
  68. package/docs/ai/prompts.mdx +25 -19
  69. package/docs/ai-chat/actions.mdx +1 -1
  70. package/docs/ai-chat/anatomy.mdx +1 -1
  71. package/docs/ai-chat/backend.mdx +24 -24
  72. package/docs/ai-chat/background-injection.mdx +5 -5
  73. package/docs/ai-chat/chat-local.mdx +2 -2
  74. package/docs/ai-chat/compaction.mdx +47 -18
  75. package/docs/ai-chat/custom-agents.mdx +4 -4
  76. package/docs/ai-chat/error-handling.mdx +4 -4
  77. package/docs/ai-chat/fast-starts.mdx +8 -7
  78. package/docs/ai-chat/frontend.mdx +6 -2
  79. package/docs/ai-chat/lifecycle-hooks.mdx +11 -11
  80. package/docs/ai-chat/migrating-from-a-route-handler.mdx +4 -4
  81. package/docs/ai-chat/migrating-from-hydrate-messages.mdx +1 -1
  82. package/docs/ai-chat/overview.mdx +1 -1
  83. package/docs/ai-chat/patterns/branching-conversations.mdx +1 -1
  84. package/docs/ai-chat/patterns/database-persistence.mdx +2 -2
  85. package/docs/ai-chat/patterns/human-in-the-loop.mdx +1 -1
  86. package/docs/ai-chat/patterns/native-compaction.mdx +29 -9
  87. package/docs/ai-chat/patterns/persistence-and-replay.mdx +1 -1
  88. package/docs/ai-chat/patterns/redacting-telemetry.mdx +2 -2
  89. package/docs/ai-chat/patterns/skills.mdx +2 -2
  90. package/docs/ai-chat/patterns/sub-agents.mdx +1 -1
  91. package/docs/ai-chat/patterns/tool-result-auditing.mdx +1 -1
  92. package/docs/ai-chat/patterns/trusted-edge-signals.mdx +2 -2
  93. package/docs/ai-chat/patterns/version-upgrades.mdx +2 -2
  94. package/docs/ai-chat/pending-messages.mdx +5 -5
  95. package/docs/ai-chat/prompt-caching.mdx +7 -7
  96. package/docs/ai-chat/quick-start.mdx +2 -2
  97. package/docs/ai-chat/testing.mdx +9 -9
  98. package/docs/ai-chat/tools.mdx +9 -5
  99. package/docs/ai-chat/transcript-storage.mdx +2 -2
  100. package/docs/ai-chat/types.mdx +10 -3
  101. package/docs/apikeys.mdx +2 -1
  102. package/docs/cli-login-commands.mdx +12 -0
  103. package/docs/concurrency.mdx +40 -0
  104. package/docs/database-connections.mdx +1 -1
  105. package/docs/deploy-environment-variables.mdx +84 -2
  106. package/docs/deployment/trusted-publishing.mdx +4 -0
  107. package/docs/errors-retrying.mdx +25 -18
  108. package/docs/github-actions.mdx +1 -1
  109. package/docs/github-integration.mdx +32 -1
  110. package/docs/how-it-works.mdx +4 -2
  111. package/docs/introduction.mdx +1 -1
  112. package/docs/mcp-introduction.mdx +7 -1
  113. package/docs/mcp-remote.mdx +90 -0
  114. package/docs/oidc/azure.mdx +81 -0
  115. package/docs/oidc/databricks.mdx +99 -0
  116. package/docs/oidc/doppler.mdx +60 -0
  117. package/docs/oidc/gcp.mdx +109 -0
  118. package/docs/oidc/hugging-face.mdx +90 -0
  119. package/docs/oidc/openrouter.mdx +91 -0
  120. package/docs/oidc/snowflake.mdx +91 -0
  121. package/docs/oidc/vault.mdx +71 -0
  122. package/docs/oidc.mdx +2 -0
  123. package/docs/runs/metadata.mdx +1 -1
  124. package/docs/self-hosting/docker.mdx +2 -0
  125. package/docs/self-hosting/env/webapp.mdx +4 -1
  126. package/docs/self-hosting/kubernetes.mdx +2 -0
  127. package/docs/tasks/schemaTask.mdx +26 -13
  128. package/docs/tasks/streams.mdx +2 -2
  129. package/docs/troubleshooting.mdx +7 -0
  130. package/docs/webhooks/channels.mdx +6 -6
  131. package/docs/webhooks/connect.mdx +9 -75
  132. package/docs/webhooks/custom-providers.mdx +87 -0
  133. package/docs/webhooks/deliveries.mdx +7 -7
  134. package/docs/webhooks/filters.mdx +2 -2
  135. package/docs/webhooks/how-it-works.mdx +58 -0
  136. package/docs/webhooks/human-in-the-loop.mdx +3 -3
  137. package/docs/webhooks/linear.mdx +5 -3
  138. package/docs/webhooks/overview.mdx +20 -71
  139. package/docs/webhooks/own-handler.mdx +62 -0
  140. package/docs/webhooks/pricing-and-limits.mdx +13 -0
  141. package/docs/webhooks/providers/attio.mdx +38 -0
  142. package/docs/webhooks/providers/autumn.mdx +40 -0
  143. package/docs/webhooks/providers/cal-com.mdx +41 -0
  144. package/docs/webhooks/providers/calendly.mdx +41 -0
  145. package/docs/webhooks/providers/clerk.mdx +40 -0
  146. package/docs/webhooks/providers/discord.mdx +45 -0
  147. package/docs/webhooks/providers/docusign.mdx +41 -0
  148. package/docs/webhooks/providers/elevenlabs.mdx +41 -0
  149. package/docs/webhooks/providers/firecrawl.mdx +54 -0
  150. package/docs/webhooks/providers/github.mdx +46 -0
  151. package/docs/webhooks/providers/gitlab.mdx +41 -0
  152. package/docs/webhooks/providers/granola.mdx +44 -0
  153. package/docs/webhooks/providers/hookdeck.mdx +98 -0
  154. package/docs/webhooks/providers/infisical.mdx +41 -0
  155. package/docs/webhooks/providers/jira.mdx +41 -0
  156. package/docs/webhooks/providers/langfuse.mdx +40 -0
  157. package/docs/webhooks/providers/linear.mdx +45 -0
  158. package/docs/webhooks/providers/loops.mdx +40 -0
  159. package/docs/webhooks/providers/metronome.mdx +42 -0
  160. package/docs/webhooks/providers/mux.mdx +40 -0
  161. package/docs/webhooks/providers/openai.mdx +40 -0
  162. package/docs/webhooks/providers/pagerduty.mdx +40 -0
  163. package/docs/webhooks/providers/plain.mdx +41 -0
  164. package/docs/webhooks/providers/planetscale.mdx +40 -0
  165. package/docs/webhooks/providers/recall-ai.mdx +54 -0
  166. package/docs/webhooks/providers/replicate.mdx +41 -0
  167. package/docs/webhooks/providers/resend.mdx +44 -0
  168. package/docs/webhooks/providers/retell.mdx +41 -0
  169. package/docs/webhooks/providers/sentry.mdx +41 -0
  170. package/docs/webhooks/providers/shopify.mdx +49 -0
  171. package/docs/webhooks/providers/slack.mdx +48 -0
  172. package/docs/webhooks/providers/square.mdx +40 -0
  173. package/docs/webhooks/providers/stripe.mdx +86 -0
  174. package/docs/webhooks/providers/supabase.mdx +41 -0
  175. package/docs/webhooks/providers/tailscale.mdx +43 -0
  176. package/docs/webhooks/providers/typeform.mdx +42 -0
  177. package/docs/webhooks/providers/vapi.mdx +47 -0
  178. package/docs/webhooks/providers/vercel.mdx +40 -0
  179. package/docs/webhooks/providers/whatsapp.mdx +45 -0
  180. package/docs/webhooks/providers/workos.mdx +41 -0
