Skip to content

kramlipi code-agent: AI That Fixes Code Until Your Tests Pass

SEO title: kramlipi code-agent: AI That Fixes Code Until Your Tests Pass
Meta description: Fix failing unit tests and CI with kramlipi code-agent. It edits your repo, runs your verify command, and only stops when checks pass. Docker + one script—local Chat UI or your pipeline.

kramlipi code-agent

Most tools that claim to “fix CI with AI” do one of two things: they talk, or they hope.

kramlipi does a third thing. It changes your code, runs the check you already trust, and refuses to call the job done until that check passes.

That is the whole product.

What you need

Three things. Nothing exotic.

  1. A Docker host (laptop or CI runner) that can pull
    ghcr.io/kramlipi/code-agent:latest
  2. An API key for a model you already use — Gemini, OpenAI, Anthropic, or similar — set on the machine or as a CI secret (GEMINI_API_KEY, OPENAI_API_KEY, …)
  3. A project folder and a verify command — the same command your pipeline already runs to decide pass or fail (pytest -q, go test ./..., npm test, …)

Optional: the one-file launcher docker-ui.sh if you want a browser UI on http://127.0.0.1:8080 instead of only the CLI.

# Linux / macOS / WSL
export GEMINI_API_KEY=your-key

curl -fsSL -o docker-ui.sh \
  https://gist.githubusercontent.com/kramlipi/d31f4f454cd127cfb552e5ed5e854af3/raw
chmod +x docker-ui.sh
bash docker-ui.sh

# Windows PowerShell
$env:GEMINI_API_KEY = "your-key"

Invoke-WebRequest -Uri "https://gist.githubusercontent.com/kramlipi/387228f78eb47e437f578f625a101707/raw" `
  -OutFile docker-ui.ps1
.\docker-ui.ps1

Gists: Linux docker-ui.sh · Windows docker-ui.ps1

Open the UI, choose a folder on your machine, and work.
Or skip the UI and run the image in CI the same way you run any other container.

What it does

In plain language:

  • Fixes failing unit tests until your test command is green
  • Fixes build, type, and lint failures the same way — against your gate
  • Raises coverage by adding tests when a coverage gate fails — not by deleting product code
  • Takes a failed CI log and produces a real fix in the repo, not a summary in Slack
  • Opens a draft pull request when you ask it to publish, so someone can review in the morning
  • Babysits a PR that keeps going red — repairs the same branch and lets CI run again
  • Shortens PR test runs by selecting only the tests that matter for the change
  • Finds missing metrics on handlers and can open an MR for instrumentation
  • Responds after deploy or to an alert with the same rule: change code, prove it, or stop

It does not replace GitHub Actions, GitLab, Jenkins, or CircleCI.
It sits beside them. Your runners still run. Your gates still decide. kramlipi proposes edits; the environment keeps the score.

It will not rewrite your workflow YAML to fake a green check. That is deliberate.

How it does it

The loop is short on purpose.

  1. Intake — a prompt, a CI log, a PR that failed, or an alert
  2. Understand — structured reading of the failure (what broke, where), not vibes
  3. Edit — real files in the project you pointed it at
  4. Verify — run the command you named
  5. Decide — exit code 0 → done (and optionally draft PR); anything else → try again or fail honestly

The model suggests.
Your verify command decides.

Same image on a laptop and in a pipeline. Same key. Same rule.
Local tonight: UI or CLI. Pipeline tomorrow: a step on failure, a follow-up job, a nightly, or a PR babysit — mount the repo, pass the key, pass the verify command.

Where it sits in a pipeline

Think of it as a worker that only speaks when it has proof.

When a job fails
Feed it the log (or a webhook). It edits the app, re-runs your verify command, and can open a draft PR. Duplicate failures that already have a fix in flight can be skipped.

When a PR keeps failing
Point it at the PR. It fixes on that branch, within a retry limit, and pushes so CI runs again. You still merge.

When CI is too slow
Ask it for the tests that match the diff. Put that shorter command in the PR job. Fall back to the full suite when selection doesn’t apply.

When a quality gate fails
Same loop — tests, build, lint, coverage — always against the command you already trust.

After release or when ops pages you
Health or metrics look wrong, or an alert fires: attempt a fix, prove it, or stop. No victory speech without a passing check.

What the pipeline must provide
The image, the secret, the checkout, and the verify command. That is the contract.

The honest pitch

Conversation is cheap.
Consequence is not.

kramlipi is for teams who are tired of AI that sounds finished while CI is still red.

What it does: fix real code and prove it.
How it does it: edit → run your check → repeat or stop.
What it needs: Docker, a model key, a project, and a verify command.

export GEMINI_API_KEY=…
bash docker-ui.sh

One image. One script if you want the UI. One rule: prove it.

kramlipi code-agent — it doesn’t get to declare victory. Your tests do.