feat: Add MiniMax M3 tool-calling, reasoning, and conformance coverage

conformance/reasoning/fixtures/minimax_m3/REASONING.batch.yaml

@@ -0,0 +1,178 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+family: minimax_m3
+model_label: MiniMax M3
+mode: batch
+refs:
+  local: &local_ref 'MiniMax M3 <mm:think> reasoning grammar; local taxonomy: lib/parsers/REASONING_CASES.md'
+cases:
+  REASONING.batch.1.a:
+    description: Empty input
+    ref: *local_ref
+    model_text: ''
+    expected:
+      dynamo:
+        reasoning_text: ''
+        normal_text: ''
+      vllm: &vllm_unavailable
+        unavailable: vLLM Python 0.23.0 in this conformance harness does not include the MiniMax M3 reasoning parser yet.
+      sglang: &sglang_unavailable
+        unavailable: SGLang 0.5.12.post1 in this conformance harness does not include the MiniMax M3 reasoning parser yet.
+  REASONING.batch.1.b:
+    description: Plain text without reasoning markers
+    ref: *local_ref
+    model_text: plain answer
+    expected:
+      dynamo:
+        reasoning_text: ''
+        normal_text: plain answer
+      vllm: *vllm_unavailable
+      sglang: *sglang_unavailable
+  REASONING.batch.2.a:
+    description: Single MiniMax M3 reasoning block
+    ref: *local_ref
+    model_text: <mm:think>thinking</mm:think>
+    expected:
+      dynamo:
+        reasoning_text: thinking
+        normal_text: ''
+      vllm: *vllm_unavailable
+      sglang: *sglang_unavailable
+  REASONING.batch.2.b:
+    description: Normal prefix before a MiniMax M3 reasoning block
+    ref: *local_ref
+    model_text: Before <mm:think>thinking</mm:think>
+    expected:
+      dynamo:
+        reasoning_text: thinking
+        normal_text: Before
+      vllm: *vllm_unavailable
+      sglang: *sglang_unavailable
+  REASONING.batch.2.c:
+    description: MiniMax M3 reasoning followed by normal answer text
+    ref: *local_ref
+    model_text: <mm:think>compute it</mm:think>The answer is 42.
+    expected:
+      dynamo:
+        reasoning_text: compute it
+        normal_text: The answer is 42.
+      vllm: *vllm_unavailable
+      sglang: *sglang_unavailable
+  REASONING.batch.2.d:
+    description: Normal text around a MiniMax M3 reasoning block
+    ref: *local_ref
+    model_text: Before <mm:think>thinking</mm:think> after
+    expected:
+      dynamo:
+        reasoning_text: thinking
+        normal_text: Before  after
+      vllm: *vllm_unavailable
+      sglang: *sglang_unavailable
+  REASONING.batch.2.e:
+    description: Empty MiniMax M3 reasoning block
+    ref: *local_ref
+    model_text: <mm:think></mm:think>Direct answer.
+    expected:
+      dynamo:
+        reasoning_text: ''
+        normal_text: Direct answer.
+      vllm: *vllm_unavailable
+      sglang: *sglang_unavailable
+  REASONING.batch.2.f:
+    description: Complex MiniMax M3 reasoning content with newlines and marker-looking text
+    ref: *local_ref
+    model_text: |-
+      <mm:think>Line 1
+      JSON-ish: {"x": [1, true]}
+      Marker-like: <mm:thinking>not a close</mm:thinking></mm:think>Done.
+    expected:
+      dynamo:
+        reasoning_text: |-
+          Line 1
+          JSON-ish: {"x": [1, true]}
+          Marker-like: <mm:thinking>not a close</mm:thinking>
+        normal_text: Done.
+      vllm: *vllm_unavailable
+      sglang: *sglang_unavailable
+  REASONING.batch.3.a:
+    description: Reasoning followed by downstream MiniMax M3 tool-call text
+    ref: *local_ref
+    model_text: |-
+      <mm:think>choose the weather tool</mm:think>]<]minimax[>[<tool_call>
