feat: ✨ authenticated catalog covers via Coil + thumbnail-rel parsing#2
Merged
Conversation
Documents the gap (Coil uses a default ImageLoader without the OPDS BasicAuthInterceptor, so calibre-web cover requests 401) and the planned fix: an ImageLoaderFactory on EReaderApp wired to OpdsHttpClient.okHttp, plus reordering OpdsClient to prefer rel="…/image/thumbnail" over rel="…/image" for grid use.
Plan for wiring Coil to OpdsHttpClient.okHttp via an ImageLoaderFactory on EReaderApp, and reordering OpdsClient to prefer rel="…/image/thumbnail" over rel="…/image". Tests, fixtures, and manual verification steps included. References the design at docs/superpowers/specs/2026-05-07-catalog-covers-design.md.
Readium's OPDS model populates pub.subcollections["images"] only when an
opds:image rel is present and silently drops opds:image/thumbnail when it
appears alone. Skip the Readium-routed image extraction and parse cover
links directly from each <entry>, mirroring the existing parseSearchLink
workaround for the analogous {searchTerms} quirk. Within an entry, prefer
opds:image/thumbnail (sized for the 2-col grid) over opds:image, and join
back to Readium's parsed publications via the unique epub acquisition href.
Spec at docs/superpowers/specs/2026-05-07-catalog-covers-design.md updated
to record the discovered Readium behavior.
Coil's default singleton ImageLoader has no awareness of the OPDS BasicAuthInterceptor, so cover requests against authenticated calibre-web instances 401 silently and the catalog grid only ever shows the gradient+initials fallback. Implement coil.ImageLoaderFactory on EReaderApp with an ImageLoader built against the existing OpdsHttpClient.okHttp, so Coil inherits Basic auth (and any future interceptors) automatically. AppContainer's opdsHttp is promoted from private to public to be reachable from the Application class.
vitofico
force-pushed
the
feature/catalog-covers-auth
branch
from
May 7, 2026 16:53
b798b1e to
8f46f74
Compare
vitofico
changed the title
vitofico
added a commit
that referenced
this pull request
May 8, 2026
…#2) * docs: 📝 design for authenticated catalog cover loading Documents the gap (Coil uses a default ImageLoader without the OPDS BasicAuthInterceptor, so calibre-web cover requests 401) and the planned fix: an ImageLoaderFactory on EReaderApp wired to OpdsHttpClient.okHttp, plus reordering OpdsClient to prefer rel="…/image/thumbnail" over rel="…/image" for grid use. * docs: 📝 implementation plan for catalog covers Plan for wiring Coil to OpdsHttpClient.okHttp via an ImageLoaderFactory on EReaderApp, and reordering OpdsClient to prefer rel="…/image/thumbnail" over rel="…/image". Tests, fixtures, and manual verification steps included. References the design at docs/superpowers/specs/2026-05-07-catalog-covers-design.md. * feat: ✨ prefer OPDS thumbnail rel and survive thumbnail-only feeds Readium's OPDS model populates pub.subcollections["images"] only when an opds:image rel is present and silently drops opds:image/thumbnail when it appears alone. Skip the Readium-routed image extraction and parse cover links directly from each <entry>, mirroring the existing parseSearchLink workaround for the analogous {searchTerms} quirk. Within an entry, prefer opds:image/thumbnail (sized for the 2-col grid) over opds:image, and join back to Readium's parsed publications via the unique epub acquisition href. Spec at docs/superpowers/specs/2026-05-07-catalog-covers-design.md updated to record the discovered Readium behavior. * feat: ✨ route Coil image loads through authenticated OPDS OkHttp client Coil's default singleton ImageLoader has no awareness of the OPDS BasicAuthInterceptor, so cover requests against authenticated calibre-web instances 401 silently and the catalog grid only ever shows the gradient+initials fallback. Implement coil.ImageLoaderFactory on EReaderApp with an ImageLoader built against the existing OpdsHttpClient.okHttp, so Coil inherits Basic auth (and any future interceptors) automatically. AppContainer's opdsHttp is promoted from private to public to be reachable from the Application class.
