The agent reads by default and writes only with permission. Log search and commit history are read-only, while reruns and bisect consume machines and need a budget.
Tool | Default permission | Why |
|---|---|---|
Log search | Read-only | It only quotes existing lines. |
Commit history | Read-only | It only lists merged changes. |
Rerun a seed | Needs approval | It spends simulator time. |
Bisect | Needs approval | It spends many simulator runs. |
Anything that runs a simulator asks first. The request names the seed, the estimated run count, and the question the runs would settle.
A human approves or trims the request. Routine small reruns can carry standing approval with a nightly cap, and anything above the cap asks again.
Log lines are data, not instructions. A failing test can print text that reads like a command, and the agent must quote it without obeying it.
It files an approval request with the seed, the count of twelve, and the suspect range. Nothing runs until the request is approved.
Nothing with it. The line is quoted as evidence if relevant, and never becomes a tool argument.