feat: refresh MiniMax skill for M3.1 Flash defaults

.claude-plugin/marketplace.json

@@ -59,6 +59,23 @@
         "perplexity"
       ]
     },
+    {
+      "name": "minimax",
+      "description": "Configure MiniMax-M3.1-Flash and MiniMax-M3 text models, regional API endpoints, reasoning behavior, and token pricing.",
+      "source": "./dist/plugins/minimax",
+      "strict": false,
+      "skills": [
+        "./skills/minimax"
+      ],
+      "category": "ai-tools",
+      "keywords": [
+        "ai",
+        "minimax",
+        "minimax-m3.1-flash",
+        "regional-endpoints",
+        "model-configuration"
+      ]
+    },
     {
       "name": "agent-md-refactor",
       "description": "Refactor bloated AGENTS.md, CLAUDE.md, or similar agent instruction files to follow progressive disclosure principles. Splits monolithic files into organized, linked documentation.",

README.md

@@ -98,6 +98,7 @@ Add skills to project knowledge or paste SKILL.md contents into the conversation
 |----------|-------|-------------|
 | 🤖 AI Tools | [codex](skills/codex/README.md) | Advanced code analysis with GPT-5.2 |
 | 🤖 AI Tools | [gemini](skills/gemini/README.md) | Large-scale review (200k+ context) |
+| 🤖 AI Tools | [minimax](skills/minimax/README.md) | Flash-first MiniMax model configuration |
 | 🤖 AI Tools | [perplexity](skills/perplexity/README.md) | Web search & research |
 | 🔮 Meta | [agent-md-refactor](skills/agent-md-refactor/README.md) | Refactor bloated agent instruction files |
 | 🔮 Meta | [command-creator](skills/command-creator/README.md) | Create Claude Code slash commands |

dist/plugins/minimax/skills/minimax/README.md

@@ -0,0 +1,33 @@
+# MiniMax Skill
+
+Configure MiniMax-M3.1-Flash and MiniMax-M3 with the correct protocol, regional API base, capability, reasoning, and token-pricing guidance.
+
+## When to Use
+
+Use this skill when a request mentions:
+
+- MiniMax API configuration
+- `MiniMax-M3.1-Flash` or `MiniMax-M3`
+- Global or China-region MiniMax endpoints
+- MiniMax model capabilities, reasoning behavior, or cost estimates
+- Preserving an existing `MiniMax-M2.7` configuration
+
+## What It Covers
+
+- Flash-first active model selection with M3 as the second active choice
+- Preservation of explicit saved, custom, and legacy model IDs
+- Global and China-region endpoint families
+- Chat Completions, native Responses, and Anthropic Messages protocol guidance
+- Context windows, multimodal inputs, and reasoning behavior
+- Input, output, and cache-token pricing
+- API-key handling through `MINIMAX_API_KEY`
+
+## Installation
+
+Install the plugin from the Agent Toolkit marketplace:
+
+```text
+/plugin install minimax@agent-toolkit
+```
+
+Then ask the agent to configure MiniMax, select a region or model, preserve an existing model ID, or estimate token costs. See `SKILL.md` for the current workflow and configuration tables.

