MealQuest/docs/ARCHITECTURE.md

# MealQuest architecture

This mirrors the architecture section of `CONTEXT.MD` in the CTXD project
`mealquest`, which is the source of record; if the two disagree, CTXD wins and
this file should be updated. It describes the **planned** application. The
repository today is the Feature 01 scaffold: `cmd/server`, `internal/config`,
`internal/server` (`/health`), `internal/frontend` (embedded PWA with SPA
fallback), the `web/` skeleton, `Makefile`, `Dockerfile` and CI.

Decision numbers (D-NNN) refer to `DECISIONS.MD` in CTXD.

## Topology

```
PWA (phone / tablet / desktop)
   │ HTTPS
   ▼
Traefik v3.6 — 10.60.1.33
   file route: mealquest-ccl.yml
   │
   ▼
docker compose project "mealquest" — 10.60.1.14
   ├── mealquest   Go service + embedded PWA, host 8090 → container 8080
   ├── searxng     sidecar for recipe search; internal compose network
   │               only, no published port
   └── Litestream  SQLite replication → Synology NAS 10.60.1.50
```

One Go service serves the API, the auth flow, media and the PWA. Its SQLite
database is replicated by Litestream to the NAS.

Traefik terminates TLS. The service honours `X-Forwarded-For` and
`X-Forwarded-Proto` only from peers listed in `MEALQUEST_TRUSTED_PROXY`
(comma-separated IPs/CIDRs; production lists `10.60.1.33`); from any other
peer they are ignored. The client is the rightmost `X-Forwarded-For` address
that is not a trusted proxy. The resolved client IP is added as `client_ip`
to every log line written while handling a request. The OIDC redirect URI
and the Secure cookies are fixed by `MEALQUEST_PUBLIC_URL`, never by request
headers, so sign-in works whichever peer the request arrives from.

## Repository layout

| Path | Purpose |
| --- | --- |
| `cmd/server` | Entry point; wiring only. |
| `internal/config` | Env-only `MEALQUEST_*` config; secrets from `*_FILE`. |
| `internal/server` | HTTP server, router, middleware. |
| `internal/frontend` | Embedded PWA, SPA fallback. |
| `internal/auth` | Authentik OIDC, sessions, CSRF. |
| `internal/db`, `internal/store` | SQLite connections, sqlc queries. |
| `internal/secretbox` | AES-256-GCM encryption for `_ct` columns. |
| `internal/household`, `internal/dinners`, `internal/recipes` | Domain. |
| `internal/ai/llm`, `internal/ai/harness`, `internal/ai/skills` | AI provider adapters, skill harness, skills. |
| `internal/suggest`, `internal/plan`, `internal/shopping`, `internal/ideas` | Feature services. |
| `internal/gcal`, `internal/notify`, `internal/push`, `internal/chat` | Integrations. |
| `internal/jobs`, `internal/scheduler` | Background jobs and schedules. |
| `migrations/` | SQL migrations (no goose or golang-migrate). |
| `web/` | React PWA. |
| `deploy/` | Deployment files. |
| `docs/` | Documentation. |
| `design/` | Requirements and UI handoff (read-only reference). |

## Backend stack and conventions

- Go 1.25, `CGO_ENABLED=0`, a single static binary.
- Router: `go-chi/chi/v5`. Logging: `log/slog` JSON.
- SQLite via `modernc.org/sqlite`, with `journal_mode=WAL`,
  `foreign_keys=ON` and `busy_timeout`. **One writer connection plus a read
  pool.** Queries generated by `sqlc`.
- IDs are ULIDs (`oklog/ulid/v2`).
- Timestamps are stored in UTC; local calendar dates are `YYYY-MM-DD` strings
  in the household timezone.
- Configuration is environment-only (`MEALQUEST_*`). Secrets come only from
  files named by `*_FILE` variables. Missing or invalid config stops startup.
- Error responses always have the body
  `{"error":{"code":"…","message":"…","retryable":false}}`.

## Authentication

- Authentik OIDC; the backend is a **confidential client**
  (`coreos/go-oidc/v3`, `golang.org/x/oauth2`).
- Session cookie `__Host-mealquest_session` (HttpOnly, Secure,
  SameSite=Lax). Sessions expire after 30 days idle or 180 days absolute.
- Non-GET requests require the CSRF header `X-MealQuest-CSRF`.
- Access requires membership of the Authentik group `mealquest-admin`.
  There is a **single role**: one shared household, no per-person profiles.

## Data model

Columns ending in `_ct` hold AES-256-GCM ciphertext (via `secretbox`).

