MiniMax Copilot

A GitHub Copilot Chat language model provider that adds MiniMax M3 and M2.7 to the Copilot model picker. Backed by a Token Plan API key.

Architecture ported from deepseek-v4-for-copilot (MIT) and adapted to the MiniMax Anthropic-compatible API.

What you get

• M3, M2.7, M2.7-highspeed in the Copilot model picker, with per-model context, output cap, and pricing shown in the tooltip.
• Image input on M3 (native multimodal) and a vision proxy fallback for M2.x so image attachments work on every model.
• Tool calling with an optional experimental cache-stabilisation switch that keeps the upstream prompt cache warm.
• Git commit message generation wired into the SCM input box. Conventional-Commits + gitmoji by default; drafts in place of whatever you had typed.
• Per-model sampling controls — set temperature, topK, etc. per model without editing code.
• Cumulative usage tracker — input / output / cache-read tokens across the whole extension lifetime, with a status command.
• Usage dashboard — a one-click webview (status-bar entry + a command) that renders today / 7-day / 30-day token usage, a 30-day bar chart, per-model breakdown, and the platform coding_plan/remains data (5h reset, weekly limit, subscription expiry) when an API key is configured. The counter updates live as new requests land.
• Diagnostics — per-request classifier, cache-hit stats, and a verbose mode that dumps every request to disk.
• Bilingual UI that follows the VS Code display language.

Requirements

• VS Code 1.111.0+
• MiniMax Token Plan subscription and API key
• VS Code Insiders is required to render MiniMax thinking blocks via the proposed languageModelThinkingPart API
• The built-in Git extension must be enabled for the commit-message generator

Quick start

1. Grab a Token Plan API key from Account / Token Plan.
2. Run MiniMax: Set API Key from the command palette.
3. Pick a model in the Copilot model picker. Run MiniMax: Show Pricing to compare costs.
4. Done — chat with the model directly in Copilot Chat, or use MiniMax: Generate Commit Message in the SCM input box.

Models

M3 1M context note: the >512K input tier is in limited rollout and the API rejects requests with max_tokens > 512_000. The effective input is capped at 512K until the rollout completes.

Historical models: M2.5 / M2.1 / M2 are no longer recommended by MiniMax and are not shipped. Power users can re-add them via minimax.modelIdOverrides + minimax.visibleModels.

Pricing (per million tokens, USD)

The prices below match the international billing site platform.minimax.io (https://api.minimax.io/anthropic). The Chinese mirror (platform.minimaxi.com, https://api.minimaxi.com/anthropic) charges in CNY at different rates — see README.zh.md for that table. The picker tooltip and MiniMax: Show Pricing both render the table that matches the active minimax.apiBaseUrl.

The M3 >512K input tier is in limited rollout and the API rejects requests with max_tokens > 512_000. The effective input is capped at 512K until the rollout completes. Token Plan subscription is billed separately.

Configuration

Commands

Per-model sampling

Two configuration knobs let you tweak the sampling parameters and request body per model ID without editing code:

// settings.json
"minimax.sampling": {
  "MiniMax-M2.7": { "temperature": 0.2, "topK": 40 },
  "MiniMax-M3":   { "topK": 80 }
},
"minimax.experimental.modelDefPresets": {
  "MiniMax-M3": { "service_tier": "auto" }
}

• temperature and topP are ignored when the model is in thinking: adaptive mode (Anthropic's constraint).
• topK and frequencyPenalty are always honoured.
• modelDefPresets is an escape hatch — any key you put there is merged into the request body verbatim (after the standard fields). Eleven reserved keys are rejected; tools is concatenated rather than replaced.

Git commit message generation

Run MiniMax: Generate Commit Message (also available from the ⋯ menu on the right of the SCM commit-message input box) to fill the input box with a draft. The generator:

• Reads the staged changes via VS Code's built-in Git extension; falls back to working-tree changes when nothing is staged.
• Caps the diff at 32 KB and the file list at 80 entries.
• Treats existing text in the input box as a draft to polish instead of starting from scratch.
• Emits Conventional-Commits-style messages (<type>(<scope>)<!>: <subject> + optional bullet body) with temperature: 0.2 and max_tokens: 256 for a reproducible first draft.

