A nightly dependency upgrader that remembers
CI and automationTeam running an agent every night to bump packages
Every night the agent tried to upgrade the same package that breaks the build, failed, and rolled back. That was 30k tokens a night on one known bad upgrade.
How do I stop my nightly upgrade agent retrying a package bump that always breaks the build?
Have the agent record a failed upgrade with ArcRoom's remember_failure and the reason. The next night, find_task returns that failure before the agent tries, and it skips the package until the failure expires or someone clears it. That ends paying for the same known bad upgrade every night. Only failures that were recorded are skipped.
Also asked as: Our upgrade agent wastes tokens every night on the same broken package. How do we stop it? · How do I make a scheduled agent remember failed upgrades? · Can an AI agent skip a package upgrade it already knows fails?
When an upgrade fails, the agent records it with remember_failure and the reason.
- Reused answer
- Saved decision
- Expiry rule
Tried before and failed (3 nights): next@15.2 → build error in app/router.
Step by step
- When an upgrade fails, the agent records it with remember_failure and the reason.
- The next night, find_task returns the failure before the agent tries.
- The agent skips that package until the failure expires or someone clears it.
What the agent sees
Tried before and failed (3 nights): next@15.2 → build error in app/router.
Skip until a new version or a team note.Compared with today's tools
A skip list in the repo, or Renovate/Dependabot ignore rules
After a bad upgrade, someone adds the package version to an ignore list in the bot config or a skip file the agent reads.
- Setup
- 15 min
- Ongoing
- Edit the list after each bad upgrade, ~5 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)
- Ongoing
- None
An ignore rule is explicit and reviewable, which is often better. ArcRoom records the failure and its reason without anyone editing config.
Where it stops
Only failures that were recorded are skipped.
Try it in your project
Works with the agents your team already uses, and any agent over MCP or the HTTP API.