vitofico
added a commit
that referenced
this pull request
May 8, 2026
…#2) * docs: 📝 design for authenticated catalog cover loading Documents the gap (Coil uses a default ImageLoader without the OPDS BasicAuthInterceptor, so calibre-web cover requests 401) and the planned fix: an ImageLoaderFactory on EReaderApp wired to OpdsHttpClient.okHttp, plus reordering OpdsClient to prefer rel="…/image/thumbnail" over rel="…/image" for grid use. * docs: 📝 implementation plan for catalog covers Plan for wiring Coil to OpdsHttpClient.okHttp via an ImageLoaderFactory on EReaderApp, and reordering OpdsClient to prefer rel="…/image/thumbnail" over rel="…/image". Tests, fixtures, and manual verification steps included. References the design at docs/superpowers/specs/2026-05-07-catalog-covers-design.md. * feat: ✨ prefer OPDS thumbnail rel and survive thumbnail-only feeds Readium's OPDS model populates pub.subcollections["images"] only when an opds:image rel is present and silently drops opds:image/thumbnail when it appears alone. Skip the Readium-routed image extraction and parse cover links directly from each <entry>, mirroring the existing parseSearchLink workaround for the analogous {searchTerms} quirk. Within an entry, prefer opds:image/thumbnail (sized for the 2-col grid) over opds:image, and join back to Readium's parsed publications via the unique epub acquisition href. Spec at docs/superpowers/specs/2026-05-07-catalog-covers-design.md updated to record the discovered Readium behavior. * feat: ✨ route Coil image loads through authenticated OPDS OkHttp client Coil's default singleton ImageLoader has no awareness of the OPDS BasicAuthInterceptor, so cover requests against authenticated calibre-web instances 401 silently and the catalog grid only ever shows the gradient+initials fallback. Implement coil.ImageLoaderFactory on EReaderApp with an ImageLoader built against the existing OpdsHttpClient.okHttp, so Coil inherits Basic auth (and any future interceptors) automatically. AppContainer's opdsHttp is promoted from private to public to be reachable from the Application class.
vitofico
added a commit
that referenced
this pull request
May 8, 2026
…#2) * docs: 📝 design for authenticated catalog cover loading Documents the gap (Coil uses a default ImageLoader without the OPDS BasicAuthInterceptor, so calibre-web cover requests 401) and the planned fix: an ImageLoaderFactory on EReaderApp wired to OpdsHttpClient.okHttp, plus reordering OpdsClient to prefer rel="…/image/thumbnail" over rel="…/image" for grid use. * docs: 📝 implementation plan for catalog covers Plan for wiring Coil to OpdsHttpClient.okHttp via an ImageLoaderFactory on EReaderApp, and reordering OpdsClient to prefer rel="…/image/thumbnail" over rel="…/image". Tests, fixtures, and manual verification steps included. References the design at docs/superpowers/specs/2026-05-07-catalog-covers-design.md. * feat: ✨ prefer OPDS thumbnail rel and survive thumbnail-only feeds Readium's OPDS model populates pub.subcollections["images"] only when an opds:image rel is present and silently drops opds:image/thumbnail when it appears alone. Skip the Readium-routed image extraction and parse cover links directly from each <entry>, mirroring the existing parseSearchLink workaround for the analogous {searchTerms} quirk. Within an entry, prefer opds:image/thumbnail (sized for the 2-col grid) over opds:image, and join back to Readium's parsed publications via the unique epub acquisition href. Spec at docs/superpowers/specs/2026-05-07-catalog-covers-design.md updated to record the discovered Readium behavior. * feat: ✨ route Coil image loads through authenticated OPDS OkHttp client Coil's default singleton ImageLoader has no awareness of the OPDS BasicAuthInterceptor, so cover requests against authenticated calibre-web instances 401 silently and the catalog grid only ever shows the gradient+initials fallback. Implement coil.ImageLoaderFactory on EReaderApp with an ImageLoader built against the existing OpdsHttpClient.okHttp, so Coil inherits Basic auth (and any future interceptors) automatically. AppContainer's opdsHttp is promoted from private to public to be reachable from the Application class.