dist/plugins/minimax/skills/minimax/SKILL.md

@@ -0,0 +1,75 @@
+---
+name: minimax
+description: Configure MiniMax-M3.1-Flash and MiniMax-M3 text models, regional API endpoints, capabilities, thinking behavior, and token pricing. Use when users mention MiniMax-M3.1-Flash, MiniMax-M3, MiniMax API setup, or MiniMax regions.
+---
+
+# MiniMax Provider Guide
+
+Use this skill when configuring a MiniMax text model or preserving an existing MiniMax configuration.
+
+## Configuration Workflow
+
+1. Preserve an explicitly saved, custom, or legacy model ID. Do not silently replace an existing `MiniMax-M2.7` selection or alter non-MiniMax settings.
+2. For a new configuration without an explicit model, select `MiniMax-M3.1-Flash` as the default. Offer `MiniMax-M3` as the second active choice, including when disabled thinking is required.
+3. Ask which region the user needs: `global_en` or `cn_zh`. Default to `global_en` only when no region is specified. Keep the region and endpoint family together.
+4. Select the client's supported protocol. Keep an existing Chat Completions or Anthropic Messages integration working; native Responses is also available on the documented global endpoint.
+5. Read the API key from `MINIMAX_API_KEY`. Never print, log, or commit it.
+6. Present the chosen model, region, protocol, endpoint, and capability assumptions before making changes.
+
+## Regional Endpoints
+
+| Region | Chat Completions base URL | Anthropic Messages base URL | Documentation |
+| --- | --- | --- | --- |
+| `global_en` | `https://api.minimax.io/v1` | `https://api.minimax.io/anthropic` | `https://platform.minimax.io/docs` |
+| `cn_zh` | `https://api.minimaxi.com/v1` | `https://api.minimaxi.com/anthropic` | `https://platform.minimaxi.com/docs` |
+
+The [native Responses contract](https://platform.minimax.io/docs/api-reference/responses-create.md) documents `POST https://api.minimax.io/v1/responses`. Verify model availability and account entitlement for the selected region. Do not infer China Responses, third-party gateway, or subscription support from this international endpoint documentation.
+
+## Active Model Selection
+
+| Order | Model | Context window | Input modalities | Thinking behavior |
+| ---: | --- | ---: | --- | --- |
+| 1 | `MiniMax-M3.1-Flash` | 1,000,000 tokens | text, image, video | adaptive only; default effort `max` |
+| 2 | `MiniMax-M3` | 1,000,000 tokens | text, image, video | adaptive or disabled; effort does not tune depth |
+
+Flash supports up to 524,288 output tokens. Set an appropriate per-request output budget rather than automatically requesting the maximum; reasoning tokens consume that budget too.
+
+### Protocol-specific thinking
+
+For Flash, accepted effort values are `low`, `medium`, `high`, `xhigh`, and `max`; omission defaults to `max`. Disabling thinking or sending effort `none` returns HTTP 400. Do not send `service_tier` or `reasoning_split: false` for Flash.
+
+| Protocol | Flash effort field | M3 thinking default and control |
+| --- | --- | --- |
+| Chat Completions | `reasoning_effort` | On by default; `thinking.type` accepts `adaptive` or `disabled`. Effort does not tune depth. |
+| Anthropic Messages | `output_config.effort` | Off by default; enable with `thinking.type: adaptive`. Effort does not tune depth. |
+| Responses | `reasoning.effort` | Off by default; a non-`none` effort enables reasoning without tuning depth. |
+
+Preserve complete assistant, tool, reasoning, and multimodal content in subsequent turns. For Chat Completions this includes `tool_calls`, `<think>` content, and Flash's `reasoning_content`; for Messages, preserve all thinking, text, and tool-use blocks. For Responses, send full client-side history, including tool calls, results, and reasoning items: `store`, `previous_response_id`, `conversation`, and automatic truncation are unsupported or ignored.
+
+See the [Chat Completions guide](https://platform.minimax.io/docs/api-reference/text-openai-api.md) and [Messages guide](https://platform.minimax.io/docs/api-reference/text-anthropic-api.md) for the current wire formats.
+
+## Legacy Compatibility
+
+Preserve existing saved, custom, or legacy model IDs, including `MiniMax-M2.7`, without silently rewriting them. Keep legacy models out of ordinary active recommendations; verify their availability on the existing route separately. M2.7's older text-only capability and always-on thinking behavior do not describe Flash or M3.
+
+## Token Pricing
+
+Prices are in USD per million tokens. These are direct API rates; gateway and regional pricing must be checked separately against the selected provider's contract.
+
+| Model / standard input tier | Input | Output | Cache read | Cache write |
+| --- | ---: | ---: | ---: | ---: |
+| `MiniMax-M3.1-Flash` | $0.15 | $0.60 | $0.015 | Not specified |
+| `MiniMax-M3`, <=512k input tokens | $0.30 | $1.20 | $0.06 | Not specified |
+| `MiniMax-M3`, >512k input tokens | $0.60 | $2.40 | $0.12 | Not specified |
+
+M3 priority admission costs 1.5 times the standard rate and uses `service_tier: priority`; this field is unsupported for Flash. Retain the provider's published `512k` notation without inventing an exact integer threshold. Confirm the [current pricing](https://platform.minimax.io/docs/pricing/overview.md), state assumptions, and calculate each token category separately. Do not invent an unspecified cache-write price.
+
+## Validation Checklist
+
+- Ordinary choices list `MiniMax-M3.1-Flash` first and `MiniMax-M3` second.
+- Explicit saved, custom, and legacy model IDs are preserved, including `MiniMax-M2.7`.
+- The endpoint belongs to the selected region and protocol.
+- Flash does not send disabled thinking, effort `none`, `service_tier`, or `reasoning_split: false`.
+- M3 thinking defaults are selected by protocol; multimodal inputs remain intact.
+- Full tool and reasoning history is retained client-side.
+- The API key remains in `MINIMAX_API_KEY` and is never included in output.

