fix: preserve Xiaomi MiMo thinking blocks

agent/anthropic_adapter.py

@@ -467,6 +467,23 @@ def _is_deepseek_anthropic_endpoint(base_url: str | None) -> bool:
     return "/anthropic" in normalized.rstrip("/").lower()
 
 
+def _is_xiaomi_mimo_anthropic_endpoint(base_url: str | None) -> bool:
+    """Return True for Xiaomi MiMo's Anthropic-compatible endpoint.
+
+    Xiaomi MiMo's ``/anthropic`` route speaks the Anthropic Messages protocol
+    but, when thinking mode is enabled, requires unsigned ``thinking`` blocks
+    synthesised from ``reasoning_content`` to round-trip on subsequent requests.
+    The match is pinned to the ``/anthropic`` path so a future OpenAI-compatible
+    Xiaomi base URL is not misclassified. See hermes-agent#24884.
+    """
+    if not base_url_host_matches(base_url or "", "xiaomimimo.com"):
+        return False
+    normalized = _normalize_base_url_text(base_url)
+    if not normalized:
+        return False
+    return "/anthropic" in normalized.rstrip("/").lower()
+
+
 def _requires_bearer_auth(base_url: str | None) -> bool:
     """Return True for Anthropic-compatible providers that require Bearer auth.
 
@@ -1745,15 +1762,17 @@ def convert_messages_to_anthropic(
     #    cache markers can interfere with signature validation.
     _THINKING_TYPES = frozenset(("thinking", "redacted_thinking"))
     _is_third_party = _is_third_party_anthropic_endpoint(base_url)
-    # Kimi /coding and DeepSeek /anthropic share a contract: both speak the
-    # Anthropic Messages protocol upstream but require that thinking blocks
-    # synthesised from reasoning_content round-trip on subsequent turns when
-    # thinking is enabled.  Signed Anthropic blocks still have to be stripped
-    # (neither endpoint can validate Anthropic's signatures); unsigned blocks
-    # are preserved.  See hermes-agent#13848 (Kimi) and #16748 (DeepSeek).
+    # Kimi /coding, DeepSeek /anthropic, and Xiaomi MiMo /anthropic share a
+    # contract: all speak the Anthropic Messages protocol upstream but require
+    # that thinking blocks synthesised from reasoning_content round-trip on
+    # subsequent turns when thinking is enabled. Signed Anthropic blocks still
+    # have to be stripped (none of these endpoints can validate Anthropic's
+    # signatures); unsigned blocks are preserved. See hermes-agent#13848 (Kimi),
+    # #16748 (DeepSeek), and #24884 (Xiaomi MiMo).
     _preserve_unsigned_thinking = (
         _is_kimi_family_endpoint(base_url, model)
         or _is_deepseek_anthropic_endpoint(base_url)
+        or _is_xiaomi_mimo_anthropic_endpoint(base_url)
     )
 
