Skip to content

Guide

Set up goal checks and self-healing for your workflows

Pairbar is a macOS app that runs scheduled Claude Code and Codex workflows, and checks and repairs them as they go. A goal check reads what the steps did and ends with GOAL MET or GOAL NOT MET; rules retry temporary errors, swap models and stop stuck steps during a run; and when something keeps failing, a read-only "doctor" proposes a change you can apply or undo.

Last updated

About 5 minutes · 7 steps

Before you start

  • A workflow in Pairbar (schedule a workflow).
  • Notifications allowed for Pairbar, so you hear about missed goals and health flags.

Steps

  1. Write the goal
    In the workflow editor's Goal section, say what success looks like in checkable terms, for example "all tests on main pass and a PR with the fix is open".
  2. Turn on the goal check
    Switch on Check the result against the goal. A last read-only step, on whichever app has more limit left, reads every step's report, looks at the folder if needed, and ends with GOAL MET or GOAL NOT MET.
  3. Choose what happens when it isn't met
    Tell me sends a notification that the goal wasn't met, and why. Try once more re-runs the steps that change files (or the last ones, if none do) with what the check found, then checks again and tells you either way.
  4. Keep Fix problems during a run on
    In When things go wrong, Fix problems during a run is on by default. It applies fixed rules (no AI) to temporary errors, missing models, stalls and low limits.
  5. Allow the other app when a plan runs low (optional)
    Turn on Use the other app when a plan runs low so a step moves to the other app with its default model instead of failing or being skipped.
  6. Pick what the doctor may do
    Under After repeated problems, choose Suggest a fix (the default), Apply small fixes, or Do nothing.
  7. Read the run, not just the result
    Each fix shows as a small badge under its step in the run, for example "Retried on Claude's default model: opus wasn't available". The goal check's verdict is the run's last line, with its own Open.

Fixes during a run

ProblemWhat happens
Temporary error: overloaded, rate limit, server or network error (judged by the tool's error output, never by the reply)the step runs again after 30 s, then 60 s: three tries in all
The chosen model isn't availablethe step runs again on the app's default model
No activity in the step's chat for 10 minutesstopped and started once more; a step in the middle of a long command isn't considered stuck
A step runs for 30 minutesstopped and failed ("Timed out"), not retried: split it in two
The plan's limit is reached or too low before a stepwith Use the other app when a plan runs low, it runs on the other app's default model
Not signed in, or the folder is gonethe run stops and says exactly what to do

Health flags

The workflow's page summarises the last 10 runs: Steps OK counts runs whose steps completed, Goal met counts checks that returned a verdict, Typical time shows the median duration, and Limit skips counts runs skipped for limits. Hover over each figure for its meaning. The figures wrap in narrow sidebars. A failed goal check or missing verdict is shown separately. A notification goes out when a new flag appears:

  • a step failed the last two runs or more;
  • the goal was missed twice in a row;
  • a step's condition skipped it in every one of the last five runs (maybe it can go);
  • three or more runs were skipped for plan limits;
  • a run took over three times as long as usual.

The doctor

When a step fails twice in a row or the goal is missed twice, a short read-only planning chat on your own Claude or Codex reads the recent runs and proposes a clearer prompt, another model or effort, a condition, a schedule, or removing a useless step. It runs at most once a day per workflow; Check it and suggest a fix in the ⋯ menu runs it now.

SettingWhat happens
Suggest a fix (default)the page shows why, each change in plain words and every rewritten prompt before and after, with Apply and Dismiss. In narrow sidebars, the comparison uses the card’s full width and the date and buttons wrap separately. Apply changes only what the doctor changed and keeps your edits since
Apply small fixeschanges of only model, effort or schedule are applied by themselves (at most once a week) with Undo; a schedule that runs more often is never small
Do nothingonly the health notifications

Undo last change in the ⋯ menu reverts one change at a time, whether it was your edit, an applied suggestion or the doctor's own; the last 10 versions are kept.

Troubleshooting

A step keeps timing out
Steps are stopped at 30 minutes and not retried. Split the work into two steps with Add a step after this.
The goal is always NOT MET
Make the goal concrete and checkable from what the steps report or the folder shows. The doctor will also suggest a change after two misses.
The doctor changed something I didn't want
Use Undo last change in the workflow's ⋯ menu, and set After repeated problems to Suggest a fix.

Frequently asked questions

Does the goal check change any files?
No. It only looks: on Claude it gets Read, Glob and Grep, on Codex the read-only sandbox.
How many times does a workflow step retry?
For temporary errors, three tries in all: after 30 seconds, then 60 seconds. Timeouts after 30 minutes aren't retried.
Can the doctor make my workflow more powerful?
No. It never raises permissions, adds steps, turns off Ask me first, turns on the goal check or changes the folder.
How do I revert a change to a workflow?
Choose Undo last change in the workflow's ⋯ menu. Each use goes back one version; the last 10 are kept.

Try Pairbar free for 7 days

One sidebar for the Claude and Codex desktop apps. Everything on for 7 days, no card needed; then $49/year ($39 first year with LAUNCH), $9/month or $129 lifetime (first 200 buyers). macOS 26 or later, Apple silicon.