Who this is for
- Who this is for
- Engineers whose GitHub Actions (or GitLab / Azure) stay red after the “simple” fix
- Teams drowning in PR comments that miss the bug and nitpick style
- Platform leads watching CI minutes climb because every PR runs everything
- Security-minded orgs that will not upload the monorepo to a multi-tenant review cloud
Who this is for
- Engineers whose GitHub Actions (or GitLab / Azure) stay red after the “simple” fix
- Teams drowning in PR comments that miss the bug and nitpick style
- Platform leads watching CI minutes climb because every PR runs everything
- Security-minded orgs that will not upload the monorepo to a multi-tenant review cloud
If you want an autonomous software engineer that replaces seniors, this is the wrong page. If you want red → verified green → draft PR, keep reading.
Problems Kramlipi solves
Failed CI builds that steal the night
Problem: The pipeline fails at 11pm. The log is four thousand lines. You grep, guess, push, wait, fail again.
What Kramlipi does: The bug-fix expert parses the log (pytest, tsc, mypy, Go, Rust, and others), edits code inside a workspace jail, and reruns your --verify-cmd. It stops when that command exits 0 — then can open a draft PR with RCA, not a paste of the raw log.
Fixed when: The same verify command CI trusts is green, and a human still reviews the diff.
code-agent experts run bug-fix --log ci.log --verify-cmd "pytest -q" --publishWe do not rewrite .github/workflows/** to make the badge green. We fix the app code that broke the build. That is a product rule, not a missing feature.
Pull requests that wait for a human to catch the obvious
Pull requests that wait for a human to catch the obvious
Problem: Review queues grow because AI-written code ships faster than seniors can read it. Linters miss product bugs. Chat-only bots say “LGTM” or dump twenty vague comments.
What Kramlipi does: First-pass line comments on the diff: scanners + repo context + an AI critic, filtered back to changed lines. Humans still own design, architecture, and risk acceptance.
Fixed when: The obvious credential, eval(), or assignment-instead-of-compare is on the PR before standup — without replacing a senior design review.
The same PR fails again after every push
The same PR fails again after every push
Problem: Fix, push, new failure, context switch all day.
What Kramlipi does: experts watch --pr N babysits checks and re-runs the fix loop until green or it hits a cap — without a pile of duplicate spam MRs.
The same PR fails again after every push
Problem: Fix, push, new failure, context switch all day.
What Kramlipi does: experts watch --pr N babysits checks and re-runs the fix loop until green or it hits a cap — without a pile of duplicate spam MRs.
code-agent experts watch --pr 42 --verify-cmd "pytest -q"Full test suite on every tiny change
Problem: One helper file changes. CI runs thousands of tests. Merge queue stalls. Finance asks why Actions minutes doubled.
What Kramlipi does: test-intel maps the git diff to impacted tests and prints a narrower verify_cmd. Python/pytest selection is live (beta). Other languages are on a collector roadmap — measure wall-clock on your repo before you promise finance a percentage.
Fixed when: PRs run the tests that match the change, with a fallback when the graph is empty or risky.
Problem: One helper file changes. CI runs thousands of tests. Merge queue stalls. Finance asks why Actions minutes doubled.
What Kramlipi does: test-intel maps the git diff to impacted tests and prints a narrower verify_cmd. Python/pytest selection is live (beta). Other languages are on a collector roadmap — measure wall-clock on your repo before you promise finance a percentage.
Fixed when: PRs run the tests that match the change, with a fallback when the graph is empty or risky.
code-agent experts run test-intel --base-branch mainCoverage gates that block the merge
Problem: Coverage is under the threshold. The temptation is to weaken the gate or delete code.
What Kramlipi does: The agent adds tests against your coverage command (--cov-fail-under=N or equivalent). It does not “fix” coverage by deleting product code.
Fixed when: The same coverage command exits 0 and a human reviews the new tests.
Flaky tests that burn reruns
Problem: The suite fails, you rerun, it passes. Teams normalize retries and hide instability.
What Kramlipi does: Fail/pass history per test node, plus skip-known-flakes on a PR with a logged reason — not silent sabotage. This is heuristic memory (~partial classification), not “100% flaky test detection.”
Fixed when: Known flakes are skipped on MR with an audit trail, and real regressions still fail the gate.
It passes locally, fails in CI
Problem: Path, Python version, missing service — local green, pipeline red.
What Kramlipi does: You give it the exact pipeline verify command (binary, Docker, or Action). It proves the fix under that command. It does not yet recreate a full CI compose stack for you — you still supply the real command and a workspace that matches CI as closely as you can.
