weekplan

/weekplan — Weekly meal plan + shopping list

Purpose

End-to-end pipeline: load profile → geocode → current supermarket offers → Cookidoo recipe selection → diet-constrained 7-day plan → per-store shopping list grouped by aisle + walking-distance route. Writes recipes directly to Cookidoo "Meine Woche" + shopping list via cookidoo-mcp write tools. Persists every weekly plan to disk under ~/.weekplan/plans/<YYYY-MM-DD>/.

Modes

• Interactive (default) — ask cadence question, ask whether to extend/replace existing plan, surface clarifying questions when ambiguous.
• Auto / unattended — triggered ONLY by explicit flag: argv contains --auto OR env WEEKPLAN_AUTO=1. Substring matches on prompt text MUST NOT activate auto mode (false-positive risk would destroy an existing curated plan without confirmation). In auto mode: NEVER call AskUserQuestion. Use defaults from profile.json. REPLACE existing plan via snapshot-then-replace (see "Snapshot before destructive ops" below). NEVER call clear_week — it is not part of the replace flow under any mode. Auto mode must complete with no user-facing prompts.

Snapshot before destructive ops

This block applies to REPLACE mode ONLY (auto, or interactive replace). EXTEND mode never enters it — it has no snapshot and no removal step, so its writes are purely additive and a partial add_to_week failure is non-fatal (record in Notes, continue).

In REPLACE mode, before any bulk remove_from_week, ALWAYS:

1. Call mcp__cookidoo__get_week_plan and serialize the full result to ~/.weekplan/snapshots/<YYYY-MM-DD>-pre-replace.json with mtime, mode (auto|interactive), and the planned new entries.
2. Only after the snapshot is on disk: perform add_to_week for the NEW recipes first.
3. Only after all add_to_week writes succeed: compute the set difference on the composite key {recipeId, dayKey} — to_remove = snapshot_entries \ new_entries, where each entry is the pair (recipeId, dayKey). Compare recipeIds as the exact normalized string the tool returns (the rNNN form). Perform remove_from_week({recipeId, dayKey}) for each pair in to_remove. A recipe that stays on the SAME day in both old and new plan is retained (never removed). A recipe that MOVES to a different day appears as an old (recipeId, oldDay) pair in to_remove (removed) plus a new (recipeId, newDay) pair added in step 2 — so the stale day is correctly cleared. Recipes the user added manually outside the snapshot are left untouched.
4. On any failure during step 2 or 3: STOP. Do not continue. Report failure with path to the snapshot file. The user can restore manually by replaying the snapshot.

This replaces the previous "clear_week first" pattern, which had no rollback path. clear_week is never used for replacement.

Inputs

Profile — ~/.weekplan/profile.json

Single source of truth. Schema: profile.schema.json in this skill folder. Example: profile.example.json.

Key fields:

• household.size, household.portionsPerMeal
• location.address (preferred) or location.zipCode; location.shoppingRadiusKm
• location.nominatimContact — REQUIRED. Your email or URL. Sent as User-Agent contact to Nominatim/Overpass (their usage policy). Without it, the supermarkets-mcp server refuses geocoding requests.
• diet.type (omnivore/pescetarian/vegetarian/vegan), diet.lowHistamine, diet.preferVeganDairyAlternatives
• diet.forbidden[] — case-insensitive substring matches in ingredient text
• diet.flaggedButAllowed[] — surface in Notes section but do not drop
• preferences.cadenceDefault — dinner-only or lunch-and-dinner
• preferences.preferredStores[] — Marktguru retailer slugs
• preferences.maxRecipesPerWeek, preferences.maxSameProteinPerWeek

Read every run via Read tool on ~/.weekplan/profile.json. If missing or invalid JSON, abort with copy-paste instructions to create from profile.example.json.

Cadence

Interactive: confirm preferences.cadenceDefault is right for this week ("Diese Woche: nur Abendessen, oder Mittag + Abend?"). Default = profile setting. Auto: use profile setting silently.

