Skip to main content
Before running tasks, configure each project’s Project settings → Integration, directly or with the Coordinator. Use the repository’s existing scripts and a committed baseline. Enter one shell command per line for setup and validation. Choose commands appropriate to your repository; the service host must already provide the language tools and credentials.

​
Validate before running workers

  1. Commit setup scripts, manifests and relevant repository guidance to the destination branch. Commit or ignore adoption metadata under your repository’s rules; untracked files can block delivery.
  2. Save Integration settings and choose Validate setup. Flowfield checks the saved branch in a fresh worktree, runs setup and validation and reports readiness. It starts no model and enables no queue.
  3. Choose a worker harness and model in Workspace settings → Workers, with supported effort/access controls. Start with one worker slot; increase capacity when independent tasks can run safely together.
  4. Have the Coordinator prepare a task, move it to Up next and choose Run queue.
Setup and checks run in the worktree directory and must leave source files unchanged. Keep dependency outputs ignored; use FLOWFIELD_RUNTIME_DIR for temporary files and FLOWFIELD_WORKSPACE for the checkout path. Checks must finish and return a failing exit code on failure. A development server belongs in Run command, not validation. Read failed validation output, fix the scripts or configuration and validate again. Revalidate after changing setup, checks or the destination branch. Setup and validation commands have configured timeouts. Follow an active worker and use Stop worker from the task conversation.

​
Choose workers

Workspace defaults apply across member projects. Task Agent settings can override them; each attempt retains the harness, model and host registration it started with. Changes affect subsequent attempts, and unavailable choices fail explicitly rather than substituting a model. The Coordinator’s choice is independent. Supported Fast mode can be enabled per task for subsequent attempts; it starts off and has no workspace-wide default. Agent setup covers native accounts, discovery and permissions; documents covers the sources selected for workers. Workers use separate Git worktrees and scoped Flowfield connections. Git isolation separates repository edits, not the whole host: use separate ports, databases and other mutable resources for concurrent tasks, or serialize them. Native access modes govern agent tools; configured setup and validation commands run directly on the service host.

​
Try a result

Try result prepares an independent copy of the selected candidate and supplies a launcher. Run it in your terminal to execute setup followed by the saved Run command. Preparing or running the copy grants no approval. Use the same resource-isolation rules for preview copies. Review the exact result before approving its delivery.