rescue

$copilot:rescue <task> — delegate an arbitrary task to Copilot

Hand a free-form task to Copilot CLI in non-interactive mode with all tool permissions granted. JSONL transcript is captured for later inspection via result.

Invocation

HOST=<host> bash "<plugin-root>/scripts/copilot-exec.sh" rescue "<task description>" [flags]

Set <host> to codex or claude. Replace <plugin-root> with the installed plugin root; in a local checkout, that is the repo root.

Flags:

• --model <id> — override Copilot's default. Useful when the task fits a specific provider (Gemini for long-context, Opus for reasoning-heavy, GPT-5.x for code-gen).
• --background — run async; print job id immediately
• --job-id <id> — caller-supplied id
• --continue — resume Copilot's most recent session
• --resume <session-id> — resume a specific session

Under the hood

copilot -p "$TASK" --allow-all-tools --output-format=json -s

--allow-all-tools is required for non-interactive mode (otherwise Copilot will hang waiting for approvals). --output-format=json produces JSONL — one JSON event per line, which result reads back. -s silences the stats banner.

Foreground vs background

• Foreground: stream output to the user as it arrives. Use for short tasks (single-file edits, focused fixes).
• Background: print the job id and return immediately. Use for anything that touches more than a few files or might run more than ~60s. Confirm the job started and point the user at status / result.

If you're unsure of scope, ask the user before launching foreground.

Resume semantics

• No --resume / --continue flag → Copilot starts a fresh session.
• --continue → most recent session for this user.
• --resume <id> → specific session by id, task id, or name (case-insensitive, exact match).

If the user trails off mid-rescue and comes back later, offer to --continue rather than starting fresh.

Output handling

For foreground runs: summarize what Copilot did, what files it touched, and what's left. Surface the session id so the user can reopen interactively via copilot --resume <id> if they want to dig in.

For background runs: confirm the job started, show the job id, and remind the user that status and result are how they retrieve it.

Failure modes

• Not authenticated → trigger setup.
• Sandbox/trust error → tell the user to run copilot interactively in this directory once.
• Quota exhausted → state this and stop. Do not silently retry.
• MCP-dependent task → Copilot CLI's -p mode does not load MCP servers (github/copilot-cli#633). Surface this and suggest an interactive Copilot session instead.