Existing week plan

Call mcp__cookidoo__get_week_plan first. Defaults to Monday of current local week. Interactive: if non-empty, ask whether to extend or replace (replace = snapshot-then-set-diff, see step 8 / "Snapshot before destructive ops"). Auto: replace silently via the same snapshot-then-set-diff flow. NEVER call clear_week as the replace mechanism.

Outputs — ~/.weekplan/plans/<YYYY-MM-DD>/

<YYYY-MM-DD> = Monday of the target week (ISO date). Files written each run:

File	Contents
plan.md	Markdown plan: header + daily table + Cookidoo confirmation
shopping-list.md	Per-store shopping list, aisle-grouped, with walking-route hint
offers.json	Snapshot of the Marktguru offers used (frozen, for later audit)
recipes.json	Snapshot of recipe IDs + titles + chosen day keys

Write order: build full content in memory → write all 4 files atomically near the end of the pipeline. Auto mode also prints plan.md to chat.

Pipeline

1. Load profile + resolve location

• Read ~/.weekplan/profile.json. Abort cleanly if missing.
• If location.address set: call mcp__supermarkets-mcp__geocode_address(address) to get lat/lon + zip. Use that zip for offer queries.
• Else use location.zipCode directly.
• Call mcp__supermarkets-mcp__find_stores_nearby(address, radiusKm=location.shoppingRadiusKm). Cache result for use in step 8.

2. Determine target meals

• dinner-only: 7 dinners. Lunch = previous-night leftovers (low-histamine: leftovers ≤24h, eaten next day cold or warmed once).
• lunch-and-dinner: 14 meals; design overlap explicitly (e.g., big-batch curry → 4 meals across 2 days).

3. Fetch supermarket offers

Call mcp__supermarkets-mcp__get_weekly_offers with address (or zipCode) from profile, stores=preferences.preferredStores. Default basket OR override based on user signals (e.g. "lust auf Asia" → asian terms).

4. Apply diet hard-filters to offers

Drop offers whose description or title contain any string from profile.diet.forbidden[] (case-insensitive).

Special rules when diet.lowHistamine == true:

• All matches in forbidden[] already cover the histamine triggers — no extra logic.
• If forbidden[] is empty but lowHistamine == true, fall back to the built-in low-histamine list (see profile.example.json for the default set).

Vegan dairy alternatives PREFERRED when preferVeganDairyAlternatives == true: prioritize oat/almond/coconut milk; if no vegan alt in offers, lactose-free dairy acceptable — surface in Notes.

5. Search Cookidoo for recipes matching offers

For top ~15 offer ingredients (post-filter), call mcp__cookidoo__search_recipes per ingredient. Aggregate ~50-80 candidates.

If profile is missing or returns no usable offers, fall back to mcp__cookidoo__random_recipe (seeded with category from profile or default basket) to fill gaps — flag in Notes.

For each candidate, call mcp__cookidoo__get_recipe for full ingredient list. Parallelize via Promise.all in a single tool batch.

6. Apply diet hard-filters to recipes

Drop recipes whose ingredients contain any diet.forbidden[] string. Flag (don't drop) items in diet.flaggedButAllowed[].

7. Optimize for low waste — ingredient overlap

Build ingredient frequency map. Greedy assembly: pick highest-rated recipe whose ingredient set maximizes overlap with already-chosen meals AND uses ≥1 offer ingredient. Diversity guard: at most preferences.maxSameProteinPerWeek of the same primary protein in 7 days.

8. Write plan to Cookidoo

For each chosen recipe + dayKey (YYYY-MM-DD, Mon-Sun of target week):

• mcp__cookidoo__add_to_week({recipeIds: [id], dayKey})
• mcp__cookidoo__add_to_shopping_list({recipeIds: [id]}) (covers all ingredients)

Write order (snapshot-then-replace, clear_week is NEVER used):

1. Call mcp__cookidoo__get_week_plan({startDate, span: 7}) and write the result + the planned new entries to ~/.weekplan/snapshots/<YYYY-MM-DD>-pre-replace.json BEFORE any write.
2. Auto mode + Interactive replace: add_to_week ALL new recipes first.
3. Only after all new add_to_week writes succeed: compute to_remove = snapshot_entries \ new_entries on the composite key {recipeId, dayKey} (recipeId compared as the exact normalized rNNN string), and remove_from_week({recipeId, dayKey}) only those pairs. A recipe on the same day in both plans is retained; a day-moved recipe has its old (recipeId, oldDay) pair removed here and its new (recipeId, newDay) pair added in step 2; manually-added recipes outside the snapshot are left untouched.
4. Interactive extend: skip steps 1 + 3, just add_to_week new recipes alongside existing.
5. Always add_to_shopping_list for the new chosen recipes (regardless of mode).

