#4834 [Engine] Anthropic errors on /v1/messages use Anthropic's error shape, so the Vercel AI SDK shows Strata's message
closed · @gdevenyi · 0 comentarios · En GitHub
Setup & installServer & APIAMD / HIPNVIDIA / CUDAModels & quantsSecurityDocumentationWindowsLinux
Descripción
## Problem and result
Errors on `/v1/messages` and `/v1/messages/count_tokens` had no top-level `type: "error"`. `@ai-sdk/anthropic` 4.0.94 requires that field. Without it the client falls back to the HTTP status text, so a person sees "Bad Request" and not Strata's explanation.
The same too-long request through `@ai-sdk/anthropic`, against the mock engine:
- Before: `AI_APICallError status 400 message: "Bad Request"`
- After: `AI_APICallError status 400 message: "request exceeds the context window: prompt (140244 tokens) leaves no room to answer in the context (131072); requests are never truncated"`
The body is now:
`{"type": "error", "error": {"type": "invalid_request_error", "message": "..."}, "request_id": "req_..."}`
The response also has a `request-id` header with the same value.
The spec is platform.claude.com/docs/en/api/errors, read on 2026-10-31. Its example shows the top-level `"type": "error"` and a `request_id`. It also says: "Every API response includes a unique `request-id` header."
Who this helps: people who point `@ai-sdk/anthropic` at Strata themselves. The `anthropic` Python SDK already raised `BadRequestError` with the full message. The OpenCode preset uses `@ai-sdk/openai-compatible`, so no shipped preset reaches the failing path. This is a small real bug.
## Related discussion
None. An earlier check for this work, on 2026-10-31, searched for "vercel", "AI SDK" and "ai-sdk" and found no issue with this symptom. It also scanned the diffs of the 52 open pull requests that touch `serve/server.py`. Only #3643 matched. It adds `not_found_error` on its own `/v1/systemone` surface, so it is not a duplicate. I did not repeat those searches for this final revision.
## Approach
One change, at the top of `_json`, for a status of 400 or more on these two paths. The path is read the same way `_post` reads it, so `/v1/v1/...`, a trailing slash and `?beta=true` are covered. Other keys in the error, such as `error.code`, stay.
`error.type` follows the status rows on the errors page:
| Status | `error.type` now | Before |
| --- | --- | --- |
| 400, 401, 429 | `invalid_request_error`, `authentication_error`, `rate_limit_error` | same |
| 403 | `permission_error` | `forbidden` |
| 404 (only a `GET` on these paths) | `not_found_error` | no type |
| 409 (not reached in normal use) | `conflict_error` | `model_busy` |
| 413 | `request_too_large` | `invalid_request_error` |
| 415 and other 4xx | `invalid_request_error` | 415 had no type |
| 500 (not reached in normal use) | `api_error` | no site sends one on these paths |
| 502, 503 | keep their type; the page has no row | same |
A 503 is never called `overloaded_error`. The Claude Code 2.1.316 binary has the check `e.status===529||(e.message?.includes('"type":"overloaded_error"')??!1)`, so that name would likely send a 503 down the overload path. I read this in the binary. I did not run that case.
The id is `req_` plus the request monitor's id when the monitor keeps a record. Otherwise it is a new random id of the same form.
Only the top-level `type` fixes the Vercel problem. `request_id`, the header and the type mapping are conformance only: no client I tested needs them. They can be dropped without losing the fix, if you prefer the smaller change.
Compatibility: this is a bug fix to error bodies on these two paths, and clients that read those bodies will see the change. Success responses and the other APIs do not change. There is no opt-in flag, because a flag would leave the bug in place by default.
## Validation
Revision: base c62d5e3a plus b0ede643. Linux, Python 3.12.3. The tests and local scripts use the mock engine and no GPU, except the real-engine section under Passed.
**Passed:**
- `python -m unittest serve.test_anthropic_base_url`: `Ran 14 tests`, `OK`. Nine are new, in class `ErrorEnvelope`. One uses the real `anthropic` SDK (2.3.0); it is skipped if the package is missing.
- The same tests with `serve/server.py` from the base: `Ran 14 tests`, `FAILED (failures=16)`, counting subtests.
- Each of these edits to the fix makes a test fail: removing the 404, 409 or 500 row, and dropping the `Retry-After` header when `request-id` is added.
- `python -m unittest serve.test_request_hardening serve.test_security serve.test_parallel serve.test_structured serve.test_request_accounting serve.test_monitor`: `Ran 110 tests`, `OK (skipped=2)`. Both skips are `jsonschema is not installed`.
- Local script, not in the repo: the same 19 requests to a base server and a branch server, with ids and timestamps normalized. 17 are identical: `/v1/messages` with and without stream and tools, `/v1/v1/messages`, `count_tokens` success, `/v1/chat/completions`, `/v1/responses`, 400s on chat and responses, 404s on `/v1/messages/batches` and `/v1/v1/chat/completions`, `GET /v1/models`, `GET /v1/nope`. 2 differ, as intended: a 400 on `/v1/messages` and a 400 on `/v1/messages/count_tokens`.
- Local script, not in the repo: `@ai-sdk/anthropic` 4.0.94 on Node 24.9.0, base and branch. The messages are quoted above. The success call worked on both, with the same usage.
- `anthropic` SDK, base and branch: `BadRequestError` on both. `e.request_id` was `None` on the base and `req_<12 hex>` on the branch. On success `_request_id` stays `None`.
