Add plan_draft schedule drafting next week (Feature 20)

Register plan.Service.DraftSchedule() as the plan_draft schedule: at the
household's plan_draft_dow/time it runs StartScheduledDraft for the ISO
week after the fire date (period plan_draft:YYYY-Www of that week), which
enqueues the same draft_week job as POST /plans/{weekStart}/draft. A
stored draft emits plan_ready:{weekStart} from the job; an AI failure is
recorded in ai_runs and emits nothing. A week that already has a plan is
left as it is.

Fake-clock tests in internal/scheduler cover drafting when absent,
skipping when present, no duplicate after restart, the AI failure path
and a settings change without restart.

Maestro task 598184a8-56fe-4ed8-887b-7780c0aebf12.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

cmd/server/main.go

@@ -30,6 +30,7 @@ import (
 	"code.cubecraftlabs.com/Maniacs_Mansion/Meal-Quest/internal/plan"
 	"code.cubecraftlabs.com/Maniacs_Mansion/Meal-Quest/internal/push"
 	"code.cubecraftlabs.com/Maniacs_Mansion/Meal-Quest/internal/recipes"
+	"code.cubecraftlabs.com/Maniacs_Mansion/Meal-Quest/internal/scheduler"
 	"code.cubecraftlabs.com/Maniacs_Mansion/Meal-Quest/internal/secretbox"
 	"code.cubecraftlabs.com/Maniacs_Mansion/Meal-Quest/internal/server"
 	"code.cubecraftlabs.com/Maniacs_Mansion/Meal-Quest/internal/shopping"
@@ -187,17 +188,29 @@ func run(cfg config.Config, logger *slog.Logger) error {
 		return err
 	}
 
+	sched, err := scheduler.New(db, settings, logger)
+	if err != nil {
+		return err
+	}
+	if err := sched.Register(suggestSvc.WeeklySchedule()); err != nil {
+		return err
+	}
+	if err := sched.Register(planSvc.DraftSchedule()); err != nil {
+		return err
+	}
+
 	router, err := server.NewRouter(logger, version, cfg.TrustedProxy, authFlow, sessions, settings, search, ai, chatSvc, dinnerSvc, suggestSvc, planSvc, shopSvc, googleSvc, pushSvc, ideasSvc, runner, photos, frontend.Assets())
 	if err != nil {
 		return err
 	}
 
 	// Start fails the jobs a previous process left running before any
-	// worker runs. The job workers and the notification worker stop, and
-	// are waited for, before the database closes.
+	// worker runs. The job workers, the notification worker and the
+	// scheduler stop, and are waited for, before the database closes.
 	workersCtx, stopWorkers := context.WithCancel(ctx)
 	defer func() {
 		stopWorkers()
+		sched.Wait()
 		runner.Wait()
 		notifier.Wait()
 	}()
@@ -213,6 +226,11 @@ func run(cfg config.Config, logger *slog.Logger) error {
 	if _, err := dinnerSvc.FailInterruptedGroundings(ctx); err != nil {
 		return err
 	}
+	// The scheduler starts last: its catch-up may enqueue jobs and
+	// notifications, so their workers must already run.
+	if err := sched.Start(workersCtx); err != nil {
+		return err
+	}
 	srv := &http.Server{
 		Addr:              cfg.HTTPAddr,
 		Handler:           router,

docs/ARCHITECTURE.md

@@ -309,20 +309,24 @@ automatic** (D-021).
   `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`,
+  left. A Side quest no recipe grounds is dropped. A `replenish` batch
+  stores each round's grounded ones before the next; a `scheduled` or
+  `manual` batch, which notifies, holds them and stores them all in one
+  transaction once its last round is done. Each is stored as 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
+  `failed`. A failed `scheduled` or `manual` batch has stored nothing; a
+  failed `replenish` batch's 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
@@ -775,8 +779,8 @@ 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. |
+| `suggestions_weekly` | Sunday 08:00 | New batch of Side quests (`suggest.Service.WeeklySchedule`). |
+| `plan_draft` | Friday 09:00 | AI drafts next week's plan (`plan.Service.DraftSchedule`). |
 | `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. |
@@ -785,6 +789,42 @@ purge are fixed and not configurable.
 