REPLACE mode only — if any add_to_week fails: STOP. Do not call remove_from_week. Report failure with snapshot path so the user can resume manually. The original plan must remain intact when our writes haven't fully succeeded. (EXTEND mode has no removal step, so a partial add_to_week failure is safe — see Failure modes.)

9. Build shopping list — cost vs distance

For each ingredient across the 7 plans:

1. Look up the offer (if any) that matched it in step 4 → preferred retailer slug.
2. Find that retailer in the find_stores_nearby result → distance.
3. Compute savings = (regular price − offer price) × quantity.
4. Group ingredients per store. If a store contributes < 2 € savings and is > 1 km farther than the primary store, consolidate to the primary store (use the primary's regular price — flag in Notes).
5. Primary store = closest of preferences.preferredStores from the nearby list.
6. Group within each store by aisle: produce / dairy-alt / fish / pantry / frozen / bakery.
7. Quantity per ingredient: sum across recipes, scale to household.portionsPerMeal.

Output a walking route hint at the top of shopping-list.md: stores ordered by distance from origin.

10. Write files + chat output

Markdown plan structure (plan.md):

• Header: week range (Mon-Sun), meal count, total estimated cost from offer prices, total estimated savings, "✓ N recipes added to Cookidoo Meine Woche"
• Daily plan table: | Day | Dinner (recipe + URL) | Lunch (leftover / new) |
• Cookidoo confirmation: "Plan added to Cookidoo. Open https://cookidoo.de/planning/de-DE to view." (no manual handoff URLs unless writes failed)
• Notes: flagged-but-allowed recipes, allergy reminders, any add_to_week write failures, consolidation decisions

shopping-list.md:

• Route hint: stores in distance order from origin
• Per-store sections with aisle subgroups + quantities

Auto mode: also print plan.md to chat. Final line: WEEKPLAN_AUTO_DONE <output-file-path> for scheduler detection.

MCP tool inventory

supermarkets-mcp (read-only)

• mcp__supermarkets-mcp__list_stores({zipCode?, address?}) — discover retailers
• mcp__supermarkets-mcp__search_offers({query, zipCode?, address?, stores?, limit?}) — keyword search
• mcp__supermarkets-mcp__get_weekly_offers({zipCode?, address?, stores?, terms?, perTermLimit?}) — fan-out basket search
• mcp__supermarkets-mcp__geocode_address({address}) — address → lat/lon/zip
• mcp__supermarkets-mcp__find_stores_nearby({address, radiusKm?}) — OSM supermarkets near origin w/ distance + retailer slug

cookidoo

Read:

• mcp__cookidoo__search_recipes({query, hitsPerPage?}) — Algolia search (empty query rejected; use random_recipe)
• mcp__cookidoo__random_recipe({category?}) — random recipe (optionally seeded)
• mcp__cookidoo__get_recipe({id}) — full detail (ingredients, instructions)
• mcp__cookidoo__get_week_plan({startDate?, span?}) — read existing plan

Write (cookidoo-mcp, ToS-risk override accepted 2026-05-06):

• mcp__cookidoo__add_to_week({recipeIds, dayKey}) — add 1+ recipes to a day
• mcp__cookidoo__remove_from_week({recipeId, dayKey})
• mcp__cookidoo__clear_week({startDate?, span?}) — wipe plan span (composite, iterates removes)
• mcp__cookidoo__add_to_shopping_list({recipeIds}) — add ingredients
• mcp__cookidoo__mark_owned({ingredientIds}) — mark pantry-owned
• mcp__cookidoo__unmark_owned({ingredientId})
• mcp__cookidoo__bookmark_recipe({recipeId}) — Merken
• mcp__cookidoo__unbookmark_recipe({recipeId})
• mcp__cookidoo__rate_recipe({recipeId, rating}) — 1-5 stars