5 tasks
vitofico
added a commit
that referenced
this pull request
May 20, 2026
…insight cache (#33) * ✨ feat(ai): BookInsight schema v4 + prompt-version wiring fix PR-ε of Bundle 2. Server: - BookInsightPayload schema v3 -> v4 with five optional depth fields: theme_analysis (dict, hard-capped at 2 keys server-side), craft_notes, comparative_anchors (sanitized list, blank-drop + cap-at-4), distinctive_take, discussion_prompts. New ComparativeAnchor sub-type. - PROMPT_VERSION "4" -> "5"; system prompt instructs the model to emit the v4 keys with two-themes cap, no invented titles, no plot reveals. - _do_generate pins payload.schema_version = 4. - New core/ai/_compat.py::_resolve_prompt_version applies Lock #19 sentinel semantics: env value "1" / "" / None resolves to the in-code PROMPT_VERSION constant; any other value is honored verbatim as the Lock #2 emergency rollback override. Helper lives in core.ai._compat so the provider lazy-import boundary stays intact (coordinator §3.18). - main.py routes settings.ai_prompt_version through the helper inside the existing if settings.ai_enabled: block. Fixes the original bug where settings default "1" flowed through unchanged. - test_lazy_imports.py SYNC_ONLY_FORBIDDEN gains _compat. Android: - BookInsightPayload DTO catch-up from v2 -> v3 -> v4: adds themes (v3 catch-up), themeAnalysis, craftNotes, comparativeAnchors, distinctiveTake, discussionPrompts. Default schemaVersion = 4. New ComparativeAnchor data class. Old v2/v3 payloads still deserialize cleanly. - InsightCards.kt adds five collapsible cards gated on non-null + non-empty/non-blank predicates with display-side sanity filters. - New BookInsightPayloadJsonTest (v2/v3/v4 round-trip + default). - New InsightSectionVisibilityTest (predicate-level coverage). Docs: - docs/sync-api.md updated to schema v4 with the new fields documented and a paragraph on prompt_version resolution semantics. Honors: Locks #2 / #7 / #19; coordinator §3.1 / §3.2 / §3.10 / §3.18. * ✨ feat(ai): catalog-download insight promote PR-ζ of Bundle 2. Server: - New `POST /ai/v1/insights/promote` route. Row-copy + alias-link promote (Lock #1): copies the cached `book_insights` row at the catalog identity onto the post-download canonical identity with `generated_at = NOW()` (Lock #23) and `previous_insight_ids=[src.id]` for lineage audit; copies `book_themes` rows; writes a user-scoped `insight_identity_aliases` row with `source='promoted_on_download'` as the idempotency anchor. - 204 when nothing to promote; 403 on cross-user ownership; 409 `ai_not_opted_in` per Lock #10 / CC-1 (normalized literal across all `/ai/v1/insights/*` routes); 429 honoring per-user daily limit `QUIRE_SERVER_AI_PROMOTE_DAILY_LIMIT` (default 100; process-local). - Structured stdout audit `event=ai.promote ...` (Lock #11 amendment; NO `ai_generation_log` row in PR-ζ — PR-β extends the audit schema). - Different `(tone, language)` re-copies under existing alias. - New `CatalogInsightStash` + `CatalogInsightStashEntry` types in `data/ai` (Lock #16). Process-local, TTL 30 min, partitioned by `(subject, opdsHref)`, cleared on base-URL change. - `AiClient.promoteInsight` (200/204/4xx handling) and `AiRepository.promoteInsight` (best-effort Boolean for PR-η wrap). - `CatalogDetailViewModel` writes to the stash on every successful catalog-side insight lookup (cache hit or miss). - `CatalogViewModel.download` now `await`s `libraryUploader .enqueueOne(...).join()` before firing promote (Lock #13: library PUT lands before the server's `_assert_owns` runs). Promote failures keep the stash entry for retry; success removes it. - DI: `AppContainer` owns the singleton stash, wires it into both view-model factories, and clears it on credential