 Runs are recorded in `schedule_runs` and executed through `jobs`.
 
+`internal/scheduler` runs them. A schedule registers a name, a rule read
+from the settings (`scheduler.Daily`, `Weekly` or `WeeklyForNextWeek`) and a
+run function; settings changes reach it through `SettingsService.Subscribe`
+and move the next fire time without a restart. Each fire time has a period
+key (`Rule.PeriodKey`): the local date for a daily rule, the ISO week of the
+fire date for a weekly one, or of the week after for `WeeklyForNextWeek`
+(`plan_draft`). Inserting `schedule_runs(schedule, period_key)` is the claim;
+a period already claimed is skipped, so a period runs at most once and a
+failed run is not retried. A claim that fails is not skipped: the pass is
+repeated every `scheduler.ClaimRetry` until the claim succeeds. A settings
+change is recorded with the instant it was saved, and fire times up to that
+instant follow the old rule and later ones the new, however late the
+scheduler applies it. At boot only each schedule's most recent fire
+time is considered: it runs if it passed within 24 hours, and is otherwise
+claimed without running (logged as missed) so it never runs later. DST
+follows `household.LocalTime`: a skipped time fires later by the gap, a
+repeated time fires once, at its first occurrence.
+
+Each schedule is built by the service that does its work, as a
+`scheduler.Schedule` value, and registered in `cmd/server` before the
+scheduler starts: `suggest.Service.WeeklySchedule()` is `suggestions_weekly`.
+Its run calls `StartScheduledBatch(period.Week)`, which starts the same
+batch and job as `POST /api/v1/suggestions/batches` (`StartBatch(manual)`),
+with the `scheduled` trigger, and returns at once; the batch runs as a job. When a
+batch is already pending or running the period starts none and logs it.
+`plan.Service.DraftSchedule()` is `plan_draft`, keyed by the ISO week
+after the fire date (`scheduler.WeeklyForNextWeek`). Its run calls
+`StartScheduledDraft(period.Week)`, which enqueues the same `draft_week`
+job as `POST /api/v1/plans/{weekStart}/draft` without confirm, so a stored
+draft emits `plan_ready:{weekStart}` from the job and a failed one (e.g.
+`ai_failed`, recorded in `ai_runs`) emits nothing. When the week already
+has a plan, in any status, it enqueues nothing; when a draft of the week
+is already queued or running it joins it. Either way it logs it.
+Fake-clock tests of a schedule live in `internal/scheduler`, where
+`SetClock` is visible to tests.
+
 ## Notifications
 
 - Events: `plan_ready`, `list_ready`, `day_of_reminder`, `new_suggestions`,
@@ -793,8 +833,9 @@ Runs are recorded in `schedule_runs` and executed through `jobs`.
 - Where each event is emitted, with its idempotency key. Each is stored by
   `notify.EmitTx` in the transaction of the change it is about, so the
   change and its notification commit together or not at all:
-  - `plan_ready:{weekStart}` (URL `/plan`) when a draft job stores the
-    AI's draft of a week. Only the first draft of a week notifies: a
+  - `plan_ready:{weekStart}` (URL `/plan`) when a draft job, asked for
+    by the API or the `plan_draft` schedule, stores the AI's draft of a
+    week. Only the first draft of a week notifies: a
     redraft (Rebuild week) has the same key, and a manual swap never
     drafts. Its title is "Next week's plan is ready" (or "This week's",
     or "The plan for the week of Oct 26") and its body the drafted
@@ -802,12 +843,19 @@ Runs are recorded in `schedule_runs` and executed through `jobs`.
   - `list_ready:{weekStart}` (URL `/plan/shopping`) when a week's Supply
     run is first created, by `POST /shopping/{weekStart}/generate` or
     after Push to Calendar; remaking it does not notify.
-  - `new_suggestions:{batchId}` (URL `/discover`) when a `scheduled` or
-    `manual` suggestion batch first adds Side quests to the deck, in the
-    transaction storing them, so they are announced even when a later
-    round or recording the batch's end fails; it names that round's Side
-    quests, and a later round's emit stores nothing. A `replenish` batch
-    never notifies.
+  - `new_suggestions:{YYYY-Www}` for a `scheduled` batch, the ISO week it
+    was scheduled for (carried in its job, so a batch caught up late is
+    still keyed by its own week), or `new_suggestions:{batchId}` for a
+    `manual` one (URL `/discover`), once, when the batch's last round is
+    done, in the one transaction that stores all its Side quests, so they
+    and their notification commit together or not at all. Its title
+    counts every Side quest the batch stored ("5 new Side quests are in")
+    and its body names them. A batch that fails at any round, as when
+    `suggest_side_quests` or a later round's `normalize_recipe` fails (the
+    failure is in `ai_runs`), stores no Side quest and sends nothing, so
+    no stored Side quest goes unannounced. A `replenish` batch never
+    notifies. This is the only place `new_suggestions` is emitted; the
+    `suggestions_weekly` schedule does not emit it itself.
   - `day_of_reminder:{date}` (URL `/dinner/{id}`) through
     `plan.Service.EmitDayOfReminder(date)`, which names the dinner planned
     on the date and emits nothing for a week with no plan. The scheduler