     last_assistant_idx = None

docs/specs/24884-xiaomi-mimo-thinking.md

@@ -0,0 +1,283 @@
+# Spec — Xiaomi MiMo Anthropic-compatible thinking-mode support
+
+- Issue: [#24884](https://github.com/NousResearch/hermes-agent/issues/24884)
+- Branch: `fix/24884-xiaomi-mimo-thinking`
+- Status: draft (pre-implementation)
+
+## Problem statement
+
+Xiaomi exposes the `mimo-v2.5-pro` model behind an Anthropic Messages-compatible
+endpoint at `https://token-plan-cn.xiaomimimo.com/anthropic`. When thinking mode
+is enabled and the conversation has at least one prior assistant turn, the
+upstream rejects replays with:
+
+```
+HTTP 400
+The reasoning_content in the thinking mode must be passed back to the API.
+```
+
+Root cause: `agent/anthropic_adapter.py::convert_messages_to_anthropic` treats
+every non-Anthropic host as a generic third-party endpoint and strips *all*
+`thinking` / `redacted_thinking` blocks from assistant messages — including the
+unsigned blocks Hermes synthesises from OpenAI-style `reasoning_content`. Kimi
+(`/coding`) and DeepSeek (`/anthropic`) already have carve-outs that preserve
+unsigned thinking while still stripping Anthropic-signed blocks (third-party
+proxies cannot validate Anthropic signatures). Xiaomi MiMo's `/anthropic` route
+follows the same contract and needs the same carve-out.
+
+## Goals
+
+1. Detect Xiaomi MiMo's Anthropic-compatible endpoint.
+2. Route it through the `_preserve_unsigned_thinking` branch alongside Kimi and
+   DeepSeek so unsigned `thinking` blocks survive the conversion.
+3. Continue stripping Anthropic-*signed* thinking blocks (Xiaomi cannot validate
+   them) and continue stripping `cache_control` from thinking blocks.
+4. Do not regress Kimi, DeepSeek, MiniMax, native Anthropic, or any other
+   provider.
+
+## Non-goals
+
+- Re-architecting the thinking-block pipeline.
+- Generic "any `/anthropic` path preserves unsigned thinking" rule — kept as
+  per-host carve-outs to fail closed.
+- Provider plugin registration, billing config, or auth changes — Xiaomi already
+  works for non-thinking requests; only the message-conversion bug is in scope.
+- Live integration testing against Xiaomi's servers.
+
+## Acceptance criteria
+
+- AC1: `_is_xiaomi_mimo_anthropic_endpoint("https://token-plan-cn.xiaomimimo.com/anthropic")`
+  returns `True`. Variants tested: trailing slash, `/anthropic/v1`, uppercase
+  host. Bare `https://xiaomimimo.com` (no `/anthropic` path) returns `False` so
+  any future OpenAI-compat base on the same domain is not misclassified.
+- AC2: Calling `convert_messages_to_anthropic(..., base_url=<Xiaomi /anthropic
+  URL>)` on a conversation with a replayed assistant tool-call message carrying
+  `reasoning_content="planning the tool call"` produces an assistant message
+  whose `content` list contains exactly one `{"type": "thinking", "thinking":
+  "planning the tool call"}` block (no `signature` field).
+- AC3: Across multiple prior assistant turns (not just the last), every one of
+  them retains its unsigned `thinking` block — Xiaomi validates history, not
+  just the latest turn.
+- AC4: A pre-existing Anthropic-signed thinking block (with `signature` or
+  `data`) on an assistant message is stripped — Xiaomi cannot validate
+  Anthropic-proprietary signatures.
+- AC5: `cache_control` keys are stripped from any `thinking` /
+  `redacted_thinking` block in the converted output.
+- AC6: Regression: an unrelated third-party `/anthropic` endpoint (use MiniMax
+  as the canary) still strips all thinking blocks. DeepSeek and Kimi tests
+  continue to pass.
+- AC7: `python -m pytest tests/agent/test_deepseek_anthropic_thinking.py
+  tests/agent/test_xiaomi_mimo_anthropic_thinking.py` passes locally and in CI.
+
+## Affected files
+
+| File | Change |
+| --- | --- |
+| `agent/anthropic_adapter.py` | Add `_is_xiaomi_mimo_anthropic_endpoint` helper; include in `_preserve_unsigned_thinking` predicate. |
+| `tests/agent/test_xiaomi_mimo_anthropic_thinking.py` | NEW. Mirror DeepSeek test file structure. |
+| `docs/specs/24884-xiaomi-mimo-thinking.md` | This document. |
+
+Nothing else (no provider registration, no model-metadata changes, no doctor /
+auth wiring) — the issue is purely a message-conversion bug for users who have
+already configured Xiaomi as a custom Anthropic-compatible provider.
+
+## Design
+
+### Detector
+
+Add next to `_is_deepseek_anthropic_endpoint`:
+
+```python
+def _is_xiaomi_mimo_anthropic_endpoint(base_url: str | None) -> bool:
+    """Return True for Xiaomi MiMo's Anthropic-compatible /anthropic route.
+
