fix(serve): a non-object json_schema answered 500 "engine failed"

[kevin9327]

Symptom

A request whose response_format.json_schema is not an object gets HTTP 500 on /v1/chat/completions and /v1/completions. Examples: a JSON-encoded string, a list, or a number. The body blames the engine. The request below is one of those payloads, and the response and log line were captured from a real coli serve on the tiny Qwen3.6 fixture:

$ curl -s localhost:8000/v1/chat/completions -d '{"model":"qwen3.6-colibri","messages":[{"role":"user","content":"x"}],
    "response_format":{"type":"json_schema","json_schema":"{\"schema\":{}}"}}'
{"error":{"message":"The colibri engine failed to process the request.","type":"server_error","param":null,"code":"engine_error"}}

server log: [api] 127.0.0.1 - request failed: 'str' object has no attribute 'get'

The request is malformed, so the answer should be a 400 that names the parameter. OpenAI SDKs retry a 5xx by default, so the malformed request is also resent before the client sees the error.

I found this by sending about 70 malformed payloads to a real coli serve (tiny Qwen3.6 fixture). This was the only one that produced a 5xx. Every other malformed field, including the neighbouring json_schema: {}, response_format: "json" and an empty GBNF grammar, already gets a 400.

Root cause

generation_options does:

schema = (response_format.get("json_schema") or {}).get("schema")

or {} covers a missing or falsy json_schema, but not a truthy one that is not a dict. The resulting AttributeError reaches do_POST's catch-all, which answers 500 engine_error.

Fix

Call .get("schema") only when json_schema is a dict. Anything else reaches the existing 400: "response_format.json_schema.schema must be an object." (param response_format). Valid schemas are forwarded exactly as before.

Tests

test_openai_server.TemplateTest.test_validates_generation_limits gains subtests for json_schema as a string, a list, a number and true. Each must raise APIError with status 400 and param response_format.

Fail-before on unfixed dev:

ERROR: test_validates_generation_limits (...) (json_schema='{"schema": {}}')
AttributeError: 'str' object has no attribute 'get'
ERROR: test_validates_generation_limits (...) (json_schema=[{...}])
AttributeError: 'list' object has no attribute 'get'
ERROR: test_validates_generation_limits (...) (json_schema=5)
AttributeError: 'int' object has no attribute 'get'
ERROR: test_validates_generation_limits (...) (json_schema=True)
AttributeError: 'bool' object has no attribute 'get'
FAILED (errors=4, skipped=1)

Verification

• tests.test_openai_server: 171 tests OK (1 skipped), in the cp949 locale and with PYTHONUTF8=1.
• Every test module that imports openai_server (18 modules; test_fp8_repack_full_family left out because it needs torch here): 375 tests OK, 26 skipped.
• End to end, Windows, coli serve on the tiny Qwen3.6 fixture (tools/make_qwen36_tiny.py + convert_qwen36.py). The same malformed-payload run now answers 400 for json_schema as a string or a list on /v1/chat/completions, and as a string on /v1/completions. Across all 62 cases of the first run: 52 × 400, 10 × 200, no 5xx, and no request failed line in the server log.

🤖 Generated with Claude Code