internal/chat/messages.go

@@ -77,13 +77,15 @@ func DayOfReminder(date store.LocalDate, dinnerID, name string) notify.Event {
 	}
 }
 
-// NewSuggestions is the new_suggestions event of suggestion batch batchID,
-// made in the week starting on week, naming its Side quests.
-func NewSuggestions(batchID string, week store.LocalDate, names []string) notify.Event {
+// NewSuggestions is the new_suggestions event keyed by id, the ISO week
+// ("2026-W41") of a scheduled suggestion batch or the batch id of a manual
+// one, of the week starting on week: how many new Side quests there are,
+// and their names.
+func NewSuggestions(id string, week store.LocalDate, names []string) notify.Event {
 	return notify.Event{
 		Type:  notify.NewSuggestions,
-		Key:   string(notify.NewSuggestions) + ":" + batchID,
-		Title: "New Side quests are in",
+		Key:   string(notify.NewSuggestions) + ":" + id,
+		Title: count(len(names), "new Side quest is in", "new Side quests are in"),
 		Body:  nameList(names),
 		URL:   "/discover",
 		Week:  week,

internal/chat/messages_test.go

@@ -110,9 +110,19 @@ func TestMessages(t *testing.T) {
 			key:   "new_suggestions:01J9BATCH",
 			url:   "/discover",
 			week:  "2026-10-12",
-			title: "New Side quests are in",
+			title: "5 new Side quests are in",
 			body:  "Lemon orzo, Bibimbap, Shakshuka + 2 more",
-			chatT: "New Side quests are in: Lemon orzo, Bibimbap, Shakshuka + 2 more. Open Discover →",
+			chatT: "5 new Side quests are in: Lemon orzo, Bibimbap, Shakshuka + 2 more. Open Discover →",
+		},
+		{
+			name:  "new suggestions of a scheduled week",
+			ev:    NewSuggestions("2026-W42", week, []string{"Pho"}),
+			key:   "new_suggestions:2026-W42",
+			url:   "/discover",
+			week:  "2026-10-12",
+			title: "1 new Side quest is in",
+			body:  "Pho",
+			chatT: "1 new Side quest is in: Pho. Open Discover →",
 		},
 		{
 			name:  "rotation review",

internal/db/querier.go

@@ -24,6 +24,9 @@ type Querier interface {
 	// Marks the oldest queued job running and returns it. Returns sql.ErrNoRows
 	// when no job is queued.
 	ClaimNextJob(ctx context.Context, now sql.NullString) (Job, error)
+	// Claims one period of a schedule. The insert is the claim: 1 row means
+	// this caller owns the period, 0 rows means it was already claimed.
+	ClaimScheduleRun(ctx context.Context, arg ClaimScheduleRunParams) (int64, error)
 	// Clears every day's calendar event id: none is on the chosen calendar.
 	ClearCalendarEvents(ctx context.Context, updatedAt string) error
 	// Records that Google refuses the stored grant: clears the refresh token and

