Scout

wiki/open-questions.md

Scout - Open Questions

[ ] What are the real fixed monthly costs for hosting, security/firewall, monitoring, support tooling, maintenance, and payment operations?
[ ] What is the measured variable cost per direct scrape, crawl page, rendered page, browser minute, screenshot, LLM extraction, and retained artifact bundle?
[ ] What target gross margin should Scout use for hosted usage?
[ ] Should the first paid model be prepaid credits, a wallet/pay-as-you-go model, or a small subscription plus overage?
[ ] What free hosted allowance is safe enough for beta without inviting abuse?
[ ] At what volume and price does hosted Scout break even and produce profit?

[ ] Is Scout primarily a developer tool, an agent skill, a local API service, or all three?
[ ] Should Scout own synthesis, or should Scout stop at evidence records and let PRISM or another agent synthesize?
[ ] What is the minimum "record quality score" required before a record can be sent to Algolia?
[ ] Should Scout include built-in domain modules, or should modules be plugins?
[ ] For job hunter mode, should Scout own the interactive intake dialogue, or should the host app collect preferences and pass a JobSearchProfile into Scout?
[ ] For PRISM mode, where should Scout stop: source discovery and records only, or should it produce a first-pass evidence summary too?

[ ] Which providers ship in v1: WebFetch host provider, Crawl4AI, browser/session, CDP/profile, saved HTML, PDF parser, ATS adapters? Update (2026-05-22): Crawl4AI/crawler, saved evidence, API scaffolds, and Scout Browser are in the platform foundation. User Browser/CDP remains pending and is now the next priority.
[ ] Should WebFetch/WebSearch be implemented only in skill mode, or represented as an abstract provider that host tools satisfy?
[ ] What is the official fallback order per use case? Update (2026-05-22): Product workflow direction is crawler first, Scout Browser for blocked/sparse/JS-heavy cases, User Browser when the user's browser can view a page that Scout cannot. Other use cases still need explicit per-use-case fallback ladders.
[ ] How should Scout report blocked or unsupported providers without looking like it silently failed? Update (2026-05-22): Browser Workbench runs now write blocked evidence, source entries, screenshots/DOM/text/link artifacts, and failed status for access-denied Scout Browser captures.

[ ] What are the canonical v1 record types?
[ ] Do product records, executive records, investor documents, job records, social posts, and page facts share a base SourceEvidence schema?
[ ] How strict should fields be when different pages expose different attributes?
[ ] How should schema versions be embedded in output artifacts?
[ ] What is the canonical schema for JobSearchProfile, TargetCompanyRecord, and CompanySocialRecord?

[ ] What package name should be used for PyPI?
[ ] Should the standalone CLI include optional extras such as scout[pdf], scout[browser], scout[ats], scout[social]?
[ ] Should the Claude/Codex skill bundle vendor the package, invoke the installed CLI, or import Scout directly?
[ ] How should users configure output directories, credentials, API keys, and provider preferences?

[ ] What is the golden test set for each use case? Update (2026-05-22): Current live UI matrix has 39 targets. It validates workflow mechanics and evidence/artifact presence, but not deep semantic field quality.
[ ] How many sites per use case are enough before claiming a module works?
[ ] What fields must be checked by exact match versus fuzzy confidence scoring?
[ ] How often should live tests run, given websites change and block behavior varies?

[ ] What should Scout refuse, warn on, or require explicit user confirmation for?
[ ] How should Scout handle authenticated pages, personal data, job applications, social platforms, and rate limits? Update (2026-05-22): User Browser mode will require explicit consent and a security boundary because it can access the user's visible browser session. Implementation is pending.
[ ] When should Scout recommend official APIs or provider integrations instead of crawling?
[ ] What provenance must be stored for legal/audit defensibility?