Copilot

Copilot Delegation

Run copilot yourself with Bash. Do not spawn a subagent to do it: wrapper agents tend to either duplicate Copilot's reading at full token cost or skip copilot entirely and do the task themselves. You compose the prompt, Copilot does the work, you verify and relay.

Invocation

A Copilot run routinely takes 10-30 minutes, longer than the foreground Bash limit, so always run it with run_in_background: true and redirect to a log. Never run it in the foreground.

copilot -C <working directory> --model gpt-5.6-sol --effort high --allow-all --no-color --log-level none -p "$(cat <<'TASK'
<task prompt>
TASK
)" 2>&1 | sed -u -E '/^(Changes|AI Credits|Tokens)[[:space:]]/d' | sed -u '/^$/d' > /tmp/copilot-<task>.log

When it finishes, read the summary first (awk '/SUMMARY/,0' /tmp/copilot-<task>.log); grep the full log only when the summary reports a failure, pulling just the error excerpt.

The run is in the background and notifies you when it finishes, so don't poll. For a mid-run peek, run a bare tail -n 20 /tmp/copilot-<task>.log (no sleep prefix; the harness blocks sleep N; tail). A growing log means it is working; a flat one only hints it may be stuck, since Copilot buffers its own stdout. If it goes silent, wrap the tool in stdbuf -oL or a PTY (script -q /dev/null copilot ...).

• --model gpt-5.5 --effort high is the default.
• --allow-all is REQUIRED for non-interactive (-p) mode; without it Copilot refuses to run tools (it enables all tools, paths, and URLs).
• For review/research/analysis, state "Do NOT modify any files" in the prompt (Copilot can edit files, so this is the read-only guard).
• sed strips the footer block (Changes …, AI Credits …, Tokens …); -u runs it unbuffered so each line flushes to the log instead of block-buffering.

Composing the task prompt

Copilot inherits nothing from this session, so the prompt must be self-contained:

• The working directory is set with -C; restating it in the prompt is fine but not required.
• Reference on-disk files by path (Copilot reads them itself); don't paste contents you'd be making it re-read. Paste in full anything NOT on disk: plan text, review findings, decisions, diffs you were handed.
• State constraints (scope, style, what not to touch) and dictate the output format so the relay is mechanical: reviews as path:line + severity (high/medium/low) + one-line rationale + verified/suspected; edits as file:lines + one-line summary each; searches as path:line + role.
• End with: "SUMMARY: files changed (path: one line each), tests run + result, anything left undone."
• For reviews, tell it to be critical and flag problems, not reassure.

Parallel tasks

Launch each task as its own background Bash call (each redirects to its own /tmp/copilot-<task>.log), end your turn, then collect summaries as they exit, verify, and re-dispatch failures with the error excerpt. Decomposition and verification stay with you, no intermediate agents. Read-only tasks always parallelize safely; write tasks conflict on the same checkout, so run them sequentially or isolate each in its own worktree/repo and merge after verification.

After Copilot finishes

• Verify edits cheaply: git diff --stat, full diffs only where it looks suspicious, plus one smoke run if the project has an obvious one. Never assume the changes are correct, but don't re-review the whole product (that recreates the duplication this skill avoids).
• Relay reviews/findings essentially verbatim; don't soften or reorder severities. Report any files modified.

Auth and known noise

• Copilot authenticates via GitHub. On "Authentication failed" the user must re-auth: run copilot interactively and use /login, or set a valid COPILOT_GITHUB_TOKEN / GH_TOKEN (needs "Copilot Requests" permission). You can't fix auth non-interactively, so surface it and stop.
• The footer (Changes, AI Credits, Tokens) is benign accounting output; the sed filter drops it.