Offline-first iOS app that aggregates AI/tech news, summarizes articles on-device (Foundation Models), and grows a visual knowledge map of concepts you've learned. Full product spec: TechPulse-Build-Spec.md. UI reference: design/.
Privacy: on-device by default, and no server of ours exists, so the "Data Not Collected" App Store label applies. Every outbound connection, and exactly what each one carries, is enumerated in PRIVACY.md — no passage of what you are reading is on that list, only words you deliberately select. The opt-in exception is your own Claude API key (Settings → AI engine), which unlocks "Go deeper", Explain and Pack generation on hardware without Apple Intelligence: "Go deeper" sends a concept's name, definition and cluster; Explain sends the word you selected plus your Pack's field and Cluster names, so your own map disambiguates the word instead of the article around it (ADR-0006, #29); generating a Pack sends the field name you typed and nothing else, because a Pack is generated to become your map and there is nothing of yours to disambiguate it with (#27). On hardware with Apple Intelligence, all three run on-device: Explain uses the surrounding sentence and stays on the phone, and a Pack is designed a cluster at a time without anything leaving.
- SwiftData models:
FeedSource,Article,Concept,LearningEvent,ConceptLink - Feed pipeline: RSS/Atom/RDF parsing, offline cache, pull-to-refresh, BackgroundTasks refresh
- On-device intelligence: Foundation Models summarization + concept extraction, NaturalLanguage fallback
- Knowledge engine: mastery scoring, embedding-based concept matching, "I know this", time decay
- Force-directed knowledge graph (Canvas + TimelineView) and Swift Charts progress views
- UI test (
Tests/UITests) drives feed → article → concept sheet → graph and saves screenshots to/tmp/techpulse_uitest. Each journey that takes screenshots declares its steps up front: a step that does not run fails the test, and its screenshot has to have been written by that run
Knowledge Pack (skill tree): ~50 pre-seeded AI-engineer concepts in 7 clusters with dependency arrows, cluster overview + detail screens, frontier detection ("ready to learn"), gap detector with feed recommendations, and the staged learning path on Progress.
M6 added: on-device quiz mode (+0.3 mastery rule), resume-seeded knowledge base, Siri App Intent ("What did I learn this week?"), cache pruning (read articles >60 days), UX pass (day groups, unread dots, chip states, related concepts, cluster filters, recenter), and a test suite: 11 Swift Testing unit tests + 2 XCUITest journeys with screenshot capture.
Post-M6 (see DEVLOG.md for dates):
- Atomic Habits reading system — 30 fresh articles/day intake cap, tiny daily goal card (ring + streak + goal-met celebration), goal picker.
- Article concept primer ("know these before you read"), word-level text selection, original-link storage with Safari/Share buttons.
- "Go deeper" — expands any concept into 3–5 linked sub-concepts on its island; on-device by default, or via your own Claude key on hardware without Apple Intelligence.
- Explain — select a word in an article and it joins your map as a Concept. On-device the surrounding sentence disambiguates it; on the BYO-key path your Pack's field and Clusters do, so nothing but the word itself leaves the device (ADR-0006).
- Full-text fetch — articles that arrive as feed snippets pull the real story from the publisher page, then regenerate the on-device summary.
- Semantic zoom on the knowledge graph — zooming spreads positions while dots/lines/labels stay constant screen size, so overview overlap always resolves; collision-culled labels; tap-a-dot glossary strip. Cluster (dependency) trees carry the same zoom, with pinch hint + recenter.
- Parser hardening — XXE / external entities disabled; entitlements file for Keychain access (BYO-key storage).
- Generate a Pack from a field you name — "Marine Biology", "Site Reliability": the app designs the map instead of you picking one, on-device where it can and with your own Claude key where it cannot, saying why rather than failing silently where it can do neither. Model output is never trusted — duplicate names, self-references, dangling dependencies and cycles are scrubbed before the same validator an imported Pack faces sees it, and you review the draft (dropping or renaming anything that does not belong) before it becomes your map. A generated Pack is labelled Generated wherever your Pack is named.
- Suggested Sources are probed before you get them — accepting a Pack's suggestions asks each one you ticked whether it is a Source at all, so a Pack carrying a dead link or a homepage does not put one in your list. A host that merely refused or timed out has said nothing about its link, so that one is still added — and then says so on its own row if the trouble persists.
- Visible Source health — a sync records what each fetch actually was, so a Source that is being throttled, that has stopped answering, or that has published nothing for six months says so on its Settings row — with what it last answered and how much of it you still have to read offline — instead of quietly producing an empty Cluster.
- Adaptive dark mode — every theme token carries a light/dark pair; mastery colors stay constant in both modes (color-blind-safe lightness order preserved).
Built for iOS 26 / Swift 6.2 / Xcode 26.3 (the spec's iOS 27 targets are adapted down; revisit reorderable containers + LanguageModel protocol when Xcode 27 ships).
xcodegen generate # regenerates TechPulse.xcodeproj from project.yml open TechPulse.xcodeproj # or: xcodebuild -scheme TechPulse -destination 'generic/platform=iOS Simulator' build
The .xcodeproj is generated — edit project.yml, not the project file.