  181. package/docs/webhooks/providers/zendesk.mdx +41 -0
  182. package/docs/webhooks/providers.mdx +53 -0
  183. package/docs/webhooks/quick-start.mdx +131 -0
  184. package/docs/webhooks/session-routing.mdx +6 -6
  185. package/docs/webhooks/signing-secrets.mdx +80 -0
  186. package/docs/webhooks/slack.mdx +11 -2
  187. package/docs/webhooks/verification.mdx +139 -0
  188. package/docs/webhooks/waiters.mdx +2 -2
  189. package/docs/webhooks/webhook-tasks.mdx +46 -0
  190. package/package.json +4 -5
  191. package/docs/migrating-from-v3.mdx +0 -332
  192. package/docs/webhooks/public-beta.mdx +0 -70
  193. package/docs/webhooks/sources.mdx +0 -252
@@ -18,6 +18,7 @@ exports.__makeChannelStreamTapForTests = makeChannelStreamTap;
18
18
  exports.__makeChannelProgressObserverForTests = makeChannelProgressObserver;
19
19
  exports.__makeChannelLifecycleReporterForTests = makeChannelLifecycleReporter;
20
20
  exports.__waitForChannelEgressForTests = __waitForChannelEgressForTests;
21
+ exports.__waitForPostTurnHooksForTests = __waitForPostTurnHooksForTests;
21
22
  const v3_1 = require("@trigger.dev/core/v3");
22
23
  const webhooks_js_1 = require("./webhooks.js");
23
24
  const modelFileDownload_js_1 = require("./modelFileDownload.js");
@@ -63,6 +64,7 @@ const externalDeploymentId_js_1 = require("./externalDeploymentId.js");
63
64
  const chatVersionSkew_js_1 = require("./chatVersionSkew.js");
64
65
  const sessions_js_1 = require("./sessions.js");
65
66
  const shared_js_1 = require("./shared.js");
67
+ const toolOutputJson_js_1 = require("./toolOutputJson.js");
66
68
  const concurrency_shared_js_1 = require("./concurrency-shared.js");
67
69
  const streams_js_1 = require("./streams.js");
68
70
  const tracer_js_1 = require("./tracer.js");
@@ -79,12 +81,21 @@ function toModelMessages(messages, context) {
79
81
  // conditional spread keeps the options object byte-identical to the no-tools
80
82
  // path when nothing was declared.
81
83
  const tools = locals_js_1.locals.get(chatResolvedToolsKey);
82
- return (0, steeringContext_js_1.convertSteeredMessages)(messages, async (batch) => (0, ai_runtime_js_1.convertToModelMessages)(batch, {
84
+ return (0, steeringContext_js_1.convertSteeredMessages)(messages, async (batch) => (0, toolOutputJson_js_1.withJsonToolResults)(await (0, ai_runtime_js_1.convertToModelMessages)(batch, {
83
85
  ignoreIncompleteToolCalls: true,
84
86
  ...(tools ? { tools } : {}),
85
- }), locals_js_1.locals.get(chatSteeringInjectionsKey) ?? new Map(), context ?? locals_js_1.locals.get(chatCurrentUIMessagesKey) ?? messages, context !== undefined || locals_js_1.locals.get(chatCurrentUIMessagesKey) !== undefined);
87
+ })), locals_js_1.locals.get(chatSteeringInjectionsKey) ?? new Map(), context ?? locals_js_1.locals.get(chatCurrentUIMessagesKey) ?? messages, context !== undefined || locals_js_1.locals.get(chatCurrentUIMessagesKey) !== undefined);
86
88
  }
87
89
  const chatTurnContextKey = locals_js_1.locals.create("chat.turnContext");
90
+ /**
91
+ * The running chat turn's `chatId`, `turn` and `clientData`: the one place every callback event
92
+ * the SDK builds takes them from, so a callback never lacks the `clientData` its siblings get.
93
+ * All three are undefined outside a turn. One locals lookup; the values are not copied.
94
+ */
95
+ function chatEventContext() {
96
+ const ctx = locals_js_1.locals.get(chatTurnContextKey);
97
+ return { chatId: ctx?.chatId, turn: ctx?.turn, clientData: ctx?.clientData };
98
+ }
88
99
  /**
89
100
  * Per-run slot holding the Session handle that backs this chat's `.in` /
90
101
  * `.out` channels. Populated at the top of `chatAgent`'s run function from
@@ -2080,6 +2091,33 @@ function emptyUsage() {
2080
2091
  outputTokenDetails: { textTokens: undefined, reasoningTokens: undefined },
2081
2092
  };
2082
2093
  }
2094
+ /**
2095
+ * The last step's usage of a `streamText` result: what the model's context window held on
2096
+ * its final call. It comes from `steps` because the result's `usage` is the last step's on
2097
+ * AI SDK 6 but the turn's sum on AI SDK 7. A custom stream source with no `steps` falls
2098
+ * back to its `usage`, which such a source supplies as the last step's. Raced against a 2s
2099
+ * timeout like the turn's `totalUsage`, since the promises can hang on a stop-abort.
2100
+ * Returns undefined when neither resolves in time.
2101
+ */
2102
+ async function readLastStepUsage(result) {
2103
+ const stepsPromise = result != null ? result.steps : undefined;
2104
+ const hasSteps = typeof stepsPromise?.then === "function";
2105
+ const source = hasSteps ? stepsPromise : result != null ? result.usage : undefined;
2106
+ if (typeof source?.then !== "function")
2107
+ return undefined;
2108
+ try {
2109
+ const resolved = await Promise.race([
2110
+ source,
2111
+ new Promise((r) => setTimeout(() => r(undefined), 2_000)),
2112
+ ]);
2113
+ return hasSteps
2114
+ ? resolved?.at(-1)?.usage
2115
+ : resolved;
2116
+ }
2117
+ catch {
2118
+ return undefined;
2119
+ }
2120
+ }
2083
2121
  function addUsage(a, b) {
2084
2122
  const add = (x, y) => x != null || y != null ? (x ?? 0) + (y ?? 0) : undefined;
2085
2123
  return {
@@ -2616,6 +2654,21 @@ function managedResponse() {
2616
2654
  }
2617
2655
  return response;
2618
2656
  }
2657
+ /**
2658
+ * The `summarize` that `chatCompact` calls between tool steps for a compaction configured on
2659
+ * `chat.agent`, a session turn or a `MessageAccumulator`: the event is built here once, so no
2660
+ * surface hands the callback less than the others do.
2661
+ * @internal
2662
+ */
2663
+ function innerSummarize(compaction, steps) {
2664
+ return (messages) => compaction.summarize({
2665
+ messages,
2666
+ usage: steps.at(-1)?.usage,
2667
+ source: "inner",
2668
+ stepNumber: steps.length - 1,
2669
+ ...chatEventContext(),
2670
+ });
2671
+ }
2619
2672
  /**
2620
2673
  * Check that no tool calls are in-flight in a step's content.
2621
2674
  * Used before compaction to avoid losing tool state mid-execution.
