Explore Offload

A cheap scout selects relevant code ranges and test-log filters by consumer focus. A fixed renderer makes raw, line-numbered bundles. The scout selects context; consumers inspect the repository and own conclusions.

Inputs

The requester supplies:

• purpose: why context is needed; each consumer's role, question, and one whitespace-free focus tag; matched watch-lists only for consumers that need them. The scout selects for these.

• seed: the task diff, named entry points, or a previous manifest with the files consumers added.

• logs: check output already captured on the current tree, or none. The scout filters logs; it never runs tests. Capture each check into the scratch directory you will give the scout:

  { <command>; } > <scratch>/<name>.log 2>&1; echo $? > <scratch>/<name>.log.exit
  

  Add the runner's options that disable color and quiet passing output when they exist.

Defaults: scout model = the cheapest fast model known capable of accurate context selection; bundle cap = 60000 characters (about 15k tokens) for the whole manifest; scout budget = 10 minutes.