| Table | Holds |
| --- | --- |
| `household_settings` | One row: timezone, dinner_time (17:30), reminder_time (16:30), review_time (09:00), plan_draft day/time (Fri 09:00), suggestions day/time (Sun 08:00), chat_autosend day/time (Sat 08:00), onboarding_completed_at, updates_seen_at. |
| `recipes` | A real recipe page: source_url (unique), site_name, title, total/prep/cook minutes (the page's JSON-LD), yield, raw_jsonld, fetched_at; and its grounding: photo_id, method, difficulty, minutes_total (normalized prep to table). |
| `recipe_ingredients` | recipe_id, position, raw_text, name, qty, unit, note. |
| `recipe_steps` | recipe_id, position, text. |
| `photos` | source_url, local_path, width, height, sha256, fetched_at. |
| `dinners` | The meal record and its lifecycle: name, normalized_name, status (`suggested` / `quest_log` / `tried` / `regular` / `archived`), origin, method, minutes_total, difficulty, cuisine, is_favorite, recipe_id, recipe_url (a pasted recipe link awaiting grounding), photo_id, grounding_status (`pending` / `grounded` / `failed` / `manual`), grounding_error (why grounding failed), grounding_revision (raised whenever a dinner is renamed, so only the latest `ground_dinner` job may write it and no candidate offered before the rename grounds it), updated_by. |
| `suggestion_batches` | trigger, status, ai_run_id. |
| `suggestions` | dinner_id, batch_id, name (the name the dish was suggested under, kept whatever its dinner is later renamed, so a Pass stays an immutable record every later batch dedupes against), reason, cited_regular_ids (JSON), status, decided_at. |
| `week_plans` | week_start (unique), status (`draft` / `synced` / `changed`), ai_note, ai_run_id, calendar_synced_at, revision (raised by every swap and stored draft). |
| `plan_days` | plan_id, date (unique per plan), dinner_id, is_new, calendar_event_id. |
| `rotation_reviews` | dinner_id + plan_date (unique), status, decision (`keep` / `drop` / `not_made`), decided_at, updated_by. |
| `shopping_lists` | week_start (unique), plan_id, title, status (`current` / `archived`), sent_to_chat_at, frozen_at. |
| `shopping_groups` / `shopping_items` | One group per dinner/day; item name, qty, needed (bool), position. |
| `ai_settings` | One row: provider, base_url, model, temperature, system_prompt, api_key_ct, api_key_last4, key_verified_at. |
| `recipe_search_settings` | One row: provider (`searxng` / `brave`), base_url (the SearXNG instance), api_key_ct and api_key_last4 (Brave). |
| `integrations_google` | One row: refresh_token_ct, calendar_id, calendar_name, connected_at, last_sync_error. |
| `chat_settings` | One row: webhook_url_ct, space_label, last_test_at, last_test_status. |
| `notification_prefs` | event (`plan_ready` / `list_ready` / `day_of_reminder` / `new_suggestions` / `rotation_review`), push, chat. |
| `push_subscriptions` | endpoint (unique), p256dh, auth, user_agent, failure_count, last_success_at. |
| `notification_outbox` | event_key + channel (unique), payload, status, attempts, next_attempt_at, last_error. |
| `jobs` | kind, status, payload, result (progressive), error, timestamps. |
| `schedule_runs` | schedule, period_key (unique). |
| `ai_runs` | skill, skill_version, provider, model, temperature, status, latency_ms, tokens_in/out, error, redacted request/response (30-day retention). |
| `sessions` | id_hash, subject, email, name, groups, csrf_token, created/last_seen/expires. |

`recipes` holds a fetched recipe page; `dinners` holds the household's meal
record and its lifecycle, optionally linked to a recipe and a photo.

## Dinner lifecycle

```
suggested (deck) ──Try it──▶ quest_log ──planned date passes──▶ tried
    │                            ▲                                │
    │                            └──────── explicit decision ─────┤
    │                                                             ├──▶ regular
    │                                                             └──▶ archived
    └──Pass──▶ archived (suggested name kept for dedupe)
```

From `tried`, an explicit decision moves the dinner to `regular`, to
`archived`, or back to `quest_log`. Promotion to `regular` is **never
automatic** (D-021).