internal/db/schedule_runs.sql.go

@@ -0,0 +1,38 @@
+// Code generated by sqlc. DO NOT EDIT.
+// versions:
+//   sqlc v1.31.1
+// source: schedule_runs.sql
+
+package db
+
+import (
+	"context"
+)
+
+const claimScheduleRun = `-- name: ClaimScheduleRun :execrows
+INSERT INTO schedule_runs (id, schedule, period_key, created_at)
+VALUES (?, ?, ?, ?)
+ON CONFLICT (schedule, period_key) DO NOTHING
+`
+
+type ClaimScheduleRunParams struct {
+	ID        string
+	Schedule  string
+	PeriodKey string
+	CreatedAt string
+}
+
+// Claims one period of a schedule. The insert is the claim: 1 row means
+// this caller owns the period, 0 rows means it was already claimed.
+func (q *Queries) ClaimScheduleRun(ctx context.Context, arg ClaimScheduleRunParams) (int64, error) {
+	result, err := q.db.ExecContext(ctx, claimScheduleRun,
+		arg.ID,
+		arg.Schedule,
+		arg.PeriodKey,
+		arg.CreatedAt,
+	)
+	if err != nil {
+		return 0, err
+	}
+	return result.RowsAffected()
+}

internal/plan/plan.go

@@ -152,8 +152,8 @@ type Service struct {
 // cal with links under publicURL, tells the household through notifier,
 // which must be over st, and registers the KindDraftWeek job on
 // runner, which must not have started yet. The kind is service-only:
-// StartDraft, which keeps one draft in flight per week, is the only way to
-// enqueue it.
+// StartDraft and StartScheduledDraft, which keep one draft in flight per
+// week, are the only way to enqueue it.
 func NewService(st *store.Store, draft Drafter, runner *jobs.Runner, cal Calendar, notifier *notify.Service, publicURL *url.URL, logger *slog.Logger) (*Service, error) {
 	switch {
 	case st == nil:
@@ -342,10 +342,27 @@ type Draft struct {
 // enqueues nothing. Drafting over a synced or changed plan needs confirm,
 // else it fails with ErrConfirmRequired and enqueues nothing.
 func (s *Service) StartDraft(ctx context.Context, weekStart store.LocalDate, confirm bool) (Draft, error) {
+	d, _, err := s.startDraft(ctx, weekStart, confirm, false)
+	return d, err
+}
+
+// StartScheduledDraft is StartDraft, without confirm, for the plan_draft
+// schedule: it enqueues the same KindDraftWeek job, or returns the one
+// queued or running, only while the week starting weekStart has no plan.
+// A week that has a plan, in any status, is left as it is: nothing is
+// enqueued and planned is true.
+func (s *Service) StartScheduledDraft(ctx context.Context, weekStart store.LocalDate) (d Draft, planned bool, err error) {
+	return s.startDraft(ctx, weekStart, false, true)
+}
+
+// startDraft is StartDraft, and when onlyUnplanned is set it enqueues
+// nothing for a week that has a plan and reports planned true.
+func (s *Service) startDraft(ctx context.Context, weekStart store.LocalDate, confirm, onlyUnplanned bool) (Draft, bool, error) {
 	if _, err := household.WeekDays(weekStart); err != nil {
-		return Draft{}, &InvalidWeekError{Value: weekStart.String()}
+		return Draft{}, false, &InvalidWeekError{Value: weekStart.String()}
 	}
 	var d Draft
+	planned := false
 	err := s.st.WithTx(ctx, func(tx *sql.Tx) error {
 		q := db.New(tx)
 		cur, err := q.GetWeekPlan(ctx, weekStart.String())
@@ -353,6 +370,9 @@ func (s *Service) StartDraft(ctx context.Context, weekStart store.LocalDate, con
 		case errors.Is(err, sql.ErrNoRows):
 		case err != nil:
 			return fmt.Errorf("plan: read week %s: %w", weekStart, err)
+		case onlyUnplanned:
+			planned = true
+			return nil
 		case cur.Status != StatusDraft && !confirm:
 			return ErrConfirmRequired
 		}
@@ -375,12 +395,12 @@ func (s *Service) StartDraft(ctx context.Context, weekStart store.LocalDate, con
 		return nil
 	})
 	if err != nil {
-		return Draft{}, err
+		return Draft{}, false, err
 	}
 	if d.Created {
 		s.draftJob.Wake()
 	}
-	return d, nil
+	return d, planned, nil
 }
 
