Working with the Harness

This page is about operator habits — how you drive an agent through a harnessed repo. These habits deliberately live here and not in AGENTS.md: they are how *you* work, not facts about the repo.

Session hygiene

• /clear between tasks. Context from a finished task is noise for the next one — it invites the agent to conflate unrelated changes. Clear it.
• One phase per session. Explore, plan, and implement are different modes. Run one phase per session so the agent's context stays focused and its output stays reviewable.
• Plan, then implement. For anything non-trivial, get a written plan first (the OpenSpec proposal is this) and review it before a single line of code is written. A wrong plan caught early is cheap; a wrong implementation is not.

When the agent goes sideways

• Re-prompt after two failed corrections. If you have corrected the same mistake twice and the agent is still wrong, stop correcting. The context is polluted. /clear, and re-prompt from a clean slate with a sharper description of what you actually want.
• Let the gate do the arguing. You do not need to convince the agent its code is wrong — run make check. The To fix: lines are the feedback loop; feed them back in rather than describing the failure yourself.

Trust the contract

The whole point of the harness is that you can review *less* and trust *more*. If make check is green, the repo satisfies its contract. Spend your attention on the diff's intent and the risky edges — not on re-checking what the gate already checked.