## API surface

- JSON API under `/api/v1`, using the error body above for every failure.
  It requires the session cookie and, for non-GET requests, the
  `X-MealQuest-CSRF` header.
- `/auth/*` — the OIDC flow.
- `/media/*` — cached photos, behind the session cookie.
  `GET /media/photos/{id}-{400|1200}.jpg` serves a stored size of a photo
  with a `photos` row, with
  `Cache-Control: private, max-age=31536000, immutable`; anything else is
  404.
- `GET`/`PATCH /api/v1/settings/recipe-search` — provider, base_url and
  the Brave key (`api_key`, write-only: responses carry only
  `api_key_set` and `api_key_last4`; an empty string removes it). The
  result must be usable: SearXNG needs a base_url, Brave a key.
  Responses also carry the read-only `default_base_url`, the deployment's
  `MEALQUEST_SEARXNG_URL` (or null), which the client pre-fills while no
  base_url is saved; only a PATCH of base_url stores a URL.
  `POST /api/v1/settings/recipe-search/test` searches "chicken tikka
  masala" and returns `{query, count}`, 409 `not_configured`, or 502
  `search_failed` with the provider's error.
- `GET`/`PATCH /api/v1/settings/ai` (`aisettings.Service`) — provider,
  base_url, model, temperature, system_prompt, plus the read-only
  `api_key_last4` and `key_verified_at`. PATCH requires a preset name as
  provider, temperature 0–1 in 0.1 steps, a non-empty model and a
  system_prompt of at most 8000 characters; a fixed-URL preset's base_url
  must be empty or its own. Provider and base_url cannot change while a
  key is stored (remove the key first), so a stored key is never sent to
  an endpoint it was not stored for.
  `PUT /api/v1/settings/ai/api-key` takes `{api_key}` (write-only, at least
  8 characters, sealed with the secret box; an empty string removes it) and
  clears `key_verified_at`; a preset without a fixed base_url takes a key
  only when its base_url passes `secureurl.Check`. `POST /api/v1/settings/ai/verify` calls
  `Provider.Verify` and records `key_verified_at`;
  `GET /api/v1/settings/ai/models` returns `{models: [{id, display_name}]}`
  live from the provider, with the key redacted from every id and
  display_name. Both work before a model is chosen
  (`presets.ForAccount`), answer 409 `not_configured` when the settings
  cannot build a provider, and 502 with the provider's typed error
  (`auth_failed`, `rate_limited`, `bad_output`, `provider_error`; 504
  `timeout`) when the call fails. `POST /api/v1/settings/ai/reset-prompt`
  restores `household.DefaultPersonaPrompt` (CTXD PROMPTS.MD §4). No
  response holds the key. `Service.HarnessConfig` reads the settings for
  each skill run, so a saved change applies to the next run.
- Dinners (`dinners.Service`). `POST /api/v1/dinners` adds a Regular from
  `{name}` or `{name, recipe_url}` (an https:// link, kept in
  `recipe_url` for grounding and never returned in a dinner response, as
  it may carry a personal token) with status and origin `regular` and
  grounding_status `pending`, answering 201 with the dinner. The name is
  always the household's: no part of a link is ever used as a dinner's
  name, and a link without a name is 400. A name
  is trimmed, at most 120 characters and must keep a word after
  `harness.Normalize`, which sets `normalized_name`. A name
  `harness.Similar` to an unarchived dinner's, or a link one already has
  (the source_url of the recipe it is grounded in, else its pasted
  `recipe_url`), is 409 `duplicate_dinner`; archived names may be added
  again.
  `GET /api/v1/dinners?status=regular|quest_log` returns
  `{dinners: [...]}` ordered by normalized name, each with `photo_url`
  (`/media/photos/{photo_id}-400.jpg`, or null) and `grounding_status`;
  any other status is 400. The Quest log (`quest_log`) is the dinners
  accepted and not yet scheduled: one on the plan for the household's
  current week (by its timezone) or a later week is left out. `PATCH /api/v1/dinners/{id}` edits `name`
  (deduplicated against the other unarchived dinners), `method`
  (`stovetop` / `oven` / `crockpot` / `grill` / `other`), `minutes_total`
  (1–1440) and `difficulty` (`easy` / `medium` / `hard`); invalid values
  are 400 `invalid_dinner` and change nothing. `DELETE /api/v1/dinners/{id}`
  sets status `archived` and keeps the row; `POST
  /api/v1/dinners/{id}/favorite` toggles `is_favorite`. Every write records
  the session's subject as `updated_by`; an unknown id is 404.
- Dinner grounding. `POST /api/v1/dinners` also enqueues a `ground_dinner`
  job and answers with its `grounding_job_id` without waiting for it
  (`GET /api/v1/jobs/{id}` reports it); if the job cannot be enqueued it
  still answers 201, with grounding_status `failed` (`grounding_error`)
  and `grounding_job_id` null. The job grounds the dinner's
  `recipe_url` directly with `GroundURL`, else its name with `Ground`
  (see Recipe grounding), and sets grounding_status `grounded` with the
  recipe's recipe_id, photo_id, method, minutes_total and difficulty,
  clearing the pasted `recipe_url` (the recipe replaces it), or
  `failed` with `grounding_error` — also the job's failure code — one of
  `no_recipe_found`, `recipe_unusable` (the pasted link), `search_not_configured`,
