milad-ahmd/bug-detective

> This skill lives in the main collection: milad-ahmd/agent-skills. Please ⭐ star it there; that's where updates land first.
> Docs: https://milad-ahmd.github.io/agent-skills/skills/bug-detective/ · Claude Code: /plugin install all-skills --marketplace milad-ahmd/agent-skills
# 🕵️ bug-detective
Debug by reproducing first, logging hypotheses and auto-bisecting the commit that broke it. An Agent Skill for Claude Code, Cline, Cursor, Codex and other AI coding agents.
Install · Usage · How it works · All skills
> Part of agent-skills, a growing collection of practical skills for AI coding agents.
## 📦 Install
One command (installs the whole collection):
bash
curl -fsSL https://raw.githubusercontent.com/milad-ahmd/agent-skills/main/install.sh | bash -s -- bug-detective
Manually (Claude Code):
bash
git clone https://github.com/milad-ahmd/bug-detective /tmp/bug-detective
mkdir -p ~/.claude/skills && cp -r /tmp/bug-detective/skills/bug-detective ~/.claude/skills/
Per project: copy skills/bug-detective into /.claude/skills/.
Any other agent: paste SKILL.md into your rules or system prompt.
## 💬 Usage
Just ask your agent:
text
This test passed last week and fails now, find the commit that broke it
The cart total is sometimes wrong, help me find out why
The full workflow, rules and output format are in SKILL.md.
## 🔎 Sample output
bisectrun.py is plain Python 3 with no dependencies. Real output on a small repo where a cart total broke somewhere in 8 commits:
text
$ python3 bisectrun.py --good v1.0 -- python3 testcart.py
Checking the repro at both ends first:
good 3a78346285 good
bad f496d7eb41 bad
Bisecting:
1673fed043 good (exit 0, run 1/1)
bd6e296b35 bad (exit 1, run 1/1)
69c5a05d10 bad (exit 1, run 1/1)
First bad commit: 69c5a05d10b22e1977410534bec22e99aa0208a5
cart: apply discount per line
Author: Dana Lee
Date: 2026-10-15 10:00:00 +0000
Steps: 3 (plus the check at --good and --bad)
cart.py | 4 ++--
notes.md | 1 +
2 files changed, 3 insertions(+), 2 deletions(-)
Read it with: git show 69c5a05
It refuses to start on a dirty tree, checks that --good passes and --bad fails before bisecting, and always restores your HEAD, Ctrl-C included. --repeat N handles flaky bugs.