skills/minimax/README.md

@@ -0,0 +1,33 @@
+# MiniMax Skill
+
+Configure MiniMax-M3.1-Flash and MiniMax-M3 with the correct protocol, regional API base, capability, reasoning, and token-pricing guidance.
+
+## When to Use
+
+Use this skill when a request mentions:
+
+- MiniMax API configuration
+- `MiniMax-M3.1-Flash` or `MiniMax-M3`
+- Global or China-region MiniMax endpoints
+- MiniMax model capabilities, reasoning behavior, or cost estimates
+- Preserving an existing `MiniMax-M2.7` configuration
+
+## What It Covers
+
+- Flash-first active model selection with M3 as the second active choice
+- Preservation of explicit saved, custom, and legacy model IDs
+- Global and China-region endpoint families
+- Chat Completions, native Responses, and Anthropic Messages protocol guidance
+- Context windows, multimodal inputs, and reasoning behavior
+- Input, output, and cache-token pricing
+- API-key handling through `MINIMAX_API_KEY`
+
+## Installation
+
+Install the plugin from the Agent Toolkit marketplace:
+
+```text
+/plugin install minimax@agent-toolkit
+```
+
+Then ask the agent to configure MiniMax, select a region or model, preserve an existing model ID, or estimate token costs. See `SKILL.md` for the current workflow and configuration tables.

