Upgrade pydantic-ai to 2.x

[benh]

Summary

Reboot pins pydantic-ai-slim==1.87.0 (reboot/requirements.in), which pins pydantic-graph==1.87.0. pydantic-ai 2.0 is stable and upstream has moved on (latest is 2.43.0). We should upgrade, both to keep up and to pick up fixes we're currently working around.

Why now

pydantic-graph 1.87.0 lets a cancelled agent run surface as anyio.ClosedResourceError instead of asyncio.CancelledError: a tracked task sends its result into a stream the cancelled run has already closed, and only BrokenResourceError is ignored there. A workflow that catches everything except cancellation then keeps running, which made //tests/reboot/examples/agent-wiki:test hang at shutdown. #179 works around it in _agent_run. The fix upstream first ships in pydantic-graph 2.2.0 (no 1.x release has it); once we're on it, the workaround in #179 can be removed.

Picking a target

pydantic-ai / pydantic-graph	ClosedResourceError fix	HTTP client	Anthropic SDK (anthropic extra)
1.87.0 (current) – 1.99.0, 2.0.0 – 2.1.0	no	httpx	0.x
2.2.0 – 2.30.x	yes	httpx	0.x
2.35.0 – 2.43.0	yes	httpx2	>= 1.3

• 2.2.0 – 2.30.x keeps reboot[anthropic] on anthropic==0.125.0 (per requirements.in / requirements-anthropic.in, Anthropic SDK 1.x rejects the httpx client pydantic-ai hands it before the httpx2 switch).
• 2.35+ also means moving the anthropic extra to SDK 1.x and httpx2.

A reasonable path is to land 2.30.x first, then take the httpx2 / Anthropic SDK 1.x step separately.

What it takes

1. Port Reboot's wrapper (reboot/agents/pydantic_ai/, ~2,300 lines) — most of the work.
   • Of the 50 names it imports from pydantic-ai, against both 2.2.0 and 2.43.0: pydantic_ai.tools.AgentBuiltinTool no longer exists and pydantic_ai.mcp.MCPServer no longer resolves; CompletedStreamedResponse moved from pydantic_ai.models.wrapper to pydantic_ai.models (deprecated in 2.43). pydantic_ai.toolsets.fastmcp wasn't checked with the fastmcp extra installed.
   • _agent.py mirrors the full signatures of run / iter / run_stream / run_stream_events, and wraps toolsets (_toolset.py) and models (_model.py). V2 is a harness-first redesign around capabilities (see the upgrade guide), so signatures, toolset wrapping and model wrapping all need checking against 2.x, not just imports.
2. Pins and lock: pydantic-ai-slim (and possibly anthropic) in reboot/requirements.in and reboot/requirements-anthropic.in; regenerate reboot/requirements_lock.txt.
3. Remove the Agents: raise a cancelled run's ClosedResourceError as a cancellation #179 workaround in reboot/agents/pydantic_ai/_run.py (_raised_while_cancelling and the ClosedResourceError conversion) and its CancellationTestCase tests, or keep the tests asserting a cancelled run raises CancelledError.
4. Docs and examples: the Python skill references (reboot/plugin/skills/python/references/agent-pydantic-ai.md, agent-tools.md, testing-external-context.md), and examples that use agents (e.g. reboot/examples/agent-wiki, whose uv.lock resolves pydantic-ai-slim).
5. Dashboard: the static analysis in reboot/dashboard/backend/code_watcher.py recognizes Agent's entry points and tool registration by name (RUN_NAMES, TOOL_REGISTRATION_NAMES, constructor and run arguments); check those still match 2.x, and its test fixture of the agents module in tests/reboot/dashboard/code_watcher_tests.py.

Testing

• //tests/reboot/agents/pydantic_ai/...
• //tests/reboot/examples/agent-wiki:test and other agent examples
• //tests/reboot/dashboard:code_watcher_tests_py