Delegate Codebase Discovery, Then Spot-Check the Handoff

Finding where an unfamiliar research repo computes pass@k, or where it filters the eval set before scoring, can take dozens of file reads before you reach the two functions that matter. If all that reading happens in your implementation session, it fills the context you need for the change.

Claude Code's common workflows describe delegating it: ask Claude to use a subagent, which reads files in its own context window and reports a summary. Ask for a fixed handoff so the summary is checkable:

Use a subagent to find where pass@k is computed and which samples it counts.
Return: the finding; file paths with line numbers; the commit (git rev-parse --short HEAD);
excerpts of at most 10 lines each; and anything you could not settle.

Spot-check before you build on it. A subagent can miss a second call site, read a generated file instead of its source, or describe code from a branch you are not on. Pick three citations from the handoff and open them. Does the line say what the summary says? Keep a tally across handoffs: after 20 checked citations, you have a rough rate at which this kind of summary holds, and an interval on it (at 17 of 20, p = 0.85 and 1.96 × √(0.85 × 0.15 / 20) ≈ 0.16). That rate tells you how much checking the next handoff needs.

Run /context before and after to see what the delegation saved in the main session. The reading still costs tokens and time in the subagent, so this manages where context goes; it does not make discovery free.

The book's delegation chapter works out when a second agent pays for itself and when the handoff costs more than it saves.

The check: each finding you act on has a file and line you opened yourself, at the commit the handoff named.