Bisecting Bugs With AI Help

When the Bug Was Born

A regression is a bug that didn't used to be there. Knowing when it appeared is more than half the fight. `git bisect` performs binary search through your history, asking you to mark each commit good or bad until it finds the first bad one. AI agents are surprisingly good at running this loop for you.

The 5-step manual bisect

# 1. Start the bisect at a known-bad and a known-good commit git bisect start git bisect bad HEAD # current commit is broken git bisect good v2.4.0 # last release was fine # Git checks out a middle commit and waits for your verdict. # 2. Test it. Either run a script, run the app, or both. npm test -- --grep "login flow" # 3. Tell git what you found git bisect bad # or `good` if the test passed # 4. Repeat (5-7 iterations for ~100 commits) # 5. When done git bisect resetManual bisect, the textbook flow. AI changes step 2 and step 3 dramatically.

Where AI plugs in

AI-assisted automated bisect

`git bisect run` takes a script and runs it on each candidate commit. The script's exit code (0 = good, non-zero = bad) decides the verdict. Have an AI agent write the script for you, then unleash bisect on it.

# Step 1: ask the agent to write a minimal repro script # Prompt: "Write a bash script test-regression.sh that exits 0 # if our login redirects to /dashboard, and 1 if it 500s. # Use curl. No browser. Save the response." # Step 2: make sure the script works on HEAD (should fail) and on v2.4.0 (should pass) chmod +x test-regression.sh ./test-regression.sh; echo $? # 1 git checkout v2.4.0 ./test-regression.sh; echo $? # 0 # Step 3: run automated bisect git bisect start HEAD v2.4.0 git bisect run ./test-regression.sh # Git crawls the history, runs the script per commit, # and prints the first bad commit when done. ~30 seconds for 100 commits.The agent writes the test, git does the search. You drink coffee.

Common bisect failures and AI fixes

• Test fails for unrelated reasons (missing dep) — agent re-runs `npm install` per commit
• Old commits don't build — agent marks those as `git bisect skip`
• Bug is intermittent — agent runs the test 5 times and votes majority
• Bug needs DB state — agent reseeds the DB before each test

# Useful prompt to a Claude Code session "I have a regression. Bug: <description>. Steps to reproduce: <steps>. 1. Find the last commit where this worked (assume it's tagged or in the last 200 commits). 2. Write a minimal bash script test-regression.sh that returns 0/1 based on whether the bug is present. 3. Validate the script works on HEAD (returns 1) and on the suspected good commit (returns 0). 4. Run `git bisect run ./test-regression.sh` and report the first bad commit. 5. Open that commit's diff and summarize, in 3 bullets, why this change introduced the bug."End-to-end automated regression hunt — works in any git project.

After the bisect

Once the agent finds the first bad commit, the next prompt is gold: "Read the diff of <commit-hash>. Explain in plain English what changed, why someone made that change, and the most surgical fix that preserves the original intent." You get history, intent, and a fix in one turn.

Bisect is the closest thing software has to a time machine.

— Linus Torvalds, more or less

The big idea: bisect plus AI agent equals a near-automated regression detective. Spend your effort writing a tight repro script and let the machinery handle the search. The first-bad-commit you find is usually so small the fix is obvious.