Recoleta
GitHub
Source note

Loop Harness Is Here

Authors
LordIsBack
Source
Hacker News
Published
Collection
Software Intelligence
Coding AgentsSoftware Engineering AutomationAgent OrchestrationCode ReviewCi Triage
Open GitHubMarkdown source

Summary

Loop-Harness is a shell-based orchestrator that runs Claude Code agents on scheduled repository tasks, checks their output with a second Claude session, and ships only approved changes. It targets unattended code maintenance tasks such as CI triage, PR review, issue grooming, dependency updates, and doc sync.

Problem

  • Coding agents with shell access can damage repositories when they run unattended; teams need isolation, narrow tool access, and a review gate.
  • Routine maintenance work needs polling, state tracking, deduplication, retries, logs, and output handling; one-off prompts do not provide those controls.
  • The problem matters because write-capable loops can commit code, open PRs, post comments, and push branches into live development workflows.

Approach

  • A scheduler picks due loops per repository using an every cadence or a 5-field cron expression.
  • Each write run gets a fresh git worktree on a loop/<name>/<ts> branch, so the agent does not modify the user’s checkout.
  • The primary claude -p session receives a task skill file and a scoped allowed_tools list. It stages commits, PR_BODY.md, or outbox files, but it cannot directly post or push.
  • A verifier claude -p session inspects the diff and staged outputs, runs cheap checks, and must print VERDICT: PASS before the orchestrator ships anything.
  • Per-loop JSON state records last_run, processed item IDs, failures, retries, and metrics, so repeated polling and duplicate triggers can be safe.

Results

  • The excerpt reports no controlled benchmark, ablation, or comparison against another agent harness.
  • The sample dashboard shows pr-reviewer with 112 runs, 98% success, 31 items processed, 64s average duration, and 2 failures on a 3-minute cadence.
  • The sample dashboard shows triage-ci with 38 runs, 95% success, 9 items processed, 241s average duration, and 2 failures on a 10-minute cadence.
  • The sample dashboard shows dependency-updater, doc-sync, and issue-groomer at 100% success, with 4, 4, and 9 runs; 6, 2, and 14 items; and 312s, 198s, and 87s average duration.
  • Concrete operational claims include one worktree per write run, a verifier gate with an exact pass verdict, scoped Claude tools, default max_retries: 2, default timeout_minutes: 15, and shutdown that waits up to 5 minutes for active loops.

Link