The model is selected by minimax.commitModel (default MiniMax-M2.7). Switch to MiniMax-M3 when the diff is hairy enough to benefit from deeper reasoning.

Thinking mode

MiniMax's Anthropic-compatible endpoint exposes a binary thinking: { type: "disabled" | "adaptive" } toggle — there is no user-facing effort knob. The extension always sends adaptive for thinking-capable models (M3) and skips the field entirely for M2.x (whose reasoning surfaces as <think>…</think> inside the text content). Forcing temperature: 1 and dropping top_p whenever thinking is on is the Anthropic constraint, not a choice.

Usage dashboard

Click the $(graph) MiniMax … item in the VS Code status bar, or run MiniMax: Open Usage Dashboard, to open a side-panel webview that combines two data sources:

• Local token accounting — every request the extension makes is recorded into a persistent counter. The dashboard aggregates it into three windows: Today, Last 7 days, and Last 30 days, with separate columns for Input, Cache read, Cache write, Output, and Requests. A 30-day bar chart and a per-model breakdown table are rendered below the windows. The counter updates live: every new request the chat provider makes bumps the day bucket and the dashboard re-renders without a manual refresh.
• Platform Token Plan — when an API key is configured, the dashboard also calls GET /v1/api/openplatform/coding_plan/remains and renders the 5-hour reset window, the weekly limit, the per-model quota table, and the subscription expiry date. Failures (401, network, malformed payload) are surfaced as a yellow banner; the local counters above remain accurate. The host (minimaxi.com vs minimax.io) is auto-picked from minimax.apiBaseUrl.

The dashboard has a Reset counters button that clears the local Memento after a confirmation prompt; the platform data cannot be reset from the extension.

Status-bar quota items

Two extra status-bar items live to the right of the daily-token counter so you can see your plan usage at a glance without opening the dashboard:

• $(bolt) 5h 73% — 5-hour reset window, remaining percent
• $(calendar) Week 11% — weekly limit, remaining percent

The colour follows the statusBarItem.remoteBackground / warningBackground / errorBackground theme tokens (green when plenty left, red when low), so the bar stays legible in both light and dark themes. Hovering shows a X / Y · resets in Hh Mm summary that mirrors the dashboard's quota card; clicking the item opens the dashboard. Without an API key both items render a muted em-dash placeholder and the tooltip nudges you to run MiniMax: Set API Key.

Both items are driven by the same in-process plan cache as the dashboard (fetchPlanUsage's 8s TTL still applies), so opening the dashboard or letting the status bar refresh does not add a second HTTP call.

Endpoint auto-selection

On first activation, if minimax.apiBaseUrl is still at its factory default, the extension picks an endpoint from the VS Code display language:

• zh* (zh-cn, zh-tw, zh-hk, zh-sg, …) → China https://api.minimaxi.com/anthropic.
• Anything else → international https://api.minimax.io/anthropic.

Once you set minimax.apiBaseUrl manually or invoke MiniMax: Switch to Global/Chinese API, the auto-selection is never re-applied.

Troubleshooting

• No models in the picker — run MiniMax: Show Provider Status to see whether an API key is set and which models are visible.
• HTTP 404 from the gateway — make sure minimax.apiBaseUrl points at a MiniMax Anthropic-compatible host (api.minimaxi.com/anthropic or api.minimax.io/anthropic), not a third-party proxy that expects the OpenAI protocol.
• "API key not configured" — run MiniMax: Set API Key; the key is stored in VS Code SecretStorage, not in settings.json.
• Empty commit-message draft — the diff may be larger than 32 KB. Run MiniMax: Show Logs to see what the generator saw.

License

SATA 2.0 (Star And Thank Author License). Chinese translation in LICENSE_zh.