Codebase Autopilot

Error Sweep

A daily reliability pass that traces your real production errors to root cause and opens a PR per verified fix.

Moderate setupShort cycleOpen loopVisible first run

Setup in Loopany

How a run flows

Before first run

Verify an error source

On schedule

Every day · 6am

Step 1 · Sweep

Group into incidents

Step 2 · Triage

Actionable vs. noise

Isolated git worktree · off main

Step 3 · Fix

Root-cause & fix

Step 4 · Ship

One PR per fix

Report

Newest report

Metric

Actionable errors

open loop — repeats on the next tick, picking up where it left off

Prerequisites

A readable error sourcerequired
Logs, an error tracker, a URL, or gh — the sweep needs somewhere real to read errors from.
Verifier setup for code changesoptional
A repo-specific harness so every fix is proven against the real app before a PR. Set it up with the verifier-setup skill

How this loop works — schedule, exit, notifications

Suggested schedule: Daily · ~6am
Loop type: Open loop — an ongoing monitor with no finish line.
Exit condition: Runs indefinitely (no finish line).
Notifications: On by default — you're alerted on each run's outcome (completion or failure). Silence with notify: never.
Ease of starting: Some setup and context are needed before the first useful run.
When you see value: As soon as real production errors exist, the first sweep root-causes them into PRs.

That morning's report looked like two regressions. Error Sweep pulled the traces and sessions, then found the one thing behind both: Safari 16.1 couldn't parse a new regex. 16 occurrences, 3 real users, modern browsers clean.

Instead of another dashboard, I got one question. Should we transpile the syntax and keep supporting the old browser, or make Safari 16.4 the floor? I answered once and we were done.

The loop's working file: what already shipped, what stays watch-only, and the one item waiting on a human decision

That's why I read this report. It follows noisy errors back to the boundary they share, then checks the counts against actual users and sessions. By the time it reaches me, I can merge a fix or answer a question. The investigation's already done.

The prompt this runs

Set up a daily loop that sweeps this project's production errors. Before creating anything, verify a concrete way to read them exists on this machine - logs, an error tracker, a URL, or gh - smoke-test it once, propose the source and sweep window, and confirm it with me; never create a blind loop. Every morning around 6am my local time, sweep the agreed window, group repeated symptoms into incidents, and separate actionable errors from noise and upstream failures. Trace each to root cause, make the smallest verified fix in a fresh git worktree off the main branch and located outside the loop folder - the loop folder is a synced content home, not a workspace - so neither my working checkout nor the sync is polluted; open one pull request per fix, never while a previous Error Sweep PR is unmerged. Never copy credentials, tokens, or personal information into reports or PRs. Write one dated report per run with front matter `type: report`, a `title`, and `date`; report the actionable-error count as a metric; nothing actionable is a clean stop. Set up the loop's dashboard at creation: an embed of the newest report plus a chart of that count.

This is the exact intent appended to the bootstrap when you create the loop — your coding agent reads it, then proposes cadence and config and confirms with you before creating anything.