- Real engine: RTX 5080 16 GB, engine 0.1.55 from release v0.2.9, Qwen3.8-Flash-Next IQ3_S, context 65536, `--no-api-key`, 127.0.0.1. Local scripts, not in the repo. `anthropic` SDK 2.3.0 and `@ai-sdk/anthropic` 4.0.94. Every run of a row gave the same result:
| Real condition | Runs, base / branch | Base | Branch |
| --- | --- | --- | --- |
| `@ai-sdk/anthropic`, overflow (150012 tokens), with and without stream | 1 / 1 each | `status 400 message: "Bad Request"` | `status 400 message: "request exceeds the context window: prompt (150012 tokens) leaves no room to answer in the context (65536); requests are never truncated"` |
| `@ai-sdk/anthropic`, `GpuBusy` 503, with and without stream | 1 / 1 each | `status 503 message: "Service Unavailable"` | `status 503 message: "the GPU is in use by another program: ..."`, `Retry-After: 15`, `isRetryable` true on both |
| `anthropic` SDK, overflow | 4 / 6, plus 1 / 3 with stream | `BadRequestError`, `e.request_id` `None`, no top-level `type`, no header | `BadRequestError`, top-level `type: "error"`, and `e.request_id`, body `request_id` and the `request-id` header are equal |
| `anthropic` SDK, 400 validation: `stop_sequences`, `temperature`, block type | 1 / 3 each, `temperature` 1 / 4 | as above, same message | as above, same message |
| `anthropic` SDK, `GpuBusy` 503 | 1 / 3 | `InternalServerError`, `server_error`, `e.request_id` `None` | `InternalServerError`, `server_error`, ids equal |
| Raw `GpuBusy` 503, without and with stream | 1 / 3 and 1 / 1 | `Retry-After: 15`, no `request-id` | `Retry-After: 15` and `request-id`; `error.type` stays `server_error` |
| 400 on `count_tokens`, `/v1/v1/messages`, a trailing slash, `?beta=true` | 1 / 1 each, `count_tokens` block type 1 / 3 | no top-level `type` | new body and header |
| 403 foreign `Origin`, on both paths | 1 / 1 each | `forbidden` | `permission_error`, new body and header |
| 404 `GET`, on both paths | 1 / 1 each | `{"error": {"message": "not found"}}` | `not_found_error`, new body and header |
| 413 `Content-Length` of 10 TB, header only, on both paths | 1 / 1 each | `invalid_request_error` | `request_too_large`, new body and header |
| 415 same `Origin`, `text/plain`, on both paths | 1 / 1 each | no type | `invalid_request_error`, new body and header |
How the rare ones were triggered: `GpuBusy` by a second program holding 300 MiB of VRAM, with `min_free_vram_mib` set to 15800 in the test config, so the unloaded model could not reload. After that program stopped, the next request reloaded the model and answered 200 on both, in 15 to 18 s.
- Real engine, other paths: `/v1/messages/batches`, the 400 on `/v1/chat/completions` and the 400 on `/v1/responses` (1 / 3 runs), and the 403 on `/v1/chat/completions` (1 / 1), have the same body on base and branch, with no `request-id`.
- Real engine, success, 1 / 1: `/v1/messages` with and without stream and `count_tokens` have the same header names on base and branch, with no `request-id`. The body keys match, and so do the first three and the last stream events. In the SDK, `_request_id` stays `None`. `@ai-sdk/anthropic` gave the same text, finish reason and usage on both. These real runs compare names and keys only. The byte comparison is the mock-engine script above.
**Failed:**
- `python -m unittest serve.test_server serve.test_responses serve.test_anthropic_base_url serve.test_chat_tool_results serve.test_context_overflow serve.test_images`: `Ran 446 tests`, `FAILED (failures=1, skipped=1)`. The one failure is `serve.test_server.GpuChoice.test_hip_ordinal` (`AssertionError: '0,1' != '1,0'`). It fails the same way on the base, so this change does not cause it.
**Not run:**
- On the real engine: a 401 (the servers ran with `--no-api-key`), a 429 (not tried), and a `GpuBusy` or `StructuredOutputError` after a stream has started (I know no way to trigger one without a patch).
- Other cards, other models and other engine versions.
- Windows and macOS.
- CUDA, HIP, SYCL and Metal builds: Python only, no engine file changed.
- Claude Code against the patched server, so its behaviour on these bodies is not shown.
- The two `serve.test_structured` tests that need `jsonschema`.
- A real 409, 500 or 502 on these paths: the tests force them by making `svc.load` raise.
## Limits
- The `request-id` header is on these error responses only. The page says every API response has one, so success responses still do not meet that sentence. This keeps success responses byte-identical.
- Error events inside a stream keep their shape, with no `request_id`. There is one older problem here. A `GpuBusy` or `StructuredOutputError` raised after a stream starts writes a raw `HTTP/1.0 503` or `502` block into the 200 stream. A local script showed this on both base and branch by raising them in `anthropic_events`; I did not get it from a real engine. This change does not fix it. On the branch the block carries the new body and a `request-id` line.
- `/v1/messages/batches` is left as it was (`{"error": {"message": "not found"}}`), outside these two paths.
- `anthropic` 2.3.0's `ErrorObject` list does not include four types sent here: `conflict_error` and `request_too_large`, which the errors page lists, and `structured_output_failed` and `server_error`, which it does not.
- A web page cannot read `request-id` across origins, because the header is not in `Access-Control-Expose-Headers`. The same id is in the body.
- The monitor's stored response for these errors now shows the new body.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
https://claude.ai/code/session_01W53iURzvDQMjnYemag8uvv
En el sitio
Enlaces a install, modelos, releases.