scout

An agent dropped into an unfamiliar directory burns 50–200K frontier tokens just orienting itself. scout does the orientation instead: it walks a corpus of code or documents, indexes it with a fast open-weight model, and answers “what’s here” and “what does it say about X” in seconds for cents. The extraction model never gets to assert an unverified claim — if a fact can’t be matched to a real quote in a real file, it doesn’t ship as a finding.

1. 01 · hydrate

   Walk the corpus

   Skips harness litter and secrets, then builds a two-tier card per file: a machine-verified skeleton — symbols via ctags, import edges, git churn, heading outlines — plus a model summary labeled a hint, never a fact. Cards are hash-keyed; only changed files cost money on the next run.

2. ›

3. 02 · route

   Ask a question

   A query routes against the index to a candidate file set, then a fast open-weight model (Cerebras gemma-4-31b) extracts candidate facts per chunk — the orientation an agent would otherwise burn 50–200K frontier tokens gathering by hand.

4. ›

5. 03 · verify

   Quote firewall

   Every finding must carry a verbatim quote that machine-checks against the source file at the claimed line. Nothing else ships. Kill-tested: 6.7% raw hallucination pressure on unfirewalled model output, 0% once the firewall runs.

   illustrative example

   ✕“uses exponential backoff with jitter” — no matching quote, dropped

   ✓“max_attempts = 3” — quote matches retry.rs:42, delivered

6. ›

7. 04 · deliver

   Cited, in seconds

   Every surviving finding is file:line addressable — cents of inference standing in for minutes of subagent exploration, with a quote attached instead of a promise.