 // Swap plans dinnerID on date of the week starting weekStart, marking the

internal/plan/schedule.go

@@ -0,0 +1,42 @@
+package plan
+
+import (
+	"context"
+	"time"
+
+	"code.cubecraftlabs.com/Maniacs_Mansion/Meal-Quest/internal/household"
+	"code.cubecraftlabs.com/Maniacs_Mansion/Meal-Quest/internal/scheduler"
+)
+
+// DraftSchedule is the plan_draft schedule: at the household's
+// plan_draft_dow and plan_draft_time it drafts the week after the fire
+// date's, its period scheduler.WeeklyForNextWeek keys plan_draft:YYYY-Www
+// for that week. The draft is the KindDraftWeek job a manual draft runs
+// (StartScheduledDraft), so a stored draft emits plan_ready keyed
+// plan_ready:{weekStart} and a draft that fails, as when the AI fails,
+// stores no plan and emits nothing. A week that already has a plan is left
+// as it is, and so is a draft of it already queued or running; either way
+// the period says so.
+func (s *Service) DraftSchedule() scheduler.Schedule {
+	return scheduler.Schedule{
+		Name: scheduler.PlanDraft,
+		Rule: func(settings household.Settings) scheduler.Rule {
+			return scheduler.WeeklyForNextWeek(time.Weekday(settings.PlanDraftDow), settings.PlanDraftTime)
+		},
+		Run: func(ctx context.Context, p scheduler.Period) error {
+			d, planned, err := s.StartScheduledDraft(ctx, p.Week)
+			if err != nil {
+				return err
+			}
+			switch {
+			case planned:
+				s.logger.InfoContext(ctx, "scheduled week draft not started: the week already has a plan",
+					"period", p.Key, "week_start", p.Week.String())
+			case !d.Created:
+				s.logger.WarnContext(ctx, "scheduled week draft not started: a draft of the week is already under way",
+					"period", p.Key, "week_start", p.Week.String(), "job_id", d.JobID)
+			}
+			return nil
+		},
+	}
+}

internal/scheduler/export_test.go

@@ -0,0 +1,9 @@
+package scheduler
+
+import "time"
+
+// SetClock makes now the clock and after the timer s fires by.
+func SetClock(s *Scheduler, now func() time.Time, after func(time.Duration) <-chan time.Time) {
+	s.now = now
+	s.after = after
+}

