Repository navigation
Releases
v2.0.4
v2.0.4 (2026-07-27)
Bug Fixes
- Coerce llm_examples_per_category to int (
bce1caa)
NumberSelector persists its value as a float (2.0), so the examples slice in select_examples raised "slice indices must be integers or None or have an index method" for anyone who had opened the Advanced options step.
Coerce on read in the coordinator so existing broken configs recover without user action, and on write in the options flow so newly saved values are stored as ints.
Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com
Detailed Changes: v2.0.3...v2.0.4
v2.0.3
v2.0.3 (2026-05-30)
Bug Fixes
- Gracefully skip cookidoo when unavailable, sync keep-only (
5c98a03)
When the Cookidoo integration is broken (e.g. auth expired), the sync now logs a warning and continues with Keep items only instead of crashing. Cookidoo completions are also skipped gracefully.
Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com
Detailed Changes: v2.0.2...v2.0.3
v2.0.2
v2.0.2 (2026-05-30)
Bug Fixes
- Use SupportsResponse.ONLY to avoid entity-matching error (
7b2667e)
HA rejects SupportsResponse.OPTIONAL when called from the UI with response data enabled but no target entities — our service uses entry_id, not entity targeting. ONLY is correct since the handler always returns a dict.
Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com
Detailed Changes: v2.0.1...v2.0.2
v2.0.1
v2.0.1 (2026-05-15)
Bug Fixes
- Crash-safe sync with journal recovery, verify+retry stale deletes (
f0adbbb)
The sync previously deleted ALL Keep items before re-adding them. A crash between delete and re-add would permanently lose the entire shopping list.
- Sync journal persists desired state before destructive ops; next sync
auto-recovers missing items if previous sync was interrupted. - Parallel deletes use return_exceptions=True (one failure won't abort batch).
- Verify+retry step after deletes catches silently surviving items.
- async_remove_item / async_add_item now log errors before re-raising.
- Cookidoo completion moved to LAST phase (after Keep is fully updated).
Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com
Detailed Changes: v2.0.0...v2.0.1
v2.0.0
v2.0.0 (2026-04-26)
Chores
-
Defensive checks, i18n entity name, ci on develop branch (
4c65423) -
Coordinator: explicit HomeAssistantError if run_sync is called for an
unknown entry_id (was a silent KeyError before). -
Skip-check extracted into a named helper for readability.
-
Categories todo entity uses translation_key + entity-translation files
instead of a hardcoded German name. -
pyproject.toml gets a comment explaining why the version is 0.0.0
(it's just tooling config, not a published package). -
Test workflow now runs on develop pushes/PRs too, so the develop
branch gets CI feedback without triggering releases. -
Strip stray triple-blank lines.
Co-Authored-By: Claude Opus 4.7 (1M context) noreply@anthropic.com
Features
-
Configurable LLM examples count, services.yaml hints, parser tests (
3fdfc0e) -
New 'Advanced' options step exposes llm_examples_per_category
(default 2, range 0-10). 0 disables the calibration block entirely. -
services.yaml entry_id now has a real description that explains
when it's needed and how to find the ID. -
Extracted parse_bulk_response and extract_speech as pure public
functions in classifier.py so the LLM-output parsing — the most
fragile spot, since different conversation agents format replies
differently — has direct unit-test coverage. -
14 new tests covering perfect format, whitespace, punctuation,
case-insensitivity, fuzzy substring matching, hallucinated
categories, out-of-range indices, partial responses and malformed
speech-extraction shapes.
Co-Authored-By: Claude Opus 4.7 (1M context) noreply@anthropic.com
- New default categories, harden review-flagged spots (
2c26f9f)
Behavior changes: - DEFAULT_CATEGORIES is now the curated 38-category list the user has been iterating on in production (Obst, Gemüse, Kartoffel, ... , Tiefkühlware, Saft, Getränke, Sonstiges). Replaces the original Rewe-style 14-category placeholder. - Service handlers now raise ServiceValidationError when called without entry_id and multiple configurations exist, instead of returning an empty result silently. - Quantities with unknown units preserve original casing in the output ('1 Tütchen + 2 Tütchen' → '3 Tütchen' instead of '3 tütchen').
Code quality: - Quantity dataclass renamed to _Quantity (private) and gained a unit_raw field to support the casing fix above. - coordinator.async_run_sync now returns a typed SyncResult TypedDict and uses a typed _CookidooGroup TypedDict instead of dict[str, dict]. - Removed dead CONF_LEARNED constant. - Promoted the lazy import of coordinator inside config_flow's async_step_learned to a top-level import (no circular dependency).
Tooling: - semantic-release now generates a CHANGELOG.md, excluding its own release commits.
Docs: - README updated to use the new English entity slug todo.cookidoo_keep_categories (no existing users to migrate).
Co-Authored-By: Claude Opus 4.7 (1M context) noreply@anthropic.com
Detailed Changes: v1.0.0...v2.0.0
v1.0.0
v1.0.0 (2026-04-26)
Refactoring
- Drop default keyword set (
adaa012)
Default keywords pointed at hardcoded category names ("Obst/Gemüse", "Tomaten Sachen", ...) but categories are user-editable via the auto-created todo entity. Once users renamed or restructured their categories, the defaults silently stopped matching anyway.
The LLM + learned cache make the keyword feature optional. Fresh installs now go straight to the LLM on first sync; everything is cached after that. Power users who want offline-only operation can still maintain keywords manually via the options flow — and now those keywords are guaranteed to match their actual categories because they wrote them themselves.
Existing users keep their previously-saved keyword config; this change only affects fresh installs.