+    Xiaomi's ``token-plan-cn.xiaomimimo.com/anthropic`` endpoint speaks the
+    Anthropic Messages protocol but, when thinking mode is enabled, requires
+    unsigned thinking blocks (synthesised from ``reasoning_content``) to
+    round-trip on every replayed assistant turn — otherwise upstream returns::
+
+        The reasoning_content in the thinking mode must be passed back to the API.
+
+    Same strip-signed / keep-unsigned policy used for Kimi /coding and
+    DeepSeek /anthropic.  Match pinned to the ``/anthropic`` path so any future
+    OpenAI-compat base on the same domain is not misclassified.
+    See hermes-agent#24884.
+    """
+    if not base_url_host_matches(base_url or "", "xiaomimimo.com"):
+        return False
+    normalized = _normalize_base_url_text(base_url)
+    if not normalized:
+        return False
+    return "/anthropic" in normalized.rstrip("/").lower()
+```
+
+`base_url_host_matches` accepts subdomains (`token-plan-cn.xiaomimimo.com`
+matches `xiaomimimo.com`), which is what we want — Xiaomi documents the
+regional host today but may add others.
+
+### Predicate wire-up
+
+In `convert_messages_to_anthropic`:
+
+```python
+_preserve_unsigned_thinking = (
+    _is_kimi_family_endpoint(base_url, model)
+    or _is_deepseek_anthropic_endpoint(base_url)
+    or _is_xiaomi_mimo_anthropic_endpoint(base_url)
+)
+```
+
+Update the inline comment to add `#24884 (Xiaomi MiMo)` after the existing
+`#13848 (Kimi) and #16748 (DeepSeek)` references so the carve-out is
+self-documenting.
+
+No other branches need to change: the rest of the thinking-block handling
+(`_THINKING_TYPES`, `cache_control` strip, image eviction) is provider-agnostic.
+
+## TDD plan
+
+Write the test file first; it must fail against `main` and pass after the
+production change.
+
+`tests/agent/test_xiaomi_mimo_anthropic_thinking.py` — mirror the DeepSeek
+test layout, replacing host and issue references:
+
+1. `test_unsigned_thinking_block_survives_replay` — parametrised over
+   `https://token-plan-cn.xiaomimimo.com/anthropic`,
+   `https://token-plan-cn.xiaomimimo.com/anthropic/`,
+   `https://token-plan-cn.xiaomimimo.com/anthropic/v1`,
+   `https://Token-Plan-CN.XiaomiMiMo.com/anthropic`. Asserts a single unsigned
+   `thinking` block with the original `reasoning_content` text and no
+   `signature` key.
+2. `test_unsigned_thinking_preserved_on_non_latest_assistant_turn` — two prior
+   tool-call turns, both must retain their unsigned thinking blocks.
+3. `test_signed_anthropic_thinking_block_is_stripped` — assistant message with
+   `signature` set; converted output must contain zero `thinking` blocks.
+4. `test_cache_control_stripped_from_thinking_block` — invariant check on
+   converted output.
+5. `test_openai_compat_xiaomi_base_is_not_matched` — direct unit test of
+   `_is_xiaomi_mimo_anthropic_endpoint`:
+   - `https://token-plan-cn.xiaomimimo.com` → `False`
+   - `https://token-plan-cn.xiaomimimo.com/v1` → `False`
+   - `https://token-plan-cn.xiaomimimo.com/anthropic` → `True`
+   - `https://token-plan-cn.xiaomimimo.com/anthropic/v1` → `True`
+   - `https://other.xiaomimimo.com/anthropic` → `True` (subdomain coverage)
+6. `test_non_xiaomi_third_party_still_strips_all_thinking` — MiniMax canary,
+   asserts zero thinking blocks remain.
+
+Use class `TestXiaomiMiMoAnthropicPreservesThinking`. Reuse the exact message
+fixtures from the DeepSeek file — the contract is identical.
+
+## Docker / containerization considerations
+
+- No new runtime dependencies, no environment variables, no schema changes.
+- Tests run inside the existing pytest container without additional setup.
+- `Dockerfile` and `docker-compose*.yml` need no edits; do not touch them.
+- Verify the change works under the project's existing Docker workflow by
+  running the pytest target in the same container the CI uses (see
+  Verification §1).
+
+## Risks
+
+- **R1 — Hostname drift.** Xiaomi could add additional regional hosts
+  (`token-plan-na.xiaomimimo.com`, `mimo.xiaomi.com`, etc.). `base_url_host_matches`
+  on `xiaomimimo.com` covers any `*.xiaomimimo.com` subdomain today; if Xiaomi
+  later serves the same API under `xiaomi.com`, a follow-up will be needed. Out
+  of scope for this issue.
+- **R2 — Future OpenAI-compat base on the same domain.** Pinning to
+  `/anthropic` in the path keeps a future `xiaomimimo.com/v1` (OpenAI wire)
+  from being misrouted. Matches the DeepSeek precedent.
+- **R3 — Signed-block regression.** If Xiaomi ever begins issuing
+  Anthropic-compatible signatures we would over-strip — but the same risk
