Automate issue triage, parallel execution worktrees, self-healing rebases, and multi-agent review verification. Seamlessly matching autonomous developer throughput to human approval bandwidth.
The repository-backed operating model: governance layers, process loops, policies, and templates. Humans decide, agents execute, Git governs.
Pillar 02 — The Knowledge LayerThe layered, vendor-neutral knowledge-ops spec with the /kb command — keeps humans and agents on the same page.
The production-ready engineering loop that schedules tasks, resolves rebase conflicts, and verifies code via tests.
Scaling multi-agent developer squads without overwhelming human managers.
Without limits, agents work on all open backlog items concurrently, opening dozens of PRs that quickly overwhelm review pipelines and lead to stale code.
Main branches move forward, leaving agent-authored PRs out-of-date or conflicting. Resolving this manually wastes developer cycles.
Running isolated agents on complex dispatches causes heavy API utilization, causing rate limits that crash execution tasks without a fallback chain.
Enforces configurable limits (e.g. `open_pr_cap_per_repo`) so agents stop creating new PRs once the human review queue is full.
Automatically detects PRs behind base branches, attempts clean Git rebases first, and falls back to LLM models to resolve conflicts safely.
Configures custom CLI pipelines (e.g. Codex -> Claude -> AGY SDK) that smoothly cascade to next models if one hits context or rate limits.
Designed to operate continuously and safely in background servers.
Uses local Git and GitHub REST APIs via `detect.py` to identify pending actions. By executing zero LLM calls during scanning, it minimizes token costs and runs instantly every minute.
Dispatches individual issues to transient systemd service units. Each runner checks out code in an isolated git worktree, preventing workspace pollution and enabling parallel builds.
Enforces local test and linter execution prior to declaring work complete. If tests fail or code style drops, the issue is kept in progress for fixing or returned to the backlog.
Maintains persistent lock files under the state directory. Prevents duplicate dispatches on the same issue across overlapping tick runs while releasing locks automatically on task exit.
How to incrementally introduce agentic development to your repositories.
Configure `TRIAGE_ENABLE_DISPATCH=0` in the service. The scanner will run and log planned actions, locks, and concurrency checks to let you inspect the loop's behavior risk-free.
Enable dispatches, but keep `automerge=false`. The agents will pick up issues assigned to them, open PRs, run code reviews, and write suggestions, but a human retains merge rights.
Enable `automerge=true` for specific repositories. PRs that are successfully authored, pass CI tests, and receive green reviews from verification agents are merged automatically.