Copilot Acp

You are the /copilot-acp skill. You delegate to GitHub Copilot CLI via the Agent Client Protocol (copilot --acp), using the official agent-client-protocol Python SDK (pip install agent-client-protocol). Every message is a Pydantic model — no hand-rolled JSON-RPC.

What ACP gives you over --output-format json:

• Native session lifecycle. initialize → session/new → session/prompt. Real session IDs that survive across the connection.
• First-class permission requests. When Copilot wants to run a tool, the server sends session/request_permission. The runner can route this to the user, auto-allow, or veto. Today the runner auto-allows for non-interactive operation.
• Per-session config. Model, reasoning effort, allow-all are session options, not CLI flags.
• Session resume. session/load is supported by the server (we capture sessionId in .session).

If you need maximum control over per-tool execution, prefer /copilot-acp. For sheer simplicity (just stream events from the CLI), use /copilot-cli. For the official wrapper, use /copilot-sdk.

Invocation patterns

On every invocation

1. Bootstrap. bash .agents/skills/copilot-acp/run.sh --bootstrap — delegates to /copilot-cli/bootstrap.sh.
2. Sync. bash .agents/skills/copilot-acp/run.sh [--model X] [--effort L] "<task>".
3. Check / cancel. bash .agents/skills/copilot-cli/check.sh / cancel.sh — the shared run-files protocol means these work uniformly.

Run files

Identical layout to /copilot-cli, run-id prefixed with -acp-:

• .jsonl — raw bidirectional JSON-RPC traffic. Every envelope wrapped in {"_send": ...} or {"_recv": ...} so client + server messages stay disambiguated.
• .md — rendered transcript (agent message chunks, tool calls, thoughts).
• .session — Copilot session id.
• .status, .task, .ask, .answer, .pid — same as /copilot-cli.

Defaults

• Model: gpt-5.6-sol (override with --model or COPILOT_MODEL; claude-opus-5 remains fully supported).
• Reasoning effort: high. Probing the live server shows it accepts none/low/medium/high/xhigh/max; high is confirmed applied via config read-back.
• Context tier: not supported. The server's only config options are mode, model, reasoning_effort, allow_all; context_tier is rejected as an unknown option. Use cli or sdk for the 1M window.
• allow_all: on (session option, applied at session/new).

What's not built yet

• bg.sh (background): the ACP runner runs synchronously today. A background variant is straightforward (same setsid pattern as /copilot-cli/bg.sh) but not shipped in the first pass.
• --multi cross-model comparison: deliberately left to /copilot-cli/multi.sh for now.
• Per-tool human approval: the runner auto-allows session/request_permission. To route to the user, add a .permission-request file write + wait-for-answer loop similar to the .ask/.answer flow.