[Dunky13]

Summary

Learned workloads opens on projects, then operations, then the existing group list, instead of one flat list of 1,800 groups. This is stacked on #79, which adds the catalog API it reads.

Projects (busiest first; ≤2 workloads fold into "Other projects")
  └─ Operations in <project>: All operations, Tests, Build, Lint, …
       └─ Groups (existing paged list, selection, details, merge)
Search: always across every project, flat list

 remote/static/profiles.js     levels, breadcrumb, focus/announce, failed-load cleanup
 native/AgentLearning.swift    same levels, Back + breadcrumb buttons, submitted query kept apart
 internal/backend
+  profileProject → NFC          one key for rows, folded set and filter
+  CLI string flags → NFC        Foundation Process passes them decomposed
+  search matches member projects, ranked after label/id matches
 internal/preview               fixture projects + --hierarchy/--project/--operation/--other-projects

Edge cases handled:

• A project folded away by a merge lists its groups directly; Back returns to Projects.
• A page past the end after a merge re-reads the last page once.

Evidence

• Before: one flat list (before-desktop-catalog.png, before-phone-catalog.png, before-native-catalog.png).
  After: projects, operations and groups at desktop 1280×900 and phone 390×844, plus native projects and groups (after-*-catalog, after-*-operations, after-*-groups, after-*-search, after-*-detail, after-native-projects, after-native-groups). Screenshots are kept out of Git and still need to be attached here.
• Checks on this content, all exit 0:
  • go test -p 2 -parallel 2 ./internal/backend ./internal/preview
  • npm run test:web and lint
  • Playwright: 35 passed (--workers=1, rebuilt preview)
  • scripts/test-native.sh
• Mutation checks:
  • Removing the search ranking fails TestProfileSearchRanksOwnLabelBeforeMemberProject.
  • Reverting the native refresh after a merge fails "A merge that folds the project at the operations level lists its groups instead of a false empty state".
  • Removing the first-open guard fails the job-link test.
• Reviews: two independent rounds (Opus 5.5). All confirmed findings are fixed except the limits listed in docs/handoffs/learning-browse.md.
• Hosted CI runs after fix(learning): stop discarding runs and undercounting CPU #78 and feat(learning): group workloads by repository, operation and tool #79 merge and this PR is retargeted to main.

Merge Danger

Door: two-way

UI and read-path only. No schema or stored data changes. Project names stored in NFD are now returned in NFC, which looks identical.

Blast Radius: learning-browse

Learned workloads on the phone dashboard and the native app. The web merge-target pickers still show the first 100 results; label matches now come first.

🤖 Generated with Claude Code