+      ]<]minimax[>[<invoke name="get_weather">
+      ]<]minimax[>[<location>NYC]<]minimax[>[</location>
+      ]<]minimax[>[</invoke>
+      ]<]minimax[>[</tool_call>
+    expected:
+      dynamo:
+        reasoning_text: choose the weather tool
+        normal_text: |-
+          ]<]minimax[>[<tool_call>
+          ]<]minimax[>[<invoke name="get_weather">
+          ]<]minimax[>[<location>NYC]<]minimax[>[</location>
+          ]<]minimax[>[</invoke>
+          ]<]minimax[>[</tool_call>
+      vllm: *vllm_unavailable
+      sglang: *sglang_unavailable
+  REASONING.batch.3.b:
+    description: Open reasoning interrupted by downstream tool-call text is not configured for MiniMax M3
+    ref: *local_ref
+    reason: MiniMax M3 uses explicit <mm:think> spans and has no configured downstream tool-call escape while a reasoning span is still open; missing-end recovery is covered by REASONING.batch.5.
+  REASONING.batch.3.c:
+    description: Recipientless Harmony commentary is not applicable to MiniMax M3
+    ref: *local_ref
+    reason: MiniMax M3 does not use Harmony commentary channels; visible text and downstream tool handoff use MiniMax-specific think and namespace-token tool-call markers.
+  REASONING.batch.3.d:
+    description: Directed Harmony commentary tool calls are not applicable to MiniMax M3
+    ref: *local_ref
+    reason: MiniMax M3 does not use Harmony `commentary to=functions.*` envelopes; downstream tool calls use MiniMax namespace-token XML markers.
+  REASONING.batch.3.e:
+    description: Directed Harmony analysis tool calls are not applicable to MiniMax M3
+    ref: *local_ref
+    reason: MiniMax M3 does not use Harmony `analysis to=functions.*` envelopes; reasoning is represented by <mm:think> tags.
+  REASONING.batch.3.f:
+    description: Harmony analysis/commentary/final channel sequencing is not applicable to MiniMax M3
+    ref: *local_ref
+    reason: MiniMax M3 has no Harmony channel protocol, so mixed analysis, commentary tool-call, and final-channel sequencing is outside its grammar.
+  REASONING.batch.4:
+    description: Dangling end marker without an opening MiniMax M3 think tag
+    ref: *local_ref
+    model_text: reason</mm:think>answer
+    expected:
+      dynamo:
+        reasoning_text: reason
+        normal_text: answer
+      vllm: *vllm_unavailable
+      sglang: *sglang_unavailable
+  REASONING.batch.5:
+    description: Missing end marker keeps the remaining text as reasoning
+    ref: *local_ref
+    model_text: <mm:think>partial reasoning
+    expected:
+      dynamo:
+        reasoning_text: partial reasoning
+        normal_text: ''
+      vllm: *vllm_unavailable
+      sglang: *sglang_unavailable
+  REASONING.batch.6.a:
+    description: Multiple MiniMax M3 reasoning spans are concatenated
+    ref: *local_ref
+    model_text: <mm:think>first</mm:think> middle <mm:think>second</mm:think> done
+    expected:
+      dynamo:
+        reasoning_text: firstsecond
+        normal_text: middle  done
+      vllm: *vllm_unavailable
+      sglang: *sglang_unavailable
+  REASONING.batch.6.b:
+    description: Stray MiniMax M3 end marker after a complete reasoning span
+    ref: *local_ref
+    model_text: <mm:think>preamble</mm:think>body</mm:think>answer
+    expected:
+      dynamo:
+        reasoning_text: preamble
+        normal_text: bodyanswer
+        reason: Parser-owned dangling close syntax is stripped from normal_text after the complete reasoning span.
+      vllm: *vllm_unavailable
+      sglang: *sglang_unavailable