internal/scheduler/plan_draft_test.go

@@ -0,0 +1,393 @@
+package scheduler_test
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"log/slog"
+	"net/url"
+	"strings"
+	"sync"
+	"testing"
+	"time"
+
+	"code.cubecraftlabs.com/Maniacs_Mansion/Meal-Quest/internal/ai/harness"
+	"code.cubecraftlabs.com/Maniacs_Mansion/Meal-Quest/internal/ai/llm"
+	"code.cubecraftlabs.com/Maniacs_Mansion/Meal-Quest/internal/ai/skills"
+	"code.cubecraftlabs.com/Maniacs_Mansion/Meal-Quest/internal/db"
+	"code.cubecraftlabs.com/Maniacs_Mansion/Meal-Quest/internal/gcal"
+	"code.cubecraftlabs.com/Maniacs_Mansion/Meal-Quest/internal/household"
+	"code.cubecraftlabs.com/Maniacs_Mansion/Meal-Quest/internal/jobs"
+	"code.cubecraftlabs.com/Maniacs_Mansion/Meal-Quest/internal/notify"
+	"code.cubecraftlabs.com/Maniacs_Mansion/Meal-Quest/internal/notify/notifytest"
+	"code.cubecraftlabs.com/Maniacs_Mansion/Meal-Quest/internal/plan"
+	"code.cubecraftlabs.com/Maniacs_Mansion/Meal-Quest/internal/scheduler"
+	"code.cubecraftlabs.com/Maniacs_Mansion/Meal-Quest/internal/store"
+)
+
+// The tests in this file run plan's plan_draft schedule on the fake clock
+// against the real plan service, job runner, harness, draft_week and
+// notification outbox, with the AI faked.
+
+// planEnv is a plan service over a scheduler env's store, the "process" a
+// scheduler runs it in. model answers draft_week.
+type planEnv struct {
+	e     *env
+	svc   *plan.Service
+	model *llm.Fake
+	ids   map[string]string
+	logs  *syncBuffer
+	stop  func()
+}
+
+// poolNames are the seven Regulars pool stores, enough for a week.
+var poolNames = []string{"Beef Chili", "Pot Roast", "Chicken Fajitas", "Spaghetti Bolognese", "Sheet Pan Salmon", "Tacos", "Pancakes"}
+
+// pool stores poolNames as Regulars and returns their ids by name.
+func (e *env) pool() map[string]string {
+	e.t.Helper()
+	ids := map[string]string{}
+	for _, name := range poolNames {
+		now := store.Now().String()
+		d, err := db.New(e.st.Writer).CreateDinner(context.Background(), db.CreateDinnerParams{
+			ID:              store.NewID(),
+			Name:            name,
+			NormalizedName:  harness.Normalize(name),
+			Status:          "regular",
+			Origin:          "regular",
+			GroundingStatus: "manual",
+			CreatedAt:       now,
+			UpdatedAt:       now,
+		})
+		if err != nil {
+			e.t.Fatal(err)
+		}
+		ids[name] = d.ID
+	}
+	return ids
+}
+
+// publicURL is MEALQUEST_PUBLIC_URL in these tests.
+var publicURL = &url.URL{Scheme: "https", Host: "meals.example.com"}
+
+// noCalendar is a Calendar the schedule must never push to.
+type noCalendar struct{ t *testing.T }
+
+func (c noCalendar) PushEvents(context.Context, []gcal.Event, func() error) error {
+	c.t.Error("plan_draft pushed to Calendar")
+	return errors.New("no calendar")
+}
+
+// plans builds a plan service drafting from the dinners ids over e with a
+// started job runner, which stop stops (and the test's end, if not before).
+func (e *env) plans(ids map[string]string) *planEnv {
+	e.t.Helper()
+	p := &planEnv{e: e, model: &llm.Fake{}, ids: ids, logs: &syncBuffer{}}
+	guard, err := harness.NewGuard(harness.DefaultMaxCalls)
+	if err != nil {
+		e.t.Fatal(err)
+	}
+	cfg := harness.Config{
+		Settings:    llm.Settings{Provider: "anthropic", Model: "claude-test", Temperature: 0.7},
+		NewProvider: func(llm.Settings, string) (llm.Provider, error) { return p.model, nil },
+		Store:       e.st,
+		Guard:       guard,
+	}
+	drafter := func(ctx context.Context, scope string, in skills.WeekRequest) (skills.Week, error) {
+		return skills.DraftWeek(ctx, cfg, scope, in)
+	}
+	logger := slog.New(slog.NewTextHandler(p.logs, nil))
+	runner, err := jobs.NewRunner(e.st, logger, 1)
+	if err != nil {
+		e.t.Fatal(err)
+	}
+	p.svc, err = plan.NewService(e.st, drafter, runner, noCalendar{e.t}, notifytest.New(e.t, e.st), publicURL, logger)
