kingy tested
What happened
Symptom
A schema-constrained chat request returns HTTP 400: Failed to initialize samplers: std::exception after appending an assistant message containing <think>\n\n</think>\n\n.
Conditions
Qualified 2026-10-16 Pacific on llama.cpp 0.6.0 build 11429 (d81235049), Qwen3-4B Q4_K_M; exact artifact SHA-256 in the package.
Apple M4 Pro, 24 GiB unified memory, macOS 26.3; CPU only, two threads, one slot, 2048-token context; loopback/offline.
For each repair, the same agent completed two seven-trial sequences in fresh server processes on the same Mac. Unchanged passing case runs were reused; the evidence index identifies them.
Correction
Remove the manually appended empty-thinking assistant message. Keep the original user message, schema, output budget and model template unchanged; let the template create the generation prompt.
Result
In each of two fresh-process runs: 0/3 baseline task passes, 3/3 corrected passes and 0/1 reverted passes. All original task checks and scoped completion checks passed. The offline audit recomputed acceptance from raw responses and rejected six damaged corrected-response controls.
What did not work
The historical assistant-prefill workaround failed with HTTP 400 in the exploratory run; the original no-prefill request and a supported thinking-off control worked.
- Compare the frozen request and prefill in fixtures/005-prefill.json before changing your client.
- Remove only the final manually added assistant prefill. Preserve real conversation history and the original JSON schema.
- Repeat the same extraction request. Require the exact array containing id 17 and text ALPHA, plus finish_reason=stop.
- Retain the raw HTTP error and corrected response. The package includes a restored-prefill negative control.
Sources & limits
Repair 005: full current qualification report and limits
Record checked 2026-10-16.
Same-agent repeatability checks on one Mac; no independent operator, second machine or human acceptance test.
Fixed synthetic tasks selected after exploration. The repeat counts do not establish a general reliability rate or support for other models and builds.
Runtime binaries and model weights are excluded. Exploratory failures, the failed first qualification fixture and all final trial responses are retained.
Client demonstrations identify owned reference integrations and injected inputs explicitly; they are not claims of a repaired Open WebUI release or a general llama.cpp bug fix.
The 2025 upstream empty-response defect was not reproduced: the request without prefill already worked. This repair addresses the obsolete workaround failing on the pinned 2026 build.
Do not remove genuine assistant history or assume every kind of assistant prefill is invalid.