Per-label pipeline files

pipelines/
  ai-feature.yaml    ← full PM→Arch→Eng→QA
  ai-fix.yaml        ← minimal Eng→Review
  ai-docs.yaml       ← DocOrchestrator
  ai-refactor.yaml   ← Engineer only
  ai-custom.yaml     ← your pipeline.yaml

main.py impact: Add --pipeline ai-feature flag.
No flag → uses default (first pipeline or a config default).
Dev can also pass --label ai-fix to simulate a label trigger.

Concurrency: Work Queue + Semaphore

          # repos.yaml / config

          max_concurrent: 1   ← safe for Ollama

          max_concurrent: 3   ← fine for OpenAI

          max_concurrent: 10  ← hosted APIs

Flow:

Watcher polls all repos (single thread)
Found issues → pushed to a work queue
N worker threads pull from queue
Each worker acquires a semaphore slot
LLM call runs; slot released when done

Full flow diagram

Watcher poll

Every N minutes

Scans repos.yaml
repos × labels

issues found

Work Queue

          { repo, label, issue_number }

          { repo, label, issue_number }

          ...

dequeue

Semaphore (max_concurrent)

Worker 1 ▶ pipelines/ai-feature.yaml
Worker 2 ▶ pipelines/ai-fix.yaml
Worker 3 ⏸ waiting...

load

Pipeline runner

Reads pipelines/<label>.yaml
Runs stages in order
Falls back to built-in if no file

Ollama-safe: max_concurrent=1

With max_concurrent: 1, only one pipeline runs at a time. The queue ensures no work is lost — issues pile up in the queue and are processed one by one. No parallel LLM calls.

Multi-Repo, Multi-Label Design