@@ -2687,15 +2740,9 @@ async function applyPrepareMessages(messages, reason) {
2687
2740
  const hook = locals_js_1.locals.get(chatPrepareMessagesKey);
2688
2741
  if (!hook)
2689
2742
  return messages;
2690
- const turnCtx = locals_js_1.locals.get(chatTurnContextKey);
2743
+ const { chatId = "", turn = 0, clientData } = chatEventContext();
2691
2744
  const prepared = await tracer_js_1.tracer.startActiveSpan("prepareMessages()", async () => {
2692
- return hook({
2693
- messages,
2694
- reason,
2695
- chatId: turnCtx?.chatId ?? "",
2696
- turn: turnCtx?.turn ?? 0,
2697
- clientData: turnCtx?.clientData,
2698
- });
2745
+ return hook({ messages, reason, chatId, turn, clientData });
2699
2746
  }, {
2700
2747
  attributes: {
2701
2748
  [v3_1.SemanticInternalAttributes.STYLE_ICON]: "task-hook-onStart",
@@ -2766,9 +2813,10 @@ function getCompactionState() {
2766
2813
  *
2767
2814
  * @example
2768
2815
  * ```ts
2769
- * prepareStep: async ({ messages, steps }) => {
2816
+ * prepareStep: async ({ messages, steps, responseMessages }) => {
2770
2817
  * // your custom logic here...
2771
2818
  * const result = await chat.compact(messages, steps, {
2819
+ * responseMessages, // needed on AI SDK 7 to rebuild the steps after a compaction
2772
2820
  * threshold: 80_000,
2773
2821
  * summarize: async (msgs) => generateText({ model, messages: msgs }).then(r => r.text),
2774
2822
  * });
@@ -2792,7 +2840,10 @@ async function chatCompact(messages, steps, options) {
2792
2840
  type: "rebuilt",
2793
2841
  messages: await applyPrepareMessages([
2794
2842
  { role: "user", content: state.summary },
2795
- ...currentStep.response.messages.slice(state.baseResponseMessageCount),
2843
+ ...(0, steeringContext_js_1.generationResponseMessages)({
2844
+ steps,
2845
+ responseMessages: options.responseMessages,
2846
+ }).slice(state.baseResponseMessageCount),
2796
2847
  ], "compaction-rebuild"),
2797
2848
  };
2798
2849
  }
@@ -2803,7 +2854,7 @@ async function chatCompact(messages, steps, options) {
2803
2854
  const totalTokens = currentStep.usage.totalTokens;
2804
2855
  const inputTokens = currentStep.usage.inputTokens;
2805
2856
  const outputTokens = currentStep.usage.outputTokens;
2806
- const turnCtx = locals_js_1.locals.get(chatTurnContextKey);
2857
+ const turnCtx = chatEventContext();
2807
2858
  const stepNumber = steps.length - 1;
2808
2859
  const shouldTrigger = options.shouldCompact
2809
2860
  ? await options.shouldCompact({
@@ -2815,9 +2866,7 @@ async function chatCompact(messages, steps, options) {
2815
2866
  source: "inner",
2816
2867
  stepNumber,
2817
2868
  steps,
2818
- chatId: turnCtx?.chatId,
2819
- turn: turnCtx?.turn,
2820
- clientData: turnCtx?.clientData,
2869
+ ...turnCtx,
2821
2870
  })
2822
2871
  : totalTokens != null && options.threshold != null && totalTokens > options.threshold;
2823
2872
  if (!shouldTrigger) {
@@ -2845,7 +2894,10 @@ async function chatCompact(messages, steps, options) {
2845
2894
  // Store state in locals for subsequent steps
2846
2895
  locals_js_1.locals.set(chatCompactionStateKey, {
2847
2896
  summary,
2848
- baseResponseMessageCount: currentStep.response.messages.length,
2897
+ baseResponseMessageCount: (0, steeringContext_js_1.generationResponseMessages)({
2898
+ steps,
2899
+ responseMessages: options.responseMessages,
2900
+ }).length,
2849
2901
  baseResponseStepCount: steps.length,
2850
2902
  });
2851
2903
  // Set model-only override — UI messages stay intact for persistence.
@@ -2871,8 +2923,7 @@ async function chatCompact(messages, steps, options) {
2871
2923
  inputTokens,
2872
2924
  outputTokens,
2873
2925
  stepNumber,
2874
- chatId: turnCtx?.chatId,
2875
- turn: turnCtx?.turn,
2926
+ ...turnCtx,
2876
2927
  writer: { write, merge },
2877
2928
  });
2878
2929
  }
@@ -2938,8 +2989,8 @@ async function chatCompact(messages, steps, options) {
2938
2989
  * ```
2939
2990
  */
2940
2991
  function chatCompactionStep(options) {
2941
- return async ({ messages, steps }) => {
2942
- const result = await chatCompact(messages, steps, options);
2992
+ return async ({ messages, steps, responseMessages }) => {
2993
+ const result = await chatCompact(messages, steps, { ...options, responseMessages });
2943
2994
  return result.type === "skipped" ? undefined : result;
2944
2995
  };
2945
2996
  }
@@ -3127,6 +3178,7 @@ async function drainSteeringQueue(config, messages, steps, queueOverride) {
3127
3178
  injectedModelMessages: injected,
3128
3179
  chatId: ctx?.chatId ?? "",
3129
3180
  turn: ctx?.turn ?? 0,
3181
+ clientData: ctx?.clientData,
3130
3182
  stepNumber,
3131
3183
  });
3132
3184
  }
@@ -3646,25 +3698,14 @@ function toStreamTextOptions(options, retain = true) {
3646
3698
  const taskCompaction = locals_js_1.locals.get(chatAgentCompactionKey);
3647
3699
  const taskPendingMessages = locals_js_1.locals.get(chatPendingMessagesKey);
3648
3700
  {
3649
- result.prepareStep = async ({ messages, steps, }) => {
3701
+ result.prepareStep = async ({ messages, steps, responseMessages, }) => {
3650
3702
  let resultMessages;
3651
3703
  // 1. Compaction
3652
3704
  if (taskCompaction) {
3653
3705
  const compactResult = await chatCompact(messages, steps, {
3706
+ responseMessages,
3654
3707
  shouldCompact: taskCompaction.shouldCompact,
3655
- summarize: (msgs) => {
3656
- const ctx = locals_js_1.locals.get(chatTurnContextKey);
3657
- const lastStep = steps.at(-1);
3658
- return taskCompaction.summarize({
3659
- messages: msgs,
3660
- usage: lastStep?.usage,
3661
- source: "inner",
3662
- stepNumber: steps.length - 1,
3663
- chatId: ctx?.chatId,
3664
- turn: ctx?.turn,
3665
- clientData: ctx?.clientData,
3666
- });
3667
- },
3708
+ summarize: innerSummarize(taskCompaction, steps),
3668
3709
  });
3669
3710
  if (compactResult.type !== "skipped") {
3670
3711
  resultMessages = compactResult.messages;
@@ -3687,13 +3728,27 @@ function toStreamTextOptions(options, retain = true) {
3687
3728
  // 4. Parallel tool results in call order, the order the next turn's history
3688
3729
  // will list them in. See `withToolResultsInCallOrder`.
3689
3730
  const base = resultMessages ?? messages;
3690
- const ordered = (0, toolResultOrder_js_1.withToolResultsInCallOrder)(base);
3691
- return resultMessages || ordered !== base ? { messages: ordered } : undefined;
3731
+ // 5. Tool results as plain JSON, which AI SDK 7 before 7.0.101 checks on
3732
+ // every step. See `withJsonToolResults`.
