Laptop work not redone by the CI agent
CI and automationSmall team with Claude Code running headless in GitHub Actions
We were moving the API from Express to Hono. A developer's agent did the routes on her laptop and stopped at 7 p.m. with the tests still red. Overnight, a CI job ran Claude Code headless to finish what was left. Without a team playbook the CI agent would have read the whole repo again and guessed what 'left' meant.
How do I stop a Claude Code job in GitHub Actions redoing work I already did on my laptop?
Add ArcRoom to the CI job: run npx arcroom init with ARCROOM_API_KEY from secrets, then claude -p with the task, then npx arcroom flush. The CI agent reads the playbook first: the laptop session's recorded steps, saved answers and team decisions, so it finishes what is left instead of starting over. Set ARCROOM_APPROVE_CHECKS so the done-checks must pass before it stops. Steps the laptop agent never saved are not passed on.
Also asked as: Our CI agent rereads the whole repo and guesses what is left. How do we hand it context? · How do I run Claude Code headless in GitHub Actions without losing what it learned? · Can a CI agent's work show up in my next local session?
The CI job runs npx arcroom init with ARCROOM_API_KEY from the repo's secrets, then runs claude -p "continue the Hono migration".
- Done-checks
- Decision
- Flush in CI
# .github/workflows/agent.yml (steps)
Step by step
- The CI job runs npx arcroom init with ARCROOM_API_KEY from the repo's secrets, then runs claude -p "continue the Hono migration".
- The CI agent gets the same project memory: the laptop session's recorded steps, saved answers and team decisions.
- The done-checks in .arcroom.json (typecheck and tests) must pass before it may stop. There's no terminal to approve them, so the job sets ARCROOM_APPROVE_CHECKS=<code>.
- The last step of the job is arcroom flush, so everything the CI agent learned reaches the playbook before the runner shuts down.
- In the morning, the developer's next session gets what the CI agent did.
What the agent sees
# .github/workflows/agent.yml (steps)
- run: npx arcroom init
env: { ARCROOM_API_KEY: ${{ secrets.ARCROOM_API_KEY }} }
- run: claude -p "continue the Hono migration"
env: { ARCROOM_APPROVE_CHECKS: ${{ secrets.ARCROOM_CHECKS_CODE }} }
- run: npx arcroom flushCompared with today's tools
A handover file in the repo plus a long prompt in the workflow
Before leaving, the developer writes what is left into HANDOFF.md and commits it. The workflow prompt tells claude -p to read that file first.
- Setup
- 15 min
- Ongoing
- Write a handover note before every overnight run, ~10 min
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 in the job)
- Ongoing
- None
The note holds only what the developer remembered to write. ArcRoom passes saved steps and answers without anyone writing them, though steps never saved don't pass either.
Where it stops
The handover comes through saved steps, answers and the project's last-run note. Steps the laptop agent never saved are not passed on.
Try it in your project
Works with the agents your team already uses, and any agent over MCP or the HTTP API.