v0.6.3

Highlights

• Full custom_tool_call support for OpenAI Responses API — request, response, and streaming. Custom tools (type: "custom") with plain-text input now pass through the gateway correctly, and cross-provider degradation synthesizes a JSON Schema fallback for Anthropic/Google
• Security fixes for the admin panel (stored XSS, login rate limiting)
• Argo provider shims with model field rewriting and thinking normalization

What's New

Added

• Full custom_tool_call support: request coercion to IR with _passthrough, response parsing, streaming (response.custom_tool_call_input.delta/done), and cross-provider degradation with synthesized parameters
• tool_type field on IR ToolCallStartEvent for correct event type dispatch
• Argo shims (argo_openai, argo_anthropic, argo_google) with model_id_field and upstream_model
• Async server-side admin test tasks
• Admin login rate limiting

Fixed

• Stored XSS in admin UI (esc() single quote escaping)
• custom_tool_call streaming type loss in gateway (from_base() missing _tool_call_types)
• Admin UI regressions (fetch models recursion, API key editing, prefix preview lag, modal abort)
• Reasoning test budget_tokens floor
• httpclient AsyncClient serialization deadlock (vendored v0.4.1)
• ty 0.0.32+ type-check compatibility

Changed

• Admin routes refactored into routes/ subpackage
• CI switched to pre-commit (ruff + ty)