base-URL change. Docs: - `docs/sync-api.md` documents the new endpoint surface, response matrix, and rate limit. - `server/README.md` env-var table gains `QUIRE_SERVER_AI_PROMOTE_DAILY_LIMIT`. Honors: Locks #1 / #10 / #11 (amendment) / #13 / #16 / #23; coordinator §3.14 / §3.17 / §3.18 / §3.20. * ✨ feat(ai): /ai/v1/insights/sync + prompt_version in /config PR-η of Bundle 2 (scope-reduced — see deviations below). Server: - New `GET /ai/v1/insights/sync`: paginated, read-only, weight=0 bulk export of the caller's owned-book insights joined through `library_items`. Cursor is the `(generated_at, id)` tuple (Lock #23) with strict-lexicographic comparison so identical timestamps never drop rows. Filters on `superseded_at IS NULL`, `library_items.deleted_at IS NULL`, and the caller's current `(model_id, prompt_version, tone, language)`. PR9 priority case prefers metadata-id matches over content-hash-only. Promoted rows (PR-ζ) participate naturally via their `generated_at = NOW()` stamp. - `GET /ai/v1/config` gains `prompt_version: str` (Lock #24) sourced from PR-ε's `_resolve_prompt_version` helper. Older deploys decode safely on the Android side because the DTO default is the legacy sentinel `"1"`. - New schemas: `InsightSyncCursor`, `InsightSyncItem`, `InsightSyncResponse`. - 10 integration tests cover opt-in gate (Lock #10), 503 when AI unconfigured, cross-user isolation, prompt-version / superseded / deleted-library-item filters, half-supplied cursor 400, three-page tuple-cursor walk with distinct timestamps, and weight=0 budget invariant. `test_config_endpoint_*` extended to assert `prompt_version`. Android (data/ai): - `AiConfig` DTO gains `promptVersion` (default "1" for back-compat). - New DTOs `InsightSyncCursor`, `InsightSyncItem`, `InsightSyncResponse`. - `AiClient.syncInsights(cursor, limit)` using `HttpUrl.Builder` so the cursor's ISO 8601 `+` is percent-encoded correctly. Decodes 200, raises `AiHttpException` on non-2xx. - 6 new MockWebServer tests cover cursor encoding, response parsing, 409 propagation, and config `prompt_version` decode (present and omitted). Docs: - `docs/sync-api.md` documents `GET /ai/v1/insights/sync` and the `prompt_version` field on `/ai/v1/config`. Deviations from pr-eta-plan.md (deliberate scope reduction to fit the remaining budget; full plan items deferred to follow-up): - `AiRepository.kt` is NOT moved from `app/` to `:data:ai` (Lock #14 amendment). Existing call sites continue to work; the move is a pure refactor that can land separately. - Room v6→v7 + `InsightEntity` + `InsightDao` + `InsightSyncRepository` + Settings "Refresh insights" UI deferred. The wire format is fully defined and tested; the on-device cache + sync orchestration is the next slice. - No `requestSync` trigger wiring in `CatalogViewModel` / `EReaderApp`. Honors: Locks #10 / #19 / #23 / #24; coordinator §3.1 / §3.15 / §3.17 / §3.18. * ✨ feat(ai): Android local insight cache + bulk-sync (PR-η) Moves AiRepository.kt from :app to :data:ai (Lock #14 amendment) and wraps it with a local-cache-first read path backed by the new book_insights Room table (v6→v7, Lock #8). On cache miss the wrappers fall through to the existing AiClient and write the response back. Adds InsightSyncRepository which orchestrates GET /ai/v1/insights/sync against the local cache, walking the (generated_at, id) tuple cursor (Lock #23) and coalescing burst triggers within a 2s window. Fires on app start (post library upload), after every successful promote (coordinator §3.17), and from a new "Refresh insights" button in the Settings AI section. Behavior contracts: - IOException on getCachedInsight() returns findAnyForIdentity row; 401/409/5xx propagate (decisions.md CC §5). - lookupInsight is