fix(claude): retain final usage when an API-error turn fails

Publish the converted result totals before the failure event so the
existing usage handler preserves work done earlier in the turn.

packages/server/src/server/agent/providers/claude/agent.error-result.test.ts

@@ -0,0 +1,190 @@
+import { expect, test } from "vitest";
+import { createTestLogger } from "../../../../test-utils/test-logger.js";
+import { collectSessionTurnEvents } from "../test-utils/session-stream-adapter.js";
+import { ClaudeAgentClient } from "./agent.js";
+import type { ClaudeQueryFactory } from "./query.js";
+
+const LIMIT = "You've hit your session limit · resets 5:20pm (Asia/Singapore)";
+
+interface ScriptedTurn {
+  // An API error reply as Claude Code writes it: a synthetic assistant message with `error`.
+  errorReply?: string;
+  // Assistant text the model wrote before or after it.
+  before?: string;
+  after?: string;
+  isError: boolean;
+  usage?: {
+    inputTokens: number;
+    cachedInputTokens: number;
+    outputTokens: number;
+    totalCostUsd: number;
+  };
+}
+
+// A query dependency that speaks Claude Code's stream interface, one scripted turn per prompt.
+function scriptedQuery(turns: ScriptedTurn[]): ClaudeQueryFactory {
+  return ({ prompt }) => {
+    let index = 0;
+    const assistant = (id: string, text: string, error?: string) => ({
+      type: "assistant",
+      session_id: "error-session",
+      parent_tool_use_id: null,
+      uuid: id,
+      message: {
+        id,
+        role: "assistant",
+        model: error ? "<synthetic>" : "claude-opus-5-5",
+        content: [{ type: "text", text }],
+        stop_reason: "stop_sequence",
+        usage: { input_tokens: 0, output_tokens: 0 },
+      },
+      ...(error ? { error } : {}),
+    });
+    const messages = (async function* () {
+      if (typeof prompt === "string") throw new Error("Expected streaming input");
+      for await (const _message of prompt) {
+        const turn = turns[index];
+        index += 1;
+        yield {
+          type: "system",
+          subtype: "init",
+          session_id: "error-session",
+          permissionMode: "default",
+          model: "claude-opus-5-5",
+        };
+        if (turn.before) yield assistant(`before-${index}`, turn.before);
+        if (turn.errorReply) yield assistant(`error-${index}`, turn.errorReply, "rate_limit");
+        if (turn.after) yield assistant(`after-${index}`, turn.after);
+        yield {
+          type: "result",
+          subtype: "success",
+          session_id: "error-session",
+          duration_ms: 1,
+          duration_api_ms: 0,
+          is_error: turn.isError,
+          num_turns: 1,
+          result: turn.isError ? (turn.errorReply ?? "") : (turn.after ?? turn.errorReply ?? ""),
+          stop_reason: "stop_sequence",
+          total_cost_usd: turn.usage?.totalCostUsd ?? 0,
+          usage: {
+            input_tokens: turn.usage?.inputTokens ?? 0,
+            output_tokens: turn.usage?.outputTokens ?? 0,
+            cache_read_input_tokens: turn.usage?.cachedInputTokens ?? 0,
+          },
+          modelUsage: {},
+          permission_denials: [],
+          uuid: `result-${index}`,
+        };
+      }
+    })();
+    return Object.assign(messages, {
+      close: () => {
+        void messages.return();
+      },
+      interrupt: async () => undefined,
