[Kip (uragoodman)]

Motivation & Context

Anthropic's provider-hosted MCP connector returns mcp_tool_result blocks whose content is parsed into list[Content]. On the next turn, _prepare_message_for_anthropic copied content.output directly into the request body, so the Anthropic SDK failed with:

TypeError: Object of type Content is not JSON serializable

The first request succeeds; every subsequent turn that replays the MCP result fails before the request is sent.

This change normalizes MCP tool result output into the shapes Anthropic accepts for mcp_tool_result.content: a string or a list of text blocks.

Description & Review Guide

• What are the major changes?
  • Added _prepare_mcp_tool_result_content to convert None, strings, top-level mappings, and Sequence[Content] | Sequence[Mapping] into Anthropic-compatible content.
  • mcp_server_tool_result now uses this helper instead of emitting content.output unchanged.
  • Text Content items become {"type": "text", "text": ...}; non-text items are skipped with a debug log because the Anthropic mcp_tool_result schema only accepts text.
  • Top-level text mappings become a one-item text-block list; other mappings are JSON-encoded.
• What is the impact of these changes?
  • Multi-turn conversations using Anthropic's hosted MCP connector now replay their tool results instead of raising TypeError.
  • Local MCP tools and normal function tools are unchanged; they use function_result.
  • Existing string and None outputs keep their previous wire shape.
• What do you want reviewers to focus on?
  • The accepted input shapes and the decision to skip non-text items, since the Anthropic schema cannot represent image/file blocks in mcp_tool_result.
  • The top-level Mapping handling and whether JSON-encoding non-text mappings is the right fallback.
  • Test coverage for parsed list content, top-level mappings, and persisted-history replay.

The diff also contains one ruff format adjustment to an existing over-wrapped call in the same file; it is required by the repository formatter.

Testing

• pytest packages/anthropic/tests -m "not integration" -q — 259 passed
• ruff format --check — pass
• ruff check --no-fix — pass
• pyright --pythonpath .venv/Scripts/python.exe packages/anthropic/agent_framework_anthropic — 0 errors
• mypy packages/anthropic/tests/test_anthropic_client.py — no issues

Full workspace build and the full unit-test suite were not run locally; CI will run them.

Related Issue

Fixes #9970

The issue is filed; opening this focused fix for maintainer review. No ready-for-implementation label was present when this PR was opened.

AI Assistance

• This is an AI-assisted contribution. I reviewed, understood, and verified all submitted content and accept responsibility for it.

AI assistance details: implementation, tests, and PR description were drafted with an AI coding assistant; I reproduced and verified the failure and the fix locally.

Contribution Checklist

• The changed package imports cleanly and the focused Anthropic test suite passes locally.
• New regression tests were added and pass.
• Ruff format, Ruff lint, Pyright source check, and Mypy test check pass locally.
• Full workspace build and full unit-test suite were not run locally; CI will run them.
• This is not a breaking change.
• The PR links to issue Python: [Bug]: Anthropic MCP tool result replay sends Content objects instead of text blocks #9970; no ready-for-implementation label was present when opened, and the PR is submitted for early direction feedback.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has insufficient remaining quota. Try requesting a review at a lower effort level.

[Eduard van Valkenburg (eavanvalkenburg)]

Kip (@uragoodman) This replay path also needs to preserve the provider’s error status. The parser does not retain is_error in provider metadata, and this block never restores it, so a failed hosted MCP call is replayed as a successful-looking result; persistence also removes the only remaining copy in raw_representation. Please persist and restore is_error, with streaming and non-streaming round-trip coverage for error results. Also, please give _prepare_mcp_tool_result_content’s output parameter a concrete union matching the supported shapes instead of Any; this helper defines a replay wire contract and should be type-checked.

[Kip (uragoodman)]

Addressed both points.

• is_error is now persisted in the Anthropic provider metadata on parse and
  restored when replaying mcp_server_tool_result. False is preserved, a
  missing field is omitted, and the value survives Message.to_dict() /
  from_dict().
• _prepare_mcp_tool_result_content now takes a concrete McpToolResultOutput
  union (str | Mapping | Sequence[Content | Mapping] | None) instead of Any.
  The generic str(output) / default=str fallback has been removed;
  unsupported shapes now raise ContentError, so the runtime contract matches
  the annotation.
• Note: the generic fallback being removed was introduced by this PR's initial
  revision, not by existing main code, so this does not change pre-existing
  behavior.

Verification on the revised commit (4d3ad2fc1):

• Focused MCP tests: 21 passed.
• Full Anthropic suite: 269 passed, 10 deselected.
• Ruff format and Ruff check on the Anthropic package: pass.
• mypy, zuban, pyrefly, ty, and Pyright source/tests: pass locally.

Pushed as 4d3ad2fc1. The main workflows are currently waiting for maintainer
approval (action_required); CLA and API compatibility have passed.