LingYun

LingYun is a VS Code extension for agentic coding work. It can inspect your workspace, edit files, run shell commands, use VS Code language features, delegate to subagents, persist sessions, compact long context, and optionally build transcript-backed memories.

Providers

LingYun currently supports three LLM backends:

• GitHub Copilot
• ChatGPT Codex Subscription
• OpenAI-compatible servers

Provider selection is controlled by lingyun.llmProvider.

Quickstart

1. Open the LingYun view.
2. Choose a provider and model from the chat header.
3. Describe a task, for example:
   • “Find where this function is defined and explain it.”
   • “Refactor this file and run tests.”
   • “Search the repo for X and update the docs.”

New tasks run with:

• lingyun.mode = build by default
• lingyun.planFirst = true by default

That means LingYun normally drafts a plan first for a new top-level task, then you can continue into execution. If you want strict read-only behavior, switch the mode to Plan.

Modes

• Build: full tool use, including edits and shell commands, subject to approvals.
• Plan: read-only. Write/edit tools are denied and auto-approve is disabled.
• Plan First: a UI behavior controlled by lingyun.planFirst. This is separate from lingyun.mode.

Built-in tools

LingYun ships with these built-in tools:

• Workspace discovery: list, glob, grep, read, read_range
• Editing: edit, write
• Code intelligence: lsp, symbols_search, symbols_peek
• Shell: bash
• Reusable instructions: skill
• Delegation: task
• Todos: todoread, todowrite
• Memory access: get_memory

Workspace-defined tools can also be loaded from:

• .vscode/agent-tools.json
• .vscode/agent-tools/*.json

Schema: schemas/agent-tools.schema.json

Approvals and path safety

• Tool approvals are driven by tool metadata plus lingyun.autoApprove.
• lingyun.autoApprove only affects Build mode.
• External paths are blocked by default, even when auto-approve is enabled.
• External-path access is controlled by lingyun.security.allowExternalPaths.

Providers and setup

GitHub Copilot

Copilot is the default provider:

{
  "lingyun.llmProvider": "copilot",
  "lingyun.model": "gpt-4o"
}

lingyun.copilot.reasoningEffort defaults to high and is used for GPT-5-family Copilot requests.

ChatGPT Codex Subscription

The Codex subscription provider uses a ChatGPT account session and exposes sign-in/sign-out in the chat header when auth UI is available.

If lingyun.model is empty or still set to the Copilot default, LingYun falls back to lingyun.codexSubscription.defaultModelId, which defaults to gpt-5.3-codex.

{
  "lingyun.llmProvider": "codexSubscription",
  "lingyun.codexSubscription.defaultModelId": "gpt-5.4"
}

OpenAI-compatible server

{
  "lingyun.llmProvider": "openaiCompatible",
  "lingyun.openaiCompatible.baseURL": "http://localhost:8080/v1",
  "lingyun.openaiCompatible.defaultModelId": "your-model-id",
  "lingyun.openaiCompatible.maxTokens": 32000
}

If your server needs a key, LingYun reads it from OPENAI_API_KEY by default. Override that env var name with lingyun.openaiCompatible.apiKeyEnv.

Sessions, compaction, and memories

Sessions

• lingyun.sessions.persist defaults to true
• Persisted sessions are workspace-scoped
• Clear them with LingYun: Clear Saved Sessions

Context compaction

LingYun automatically compacts long sessions by default:

{
  "lingyun.compaction.auto": true,
  "lingyun.compaction.prune": true,
  "lingyun.compaction.toolOutputMode": "afterToolCall"
}

You can also compact the current session manually with LingYun: Compact Session.

Per-model context limits live under lingyun.modelLimits. Keys can be either plain model ids or provider-scoped keys:

{
  "lingyun.modelLimits": {
    "gpt-4o": { "context": 128000, "output": 32000 },
    "codexSubscription:gpt-5.4": { "context": 272000, "output": 32000 }
  }
}

Provider-scoped entries take precedence, and plain model-id entries still work as the backward-compatible fallback.

Memories

The memory pipeline is enabled by default:

• lingyun.features.memories = true
• lingyun.memories.autoRecall = true

LingYun can extract transcript-backed memory from persisted sessions, auto-recall relevant context before top-level turns, and expose memory artifacts through the get_memory tool.

Useful commands:

• LingYun: Update Memories
• LingYun: Drop Memories

Skills

LingYun supports reusable SKILL.md files.

Default search paths include:

• Workspace-relative: .lingyun/skills, .claude/skills, .opencode/skill, .opencode/skills
• Home-directory paths: ~/.config/lingyun/skills, ~/.agent/skills, ~/.agents/skills, ~/.codex/skills, ~/.claude/skills

Home-directory paths are ignored unless lingyun.security.allowExternalPaths = true.

Mention $skill-name in a prompt to auto-inject that skill for the current turn.

Relevant settings:

{
  "lingyun.skills.enabled": true,
  "lingyun.skills.maxPromptSkills": 50,
  "lingyun.skills.maxInjectSkills": 5,
  "lingyun.skills.maxInjectChars": 20000
}

Subagents

LingYun can delegate work through the task tool. It also supports an optional automatic explore prepass before each user turn:

{
  "lingyun.subagents.explorePrepass.enabled": true,
  "lingyun.subagents.explorePrepass.maxChars": 8000,
  "lingyun.subagents.model": ""
}

Task-subagent output injected back into the main session is capped by lingyun.subagents.task.maxOutputChars.

Commands

Main commands:

• LingYun: Start Task
• LingYun: Open Agent
• LingYun: Open Office
• LingYun: Abort
• LingYun: Clear History
• LingYun: Show Logs

Session and context commands:

• LingYun: Clear Saved Sessions
• LingYun: Compact Session
• LingYun: Update Memories
• LingYun: Drop Memories

History commands:

• LingYun: Undo
• LingYun: Redo

Tooling commands:

• LingYun: List Tools
• LingYun: Create Tools Config
• LingYun: Run Tool

Default keybinding:

• Cmd+Shift+. on macOS
• Ctrl+Shift+. elsewhere

Troubleshooting

• OpenAI-compatible models do not load: make sure lingyun.openaiCompatible.baseURL includes /v1 and the server answers GET /v1/models.
• Responses cut off: increase lingyun.openaiCompatible.maxTokens or configure lingyun.modelLimits.
• Compaction is too aggressive or not aggressive enough: set lingyun.modelLimits for the exact model, or use a provider-scoped key like codexSubscription:gpt-5.4.
• Need more visibility: use LingYun: Show Logs and enable lingyun.debug.details for fuller logs with visible paths/URLs/hosts, or turn on lingyun.debug.llm, lingyun.debug.tools, or lingyun.debug.plugins individually.

Advanced

• Workspace tools: .vscode/agent-tools.json and .vscode/agent-tools/*.json
• Workspace plugins: .lingyun/plugin/*.js
• Skills docs: ../../docs/SKILLS.md
• Plugin docs: ../../docs/PLUGINS.md
• Development and architecture notes: ../../AGENTS.md

License

Apache License 2.0