+      setModel: async () => undefined,
+      setPermissionMode: async () => undefined,
+      supportedModels: async () => [],
+      supportedCommands: async () => [],
+    }) as ReturnType<ClaudeQueryFactory>;
+  };
+}
+
+async function runTurns(turns: ScriptedTurn[]) {
+  const client = new ClaudeAgentClient({
+    logger: createTestLogger(),
+    queryFactory: scriptedQuery(turns),
+    resolveBinary: async () => process.execPath,
+  });
+  const session = await client.createSession({
+    provider: "claude",
+    cwd: process.cwd(),
+  });
+  try {
+    const results = [];
+    for (const [index] of turns.entries()) {
+      results.push(await collectSessionTurnEvents(session, `turn ${index + 1}`));
+    }
+    return results;
+  } finally {
+    await session.close();
+  }
+}
+
+function assistantTexts(events: Awaited<ReturnType<typeof runTurns>>[number]): string[] {
+  return events.flatMap((event) =>
+    event.type === "timeline" && event.item.type === "assistant_message" ? [event.item.text] : [],
+  );
+}
+
+function terminal(events: Awaited<ReturnType<typeof runTurns>>[number]) {
+  return events.find(
+    (event) =>
+      event.type === "turn_completed" ||
+      event.type === "turn_failed" ||
+      event.type === "turn_canceled",
+  );
+}
+
+test("a turn Claude Code ends on an API error fails with the error, shown once", async () => {
+  const [limited, overloaded, next] = await runTurns([
+    { errorReply: LIMIT, isError: true },
+    {
+      before: "Checking the deploy doc.",
+      errorReply: "API Error: 529 Overloaded",
+      isError: true,
+    },
+    { after: "Done.", isError: false },
+  ]);
+
+  expect(terminal(limited)).toMatchObject({
+    type: "turn_failed",
+    error: LIMIT,
+  });
+  expect(assistantTexts(limited)).toEqual([]);
+
+  expect(terminal(overloaded)).toMatchObject({
+    type: "turn_failed",
+    error: "API Error: 529 Overloaded",
+  });
+  expect(assistantTexts(overloaded)).toEqual(["Checking the deploy doc."]);
+
+  // The session keeps working after a failed turn.
+  expect(terminal(next)).toMatchObject({ type: "turn_completed" });
+  expect(assistantTexts(next)).toEqual(["Done."]);
+});
+
+test("a failed turn retains its final token counts and cost", async () => {
+  const usage = {
+    inputTokens: 120,
+    cachedInputTokens: 80,
+    outputTokens: 30,
+    totalCostUsd: 0.012,
+  };
+  const [failed] = await runTurns([
+    { before: "Checking the deploy doc.", errorReply: LIMIT, isError: true, usage },
+  ]);
+
+  expect(failed).toContainEqual(
+    expect.objectContaining({ type: "usage_updated", usage: expect.objectContaining(usage) }),
+  );
+  expect(terminal(failed)).toMatchObject({ type: "turn_failed", error: LIMIT });
+  expect(assistantTexts(failed)).toEqual(["Checking the deploy doc."]);
+});
+
+test("an error reply the turn moves past is still shown", async () => {
+  const [recovered, completed] = await runTurns([
+    {
+      errorReply: "API Error: 529 Overloaded",
+      after: "Recovered.",
+      isError: false,
+    },
+    { errorReply: LIMIT, isError: false },
+  ]);
+
+  expect(terminal(recovered)).toMatchObject({ type: "turn_completed" });
+  expect(assistantTexts(recovered)).toEqual(["API Error: 529 Overloaded", "Recovered."]);
+
+  expect(terminal(completed)).toMatchObject({ type: "turn_completed" });
+  expect(assistantTexts(completed)).toEqual([LIMIT]);
+});

