Post

You can do half of this in a stock CLI: keep the noise out, but not collapse what's already in. I wrap noisy commands so full output goes to a file and the transcript sees exit code and tail. One line in the rules file makes agents stick to it. Retroactive collapse needs your own tool.

Agent transcripts have an awkward double duty: record what happened and steer what happens next. Managing their size sounds like a core part of the interaction design.

I think that's a major benefit of building my own orchestration tool. It's not as easy to do (or is it even possible) using a stock coding agent app/cli. In this case I make it collapse the cli callback which has excessive information. From vs to:

Collapsing the callback is what a stock CLI leaves to the model. The full output enters context once, then every later turn carries it. Cutting it before it lands is a cost decision as much as a readability one.