A test flakes and CI goes red at 2am. By standup the fix is already a merge request, pipeline green, waiting on your review.

Starts with red pipelines. Takes the whole backlog.

Pipeline fixing & CI triage

Diagnose the failing job straight from the logs
Patch the flake, not just retry it
Open a fix MR with CI back to green

Watch a real run

spec/auth_spec.rb+12 −5

@@ -41,4 +41,4 @@
- sleep 0.1
- expect(token).to be_present
+ wait_for { token }
+ retry_on_timeout(3)

Review comments, handled

A review comment becomes a follow-up commit
You approve diffs, not babysit threads

reviewer extract this into a helper and cover the nil case?

vemlor Done in a3f21c. Helper + 2 specs. CI green.

Dependency updates

Reads the changelog, fixes the call sites
Majors land green, not as red bot PRs

OpenCI passed

chore: axios 1.6 → 2.0, migrate interceptors

7 files+64−41

Codemod migrations

One mechanical refactor across every repo
One reviewable MR per repo, not a mega-diff

api!412 merged
web!388 merged
infra!72 running

Issue triage + bug fixing

Add reproduction steps and context to new bugs
Answer any @vemlor comment with an MR

you @vemlor fix the flake in auth_spec.rb

vemlor Pushed fix/auth-flake. CI green · !1248

Scheduled chores & reports

Docs, lint, and test gaps on a cadence
Digest of every merged MR, posted for you

Daily digest09:00

12 merge requests merged
3 repositories updated
posted to #engineering

How the loop works.

It starts with a comment.

A push, an MR, a red pipeline, or @vemlor fix the flake. Your GitLab or GitHub fires a signed webhook and vemlor queues the task.

POST/webhook/gitlabsigned

eventnote

comment"@vemlor fix the flake"

refmain

2m ago · queued task_1247

A runner boots in your cluster.

An ephemeral pod starts, checks out the repo warm, and authorizes your own agent subscription. Tokens and code never leave your side.

runner-eu-3eu-west

workspace acquired 0.4s
repo cloned warm
agent authorized 0.2s
claude-code running 4.1s

The agent does the work.

It reads the failure, edits the code, and re-runs the checks until they pass. No babysitting, and never a half-fix pushed to main.

spec/auth_spec.rb+12 −5

@@ -41,5 +41,6 @@ describe 'token refresh'
- sleep 0.1
- expect(token).to be_present
+ wait_for { token }.to be_present
+ retry_on_timeout(3)

3 retries stabilized · CI green

You get a reviewable MR.

A branch is pushed and an MR opened with the diff and CI status tracked back. You review, you merge. The same shape your team already trusts.

Open vemlorhq/api · !1248

fix(ci): stabilize flaky auth spec

4 files +247−246 CI passed

Any agent. Any model. Any platform.

Bring your own Claude Code or Codex subscription and your own model keys. vemlor runs them, you keep the bill.

Agents

Claude Code
Codex

Models

Z.ai
DeepSeek
Gemini
Grok
Kimi

Platforms

GitHub
GitLab
Bitbucket soon

Your autonomousPR engineer.