Summary
I started a build for one feature in a repository. A few minutes later, as the first build was still going, I started another build for the same repository. I'm not sure if that's actually relevant to the bug, but thought it might be worth mentioning.
After this, agents in both builds started getting input describing PRDs and acceptance criteria for the feature in the other build. Eventually, the Product Manager and Architect agents in each build started failing, and finally both builds failed completely.
Steps to Reproduce
- Start a build in a repo for one feature.
- Start a second build in that repo for an unrelated feature.
Expected Behavior
Agents in each build should be scoped to that build only, allowing multiple siloed builds to be run simultaneously.
Actual Behavior
Agents from each build were getting input from agents in the other build, eventually causing both builds to fail.
Environment
OS: Ubuntu 24.04
Deployed from Docker Compose file shipped in the repo.
Provider: locally-hosted vLLM instances.
Logs or Trace
Summary
I started a build for one feature in a repository. A few minutes later, as the first build was still going, I started another build for the same repository. I'm not sure if that's actually relevant to the bug, but thought it might be worth mentioning.
After this, agents in both builds started getting input describing PRDs and acceptance criteria for the feature in the other build. Eventually, the Product Manager and Architect agents in each build started failing, and finally both builds failed completely.
Steps to Reproduce
Expected Behavior
Agents in each build should be scoped to that build only, allowing multiple siloed builds to be run simultaneously.
Actual Behavior
Agents from each build were getting input from agents in the other build, eventually causing both builds to fail.
Environment
OS: Ubuntu 24.04
Deployed from Docker Compose file shipped in the repo.
Provider: locally-hosted vLLM instances.
Logs or Trace