Skip to content

Latest commit

 

History

History
70 lines (52 loc) · 2.29 KB

team-workflow.md

File metadata and controls

70 lines (52 loc) · 2.29 KB

Team Workflow

This page explains how Baudbot handles work from first request to final delivery so teams get quick acknowledgment, autonomous execution, and clear status updates.

High-level lifecycle

  1. Intake
    • Request arrives from Slack or direct chat (email is experimental-only).
    • Control agent treats all external content as untrusted and extracts intent.
  2. Acknowledge
    • Immediate response in the originating channel/thread ("On it", "Looking now").
  3. Track
    • Create a todo with source context (channel/thread metadata; include email metadata only when experimental email is enabled).
  4. Plan + route
    • Choose the correct repository (or multiple repos if needed).
  5. Execute
    • Spawn one or more task-scoped dev-agent sessions in git worktrees.
  6. Iterate
    • Dev agent runs code/test/PR/CI loops and reports milestones.
  7. Relay
    • Control agent posts progress and outcomes back to the same thread.
  8. Close out
    • Mark todo done, capture results, clean up sessions/worktrees.

Single-repo flow

Slack thread
  → control-agent todo + ack
  → spawn dev-agent in worktree
  → code + tests + PR + CI fixes
  → completion report
  → control-agent posts PR/summary in same thread

Multi-repo flow

Use either:

  • Sequential mode for dependent tasks (repo B needs output from repo A)
  • Parallel mode for independent tasks

Control agent coordinates fan-out/fan-in and keeps the user updated from one thread.

Follow-up handling

When users add requirements mid-task ("also do X"):

  • Keep the same todo and thread
  • Forward the update to the active dev-agent
  • Continue CI/review loop
  • Report updated status

Failure and escalation policy

If CI keeps failing or work gets stuck:

  • Avoid infinite loops
  • Post a clear status update with blocker context
  • Ask the user to decide next step (scope cut, rollback, manual handoff)
  • Preserve state in todo notes

Operational expectations

  • Always reply in-thread for Slack work
  • Keep user-facing updates concise and frequent
  • Prefer practical artifacts in status updates (PR link, failing check, preview URL)
  • Clean up ephemeral resources after completion

For role-level behavior details, see agents.md. For commands and runbook tasks, see operations.md.