3733
+ const prepared = await (0, toolOutputJson_js_1.withJsonToolResults)((0, toolResultOrder_js_1.withToolResultsInCallOrder)(base));
3734
+ return resultMessages || prepared !== base ? { messages: prepared } : undefined;
3692
3735
  };
3693
3736
  }
3694
3737
  if (retain && typeof result.prepareStep === "function") {
3695
3738
  result.prepareStep = (0, steeringContext_js_1.retainStepMessages)(result.prepareStep);
3696
3739
  }
3740
+ /**
3741
+ * A steer that `pendingMessages.prepare` turns into a `system` message lands in
3742
+ * the model lane in that form, and so replays in every later turn's `messages`.
3743
+ * AI SDK 7 rejects a system message in `messages` unless this is set (its default
3744
+ * is `false`), so without it the step fails with an invalid-prompt error on the
3745
+ * turn that was steered and on every turn after it. Routing the note to
3746
+ * `instructions` for one step would not help the later turns, because the lane
3747
+ * is what the caller's own `streamText({ messages })` reads. A caller that wants
3748
+ * the strict behaviour passes `allowSystemInMessages: false` after the spread.
3749
+ * AI SDK 5 and 6 ignore the key and accept system messages already.
3750
+ */
3751
+ result.allowSystemInMessages = true;
3697
3752
  const channelDownload = locals_js_1.locals.get(chatChannelDownloadKey);
3698
3753
  if (channelDownload) {
3699
3754
  result.experimental_download = channelAwareDownload(channelDownload);
@@ -3747,17 +3802,35 @@ function isActionTurn(value) {
3747
3802
  * tail does not match the old message's conversion, nothing is changed and
3748
3803
  * `false` is returned so the caller can fall back to a full reconversion.
3749
3804
  */
3750
- async function replaceModelRun(lane, oldUi, newUi, tailAfter, knownOldRun) {
3805
+ async function replaceModelRun(lane, oldUi, newUi, tailAfter, knownOldRun, oldIsLastUiMessage = false) {
3751
3806
  const oldRun = knownOldRun ?? (await toModelMessages([stripProviderMetadata(oldUi)]));
3752
3807
  const newRun = await toModelMessages([stripProviderMetadata(newUi)]);
3753
- // A message that converts to nothing (a pending tool call with no output yet,
3754
- // which `ignoreIncompleteToolCalls` drops) locates no run in the lane. Matching
3755
- // an empty slice would splice the new run in without removing what the message
3756
- // actually contributed, such as a spliced head-start partial, and the lane would
3757
- // then carry the same tool call twice.
3758
- if (oldRun.length === 0)
3759
- return false;
3760
3808
  const end = lane.length - tailAfter;
3809
+ // A message that converts to nothing (a tool call awaiting approval on AI SDK 7,
3810
+ // or a pending one with no output yet, which `ignoreIncompleteToolCalls` drops)
3811
+ // locates no run in the lane. When it is the last UI message and the lane holds
3812
+ // none of its tool calls, it contributed nothing, so the new run goes in at the
3813
+ // tail. A message that is not the last one has later turns after its place, which
3814
+ // the tail is not. When the lane does hold one of its calls, such as a spliced
3815
+ // head-start partial, matching an empty slice would leave the old form behind and
3816
+ // the lane would carry the same tool call twice.
3817
+ if (oldRun.length === 0) {
3818
+ if (!oldIsLastUiMessage || end < 0 || end > lane.length)
3819
+ return false;
3820
+ const oldCallIds = new Set();
3821
+ for (const part of oldUi.parts) {
3822
+ const id = part.toolCallId;
3823
+ if (typeof id === "string")
3824
+ oldCallIds.add(id);
3825
+ }
3826
+ // Tool-call, tool-result and approval parts of a lane message all carry `toolCallId`.
3827
+ const holdsOldCall = lane.some((message) => Array.isArray(message.content) &&
3828
+ message.content.some((part) => oldCallIds.has(part.toolCallId)));
3829
+ if (holdsOldCall)
3830
+ return false;
3831
+ lane.splice(end, 0, ...newRun);
3832
+ return true;
3833
+ }
3761
3834
  const start = end - oldRun.length;
3762
3835
  if (start < 0 || end > lane.length)
3763
3836
  return false;
@@ -4051,6 +4124,8 @@ function makeChannelDownloader(connectors, handleFor) {
4051
4124
  return null;
4052
4125
  };
4053
4126
  return (url) => {
4127
+ if (isInlineDataUrl(url))
4128
+ return Promise.resolve(null);
4054
4129
  const key = url.href;
4055
4130
  const cached = cache.get(key);
4056
4131
  if (cached)
@@ -4073,17 +4148,32 @@ function makeChannelDownloader(connectors, handleFor) {
4073
4148
  }
4074
4149
  function fileUrlOf(part) {
4075
4150
  const p = part;
4076
- const raw = p?.type === "file" ? p.data : p?.type === "image" ? p.image : undefined;
4077
- if (raw instanceof URL)
4078
- return raw;
4079
- if (typeof raw !== "string" || !/^https?:\/\//i.test(raw))
4080
- return undefined;
4081
- try {
4082
- return new URL(raw);
4151
+ let raw = p?.type === "file" ? p.data : p?.type === "image" ? p.image : undefined;
4152
+ // AI SDK 7 tags the data (`{ type: "url", url }`); AI SDK 6 puts a URL or string there directly.
4153
+ if (raw && typeof raw === "object" && !(raw instanceof URL)) {
4154
+ const tagged = raw;
4155
+ raw = tagged.type === "url" ? tagged.url : undefined;
4083
4156
  }
4084
- catch {
4157
+ let url;
4158
+ if (raw instanceof URL) {
4159
+ url = raw;
4160
+ }
4161
+ else if (typeof raw === "string" && /^https?:\/\//i.test(raw)) {
4162
+ try {
4163
+ url = new URL(raw);
4164
+ }
4165
+ catch {
4166
+ return undefined;
4167
+ }
4168
+ }
4169
+ else {
4085
4170
  return undefined;
4086
4171
  }
4172
+ return isInlineDataUrl(url) ? undefined : url;
4173
+ }
4174
+ /** An inline `data:` upload carries its bytes in the URL, so it is never handed to a connector, cached or logged. */
4175
+ function isInlineDataUrl(url) {
4176
+ return url.protocol === "data:";
4087
4177
  }
4088
4178
  /**
4089
4179
  * Fetch the channel-owned files in the messages a turn hands to `run()`, so the model call reads
@@ -4150,7 +4240,8 @@ async function resolveChannelFiles(messages, download) {
4150
4240
  /**
4151
4241
  * The AI SDK `experimental_download` for a turn: channel-owned URLs come from the connectors, and
4152
4242
  * every other URL goes to `fallback`, or to the AI SDK's default behaviour (left for a model that
4153
- * reads the URL itself, fetched from a public host otherwise).
4243
+ * reads the URL itself, fetched from a public host otherwise). An inline `data:` URL is neither: it
4244
+ * is left to the AI SDK to decode, and reaches no connector, fallback or fetch.
4154
4245
  */
4155
4246
  function channelAwareDownload(download, fallback) {
4156
4247
  return async (requests) => {
@@ -4158,7 +4249,9 @@ function channelAwareDownload(download, fallback) {
4158
4249
  const file = await download(request.url);
4159
4250
  return file ? { data: file.data, mediaType: file.mediaType } : null;
4160
4251
  }));
4161
- const misses = requests.flatMap((_, i) => (results[i] ? [] : [i]));
4252
+ // An inline `data:` file is left to the AI SDK, which decodes it itself: it is no connector's
4253
+ // file, and `fetchPublicFile` would refuse it with the whole upload in the error message.
