| name | scout |
| description | Context-loading discipline for codebases. Use BEFORE opening, reading, or exploring files — and for monorepo/broad greps (TODO sweeps, 'where is X'): search first, read narrow, one repo map per session. Pair with dispatch path A when the sweep would pollute main context. Prevents the 20k-token 'let me look around' burn. |
scout — Context Discipline
The cheapest token is the one never loaded. Explore structure, not content.
The ladder — always start at the top
- Memory first. Project already known? Ask the memory MCP (
load_project
or search) before touching the filesystem — a 300-token briefing beats a
20k-token re-exploration.
- Map second. Need repo orientation? Build/read ONE overview: file tree
(
ls/glob), or a generated map (codesight CONTEXT.md, aider-style
repo-map) if available. Cache it mentally — never regenerate mid-session.
- Search third. Locate by symbol/pattern (grep, ast-grep, Serena
find_symbol) — get file:line, not file contents.
- Read last, read narrow. Open only the located range (offset/limit around
the hit). Full-file reads only when the file is small or the task is a
whole-file rewrite.
Hard rules
- Never read a file to find something — that's what search is for.
- Never re-read a file you just edited — you know what's in it.
- Never load generated artifacts (lockfiles, dist/, minified, fixtures)
unless they ARE the subject of the task.
- Broad sweep needed? (many files, unknown naming, "where is X handled?")
→ don't do it inline; hand it to
dispatch path A and receive the conclusion
only. External CLI coding workers with a team-up roster → dispatch path B
(mailbox + watcher), not inline exploration.
- Docs/logs over ~200 lines: extract the relevant slice (grep the error ID,
read the section) — never the whole thing.
Symbol work
If the Serena MCP (or an LSP-backed tool) is installed, symbol-level lookups and
edits go through it: find_symbol / find_references / targeted insert replace
an 8-step grep-and-read loop with one call. Scout owns the overview; Serena
owns the symbols. Never run both for the same question.
Budget instinct
Before any read, estimate: "does the value of this content exceed its rent?"
Every loaded token is paid again on every subsequent turn until compaction.
A 5k-token file read that answers a yes/no question is a bad trade — search
for the one line that answers it.