skills/minimax/SKILL.md

@@ -0,0 +1,75 @@
+---
+name: minimax
+description: Configure MiniMax-M3.1-Flash and MiniMax-M3 text models, regional API endpoints, capabilities, thinking behavior, and token pricing. Use when users mention MiniMax-M3.1-Flash, MiniMax-M3, MiniMax API setup, or MiniMax regions.
+---
+
+# MiniMax Provider Guide
+
+Use this skill when configuring a MiniMax text model or preserving an existing MiniMax configuration.
+
+## Configuration Workflow
+
+1. Preserve an explicitly saved, custom, or legacy model ID. Do not silently replace an existing `MiniMax-M2.7` selection or alter non-MiniMax settings.
+2. For a new configuration without an explicit model, select `MiniMax-M3.1-Flash` as the default. Offer `MiniMax-M3` as the second active choice, including when disabled thinking is required.
+3. Ask which region the user needs: `global_en` or `cn_zh`. Default to `global_en` only when no region is specified. Keep the region and endpoint family together.
+4. Select the client's supported protocol. Keep an existing Chat Completions or Anthropic Messages integration working; native Responses is also available on the documented global endpoint.
+5. Read the API key from `MINIMAX_API_KEY`. Never print, log, or commit it.
+6. Present the chosen model, region, protocol, endpoint, and capability assumptions before making changes.
+
+## Regional Endpoints
+
+| Region | Chat Completions base URL | Anthropic Messages base URL | Documentation |
+| --- | --- | --- | --- |
+| `global_en` | `https://api.minimax.io/v1` | `https://api.minimax.io/anthropic` | `https://platform.minimax.io/docs` |
+| `cn_zh` | `https://api.minimaxi.com/v1` | `https://api.minimaxi.com/anthropic` | `https://platform.minimaxi.com/docs` |
+
+The [native Responses contract](https://platform.minimax.io/docs/api-reference/responses-create.md) documents `POST https://api.minimax.io/v1/responses`. Verify model availability and account entitlement for the selected region. Do not infer China Responses, third-party gateway, or subscription support from this international endpoint documentation.
+
+## Active Model Selection
+
+| Order | Model | Context window | Input modalities | Thinking behavior |
+| ---: | --- | ---: | --- | --- |
+| 1 | `MiniMax-M3.1-Flash` | 1,000,000 tokens | text, image, video | adaptive only; default effort `max` |
+| 2 | `MiniMax-M3` | 1,000,000 tokens | text, image, video | adaptive or disabled; effort does not tune depth |
+
+Flash supports up to 524,288 output tokens. Set an appropriate per-request output budget rather than automatically requesting the maximum; reasoning tokens consume that budget too.
+
+### Protocol-specific thinking
+
+For Flash, accepted effort values are `low`, `medium`, `high`, `xhigh`, and `max`; omission defaults to `max`. Disabling thinking or sending effort `none` returns HTTP 400. Do not send `service_tier` or `reasoning_split: false` for Flash.
+
+| Protocol | Flash effort field | M3 thinking default and control |
+| --- | --- | --- |
+| Chat Completions | `reasoning_effort` | On by default; `thinking.type` accepts `adaptive` or `disabled`. Effort does not tune depth. |
+| Anthropic Messages | `output_config.effort` | Off by default; enable with `thinking.type: adaptive`. Effort does not tune depth. |
+| Responses | `reasoning.effort` | Off by default; a non-`none` effort enables reasoning without tuning depth. |
+
+Preserve complete assistant, tool, reasoning, and multimodal content in subsequent turns. For Chat Completions this includes `tool_calls`, `<think>` content, and Flash's `reasoning_content`; for Messages, preserve all thinking, text, and tool-use blocks. For Responses, send full client-side history, including tool calls, results, and reasoning items: `store`, `previous_response_id`, `conversation`, and automatic truncation are unsupported or ignored.
+
+See the [Chat Completions guide](https://platform.minimax.io/docs/api-reference/text-openai-api.md) and [Messages guide](https://platform.minimax.io/docs/api-reference/text-anthropic-api.md) for the current wire formats.
+
+## Legacy Compatibility
+
+Preserve existing saved, custom, or legacy model IDs, including `MiniMax-M2.7`, without silently rewriting them. Keep legacy models out of ordinary active recommendations; verify their availability on the existing route separately. M2.7's older text-only capability and always-on thinking behavior do not describe Flash or M3.
+
+## Token Pricing
+
+Prices are in USD per million tokens. These are direct API rates; gateway and regional pricing must be checked separately against the selected provider's contract.
+
+| Model / standard input tier | Input | Output | Cache read | Cache write |
+| --- | ---: | ---: | ---: | ---: |
+| `MiniMax-M3.1-Flash` | $0.15 | $0.60 | $0.015 | Not specified |
+| `MiniMax-M3`, <=512k input tokens | $0.30 | $1.20 | $0.06 | Not specified |
+| `MiniMax-M3`, >512k input tokens | $0.60 | $2.40 | $0.12 | Not specified |
+
+M3 priority admission costs 1.5 times the standard rate and uses `service_tier: priority`; this field is unsupported for Flash. Retain the provider's published `512k` notation without inventing an exact integer threshold. Confirm the [current pricing](https://platform.minimax.io/docs/pricing/overview.md), state assumptions, and calculate each token category separately. Do not invent an unspecified cache-write price.
+
+## Validation Checklist
+
+- Ordinary choices list `MiniMax-M3.1-Flash` first and `MiniMax-M3` second.
+- Explicit saved, custom, and legacy model IDs are preserved, including `MiniMax-M2.7`.
+- The endpoint belongs to the selected region and protocol.
+- Flash does not send disabled thinking, effort `none`, `service_tier`, or `reasoning_split: false`.
+- M3 thinking defaults are selected by protocol; multimodal inputs remain intact.
+- Full tool and reasoning history is retained client-side.
+- The API key remains in `MINIMAX_API_KEY` and is never included in output.