v0.10.0 — Responses adapter tuning knobs and model-scoped sentinel pruning

Highlights

• Default Responses reasoning effort — CLLAMA_RESPONSES_DEFAULT_REASONING_EFFORT supplies a reasoning effort for Responses-adapted requests that omit one; explicit caller values always win. Values are validated (none/minimal/low/medium/high); an invalid configuration skips injection and emits a responses_reasoning_effort_invalid intervention instead of failing every adapted request. (#44, #48)
• Opt-in required→auto tool-choice relaxation — CLLAMA_RESPONSES_REQUIRED_TOOL_CHOICE_AS_AUTO relaxes a caller's tool_choice: "required" at the Responses adapter boundary, for runtimes that apply chat-only tool policy to scheduled turns. Every applied rewrite emits a responses_required_tool_choice_relaxed intervention — the mutation is never silent. (#44, #48)
• Model-scoped sentinel pruning — CLLAMA_TOOL_ARG_PRUNE_SENTINELS takes <provider>/<model-prefix> entries selecting models whose optional tool arguments are dropped when set exactly to the schema minimum (a filler habit of specific model families). Pruning is per-model, independent of schema validation, and every prune emits a managed_tool_args_pruned intervention. (#45, #46)

Governance note

All three knobs follow the proxy's intervention contract: request mutations are opt-in, validated, and observable in the audit stream.

Image: ghcr.io/mostlydev/cllama:v0.10.0 (linux/amd64, linux/arm64)