copilot-sdk（GitHub Copilot SDK 测试）

GitHub Copilot SDK

Overview

The GitHub Copilot SDK exposes the same Copilot CLI agent runtime over JSON-RPC, so apps can drive Copilot programmatically instead of building their own orchestration layer.

Status: Public preview SDKs: Node.js/TypeScript, Python, Go,.NET, Java Architecture: Application -> SDK client -> JSON-RPC -> Copilot CLI

How to use this skill

When helping with the Copilot SDK:

1. Prefer the official docs index and the language-specific README over memory.
2. Treat the top-level SDK README plus docs/ as the source of truth for shared behavior.
3. Call out preview status when stability or breaking changes matter.
4. Avoid hardcoding model lists when runtime discovery via listModels() is available.
5. Watch for stale guidance around permissions, lifecycle methods, and event names.

Current source of truth

Core SDK docs

• GitHub Copilot SDK repository
• Documentation index
• Getting started guide
• Setup guides
• Local CLI setup
• Bundled CLI setup
• Backend services setup
• Scaling and multi-tenancy
• Azure Managed Identity with BYOK
• Authentication
• BYOK
• Features index
• Image input
• Steering and queueing
• OpenTelemetry instrumentation
• Troubleshooting
• SDK/CLI compatibility

Language-specific docs

• Node.js / TypeScript README
• Python README
• Go README
• .NET README
• Java SDK repository

Copilot CLI and GitHub Docs

• About GitHub Copilot CLI
• Using GitHub Copilot CLI
• Custom agents configuration reference
• Enhancing agent mode with MCP
• Supported models

Recipes and examples

• Copilot SDK cookbook
• Node.js samples
• Go samples
• .NET samples

High-value facts

Authentication and prerequisites

• A GitHub Copilot subscription is required for normal SDK use.
• BYOK is supported and does not require GitHub Copilot authentication.
• Node.js, Python, and.NET bundle the Copilot CLI automatically.
• Go can use an installed CLI or embed/bundle one with the go tool bundler workflow.
• Java currently lives in github/copilot-sdk-java and expects the CLI to be installed separately.
• Azure Managed Identity / Entra auth is supported as a documented BYOK pattern by passing short-lived bearer tokens from DefaultAzureCredential.

Permissions

• The SDK uses a deny-by-default permission model.
• In practice, create/resume flows should provide an explicit permission handler such as:

- TypeScript: approveAll - Python: PermissionHandler.approve_all - Go: copilot.PermissionHandler.ApproveAll - .NET: PermissionHandler.ApproveAll - Java: PermissionHandler.APPROVE_ALL

Session lifecycle

• Preferred cleanup method: disconnect()
• Deprecated cleanup method: destroy()
• To resume sessions reliably, provide your own sessionId when creating them.
• BYOK provider configuration must be provided again when resuming because keys are not persisted.

Transport and deployment

• Default transport is stdio with an SDK-managed CLI process.
• You can connect to an external headless CLI server via cliUrl.
• Current external server docs use:

copilot --headless --port 4321

Models

• Do not hardcode model support unless the user specifically needs a fixed list.
• Prefer client.listModels() and the official supported-models page.
• reasoningEffort exists for models that support it.

Installation

Setup and deployment choices

Pick the setup that matches the application shape:

• Local CLI - simplest path for personal tools and development.
• Bundled CLI - ship a CLI binary with your app for desktop/distributable tooling.
• Backend services - run the CLI in headless mode and connect with cliUrl.
• Scaling and multi-tenancy - shared CLI vs CLI-per-user, shared storage, and session locking.
• Azure Managed Identity - use BYOK with short-lived bearer tokens instead of static API keys when Azure auth is the real requirement.

Quick start pattern

Use the same mental model in every language:

1. Create/start the client.
2. Create a session with a permission handler.
3. Register event handlers before send() if you need streaming or progress.
4. Send with send() or sendAndWait().
5. Wait for session.idle or the returned final message.
6. disconnect() the session and stop/dispose the client.

