Showcase / Made at Kane CLI Hack Day
AI agent
BUG QUEST
A DAG lost half its rows and Kane said PASS. Bug Quest writes the precise checks that catch it, fixes the code, and has Kane verify the fix.
Data pipelines fail silently: a DAG shows Success while rows go missing. In my Airflow-style Pipeline Monitor, load processed 500 of 1,000 rows, and a basic Kane check ("check the pipeline ran") said PASS, even with bug detection on. Bug Quest is an AI agent that fixes this. You type a short instruction; OpenAI exp…
1 builder credited · submitted Oct 24, 6:28 PM PDT
Kane · proof view
- Load step silently drops half the rows.
- Expand instruction into precise row and status checks.
- Kane fails the check: code 740, 1,000 vs 500.
- Fix code, Kane re-checks; pass ships, fail reverts.
- Open the Pipeline Monitor. Verify load task rows equal extract task rows (1000).Suspected · failed
- Open the Pipeline Monitor. Verify load task rows equal extract task rows (1000).Suspected · blocked
- Open the Pipeline Monitor. Verify load task rows equal extract task rows (1000).Suspected · failed
Product
AI agent
d3nl4wy7jovhq0.cloudfront.net
Harness
3 Kane parts
kane-cli run · sealed evidence packs · Test Manager
Evidence
3 runs
0 verified · 3 detected
Repo
4e19d25
github.com/EstrellaAce29/bug-quest
01 / The creation
What it is, and why.
Data pipelines fail silently: a DAG shows Success while rows go missing. A short check like "check the pipeline ran" passes, even with bug detection on, because it only verifies the green status.
Data pipelines fail silently: a DAG shows Success while rows go missing. In my Airflow-style Pipeline Monitor, load processed 500 of 1,000 rows, and a basic Kane check ("check the pipeline ran") said PASS, even with bug detection on. Bug Quest is an AI agent that fixes this. You type a short instruction; OpenAI expands it into precise Kane objectives with exact expected values (e.g. load rows must equal extract rows). Kane CLI runs each check in a real browser and catches the bug (code 740). Bug Quest then auto-fixes the code with git checkpoints, re-runs the same Kane check to verify the fix (rolling back after up to 3 failed attempts), and shows the results, code diff, and Test Manager evidence live on the /quest page. Built with Next.js, OpenAI, and Kane CLI. Repo starts with the bug on purpose so you can watch the full loop: catch, fix, verify.
- 01User names an agent and types a short instruction
- 02OpenAI expands it into precise Kane objectives with exact expected values
- 03Kane runs the original instruction and each objective in Chrome
- 04Bug Quest parses run_end: status, result_code, Test Manager link
- 05Silent bugs are scored; the basic check passes while precise checks fail
- 06
02 / The harness
How Kane checks it.
Bug Quest takes a short testing instruction and calls the OpenAI API to expand it into precise Kane objectives with exact expected values (e.g. "verify load rows equal extract rows (1000)"). It runs the original instruction and each objective with kane-cli run --agent against an Airflow-style Pipeline Monitor, parses the run_end result (status, result_code, Test Manager link), streams progress live, and scores silent bugs. In our run, the basic instruction passed (code 100) while the precise check failed with code 740 (confirmed bug: 1,000 extracted vs 500 loaded). After the code fix, the same Kane check was re-run with the bug switch still on and passed.
- kane-cli run
- sealed evidence packs
- Test Manager
- 01Load step silently drops half the rows.
- 02Expand instruction into precise row and status checks.
- 03Kane fails the check: code 740, 1,000 vs 500.
- 04Fix code, Kane re-checks; pass ships, fail reverts.
03 / Evidence
What actually ran.
Verified means a passing run reported through a paired, verified channel. A failing run is a detection, not a confirmed bug. Everything else is as the team reported it.
- Suspected · failedOct 24, 4:35 PM PDT
Open the Pipeline Monitor. Verify load task rows equal extract task rows (1000).
kanehack_cli · kanehack-6c3095c3-e442-4e07-89b6-93aaa9d97ba3 · exit 1
- Suspected · blockedOct 24, 6:10 PM PDT
Open the Pipeline Monitor. Verify load task rows equal extract task rows (1000).
kanehack_cli · kanehack-83702145-6cc8-4fcb-8506-82a2cfe45c75 · exit 2
- Suspected · failedOct 24, 6:11 PM PDT
Open the Pipeline Monitor. Verify load task rows equal extract task rows (1000).
kanehack_cli · 67daa20d-468c-429b-97e2-e4ed18793ec7 · exit 1
04 / Outcome
What changed.
1bug
Silent row-mismatch bug caught
Baseline 0 bug
Source: KR-261; Test Manager runs 01–03; extract 1,000 vs load 500 (team reported)
1attempt
Auto-fix attempts needed for a Kane-verified fix
Baseline 3 attempt
Source: auto-fix commit e8560fe; Kane re-ran the same check and passed (code 100); rowMismatchBug stayed true (team reported)
0false alarms
False alarms after auto-fix
Source: Live quest run on the auto-fixed code: all checks passed (team reported)
Known limitations
- Pipeline is simulated, no real Airflow or warehouse
- Auto-fix is verified on the row-mismatch bug only; other bug types not yet tested
- Auto-fix edits lib/pipeline.ts only
- Live mode runs Kane checks sequentially (about 2-3 min per quest)
- Live mode needs a local machine with Kane CLI, Chrome, and an OpenAI key
- Recorded view replays earlier real runs
05 / Ship it forward
The code. The proof.
The story.
- Prashansa Chaudhary
Also submitted
Run it
npm install && cp .env.example .env.local && npm run dev