fix(research): align evidence query limits

[iamjr15]

Why

Live production tracing of the research Markdown/PDF acceptance flow found that workflow queries feed three Exa fields: the main query accepts 2,000 characters, while highlightQuery and summaryQuery accept 1,000. The workflow used the broader limit, so detailed topics produced deterministic Zod failures before evidence retrieval.

The same natural prompt also showed the model selecting fan-out for a three-source comparison and sending entities as one combined string.

What changed

• bound workflow-generated evidence queries to the strictest shared 1,000-character contract
• retain the distinct angle/entity suffix when shortening a verbose topic
• describe fan-out entities explicitly as an array of separate short names
• route comparisons of a few sources or standards within one topic to focused research
• reserve fan-out for four or more independent entities or genuinely separate angles

Architecture and migration impact

No database, deployment-topology, or public API changes. This aligns an internal workflow with its downstream provider schema and sharpens agent routing instructions.

Verification

• pnpm lint
• pnpm typecheck
• pnpm turbo build --force
• pnpm deadcode
• pnpm architecture:check
• pnpm turbo skills:build
• parsed maximum-length deep and fan-out queries through ExaSearchInputSchema with query, highlightQuery, and summaryQuery
• asserted all generated queries remain distinct and preserve their angle/entity suffix
• production trace identified the original executeExaSearch Zod failure; the same natural-language flow will be rerun after deployment

The local machine reports the existing Node engine warning (repo pins 24.18.0; local Node is 26.4.0); all checks completed successfully.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment

Project	Deployment	Actions	Updated (UTC)
cheatcode	Skipped		Aug 5, 2026 5:33am