4254
+ const misses = requests.flatMap((request, i) => results[i] || request.url.protocol === "data:" ? [] : [i]);
4162
4255
  if (misses.length === 0)
4163
4256
  return results;
4164
4257
  const rest = fallback
@@ -4785,28 +4878,46 @@ function makeChannelStreamTap(editor, options = {}) {
4785
4878
  return new TransformStream(transformer);
4786
4879
  }
4787
4880
  /**
4788
- * Per-chat latch for the channel egress that runs AFTER `trigger:turn-complete` is written (the
4789
- * final `send()` and the done/error reactions). The run loop opens it before writing the chunk and
4790
- * settles it once that egress has finished, so a caller that saw the chunk can wait for the
4791
- * connector to be quiet instead of guessing with timers. One latch per chat at a time; a chat with
4792
- * no open latch is quiet.
4881
+ * Per-chat latches for work the run loop does AFTER it writes `trigger:turn-complete`. The loop
4882
+ * opens a latch before writing the chunk and settles it once that work has finished, so a caller
4883
+ * that saw the chunk can wait for it instead of guessing with timers. One latch per chat at a
4884
+ * time (opening an open one keeps it); a chat with no open latch is quiet.
4793
4885
  */
4794
- const channelEgressLatches = new Map();
4795
- function openChannelEgressLatch(chatId) {
4796
- let settle;
4797
- const promise = new Promise((resolve) => (settle = resolve));
4798
- channelEgressLatches.set(chatId, { promise, settle });
4799
- }
4800
- function settleChannelEgressLatch(chatId) {
4801
- const latch = channelEgressLatches.get(chatId);
4802
- if (!latch)
4803
- return;
4804
- channelEgressLatches.delete(chatId);
4805
- latch.settle();
4886
+ function createChatLatches() {
4887
+ const latches = new Map();
4888
+ return {
4889
+ open(chatId) {
4890
+ if (latches.has(chatId))
4891
+ return;
4892
+ let settle;
4893
+ const promise = new Promise((resolve) => (settle = resolve));
4894
+ latches.set(chatId, { promise, settle });
4895
+ },
4896
+ /** `"run-ended"` marks a settle made as the run exits, so a waiter knows the run itself is over. */
4897
+ settle(chatId, reason = "settled") {
4898
+ const latch = latches.get(chatId);
4899
+ if (!latch)
4900
+ return;
4901
+ latches.delete(chatId);
4902
+ latch.settle(reason);
4903
+ },
4904
+ /** Why the latch settled; `undefined` when nothing is open. */
4905
+ async wait(chatId) {
4906
+ return latches.get(chatId)?.promise;
4907
+ },
4908
+ };
4806
4909
  }
4910
+ /** The channel egress after the chunk: the final `send()` and the done/error reactions. */
4911
+ const channelEgressLatches = createChatLatches();
4912
+ /** The post-turn hooks after the chunk: `onTurnComplete` and the turn-complete transcript save. */
4913
+ const postTurnHookLatches = createChatLatches();
4807
4914
  /** Resolves once the post-turn channel egress for `chatId` has finished (immediately when none is running). */
4808
4915
  async function __waitForChannelEgressForTests(chatId) {
4809
- await channelEgressLatches.get(chatId)?.promise;
4916
+ return channelEgressLatches.wait(chatId);
4917
+ }
4918
+ /** Resolves once the post-turn hooks for `chatId` have finished (immediately when none are running). */
4919
+ async function __waitForPostTurnHooksForTests(chatId) {
4920
+ return postTurnHookLatches.wait(chatId);
4810
4921
  }
4811
4922
  function chatCustomAgent(options) {
4812
4923
  const { clientDataSchema, onClientDataValidationError, clientDataReportErrorAt, run: userRun, ...restOptions } = options;
@@ -6709,6 +6820,7 @@ function chatAgent(options) {
6709
6820
  message: lastUIMessage,
6710
6821
  chatId: currentWirePayload.chatId,
6711
6822
  turn,
6823
+ clientData: chatEventContext().clientData,
6712
6824
  });
6713
6825
  }
6714
6826
  catch {
@@ -6924,6 +7036,7 @@ function chatAgent(options) {
6924
7036
  messages: cleanedUIMessages,
6925
7037
  chatId: currentWirePayload.chatId,
6926
7038
  turn,
7039
+ clientData: chatEventContext().clientData,
6927
7040
  trigger: currentWirePayload.trigger,
6928
7041
  });
6929
7042
  }, {
@@ -7063,7 +7176,11 @@ function chatAgent(options) {
7063
7176
  if (idx !== -1) {
7064
7177
  const previous = accumulatedUIMessages[idx];
7065
7178
  accumulatedUIMessages[idx] = mergeIncomingIntoHydrated(previous, incoming);
7066
- replacedPairs.push({ previous, merged: accumulatedUIMessages[idx] });
7179
+ replacedPairs.push({
7180
+ previous,
7181
+ merged: accumulatedUIMessages[idx],
7182
+ isLast: idx === accumulatedUIMessages.length - 1,
7183
+ });
7067
7184
  replaced = true;
7068
7185
  }
7069
7186
  else {
@@ -7073,8 +7190,8 @@ function chatAgent(options) {
7073
7190
  }
7074
7191
  if (replaced) {
7075
7192
  let inPlace = true;
7076
- for (const { previous, merged } of replacedPairs) {
7077
- if (!(await replaceModelRun(accumulatedMessages, previous, merged, 0))) {
7193
+ for (const { previous, merged, isLast } of replacedPairs) {
7194
+ if (!(await replaceModelRun(accumulatedMessages, previous, merged, 0, undefined, isLast))) {
7078
7195
  inPlace = false;
7079
7196
  break;
7080
7197
  }
@@ -7608,6 +7725,21 @@ function chatAgent(options) {
7608
7725
  new Promise((r) => setTimeout(r, 2_000)),
7609
7726
  ]);
7610
7727
  }
7728
+ /**
7729
+ * A stream-level failure (an `error` chunk) is the turn's finish
7730
+ * reason. A step rejected before the model runs (an invalid prompt,
7731
+ * a throwing `prepareStep`) ends the stream with one but no finish
7732
+ * reason of its own, or the previous step's. Tool errors do not
7733
+ * count: they are `tool-input-error` / `tool-output-error` parts
7734
+ * that go back to the model, and the turn finishes normally.
7735
+ */
7736
+ // Looked up once per turn; the buffer is only scanned again from
7737
+ // `errorScanLength` on, in case a chunk lands after this point.
7738
+ const errorScanLength = turnBufferedChunks.length;
7739
+ const earlyErrorChunk = turnBufferedChunks.find(isErrorUIMessageChunk);
7740
+ if (earlyErrorChunk) {
7741
+ capturedFinishReason = "error";
7742
+ }
7611
7743
  capturedResponseMessage =
7612
7744
  (await locals_js_1.locals.get(chatManagedResponseKey)?.snapshot()) ?? capturedResponseMessage;
7613
7745
  if (capturedResponseMessage)
@@ -7632,18 +7764,7 @@ function chatAgent(options) {
7632
7764
  /* non-fatal — usage capture failed */
7633
7765
  }
7634
7766
  }
7635
- const lastStepUsagePromise = runResult != null ? runResult.usage : undefined;
7636
- if (typeof lastStepUsagePromise?.then === "function") {
7637
- try {
7638
- lastStepUsage = (await Promise.race([
7639
- lastStepUsagePromise,
7640
- new Promise((r) => setTimeout(() => r(undefined), 2_000)),
7641
- ]));
7642
- }
7643
- catch {
7644
- /* non-fatal — usage capture failed */
7645
- }
7646
- }
7767
+ lastStepUsage = await readLastStepUsage(runResult);
7647
7768
  if (turnUsage) {
7648
7769
  cumulativeUsage = addUsage(cumulativeUsage, turnUsage);
7649
7770
  previousTurnUsage = turnUsage;
@@ -7834,7 +7955,8 @@ function chatAgent(options) {
7834
7955
  if (capturedResponseMessage) {
7835
7956
  responseCommitted = true;
7836
7957
  capturedPartialResponse = capturedResponseMessage;
7837
- const errorChunk = turnBufferedChunks.find((chunk) => chunk.type === "error");
7958
+ const errorChunk = (earlyErrorChunk ??
