diff --git a/.claude-plugin/marketplace.json b/.claude-plugin/marketplace.json index 28a81243..73d893e1 100644 --- a/.claude-plugin/marketplace.json +++ b/.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.", diff --git a/README.md b/README.md index dd457d8f..32d324c0 100644 --- a/README.md +++ b/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 | diff --git a/dist/plugins/minimax/skills/minimax/README.md b/dist/plugins/minimax/skills/minimax/README.md new file mode 100644 index 00000000..70ca1afd --- /dev/null +++ b/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. diff --git a/dist/plugins/minimax/skills/minimax/SKILL.md b/dist/plugins/minimax/skills/minimax/SKILL.md new file mode 100644 index 00000000..57c5867b --- /dev/null +++ b/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`, `` 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. diff --git a/skills/minimax/README.md b/skills/minimax/README.md new file mode 100644 index 00000000..70ca1afd --- /dev/null +++ b/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. diff --git a/skills/minimax/SKILL.md b/skills/minimax/SKILL.md new file mode 100644 index 00000000..57c5867b --- /dev/null +++ b/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`, `` 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.