copilot-sdk

name	copilot-sdk
description	Delegate to Copilot via the official Go SDK (github.com/github/copilot-sdk/go). Source ships under scripts/; first invocation builds bin/runner. Sibling of /copilot-cli (JSONL) and /copilot-acp (ACP).
allowed-tools	Bash, Read, Write, Glob

You are the /copilot-sdk skill. You delegate to GitHub Copilot using the official Go SDK (github.com/github/copilot-sdk/go).

Why this skill exists alongside /copilot-cli and /copilot-acp:

• The SDK is the officially supported integration surface. It shields you from CLI flag churn.
• The SDK speaks JSON-RPC to the bundled CLI under the hood — same wire as /copilot-acp, but with typed events instead of raw envelopes.
• This skill is a comparison harness: run the same task through /copilot-cli, /copilot-acp, and /copilot-sdk to see which integration style works best for your use case.

Source distribution

The Go source lives in scripts/:

.agents/skills/copilot-sdk/
├── SKILL.md
├── run.sh           # entrypoint
├── build.sh         # builds scripts/main.go → bin/runner
├── scripts/
│   ├── go.mod
│   └── main.go      # imports github.com/github/copilot-sdk/go
└── bin/runner       # built on first use (gitignored)

bin/runner is built on first invocation via go build. Subsequent runs reuse the built binary; the build is skipped when binary mtime is newer than source.

Invocation patterns

User input	What to do
/copilot-sdk --bootstrap	Run shared copilot bootstrap + build the Go runner.
/copilot-sdk <task>	Sync SDK run.
/copilot-sdk --model claude-opus-4.7 <task>	Override model.
/copilot-sdk-check <run-id>	Use shared check.sh (same as /copilot-cli-check).
/copilot-sdk-cancel <run-id>	Use shared cancel.sh.

On every invocation

1. Bootstrap. bash .agents/skills/copilot-sdk/run.sh --bootstrap — runs the shared copilot installer/auth check, then builds the Go runner.
2. Sync. bash .agents/skills/copilot-sdk/run.sh [--model X] "<task>". If the binary isn't built yet, run.sh auto-builds.
3. Check / cancel. Use bash .agents/skills/copilot-cli/check.sh and cancel.sh — the shared run-files protocol means they work uniformly.

Run files

Identical layout to the sibling skills, run-id prefixed with -sdk-:

• .jsonl — raw event dump from SDK callbacks (the SDK abstracts away the wire, so the JSONL is best-effort serialization of typed event payloads).
• .md — rendered transcript.
• .session — Copilot session id.
• .status, .task — shared protocol.

Defaults

• Model: claude-opus-4.7 (env: COPILOT_MODEL, flag: --model).
• Permission handler: PermissionHandler.ApproveAll (auto-approves tool requests). Swap this in scripts/main.go if you want manual control.
• Reasoning effort: xhigh (env: COPILOT_REASONING_EFFORT, flag: --effort). The SDK passes this through SessionConfig.ReasoningEffort; the upstream server caps per-model and may reject unsupported values.
• Timeout: 1 hour wall-clock cap on SessionIdle (env: COPILOT_TIMEOUT_S, flag: --timeout; set 0 to disable).

Status: public preview

The copilot-sdk is GitHub's official public preview as of this writing. Expect API churn. The skill builds from source, so when the SDK updates we'll catch breakage at build time, not runtime.