Feature: Containerized LangGraph reviewer pipeline for CI/CD (multi-agent automation for PR audits)

[ojfbot]

Overview

Build a containerized LangGraph application to automate running a series of specialized reviewer agents against this repository. The reviewer workflow is inspired by our existing pr-educator and blog-article-evaluator agents, and will be extended to include new ones such as a11y-audit and more.

---

Motivation

To ensure consistent, scalable, and automated review of code and content changes, we want an orchestrated workflow that runs in CI/CD, yet is flexible enough to trigger via both automated (event-based) and human/manual mechanisms.

---

Requirements

• Containerized LangGraph App: Build and maintain a Dockerized LangGraph agent pipeline tailored for multi-agent reviewer workflows, leveraging our patterns from pr-educator and blog-article-evaluator. Agents include:
  • Existing: pr-educator, blog-article-evaluator
  • New: a11y-audit
  • Others: Framework extensible for future agents (TBD)
• Trigger Conditions:
  • On first open (draft/ready for review)
  • On change from draft ➔ ready for review
  • On post-merge (merged to main)
  • On explicit manual trigger (human or agent-initiated)
• Workflow:
  • Scripted sequence: Each agent runs in prescribed order with shared state and can output actions, comments, PR suggestions, or audit reports.
  • Long-running/timed runs must not slow core CI; output should be clearly separated (logs, comments, or artifacts)
• Security & Resource Management:
  • Ensure secrets and API keys remain securely scoped (see CLAUDE.md)
  • Limit compute impact and idle runtime in CI
  • Ensure runs are idempotent and debounced from redundant triggers
• Documentation:
  • Document trigger points and manual override procedures
  • Add clear agent definition/extension steps for future reviewer agents
• Extensibility:
  • Architecture should allow adding future agents with minimal overhead

---

References

• Existing reviewer agents: see pr-educator, blog-article-evaluator (see repo)
• Agent orchestration: see multi-agent description in ARCHITECTURE.md
• Security/process: refer to CLAUDE.md and SECURITY.md

---

Acceptance Criteria

• Dockerized LangGraph app for agents exists with CI/CD entrypoint
• Triggers on the listed events as defined above
• Supports manual (human or agent) trigger
• Clearly documented and extensible for new reviewer agents
• Security practices followed for secrets and resource use
• Outputs are organized and do not interfere with core CI

---

/cc @ojfbot