  `search_failed`, `ai_failed`, `interrupted` (the server stopped) or
  `grounding_error`. Each job carries the dinner's `grounding_revision`
  from when it was enqueued; renaming a dinner raises the revision, and
  renaming a `pending` one enqueues a job for the new name, clearing its
  `recipe_url` when the normalized name changes.
  A job for a dinner archived, no longer `pending` or at another revision
  does no work, and a job only writes a dinner still unarchived, `pending`
  and at its revision, so a fix-up made meanwhile wins and an earlier job
  for an old name — queued or running — never decides the outcome.
  Archiving or renaming a dinner also stops its running job where it is,
  so no further search, page fetch, photo download or model call is made
  for it. A
  method, minutes_total or difficulty a member saved while the dinner was
  `pending` is kept; grounding fills only those left empty. At boot, after
  the job runner fails the jobs a previous process left running, every
  unarchived dinner still `pending` with no `ground_dinner` job for its revision
  queued or running is set `failed` with `interrupted`, so no dinner stays
  pending with nothing left to ground it.
  `GET /api/v1/dinners/{id}/recipe-candidates`
  searches for the dinner's name and returns `{candidates: [{candidate_id,
  title, site_name, photo_url, minutes}]}`, at most 3, ranked as `Ground`
  ranks them; it fetches pages but downloads no photo and runs no model,
  so `photo_url` is set only for a page already stored. `candidate_id` is
  an opaque id the server resolves to the page it last offered that dinner
  (kept in memory, so a restart means fetching the candidates again) at
  the `grounding_revision` it was searched at, so a candidate offered
  before any rename is refused even if the name is changed back; the
  page's URL, which may carry a token, is never returned or cached. An
  unknown dinner is 404, recipe search not set up 409 `not_configured` and
  a search failure 502 `search_failed`; any other failure, such as a
  stored recipe that cannot be read, is 500, logged with a fixed reason
  and the dinner id only — never a candidate's URL or dependency error
  text.
  `POST /api/v1/dinners/{id}/ground` takes exactly one of `{candidate_id}`,
  `{recipe_url}` (both grounded with `GroundURL`, setting `grounded`) or `{manual: {ingredients[], steps[], minutes,
  method}}` (1–100 ingredient lines, up to 100 steps, each non-blank and at
  most 500 characters; minutes 1–1440). A manual recipe is stored as typed
  — no model runs — as a `recipes` row whose source_url is
  `mealquest:manual:{recipe_id}` and whose raw_jsonld is a schema.org
  Recipe built from it, and sets `manual` with its method and minutes, no
  photo and the dinner's difficulty unchanged. Either clears the dinner's
  pasted `recipe_url`: the recipe chosen replaces it. It answers 200 with
  the dinner; an invalid body is 400, an unknown dinner 404, a page that cannot
  ground it 422 `ungroundable` (the dinner is unchanged), recipe search not
  set up 409 `not_configured`, a search or model failure 502
  `search_failed` / `ai_failed` (only the reason code is logged, never the
  provider's text), and grounding a candidate or link that runs past 90
  seconds (fetch, photo and `normalize_recipe`, all in the request) 504
  `ground_timeout`, leaving the dinner unchanged. An archived dinner is 409
  `dinner_archived`. A candidate or link is written only if the dinner is
  still as it was before grounding began: one renamed or grounded some
  other way meanwhile is 409 `ground_conflict`, and one archived meanwhile
  409 `dinner_archived`, either leaving the dinner unchanged. A method,
  minutes or difficulty a member edits meanwhile is kept over the recipe's.
  Any other failure grounding a candidate or link, or recording it on the
  dinner, is 500, logged with a fixed reason and the dinner id only —
  never the link, which can carry a personal token, or dependency error
  text.
- Side quest suggestions (`suggest.Service`). `POST
  /api/v1/suggestions/batches` creates a `manual` batch and enqueues its
  `suggestion_batch` job, answering 202 `{batch_id, job_id, trigger,
  status}` without waiting; while a batch is `pending` or `running` it
  answers 202 with that batch instead, so at most one runs at a time (the
  check, the batch and its job share one write transaction). The job reads the
  household context, the deck's names and every declined name, runs
  `suggest_side_quests` for 5 Side quests (scope: the batch id), then
  grounds the survivors in the model's order with `Grounder.GroundMany`,
  as many as are still wanted at a time, until 5 are kept or none are
  left. A Side quest no recipe grounds is dropped; each round's grounded
  ones are stored before the next — a dinner with status `suggested`,
  origin `side_quest`, grounding_status `grounded` with the recipe's
  recipe_id, photo_id, method, minutes_total and difficulty and the
  model's cuisine, and a `pending` suggestion with the name it is
  suggested under, the reason and `cited_regular_ids` — unless, since the
  job read its input, a dinner it repeats was added, a suggestion it
  repeats passed, or a Regular it cites renamed or archived (checked in
