reader-context-hygienelisted

# Reader Context Hygiene ## What a reader is, and when to invoke this skill A reader (scout) agent ingests an external source — a design board, a live site, a large document — on behalf of a consumer that will make decisions from the result. The reader's return message IS its interface; everything the orchestrator will act on must be *in* that message, and everything it won't act on must stay *out* of it. Invoke this skill when authoring or reviewing any fan-out reader agent, before it ships. This is authoring-time doctrine (like `loop-engineering`): current qa testers follow their own per-scenario contract and are exempt — see the caveat below. ## The minimum bar (MUST) 1. **Bulk to disk, signals inline.** The full artifact is written to disk (a gitignored workspace, or a *subdirectory* of a report directory — never at a path a consumer globs for reports, e.g. `docs/testing/reports/*.md`); the return message carries the path + a 3–5 sentence summary + top-N takeaways **plus every decision-relevant status as a named inline field** (e.g. `items_extracted=214`, `MISSING_DESIGN=<screen>`, `needs_escalation=yes|no`, coverage gaps). A gate flag buried in the artifact file is a gate the orchestrator never sees — the exact failure this contract exists to prevent. 2. **Never inline bulk.** No base64, no full dumps; verbose evidence (screenshots, response bodies, board dumps) is referenced by path. 3. **Fail-closed access.** No access / auth failure → STOP with a diagnostic; never