[Atharva Vichare (atty57)]

Motivation & Context

agent-framework-anthropic drops redacted_thinking content blocks while parsing a Claude response. The normal function-invocation loop then replays the assistant turn without that block, so the follow-up request carrying the tool_result is missing reasoning that Anthropic requires to be returned unmodified. Only the pause_turn continuation looked at the dropped-content bookkeeping, so ordinary tool loops lost the block silently.

Description & Review Guide

• What are the major changes?

  • _parse_contents_from_anthropic now maps redacted_thinking to its own reasoning item: Content.protected_data carries the opaque data payload and additional_properties["anthropic"]["reasoning_block"] records the native block type. This mirrors the existing Bedrock redacted-reasoning handling and spec 005's rule that opaque provider reasoning payloads use Content.protected_data.
  • _prepare_message_for_anthropic serializes a marked item back as {"type": "redacted_thinking", "data": ...}, so [thinking, redacted_thinking, tool_use] and [redacted_thinking, tool_use] are replayed in order. A marked item without its payload raises ContentError before the request instead of being dropped silently.
  • Each parsed reasoning block gets its own identity so response aggregation does not fold a redacted_thinking block into an adjacent thinking block when a stream is finalized, and holds its own block in non-streaming responses too.
  • The pause_turn replay guard treats a marked redacted item as replayable only when it has a payload, so continuations over redacted reasoning now proceed instead of stopping early.
  • Native reasoning items are marked so the generated identity does not route an unsigned Anthropic thinking block through the provider-summary path that serializes reasoning without a payload as text.

• What is the impact of these changes?

  • Redacted reasoning round-trips unchanged in streaming and non-streaming runs, through provider request replay, and after persistence (raw_representation is not required).
  • pause_turn continuations that previously stopped because redacted reasoning was dropped now continue with a faithful transcript.
  • No public API change and no new content type; behavior matches the existing Bedrock and spec 005 provider-metadata conventions.

• What do you want reviewers to focus on?

  The identity assignment on reasoning items (needed so streamed thinking and redacted_thinking blocks stay separate), and the choice to fail loudly rather than skip a redacted item whose payload is missing.

Related Issue

Fixes #10104

The issue is triaged, reproduced, and assigned, but I could not confirm a ready-for-implementation label or a maintainer comment agreeing with this direction yet. The change follows the direction proposed in the issue and the existing Bedrock redacted-reasoning implementation; I'm happy to adjust if reviewers prefer a different approach. No other open PR addresses this issue.

AI Assistance

• No material AI assistance was used.
• 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 this PR description were drafted with AI assistance. I reviewed and verified the diff, ran the package test suite, and confirmed the reproduction from the issue now passes.

Contribution Checklist

• The code builds clean without any errors or warnings
• All unit tests pass, and I have added new tests where possible
• The PR follows the Contribution Guidelines
• This PR links to an agreed issue with no competing open PR, or the Related Issue section documents a trivial-change or repository-automation exception.
• This is not a breaking change. If it is a breaking change, add the breaking change label (or add "[BREAKING]" to the title prefix, before or after any language prefix) — a workflow keeps the label and title prefix in sync automatically.

[Copilot]

🔵 Needs a closer look

Core-team agreement required by Spec 004 remains unconfirmed for the provider/history serialization and continuation changes, so a maintainer must confirm the intended scope and direction.

No open issues 🎉

What changed in this PR

This PR addresses #10104 by preserving Anthropic redacted_thinking blocks in Agent Framework’s Python integration, allowing tool loops and paused turns to replay assistant reasoning faithfully.

Changes:

• Stores opaque redacted payloads in Content.protected_data with provider metadata and separate reasoning identities.
• Replays redacted blocks in order, rejects missing payloads, and enables faithful pause_turn continuation.
• Adds regression coverage for streaming, non-streaming, and persisted-history replay, and updates the related specifications.

File	Description
python/​packages/​anthropic/​tests/​test_anthropic_client.py	Adds redacted-reasoning replay, persistence, validation, and continuation tests.
python/​packages/​anthropic/​agent_framework_anthropic/​_chat_client.py	Preserves and serializes redacted blocks, separates reasoning identities, and updates replay eligibility.
docs/​specs/​005-python-content-identity-and-protocol-metadata.md	Documents the Anthropic redacted-thinking marker and regression coverage.
docs/​specs/​004-python-function-calling-loop.md	Updates pause-turn behavior and test references to reflect supported redacted reasoning.

_{Review effort: Balanced}