Uploading source to someone else’s cloud
8. Uploading source to someone else’s cloud
Problem: Review SaaS wants the monorepo. Procurement and security stall.
What Kramlipi does: Customer-hosted binary / Action / VS Code. BYO LLM (Gemini, Claude, OpenAI, local). We do not train on your codebase in our cloud. Keys stay in environment variables, not committed config.
Features (what you can run today)
| Feature | Surface | What “done” means |
|---|---|---|
| Verify-gated repair | bug-fix, run | Your verify_cmd exits 0 or the run fails closed |
| Draft PR / MR | --publish | Human reviews a diff, not a chat transcript |
| First-pass PR review | code-review | Inline comments with evidence on changed lines |
| PR babysit | experts watch | Re-fix until green or max attempts |
| Diff-aware tests | test-intel | Impacted pytest (beta); other languages collecting |
| Coverage growth | run + cov verify | Tests added; gate not lowered |
| Flake memory | memory + skip-on-MR | History + optional skip, with a reason |
| Forbidden workflows | policy | .github/workflows/** not auto-edited |
| Local / CI / editor | binary, Action, VS Code, web chat | Same loop, your runner |
| Economy mode | default caps | Fewer empty LLM loops; subprocess is the grade |
Where it runs: local binary, code-agent web serve, CLI run, GitHub Action, VS Code extension, Docker. Bring Gemini, Claude, OpenAI, or a local model.
Features (what you can run today)
| Feature | Surface | What “done” means |
|---|---|---|
| Verify-gated repair | bug-fix, run | Your verify_cmd exits 0 or the run fails closed |
| Draft PR / MR | --publish | Human reviews a diff, not a chat transcript |
| First-pass PR review | code-review | Inline comments with evidence on changed lines |
| PR babysit | experts watch | Re-fix until green or max attempts |
| Diff-aware tests | test-intel | Impacted pytest (beta); other languages collecting |
| Coverage growth | run + cov verify | Tests added; gate not lowered |
| Flake memory | memory + skip-on-MR | History + optional skip, with a reason |
| Forbidden workflows | policy | .github/workflows/** not auto-edited |
| Local / CI / editor | binary, Action, VS Code, web chat | Same loop, your runner |
| Economy mode | default caps | Fewer empty LLM loops; subprocess is the grade |
Where it runs: local binary, code-agent web serve, CLI run, GitHub Action, VS Code extension, Docker. Bring Gemini, Claude, OpenAI, or a local model.
How the loop works (mechanism, not magic)
- Read the failure log or the PR diff — parsers first, not a model guessing from a screenshot.
- Edit only under a workspace jail (path allowlist, file caps).
- Verify with your command. The model does not mark its own homework.
- Publish a draft PR or line comments when the gate is green (or review-only when you asked for comments).
- Stop honestly if verify stays red. No silent pass.
That is why we say verify-gated. It is also why we refuse to “repair CI” by rewriting pipeline YAML.
What Kramlipi is not
Be precise so searchers who need a different product leave before they bounce:
- Not a full senior design-review SaaS (CodeRabbit-class breadth of comments is not the job)
- Not a CircleCI replacement and not a Chunk lock-in — we sit on GitHub Actions and your runner
- Not a guaranteed-fix SLA — we promise the gate, not a percentage
- Not SOC 2 / HIPAA certified (do not claim it)
- Not an agent that edits GitHub Actions YAML to fake green
Honest stack: many teams keep a comment bot and Kramlipi. One talks about the diff. The other will not ship until your tests pass.
FAQ
Is Kramlipi an AI code review tool?
Yes — first-pass, diff-aware line comments with scanners and a critic. It is also CI repair. Review without a verify gate is only half the job.
Can it fix a failing GitHub Actions build?
It can parse the failed log, change application code, and rerun your test/build command. It will not edit workflow YAML to force a green check.
Does it detect flaky tests?
It keeps fail/pass history per test node and can skip known flakes on a PR. Treat that as evidence-backed memory, not a perfect classifier.
Do you host our source?
No. Binary, Action, or editor on your side. You bring the LLM key.
Is there a free tier?
Yes — local free tier with caps, no credit card required. Production CI at team pricing is on kramlipi.com.
CodeRabbit alternative?
If you only want more review comments, stay with a review bot. If you want failed build → your tests green → draft PR, Kramlipi is the other job.
Get started
- Download the binary: code-agent-binaries
code-agent setupthencode-agent doctor --provider-test- Repair:
code-agent experts run bug-fix --log ci.log --verify-cmd "pytest -q" - Review in CI: pin the Marketplace Action
Docs: Get started · Positioning: verify-cmd · draft PR · customer-hosted · BYO key · forbidden workflows · receipt