packages/server/src/server/agent/providers/claude/agent.ts

@@ -2389,6 +2389,12 @@ class ClaudeAgentSession implements AgentSession {
   private mainTurnInFlight = false;
   private foregroundHasVisibleActivity = false;
   private activeTurnHasAssistantText = false;
+  /**
+   * The reply Claude Code writes when a turn stops on an API error ("You've hit your session
+   * limit", "API Error: 529 ..."), held back until the result says whether the turn failed: a
+   * failed turn shows the same text as its error, so showing both would repeat it.
+   */
+  private withheldErrorReply: AgentTimelineItem[] = [];
   private readonly contextUsage: ClaudeContextUsageState;
   private userMessageIds: string[] = [];
   private readonly emittedUserMessageIds = new Set<string>();
@@ -2531,6 +2537,7 @@ class ClaudeAgentSession implements AgentSession {
     this.activeForegroundTurnId = turnId;
     this.foregroundHasVisibleActivity = false;
     this.activeTurnHasAssistantText = false;
+    this.withheldErrorReply = [];
     this.contextUsage.beginTurn();
     this.transitionTurnState("foreground", "foreground turn started");
     this.clearRecentStderr();
@@ -4419,6 +4426,21 @@ class ClaudeAgentSession implements AgentSession {
               }) satisfies AgentStreamEvent,
           );
 
+    if (!readClaudeParentToolUseId(message)) {
+      if (message.type === "assistant" && message.error) {
+        this.withheldErrorReply.push(...assistantTimelineEvents.map((event) => event.item));
+        return messageEvents;
+      }
+      if (
+        message.type === "assistant" ||
+        message.type === "user" ||
+        message.type === "stream_event"
+      ) {
+        const released: AgentStreamEvent[] = [];
+        this.releaseWithheldErrorReply(released);
+        return [...released, ...messageEvents, ...assistantTimelineEvents];
+      }
+    }
     return [...messageEvents, ...assistantTimelineEvents];
   }
 
@@ -4966,6 +4988,23 @@ class ClaudeAgentSession implements AgentSession {
     events: AgentStreamEvent[],
   ): void {
     const usage = this.convertUsage(message);
+    // Claude Code reports a turn that stopped on an API error (a usage limit, an overloaded or
+    // rejected request) as `success` with `is_error`, carrying the error text in `result`.
+    if (message.subtype === "success" && message.is_error) {
+      const withheld = this.withheldErrorReply;
+      this.withheldErrorReply = [];
+      const resultText = typeof message.result === "string" ? message.result.trim() : "";
+      const withheldText = withheld
+        .map((item) => (item.type === "assistant_message" ? item.text : ""))
+        .join("\n")
+        .trim();
+      if (usage) events.push({ type: "usage_updated", provider: "claude", usage });
+      events.push(...this.sidechainTracker.finishAll("failed"));
+      events.push(this.buildTurnFailedEvent(resultText || withheldText || "Claude run failed"));
+      return;
+    }
+    const releasedErrorReply = this.withheldErrorReply.length > 0;
+    this.releaseWithheldErrorReply(events);
     if (message.subtype === "success") {
       events.push(...this.sidechainTracker.finishAll("completed"));
       // Built-in slash commands (e.g. /voice, /usage, "Unknown command: …")
@@ -4975,7 +5014,12 @@ class ClaudeAgentSession implements AgentSession {
       // accounting from provider gateways does not duplicate streamed output.
       const resultText = typeof message.result === "string" ? message.result.trim() : "";
       const outputTokens = message.usage?.output_tokens;
-      if (resultText.length > 0 && outputTokens === 0 && !this.activeTurnHasAssistantText) {
+      if (
+        resultText.length > 0 &&
+        outputTokens === 0 &&
+        !this.activeTurnHasAssistantText &&
+        !releasedErrorReply
+      ) {
         events.push({
           type: "timeline",
           provider: "claude",
@@ -4997,6 +5041,14 @@ class ClaudeAgentSession implements AgentSession {
     events.push(this.buildTurnFailedEvent(errorMessage));
   }
 
+  private releaseWithheldErrorReply(events: AgentStreamEvent[]): void {
+    if (this.withheldErrorReply.length === 0) return;
+    for (const item of this.withheldErrorReply) {
+      events.push({ type: "timeline", item, provider: "claude" });
+    }
+    this.withheldErrorReply = [];
+  }
+
   private createClaudeSessionChangedNotice(
     oldSessionId: string,
     newSessionId: string,