fix(anthropic): populate reasoning_content when converting thinking blocks to OpenAI format

[anisen943]

Fixes #27946.

Problem

When LiteLLM's Anthropic-Messages → OpenAI Chat-Completions adapter
(litellm/llms/anthropic/experimental_pass_through/adapters/transformation.py::translate_anthropic_messages_to_openai)
converts an assistant message that contains thinking content blocks,
it stores them in the custom thinking_blocks field but never sets the
standard reasoning_content field. Reasoning models hosted behind
OpenAI-compatible endpoints — Moonshot Kimi K2.5+, DeepSeek R1,
OpenAI o-series — reject multi-turn requests whose prior-turn assistant
messages lack reasoning_content with:

The 'reasoning_content' in the thinking mode must be passed back to the API.

(Moonshot's docs make this explicit at https://platform.kimi.ai/docs/guide/use-kimi-k2-thinking-model FAQ Q1; DeepSeek's reasoning-model guide says the same.)

Fix

When thinking_blocks is populated, also set reasoning_content from
the concatenated text of every thinking-type block (newline-joined,
preserving order). Opaque redacted_thinking blocks are skipped —
they're not text the upstream model can consume.

if len(thinking_blocks) > 0:
    assistant_message["thinking_blocks"] = thinking_blocks
    reasoning_text = "\n".join(
        block.get("thinking", "")
        for block in thinking_blocks
        if isinstance(block, dict)
        and block.get("type") == "thinking"
        and block.get("thinking")
    )
    if reasoning_text:
        assistant_message["reasoning_content"] = reasoning_text

This improves on the prior approach in #27947, which only inlined the
first thinking block — a Greptile review on that PR
specifically flagged the single-block truncation as the remaining gap.
Concatenation preserves the full chain across multi-step agent turns
where models emit several thinking blocks per turn.

Test plan

• New: test_translate_anthropic_messages_to_openai_populates_reasoning_content
  — assistant message with two real thinking blocks interleaved with a
  redacted_thinking block + a text block. Asserts: (a) reasoning_content
  equals the newline-joined real-thinking text, (b) thinking_blocks
  still contains all three blocks in original order, (c) redacted block
  is not inlined.
• New: test_translate_anthropic_messages_to_openai_skips_reasoning_content_without_thinking
  — guard against emitting an empty reasoning_content key when the
  assistant message has no thinking. Providers that reject unknown
  fields would 400.
• Existing: all 70 tests in
  tests/test_litellm/llms/anthropic/experimental_pass_through/adapters/test_anthropic_experimental_pass_through_adapters_transformation.py
  pass unchanged — make test-unit clean.

End-to-end repro that motivated the patch

Verified live with the Claude Agent SDK (Python) → bundled claude
CLI → LiteLLM proxy 1.85.0 → Moonshot Kimi K2.5. The Claude
Agent SDK path is what catches this bug in practice — it always
round-trips through the Anthropic Messages adapter even when the
upstream model is OpenAI-shape.

Turn	Before fix	After fix
T2 "list 5 accounts"	44.2 s, 3 round-trips, ends with full chart-of-accounts	44.2 s (unchanged, no prior reasoning)
T3 "which is the bank/cash account?"	47.7 s, 4 round-trips — Kimi re-fetches the chart from scratch because it has no record of T2's reasoning	13.6 s, 1 round-trip — Kimi references "the list I just showed you"

Before the fix: the LiteLLM proxy logged Missing reasoning_content for assistant message on every Moonshot call (MoonshotChatConfig.fill_reasoning_content falls back to a " " placeholder when reasoning_content is absent, which Kimi treats as empty). After the fix: zero such warnings across a 26-call session.

Scope

This PR is intentionally scoped to the single behaviour fix in
transformation.py plus its two regression tests — nothing else.
(Earlier #27947 bundled unrelated security work; per CONTRIBUTING.md
"Keep scope isolated", this version separates the concerns.)

I will sign the CLA on first PR-bot prompt.

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[greptile-apps]

Greptile Summary

This PR fixes the Anthropic→OpenAI adapter to also set reasoning_content on assistant messages that contain thinking blocks, enabling correct multi-turn round-trips to reasoning models (Moonshot Kimi, DeepSeek R1, OpenAI o-series) that 400 without it.

• Concatenates the text from all thinking-type blocks with \n; redacted_thinking blocks are intentionally skipped as they are opaque to upstream models.
• Adds two unit tests: one for the multi-block concatenation path (including a mixed thinking/redacted_thinking/thinking sequence), and one asserting that reasoning_content is absent when no thinking blocks exist.

Confidence Score: 5/5

Safe to merge — the change is isolated to a single conditional branch in the Anthropic→OpenAI adapter, adds no new code paths for non-thinking messages, and is covered by focused unit tests.

The transformation logic is correct: ChatCompletionThinkingBlock and ChatCompletionRedactedThinkingBlock are TypedDicts (which subclass dict), so isinstance(block, dict) and .get() behave as expected. The guard if reasoning_text prevents emitting an empty key. The two new tests directly exercise both the happy path and the no-thinking guard, and no existing tests were modified.

No files require special attention.

Important Files Changed

Filename	Overview
litellm/llms/anthropic/experimental_pass_through/adapters/transformation.py	Adds reasoning_content population from concatenated thinking-type blocks alongside the existing thinking_blocks field; redacted blocks are correctly excluded. Logic is sound given TypedDict inherits from dict so .get() and isinstance(block, dict) both work correctly.
tests/test_litellm/llms/anthropic/experimental_pass_through/adapters/test_anthropic_experimental_pass_through_adapters_transformation.py	Adds two targeted unit tests: one verifies multi-block reasoning_content concatenation (with a redacted block correctly omitted), the other guards against emitting reasoning_content on plain-text-only messages. No real network calls; pure in-memory transformation tests.

_{Reviews (1): Last reviewed commit: "fix(anthropic): populate reasoning_conte..." | Re-trigger Greptile}

[oss-pr-review-agent-shin]

🤖 litellm-agent: This PR is currently BLOCKED from merge.

Score: 0/5 ❌

Why blocked:

• all Phase B agent checks non-approving (phase_b_none_approved, -5 pts)

Details: Score docked for: all Phase B agent checks non-approving (karpathy + security + coverage gap).

Fix the issues above and push an update — the bot will re-review automatically.

Note: This bot is still in beta and might not always work as expected. Please share any feedback via Slack.