#5930 [GUI] Return HTTP 400 for malformed inference JSON

closed · @djmarker287-code · 0 评论 · 去 GitHub 看

Setup & installServer & APIAMD / HIPModels & quantsSecurityWindows

说明

## Problem and result

`local-ai-worker` checks whether Responses is available by sending one byte, `{` (`0x7b`), as an uncompressed JSON request body. It expects HTTP 400 or 422. With the LAN route from #5925 applied, Strata returned HTTP 500 with `Request failed`, so the worker stopped at readiness. The same request now returns HTTP 400 with `Invalid JSON in request body`.

## Related discussion

#5925 enables the Responses route on the desktop LAN listener. This fix is based directly on `main` and also applies to inference routes already available there.

## Approach

Give JSON parse errors a distinct type and catch them in the shared inference handler. Return the error in the format used by the requested API. API-key checks and unknown-route handling retain their existing status codes.

## Validation

**Passed**

Live test on Strata Desktop 0.2.13, Windows 11, Radeon AI PRO R9700 (32 GB), HIP engine 0.1.59, with `iq3_s` loaded. The installed host had the LAN change from #5925. Requests were sent from the Strata PC to its LAN address. The curl equivalent of the readiness probe is below. Set `LAN_BASE` to the shared `/v1` URL and add the authorization header if a key is configured.

```sh
curl -i --noproxy '*' -H 'Content-Type: application/json' \
  --data-binary '{' "$LAN_BASE/responses"
```

| Request | Before | After |
| --- | --- | --- |
| LAN `POST /v1/responses` with `{` | `500 Request failed` | `400 Invalid JSON in request body` |
| Loopback `POST /v1/responses` with `{` | `500 Body must be JSON` | `400 Invalid JSON in request body` |
| Unknown route with `{` | `404` | `404` |

The live test also sent these two payloads using Node fetch. These curl commands reproduce them.

```sh
curl -i -H 'Content-Type: application/json' \
  --data-binary '{"model":"iq3_s","input":"Reply with exactly OK.","reasoning":{"effort":"none"},"max_output_tokens":24,"store":false,"stream":false}' \
  "$LAN_BASE/responses"

curl -iN -H 'Content-Type: application/json' \
  --data-binary '{"model":"iq3_s","input":"Reply with exactly OK.","reasoning":{"effort":"none"},"max_output_tokens":24,"store":false,"stream":true}' \
  "$LAN_BASE/responses"
```

The JSON request returned HTTP 200, `status: completed`, and `OK`. The streaming request returned HTTP 200 as `text/event-stream`, streamed `OK`, and emitted `response.completed`.

On commit `735d1e0` based on `13eedf5`, Windows 11 and Node 24.12.0, all 20 targeted tests passed from `apps/host`.

```sh
node --test tests/bad-host-header.test.mjs tests/native-gateway.test.mjs tests/cors-share.test.mjs tests/serving-key-compare.test.mjs tests/connect-urls.test.mjs
npm run typecheck
npm run build
```

**Failed**

The new regression test returned 500 on the unpatched base, where it expects 400.

**Not run**

The full host suite and `local-ai-worker ready` on the worker machine. The HTTP probe and its unknown-route control were tested directly.

## Limits

LAN Responses requires the routing change in #5925. The worker export captured a readiness failure before inference. The original Windows Codex error remains unverified.

本站相关内容

相关页面的快捷入口。