Bisect — AI skill for Claude Code

Your agent started failing. Find the combination behind it.

You added a skill. Changed a rule. Connected another tool.

Now your agent skips a step it used to get right.

**Bisect tests smaller sets of add-ons until it finds a combination that still brings the same bug back.** You get the selected add-ons, the test history, and a file you can replay.

One add-on can pass on its own. Another can pass on its own. Together, they can fail. Bisect checks that too.

See it happen

npx --yes github:automatedcatso/Bisect demo

Requires **Node.js 22+ and Git**. No account, API key, or model call for this demo.

16 components
       ↓
fast-finish + checkpoint

fast-finish alone    PASS
checkpoint alone     PASS
together             FAIL

This is an invented, repeatable demo. Both example components work alone; together they skip validation. Bisect runs the tests and checks the reduced result again. Exit code `1` is expected: it found the demo failure.


Prefer a local checkout? There is no build step:

git clone https://github.com/automatedcatso/Bisect.git
cd Bisect
node bin/harness-bisect.mjs demo

A bug report someone can use

Instead of “my agent sometimes acts weird,” bring:

• The combination: which selected add-ons still trigger the failure.
• **The