Auto-checklist (use TaskCreate)

1. Detect mode (interactive vs auto — explicit --auto or WEEKPLAN_AUTO=1 only; no substring sniffing)
2. Read profile.json — abort if missing/invalid
3. Geocode address (if set) + find_stores_nearby
4. Cadence: ask interactively, default from profile in auto
5. Read existing Cookidoo week plan
6. If replace mode: snapshot existing plan to ~/.weekplan/snapshots/<YYYY-MM-DD>-pre-replace.json (clear_week is never called)
7. Fetch weekly offers (using resolved address/zip + preferredStores)
8. Apply diet filter to offers
9. Search Cookidoo for top offer-ingredients (use random_recipe to fill gaps)
10. Fetch recipe details in parallel
11. Apply diet filter to recipes
12. Greedy assembly with overlap optimization
13. add_to_week per chosen recipe + dayKey — in REPLACE mode, STOP entire pipeline if any write fails (do not run step 14); in EXTEND mode, a failed write is non-fatal (no removal follows) — record it in Notes and continue
14. Only after all new add_to_week writes succeed AND mode is replace: remove_from_week({recipeId, dayKey}) each pair in to_remove (= snapshot {recipeId,dayKey} pairs minus new-plan pairs; same-day retained recipes stay, day-moved recipes get their old day cleared)
15. add_to_shopping_list for all chosen recipes
16. Build cost-vs-distance shopping plan using find_stores_nearby data
17. Write 4 output files to ~/.weekplan/plans/<YYYY-MM-DD>/
18. Auto mode: also print plan.md + emit WEEKPLAN_AUTO_DONE line

Failure modes

• Profile missing/invalid → instruct user to copy profile.example.json → ~/.weekplan/profile.json and edit. Exit cleanly.
• Geocode fails → fall back to location.zipCode if set. Skip nearby-stores step; cost-vs-distance degrades to cost-only. Note in output.
• Overpass query fails or times out → skip nearby-stores, cost-only mode, note.
• Cookidoo cookies expired → instruct user to re-run playwright-cli login + import-state. Show exact commands.
• Marktguru offer endpoint TLS or rate-limit → degrade to generic seasonal basket without offer matching; warn user.
• Histamine list too restrictive → if <7 valid recipes survive after 2 search rounds, broaden basket terms and re-search before failing.
• No vegan dairy alt in offers → fall back to lactose-free dairy at primary store; surface in notes.
• add_to_week fails in EXTEND mode (no removal step) → continue pipeline, list failed (recipe, dayKey, error) in Notes section so user can manually add via URL fallback. Safe because nothing gets removed.
• add_to_week fails in REPLACE mode → STOP. Do NOT run remove_from_week. Surface the snapshot path so the user can resume manually. The partial new writes simply append to the existing plan (nothing destroyed). Note it in output.

After completing

Interactive mode: provide 3 numbered next steps:

1. Open https://cookidoo.de/planning/de-DE to verify plan
2. Open ~/.weekplan/plans/<YYYY-MM-DD>/shopping-list.md on phone
3. Rate recipes after cooking via mcp__cookidoo__rate_recipe so memory tracks preferences

Auto mode: skip next-steps section. Final line of output: WEEKPLAN_AUTO_DONE <output-file-path> for scheduler detection.

Scheduling auto mode

claude -p "/weekplan --auto" --dangerously-skip-permissions
WEEKPLAN_AUTO=1 claude -p "/weekplan"

Cron (Sunday 18:00 local):

0 18 * * 0 cd /Users/marten/Code/ade_anima/food-planner && /usr/local/bin/claude -p "/weekplan --auto" --dangerously-skip-permissions >> ~/.weekplan/cron.log 2>&1

Cookies must be present at ~/.cookidoo-mcp/cookies.txt. Refresh manually via playwright-cli login + import-state when expired.