v2026.6.7

Changed

• Upgraded to PydanticAI v2 (pydantic-ai>=2.0.0). Every pydantic-ai-slim and pydantic-evals pin moved from >=1.107.0 to >=2.0.0. The base install now uses the mcp extra in place of the removed fastmcp back-compat extra (which previously supplied pydantic_ai.mcp.MCPToolset); InitRunner's own FastMCP server dependency (fastmcp>=3.2.0) is unrelated and unchanged. The a2a extra now pulls the upstream fasta2a package directly, since pydantic-ai-slim[a2a] was removed in v2.
• end_strategy now defaults to graceful (previously early), matching PydanticAI v2. A role with no explicit spec.execution.end_strategy will now also run the function-tool calls the model requested alongside a successful output, instead of stopping at the first final output. Set end_strategy: early to restore the old behavior. The new graceful value is also accepted by the --agent-spec importer and emitted by the exporter.
• The WebSearch capability is native-only. Under v2 it uses the provider's native web search and raises on models that lack one, rather than silently adapting. WebFetch still receives InitRunner's injected SSRF-protected local fallback.
• The BuiltinTool capability is renamed to NativeTool. PydanticAI v2 renamed the underlying class, so a role using - BuiltinTool under spec.capabilities must switch to - NativeTool.
• Custom tool builders that register a context-free function must use @toolset.tool_plain. PydanticAI v2 requires the @toolset.tool decorator's function to take a RunContext as its first parameter; registering a context-free function with @toolset.tool now raises at schema-generation time. InitRunner's built-in tools already use tool_plain; the tool-creation guide and examples are updated to match.
• Agent instrumentation is wired as a pydantic_ai.capabilities.Instrumentation capability instead of the removed Agent(instrument=...) constructor argument. The spec.observability YAML block and its behavior are unchanged.

Fixed

• Streaming runs use the v2 run_stream_events() context-manager form. v2 makes Agent.run_stream_events() an async context manager only; the executor now drives it with async with ... as events, restoring event streaming (token callbacks and the audit timeline) under v2.
• Tool-mode structured-output calls are recorded in the run-event timeline again. v2 routes output-tool calls and results through dedicated OutputToolCallEvent / OutputToolResultEvent rather than the function-tool events; the audit and eval timeline now captures these as output_tool_call / output_tool_result entries.
• Streaming a text-output role with an approval-gated tool now pauses instead of crashing. When such a role's model calls an approval: required tool, the run output becomes a DeferredToolRequests; v2's stream_text() raises on non-text output, which previously escaped uncaught from the streaming run (CLI --stream and the OpenAI-compatible streaming API). The streaming executor now catches that and routes to the human-in-the-loop pause (status: paused), matching the non-streaming path.

Dependencies

• Bumped pydantic-settings from 2.12.0 to 2.14.2 (#165).
• Bumped posthog-js from 1.386.6 to 1.393.4 in the dashboard frontend (#166).
• Bumped the actions/checkout CI action from 6 to 7 (#161).