  the storing transaction). The job publishes
  `{batch_id, stage, suggestion_ids, dropped}` as it goes (`suggesting`,
  `grounding`, `done`). It fails with `no_regulars`, `ai_failed` (the
  skill or `normalize_recipe` failed; no Side quest is made up instead,
  D-017), `search_not_configured` or `search_failed`, and the batch is set
  `failed`; Side quests stored before the failure stay in the deck. If
  recording the batch's end fails, its job fails; the job runner retries
  a job's final status write with backoff until it lands, each attempt
  bounded at 5 seconds; at shutdown an attempt that fails or times out is
  not retried, and the next start fails the job as interrupted. Recording
  the batch's end is bounded the same way (one attempt, 5 seconds, made
  even during shutdown), so a busy writer cannot hold a stopping worker;
  a batch whose end timed out is left `running` with its job. Each time a
  batch is started, before the active batch is selected, every batch
  still `pending` or `running` with no job queued or running (the process
  stopped, or recording its end failed) is set `failed`, so it does not
  block the next.
  `GET /api/v1/suggestions/deck` returns `{suggestions: [{id, dinner_id,
  name, photo_url, difficulty, method, minutes_total, cuisine, reason,
  cited_regular_ids, status, dinner_status, decided_at, created_at}],
  active_batch, failed_batch}`: the `pending` suggestions whose dinner is
  still `suggested`, oldest first; the batch `pending` or `running`
  (`{batch_id, job_id, trigger, status}`, or null); and the newest batch
  when it `failed` (same shape, or null; a later batch, done or under
  way, supersedes the failure), read in one snapshot, so a client that
  did not start the batch can follow its job and learns of a failure it
  was not following. At most one of the two is set. `POST /api/v1/suggestions/{id}/accept` (Try it) sets the
  suggestion `accepted` and moves its dinner to `quest_log`; `POST
  /api/v1/suggestions/{id}/decline` (Pass) sets it `declined` and archives
  the dinner; every later batch, and the household context's passed list,
  reads the suggestion's own name as declined, so renaming the archived
  dinner changes nothing. Both record
  the session's subject as the dinner's `updated_by` and answer 200
  `{suggestion, replenish_job_id}`. Repeating a decision changes nothing
  and answers 200; the opposite decision is 409 `suggestion_decided`, a
  suggestion whose dinner left the deck some other way 409
  `dinner_not_suggested`, and an unknown id 404. A decision that leaves
  fewer than 5 in the deck starts a `replenish` batch, or joins the batch
  already pending or running, and returns that batch's job id either way
  (`replenish_job_id` is null when the deck keeps 5 or more). The batch and its job are written in
  the decision's own transaction, so a decision is never recorded without
  the replenish it calls for; when they cannot be written, neither is the
  decision, and a retry decides again.
- Weekly plans (`plan.Service`). `{weekStart}` is a household-local Monday
  (`YYYY-MM-DD`); any other value is 400 `invalid_week`. `GET
  /api/v1/plans/{weekStart}` returns `{id, week_start, week_label, status,
  ai_note, calendar_synced_at, updated_at, days}`, or 404 when the week has
  no plan; `days` are the seven days Monday first, each `{date, dinner: {id,
  name, photo, minutes, method, origin}, is_new, start_by}`. `is_new` means
  the dinner was planned from the Quest log; `start_by` is the household's
  dinner time on that date minus `minutes_total`, computed on read, for a
  crock-pot dinner with known minutes, else null. `POST
  /api/v1/plans/{weekStart}/draft` takes `{confirm?}` and answers 202
  `{jobId}` for a `draft_week` job, or for the one already queued or
  running for the week — a confirmed one when the draft is confirmed, so a
  confirmed draft never joins an unconfirmed job. The job runs the `draft_week` skill on every Regular
  and every Quest log dinner no other plan for the current or a later week
  has, with the household context and last week's dinners, and stores the
  week: a new plan is `draft`; a redraft keeps each day's
  `calendar_event_id`, and a `synced` or `changed` plan becomes `changed`.
  Drafting over a `synced` or `changed` plan needs `{"confirm": true}`, else
  409 `confirm_required`; a plan synced while an unconfirmed draft runs is
  left as it is and the job fails `confirm_required`. The job reads the
  plan's revision before draft_week runs; if the plan was edited (a swap,
  or another draft stored) by the time the draft is stored, the plan is
  left as it is and the job fails `plan_changed`. A pool under seven
  dinners fails the job `not_enough_dinners`, draft_week failing
  `ai_failed`. `PUT /api/v1/plans/{weekStart}/days/{date}` takes
  `{dinner_id}`, a Regular or Quest log dinner, swaps the day, recomputes
  `is_new`, moves a `synced` plan to `changed`, and answers 200 with the
  plan. A date outside the week is 400 `invalid_date`, an unknown dinner or
  a week with no plan 404, and any other dinner 409 `dinner_not_plannable`.
- `/health` — liveness.
- Everything else serves the embedded PWA with SPA fallback.

