feat(mcp): move to mcp 2.x and pydantic-ai 2.x

[cardosofede]

Summary

Moves both MCP servers and the pydantic-ai agent client to the 2.x SDKs (mcp>=2.2,<3, pydantic-ai>=2.52,<3), and fixes a turn-teardown bug in the pydantic-ai client found along the way.

MCP / pydantic-ai 2.x (FEAT-130)

• 0b74ed96 freezes the v1 tool surface of both servers as a fixture (tests/fixtures/mcp_tool_surface_v1.json).
• 8e6489f7 moves to mcp 2.x and pydantic-ai 2.x.
• 85cc2e3d pins that the v2 tool surface matches the frozen v1 one, so the upgrade changes no tool name or schema the model sees.
• 11bbfdd4 keeps a failing tool telling the model why it failed.
• 7b957fa8 ports the pydantic-ai client to the 2.x MCP toolset.

Abandoned pydantic-ai turn (CORR-714)

• 829b56e8: prompt_stream yielded from inside the request slot, agent.iter() and the usage fold. A reader that stopped mid-turn (a WS drop, a page reload, a break) left that stack to be unwound by closing the generator, which pydantic-ai's run cannot survive: aclose() raised, the per-backend request slot stayed held until the generator was garbage collected, and the unwind raised again in the finalizer.
• The run now lives in a task the client owns and prompt_stream relays its events through a queue, in step with the reader, so abort_prompt still stops the run where the reader is. When the reader closes, drops or is cancelled out of the stream, the task is cancelled and awaited, the usage is folded and the slot is released there.
• PromptDone no longer waits for the reader to come back, so the slot is free as soon as the turn is over.

Known gap

The client only learns the reader left when the stream is closed, dropped, or cancelled while waiting for the next event. A reader cancelled inside its own loop body whose frame stays referenced (a kept task or traceback) still holds the slot until that frame is released. The consumers in sessions.py, agent_run.py and engine.py are unchanged; closing this needs aclosing through the whole generator chain, ACP included.

Test plan

• uv run pytest -W error::pytest.PytestUnraisableExceptionWarning — 6446 passed, 2 skipped (run at 829b56e8, before merging the 10 commits main has gained since)
• black and isort clean on the touched files
• tests/test_mcp_tool_schema_parity.py: v2 tool surface equals the frozen v1 one
• tests/test_agents.py: closing, dropping or cancelling the stream mid-turn returns the slot and leaves no turn task
• Live chat and a loop tick on a local backend (LM Studio / Ollama) against the 2.x SDKs

[greptile-apps]

Confidence Score: 4/5

[High risk] Upgrades MCP and pydantic-ai to major versions 2.x.

The PR appears safe to merge; the remaining feedback is a non-blocking regression-test gap.

Findings

1.  Null-content behavior lacks coverage ▶

Summary

The PR migrates both MCP servers and the pydantic-ai client to their 2.x SDKs and moves streamed turns into client-owned tasks for orderly cancellation.

• It freezes and checks the advertised MCP tool surface, preserves tool-error messages, and updates provider and lifecycle tests.
• The null-content compatibility override would benefit from a behavioral regression test.

Diagram

sequenceDiagram
  participant Reader
  participant Stream as prompt_stream
  participant Turn as Owned turn task
  participant Agent as pydantic-ai run
  Reader->>Stream: Request next event
  Stream->>Turn: Start turn
  Turn->>Agent: Iterate run
  Turn-->>Stream: Queue event
  Stream-->>Reader: Yield event
  Reader->>Stream: Resume or close
  alt Resume
    Stream-->>Turn: Acknowledge event
  else Close
    Stream->>Turn: Cancel and await
    Turn->>Agent: Unwind run and release slot
  end

Loading

_{Reviews (1) · Last reviewed commit: "fix(agents): a pydantic-ai turn ends whe..."}

[greptile-apps]

Null-content behavior lacks coverage The 2.x migration changes the base class of the workaround for Ollama and LM Studio, but the new test checks only the model’s type. Please test that a tool-call-only assistant message is mapped with content: ""; otherwise a change to the SDK’s private mapping method could go unnoticed and restore the content: null requests these backends reject.