conformance/reasoning/fixtures/minimax_m3/REASONING.stream.yaml

@@ -0,0 +1,132 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+family: minimax_m3
+model_label: MiniMax M3
+mode: stream
+refs:
+  local: &local_ref 'MiniMax M3 <mm:think> reasoning grammar; local taxonomy: lib/parsers/REASONING_CASES.md'
+cases:
+  REASONING.stream.1.a:
+    description: Empty stream chunk
+    ref: *local_ref
+    chunks:
+    - ''
+    expected:
+      dynamo:
+        reasoning_text: ''
+        normal_text: ''
+      vllm: &vllm_unavailable
+        unavailable: vLLM Python 0.23.0 in this conformance harness does not include the MiniMax M3 reasoning parser yet.
+      sglang: &sglang_unavailable
+        unavailable: SGLang 0.5.12.post1 in this conformance harness does not include the MiniMax M3 reasoning parser yet.
+  REASONING.stream.1.b:
+    description: Plain text without reasoning markers
+    ref: *local_ref
+    chunks:
+    - plain
+    - ' answer'
+    expected:
+      dynamo:
+        reasoning_text: ''
+        normal_text: plain answer
+      vllm: *vllm_unavailable
+      sglang: *sglang_unavailable
+  REASONING.stream.2.a:
+    description: MiniMax M3 reasoning split across chunks
+    ref: *local_ref
+    chunks:
+    - <mm:think>rea
+    - son</mm:think>answer
+    expected:
+      dynamo:
+        reasoning_text: reason
+        normal_text: answer
+      vllm: *vllm_unavailable
+      sglang: *sglang_unavailable
+  REASONING.stream.2.b:
+    description: Multiple MiniMax M3 reasoning spans across chunks
+    ref: *local_ref
+    chunks:
+    - <mm:think>first</mm:think>
+    - ' middle '
+    - <mm:think>second</mm:think> done
+    expected:
+      dynamo:
+        reasoning_text: firstsecond
+        normal_text: ' middle  done'
+      vllm: *vllm_unavailable
+      sglang: *sglang_unavailable
+  REASONING.stream.2.c:
+    description: Prompt-prefilled close marker after a complete MiniMax M3 span
+    ref: *local_ref
+    chunks:
+    - <mm:think>preamble</mm:think>
+    - body</mm:think>
+    - answer
+    expected:
+      dynamo:
+        reasoning_text: preamblebody
+        normal_text: answer
+        reason: MiniMax M3 enables dangling-end recovery, so a later unpaired </mm:think> chunk is treated as a prompt-prefilled reasoning close.
+      vllm: *vllm_unavailable
+      sglang: *sglang_unavailable
+  REASONING.stream.3.a:
+    description: MiniMax M3 start marker split across chunks
+    ref: *local_ref
+    chunks:
+    - <mm:thi
+    - nk>thinking</mm:think>answer
+    expected:
+      dynamo:
+        reasoning_text: thinking
+        normal_text: answer
+      vllm: *vllm_unavailable
+      sglang: *sglang_unavailable
+  REASONING.stream.3.b:
+    description: MiniMax M3 end marker split across chunks
+    ref: *local_ref
+    chunks:
+    - <mm:think>thinking</mm:
+    - think>answer
+    expected:
+      dynamo:
+        reasoning_text: thinking
+        normal_text: answer
+      vllm: *vllm_unavailable
+      sglang: *sglang_unavailable
+  REASONING.stream.3.c:
+    description: Partial MiniMax M3 start-marker prefix later becomes normal text
+    ref: *local_ref
+    chunks:
+    - plain <mm:th
+    - esis answer
+    expected:
+      dynamo:
+        reasoning_text: ''
+        normal_text: plain <mm:thesis answer
+      vllm: *vllm_unavailable
+      sglang: *sglang_unavailable
+  REASONING.stream.4:
+    description: Dangling MiniMax M3 end marker split across chunks
+    ref: *local_ref
+    chunks:
+    - reason</mm:
+    - think>answer
+    expected:
+      dynamo:
+        reasoning_text: reason
+        normal_text: answer
+      vllm: *vllm_unavailable
+      sglang: *sglang_unavailable
+  REASONING.stream.4.a:
+    description: Close marker only is stripped as prompt-prefilled reasoning close
+    ref: *local_ref
+    chunks:
+    - </mm:think>I'll check the content of both files.
+    expected:
+      dynamo:
+        reasoning_text: ''
+        normal_text: I'll check the content of both files.
+      vllm: *vllm_unavailable
+      sglang: *sglang_unavailable

conformance/toolcalling/fixtures-stream-v2/minimax_m3/TOOLCALLING.streamv2.1.yaml

