Delegate

Use when the work is self-contained and iteration isn't needed: codebase exploration, bulk research across many files, independent subtasks that can run in parallel, one-shot refactors with a clear spec.

Do not use for iterative work needing back-and-forth, or for tasks already mid-flight in the main context.

The subprocess inherits OPENAI_API_KEY, MODEL, OPENAI_BASE_URL and has no prior history. The prompt must be fully self-contained:

• Absolute paths to every relevant file or directory
• The goal stated in one line
• Constraints (what not to touch, style, scope limits)
• Expected output format (e.g. "print a bulleted list of file:line references", "write results to /tmp/mi-.out and print its path")

Spawning subagents: use node "$MI_PATH" — the harness sets MI_PATH automatically:

node "$MI_PATH" -p '<prompt>'

Sequential (one task, wait for result):

node "$MI_PATH" -p 'Read /abs/path/foo.py and list every function that touches the database. Print one per line as file:line name.'

stdout becomes the tool result. Use timeout= on the bash call if the task could hang.

Parallel (multiple independent tasks):

node "$MI_PATH" -p '<prompt A>'   # with bg=truthy -> pid:A log:/tmp/mi-A.log
node "$MI_PATH" -p '<prompt B>'   # with bg=truthy -> pid:B log:/tmp/mi-B.log

Collect each pid and log. Background children are detached (the harness calls unref) so wait will not find them — poll with kill -0 <pid> 2>/dev/null instead (exit 0 = still running, exit 1 = finished).

Reading long logs: subagent logs grow large. Do NOT cat the full log blindly — instead:

• tail -n 50 /tmp/mi-A.log to see the final output and whether the agent concluded
• grep -n "RESULT\|ERROR\|DONE\|Traceback" /tmp/mi-A.log to surface key lines quickly
• wc -l /tmp/mi-A.log to know total size before committing to a full read
• If the agent wrote a compact result file (e.g. /tmp/mi-A.out), read that instead — it's why you ask for one

Always prefer telling each subprocess to write a compact result file under /tmp/mi-* so you don't have to parse transcript noise.

Keep prompts short and specific. A vague delegation wastes a whole subprocess.