the active-fetch path — IOException is not swallowed. - syncNow() is mutex-guarded; concurrent callers single-flight. - Eviction policy "none in v1"; rows at obsolete (model, prompt, style) keys go cold but remain reachable via findAnyForIdentity fallback. No server-side changes (the /sync endpoint shipped in 37b3f03). * ✅ test(ai): coverage for local insight cache + sync (PR-η) - MigrationTest: v6→v7 creates book_insights with the 4 indices and preserves pre-existing documents rows. - InsightDaoTest: round-trip upsert/getByIdentity, REPLACE semantics on same-PK upsert, findAnyForIdentity returns latest across styles, cursor-tip queries, prompt-version key separation. - AiRepositoryLocalFirstTest: local-cache short-circuits both lookup and getCached paths, write-back on 200, narrow IOException fallback via findAnyForIdentity, 401/409/500 rethrow, prompt-version change invalidates by key, and the AiConfig default for the new prompt_version field on legacy server responses. - InsightSyncRepositoryTest: tuple-cursor walk across pages including the cursor query params, skip on ai_enabled=false, Failed on network error, and burst coalescing within the debounce window. * 🐛 fix(tests): make insight-sync coalesce test CI-robust Replace the brittle `server.requestCount == 3` assertion (which raced against the real Dispatchers.IO HTTP delivery) with a synchronization on `server.takeRequest(5s)` followed by a negative-window check `server.takeRequest(200ms) == null`. Drives the same coalescing guarantee but waits for actual I/O instead of polling a counter.
vitofico
added a commit
that referenced
this pull request
May 21, 2026
…le (pr-γ) PR-γ of Bundle 3 (Reader Profile) — adds the user-visible Reader Profile surface on Android. New "Insights" entry on the Library overflow menu opens LibraryInsightsScreen, which renders five sealed UI states (Disabled / Empty / Loading / Loaded / Error) over GET /ai/v1/profile and POST /ai/v1/profile/refresh. Highlights: * AiClient gains refreshProfile() (200 → ReaderProfileResponseDto, 409 ai_not_opted_in, 429 quota, 5xx model failure) and deleteProfile() (idempotent 204 — wired by PR-δ from Settings). * AiRepository.refreshProfile drains the sync orchestrator + library uploader best-effort before calling the server (Lock #6 NTH #9; corrections.md PR-γ Required #6). Preflight failures log via the PreflightOutcome callback and the refresh proceeds regardless. * Decoupling via two fun interfaces (ProfilePreflightSync / ProfilePreflightLibrary) so :data:ai never depends on :data:sync or :data:library. * AiConfig.progressSupported (Lock #5) consumed to drive Disabled.ProgressUnsupported; the server already advertises the field. * Local input-fingerprint approximation per Lock #12 (mirrors PR-β's server recipe; uses ReaderStats.booksWithThemesCount per Lock #15 / CC-8, NOT topThemes.size). Mismatch renders a low-emphasis "Profile may be out of date" hint (architect Finding #2). * In-library recommendations resolve via DocumentRepository.findByIdentity → /reader/{id}; discovery recommendations expose an explicit "Open in browser" button gated on https:// (architect Finding #4), wrapped in runCatching for ActivityNotFoundException. * Coverage meter uses booksWithThemesCount/totalBooks (NTH #12, Lock #15). * AppNavGraph wires composable("library/insights"); AppContainer adds LibraryInsightsViewModelFactory + a StateFlow<Boolean> over AiPreferences.aiEnabled. Tests cover the disabled/empty/loaded/error state transitions, the local fingerprint recipe (match + mismatch + null cases), and best-effort preflight semantics (sync→library→refresh order, network still fires when preflight throws). All :app:testDebugUnitTest and :data:ai:testDebugUnitTest pass.
