v0.12.13 — Reliability & hygiene

0.12.13 (2026-06-27)

A reliability-focused release: a new capability-first tools gate so models that can't do function calling degrade gracefully instead of failing the turn, a major Windows sandbox fix, and a round of audited provider- and config-layer hardening.

Highlights

• Tool-incapable models just work now — across local and cloud backends. Point Wayland Core at a model that doesn't support function calling and the turn no longer dies on a raw provider error. Ollama models are detected up front via /api/show and have their tools array dropped before the request is even sent. Any backend that rejects tools with a 400 — llama.cpp started without --jinja (tools param requires --jinja flag), or an Ollama model that 400s with does not support tools — is caught, retried without tools, and remembered, so every later turn for that model skips tools pre-emptively. Tool-incapable Bedrock models (DeepSeek-R1 reasoning, Stability image, Titan/Cohere embedding) are name-gated the same way. Tool-capable models are unaffected — they keep their tools and call them exactly as before. (#389)
• The Windows sandbox runs real subprocesses again. The AppContainer backend no longer caps active processes too aggressively (ActiveProcessLimit raised to 512), resolves the launch shell correctly, and emits clearer diagnostics when a shell can't be found — so multi-step tool use works under the sandbox on Windows. (#321, #322, #323, #324)

Provider reliability

• Anthropic errors are classified correctly. Non-credit Anthropic API errors are no longer misread as out-of-credit / billing failures, so genuine transient errors surface instead of a misleading "purchase credits" signal. (#329)
• Flux reasoning summaries render as thinking. A FluxRouter reasoning_summary is now decoded into a per-turn thinking subject, so reasoning summaries appear as proper thinking content. (#318)

Configuration & hygiene

• Config surface tightened. env_passthrough is now wired through, unknown configuration keys produce a warning (via serde_ignored) instead of being silently dropped, and the sandbox configuration surface is exposed as a toggle. (#325, #326, #327)