TypeScript example

import { CopilotClient, approveAll } from "@github/copilot-sdk";

const client = new CopilotClient();
await client.start();

const session = await client.createSession({
  model: "gpt-5",
  streaming: true,
  onPermissionRequest: approveAll,
});

session.on("assistant.message_delta", (event) => {
  process.stdout.write(event.data.deltaContent ?? "");
});

await session.sendAndWait({ prompt: "What is 2+2?" });

await session.disconnect();
await client.stop();

Core capabilities to remember

Client and session APIs

Common operations across SDKs:

• Client lifecycle: start(), stop(), forceStop()
• Session lifecycle: createSession(), resumeSession(), disconnect()
• Messaging: send(), sendAndWait(), abort(), getMessages()
• Discovery: listModels(), listSessions(), getStatus() / ping()

Events and streaming

• Final assistant output arrives in assistant.message.
• Streaming text arrives in assistant.message_delta.
• session.idle is the reliable "turn complete" signal.
• The event system now includes reasoning, tool progress, permission, elicitation, sub-agent, and skill events.

See references/event-system.md.

Custom tools

• Node uses defineTool(...) with Zod or raw JSON Schema.
• Python uses @define_tool with Pydantic models.
• Go prefers DefineTool(...).
• .NET uses AIFunctionFactory.Create(...).
• Overriding built-ins always requires explicit opt-in:

- TypeScript: overridesBuiltInTool: true - Python: overrides_built_in_tool=True - Go: OverridesBuiltInTool = true - .NET: AdditionalProperties["is_override"] = true

• Custom tools can also opt into skipPermission.

Custom agents, MCP, hooks, and skills

• customAgents lets you define sub-agents per session.
• mcpServers attaches local or remote MCP servers.
• Hooks provide control points such as onPreToolUse, onPostToolUse, onUserPromptSubmitted, and lifecycle/error hooks.
• Skills are loaded with skillDirectories; disable selectively with disabledSkills.

See references/cli-agents-mcp.md.

Attachments, commands, and interaction

• Sessions can send file, directory, and image attachments.
• Image input supports both file and blob attachments, and vision should be checked through model capabilities.
• In-flight messaging supports mode: "immediate" for steering and mode: "enqueue" for queueing.
• The SDK can register custom slash commands.
• Apps can answer user questions with onUserInputRequest.
• Rich UI prompts are available through elicitation handlers and session.ui when the connected client supports them.

Telemetry and observability

• The SDK supports OpenTelemetry configuration through TelemetryConfig.
• Trace context propagation is built in, with Node using an explicit onGetTraceContext callback for outbound propagation.

Persistence and long-running work

• Use a stable sessionId for resumable sessions.
• Use infiniteSessions for long-running workflows that may need compaction.
• Session state is stored under ~/.copilot/session-state/ unless configuration overrides it.

SDK vs. CLI-only features

• The SDK exposes programmatic surfaces for sessions, models, plans, mode switching, workspace files, custom agents, hooks, MCP, skills, and telemetry.
• Many terminal UX features remain CLI-only, such as most slash-command workflows, interactive pickers, and export/share commands.
• When translating a CLI workflow into app code, check the compatibility guide before assuming a slash command has an SDK equivalent.

Language conventions

Common gotchas

• The SDK is public preview, so older examples drift quickly.
• Hardcoded model tables get stale; prefer runtime discovery.
• destroy() still appears in older examples but disconnect() is the current method.
• A missing permission handler causes confusion fast; treat it as required for real sessions.
• assistant.message and assistant.message_delta use event.data.*, not top-level event.content.
• Streaming/event subscriptions should be attached before send().
• Session resumption without a caller-provided sessionId is awkward to operationalize.

Local reference files in this skill

• references/working-examples.md - current starter examples, including tools and resume patterns
• references/event-system.md - event names, lifecycle, and language access patterns
• references/cli-agents-mcp.md - custom agents, skills, MCP, headless CLI, and config locations
• references/troubleshooting.md - common failures, debug logging, auth, permissions, and transport issues