@@ -0,0 +1,38 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+# Streaming fixture (per-chunk, per-impl). MiniMax M3 is present in the
+# stream-v2 matrix, but no pinned peer/local v2 stream parser is available yet.
+
+family: minimax_m3
+model_label: 'MiniMax M3'
+mode: streamv2
+captured_with:
+  vllm_rust: 'v0.23.0 78743ab5bffd381e88f97e1c8ba20473b0ae6d75'
+  vllm_python: '0.23.0'
+  sglang_python: '0.5.12.post1'
+cases:
+  TOOLCALLING.streamv2.1:
+    description: 'Single tool call (happy path)'
+    ref: 'derived from TOOLCALLING.batch.1'
+    tools:
+    - name: get_weather
+      parameters:
+        type: object
+        properties:
+          location:
+            type: string
+    unavailable:
+      dynamo_rust: 'Dynamo parser v2 TC streaming not yet implemented for this family; vLLM Python/SGLang per-chunk output is the target to match.'
+      vllm_rust: 'vLLM Rust v0.23.0 in this conformance harness does not include the MiniMax M3 tool parser yet.'
+      vllm_python: 'vLLM Python 0.23.0 in this conformance harness does not include the MiniMax M3 tool parser yet.'
+      sglang_python: 'SGLang 0.5.12.post1 in this conformance harness does not include the MiniMax M3 detector yet.'
+    chunks:
+    - delta_text: |-
+        ]<]minimax[>[<tool_call>
+        ]<]minimax[>[<invoke name="get_weather">
+        ]<]minimax[>[<location>NYC]<]minimax[>[</location>
+        ]<]minimax[>[</invoke>
+        ]<]minimax[>[</tool_call>
+    - delta_text: ''
+      finish_reason: tool_calls

conformance/toolcalling/fixtures-stream-v2/minimax_m3/TOOLCALLING.streamv2.10.yaml

@@ -0,0 +1,38 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+family: minimax_m3
+model_label: 'MiniMax M3'
+mode: streamv2
+captured_with:
+  vllm_rust: 'v0.23.0 78743ab5bffd381e88f97e1c8ba20473b0ae6d75'
+  vllm_python: '0.23.0'
+  sglang_python: '0.5.12.post1'
+cases:
+  TOOLCALLING.streamv2.10:
+    description: 'Duplicate calls (same name twice)'
+    ref: 'derived from TOOLCALLING.batch.10'
+    tools:
+    - name: get_weather
+      parameters:
+        type: object
+        properties:
+          location:
+            type: string
+    unavailable:
+      dynamo_rust: 'Dynamo parser v2 TC streaming not yet implemented for this family; vLLM Python/SGLang per-chunk output is the target to match.'
+      vllm_rust: 'vLLM Rust v0.23.0 in this conformance harness does not include the MiniMax M3 tool parser yet.'
+      vllm_python: 'vLLM Python 0.23.0 in this conformance harness does not include the MiniMax M3 tool parser yet.'
+      sglang_python: 'SGLang 0.5.12.post1 in this conformance harness does not include the MiniMax M3 detector yet.'
+    chunks:
+    - delta_text: |-
+        ]<]minimax[>[<tool_call>
+        ]<]minimax[>[<invoke name="get_weather">
+        ]<]minimax[>[<location>NYC]<]minimax[>[</location>
+        ]<]minimax[>[</invoke>
+        ]<]minimax[>[<invoke name="get_weather">
+        ]<]minimax[>[<location>LA]<]minimax[>[</location>
+        ]<]minimax[>[</invoke>
+        ]<]minimax[>[</tool_call>
+    - delta_text: ''
+      finish_reason: tool_calls