Skip to content
#

Work on main with a coding agent

Check every commit your coding agent makes on main, or a whole span of commits when you ask.

#

Where the check fits

You and your coding agent commit straight to main, so there's no pull request to hang a check on. Start it from the commit instead. The check runs in RunStory's sandbox while you keep working, and it never blocks a commit.

Every commit, or one span Every commit 5 commits, 5 checks main a1c4e07 7be29d3 30f9a6c c8d1b52 e55a2b9 check 57 no problems check 58 no problems check 59 1 problem found check 60 no new problems check 61 no new problems A span of commits 4 commits, 1 check main a1c4e07 7be29d3 30f9a6c c8d1b52 e55a2b9 check 57 1 problem found runstory run a1c4e07.. checked earlier built and run

The same five commits, checked two ways: one check for each commit, or one check for the four commits after a1c4e07.

Every commitA span of commits
StartsBy itself, after each commitWhen you or your coding agent run one command
Tells youWhich commit broke somethingWhat's broken at the end of the span
ReadsThe commit, its diff, and the coding requests behind itThe commits in the span and their combined diff
CostsOne check for each commit that changes codeOne check
#

Check every commit

  1. Set up the repository once. runstory init installs the commit hook, plus a hook that hands findings to your coding agent. runstory setup starts your app in the sandbox, so you can see it boot.

    runstory init
    runstory setup
    
  2. Turn on automatic checks in your own terminal.

    runstory settings --automatic on
    

    RunStory shows the price, like "Each commit starts a check of 6 credits, paid by your personal account", and asks you to type yes.

  3. Keep working. Each commit that changes code now starts a check in the background.

Your coding agent can't turn this on. If it runs the same command, nothing changes.

RunStory skips what it can. Once you have a first result, a commit that only touches files that don't run, like a README, is skipped and costs nothing. A quick burst of commits, like a rebase, gets one check of the newest.

#

Check a span of commits

Leave automatic checks off, and start one check when a piece of work is done. Name the commit the work started from, then two dots.

runstory run a1c4e07..
You typeRunStory checks
runstory runThe newest commit
runstory run HEAD~4..The last 4 commits, as one change
runstory run a1c4e07..Every commit after a1c4e07
runstory run v1.2..Every commit since the tag v1.2

RunStory builds and runs the newest commit in the span, and studies the whole change. Add --dry-run to see what a check would cover before you spend anything.

$ runstory run a1c4e07.. --dry-run
Checking a1c4e07.., 4 commits from a1c4e07 to main (e55a2b9), with 6 files changed
...
This check runs 2 agents and costs 6 credits. You have 41 credits left.

Nothing was sent, because you passed --dry-run.

Ask your coding agent to do it

Your coding agent starts a check only when you ask. runstory init tells it so, in a short section it adds to AGENTS.md. So ask in the same request as the work, and name the commit the work starts from.

Ask your agent

Add gift cards to checkout, and commit as you go. When you're done, run one RunStory check of every commit since a1c4e07, then fix what it finds.

The agent runs runstory run a1c4e07... It tells you the price first, and starts the check once you agree.

#

What comes back

A check can run for up to an hour, so the result arrives while you're on to the next thing.

  • To your coding agent. At the end of its turn, it gets each new finding. This works in Claude Code, Codex, and Cursor.
  • To you. runstory status shows what finished. Each check also has a web page, with a video of the browser when your app is a web app.
  • With proof. Each finding has an id, a script that repeats the failure, and what that script did on the commit before your change.

When your agent commits a fix, it names the finding's id in the commit message. Then it runs the finding's script again, 3 times, in the sandbox. That retest costs no credits, and 3 passes close the finding.

runstory findings retest f-3a9c41d27b6e

See the whole loop, from a finding to its retest.

#

Example: a checkout change on main

A developer at Acme, an online furniture store, works on main with Claude Code. Automatic checks are on.

  1. The developer asks: "Let customers edit their delivery address during checkout."
  2. Claude Code changes the checkout code and commits 30f9a6c. The commit hook starts check 59.
  3. RunStory builds the store from 30f9a6c and starts it. Its agents add 2 items to the cart, change the delivery address during checkout, and find the cart empty.
  4. The failure's script repeats it 3 more times, and a second agent confirms it. The same script passed on 7be29d3, the commit before.
  5. At the end of its next turn, Claude Code gets finding f-3a9c41d27b6e and its script.
  6. It fixes the code and commits 9b2f4d1, "Keep the cart when the address changes (f-3a9c41d27b6e)". Then it runs the retest.
$ runstory findings retest f-3a9c41d27b6e
Re-test, setup 6, at 9b2f4d1
  f-3a9c41d27b6e  Its test passed 3 times at 9b2f4d1, so it is closed.
#

What it costs

A default check uses 2 frontier models and costs 6 credits. With 1 model, it costs 3.

ToRun
Cap this project's checks at 30 credits a dayrunstory settings --daily-limit 30
Check with 1 modelrunstory settings --providers anthropic
Check a commit that would be skipped, like a README editrunstory run --full
Pause automatic checksrunstory settings --automatic off

See RunStory pricing.