Skip to content

How It Works

Barbossa runs AI agents in Docker. Each agent has a role and runs on a schedule.

Pipeline

  ┌───────────┐   ┌─────────┐
  │ Discovery │   │ Product │
  └─────┬─────┘   └────┬────┘
        └──────┬───────┘
         ┌──────────┐
         │ Backlog  │  ← GitHub Issues
         └────┬─────┘
         ┌──────────┐
         │ Engineer │  → creates PR
         └────┬─────┘
         ┌──────────┐
         │Tech Lead │  → merge or reject
         └──────────┘

Agents

Engineer

Implements issues labeled backlog.

  • Branch: barbossa/{issue-id}-{slug}
  • Uses issue context + CLAUDE.md
  • One issue at a time per repo

Tech Lead

Reviews PRs.

  • Waits for CI to pass
  • 8 dimensions: quality, tests, security, performance, complexity, UI/UX, integration, bloat
  • 3-strikes: closes after 3 failed reviews
  • Merges approved PRs

Discovery

Finds technical debt.

  • TODO/FIXME comments
  • Missing tests
  • Console.logs
  • Creates issues with evidence

Product Manager

Suggests features.

  • Reads CLAUDE.md
  • Creates issues with acceptance criteria

Auditor

Weekly health check.

  • Success rates
  • PR outcomes
  • Creates issues for problems

Spec Generator

Cross-repo specs (Spec Mode only).

  • Groups repos into products
  • Creates parent spec + child tickets

Security

Container runs as non-root (barbossa, UID 1000).

Action Allowed
Read code Yes
Create branches Yes (barbossa/ prefix)
Create/merge PRs Yes (own only)
Create issues Yes
Force push No
Other repos No

Protected files:

{ "do_not_touch": [".env*", "src/auth/**"] }

Schedule

Agents offset to avoid contention:

Agent When Purpose
Engineer :00 Creates PRs
Tech Lead +1h Reviews PRs
Discovery 6x/day Stocks backlog
Product 3x/day Suggests features
Auditor Daily Health check

Manual run

docker exec barbossa barbossa run <agent>
# engineer, tech-lead, discovery, product, auditor, spec