v2026.4.2

Added

• PydanticAI agent import -- initrunner new --pydantic-ai <file> converts PydanticAI scripts to role.yaml via CLI or dashboard. AST extraction of Agent() model config, system_prompt/instructions, @agent.tool/@agent.tool_plain/FunctionToolset tools, output_type (including NativeOutput/ToolOutput/PromptedOutput unwrapping), ModelSettings, UsageLimits. Sidecar module with RunContext parameter stripping and body-reference TODO insertion. Warns on pydantic_graph, logfire, MCP servers, builtin_tools, @agent.output_validator
• Dashboard Import mode redesign -- single "Import" card with pill toggle for LangChain/PydanticAI framework selection. Separate source buffers per framework
• Dashboard structured output fallback -- SSE streaming auto-falls back to non-streaming execution for output.type: json_schema roles instead of crashing
• Shared sidecar validation -- _sidecar_common.py extracts blocked-module validation used by both LangChain and PydanticAI importers
• PydanticAI import documentation -- docs/getting-started/pydanticai-import.md with real before/after example

Fixed

• Provider auto-detection crashes on missing SDK -- detect_provider_and_model() and list_available_providers() now skip providers whose Python SDK is not installed, falling through to the next available provider
• Async tool RunContext stripping -- regex patterns now handle async def functions
• Dashboard starter override -- manual mode selection clears pendingStarter so ?starter= URL params don't override user choice