Skip to content
BranchesTags

Repository files navigation

OmniCopilot

中文版

A VS Code extension that lets you use models from multiple LLM platforms in GitHub Copilot Chat as language model providers.

Supported Providers

Provider Vendor ID Models
DeepSeek deepseek deepseek-v4-flash, deepseek-v4-pro, deepseek-v4-flash-vision-exp
GLM Coding Plan CN glm-coding-plan-cn GLM-5.3, GLM-5.3-Flash
Kimi Code Plan moonshot k3, k3-256k, kimi-for-coding, kimi-for-coding-highspeed
Moonshot (Open Platform) moonshot-open kimi-k3, kimi-k2.7-code, kimi-k2.7-code-highspeed, kimi-k2.6
Qwen Token Plan qwen qwen3.8-max, qwen3.8-flash, qwen3.7-max, qwen3.7-plus, qwen3.6-flash, glm-5.2, deepseek-v4-pro(-0813), deepseek-v4-flash-0731
MiniMax Token Plan CN minimax MiniMax-M3, MiniMax-M2.7, MiniMax-M2.7-highspeed, MiniMax-M2.5
Volcengine Coding Plan CN volcengine doubao-seed-2.1-turbo, doubao-seed-evolving, doubao-seed-2.0-lite, kimi-k2.7-code, minimax-m3, deepseek-v4-flash, deepseek-v4-pro, glm-5.3, glm-5.3-flash
Volcengine Agent Plan CN volcengine-agent-plan the Coding Plan set plus doubao-seed-2.0-mini and kimi-k3

Tested & Working

The following platforms have been tested and confirmed working:

  • DeepSeek Open Platform (platform.deepseek.com)
  • Kimi Code (Kimi coding model)
  • MiniMax Token Plan CN (platform.minimaxi.com)
  • GLM Coding Plan CN (open.bigmodel.cn Coding API — re-verification pending after the vendor ID change)

Note

GLM Coding Plan billing: per Zhipu's docs, the Coding endpoint (open.bigmodel.cn/api/coding/paas/v4) only counts toward the Coding Plan quota when called from officially supported tools (Claude Code, Kilo Code, OpenCode, TRAE, CodeBuddy, etc.). VS Code Copilot Chat is not on that list — success is not guaranteed, usage may be billed at pay-as-you-go API rates instead of your plan's credits, and Zhipu's usage notes treat non-listed-tool calls as a violation that may lead to throttling or account restrictions. Keep an eye on your billing and account status.

Warning

Qwen Token Plan terms and endpoint: the extension points at the Token Plan endpoint (token-plan.cn-beijing.maas.aliyuncs.com/compatible-mode/v1) and expects the subscription key (sk-sp-…) from platform.qianwenai.com — Token Plan and pay-as-you-go credentials/endpoints are fully isolated and must not be mixed (pay-as-you-go keys (sk-ws-…) belong to dashscope.aliyuncs.com/compatible-mode/v1). The key is also restricted to interactive coding/agent tools (Claude Code, Cursor, Qwen Code, Qoder, OpenClaw, etc.) — the docs explicitly forbid generic API usage and state that violations may suspend the subscription or ban the API key. VS Code Copilot Chat is not on the official tool list, so use at your own discretion and watch your account status.

TODO

  • Test Volcengine Coding Plan / Agent Plan
  • Test Qwen Token Plan
  • Test Kimi Open Platform
  • Support SiliconFlow
  • Support MiniMax International
  • Support GLM International
  • Support SiliconFlow International
  • Verify thinking effort levels (DeepSeek None/High/Max; others None/Low/Medium/High or None/On) actually take effect across providers
  • To be continued…

Features

  • Multiple Providers: Access models from major LLM platforms
  • Per-Model Thinking Effort: Hover any thinking-capable model in the Copilot picker to pick the effort level for the next turn — no need to flip a global switch
    • DeepSeek V4 menu: None / Low / High / Max (matches the V4 API's reasoning_effort domain; thinking is on by default, None disables it explicitly)
    • Kimi K3 (Code Plan k3 / k3-256k, Open Platform kimi-k3): Low / High / Max — no None option, thinking is always on; effort maps to reasoning_effort on both endpoints
    • GLM-5.3 / GLM-5.3-Flash menu: Low / High / Max — no None option, thinking is always on (the Coding endpoint routes old GLM IDs like glm-5.1 / glm-4.7 to these two models)
    • 4-level menu (None / Low / Medium / High) for Qwen reasoning models
    • 2-level menu (None / On) for models that only expose a thinking on/off knob (Kimi K2.6, MiniMax-M3, pre-5.3 GLM, Volcengine reasoning models) — MiniMax-M3's None genuinely disables thinking
    • Thinking-locked models expose no menu at all: K2.7 Code (Code Plan kimi-for-coding(-highspeed), Open Platform kimi-k2.7-code(-highspeed)) and MiniMax M2.x — their "None" would silently reroute the model or keep thinking on anyway
  • Thinking UI: Models with reasoning capabilities show collapsible thinking sections via LanguageModelThinkingPart
  • Context Gauge: Streams stream_options: { include_usage: true } and reports the real token usage back to Copilot Chat, so the context-window indicator shows actual usage instead of 0; falls back to CJK-aware token estimation (Chinese ≈ 1 token/char) before the first real usage arrives
  • Vision Support: Vision-capable models (deepseek-v4-flash-vision-exp, glm-5.3-flash, kimi-for-coding, MiniMax-M3, qwen3.8-max, qwen3.8-flash, qwen3.7-plus, qwen3.6-flash) can read images attached in Copilot Chat
  • Tool Calling: Function calling support for compatible models

