Scan & Snap Phase 0+1 shipped (v0.1.29)

Where this entry picks up

Continues from the Scan & Snap design. The design’s Phase 0 and Phase 1 are now built, tested, and shipped as v0.1.29 (BRAT). This is the first time a user can scan their vault and get drafted rules without writing regex.

Built across three delegated subagents (each verified independently by the main agent), with the release gate — the real-Obsidian E2E run + version bump — kept under direct control. That gate earned its keep: it caught a failure the unit tests couldn’t.

What shipped

Phase 0 — engine refactors (commit 847f58b, no version bump)

Pure, internal, zero user-facing change, so the builder can reuse engine code:

• Exported inferFolderScheme / inferFolderNaming from inferTyped.ts (were private).
• Lifted AXIS_CONVENTIONS (axis→marker table) into a shared src/engine/axisConventions.ts + added slugifyToTagEntry.
• Deduped the copy-pasted emoji/JD normalizer (matchesNormalized et al.) into src/engine/folderNormalize.ts; both detectPacks.ts and detectionTree.ts import it.
• Added a derive → infer → derive round-trip test proving folderAnchor survives when carried explicitly, plus a documented test showing under: anchors are not recoverable from regex shape alone (why the future synthesizer must carry anchors, never re-infer them).
• Known quirk pinned, not fixed: stripEmojiAndJD doesn’t fully normalize emoji+JD folders at inner path segments (Output/📁 01 - Projects/Web → Output/ 01 - Projects/Web). Preserved verbatim (silently “fixing” would shift detection) and locked with a KNOWN QUIRK test. Only affects path-scoped signals on deeply-nested decorated vaults — revisit deliberately when next touching detection.

Phase 1a — the pure planner (commit 937ebce, no version bump)

src/engine/scanAndSnapPlan.ts — buildScanAndSnapPlan(input) turns detected packs + vault folders into CandidateRow[]. For each detected pack’s anchored instance, it scopeRules the pack’s rules onto the anchor and computes per row: coverage (previewRule), bijectivity verdict (layered over template/typed/flag shapes, falling back to 'unknown'), and a conflict analysis computed over candidates ∪ existingRules that distinguishes benign candidate-vs-candidate overlap from the dangerous collides-with-an-installed-rule case, naming the predicted runtime winner via findBestMatch. Plus sortCandidatesByNoise / sortCandidatesByConflict triage helpers. 19 unit tests.

Phase 1b — the modal (commit ef54e65) + bump (7edd517 → v0.1.29)

src/ui/ScanAndSnapModal.ts + the scan-and-snap-draft-rules command (“Scan & snap: draft rules from my vault”). On open: walk folders → detectPacks → load detected packs’ JSON from disk → buildScanAndSnapPlan → render a candidate table. Per the locked decisions: every candidate checked by default, junk (0-match) sorted to the top, read-only (never creates/moves folders). Each row shows a coverage chip, a bijectivity chip (Lossy / 1:1 / Conditional), and a conflict badge (red “Overlaps an existing rule” vs soft “Overlaps another candidate”, with the predicted winner in the tooltip) plus sample folder→tag emissions. “Add N rules” → confirm modal → dedupe-by-id merge into settings.

The E2E gotcha the release gate caught

The first real-Obsidian E2E run failed 4/6: the modal opened but rendered zero candidate rows (empty state). Root cause: the wdio install only copies main.js / manifest.json / styles.css — not the rule-packs/ folder. Detection still works (the manifest is bundled into the build), so the modal opens; but the modal loads pack rules from disk to build candidates, finds none, and shows the empty state.

This is a test-environment gap, not a product bug — real BRAT/community installs ship the full plugin directory, so the pack files are present. The fix (in scan-and-snap.e2e.ts, same approach as typed-model.e2e.ts): the before hook now reads the repo’s rule-packs/*.json (Node side) and stages them into the test vault’s plugin dir via the adapter before the modal runs. Re-run: 6/6 green.

Worth noting: the earlier scope-detect E2E never exercised disk pack-loading (it only renders the detection tree, which needs just the bundled manifest + folder list). Scan & Snap is the first E2E to load pack files on open, so it surfaced the gap. Latent robustness question for later: consider bundling pack rules at build time too (the way the manifest already is), so the feature degrades gracefully even on a stripped install.

Verification (all independently re-run by the main agent)

• bun run build clean · bun test src 843 pass / 0 fail · npm run lint 0 errors (submission-readiness held) · scan-and-snap.e2e.ts 6/6 pass in real Obsidian · no AI attribution in any commit.

Next: Phase 2

The SEACOW axis legend with open “snap” slots — activating the compatibleWith / axes pack metadata that’s currently parsed but consumed by zero app code, so a user sees “your Work axis is an open slot” and can snap a compatible block (PARA / JD / Zettelkasten) onto it, with a hover scope-tint preview. Then Phase 3 (the raw-structure synthesizer) and Phase 4 (establish + save-as-pack). See the design entry for the full plan.