Use case · Developers

Debug at the last good message.

A debugging conversation goes wrong the same way code does: one bad assumption early, and everything after inherits it. Branch at the moment before the wrong turn and the fix costs a click, not a restart.

The shape of the problem

Three candidate fixes. One polluted thread, or three clean branches.

01 / SCENARIO

Debugging without losing context

When the model starts guessing, do not argue with it in-line. Branch at the last message that was still on solid ground and run each candidate fix in its own path: the race condition theory, the cache theory, the config theory. Each branch inherits the full setup and none of the noise from the others.

  • Branch at the last good message, not the latest one
  • One candidate fix per branch, run in parallel paths
  • Dead-end branches stay visible as the record for code review
02 / SCENARIO

Comparing models on real code

Same repository context, same question, two brains: keep the main thread on your daily model and let a branch put a different one on the case. Ask the identical question verbatim, then read the disagreement as a diff instead of scrolling two long answers.

  • Switch the model on a branch; context carries over identically
  • Ask the same question word for word for a fair comparison
  • Two premium runs bill credits twice; that is the honest cost of certainty
03 / SCENARIO

The machine itself, on a branch

With Graft, a pip-installed bridge runs code and reads files on your machine while every tool call lands on the tree. Replay a session, compare two approaches to the same failing test, and merge the one that passed.

  • Local code execution recorded as branches
  • Outbound connection only; nothing on your machine listens
  • Branch straight out of ChatGPT with the Chrome extension
The features doing the work

What this workflow leans on.

  • Branching on every plan: the debug loop is free forever
  • Compare as a diff: candidate fixes read as red and green, not prose
  • A model per branch: Claude on one path, GPT-5 on the other
  • Graft local bridge: runs and files on the tree, replayable
  • Projects: one per repo, with standing context that briefs the model
  • Export: the winning branch becomes the PR description

Compare and merge are Pro features; branching and Explore are free on every plan. See pricing for the full split.

Get started

Never re-explain yourself.

Start free in under a minute. No card required. Use built-in credits or bring your own keys.