Bot-driven sessions that don't redo the last one's work
Long-running agentsDeveloper who runs claude -p from a chat bot on a home PC
I wrote a small Telegram bot that passes my messages to claude -p in a project folder on my desktop. On the train I send 'add the export button', and an hour later 'now fix the test you broke'. Each message is a new headless session, so before ArcRoom the second one had no idea what the first one did.
How do I stop each claude -p call from my chat bot redoing what the last one did?
Set up ArcRoom by running npx arcroom init --agent claude once in the folder your bot uses. Each claude -p call then reads the playbook first: the steps the last session recorded and the last-run note, with what it asked, edited and whether it stopped mid-way. Approve the done-checks once at your desk with arcroom checks approve <code>, so every bot session must pass them before it stops.
Also asked as: Each headless Claude Code session from my Telegram bot starts with no memory. How do I fix it? · How do I pass context between separate claude -p runs? · Can Cursor or Codex pick up what my bot's Claude Code session did?
Run npx arcroom init --agent claude once in the folder the bot uses.
- Last-run note
- Done-checks
Last session here (Claude Code, 52 min ago):
Step by step
- Run npx arcroom init --agent claude once in the folder the bot uses.
- Approve the done-checks once at the desk with arcroom checks approve <code>; every bot session must pass them before it stops.
- Each claude -p call gets the last-run note: what the last session asked, edited, and whether it stopped mid-way.
- Back at the desk, Cursor or Codex in the same folder gets the same note.
What the agent sees
Last session here (Claude Code, 52 min ago):
asked: add the export button to /reports
edited: src/reports/Export.tsx, src/reports/api.ts
stopped mid-way: pnpm test failed (reports.spec.ts)Compared with today's tools
Claude Code auto memory + --continue
The bot passes --continue so each message resumes the last session, and Claude Code keeps its own memory notes.
- Setup
- 5 min
- Ongoing
- None, until you switch agent
ArcRoom
Every agent reads the team playbook before it works and writes to it when it finishes. Nothing to update by hand.
- Setup
- 2 min (npx arcroom init)
- Ongoing
- None
Resuming one session works well inside Claude Code; it doesn't reach Codex or Cursor at the desk, and it doesn't run done-checks before stopping.
Where it stops
ArcRoom sees only the Claude Code sessions, not the bot or your chat history. Prompts leave the machine (first 1,000 characters, redacted); check with arcroom sent.
Try it in your project
Works with the agents your team already uses, and any agent over MCP or the HTTP API.
More in long-running agents
- A three-day migration where no part is done twiceDeveloper leaving Claude Code on a long migration over a weekend
- Three scheduled agents, nothing done twiceSolo founder scheduling headless agents on a home machine
- Coding agents in containers, one per issueTeam running Claude Code headless in containers on its own server