7959
+ turnBufferedChunks.slice(errorScanLength).find(isErrorUIMessageChunk));
7838
7960
  channelStreamErrorText = errorChunk?.errorText;
7839
7961
  channelTurnText = turnBufferedChunks
7840
7962
  .map((chunk) => {
@@ -7873,9 +7995,7 @@ function chatAgent(options) {
7873
7995
  usage: contextUsage,
7874
7996
  turnUsage,
7875
7997
  totalUsage: cumulativeUsage,
7876
- chatId: currentWirePayload.chatId,
7877
- turn,
7878
- clientData,
7998
+ ...chatEventContext(),
7879
7999
  source: "outer",
7880
8000
  });
7881
8001
  if (shouldTrigger) {
@@ -7895,9 +8015,7 @@ function chatAgent(options) {
7895
8015
  messages: accumulatedMessages,
7896
8016
  usage: turnUsage,
7897
8017
  totalUsage: cumulativeUsage,
7898
- chatId: currentWirePayload.chatId,
7899
- turn,
7900
- clientData,
8018
+ ...chatEventContext(),
7901
8019
  source: "outer",
7902
8020
  });
7903
8021
  // Apply compactModelMessages/compactUIMessages callbacks, or defaults.
@@ -7943,8 +8061,7 @@ function chatAgent(options) {
7943
8061
  inputTokens: turnUsage.inputTokens,
7944
8062
  outputTokens: turnUsage.outputTokens,
7945
8063
  stepNumber: -1, // outer loop, not a step
7946
- chatId: currentWirePayload.chatId,
7947
- turn,
8064
+ ...chatEventContext(),
7948
8065
  writer: { write, merge },
7949
8066
  });
7950
8067
  }
@@ -8063,7 +8180,8 @@ function chatAgent(options) {
8063
8180
  }
8064
8181
  }
8065
8182
  settleRecoveredTurn(currentWirePayload);
8066
- openChannelEgressLatch(currentWirePayload.chatId);
8183
+ channelEgressLatches.open(currentWirePayload.chatId);
8184
+ postTurnHookLatches.open(currentWirePayload.chatId);
8067
8185
  // Write turn-complete control chunk — closes the frontend stream.
8068
8186
  const turnCompleteResult = await writeTurnCompleteChunk(currentWirePayload.chatId, turnAccessToken);
8069
8187
  // Tool calls that have not had controls posted yet; a resume with the rest still
@@ -8262,7 +8380,7 @@ function chatAgent(options) {
8262
8380
  });
8263
8381
  }
8264
8382
  }
8265
- settleChannelEgressLatch(currentWirePayload.chatId);
8383
+ channelEgressLatches.settle(currentWirePayload.chatId);
8266
8384
  // Fire onTurnComplete — stream is closed, use for persistence.
8267
8385
  if (onTurnComplete) {
8268
8386
  await tracer_js_1.tracer.startActiveSpan("onTurnComplete()", async () => {
@@ -8361,6 +8479,7 @@ function chatAgent(options) {
8361
8479
  });
8362
8480
  }
8363
8481
  }
8482
+ postTurnHookLatches.settle(currentWirePayload.chatId);
8364
8483
  } // end if (!isAction)
8365
8484
  // NOTE: We intentionally do NOT await deferred work from onTurnComplete here.
8366
8485
  // Promises deferred in onTurnComplete (e.g. background self-review via
@@ -8474,12 +8593,13 @@ function chatAgent(options) {
8474
8593
  if (turnError instanceof Error &&
8475
8594
  turnError.name === "AbortError" &&
8476
8595
  runSignal.aborted) {
8477
- settleChannelEgressLatch(currentWirePayload.chatId);
8596
+ channelEgressLatches.settle(currentWirePayload.chatId);
8597
+ postTurnHookLatches.settle(currentWirePayload.chatId);
8478
8598
  channelLifecycle.dispose();
8479
8599
  // Full run cancellation — exit immediately
8480
8600
  throw turnError;
8481
8601
  }
8482
- settleChannelEgressLatch(currentWirePayload.chatId);
8602
+ channelEgressLatches.settle(currentWirePayload.chatId);
8483
8603
  // OOM errors must escape the turn loop so the task runtime can
8484
8604
  // honor `retry.outOfMemory.machine` (set on chat.agent via
8485
8605
  // `oomMachine`). Catching them here would keep the dead worker
@@ -8600,6 +8720,7 @@ function chatAgent(options) {
8600
8720
  writer.write({ type: "error", errorText });
8601
8721
  });
8602
8722
  // Signal turn complete so the client knows this turn is done
8723
+ postTurnHookLatches.open(currentWirePayload.chatId);
8603
8724
  errorTurnCompleteResult = await writeTurnCompleteChunk(currentWirePayload.chatId);
8604
8725
  // A later action's snapshot reuses this cursor, so it has to move
8605
8726
  // here too or that snapshot resumes from before the failed turn.
@@ -8800,6 +8921,7 @@ function chatAgent(options) {
8800
8921
  });
8801
8922
  }
8802
8923
  }
8924
+ postTurnHookLatches.settle(currentWirePayload.chatId);
8803
8925
  if (locals_js_1.locals.get(chatCloseRequestedKey)) {
8804
8926
  await performChatClose();
8805
8927
  return;
@@ -8855,6 +8977,12 @@ function chatAgent(options) {
8855
8977
  // reaching one of the loop's close checks (a turn timeout, an OOM
8856
8978
  // re-throw). `performChatClose` is idempotent, so the ordinary path
8857
8979
  // having already run it costs nothing here.
8980
+ // Any exit that skips a turn's own settle (an OOM re-throw, a throw from the turn's
8981
+ // error handler) leaves its post-turn latches open, and the module-level entries would
8982
+ // outlive this run and block the next one for the same chat. Settling is a no-op when
8983
+ // nothing is open, and runs first so a throwing close cannot skip it.
8984
+ channelEgressLatches.settle(currentWirePayload.chatId, "run-ended");
8985
+ postTurnHookLatches.settle(currentWirePayload.chatId, "run-ended");
8858
8986
  await performChatClose();
8859
8987
  // `stopSub` is registered post-preload so the close-during-preload
8860
8988
  // early-return path may exit before it ever attached. Guard the
@@ -9759,6 +9887,10 @@ async function flushDeferredChatCustomAgentClientDataErrors() {
9759
9887
  }
9760
9888
  }
9761
9889
  }
9890
+ /** Whether a buffered chunk is a stream-level `error` chunk. */
9891
+ function isErrorUIMessageChunk(chunk) {
9892
+ return chunk.type === "error";
9893
+ }
9762
9894
  /**
9763
9895
  * Pass every chunk through untouched while recording it in `buffer`. Handles
9764
9896
  * both the `AsyncIterable` and `ReadableStream` shapes `toUIMessageStream()`
@@ -9940,6 +10072,7 @@ class ChatMessageAccumulator {
9940
10072
  const ui = replace ? [] : [...this.uiMessages];
9941
10073
  const lane = replace ? [] : [...this.modelMessages];
9942
10074
  let appendedModelCount = 0;
10075
+ let appendedUiCount = 0;
9943
10076
  let rebuild = replace;
9944
10077
  for (const raw of messages) {
9945
10078
  const message = raw.role === "assistant" ? cleanupAbortedParts(raw) : raw;
@@ -9952,13 +10085,18 @@ class ChatMessageAccumulator {
9952
10085
  const idx = ui.findIndex((m) => m.id === resolved.id);
9953
10086
  const previous = ui[idx];
9954
10087
  ui[idx] = mergeIncomingIntoHydrated(previous, resolved);
9955
- if (!rebuild && !(await replaceModelRun(lane, previous, ui[idx], appendedModelCount))) {
10088
+ // The lane's tail is only this message's place when nothing came after it
10089
+ // before this call's own appends.
