- update_prep_task: move whitelist guard above filter so invalid column
check runs on raw kwargs (was dead code — set(filtered) - allowed is
always empty); fixes latent SQL injection path for future callers
- main.py: move register_kiwi_programs() into lifespan context manager
so it runs once at startup, not at module import time
- MealPlanView.vue: remove debug console.log stubs from onSlotClick and
onAddMealType (follow-up issue handlers, not ready for production)
- Add GET /{plan_id}/prep-session endpoint so frontend can retrieve existing sessions without creating
- Fix list_plans response_model from list[dict] to list[PlanSummary] with proper _plan_summary() mapping
- Replace assert in store.update_prep_task with ValueError (assert is stripped under python -O)
- Add day_of_week 0-6 validation to upsert_slot endpoint
- Remove MagicMock sqlite artifact files left by pytest (already in .gitignore)
- upsert_slot: raise 422 immediately if meal_type not in VALID_MEAL_TYPES
- update_prep_task: assert whitelist safety contract after dict comprehension
- CreatePlanRequest: week_start typed as date with must_be_monday validator; str() cast at call site
- PrepTask: frozen=True; build_prep_tasks rewired to use (priority, kwargs) tuples so frozen instances are built with correct sequence_order in one pass (no post-construction mutation)
- Move deferred import json to file-level in meal_plans.py
- Fix test dates: "2026-04-14" was a Tuesday; updated request bodies to "2026-04-13" (Monday)
Hard Day Mode now prioritises results by effort tier before match_count:
Tier 0 (premade): frozen/instant title keywords, or ≤2 ingredients with
heat/microwave-only steps (frozen dinner, heat-and-eat, microwave meal)
Tier 1 (super simple): ≤3 ingredients + any easy method (quesadilla,
cheese toast, scrambled eggs)
Tier 2 (easy/moderate): everything else that passed the 'involved' filter
Assembly templates default to tier 1 (inherently simple). Normal mode sort
is unchanged — match_count only.
- Add migration 021: recipe_browser_fts FTS5 table on category + keywords
columns, eliminating LIKE '%keyword%' full sequential scans on 3.1M rows
- _count_recipes_for_keywords now uses FTS5 MATCH (O(log N) vs O(N))
- browse_recipes reuses cached count, eliminating the second COUNT(*) scan
per page request; ORDER BY r.id replaces the unindexed ORDER BY title sort
- Module-level _COUNT_CACHE keyed by (db_path, keywords) means domain-switch
category counts are computed once per process lifetime
feat(find): dietary preset grid, Big 9 allergen pills, Hard Day Mode surface
- Dietary constraints replaced with toggle-button preset grid (8 options)
+ free-text "Other" field; removes dense freeform text input
- Allergies replaced with Big 9 pill picker (peanuts, tree nuts, shellfish,
fish, milk, eggs, wheat, soy, sesame) + "Other" for custom entries
- Hard Day Mode surfaced as a standalone aria-pressed button above the
dietary collapsible; no longer buried inside a collapsed section
- Active-state dot indicators on both collapsibles show filter engagement
at a glance without expanding
fix(a11y): aria-describedby wiring for wildcard checkbox and tag inputs (#40)
- Persistent hint spans replace placeholder-only instructions for constraint
and allergy fields (WCAG 3.3.2)
fix(browse): auto-select highest-count category on domain switch (#41)
- Eliminates the 3-decision cold start (domain → category → content)
- Surprise Me button added for zero-decision random navigation
Replace hand-rolled feedback.py with make_feedback_router() from
circuitforge_core.api.feedback. Update tests to mount the shared
router on a minimal FastAPI app and mock at the core module level.
Patch _user_db_path tests to monkeypatch CLOUD_DATA_ROOT onto a
tmp_path so they never touch /devl or any real filesystem path.
Remove duplicate X-Real-IP comment block in cloud_session.get_session.
Add household_id and is_household_owner fields to CloudUser dataclass.
Update _user_db_path to route household members to a shared DB path.
Update _fetch_cloud_tier to return a 3-tuple and cache a dict.
Update get_session to unpack and propagate household fields.
- feedback.py: add GET /feedback/status endpoint (returns {enabled: bool})
so frontend can probe on mount instead of optimistic-enable; remove
unused get_db import
- FeedbackButton.vue: probe /feedback/status on mount, start hidden;
drop redundant 503-hide path (status probe makes it redundant)
- pyproject.toml: declare requests>=2.31 (used by feedback.py Forgejo calls)
- tests/api/test_feedback.py: 7 tests — status endpoint (no-token, token,
demo mode), POST 503/403, happy path with mocked Forgejo, 502 on error
- .env.example: document ANTHROPIC_API_KEY, OPENAI_API_KEY, OLLAMA_HOST,
OLLAMA_MODEL, CF_ORCH_URL, CF_LICENSE_KEY with usage comments
- config.py: expose CF_LICENSE_KEY in Settings for startup visibility
- pyproject.toml: pin circuitforge-core >= 0.6.0 (env-var auto-config +
CFOrchClient bearer auth land in 0.6.0)
Bare-metal self-hosters can now run Kiwi with only OLLAMA_HOST set and
zero yaml config. Paid+ users set CF_ORCH_URL + CF_LICENSE_KEY for
managed cloud GPU inference.
Migration 015 did a one-time rebuild of recipes_fts at creation time but
omitted triggers, so rows inserted after that point were invisible to MATCH
queries. Adds AFTER INSERT/UPDATE/DELETE triggers to 015 (fresh DBs / tests)
and migration 016 to backfill them on existing databases.
Fixes 3 failing tests: test_search_recipes_by_ingredient_names,
test_level1_returns_ranked_suggestions, test_level2_returns_swap_candidates.
_parse_json_from_text always returns a dict (never None), so the
previous `if parsed is not None` guard was permanently true — garbled
docuvision output would return an empty skeleton instead of falling
through to the local VLM. Replace the check with a meaningful-content
test (items or merchant present). Add two tests: one that asserts the
fallthrough behavior on an empty parse, one that confirms the fast path
is taken when parsing succeeds.
Introduces a thin HTTP client for the cf-docuvision service and wires it
as a fast path in VisionLanguageOCR.extract_receipt_data(). When CF_ORCH_URL
is set, the pipeline attempts docuvision allocation via CFOrchClient before
loading the heavy local VLM; falls back gracefully if unavailable.
