v0.9.0 — Responses API adapter, ordered failover chains, terminal-on-success tools

Highlights

• Responses API adapter — models reachable only through OpenAI's /v1/responses endpoint (built-in: openai/gpt-5.6*, openai/gpt-5-pro*) are translated at the provider boundary. Agents keep the /v1/chat/completions contract; audit, session history, budgets, tool mediation, and failover all observe the unchanged shape. Extend with CLLAMA_RESPONSES_API_MODELS, disable with CLLAMA_RESPONSES_API_DISABLED. Automatic one-shot retry through the adapter when an upstream rejection names /v1/responses. (#31, #36)
• Ordered failover chains — declared-model failover now traverses every fallback-slot entry of the compiled model policy in declared order, and advances to declared fallbacks on managed response-body read errors, preserving key cooldowns across the traversal. (#28)
• Terminal-on-success managed tools — tools annotated "x-claw.terminalOnSuccess": true end the mediated turn after a successful call with a protocol-valid empty assistant response. Order-invalid calls are rejected before any tool executes. (#30)
• Reasoning replay preservation — Responses-adapter reasoning items are preserved across managed tool rounds; mismatched tool-call IDs in replay are rejected. (#33)
• Current model pricing — adds Claude Fable 5 / Opus 5 / Sonnet 5, Gemini 3.6 Flash (direct + OpenRouter routes), corrects Claude Haiku 4.5 to its current rate, and prices the Responses-adapter models. Sonnet 5 uses the standard 3/15 ceiling so budget enforcement does not undercount when promotional pricing ends 2026-08-31. (#36, #40)

Versioning note

This release follows v0.7.8. The version jumps to v0.9.0 because a stale v0.8.0 tag/release from June (a duplicate of v0.7.5) already occupies that number; skipping it avoids ambiguity. The v0.7.8 tag was never published as a release or image.

Image: ghcr.io/mostlydev/cllama:v0.9.0 (linux/amd64, linux/arm64)