SHIPIT

Interleaved thinking & context editing

Let Claude think between tool calls with interleaved_thinking + thinking_budget_tokens, and have Anthropic clear stale tool results server-side with context_management. Thinking blocks and budget surface in LLMResponse.metadata. New in v1.0.12.

2 min read
5 sections
Edit this page

New in v1.0.12, AnthropicChatLLM exposes two more Claude-API power features: interleaved thinking (the model reasons between tool calls, not just before its first answer) and server-side context editing (Anthropic clears stale tool results for you, on its side of the wire).

Interleaved thinking

Extended thinking lets Claude produce thinking blocks before answering. Interleaved thinking extends that across a tool loop — the model can think, call a tool, see the result, think again, and so on. It's enabled by pairing two kwargs on AnthropicChatLLM:

python
from shipit_agent.llms import AnthropicChatLLM

llm = AnthropicChatLLM(
    model="claude-sonnet-4-20250514",
    thinking_budget_tokens=2048,     # turns on extended thinking
    interleaved_thinking=True,       # allow thinking between tool calls
)

The beta header interleaved-thinking-2025-05-14 is attached only when interleaved_thinking=True and thinking_budget_tokens is set — interleaved thinking requires extended thinking to be on. With either flag off, the request stays on the GA endpoint, unchanged.

Honest scope. The adapter side of the round-trip is wired: response thinking blocks (with their signatures) are surfaced in LLMResponse.metadata["thinking_blocks"] and re-emitted first on assistant messages whose metadata carries them. Completing the round-trip across a multi-turn tool loop also needs the runtime to copy that metadata onto the next assistant message; until that propagation lands, treat interleaved thinking as a degrade-gracefully passthrough — it never breaks a run, and single-response thinking is fully captured.

What surfaces in metadata

Metadata keyWhenContents
metadata["thinking_budget_tokens"]thinking_budget_tokens setThe configured budget.
metadata["thinking_blocks"]interleaved_thinking=True and blocks presentThe full thinking / redacted_thinking blocks, including signatures, for round-tripping.
response.reasoning_contentany thinking presentThe concatenated thinking text, which the runtime emits as reasoning_started / reasoning_completed events.

Context editing

Long tool loops bloat the context window with old tool results. context_management hands that cleanup to Anthropic: pass a context-management config and the API clears stale tool results server-side, so you don't have to truncate the transcript yourself.

python
from shipit_agent.llms import AnthropicChatLLM

llm = AnthropicChatLLM(
    model="claude-sonnet-4-20250514",
    context_management={"edits": [{"type": "clear_tool_uses_20250919"}]},
)

When set, the config is forwarded as the context_management request param and the context-management-2025-06-27 beta header is attached automatically; the call is routed to the beta endpoint. With context_management=None (the default) nothing changes and the request is identical to legacy. This composes with shipit's own client-side context management — use either, or both.

Provider note

Extended and interleaved thinking, and server-side context editing, are Anthropic features (also reachable for Anthropic models via Bedrock / LiteLLM). The equivalents on other providers are their reasoning modes: OpenAI's reasoning models (o-series / GPT-5-class, via reasoning_effort) and Gemini's thinking. shipit already captures reasoning content from all of them into response.reasoning_content (see Reasoning & thinking), but the interleaved_thinking / context_management kwargs are Anthropic-specific.

See also