## AI layer

- **Provider interface** (`ai/llm`) with adapters for OpenAI-compatible,
  Anthropic and Ollama endpoints, called over plain HTTP — no vendor SDKs.
- **Skill harness** (`ai/harness`): prompt layering, validation of the
  structured output, and **one repair retry**; if the repair also fails the
  run fails — no silent fallback. Every run writes one `ai_runs` row when it
  ends, covering the repair turn too, with status `ok`, `repaired`,
  `bad_output` or `provider_error`; a prompt that cannot be rendered is
  `bad_output` and a provider that cannot be built is `provider_error`, both
  with no turns. The harness builds the provider itself from the settings
  and the decrypted API key (`presets.FromSettings`), so the key it redacts
  is always the key the provider sends. Request and response
  bodies are redacted before they are stored: the decrypted API key,
  Authorization and API-key header values, Google Chat webhook URLs and URLs
  with key-like query parameters are removed, inside JSON strings too; the
  skill, provider and model columns are redacted the same way. Writing the
  row is bounded: a run whose work ended in time must write its row before
  the skill's deadline or it fails with `llm.ErrTimeout` and its row records
  the timeout; a failed run's row has a further fixed bound, and if even
  that write fails the run returns the write error with its own and no row
  is stored. A stored row's status therefore always matches what the caller
  received.
  `harness.Purge` deletes rows older than `harness.RunRetention` (30 days),
  clearing the `suggestion_batches` and `week_plans` references first. The
  harness does not call it: the scheduler's daily `ai_runs` purge
  (Feature 20) must call `harness.Purge` with `now - RunRetention`.
- **Concurrency guards** (`harness.Guard`, one per process): `harness.Run`
  takes a scope key, and concurrent runs of one skill with one scope share a
  single in-flight run and its result, error included (single-flight); runs
  with different scopes are independent. A semaphore lets at most
  `harness.DefaultMaxCalls` (2, set by `NewGuard`) LLM calls run at once;
  waiting for a slot honours the context. Each skill declares a timeout
  (`harness.DefaultTimeout`, 60 s, unless it needs another) bounding the
  whole run, slot waits and repair turn included; a run past it fails with
  `llm.ErrTimeout`. A caller whose context ends gets the context error; the
  shared run is cancelled, freeing its slot, once no caller waits for it.
- **Prompt layering** (D-007): each skill version's prompt is a
  `text/template` file embedded from `ai/skills/prompts/`
  (`<skill>.v<version>.tmpl`). The system prompt is always, in order:
  [1] the admin persona (`ai_settings.system_prompt`), [2] the skill's
  instructions, [3] the household context block, [4] the output contract
  with the JSON Schema — the contract is always last. The user message comes
  from the skill's input. `checksums.txt` pins every template; changing one
  without a version bump fails the tests.
- **Household context** (`harness.BuildContext`): layer 3 holds today's date
  and weekday in the household timezone, the Regulars, the Quest log, the
  last 60 passed Side quests and the plans of the last four ISO weeks. A
  character budget trims the oldest plans first, then the oldest passes;
  Regulars and the Quest log are never trimmed.
- **Name dedupe** (`harness.Normalize`, `harness.Similar`, `harness.Deduper`):
  a dinner name is normalized by lowercasing, dropping combining marks and
  folding precomposed letters to their base letters through a table generated
  from Unicode data (`gen_folds.py`), so composed and decomposed spellings
  match, stripping punctuation, dropping filler words (easy, best, homemade, quick, simple,
  the, a, recipe, classic) and collapsing whitespace. Two names are the same
  dinner when the token-set (Jaccard) similarity of their words is at least
  0.85. A `Deduper` over the Regulars, Quest log, pending suggestions and
  passed names rejects a suggestion that repeats any of them.
- **Side quest suggestions** (`skills.SuggestSideQuests`, skill
  `suggest_side_quests` v1): asks for twice the Side quests needed, at the
  household temperature, with the household context plus the pending Side
  quests. Malformed replies (bad JSON, a missing field, a bad enum, an
  empty name) get the one repair turn. Well-formed Side quests that break a
  rule are dropped, not repaired, and listed in `Dropped` with the rule:
  one that repeats a Regular, Quest log dinner, pending Side quest, any
  passed name (the whole history, not only the 60 in the context) or an
  earlier Side quest of the batch; one citing no Regular or a name that is
  not a Regular; and one whose reason is over 180 characters or names none
  of its cited Regulars. Cited names are rewritten as the household writes
  them. Grounding each survivor in a recipe is the caller's next step.