10090
+ const isLast = idx === ui.length - 1 - appendedUiCount;
10091
+ if (!rebuild &&
10092
+ !(await replaceModelRun(lane, previous, ui[idx], appendedModelCount, undefined, isLast))) {
9956
10093
  rebuild = true;
9957
10094
  }
9958
10095
  continue;
9959
10096
  }
9960
10097
  ui.push(message);
9961
10098
  if (!rebuild) {
10099
+ appendedUiCount++;
9962
10100
  const converted = await toModelMessages([message]);
9963
10101
  lane.push(...converted);
9964
10102
  appendedModelCount += converted.length;
@@ -10131,13 +10269,14 @@ class ChatMessageAccumulator {
10131
10269
  const comp = this._compaction;
10132
10270
  const pm = this._pendingMessages;
10133
10271
  const queue = this._steeringQueue;
10134
- return (0, steeringContext_js_1.retainStepMessages)(async ({ messages, steps }) => {
10272
+ return (0, steeringContext_js_1.retainStepMessages)(async ({ messages, steps, responseMessages, }) => {
10135
10273
  let resultMessages;
10136
10274
  // 1. Compaction
10137
10275
  if (comp) {
10138
10276
  const result = await chatCompact(messages, steps, {
10277
+ responseMessages,
10139
10278
  shouldCompact: comp.shouldCompact,
10140
- summarize: (msgs) => comp.summarize({ messages: msgs, source: "inner" }),
10279
+ summarize: innerSummarize(comp, steps),
10141
10280
  });
10142
10281
  if (result.type !== "skipped") {
10143
10282
  resultMessages = result.messages;
@@ -10156,9 +10295,11 @@ class ChatMessageAccumulator {
10156
10295
  }
10157
10296
  /**
10158
10297
  * Run outer-loop compaction if needed. Call after adding the response
10159
- * and capturing usage. Pass the LAST step's usage (`result.usage`), which is
10160
- * the context the model held on its final call; `result.totalUsage` sums every
10161
- * step of a tool-using turn and belongs in `context.turnUsage`. Applies
10298
+ * and capturing usage. Pass the LAST step's usage
10299
+ * (`(await result.steps).at(-1)?.usage`), which is the context the model held on
10300
+ * its final call. `result.usage` is that on AI SDK 6 but the turn's sum on AI SDK 7,
10301
+ * so read `steps`. The turn's sum (`result.totalUsage` on AI SDK 6, `result.usage`
10302
+ * on AI SDK 7) belongs in `context.turnUsage`. Applies
10162
10303
  * `compactModelMessages` and `compactUIMessages` callbacks if configured.
10163
10304
  *
10164
10305
  * @returns `true` if compaction was performed, `false` otherwise.
@@ -10166,6 +10307,13 @@ class ChatMessageAccumulator {
10166
10307
  async compactIfNeeded(usage, context) {
10167
10308
  if (!this._compaction || !usage)
10168
10309
  return false;
10310
+ // An explicit context wins; inside a chat turn the turn's own fills in what it leaves out.
10311
+ const turnContext = chatEventContext();
10312
+ const eventContext = {
10313
+ chatId: context?.chatId ?? turnContext.chatId,
10314
+ turn: context?.turn ?? turnContext.turn,
10315
+ clientData: context?.clientData ?? turnContext.clientData,
10316
+ };
10169
10317
  const shouldTrigger = await this._compaction.shouldCompact({
10170
10318
  messages: this.modelMessages,
10171
10319
  totalTokens: usage.totalTokens,
@@ -10174,9 +10322,7 @@ class ChatMessageAccumulator {
10174
10322
  usage,
10175
10323
  turnUsage: context?.turnUsage,
10176
10324
  totalUsage: context?.totalUsage,
10177
- chatId: context?.chatId,
10178
- turn: context?.turn,
10179
- clientData: context?.clientData,
10325
+ ...eventContext,
10180
10326
  source: "outer",
10181
10327
  });
10182
10328
  if (!shouldTrigger)
@@ -10185,18 +10331,16 @@ class ChatMessageAccumulator {
10185
10331
  messages: this.modelMessages,
10186
10332
  usage,
10187
10333
  totalUsage: context?.totalUsage,
10188
- chatId: context?.chatId,
10189
- turn: context?.turn,
10190
- clientData: context?.clientData,
10334
+ ...eventContext,
10191
10335
  source: "outer",
10192
10336
  });
10193
10337
  const compactEvent = {
10194
10338
  summary,
10195
10339
  uiMessages: this.uiMessages,
10196
10340
  modelMessages: this.modelMessages,
10197
- chatId: context?.chatId ?? "",
10198
- turn: context?.turn ?? 0,
10199
- clientData: context?.clientData,
10341
+ chatId: eventContext.chatId ?? "",
10342
+ turn: eventContext.turn ?? 0,
10343
+ clientData: eventContext.clientData,
10200
10344
  source: "outer",
10201
10345
  };
10202
10346
  this.modelMessages = this._compaction.compactModelMessages
@@ -10494,6 +10638,7 @@ function createChatSession(payload, options) {
10494
10638
  message: lastUIMessage,
10495
10639
  chatId: currentPayload.chatId,
10496
10640
  turn,
10641
+ clientData: chatEventContext().clientData,
10497
10642
  });
10498
10643
  }
10499
10644
  catch {
@@ -10650,18 +10795,7 @@ function createChatSession(payload, options) {
10650
10795
  /* non-fatal */
10651
10796
  }
10652
10797
  }
10653
- const lastStepUsagePromise = source.usage;
10654
- if (typeof lastStepUsagePromise?.then === "function") {
10655
- try {
10656
- lastStepUsage = (await Promise.race([
10657
- lastStepUsagePromise,
10658
- new Promise((r) => setTimeout(() => r(undefined), 2_000)),
10659
- ]));
10660
- }
10661
- catch {
10662
- /* non-fatal */
10663
- }
10664
- }
10798
+ lastStepUsage = await readLastStepUsage(source);
10665
10799
  // Outer-loop compaction (same logic as chat.agent)
10666
10800
  if (sessionCompaction && turnUsage && !turnObj.stopped) {
10667
10801
  const contextUsage = lastStepUsage ?? turnUsage;
@@ -10673,9 +10807,7 @@ function createChatSession(payload, options) {
10673
10807
  usage: contextUsage,
10674
10808
  turnUsage,
10675
10809
  totalUsage: cumulativeUsage,
10676
- chatId: currentPayload.chatId,
10677
- turn,
10678
- clientData: currentPayload.metadata,
10810
+ ...chatEventContext(),
10679
10811
  source: "outer",
10680
10812
  });
10681
10813
  if (shouldTrigger) {
@@ -10683,9 +10815,7 @@ function createChatSession(payload, options) {
10683
10815
  messages: accumulator.modelMessages,
10684
10816
  usage: turnUsage,
10685
10817
  totalUsage: cumulativeUsage,
10686
- chatId: currentPayload.chatId,
10687
- turn,
10688
- clientData: currentPayload.metadata,
10818
+ ...chatEventContext(),
10689
10819
  source: "outer",
10690
10820
  });
10691
10821
  const compactEvent = {
@@ -10729,12 +10859,13 @@ function createChatSession(payload, options) {
10729
10859
  const hasPending = !!sessionPendingMessages;
