Validate Cinference tool loops in Hermes, OMP and pi.dev

[janitor-manager]

Fast Qwen vanilla and Huihui passed a direct non-strict tool call and result continuation, but that does not prove full agent harness compatibility. Track the rollout gate for ordinary coding and voice tool loops.

Audit actual Hermes (default/Mira, including LiveKit), OMP and pi.dev requests on the installed machines. Use the existing provider compatibility knobs (OMP/pi supportsStrictMode=false), preserve schema validation in the tool executor, and avoid silently dropping guarantees a workflow needs. Cover streaming argument assembly, multiple tools, tool-result replay, invalid arguments with repair, reasoning on/off, cancellation and long-context continuation for both fast model aliases.

The pinned runtime rejects strict=true, required/named tool_choice, parallel_tool_calls=false with callable tools, and constrained JSON. Record which clients send each option and provide minimal redacted reproductions. Use automatic tool choice where semantically acceptable. A workflow requiring enforcement must remain unsupported until the separate engine work is done.

Acceptance: per-harness results, exact versions/settings and regression tests or bounded probes, updated PRIMARY_MODEL_SWITCH.md, and a clear remaining-limit matrix. Respect the 6000 borrowing guide; no installation/default change is authorized by this tracking issue.

Related: #28 and #29. Engine source: satellitedown/cinference b74044fb0a319cd2a737cb7108012e6344b96dac, docs/serving.md.

API prerequisite: https://github.com/kortexa-ai/api.server/issues/95. Separate engine work: #32 (schemas) and #33 (tool-choice enforcement).

[janitor-manager]

The prerequisite API adapter is delivered in api.server 654ac12 and deployed on Smarty under https://github.com/kortexa-ai/api.server/issues/95. It advertises Cinference capabilities and the Messages/Responses generation surface, retains native token-count routes and error semantics, and explicitly reports unsupported guarantees. All API validation/build checks passed on both machines (608 tests passed, 17 skipped), with CPU HTTP proxy fixtures for both model aliases.

Documentation delivered in models.server 6617034, clean and synchronized on Snappy/Smarty. PRIMARY_MODEL_SWITCH.md now points to the API contract and the separate engine work: #32 for strict schemas/constrained JSON; #33 for required/named/single-call selection. No harness settings/defaults changed, no fast-model service was installed, and the GPU handoff was not disturbed.

This issue stays open for real agent/harness qualification and native fast-model gateway inference during a later authorized GPU window. CPU fixtures are proxy evidence only; they do not complete that acceptance.