Find bugs and prove each one with a failing test
Run taskwing hunt on a part of your code. A session writes a test for every bug it suspects and keeps only the tests that fail on today's code. You confirm the proven bugs. Each fixer agent gets the bug but not its test, and the run ends as one draft pull request with a bugs table. On Osprey, a hunt proved and fixed 4 bugs in 8 minutes 58 seconds.
When this fits
A part of your code that you suspect, but cannot name a bug in: a scoring path, an event bus, a date-handling layer, a payment flow. You want real bugs with proof, not a list of possible issues to triage.
How a hunt goes
taskwing hunt "the payment scoring code" --no-pr # or no area: the whole repository
- Hunt. A session reads the area and writes a failing test for each bug it suspects. A test that passes on today's code means no bug, and the suspect is dropped.
- Pick. You see the proven bugs, each with its failing test, and press
yto fix them. - Fix. One agent per bug, each in its own sandbox. It gets the bug in plain words, not the test.
- Exam. On the joined work, every bug's test must pass and every existing test must still pass.
- Pull request. One draft pull request with a table of the bugs, the proof and the fix.
A real hunt, with the receipt
On 15 November 2026 we ran a hunt on Osprey, an open source transaction monitoring engine written in Go.
| Measure | Result |
|---|---|
| Bugs proven by a failing test | 4 |
| Fixer agents | 4 |
| Time | 8 min 58 s |
| Exam | 5 of 5 checks passed |
| Shipped | 2 fixes, in Osprey pull request #120 (merged) |
Why only 2 shipped: the other 2 bugs were real, but main had already fixed them in two earlier pull requests. The hunt had run on an older checkout. One of the shipped bugs is easy to picture: after Unsubscribe, the in-memory bus still delivered messages, and in 300 rounds the handler ran 289 times when it should have run 0. The write-up is in a bug hunt on Osprey.
Limits today
- Hunts run on your current checkout, so pull main first.
- The "existing tests still pass" check currently runs with the hidden bug tests, so one stuck fixer can look like a broken old test. Splitting them is planned.
- Fixes that touch scoring, a public API or migrations are not flagged for extra review yet. Read those diffs closely.
Questions
How is this different from asking an agent to find bugs?
An agent asked to find bugs returns opinions. A hunt returns only bugs that a test proves on today's code, and a suspect whose test passes is dropped. You never review a guess.
Why don't the fixers see the failing test?
So they fix the bug instead of the test. A fixer that can read the check can make it pass without fixing anything; the exam then runs the hidden test on the joined work.
How many bugs does one hunt keep?
Up to --max-cards, 4 by default. If two bugs need the same files, the second waits for the next hunt.
Can a hunt report bugs that are already fixed?
Yes, today. A hunt runs on your current checkout. In our Osprey hunt, 2 of 4 proven bugs were already fixed on main. Hunting on the latest main and skipping bugs with open pull requests is planned.
Does a hunt need a sandbox account?
For the fixers, yes: boat today. You can try the flow with --provider local and --no-pr, which uses folders on your machine and stops before any pull request.
Try the proof of concept
TaskWing is open source (AGPL-3.0) and early. Today it runs Claude Code agents in boat sandboxes, or in local folders to try the flow.
go install github.com/josephgoksu/TaskWing/cmd/taskwing@main
taskwing run "add a /health endpoint with a test" --provider local --no-pr