[ScottMansfield]

Adds extended thinking support to AnthropicLlm and Claude in @google/adk-integrations. It also skips request parts that the Anthropic Messages API rejects. Before this change, a session that ran a Gemini turn and then switched to Claude failed on every later turn, because partToMessageBlock threw on thought parts.

Files changed

5 files changed, 838 insertions(+), 21 deletions(-):

• integrations/src/models/anthropic_utils.ts (+158 / -9)
• integrations/src/models/anthropic_llm.ts (+6 / -3)
• integrations/test/models/anthropic_utils_test.ts (+433 / -7)
• integrations/test/models/anthropic_llm_test.ts (+185 / -0)
• docs/guides/integrations/models/anthropic_llm/index.md (+56 / -2)

The deletions are the replaced throw on thought parts, the rewritten test that expected that throw, the unsupported-block test fixture (it used thinking, which is now supported), the request assembly lines in AnthropicLlm, and the outdated Limitations bullets in the guide.

Ported adk-python behavior

• thinkingConfig.thinkingBudget maps to the Anthropic thinking parameter: 0 gives disabled, a negative value gives {type: 'adaptive', display: 'summarized'}, a positive value gives {type: 'enabled', budget_tokens}, and an unset budget throws.
• thinking and redacted_thinking response blocks become thought: true parts, and these parts convert back to the same blocks on later turns.
• A part that carries only a thoughtSignature from another model is dropped with a warning. A turn left with no blocks is not sent.
• Empty text parts are skipped. A user turn left empty only by skipped empty text sends one '.' text block.
• Empty and whitespace-only response text blocks are skipped.
• usage.output_tokens_details.thinking_tokens is reported as thoughtsTokenCount and subtracted from candidatesTokenCount. totalTokenCount stays input plus output.

Interface differences

Member	adk-python	adk-js	Why
thought_signature / thoughtSignature	bytes (UTF-8 encoded)	string	@google/genai types Part.thoughtSignature as string, so the code stores Anthropic signature and data as-is.
_is_empty_text_part / isEmptyTextPart	Checks model_fields_set	Checks text === '', no signature, and no content payload field	Part is a plain interface with no record of which fields are set.
_build_anthropic_thinking_param / buildAnthropicThinkingParam	Returns a dict or None	Returns an @anthropic-ai/sdk ThinkingConfig* union or undefined	Native SDK types.

Verification

• npx vitest run --project unit:integrations: 3 files and 172 tests passed (104 in anthropic_utils_test.ts, 66 in anthropic_llm_test.ts, 2 in version_test.ts).
• npm run build, ts:check, ts:check:samples, lint, format:check, docs:check, and npx secretlint "**/*": all exited 0.
• npm run test:coverage on Linux: 6558 tests passed. Three files failed only because of the local machine: the Playwright browser is not installed, and the local npm wrapper writes notices to stderr.
• Mutation check: changing display: 'summarized' to 'omitted' made the 2 adaptive thinking tests fail.

What the reviewer must decide

• Is a single '.' text block the right way to keep a user turn that held only empty text parts? Without it, the conversation ends on an assistant turn, which Anthropic treats as prefill.
• Is it right to store Anthropic thinking.signature and redacted_thinking.data verbatim in Part.thoughtSignature?
• A thinking block with empty text but a signature comes back as redacted_thinking on the next turn. adk-python has the same gap. Should a follow-up change fix it?

Produced by Foundry (task 4d0ea73c-114a-4942-8df3-1f045a4f6a06). Reviewed by an automated
reviewer before upload; still needs human review.

[ScottMansfield]

Scored 2026-10-28 16:56 UTC at 8d65696, round 1. Earlier runs are in this comment's edit history.

Foundry merge critic: APPROVE

Add extended thinking support to AnthropicLlm and Claude and skip empty text parts and foreign thought signatures that the Anthropic API rejects.

Rubric — the default is approve; each rule is a reason a human should read this change first.

[no ] public_api — leaves package exports in integrations/src/index.ts and AnthropicLlm and Claude signatures unchanged; new helpers and union members in anthropic_utils.ts are additive
[no ] weakened_guarantee — adds thinkingBudget validation and converts thinking parts while preserving unsupported-part throws on bare thought and unknown parts
[no ] sensitive_surface — touches only message, thinking config, and token usage conversion in integrations/src/models/ and docs
[no ] release_surface — does not modify package.json, lockfiles, build scripts, or CI workflows
[no ] test_removed — adds 21 unit tests and updates two unsupported-kind test fixtures in anthropic_utils_test.ts so they still assert the unsupported error branches
[no ] inexplicable — all five modified files directly implement, test, and document extended thinking and part filtering for Claude

Posted by Foundry's merge critic as evidence rather than a request, so nothing here is waiting on you. Advisory only: Foundry has given no approval.