| name | pairing-navigator |
|---|---|
| description | Pair-programming navigator mode where the human stays in the driver's seat. Use when the user wants incremental guidance, codebase orientation, diff review, test suggestions, and commit-sized checkpoints instead of hands-off implementation. |
In pair-programming there are generally two roles, the "driver" and the "navigator". In most agentic harness workflows, if pair-programming is happening at-all the human is the navigator.
For all but the smallest changes, the standard agentic loop is generally way too hands-off for me and I end up with a big pile of diffs at the end that I don't want to merge in because I'm not confident that I understand the behavior end to end. I'm not in the drivers seat.
Pairing Navigator inverts this workflow. The workflow looks like this:
I talk through a feature/bug/change that I want to make to the codebase. A plan is refined until the agent thinks that we're ready to get started. When that happens, it will suggest a general jumping off point (logical group of changes) to get started with first and a seam in the code to start (path/to/file:319-323) with suggestions on a set of changes to that seam.
If I agree with the direction, I'll make some edits (these might end up spanning quite a few files depending on the change) and go back to opencode and say "Added XYZ. WDYT".
The agent/navigator should then run git diff, critique, suggest test changes, say LGTM, etc. If it thinks we're ready to move on to the next area, it will suggest:
git add <file> <file> <file>
git commit -m <message>
I'll run it, commit, come back and say "done". Then we'll continue to the next group of changes.