Usage

  1. Install the extension
  2. Open Copilot Chat → Manage Models → Add Model
  3. Select a provider and enter your API key
  4. Start chatting with the selected model

Configuration

Thinking effort is now selected per model, per turn via the Copilot model picker's hover menu — there is no global thinking-effort setting.

Setting Description Default
omniCopilot.contextLength Max input context length (4K–1M presets, or custom) default
omniCopilot.customContextLength Custom max input context (used when contextLength is custom) 131072
omniCopilot.enableVision Enable vision for supported models true

Development

Prerequisites

Setup

git clone https://github.com/LyaQanYi/OmniCopilot.git
cd OmniCopilot
npm install

Build & Run

# Compile TypeScript
npm run compile

# Watch mode (auto-recompile on changes)
npm run watch

To debug the extension, press F5 in VS Code to launch an Extension Development Host with the extension loaded.

Package .vsix

npx @vscode/vsce package --no-dependencies

Project Structure

src/
├── extension.ts   # Extension entry point, activation & commands
├── provider.ts    # Language model provider implementation
├── api.ts         # API call logic (streaming, thinking, vision)
├── models.ts      # Preset model definitions per vendor
└── types.ts       # Shared TypeScript interfaces

Contributing

Contributions are welcome! Here's how you can help:

  1. Fork the repository
  2. Create a feature branch: git checkout -b feat/my-feature
  3. Commit your changes: git commit -m "feat: add my feature"
  4. Push to the branch: git push origin feat/my-feature
  5. Open a Pull Request

Guidelines

  • Follow existing code style (TypeScript strict mode)
  • Test with at least one provider before submitting
  • Keep commit messages clear and descriptive
  • One feature/fix per PR when possible

Ways to Contribute

  • Test a provider — Pick an untested platform from the TODO list, test it, and report results
  • Add a new provider — Add model definitions in models.ts and register in extension.ts
  • Fix bugs — Check Issues for reported problems
  • Improve docs — Help with documentation or translations

Requirements

  • VS Code 1.108.0+
  • GitHub Copilot extension

Changelog

0.4.1 — 2026-09-03

  • Fix: answer text leaking into the thinking block — when a model's answer quoted literal <think> / </think> strings (e.g. while reviewing this very codebase), the tag parser mistook them for delimiters and rerouted answer segments into the collapsible thinking section. Thinking now flows through two dedicated paths: reasoning_content deltas map straight to LanguageModelThinkingPart, content deltas stream verbatim as text; literal-tag parsing only runs for vendors that inline thinking inside content (MiniMax native API, opt-in via inlineThinkTags)
  • Context gauge no longer stuck at 0: streams stream_options: { include_usage: true } (with an automatic retry without the flag on 400/422) and reports the real usage to Copilot Chat via a LanguageModelDataPart (usage mime), lighting up the token indicator for extension-contributed models
  • CJK-aware token estimation in provideTokenCount: Chinese/Japanese/Korean text counts ≈ 1 token per character instead of length / 4, fixing 4-6x undercounts in Chinese-heavy conversations

0.4.0 — 2026-08-30

Pre-release: every provider re-verified against official docs. GLM API keys must be re-entered (vendor ID changed).

