A long-running shell-based AI agent that uses Codex CLI and GitHub CLI (gh) to continuously triage repository work.
The project uses a structured task tracking system located in the .tasks/ directory:
.tasks/manifest.json: A central registry of all tasks, their statuses, and associated GitHub Issue IDs..tasks/<task-slug>.md: Detailed information for each task, following the lifecycle defined inspec.md:- Research
- Plan
- Implementation Decision (Proceed/Refuse)
- Implementation Log
- Test Results
Status codes:
todo: Task is in the backlog.in-progress: Task is currently being worked on.done: Task is completed and verified.refused: Task was intentionally not implemented.failed: Task implementation failed.
Following the spec.md recommendation:
doctor.sh- supervisor loop + logging + locks
- sync-main worker
- project issue picker
- issue workflow with research and plan checkpoints
- implement-or-refuse decision gate
- refusal worker and Backlog transition
- implement stage
- test stage
- comment scanning + reactions
- PR conflict resolution
- cleanup and retry polish