kingy tested
What happened
Symptom
The server returns HTTP 500: tools param requires --jinja flag.
Conditions
Qualified October 15, 2026 on llama.cpp 0.6.0 build 11429 (d81235049384534c167caea52b85a694f6103d14), Qwen3-4B Q4_K_M.
macOS 26.3 build 25D125, Apple M4 Pro / Mac16,7, 24 GiB unified memory; CPU-only inference.
Separate-agent rerun from a fresh extraction and fresh server processes on the same physical Mac.
Correction
In the tested llama-server command, replace the explicit --no-jinja flag with --jinja. Restart that local server with the same model and settings, then validate the complete tool-call task.
Result
For both the author and separate-agent qualifying runs: 0/3 baseline passes, 3/3 corrected full-task passes and 0/1 reverted passes. All qualifying corrected responses stopped normally and reported zero cached tokens.
What did not work
The original corrected fixture executed the right tool but returned Stock for SKU: 37, failing the strict final-answer check. That failed qualification remains retained.
- Save the complete original command and error before changing it. This repair applies to explicitly disabled Jinja; Jinja is already enabled by default in this build.
- Replace --no-jinja with --jinja. Keep the model file, prompt, tool definition and remaining settings unchanged.
- Check the full flow: the model calls lookup_stock with sku KINGY-017, the synthetic client returns stock 37, and the final answer is exactly Stock for KINGY-017: 37.
- Read README.md in the evidence ZIP for the frozen runner. Its documented independent-audit command inspects retained evidence without loading a model.
Sources & limits
Repair 001: full qualification report and limits
Record checked 2026-10-15.
Same-host separate-agent reproduction, not independent human testing or a second-machine result.
These fixed synthetic tasks and repeated settings do not establish a general reliability rate, other-model behavior or minimum hardware requirements.
Original failures, failed correction proposals, prompts, outputs, settings and checksums remain in the evidence package. Model weights and runtime executables are not included.
Qualification ran October 15, 2026. Publication does not make this a test of later software or model revisions.
This restores one synthetic tool-backed task with no external inventory action; it is not a general solution to tool-calling failures.