authorgravatar for 71498452+StiensWout@users.noreply.github.comWout Stiens <71498452+StiensWout@users.noreply.github.com> 2026-09-07 23:58:28+02:00
committergravatar for noreply@github.comGitHub <noreply@github.com> 2026-09-07 14:58:28-07:00
log3bf74eb6dce3e1372cfafc2ca84825f4b13bae6a
treee988639a808872aea02c84aa2d1f09d3f8492f9d
parent0d34579d674920cc47fc5c908494f51ed3895204
signature Signed by PGP key B5690EEEBB952194

fix(claude): report usage limits on retried turns (#10549)

2 files changed, 180 insertions(+), 2 deletions(-)

apps/server/src/provider/Layers/ClaudeAdapter.test.ts+173-1
......@@ -2367,6 +2367,139 @@ describe("ClaudeAdapterLive", () => {
23672367 );
23682368 });
23692369
2370 const usageLimitMessage =
2371 "Claude usage limit reached. Send the message again once the limit resets.";
2372 const genericApiErrorMessage = "Claude gave up after repeated API errors.";
2373 const rateLimitAssistant = {
2374 type: "assistant",
2375 session_id: "sdk-session-limit",
2376 uuid: "assistant-limit",
2377 parent_tool_use_id: null,
2378 error: "rate_limit",
2379 message: {
2380 id: "assistant-message-limit",
2381 model: "<synthetic>",
2382 content: [{ type: "text", text: "You've hit your session limit" }],
2383 },
2384 };
2385 const rateLimitResult = {
2386 type: "result",
2387 subtype: "success",
2388 is_error: true,
2389 terminal_reason: "api_error",
2390 session_id: "sdk-session-limit",
2391 uuid: "result-limit",
2392 };
2393
2394 it.effect.each([
2395 {
2396 name: "an assistant-only rate limit",
2397 messages: [rateLimitAssistant],
2398 expected: usageLimitMessage,
2399 },
2400 {
2401 name: "a normal parent response after a rate limit",
2402 messages: [rateLimitAssistant, { ...rateLimitAssistant, error: undefined }],
2403 expected: genericApiErrorMessage,
2404 },
2405 {
2406 name: "a server error after a rate limit",
2407 messages: [rateLimitAssistant, { ...rateLimitAssistant, error: "server_error" }],
2408 expected: genericApiErrorMessage,
2409 },
2410 {
2411 name: "a subagent rate limit",
2412 messages: [{ ...rateLimitAssistant, parent_tool_use_id: "nested-tool" }],
2413 expected: genericApiErrorMessage,
2414 },
2415 {
2416 name: "a subagent response after a parent rate limit",
2417 messages: [
2418 rateLimitAssistant,
2419 { ...rateLimitAssistant, error: undefined, parent_tool_use_id: "nested-tool" },
2420 ],
2421 expected: usageLimitMessage,
2422 },
2423 ])("classifies the terminal API failure after $name", ({ messages, expected }) => {
2424 const harness = makeHarness();
2425 return Effect.gen(function* () {
2426 const adapter = yield* ClaudeAdapter;
2427 const eventsFiber = yield* adapter.streamEvents.pipe(
2428 Stream.takeUntil((event) => event.type === "turn.completed"),
2429 Stream.runCollect,
2430 Effect.forkChild,
2431 );
2432 const session = yield* adapter.startSession({
2433 threadId: THREAD_ID,
2434 provider: ProviderDriverKind.make("claudeAgent"),
2435 runtimeMode: "full-access",
2436 });
2437 yield* adapter.sendTurn({ threadId: session.threadId, input: "hello", attachments: [] });
2438 for (const [index, message] of messages.entries()) {
2439 harness.query.emit({ ...message, uuid: `assistant-${index}` } as unknown as SDKMessage);
2440 }
2441 harness.query.emit(rateLimitResult as unknown as SDKMessage);
2442
2443 const events = Array.from(yield* Fiber.join(eventsFiber));
2444 const errors = events.filter((event) => event.type === "runtime.error");
2445 assert.equal(errors.length, 1);
2446 assert.equal(errors[0]?.payload.message, expected);
2447 assert.equal(completedTurn(events).state, "failed");
2448 assert.equal(completedTurn(events).errorMessage, expected);
2449 }).pipe(
2450 Effect.provideService(Random.Random, makeDeterministicRandomService()),
2451 Effect.provide(harness.layer),
2452 );
2453 });
2454
2455 it.effect("names repeated usage limits without carrying them into a later turn", () => {
2456 const harness = makeHarness();
2457 return Effect.gen(function* () {
2458 const adapter = yield* ClaudeAdapter;
2459 const session = yield* adapter.startSession({
2460 threadId: THREAD_ID,
2461 provider: ProviderDriverKind.make("claudeAgent"),
2462 runtimeMode: "full-access",
2463 });
2464 for (const [index, expected] of [
2465 usageLimitMessage,
2466 usageLimitMessage,
2467 genericApiErrorMessage,
2468 ].entries()) {
2469 const eventsFiber = yield* adapter.streamEvents.pipe(
2470 Stream.takeUntil((event) => event.type === "turn.completed"),
2471 Stream.runCollect,
2472 Effect.forkChild,
2473 );
2474 yield* adapter.sendTurn({ threadId: session.threadId, input: "again", attachments: [] });
2475 if (index === 0) {
2476 harness.query.emit({
2477 type: "rate_limit_event",
2478 rate_limit_info: { status: "rejected", rateLimitType: "five_hour" },
