Nothing to do
NO_ACTION
Transient, external, already fixed, or already covered.
Ends here, silently.How a production error fixes itself
Every Beeswax production error flows through this pipeline. It deploys a fix alone only when the change is small, tested and proven by CI. Everything else waits for a person.
- Beeswax and Airbrake
- n8n
- Autofix server
- GitHub
- Heroku
An error happens in production
Airbrake groups it and sends it to n8n.
Noise is filtered out
n8n normalises the error and passes it to the autofix server.
Dropped: shutdown signals, and errors with no application code in the stack.
The job joins a queue
Jobs run one at a time on a clean copy of the code.
Skipped: a repeat of a recent error, or one that already has an open PR.
Claude investigates and picks one verdict
It reads the code, checks for duplicates, and if it writes a fix, adds a regression test and runs it.
Needs a person to think
ANALYSIS
The root cause is unclear, or the fix needs a human decision.
Draft PR with a write-up. Email sent.Fix, but your call
FIX_PR
A confident fix that touches business behaviour.
PR with an open question. Email sent. Never merged alone.Obvious defect
MECHANICAL_FIX
The bug and the correction are clear from the code alone.
Continues below.The change must be small and safe
- Claude's tests ran and passed
- At most 2 non-test files
- At most 60 non-test lines
- A test file in the change
- No auth, permissions, API, schema, config or front-end files
- Auto-merge switched on
Any miss: it becomes a “Fix, but your call” PR and you get an email.
CI must prove it on that exact commit
The server waits up to 30 minutes for independent proof, because Claude's own test report is not enough.
- Every workflow passed
- The Minitest suite really ran, and passed
- The commit has not moved
Failed, timed out or moved: nothing merges. The PR stays open for you.
Merged and deployed
The server squash-merges that commit into main, and Heroku deploys it automatically.
You hear about it, and the error closes
n8n emails “Fix deployed” and resolves the error in Airbrake.
When things go wrong
Claude's usage limit is hit
n8n waits five hours and sends the job again.
The autofix server can't be reached
n8n sends an alert email, at most once a day.
The first model is unavailable
The job falls back to a second trusted Claude model.
You want it to stop deploying
One setting turns auto-merge off. Fixes then wait as PRs.
Also on this server
Bug reports from Beeswax users and feature requests use the same server. Feature requests never auto-merge, and bug-report fixes don't either unless that is switched on.