GitHub Copilot

GitHub Copilot quickstart

Install the crew, verify the payload, then run a safe smoke test.

Shipmates emits Copilot custom agents under .github/agents/ and Agent Skills under the shared .agents/skills/ tree.

Install in a sandbox

Keep this first run outside a production repository.

shipmates install --harness github-copilot --dir /path/to/sandbox --with-tools none

The install writes twelve custom agents at /path/to/sandbox/.github/agents/*.agent.md and thirteen skills at /path/to/sandbox/.agents/skills/*/SKILL.md. The .agent.md suffix is required by Copilot discovery.

Restart or start a new Copilot session after installing. The current adapter emits the documented Copilot frontmatter and maps semantic capabilities to read, search, edit, execute, web, and agent. These are payload tool names, not necessarily names accepted by the CLI approval flags.

Verify payload fidelity

This is deterministic validation, not a live Copilot claim.

cargo run -- check --target github-copilot
ls /path/to/sandbox/.github/agents
ls /path/to/sandbox/.agents/skills/ship-issue

CI compares every file produced by build --target github-copilot with the checked-in golden digest. The fixture covers both dot-directories and rejects missing or unexpected files.

Passing this check proves adapter output and registration shape. It does not prove that a running Copilot session discovers agents, loads skills, passes arguments, or completes a workflow.

Optional non-destructive smoke test

Runtime smoke is unavailable in current validation; keep this recipe manual and disposable.

copilot --agent security-engineer \
  --prompt "Run the installed /shipmates-harden order in report mode against this sandbox. Read AGENTS.md, report findings, and do not edit files or open a PR."

If runtime prerequisites become available, expected result is Copilot resolving security-engineer.agent.md, selecting the installed read-only harden skill, reporting findings, and changing no files. No such result is recorded here; this command is a recipe, not evidence of runtime support.

Custom-agent tools frontmatter describes the model's tool surface. Headless CLI approval is a separate layer and may use different patterns; consult copilot help permissions for the installed CLI. Do not translate read/search blindly into --allow-tool, and do not use --allow-all. Unavailable CLI, authentication, approval, or model selection means not run.

/ship-issue Tier B boundary

The adapter provides the dispatch surface; it does not provide the external runner.

Validated statically: the shared ship-issue skill is emitted, and the custom agents it names are emitted with Copilot-native filenames and least-privilege tools.
Not validated here: a real issue-to-PR run, scripted acceptance board, CI polling, or bounded fix loop through copilot --agent.
Reason: those behaviours belong to the external orchestrator described by issue #13; the current Copilot adapter only compiles and installs payload files.

Model note. effort: false means Copilot has no documented install-time per-agent effort field; the adapter emits no model: or effort key. That is separate from runtime model selection: the orchestrator chooses a model at spawn where supported, otherwise the session model is inherited.

Therefore GitHub Copilot remains builds, not runtime-verified. This page intentionally does not call the target tested, Tier B-complete, or flagship-complete.

More docs

All docs