10730
10860
  if (!hasCompaction && !hasPending)
10731
10861
  return undefined;
10732
- return (0, steeringContext_js_1.retainStepMessages)(async ({ messages: stepMsgs, steps, }) => {
10862
+ return (0, steeringContext_js_1.retainStepMessages)(async ({ messages: stepMsgs, steps, responseMessages, }) => {
10733
10863
  let resultMessages;
10734
10864
  if (sessionCompaction) {
10735
10865
  const compactResult = await chatCompact(stepMsgs, steps, {
10866
+ responseMessages,
10736
10867
  shouldCompact: sessionCompaction.shouldCompact,
10737
- summarize: (msgs) => sessionCompaction.summarize({ messages: msgs, source: "inner" }),
10868
+ summarize: innerSummarize(sessionCompaction, steps),
10738
10869
  });
10739
10870
  if (compactResult.type !== "skipped") {
10740
10871
  resultMessages = compactResult.messages;
@@ -11130,7 +11261,7 @@ function overrideRequestHeaders(authHeaders) {
11130
11261
  async function callSessionsCreateWithOverride(args) {
11131
11262
  const apiClient = v3_1.apiClientManager.client;
11132
11263
  if (!apiClient) {
11133
- throw new Error("chat.createStartSessionAction: no API access token configured. Set TRIGGER_SECRET_KEY or call apiClientManager.setGlobalAPIClientConfiguration before invoking the action.");
11264
+ throw new Error("chat.createStartSessionAction: no API access token configured. Call configure({ accessToken }) or pass apiClient: { accessToken } to createStartSessionAction before invoking the action.");
11134
11265
  }
11135
11266
  const authHeaders = apiClient.getAuthenticationHeaders("createSession");
11136
11267
  const ctx = { endpoint: "sessions", chatId: args.chatId };
@@ -11154,55 +11285,60 @@ async function callSessionsCreateWithOverride(args) {
11154
11285
  async function mintPublicTokenWithOverride(args) {
11155
11286
  const apiClient = v3_1.apiClientManager.client;
11156
11287
  if (!apiClient) {
11157
- throw new Error("chat.createStartSessionAction: no API access token configured for JWT mint.");
11288
+ throw new Error("chat.createStartSessionAction: no API access token configured for public-token minting. Call configure({ accessToken }) or pass apiClient: { accessToken } to createStartSessionAction.");
11158
11289
  }
11159
11290
  const authHeaders = apiClient.getAuthenticationHeaders("createPublicToken");
11160
- const accessToken = authHeaders.Authorization.slice("Bearer ".length);
11161
11291
  const ctx = { endpoint: "auth", chatId: args.chatId };
11162
11292
  const scopes = [`read:sessions:${args.chatId}`, `write:sessions:${args.chatId}`];
11163
- const serverMint = (0, v3_1.isAdditionalApiKey)(accessToken) || (0, v3_1.isRunToken)(accessToken);
11164
- const endpoint = serverMint ? "/api/v1/auth/public-tokens" : "/api/v1/auth/jwt/claims";
11165
- const url = `${resolveChatStartBaseURL("auth", args.chatId, args.baseURLOption)}${endpoint}`;
11293
+ const baseURL = resolveChatStartBaseURL("auth", args.chatId, args.baseURLOption);
11294
+ const url = `${baseURL}/api/v1/auth/public-tokens`;
11166
11295
  const init = {
11167
11296
  method: "POST",
11168
11297
  headers: overrideRequestHeaders(authHeaders),
11169
- ...(serverMint
11170
- ? {
11171
- body: JSON.stringify({
11172
- scopes,
11173
- expirationTime: args.expirationTime instanceof Date
11174
- ? Math.floor(args.expirationTime.getTime() / 1000)
11175
- : args.expirationTime,
11176
- }),
11177
- }
11178
- : {}),
11298
+ body: JSON.stringify({
11299
+ scopes,
11300
+ expirationTime: args.expirationTime instanceof Date
11301
+ ? Math.floor(args.expirationTime.getTime() / 1000)
11302
+ : args.expirationTime,
11303
+ }),
11179
11304
  };
11180
11305
  const response = args.fetchOverride
11181
11306
  ? await args.fetchOverride(url, init, ctx)
11182
11307
  : await fetch(url, init);
11183
11308
  if (!response.ok) {
11184
- if (serverMint && response.status === 404) {
11185
- const credentialType = (0, v3_1.isRunToken)(accessToken) ? "run token" : "additional API key";
11186
- throw new Error(`This ${credentialType} cannot self-sign public tokens, and the server does not support public-token minting. Upgrade the server or use the root API key.`);
11309
+ const accessToken = authHeaders.Authorization.slice("Bearer ".length);
11310
+ if ((0, v3_1.canFallbackToLocalPublicToken)(accessToken, scopes, response.status)) {
11311
+ await response.arrayBuffer();
11312
+ const claimsURL = `${baseURL}/api/v1/auth/jwt/claims`;
11313
+ const claimsInit = {
11314
+ method: "POST",
11315
+ headers: overrideRequestHeaders(authHeaders),
11316
+ };
11317
+ const claimsResponse = args.fetchOverride
11318
+ ? await args.fetchOverride(claimsURL, claimsInit, ctx)
11319
+ : await fetch(claimsURL, claimsInit);
11320
+ if (!claimsResponse.ok) {
11321
+ const text = await claimsResponse.text().catch(() => "");
11322
+ throw new Error(`auth.generateJWTClaims failed: ${claimsResponse.status} ${text}`);
11323
+ }
11324
+ const claims = (await claimsResponse.json());
11325
+ return (0, v3_1.generateJWT)({
11326
+ secretKey: accessToken,
11327
+ payload: { ...claims, scopes },
11328
+ expirationTime: args.expirationTime,
11329
+ });
11330
+ }
11331
+ if (response.status === 404) {
11332
+ throw new Error("The server does not support public-token minting for this credential. Upgrade your self-hosted server.");
11187
11333
  }
11188
11334
  const text = await response.text().catch(() => "");
11189
11335
  throw new Error(`auth.createPublicToken failed: ${response.status} ${text}`);
11190
11336
  }
11191
11337
  const responseBody = (await response.json());
11192
- if (serverMint) {
11193
- if (typeof responseBody.token !== "string") {
11194
- throw new Error("auth.createPublicToken failed: server response did not include a token");
11195
- }
11196
- return responseBody.token;
11338
+ if (typeof responseBody.token !== "string") {
11339
+ throw new Error("auth.createPublicToken failed: server response did not include a token");
11197
11340
  }
11198
- return (0, v3_1.generateJWT)({
11199
- secretKey: accessToken,
11200
- payload: {
11201
- ...responseBody,
11202
- scopes,
11203
- },
11204
- expirationTime: args.expirationTime,
11205
- });
11341
+ return responseBody.token;
11206
11342
  }
11207
11343
  /**
11208
11344
  * Creates a server-side helper that reads a conversation from a transcript