+  applies to Kimi and DeepSeek and has not materialised; the strip-signed
+  branch matches the documented compatibility behaviour.
+- **R4 — Live verification.** Without Xiaomi API credentials we rely on the
+  unit-level contract. The fix mirrors a pattern proven against two other
+  providers, which materially lowers this risk.
+
+## Senior Engineer Review
+
+Reviewing my own plan with a critical eye.
+
+**What's right**
+
+- Smallest viable change: one detector + one predicate slot. No new abstraction,
+  no premature "third-party Anthropic registry" refactor. Three carve-outs is
+  still under the rule-of-three threshold — converting to a table now would be
+  speculative.
+- Path-pinned detector (matches DeepSeek precedent) avoids the failure mode
+  where a same-domain OpenAI-compat base gets misclassified.
+- Test coverage mirrors an existing, proven file, so reviewers can diff the two
+  to spot omissions.
+- AC list pins both the new behaviour and the non-regression behaviour (MiniMax
+  canary, DeepSeek/Kimi suites green).
+
+**Concerns and adjustments**
+
+1. *Substring `/anthropic` could match `/anthropic-legacy` or `/anthropic2`.*
+   The DeepSeek detector has the same property and it has not bitten in
+   production. Tightening to a path-segment check (`startswith("/anthropic")`
+   after stripping the host, plus boundary check) would be safer but
+   inconsistent with DeepSeek. **Adjustment:** keep parity with DeepSeek for
+   now; if we tighten, do it for all three carve-outs in a follow-up. Noted as
+   R2 above.
+2. *Subdomain-matching surface area.* `base_url_host_matches("xiaomimimo.com")`
+   will accept *any* subdomain. That is desired for regional hosts but means a
+   hypothetical `staging.xiaomimimo.com` proxy with different semantics would
+   also be routed through this branch. **Adjustment:** acceptable — `/anthropic`
+   path pin narrows the surface, and Xiaomi-controlled subdomains are
+   presumed-honest by definition.
+3. *Should this be a generic config flag instead?* A `preserve_unsigned_thinking:
+   true` per-provider config would scale better. **Adjustment:** out of scope.
+   Users on `fix/24884` need the bug gone today; the config refactor is a
+   separate ticket and would require schema, docs, and migration. Three
+   carve-outs is still cheap.
+4. *No live integration test.* We can't hit Xiaomi from CI. The unit-test
+   contract is what the existing two providers ship with — accept the same
+   bar. The user reporting the issue will validate end-to-end against their
+   credentials before close.
+5. *Model name field — should `_preserve_unsigned_thinking` also detect Xiaomi
+   by model name (`mimo-*`) the way Kimi does?* Kimi has that branch because
+   private gateways front Kimi with custom hostnames. Xiaomi MiMo has no
+   evidence of that deployment pattern today. **Adjustment:** host-only match
+   for now; revisit if a private-gateway report comes in.
+6. *Thinking kwarg suppression.* Kimi has a second carve-out in `_build_kwargs`
+   that drops the `thinking` parameter entirely (Kimi drives thinking
+   server-side). Xiaomi's bug message asks us to *pass thinking content back*
+   — the kwarg itself is fine; the issue is in message replay. **Adjustment:**
+   no change to the `thinking` kwarg branch.
+
+**Verdict:** plan is minimal, mirrors precedent, has clear acceptance criteria
+and a non-regression canary. Proceed.
+
+## Verification commands
+
+Run from the repo root after implementation:
+
+```bash
+# 1. Targeted unit tests — must pass.
+python -m pytest -q \
+  tests/agent/test_xiaomi_mimo_anthropic_thinking.py \
+  tests/agent/test_deepseek_anthropic_thinking.py
+
+# 2. Broader adapter regression net.
+python -m pytest -q tests/agent/
+
+# 3. Lint / type-check parity with CI (only if the repo's lint target exists).
+ruff check agent/anthropic_adapter.py tests/agent/test_xiaomi_mimo_anthropic_thinking.py
+
+# 4. Confirm the new test file fails on main (TDD red).
+git stash && python -m pytest -q tests/agent/test_xiaomi_mimo_anthropic_thinking.py ; git stash pop
+```
+
+(Step 4 is informational — once the production patch is applied the test must
+go green.)
+
+## Out of scope (follow-ups, not this PR)
+
+- Live e2e test against `mimo-v2.5-pro`.
+- Consolidating Kimi/DeepSeek/Xiaomi into a config-driven registry.
+- Tightening `/anthropic` path detection to a strict segment check across all
+  three providers.
+- Adding Xiaomi to the provider picker / `model_metadata.py` if/when Hermes
+  ships it as a first-class provider.