[woai3c]

Summary

• Add a Desktop-first workflow for selecting and comparing any two compatible analysis records from History, without requiring a pinned reference.
• Persist versioned capture conditions and fail closed when page health, routes, viewports, or capture settings are not comparable.
• Compare supported design tokens plus conservatively matched layout, interaction-state, and responsive observations, with evidence references and explicit partial-coverage limitations.
• Add reviewable token governance, privacy-safe Desktop persistence, CLI reliability improvements, and standards-compliant MCP stdio behavior.

Desktop comparison

• Provide one Compare two analyses action with a clear earlier/later record picker and detected site names.
• Report each category as changed, unchanged, or inconclusive; partial reliable evidence remains usable instead of invalidating unrelated categories.
• Link results to captured evidence and use conservative cross-capture entity matching. Ambiguous or unmatched entities are excluded rather than reported as design drift.
• Treat layout, interaction-state, and responsive comparison as observed evidence, not pixel regression or approved token-contract rules.
• Allow reviewed, supported token changes to create versioned partial Design Contract revisions with audit history.

Capture reliability and privacy

• Add a versioned Analysis Request and Capture Manifest covering viewports, access/depth/discovery settings, browser/emulation data, stabilization, coverage, and page health.
• Treat maxPages as an upper bound: a site that exposes one healthy page with all planned viewports is complete even when the configured limit is higher. Actual selected-page or viewport failures still fail closed.
• Sanitize persisted, logged, and displayed Desktop URLs while keeping SQLite migrations additive and idempotent.
• Treat legacy records without the required manifest and health evidence as inconclusive.

Shared entry points

• Keep the primary comparison workflow in the Desktop app.
• Improve CLI diagnostics with actual browser-launch validation, explicit browser paths, structured exit codes, cancellation code 130, and parseable stdout.
• Keep MCP stdio output to one JSON-RPC object per line and share the core analysis request schema.

Verification

• pnpm test — 46 files / 300 tests passed
• pnpm test:design-evidence — 2 files / 24 real-browser regression tests passed
• pnpm test:e2e — Desktop and CLI builds plus 68 E2E tests passed
• Desktop TypeScript check passed
• CLI/MCP TypeScript check passed
• ESLint and Prettier checks passed

Scope and limitations

• Comparison is not pixel regression and does not classify observed changes as defects.
• Controlled fixtures validate behavior and fail-closed rules; they do not prove universal accuracy on arbitrary live sites.
• Section and component identity matching remains conservative. Ambiguous and unmatched evidence is not automatically treated as drift.
• Layout, interaction-state, and responsive observations are not approved Design Contract rules.
• Authorized live-site and end-user workflow validation is still required before claiming full product exit criteria.

Reviewer focus

• Capture comparability and fail-closed behavior, especially page-limit upper-bound semantics
• SQLite migration and URL-sanitization idempotence
• Cross-capture entity matching and partial-coverage behavior
• Desktop selection/dialog UX and preload/IPC boundaries
• CLI exit behavior and MCP stdio framing