2479 session_id: "sdk-session-limit",
2480 uuid: "limit-rejected",
2481 } as unknown as SDKMessage);
2482 }
2483 if (index < 2) {
2484 harness.query.emit({
2485 ...rateLimitAssistant,
2486 uuid: `assistant-limit-${index}`,
2487 } as unknown as SDKMessage);
2488 }
2489 harness.query.emit({
2490 ...rateLimitResult,
2491 uuid: `result-limit-${index}`,
2492 } as unknown as SDKMessage);
2493 const payload = completedTurn(Array.from(yield* Fiber.join(eventsFiber)));
2494 assert.equal(payload.state, "failed");
2495 assert.equal(payload.errorMessage, expected);
2496 }
2497 }).pipe(
2498 Effect.provideService(Random.Random, makeDeterministicRandomService()),
2499 Effect.provide(harness.layer),
2500 );
2501 });
2502
23702503 it.effect.each([
23712504 {
23722505 name: "listed error with api_error",
......@@ -2418,6 +2551,43 @@ describe("ClaudeAdapterLive", () => {
24182551 expected: /repeated API errors/,
24192552 expectedState: "failed",
24202553 },
2554 {
2555 name: "assistant rate limit followed by overload",
2556 evidence: "assistant-rate-limit",
2557 result: {
2558 subtype: "success",
2559 is_error: true,
2560 terminal_reason: "api_error",
2561 api_error_status: 529,
2562 errors: [],
2563 },
2564 expected: /overloaded \(529\)/,
2565 expectedState: "failed",
2566 },
2567 {
2568 name: "assistant rate limit followed by a listed error",
2569 evidence: "assistant-rate-limit",
2570 result: {
2571 subtype: "success",
2572 is_error: true,
2573 terminal_reason: "api_error",
2574 errors: ["Tool execution failed: EACCES"],
2575 },
2576 expected: /EACCES/,
2577 expectedState: "failed",
2578 },
2579 {
2580 name: "assistant rate limit followed by an interrupt",
2581 evidence: "assistant-rate-limit",
2582 result: {
2583 subtype: "error_during_execution",
2584 is_error: true,
2585 terminal_reason: "aborted_tools",
2586 errors: [],
2587 },
2588 expected: undefined,
2589 expectedState: "interrupted",
2590 },
24212591 ...[
24222592 "recovered-missing-reset",
24232593 "recovered-next-reset",
......@@ -2465,7 +2635,9 @@ describe("ClaudeAdapterLive", () => {
24652635 input: "synthetic hello",
24662636 attachments: [],
24672637 });
2468 if (evidence === "auth" || evidence === "nested-auth") {
2638 if (evidence === "assistant-rate-limit") {
2639 harness.query.emit(rateLimitAssistant as unknown as SDKMessage);
2640 } else if (evidence === "auth" || evidence === "nested-auth") {
24692641 harness.query.emit({
24702642 type: "assistant",
24712643 session_id: "sdk-audit",
apps/server/src/provider/Layers/ClaudeAdapter.ts+7-1
......@@ -161,6 +161,7 @@ interface ClaudeTurnState {
161161 nextSyntheticAssistantBlockIndex: number;
162162 authenticationFailureMessage: string | undefined;
163163 rejectedRateLimitTypes: Set<string>;
164 latestAssistantRateLimited: boolean;
164165}
165166
166167interface AssistantTextBlockState {
......@@ -3168,6 +3169,7 @@ export const makeClaudeAdapter = Effect.fn("makeClaudeAdapter")(function* (
31683169 nextSyntheticAssistantBlockIndex: -1,
31693170 authenticationFailureMessage: undefined,
31703171 rejectedRateLimitTypes: new Set(),
3172 latestAssistantRateLimited: false,
31713173 };
31723174 context.session = {
31733175 ...context.session,
......@@ -3226,6 +3228,9 @@ export const makeClaudeAdapter = Effect.fn("makeClaudeAdapter")(function* (
32263228 }
32273229
32283230 if (context.turnState) {
3231 // Limited retries may only carry an assistant error, without a new window
3232 // event. Later parent responses replace this evidence if the turn recovers.
3233 context.turnState.latestAssistantRateLimited = message.error === "rate_limit";
32293234 // The CLI can report authentication failure before ending the turn as a
32303235 // generic API error, so retain that evidence for the result fallback.
32313236 if (message.error === "authentication_failed") {
......@@ -3263,7 +3268,7 @@ export const makeClaudeAdapter = Effect.fn("makeClaudeAdapter")(function* (
32633268 const turn = context.turnState;
32643269 const failureHint =
32653270 turn?.authenticationFailureMessage ??
3266 (turn && turn.rejectedRateLimitTypes.size > 0
3271 (turn && (turn.rejectedRateLimitTypes.size > 0 || turn.latestAssistantRateLimited)
32673272 ? "Claude usage limit reached. Send the message again once the limit resets."
32683273 : undefined);
32693274 const { status, errorMessage } = resultOutcome(message, failureHint);
......@@ -4949,6 +4954,7 @@ export const makeClaudeAdapter = Effect.fn("makeClaudeAdapter")(function* (
49494954 nextSyntheticAssistantBlockIndex: -1,
49504955 authenticationFailureMessage: undefined,
49514956 rejectedRateLimitTypes: new Set(),
4957 latestAssistantRateLimited: false,
49524958 };
49534959
49544960 const updatedAt = yield* nowIso;