booting
When production breaks, Cicatrixa reproduces the crash as a failing test, patches until your whole suite is green, and ships through a canary — before your on-call wakes up.
▼ HOW IT REFUSES TO GUESS
The objection, answered
Yes — the model in your terminal can debug. But between "a model that can debug" and "a production system that heals itself" sits everything that isn't intelligence. Nobody says "why do you need CI — you can run the tests by hand." We're CI for incident response.
Nobody asks at 3:07am. The incident must be noticed, triaged against noise, and acted on while the human is asleep. A chat window has no pager.
Context assembly is the hard 80%. Traceback, the right slice of source, recent deploys, somewhere to run code. Engineers rebuild this by hand every incident; we assemble it automatically, every time.
A sandbox with a burden of proof. Ad-hoc model use produces a plausible patch. Our harness physically cannot ship anything without a previously-failing, now-passing test and a green suite. The rule is enforced by the pipeline, not by prompt discipline.
Deploy rails. A model can't put itself on 20% of live traffic and roll itself back. We own the last mile — which is where the actual risk lives.
No reproduction, no patch. Ever.
— THE ONE RULE THE AGENT CANNOT BREAK
The healing loop
The same bar you hold human engineers to — evidence first, ship second.
Tails logs and health checks on the live service. Catches error spikes and crashes with full tracebacks.
Reads the traceback against your indexed source and pins the root cause — file, line, and the exact input that triggers it.
The step nobody else does. Writes a failing test in an isolated sandbox that proves the bug. Can't prove it? The pipeline stops. Guessing is not an option it has.
Patches until the reproduction test passes and the full existing suite stays green — verified in sandboxes, never on prod.
Optional one-click human approval on the verified diff. Autonomy is a dial you turn as trust accrues, not a leap of faith.
Ships to a slice of live traffic, watches the error rate, auto-promotes or rolls back. The regression test stays in your repo forever.
The product, live
This is the control plane your team sees — a real healing run, stage by stage, from crash to promoted canary.
Get running
The managed cloud needs nothing installed — two browser authorizations and incidents start arriving as proven PRs. Prefer to hold it in your hands first? The full healing loop ships as a self-hosted stack you can run on a laptop.
Paste one webhook into Sentry, Datadog, or CloudWatch — the tools you already run. Five minutes, reversible any time.
The same authorize button you've clicked for CI. Repo read, pull-request write, nothing else ever touches your infrastructure.
Each incident becomes a PR: failing test + minimal fix + green suite + canary plan.
Where we sit
| strong model in a terminal | AI SRE assistants | cicatrixa | |
|---|---|---|---|
| Notices the incident | ✗ you do | ✓ alerts | ✓ alerts |
| Assembles the context | ✗ you do | ✓ | ✓ |
| Reproduces as a failing test | ± if you insist | ✗ | ✓ always, or halts |
| Writes a verified fix | ± plausible patch | ✗ explains only | ✓ repro + suite green |
| Ships it & rolls back | ✗ | ✗ | ✓ canary, auto |
| Leaves an artifact behind | ✗ chat scrollback | ± incident report | ✓ permanent regression test |
What lands in your repo
This is the artifact your team reviews — the same bar you hold humans to: the test that proves the bug, the minimal fix, and the checks that gate the merge.
Questions, answered straight
We read your repo through the GitHub App scope you grant — nothing broader. Reproduction runs in isolated, network-disabled sandboxes that are destroyed after each run. Your code is never used to train models.
The pipeline halts — by design. You get the diagnosis and the attempted test, clearly marked unproven. What you never get is a blind patch. Refusing to guess is the product.
It can't touch it. The output is a pull request; nothing ships until your existing CD ships it after merge. Teams that turn the autonomy dial up get canary deploys with automatic rollback — a fix reaches 100% of traffic only after proving itself on a slice.
Autofix tools patch from a stack trace and hope — a stack trace tells you where code died, not what correct behavior is. We require a failing test that proves the bug before any patch is written, then a green suite, then a canary. Proof, not plausibility.
The beta targets Python services on GitHub, with Sentry, Datadog, and CloudWatch as alert sources. More languages follow the same harness — ask us about yours.
Early access
We're onboarding a small group of design partners. Bring us one real production bug — we'll heal it free, failing test included.
Contact
Questions, pilots, press, or a production bug you want healed — one address reaches the founders directly.