Manage context effectively in large codebase exploration

Keep exploring in one long session so nothing has to be re-established.

Late in a long session answers drift toward generic patterns and away from the specific classes found earlier — and it still sounds authoritative.

Write findings to a scratchpad and refer back to it, so knowledge outlives the window.

Have the main agent read the files itself to keep everything in one place.

Verbose discovery fills the context that coordination needs, so the agent loses the overview it was there to hold.

Delegate specific investigations to subagents; the main agent keeps the map, not the terrain.

Carry the raw exploration output into the next phase so nothing is lost.

The next phase begins with its budget already spent on the previous phase's transcript.

Summarise the phase, then inject the summary as the next phase's starting context.

Handle a crash by starting the exploration again.

Every agent's work is lost, and on a large codebase the re-run is the expensive part.

Have each agent export state to a known location and reload the manifest on resume.

Compact aggressively as soon as the context fills.

Compaction is lossy, so specifics established early are the first to go — and they are usually the reason the session was still running.

Persist what must survive first, then compact. Compaction is not a substitute for writing things down.