fix(agent): preserve MiMo thinking blocks on Anthropic replay

[CallMe1101]

Problem

Xiaomi MiMo's Anthropic-compatible endpoint recently tightened thinking-mode validation. When thinking mode is enabled, prior assistant turns containing tool_use must replay their full reasoning_content on subsequent requests. If omitted, MiMo returns HTTP 400:

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

Generic third-party Anthropic handling strips all thinking blocks, breaking MiMo replays — same root cause as Kimi (#13848) and DeepSeek (#16748).

Solution

MiMo follows the same replay contract as Kimi and DeepSeek: strip Anthropic-signed thinking blocks (third-party endpoints cannot validate signatures) but preserve unsigned thinking blocks synthesized from reasoning_content.

Changes to agent/anthropic_adapter.py

• Add _XIAOMI_MIMO_MODEL_PREFIXES — model name prefixes (mimo-*, mimo_*, xiaomi-mimo-*, xiaomi_mimo_*)
• Add _model_name_is_xiaomi_mimo() — detects MiMo model names (strips vendor prefix, handles deep-namespaced forms)
• Add _is_xiaomi_mimo_anthropic_endpoint() — matches by model name or base_url host (xiaomimimo.com, mimo.com)
• Update _preserve_unsigned_thinking to include MiMo alongside Kimi/DeepSeek
• Inject single-space placeholder thinking block when an assistant tool-use turn has no thinking block at all — handles migrated/cross-provider history (reproduces akaDRJ's live failure from fix: pad Xiaomi MiMo reasoning_content replay #24726)

Tests

• tests/agent/test_xiaomi_mimo_anthropic_thinking.py (277 lines): unsigned thinking preservation, signed block stripping, generic third-party behavior, placeholder injection, lookalike model names, deep-namespaced models
• tests/run_agent/test_xiaomi_mimo_reasoning_replay.py (138 lines): reasoning padding for MiMo providers, missing/empty reasoning_content handling, lookalike detection

Notes

• run_agent.py already has MiMo reasoning support (merged via fix(agent): add Xiaomi MiMo to reasoning_content echo-back providers (#25358) #25484). This PR completes the picture by adding the corresponding anthropic_adapter.py changes.
• PR feat: support Xiaomi MiMo thinking-mode reasoning_content replay #24465 covered both files but has merge conflicts. This PR is rebased on latest main and only touches anthropic_adapter.py + tests.
• PR fix: preserve thinking blocks for Xiaomi MiMo models via Anthropic relay #26775 was a previous attempt closed as duplicate.

Testing

All 38 tests pass:

python -m pytest tests/agent/test_xiaomi_mimo_anthropic_thinking.py tests/run_agent/test_xiaomi_mimo_reasoning_replay.py -v
# 38 passed

Related

• Closes fix: pad Xiaomi MiMo reasoning_content replay #24726 (akaDRJ live repro)
• Refs feat: support Xiaomi MiMo thinking-mode reasoning_content replay #24465 (original PR, conflicted)
• Refs fix(agent): add Xiaomi MiMo to reasoning_content echo-back providers (#25358) #25484 (run_agent.py MiMo fix, merged)
• Refs [Bug]: 400 error for kimi-for-coding #13848 (Kimi thinking replay)
• Refs DeepSeek /anthropic (V4 thinking): stripped thinking blocks cause HTTP 400 on replay #16748 (DeepSeek thinking replay)
• Refs fix: preserve thinking blocks for Xiaomi MiMo models via Anthropic relay #26775 (previous attempt, closed as duplicate)

[alt-glitch]

Duplicate of #24465 (earliest open PR for Xiaomi MiMo reasoning_content echo-back, 26 tests). This is the 12th+ competing PR for the same feature. See also #24784 (which adds gateway retry prevention as net-new scope).

Related: #24603, #24887, #25379, #24893 (closed), #26775 (closed, same author).

[teknium1]

Thanks for tracing the replay failure and covering signed versus unsigned thinking blocks. The current main path still lacks Xiaomi-specific unsigned-thinking preservation at agent/anthropic_adapter.py:2276-2279, so the underlying issue remains relevant.

Problems

• agent/anthropic_adapter.py:509-513 makes a MiMo model slug sufficient to select Xiaomi Anthropic replay behavior. The added tests deliberately accept https://llm.example.com/anthropic for xiaomi-mimo-* and xiaomi/mimo-* models (tests/agent/test_xiaomi_mimo_anthropic_thinking.py:31-32). A model name does not establish that an arbitrary proxy accepts Xiaomi's unsigned-thinking contract.

Suggested changes

• Scope detection to Xiaomi's xiaomimimo.com host and an /anthropic route; add negative cases for OpenAI-compatible Xiaomi URLs and arbitrary proxy/aggregator slugs. Commit 06d94943d671ad85119761d8657a7daa07eaea0a contains that narrower salvage direction, but is not on current main.

Automated hermes-sweeper review.

[teknium1]

A MiMo-named model alone does not prove that an arbitrary Anthropic-compatible proxy implements Xiaomi's replay contract. Please gate this on Xiaomi's host plus the /anthropic route; the current tests at tests/agent/test_xiaomi_mimo_anthropic_thinking.py:31-32 lock in the overbroad proxy behavior.