Skip to content

backlog.md

Latest commit

 

History

History
116 lines (91 loc) · 5.74 KB

backlog.md

File metadata and controls

116 lines (91 loc) · 5.74 KB
Raw

Backlog

Forward-looking only: remaining work + recorded decisions. Shipped history lives in CHANGELOG.md and the code — it's not duplicated here. Design notes for the big v2 features are in .claude/memory/project_v2_vision.md.

Nothing below is a committed schedule; order shifts with what becomes urgent.

Status — v1 is feature-complete

The shipping app does everything v1 set out to do (plus a large slice of the original v1.5/v2 plan). What remains is optional polish and future vision — no release blockers.

Near-term polish (nice-to-have)

  • Tag editing — drag-to-reparent & inline rename. Reparenting and rename already work via TagEditDialog (and delete shows an affected-counts warning). Missing only: drag a row onto another to reparent, and rename-in-place on the row instead of opening the dialog. (Today's draggable is tag→activity, not tag→tag.)
  • One-click prune of orphan store folders. The integrity check already detects orphan store/<id>/ dirs (GET /api/admin/integrity); add a guarded "remove" action. (Supersedes the old pruneOrphans no-op.)
  • Video first-frame poster (ffmpeg if on PATH). Videos are accepted and shown, but there's no generated poster frame yet.
  • Optional UI unlock for encryption instead of env vars (TOURGAZE_STORE_*). Lower priority — the env-var model (matros-style) works and is the documented path.

v1.5 — map/AI enrichment (future)

Shipped already: Nominatim reverse-geo + similar-ride tag voting, Open-Meteo weather auto-fill, persisted start/end location, OSM Overpass passes + named peaks. Remaining:

  • LLM-summarised ride description — local Ollama (or Anthropic API behind an opt-in) over stats + weather + tags + locations → 1–2 sentence description; "Suggest description" button on EditTour.
  • Strava-style segment matching — popular climbs auto-named.
  • Wikipedia / Wikidata landmark cards at start/end/waypoints. Hook exists: geo_feature.wikidata is already stored.

v2 — comparison-replay vision

Foundation + ghost-chase replay + media are shipped (route fingerprint, findSimilar, coloured ghosts, finish-together toggle, inline race bars, off-screen edge arrows, gallery, EXIF placement, map media markers). Remaining:

  • Hausrunde auto-detection — cluster activities by start proximity + fingerprint Jaccard; banner "looks like your Hausrunde — tag them all?" on clusters ≥3, then auto-tag future imports once a cluster name is confirmed. (PredictionService already does Hausrunde-style start clustering for geocoding; the cluster-naming/banner feature is the new part.)
  • Long-segment compare (Strava-style) — compare currently spans whole tracks; detect the longest contiguous overlap window and zoom the chart + timeline to it.

v3 — LLM-curated "Hollywood" replay (vision)

Goal: replay-with-media becomes a short film — map glides, camera slows at media points, photo/video fades in (Ken-Burns on stills), caption as subtitle. The smooth-replay pipeline (sub-point interp + camera strategies + dead-end hold) is already in place; this is gated on LLM curation quality.

  • LLM photo description + highlight rank — local vision model (Ollama llava/qwen2-vl) preferred, Anthropic API as opt-in fallback.
  • Auto-cut shortlist — top-N by score, spread across the ride duration.
  • Cinematic overlay during replay — corner photo card + Ken-Burns, PiP for video, caption subtitle (local time via app.timezone).
  • Music sync (stretch) — align cuts to the beat of a chosen track.
  • v3 prep columns — activity_media.highlight_score, caption_ai, media_tags. Add when the LLM pass lands (not present yet).

Guardrails (don't break before v3):

  • The per-activity track JSON files in cache/ are load-bearing for the smooth-coords pipeline the cinematic overlay needs — don't drop them.
  • Encryption must keep covering media under store/<id>/media/, not just tracks.
  • media.captured_at stays UTC; render local time via app.timezone at display.

Out of scope (decisions, not "later")

Consciously dropped — recorded so future sessions don't re-litigate.

  • Auto-tag by region / country / island on import. Geo is derived data, not user intent: materialising region/country as tags polluted the hand-made tag space (epic, commute) with Balearische Inseln / ES noise. The facts already live as first-class activity fields (startLocation/startCountry/ endLocation/endCountry), enough to browse/group by location. Import no longer proposes geo tags. Hook if this ever returns: a hierarchical Location root with rename-safe matching needs a content-derived stable key on tag (a sourceKey, e.g. loc:es:illes-balears) — the row's ULID id can't dedupe re-derived text, and matching by display name breaks on rename.
  • Re-encryption migration when enabling crypto on a populated install. matros has none either: new files follow the current setting, old plaintext stays readable via fallback. Not worth building.
  • 3D extruded terrain (setTerrain) — blacked out the map on some GPUs and let the pitched replay camera clip under the mesh. Ships 2D hillshade only; revisit with camera-collision guards only if there's demand.
  • Dexie / IndexedDB pre-cache of all tracks — on localhost the browser HTTP cache + TanStack Query already cover this. Not worth the ceremony.
  • JWT auth + user passwords — tourgaze is a personal-local app; no auth + optional crypto-at-rest is the right trade-off. Revisit only if network exposure becomes a goal.