  • Renames: Bigmodel Plan → GLM Coding Plan CN (vendor ID now glm-coding-plan-cn), Qwen → Qwen Token Plan, MiniMax → MiniMax Token Plan CN
  • Kimi split into Kimi Code Plan (moonshot) and Moonshot Open Platform (moonshot-open): k3, k3-256k, kimi-for-coding(-highspeed), kimi-k3, kimi-k2.7-code(-highspeed), kimi-k2.6
  • Volcengine split into Volcengine Coding Plan CN (volcengine) and Volcengine Agent Plan CN (volcengine-agent-plan): Doubao Seed 2.1 Turbo / Seed Evolving / 2.0 Lite (plus 2.0 Mini and Kimi K3 on Agent Plan) hosted alongside kimi-k2.7-code, minimax-m3, deepseek-v4-flash/pro and glm-5.3(-flash); doubao-seed-2.0-pro, ark-code-latest and the stale third-party IDs are gone
  • New models: deepseek-v4-flash-vision-exp, qwen3.7-max, qwen3.6-flash, glm-5.3 / glm-5.3-flash; lineups trimmed to what each platform actually serves (GLM Coding endpoint keeps only 5.3 / 5.3-Flash)
  • Thinking semantics overhauled: new thinkingLocked (no picker menu) for K2.7 Code and MiniMax M2.x whose thinking cannot be disabled; three-level Low/High/Max menu for Kimi K3 and GLM-5.3(-Flash); Low added to DeepSeek where None genuinely disables; MiniMax-M3 and Volcengine "None" now send an explicit disable
  • Doc-verified effort knobs: DeepSeek reasoning_effort (low/high/max), Qwen thinking_budget (max 32768), DashScope-hosted GLM/DeepSeek on reasoning_effort, Zhipu tool_stream for streaming tool calls, MiniMax max_completion_tokens
  • Reliability: reasoning_content backfilled for tool loops on DeepSeek, GLM and Kimi Open Platform; vision lists refreshed (deepseek-v4-flash-vision-exp, glm-5.3-flash, MiniMax-M3, five Qwen models)
  • Custom model support removed: the custom-openai provider, the per-vendor customModelIds settings and the Add Custom Model ID command are gone — VS Code's built-in custom model flow covers this

0.3.0 — 2026-05-08

  • Per-model Thinking Effort picker in Copilot model selector — hover a thinking-capable model and choose effort for the next turn, no global setting needed
    • DeepSeek V4 menu: None / High / Max (matches the V4 API's reasoning_effort domain)
    • 4-level menu (None / Low / Medium / High) for Qwen reasoning models
    • 2-level menu (None / On) for GLM, Kimi, MiniMax, and Volcengine reasoning models
  • DeepSeek model list updated from deepseek-chat / deepseek-reasoner to deepseek-v4-flash / deepseek-v4-pro (1M input, 384K output, both reasoning-capable)
  • Removed global omniCopilot.enableThinking and omniCopilot.thinkingEffort settings, the matching status-bar items, and OmniCopilot: Toggle Thinking Mode / Set Thinking Effort commands — picker covers all cases now
  • Vendor-specific reasoning mapping reworked to handle the full None / On / Low / Medium / High / Max space:
    • DeepSeek: reasoning_effort: high|max when enabled, omitted when None
    • Qwen: enable_thinking + thinking_budget (1024 / 4096 / 16384 tokens; max → 16384)
    • Moonshot: explicit thinking: { type: "enabled"|"disabled" }
    • Volcengine: thinking: { type: "enabled" } only when enabled
    • Zhipu / MiniMax: no API knob, picker only controls output stripping

0.2.0 — 2026-04-30

  • Add user-configurable max input context length cap (presets 4K–1M + custom 1K-2M tokens), shown in status bar

0.1.3 — 2026-04-12

  • Add Volcengine Plan provider with 8 models (doubao-seed, minimax-m2.5, glm-4.7, deepseek-v3.2, kimi-k2.5)
  • Gate reasoning_content field by vendor capability — only send it for DeepSeek, Qwen, Moonshot/Kimi, and Zhipu; avoids request rejection on strict backends (Volcengine, MiniMax, custom)
  • Extract shared buildOpenAIMessages helper to deduplicate message-serialization logic between MultiModelChatProvider and CustomOpenAIProvider
  • Add reasoning_content handling in CustomOpenAIProvider

0.1.2 — 2026-04-09

  • Add Bigmodel Plan (Zhipu) provider with GLM-5.1, GLM-5-Turbo, GLM-4.7, GLM-4.5-Air models
  • Enable thinking capability for Zhipu models
  • Normalize Zhipu model IDs

0.1.1 — 2026-04-07

  • Enable thinking capability for MiniMax vendor
  • Enhance thinking support logic in both MultiModelChatProvider and CustomOpenAIProvider
  • Fix: flush remaining thinking buffer and pending tool calls after stream ends
  • Add MIT LICENSE file
  • Fix repository URL in package.json

0.1.0 — 2026-04-06

  • Initial release
  • Multi-model provider architecture with DeepSeek, Moonshot (Kimi), Qwen, MiniMax support
  • Custom OpenAI-compatible provider for any endpoint
  • Thinking mode with <think> tag parsing and collapsible UI (via LanguageModelThinkingPart)
  • Vision support for image-capable models
  • Tool calling / function calling support
  • Custom model ID management via settings and command palette
  • Configurable thinking effort (low / medium / high)

License

MIT

About

A versatile model provider for Copilot that enables integration with various LLMs (such as Kimi, DeepSeek, and more) through a unified API adapter.

Resources

Readme
MIT license
Activity

Stars

6 stars

Watchers

0 watching

Forks

4 forks
Report repository

Releases

Packages

Contributors

Languages