## Recipe grounding

- Dinners are grounded in real recipe pages found through recipe search
  (`searxng` or `brave`, per `recipe_search_settings`).
- Fetching is **SSRF-safe** (`recipes.Fetcher`): https:// only, at every
  redirect hop. Its dialer resolves the host once and refuses unless every
  answer is public — every non-globally-reachable IPv4 special-purpose
  range (loopback, RFC 1918, link-local, carrier-grade NAT, documentation
  TEST-NETs, benchmarking 198.18/15, multicast, reserved and more) is
  refused, IPv4-mapped forms included, and IPv6 must be global unicast
  (2000::/3) outside documentation, 6to4, Teredo and other special ranges —
  then dials the address it checked, so a
  rebinding DNS answer cannot swap in a private one. Keep-alives and
  environment proxies are off, so every request dials and checks afresh. At
  most 3 redirects; a body over 2 MB is an error, never truncated; 10 s for
  the whole fetch; a descriptive `User-Agent`; requests to one host at least
  1 s apart, counted from requests actually sent (an abandoned wait holds no
  slot); 2xx pages cached by URL for an hour (64 pages, oldest evicted), and
  callers get their own copy of a cached body.
- Recipe data is parsed from schema.org `Recipe` JSON-LD
  (`recipes.Extract`): a small targeted scan of the page's
  `application/ld+json` scripts and `og:site_name` (no HTML parser
  dependency), finding the Recipe node at a script's top level, in a
  top-level array or in an `@graph`, with `@type` a string or an array of
  terms or absolute schema.org IRIs (`https://schema.org/Recipe`). An image
  or publisher given as an `{"@id": ...}` reference resolves to the node
  with that `@id` anywhere in the page's JSON-LD. The largest image (by ImageObject width × height; the first of equals) is
  resolved to an absolute URL; instructions — one string of lines, or a
  list of strings, HowToSteps and HowToSections — are flattened in order;
  ISO-8601 prep/cook/total times become minutes, total derived as prep +
  cook when only those are given; the site name is the publisher's, else
  `og:site_name`, else the host. A page with no Recipe is `ErrNoRecipe`,
  with two is `ErrManyRecipes`; a shape schema.org does not allow — such as
  a non-ISO-8601 duration like "20 minutes" or an empty one, or an image
  `@id` no node has — is `ErrMalformedRecipe`, never guessed at.
- Recipe search (`recipes.SearchProvider`: `Search(ctx, query, n)` →
  `[]Candidate{URL, Title, Snippet}`, n 1–20) appends " recipe" to a query
  without that word. SearXNG is called at
  `{recipe_search_settings.base_url}/search?q=…&format=json&categories=general` —
  directly, not through the Fetcher, since this configured endpoint may be
  on the internal network; Brave calls the Brave Web Search API with the
  key decrypted from `recipe_search_settings.api_key_ct`
  (`recipes.NewSearchProvider`). Redirects are not followed; 401/403 is
  `ErrSearchRefused`, 429 `ErrSearchRateLimited`, other non-2xx a
  `SearchStatusError`, and a body that is not the provider's JSON (SearXNG
  without a `results` array; Brave without `"type": "search"` or with a
  `web` object lacking `results` — Brave omits `web` when nothing matched),
  over 1 MB or with a result lacking an absolute http(s) URL
  `ErrSearchResponse`.
- Ranking (`recipes.Rank`) fetches and extracts the top 5 candidates
  concurrently, rejects any whose recipe lacks ingredients or an image, and
  orders the rest by the token similarity of the recipe's title to the dish
  name (`harness.Similarity`, 0–1) plus a quarter for each of ingredients,
  image, steps and total minutes; ties keep search order. No unranked
  candidate is dropped silently: when some are ranked, they are returned
  with `ErrCandidatesUnranked`; when none are, the error is
  `ErrNoUsableRecipe`. Either is joined with each unranked candidate's URL
  and reason (fetch failure, no valid recipe, missing ingredients or
  image), so the caller decides per item what a failure means.
- Ingredient lines are normalized by the `normalize_recipe` v1 skill
  (`skills.NormalizeRecipe`), always at temperature 0 whatever the
  household's setting. From the title, raw ingredient lines, steps and
  JSON-LD times it returns `{ingredients: [{name, qty, unit, note}], method,
  difficulty, minutes_total}`. Validation requires exactly one ingredient
  per input line, in order (a name matching another line but not its own is
  out of order), each with a name; a mismatch gets the harness's repair
  turn. Code-derived values win: a positive JSON-LD total is
  `minutes_total`; "slow cooker" or "crock pot" in the title or a step makes
  the method `crockpot`; a difficulty the model leaves null is derived —
  hard over 10 steps or 90 minutes, easy at 5 steps or fewer and 45 minutes
  or less, else medium, a crockpot's minutes not counting.
- Photos are cached locally (`photos`) and processed with
  `golang.org/x/image`. `recipes.PhotoCache.DownloadPhoto` downloads through
  the Fetcher's `FetchImage` (JPEG, PNG, WebP or GIF content types, 8 MB
  cap, not cached), decodes the image (a GIF's first frame) and refuses a
  decoded image under 300 px on its short side, recording nothing.
  `CachePhoto` then records it in the caller's transaction: it scales the
  image so its long side is 1200 px and 400 px, and stores each as JPEG
  quality 82 at `$MEALQUEST_DATA_DIR/media/photos/{id}-{size}.jpg`. Images
  are deduplicated by the SHA-256 of the downloaded bytes; the writer's
  single connection serialises transactions, so one image is never
  recorded twice. Files written for a transaction that rolls back are
  removed by the caller (`RemovePhotoFiles`).
- Grounding (`recipes.Grounder`) ties these together.
  `Ground(ctx, dishName, query)` searches for the top 5 candidates and ranks
  them all, stored and new alike: a candidate whose `source_url` is already
  stored is scored on its stored recipe (title, ingredients, steps, times,
  photo source) and its page is never refetched; every other candidate is
  fetched and extracted. Ranking alone decides between them. It then tries
  each ranked candidate best first. A stored one is returned as it is, so a
  stored recipe is reused only when it ranks above every new candidate that
  grounds. A new one is grounded in this order: (1) `DownloadPhoto` —
  the download and validation step of photo caching — (2)
  `normalize_recipe`, then (3) one **transaction** that runs `CachePhoto`
  (scales and writes the photo files and records the `photos` row) and
  stores the recipe, ingredients and steps (an existing `source_url` is
  updated in place); files written for a rolled-back transaction are
  removed. It returns `Grounded{recipe_id,
  photo_id, method, minutes_total, difficulty, source_url, site_name}`. A
  candidate whose photo fails or whose model output is unusable
  (`llm.ErrBadOutput`) is passed over; when none grounds, the error is a
  typed `*recipes.ErrUngroundable` listing every reason (no candidates,
  Rank's per-candidate reasons, photo or normalize failures). Failures
  that are not the candidates' — search, model provider, database — are
  returned as they are. `GroundURL(ctx, url)` does the same for one pasted
  URL. `Candidates(ctx, dishName, query, n)` returns the first n of the
  candidates `Ground` would try, without grounding any. Search failures
  wrap `recipes.ErrSearchFailed` and model provider failures
  `recipes.ErrNormalizeFailed`. `GroundMany` grounds items 3 at a time and returns one result per
  item, so one failure never affects its siblings. A recipe is usable only
  with ingredients, an image and, if given, a positive total time.

## Scheduler

Times are in the household timezone. The times of `day_of_reminder`
(`reminder_time`) and `rotation_check` (`review_time`), and the day and time
of `suggestions_weekly`, `plan_draft` and `shopping_chat_autosend`, are
household settings in `household_settings` (`dinner_time` is also stored
there); the table shows their defaults. `week_rollover` and the `ai_runs`
purge are fixed and not configurable.

| Schedule | When | Does |
| --- | --- | --- |
| `suggestions_weekly` | Sunday 08:00 | New batch of Side quests. |
| `plan_draft` | Friday 09:00 | AI drafts next week's plan. |
| `shopping_chat_autosend` | Saturday 08:00 | Sends the Supply run to Google Chat. |
| `day_of_reminder` | Daily 16:30 | Tonight's dinner reminder. |
| `rotation_check` | Daily 09:00 | Asks "Keep it in rotation?" for tried dinners. |
| `week_rollover` | Monday 00:05 (fixed) | Freezes last week's Supply run into the Supply archive. |
| `ai_runs` purge | Daily (fixed) | Calls `harness.Purge` to delete `ai_runs` rows past 30-day retention (Feature 20). |

Runs are recorded in `schedule_runs` and executed through `jobs`.

