Show HN: Agentic coding workflows built on Git worktrees and task evidence

Summary

GlueRun-go is a local orchestration engine for running multiple AI coding agents against one Git repository with worktree isolation, leases, gates, audits, and recorded task evidence.

Problem

Parallel coding agents can overwrite each other, leave stale work behind, or finish without enough evidence for a human or another agent to judge the change.
Long-running agent tasks can block repo control operations such as import, integrate, status, and stop.
Per-repo script copies make upgrades and project-specific behavior hard to manage across many repositories.

Approach

The engine uses a three-tier scheduler: L0 origin loop, L1 area planners, and L2 worker agents.
Each task runs in its own Git worktree and holds a JSON lease that records owner, retry count, and expiry.
Workers write structured state packets with owned files, changed files, commands, tests, and evidence; an auditor checks the packet and gate result.
A deterministic decider maps failure class and retries left to retry, amend-scope, escalate, or park, with a model call only as fallback.
Repos pin an installed engine version and keep repo-specific behavior in config, local overrides, or opt-in modules.

Results

The excerpt reports engineering claims, not benchmark results on a standard dataset.
Detached dispatch is on by default and makes gluerun reconcile --actuate return within seconds while workers continue in background; the legacy synchronous path waits for every worker.
Crash detection improves from a 60-minute stale-lease window to about one reconcile cycle by using dispatch records, worker exit files, and PID liveness checks.
The system includes 23 regression tests run by bash tests/run.sh.
Runtime session resume uses 10 staleness gates and falls back to a fresh run if any gate fails or the runner refuses resume.

Link

https://github.com/alex-reysa/glueRun-go