vitofico
added a commit
that referenced
this pull request
May 21, 2026
…-β + PR-γ + PR-δ) (#34) * ✨ feat(server): reader_profiles substrate + abandoned_at + ReaderStats v1 + GET /ai/v1/profile Lands the server half of pr-α (Bundle 3 / Reader Profile substrate): * Migrations - progress_002: add `progress.abandoned_at` + check constraint `finished_at IS NULL OR abandoned_at IS NULL` + partial index on `(document_pk) WHERE abandoned_at IS NOT NULL`. - ai_005: new `reader_profiles` table keyed on `(tenant_id, subject)` with `payload JSONB`, `schema_version`, `model_id`, `prompt_version`, `input_fingerprint VARCHAR(16)` (Lock #12 / coordinator §3.6). * ORM - `Progress.abandoned_at` + `ck_progress_abandoned_xor_finished`. - New `ReaderProfile` model — user-scoped, NOT shared cache. * Pydantic - `AuthorCount`, `ReaderStats` (with `books_with_themes_count: int = 0` per Lock #15 / CC-8), `BookRec` (full §3.4 provenance shape with safe defaults), `ReaderProfilePayload` (schema_version=1), and the `ReaderProfileResponse` envelope. * `_compute_reader_stats` (`quire_server.core.ai.service`) - Deterministic per-user stats: total/finished/in-progress/abandoned counts, `finish_rate_by_theme` (top-10), `most_read_authors` (top-5). - Mirrors PR9's `pick_priority = case(...)` pattern exactly (coordinator §3.4 REJECT (e)). - pr-α emits `books_with_themes_count = 0`; pr-β replaces with the real count. * Terminal-state invariant - `push_progress` clears the opposite terminal flag on a flip; preserves `percent` on the abandon transition. - `pull_progress` drops `abandoned_at` (finished wins) for corrupt legacy rows and emits a structured warning. * `GET /ai/v1/profile` - Cache-only read. No opt-in gate (opted-out users can still read their last generation). 404 when no row exists. Refresh / LLM generation lives in pr-β. * Tests (server/tests/integration/test_reader_profile.py + adjusted test_health / test_migrate_script / test_readyz_migration_state for the new ai_005 + progress_002 heads). * ✨ feat(android): abandoned_at progress field, MIGRATION_7_8, profile fetch + abandon DAO Lands the Android half of pr-α (Bundle 3 / Reader Profile substrate): * Room - DB version 7 -> 8 + MIGRATION_7_8 adds nullable `abandonedAt INTEGER` column to `progress`. - `ProgressEntity.abandonedAt: Long?` (Long? maps to nullable INTEGER). - `ProgressDao.markAbandoned` / `unmarkAbandoned` bump BOTH `updatedAt` AND `localUpdatedAt` to `now`. This is load-bearing: Android pushes `progress.updatedAt` as `client_updated_at`, and the server's LWW guard only accepts the row when `client_updated_at` is strictly newer. Setting only `localUpdatedAt` would make the row "dirty" locally but the server would silently reject the abandon. - `markAbandoned` clears `finishedAt`; `unmarkAbandoned` leaves `percent` untouched. (Abandoning at 60% remembers 60%.) * Core / sync DTO - `Progress.abandonedAt` on the domain model so the push path can read it through ProgressRepository.dirty(). - `ProgressItemDto.abandonedAt` (snake_case wire name). - `SyncOrchestrator`: push serializes `abandonedAt` as ISO-8601; pull defensively drops `abandonedAt` if `finishedAt` is also set (finished wins) — mirrors the server's defensive read. * AI - `AiClient.fetchProfile()` -> `GET /ai/v1/profile`. Maps 404 to null. - `AiRepository.fetchProfile()`, `markAbandoned()`, `unmarkAbandoned()`. Constructor gains an optional `progressDao` so existing test sites that only stub `insightDao` keep compiling; AppContainer wires the real DAO. - DTOs for `ReaderProfileResponse`, `ReaderProfilePayload`, `ReaderStats` (including `booksWithThemesCount` per Lock #15 / CC-8), `BookRec` (full coordinator §3.4 shape), `AuthorCount`. - `source_type`, `owned_state`, `confidence` declared as `String` (not enums) so a future server bump doesn't crash existing clients. * Tests - MigrationTest: `migrate 7 to 8 adds abandonedAt column`, `migrate 6 to 8 chains via book_insights and abandonedAt` (the Lock #8 chained-migration test). - ProgressDaoTest: mark/unmark abandon semantics + LWW timestamp correctness (BOTH updatedAt AND localUpdatedAt