| name | autopilot |
| description | Set up autopilot mode for a channelhub project. Use when the user wants to configure how their project answers questions on their behalf — specifically which preferences the proxy should honor. |
Autopilot Mode
When a channelhub session is in autopilot mode, the daemon fires /btw inside
the session on every user-facing question. The proxy answers using the
session's own conversation context plus the preferences in autopilot.md.
Set up autopilot.md
Create autopilot.md in your project root. Use markdown freeform — the proxy
will read it verbatim. A good starter:
# Preferences for autopilot
- Prefer Bun over Node
- Prefer minimal dependencies; avoid adding new ones unless unavoidable
- Always TDD: test first, then implement
- Prefer explicit over clever
- For UI decisions, pick the simpler / more-accessible option
- Never add analytics / tracking without asking
Enable autopilot
- Telegram:
/autopilot <session-name> on
- Web: click the "Autopilot" toggle in the session row
- CLI:
bun run src/cli.ts autopilot <session-name> on
What gets escalated
The proxy will skip the /btw and ping you directly when:
- Your question contains a risk keyword (
delete, force push, production, billing, api key, etc.)
- The proxy answers
ESCALATE: <reason> (for irreversible or out-of-scope choices)
- The /btw overlay fails to parse or times out
Veto window
By default, autopilot shows the draft answer for 30 seconds before auto-sending.
You can Send it as-is, Edit the text, or Cancel to skip this one and
handle it manually. If you take too long, the answer is sent anyway — this keeps
the workflow moving without needing you to babysit every interaction.