@Peek-A-Booo @alt-glitch — flagging three additions from #24603 (duplicate, will defer to this PR as the earliest) that may be worth folding in before merge. All are battle-tested with hermetic tests in #24603 and can be cherry-picked or rewritten standalone.

1. Third-party namespaced catalog form /mimo-

This PR's _model_name_is_xiaomi_mimo (and the run_agent.py counterpart) match mimo-* and xiaomi-mimo-* as bare/catalog prefixes after rsplit("/", 1). That covers openrouter/mimo-v2.5-pro (because the rsplit-tail starts with mimo-), but not deeper-namespaced slots like vendor/sub/mimo-v3 or nous-portal/mimo-v2 reliably across both paths (the run_agent.py detector here has only prefix matching, not rsplit). #24603 adds an explicit "/mimo-" in model check so any depth of namespacing works on both protocol paths.

2. Explicit negative guards against lookalike model families

Neither #24465 nor #24605 has tests asserting that the matcher does not trigger on MiniMax (minimax-text-01), Mistral (mistralai/mistral-medium), Microsoft Phi-4 (microsoft/phi-4), mimowave-7b, or phi-4-mimo-style. These are realistic adjacent model families that a future careless detector change could accidentally enroll into MiMo's enforcement path — which on the Anthropic side would break MiniMax replays outright (it rejects unsigned thinking blocks). #24603 pins these as explicit regression tests in both tests/run_agent/test_mimo_reasoning_content_echo.py and tests/agent/test_mimo_anthropic_thinking.py.

3. Legacy reasoning_content="" upgrade path

Sessions persisted before #17341 stored reasoning_content="" literally at creation time. On replay, MiMo (like DeepSeek V4 Pro) rejects the empty string with HTTP 400. #24603's _copy_reasoning_content_for_api upgrades "" → " " when the active provider enforces the echo and the source's literal reasoning_content is the empty string. Without this, users with pre-#17341 history hit 400 on their first MiMo turn even after this PR merges.

Offer

If useful: the negative-guard tests and /mimo- clause are small, self-contained additions; happy to open a follow-up against this branch (or against the merge commit) so the canonical PR carries the full coverage. Tests in #24603 are hermetic and pass alongside the existing DeepSeek/Kimi suites (101/101 green locally).

Closing #24603 in favour of this PR once merged. CC @wesleysimplicio (#24605) in case the same additions are useful there.

Adding one live repro datapoint from #24726 before I close/defer it here.

I hit the same 400 in a real Hermes gateway session after the primary sub2api-openai provider failed over to Xiaomi MiMo:

Provider: xiaomi
Model: mimo-v2.5-pro
Endpoint: https://token-plan-cn.xiaomimimo.com/anthropic
Error: HTTP 400: Param Incorrect
Details: The reasoning_content in the thinking mode must be passed back to the API.

The request dump shape was slightly different from plain OpenAI tool_calls: the replayed assistant turns were already in Anthropic content-block form:

{"role":"assistant","content":[{"type":"tool_use","id":"call_...","name":"terminal","input":{...}}]}

In that dump there were 29 assistant tool-use turns and 0 had reasoning_content/thinking replay data. A local regression that made the failing dump pass was:

• detect xiaomimimo.com / mimo-* on the Anthropic adapter path
• preserve unsigned thinking blocks for MiMo, like Kimi/DeepSeek
• when a MiMo Anthropic replay assistant turn has tool_use blocks but no preserved thinking, prepend {"type":"thinking","thinking":" "}
• do not send the top-level Anthropic thinking parameter to MiMo's /anthropic fallback endpoint for that replay shape

Offline verification against the failing dump after that change:

assistant_count 29
tool_use_missing_placeholder []
top_level_has_thinking False
first_assistant_blocks ['thinking', 'tool_use']

I have no need to keep #24726 competing with this PR. If this live replay shape is already covered here, great. If not, it may be worth adding as a small extra regression before merge.