## Notifications

- Events: `plan_ready`, `list_ready`, `day_of_reminder`, `new_suggestions`,
  `rotation_review`; each can go to Web Push and/or Google Chat per
  `notification_prefs`.
- Delivery goes through `notification_outbox`.
- Web Push uses VAPID (`webpush-go`). Google Chat uses the stored incoming
  webhook.

## Google Calendar

- One-way push: the app is the source of truth; edits in Google Calendar do
  not come back.
- Connected once with Google OAuth in Settings; one event per planned
  dinner.
- **Deterministic event ids**; each plan day records its
  `calendar_event_id`.

## Frontend

- React 19, TypeScript, Vite, React Router 7, TanStack Query 5,
  `vite-plugin-pwa`, `lucide-react` (stroke 2.75).
- CSS Modules over `tokens.css`; no Tailwind, no Framer Motion, no
  hard-coded hex. Self-hosted Bagel Fat One and Figtree.
- Tests: Vitest, Testing Library, jsdom.
- Offline cache for the current plan, shopping list and recipes.
- Logout (`POST /auth/logout`) answers with `Clear-Site-Data: "storage",
  "cache"`, wiping the offline marker, the service worker and its caches; the
  worker leaves that request to the browser, as it does every network-only one.

| Route | Screen |
| --- | --- |
| `/` | Home |
| `/onboarding` | Add regular dinners |
| `/plan` | Weekly plan (Draft board) |
| `/plan/shopping` | Shopping list (Supply run) |
| `/plan/shopping/archive` | Supply archive |
| `/discover` | Swipe suggestions (Side quests) |
| `/quick` | Quick dinners |
| `/cook-from-pantry` | Cook from what I have |
| `/dinner/:id` | Recipe / dinner detail (Kitchen reading mode on tablet) |
| `/settings` | Settings home |
| `/settings/household` | Household settings |
| `/settings/regulars` | Regular dinners |
| `/settings/calendar` | Google Calendar connection |
| `/settings/notifications` | Notifications and Google Chat |
| `/settings/ai` | Admin · AI assistant |

`/settings/household`, `/settings/regulars` and `/settings/calendar` come from
CTXD; the other routes are defined in the UI handoff.

First run: while `household_settings.onboarding_completed_at` (read from
`GET /api/v1/settings/household`) is null, every route redirects to
`/onboarding`. Onboarding is a full screen outside the app shell, in three
steps: the welcome (`/onboarding`), the Regulars (`/onboarding/regulars`,
mock 1h, adding and archiving through the dinners API optimistically; Next
waits for three saved Regulars, and a common pick is not offered once a
Regular has a name `harness.Similar` to it — the PWA's copy of the name
rules in `web/src/dinners/names.ts` folds letters with
`web/src/dinners/letterFolds.json`, which `gen_folds.py` writes from the
same table as the server's, and runs the same cases,
`web/src/dinners/testdata/dedupe_cases.json`, as the server's) and
notifications (`/onboarding/notifications`: what MealQuest notifies about,
the iOS install hint, "Turn on notifications" and "Maybe later"). "Maybe
later", and "Turn on notifications" once notifications are enabled, send
`PATCH /api/v1/settings/household` with `{"onboarding_completed": true}`; the server sets `onboarding_completed_at`
to the time of that update (keeping the first time if it is already set;
the field itself is read-only and `false` is refused) and the app goes Home.
Completing needs at least three Regulars (`household.MinRegulars`); with
fewer the server answers 400 `invalid_settings` naming
`onboarding_completed` and changes nothing. Step 3 can be reached with
fewer — by its URL, or after another member removes a Regular — so it then
shows the server's reason with an "Add Regulars" button back to step 2.
Notification permission is never asked for on page load; until Feature 19
wires Web Push, "Turn on notifications" asks for nothing and says plainly
that notifications are not on yet, offering "Finish setting up" (which
completes onboarding as above) instead of a silent success. Offline on the
saved pages the settings cannot be read (the service worker never caches `/api/v1/settings`),
so the saved pages open as they are.

Regular dinners (`/settings/regulars`) lists the Regulars from
`GET /api/v1/dinners?status=regular`, each with its photo, "method ·
minutes", favorite heart and grounding badge ("Finding recipe…" while
`pending`, "Recipe found" when `grounded` or `manual`, "Needs a recipe" when
`failed`). The add field takes a name or a pasted `https://` recipe link.
Tapping a Regular opens an edit sheet (name, method, minutes, difficulty;
remove archives it after a confirmation); tapping one that needs a recipe
opens "Fix recipe" (pick one of the recipe candidates, paste a link or type
the ingredients and steps in). Adds, removals and the heart show at once and
are undone if the server refuses them. While any Regular is `pending`, the
list is read again every 3 seconds (not while a change is in flight), and
polling stops once none is.

## Boundaries and gotchas

- **Secrets never reach the client** — API keys, OAuth tokens and the Chat
  webhook stay server-side, encrypted at rest.
- **The app is the source of truth for the plan**; Google Calendar is a
  one-way mirror.
- **Never auto-promote to Regular** (D-021).
- **Quick** means cooked in a crock pot **or** ≤ 30 minutes prep to table.
- **Weeks run Monday–Sunday** in the household timezone.
- **No silent fallbacks** — fail fast and report.
- **iOS Web Push requires the PWA to be installed to the Home Screen**; show
  an install hint.