The readiness advisor failed capable local models for adapter and probe
defects rather than model incapability (#1624 Ollama, #1614 llama.cpp):
- Send num_ctx sized to the haystack fixtures (clamped to the model's
trained window) so Ollama no longer truncates ~25KB prompts at its
4096-token server default; the trained-window guard alone passed while
the runtime request was being truncated.
- Forward an explicitly pinned temperature 0 instead of dropping it to
Ollama's 0.8 default against a nonce-exact validator (ChatRequest gains
TemperatureSet; Ollama options temperature is now a pointer).
- Raise the probe generation cap from 256 to 2048 tokens so qwen3-style
<think> reasoning cannot exhaust the budget before the tool call, and
surface the provider done_reason when validation fails.
- Synthesise tool-call IDs in the OpenAI-compatible adapter (streaming
finalizer and buffered path) when the server omits them, as llama.cpp
commonly does, mirroring the Ollama adapter instead of failing tool
protocol 0/3 on transport shape.
- Probe with the Patrol loop's 60s stream stall allowance instead of the
12s chat default (chat.PatrolProviderStreamIdleTimeout is now exported).
- Stop discarding probe and validator errors: log them, carry them in a
new PatrolModelReadinessResult.Details field surfaced through the API
snapshot and Settings UI, and keep a transport-level probe failure's
specific diagnosis instead of overwriting it with the generic
capability wording.
Builds on 4a2335ce7, which already reclassifies protocol failure as
"provider connected; Patrol capability not verified".
Fixes#1624Fixes#1614
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
courtmanr@gmail.com
eb665b2a0d