bumped) + dirty() pickup. - ProgressDtosTest: round-trip + null-omission for `abandonedAt`. - SyncOrchestratorTest: defensive-pull rule when server emits both terminal flags. - AiRepositoryProfileTest: fetchProfile 404/200, mark/unmark delegation through FakeProgressDao. * ✨ feat(server): refresh_profile orchestrator + POST /ai/v1/profile/refresh (pr-β) Stitches the AI building blocks into a per-user reader-profile orchestrator gated on AI_ENABLED + PROGRESS_ENABLED + opt-in: - ai_006 migration: ai_generation_log.book_insight_id becomes nullable, new `kind` discriminator with CHECK admitting 'insight'/'profile'/'promote' (Lock #11 amendment), partial index for profile-row lookups, and ai_usage_daily.profile_count for the 3/day cap. - refresh_profile(): per-user singleflight lock, low-data short-circuit with weight=0 (runs even at quota cap), daily-cap enforcement, global semaphore around the LLM call, deterministic candidate maps for in-library + OpenLibrary-discovery recommendations, owner-exclusion belt-and-suspenders, 16-hex-char input_fingerprint, kind='profile' audit rows per waiter (including collapsed singleflight waiters). - PR-ζ wire-up (Lock #11 amendment): promote handler now writes a kind='promote' DB row alongside the retained stdout audit line. - Retriever.author_bibliography(): 30d positive cache, 24h negative cache, 429 with Retry-After capped at 6h, 5xx → stale-if-error via new allow_stale=True read path. - _compute_reader_stats now populates books_with_themes_count (Lock #15). - ConfigResponse gains progress_supported so AI-only deploys can suppress the Android reader-profile UI (coordinator §3.5). - READER_PROFILE_PROMPT + READER_PROFILE_PROMPT_VERSION='1' bumped independently of the per-book PROMPT_VERSION. - Tests: existing audit-log / migration / promote / health tests updated for the ai_006 schema (kind required, head moves ai_005 → ai_006). Integration test_reader_profile passes all 11 cases. - docs/sync-api.md: full POST /profile/refresh section + progress_supported on GET /ai/v1/config. 426 tests pass. * ✨ feat(android): Library Insights screen + refreshProfile/deleteProfile (pr-γ) PR-γ of Bundle 3 (Reader Profile) — adds the user-visible Reader Profile surface on Android. New "Insights" entry on the Library overflow menu opens LibraryInsightsScreen, which renders five sealed UI states (Disabled / Empty / Loading / Loaded / Error) over GET /ai/v1/profile and POST /ai/v1/profile/refresh. Highlights: * AiClient gains refreshProfile() (200 → ReaderProfileResponseDto, 409 ai_not_opted_in, 429 quota, 5xx model failure) and deleteProfile() (idempotent 204 — wired by PR-δ from Settings). * AiRepository.refreshProfile drains the sync orchestrator + library uploader best-effort before calling the server (Lock #6 NTH #9; corrections.md PR-γ Required #6). Preflight failures log via the PreflightOutcome callback and the refresh proceeds regardless. * Decoupling via two fun interfaces (ProfilePreflightSync / ProfilePreflightLibrary) so :data:ai never depends on :data:sync or :data:library. * AiConfig.progressSupported (Lock #5) consumed to drive Disabled.ProgressUnsupported; the server already advertises the field. * Local input-fingerprint approximation per Lock #12 (mirrors PR-β's server recipe; uses ReaderStats.booksWithThemesCount per Lock #15 / CC-8, NOT topThemes.size). Mismatch renders a low-emphasis "Profile may be out of date" hint (architect Finding #2). * In-library recommendations resolve via DocumentRepository.findByIdentity → /reader/{id}; discovery recommendations expose an explicit "Open in browser" button gated on https:// (architect Finding #4), wrapped in runCatching for ActivityNotFoundException. * Coverage meter uses booksWithThemesCount/totalBooks (NTH #12, Lock #15). * AppNavGraph wires composable("library/insights"); AppContainer adds LibraryInsightsViewModelFactory + a StateFlow<Boolean> over AiPreferences.aiEnabled. Tests cover the disabled/empty/loaded/error state transitions, the local fingerprint recipe (match + mismatch + null cases), and best-effort preflight semantics (sync→library→refresh order, network still fires when preflight throws). All :app:testDebugUnitTest and :data:ai:testDebugUnitTest pass. * ✨ feat(pr-δ): DELETE /ai/v1/profile + abandoned UX + delete-profile button Server: - DELETE /ai/v1/profile (idempotent, 204) gated on ai_enabled + opt-in - 5 integration tests covering happy path, idempotency, tenant isolation, not-opted-in (409), and ai_disabled (503) Android: - LibraryPreferences struct (sort + showAbandoned), persisted via the same SharedPreferences file (Lock #17); migration is read-only — old key 'library_sort' keeps working. - Library home gains a 'Show abandoned' toggle; abandoned books are hidden by default. - Settings: 'Delete reader profile' button calls AiRepository.deleteProfile() with snackbar feedback (success / failure) and a re-entrancy guard. - ProgressRepository: expose abandonedAt for the library filter. * ✨ fix(server): POST /profile/refresh returns the ReaderProfileResponse envelope `POST /ai/v1/profile/refresh` used to return the raw `ReaderProfilePayload` while the Android client decodes the wrapped `ReaderProfileResponseDto` envelope — every real refresh would have failed client-side deserialization. The bug was masked in `AiRepositoryRefreshTest` because the fixture body already used the envelope shape. Refresh now returns the same envelope as `GET /profile` so the server is internally consistent and the Android client decodes one DTO for both endpoints. Added an integration test that exercises the low-data short-circuit path (no LLM call required) and asserts the envelope top-level fields are present. * ✨ fix: canonicalize reader profile fingerprint as epoch millis Server hashed `datetime.isoformat()` (UTC → `+00:00`) while Android hashed `Instant.toString()` (UTC → `Z`). The two representations diverge for the same instant, so the fingerprint compare reported spurious staleness on every refresh and the "Profile may be out of date" banner would have been permanently on. Both ends now serialize `latest_progress_updated_at` as epoch milliseconds (or the literal `"none"` when absent) before hashing. Docstrings on both sides call out the contract so a future change on either side cannot drift again. Existing VM fingerprint test updated to the canonical token and a second test covers the no-progress path. * ✨ feat(android): surface ai_suggested_recommendations in Library Insights The server `ReaderProfilePayload` carries an `ai_suggested_recommendations` list alongside the in-library and discovery buckets, but the Android DTO didn't declare the field and the screen never rendered the bucket — so AI-only suggestions (no candidate match, no Open Library URL) were silently dropped from the UI. Added the field to `ReaderProfilePayloadDto` (defaulted, so older server builds round-trip cleanly) and added a "Suggested by AI" LazyRow section between Discovery and the footer. `BookRecCard` already handles `source_type == "ai_suggested"` by showing title/author/rationale with no link affordance, so no card-level changes were needed.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
ImageLoaderto the existing OPDSOkHttpClientvia anImageLoaderFactoryonEReaderApp, so cover requests inherit theBasicAuthInterceptor. Without this, calibre-web cover requests 401'd silently and the catalog grid only ever showed the gradient/initials fallback.OpdsClientwith a small XML scan over each<entry>, joined back to publications by acquisition href. Readium's OPDS model populatespub.subcollections["images"]only whenopds:imageis present and silently dropsopds:image/thumbnailwhen it appears alone — reordering rels couldn't fix the thumbnail-only case. The new scan prefersopds:image/thumbnailoveropds:image(saves bandwidth on a 2-col grid) and mirrors the existingparseSearchLinkworkaround in the same file.docs/superpowers/.Test plan
scripts/dgradle :data:opds:testDebugUnitTest— new parser tests cover both-rels (thumbnail wins), thumbnail-only, image-only fallback; existing tests unchanged.scripts/dgradle :app:testDebugUnitTestscripts/dgradle :app:assembleDebugproduces a working APK.