Co-Authored-By: Claude Opus 4.7 (1M context) noreply@anthropic.com
- Drop keyword feature, send learned examples to LLM, fix dead code (
eae2bd9)
Behavior changes: - The keyword-substring matcher is gone. The LLM + persistent cache cover the same use cases, with the cache being checked first (instant), and the LLM only consulted for genuinely new items. The cache is editable via the options flow, so manual overrides still work — even for offline-only setups, the user just populates the cache themselves instead of curating keywords. - Bulk LLM calls now include up to 2 already-learned examples per category as calibration, so the LLM adopts the user's personal category schema and stops repeating earlier mistakes. - Removed CONF_CLEAR_BEFORE_SYNC: it was no longer wired up since the merge-and-rewrite refactor and was confusing dead UI.
Code quality: - Multi-entry service handler now resolves the target entry_id at call time (single registered entry → that one; multiple → require explicit entry_id). Previously the closure captured the first entry and ignored later ones. - Phase comments in the coordinator's run_sync are now numbered consistently (1..6 instead of overlapping 1, 2, 3a, 3b, 3, 4, 5...) - async_complete_item now catches HomeAssistantError specifically, not bare Exception. - _text_to_mapping has an explicit lowercase_keys flag instead of a hardcoded behavior tied to one caller. - Removed unused _LOGGER import in todo.py.
Docs: - README updated for the new flow (no more keyword references, quantity aggregation explained, cache editor mentioned). - services.yaml description rewritten.
Co-Authored-By: Claude Opus 4.7 (1M context) noreply@anthropic.com
Detailed Changes: v0.4.1...v1.0.0
v0.4.1
v0.4.1 (2026-04-26)
Bug Fixes
- Serialize cookidoo completes to avoid lost updates (
8247164)
The Cookidoo integration (miaucl/ha-cookidoo) loses ~90% of todo.update_item calls when they fire in parallel — a race somewhere on its API client. Switching the completes to sequential while keeping the keep-deletes parallel restores reliability.
Net cost: a few extra hundreds of milliseconds for the cookidoo completes phase, which is acceptable.
Co-Authored-By: Claude Opus 4.7 (1M context) noreply@anthropic.com
Detailed Changes: v0.4.0...v0.4.1
v0.4.0
v0.4.0 (2026-04-26)
Continuous Integration
- Bump checkout and setup-python to Node 24-compatible versions (
35cd43d)
Features
- Smarter classification + editable learned cache (
87b90bf)
Two improvements that together remove the need for hand-crafting tons of keyword specifiers like 'edelsüß = Gewürze':
-
Items containing a comma (Cookidoo's variant syntax, e.g. "Paprika, edelsüß") now skip the keyword match entirely and go directly to the LLM. Substring matching is too primitive to handle these — "paprika" would always win, regardless of qualifier. Once the LLM classifies them, they are cached and resolved instantly on subsequent syncs.
-
The learned cache is now editable through the integration's options flow (Configure → Learned mappings). Manual entries override the LLM since the cache is checked before any classifier runs.
Also: classification now runs on the clean item name (without the appended " (qty)") so cache keys stay stable when only the quantity changes between syncs.
Tests for both behaviors added; classifier import made lazy so the pure functions are testable without HA installed.
Co-Authored-By: Claude Opus 4.7 (1M context) noreply@anthropic.com
Detailed Changes: v0.3.0...v0.4.0
v0.3.0
v0.3.0 (2026-04-26)
Bug Fixes
- Complete cookidoo items even when they are duplicates of existing keep items (
eb1d7c7)
Previously the cookidoo completion path was tied to whether an item went through the add-to-keep loop. With the merge dedup logic, a cookidoo item that already existed in keep was deduped out of the add path and never completed in cookidoo, causing it to reappear on every subsequent sync.
Now every open cookidoo item is unconditionally queued for completion, keyed by UID (preferred, more reliable for todo.update_item) with summary fallback. Also upgraded the silent-fail log to WARNING so future issues are visible.
Co-Authored-By: Claude Opus 4.7 (1M context) noreply@anthropic.com
Features
- Aggregate quantities, add tests + semantic-release CI (
23887e9)
Cookidoo items with the same name (e.g. "Zwiebeln" appearing in two recipes with 70 g and 1 Stk) are now grouped, their quantities parsed and aggregated. Compatible units are converted (e.g. 800 g + 0.5 kg = 1.3 kg, 1 EL + 2 TL = 1.67 EL). Incompatible units are shown side by side ("70 g + 1 Stk"). The result is appended to the Keep entry as "Name (qty)".
Also handles deduplication with manually-added Keep entries: a Keep "Tomaten" and a Cookidoo "Tomaten (200 g)" are now treated as the same item (Cookidoo wins, carrying the quantity).
Add a real test suite (pytest, 57 tests) for the quantity parser and aggregator. Set up GitHub Actions: tests run on push/PR; on push to main, python-semantic-release auto-bumps the version in manifest.json and publishes a GitHub release based on Conventional Commits.
Co-Authored-By: Claude Opus 4.7 (1M context) noreply@anthropic.com
Detailed Changes: v0.2.0...v0.3.0
v0.2.0 - faster sync
Performance
- Parallelize Keep deletes + Cookidoo completes — order-independent calls now run concurrently via `asyncio.gather`. Roughly halves wall-clock time for typical syncs.
- Skip entirely when the Keep list is already sorted and there are no open Cookidoo items left to copy.
- When Keep is already aligned but Cookidoo has open items, only the Cookidoo completes run (no needless Keep clear+rewrite).