ACTIVEagent

scout

Reads the codebase before anything is written, and says out loud whether the premise is already stale.

Model
sonnet
Tools
ReadGrepGlobBash(git *)Bash(ls *)Bash(cat *)Bash(rg *)Bash(pnpm ls *)
Cost
One Sonnet pass over a few dozen files. The cheapest stage, and the one that saves the most.

WHEN IT RUNS

Read-only codebase investigation for the SPID pipeline. Use at the start of non-trivial work, or whenever a cached scout-report may be stale. Produces .buildloop/scout-report.md and never edits anything.

CONTEXT

The task, plus whatever it chooses to read: CLAUDE.md at every level, package.json and turbo.json for the real commands, the files the task will touch, and the tests that constrain them.

MEMORY

None between runs. It writes .buildloop/scout-report.md and that file is the only thing that survives — deliberately, because a report that describes a codebase at a moment should be re-derived rather than trusted twice.

EVALUATION

Every claim in the report carries a file:line anchor, and anything it could not confirm is marked UNVERIFIED. A report with unanchored claims failed, whatever it concluded.

WHERE IT FAILS · 3

  • Summarises the code it read rather than the code the task will touch, which produces a long report that answers nothing.
  • Misses upstream drift when the branch is behind, and every downstream stage then inherits a false premise that still compiles.
  • Over-reads on a small task: the scout stage costs more than the edit for a one-line change.

LESSONS · 1

  • The single most valuable line it produces is 'upstream moved' — because that is the one failure the gates cannot catch.

CONNECTED · 2

SEE THE GRAPH

← ALL AGENTS