+	if err != nil {
+		e.t.Fatal(err)
+	}
+	ctx, cancel := context.WithCancel(context.Background())
+	if err := runner.Start(ctx); err != nil {
+		cancel()
+		e.t.Fatal(err)
+	}
+	var once sync.Once
+	p.stop = func() {
+		once.Do(func() {
+			cancel()
+			runner.Wait()
+		})
+	}
+	e.t.Cleanup(p.stop)
+	return p
+}
+
+// reply scripts the next draft_week answer: poolNames on the days of the
+// week starting weekStart, Monday to Sunday.
+func (p *planEnv) reply(weekStart store.LocalDate) {
+	p.e.t.Helper()
+	dates, err := household.WeekDays(weekStart)
+	if err != nil {
+		p.e.t.Fatal(err)
+	}
+	var days []string
+	for i, name := range poolNames {
+		days = append(days, fmt.Sprintf(`{"date":%q,"dinner_id":%q}`, dates[i], p.ids[name]))
+	}
+	js := fmt.Sprintf(`{"days":[%s],"note":""}`, strings.Join(days, ","))
+	p.model.QueueGenerate(llm.GenerateResult{Response: llm.Response{Text: js, JSON: json.RawMessage(js)}})
+}
+
+// draftJob is a draft_week jobs row.
+type draftJob struct {
+	Status, Error string
+}
+
+// drafts waits until no draft_week job is queued or running and returns
+// them all, oldest first.
+func (e *env) drafts() []draftJob {
+	e.t.Helper()
+	deadline := time.Now().Add(10 * time.Second)
+	for {
+		rows, err := e.st.Reader.Query(`SELECT status, COALESCE(error, '') FROM jobs WHERE kind = ? ORDER BY created_at, id`, plan.KindDraftWeek)
+		if err != nil {
+			e.t.Fatal(err)
+		}
+		var out []draftJob
+		busy := false
+		for rows.Next() {
+			var j draftJob
+			if err := rows.Scan(&j.Status, &j.Error); err != nil {
+				e.t.Fatal(err)
+			}
+			busy = busy || j.Status == string(jobs.StatusQueued) || j.Status == string(jobs.StatusRunning)
+			out = append(out, j)
+		}
+		if err := rows.Err(); err != nil {
+			e.t.Fatal(err)
+		}
+		rows.Close()
+		if !busy {
+			return out
+		}
+		if time.Now().After(deadline) {
+			e.t.Fatalf("draft jobs still under way: %+v", out)
+		}
+		time.Sleep(5 * time.Millisecond)
+	}
+}
+
+// wantDrafts fails unless the draft_week jobs are got.
+func wantDrafts(t *testing.T, what string, got []draftJob, want ...draftJob) {
+	t.Helper()
+	if fmt.Sprint(got) != fmt.Sprint(want) {
+		t.Fatalf("%s: draft jobs = %+v, want %+v", what, got, want)
+	}
+}
+
+// weekPlans returns the week_start of every plan, oldest first.
+func (e *env) weekPlans() []string {
+	e.t.Helper()
+	rows, err := e.st.Reader.Query(`SELECT week_start FROM week_plans ORDER BY week_start`)
+	if err != nil {
+		e.t.Fatal(err)
+	}
+	defer rows.Close()
+	var out []string
+	for rows.Next() {
+		var w string
+		if err := rows.Scan(&w); err != nil {
+			e.t.Fatal(err)
+		}
+		out = append(out, w)
+	}
+	if err := rows.Err(); err != nil {
+		e.t.Fatal(err)
+	}
+	return out
+}
+
+// planReady returns the distinct plan_ready keys stored, oldest first.
+func (e *env) planReady() []string {
+	e.t.Helper()
+	var keys []string
+	for _, r := range notifytest.Rows(e.t, e.st, notify.PlanReady) {
+		if len(keys) == 0 || keys[len(keys)-1] != r.Key {
+			keys = append(keys, r.Key)
+		}
+	}
+	return keys
+}
+
+var (
+	succeeded = draftJob{Status: string(jobs.StatusSucceeded)}
+	aiFailed  = draftJob{Status: string(jobs.StatusFailed), Error: plan.ReasonAIFailed}
+	// w43 is the Monday of 2026-W43, the week after Friday Oct 16.
+	w43 = mustLocalDate("2026-10-19")
+	w44 = mustLocalDate("2026-10-26")
+)
+
+func mustLocalDate(s string) store.LocalDate {
+	d, err := store.ParseLocalDate(s)
+	if err != nil {
+		panic(err)
+	}
+	return d
+}
+
+// TestPlanDraftDraftsNextWeek: at the seeded Friday 09:00 the schedule
+// drafts the next ISO week through the plan service's draft job, period
+// plan_draft:YYYY-Www of that week, and the draft emits plan_ready keyed
+// by its Monday once.
+func TestPlanDraftDraftsNextWeek(t *testing.T) {
+	e := newEnv(t)
+	p := e.plans(e.pool())
+	p.reply(w43)
+	p.reply(w44)
+
+	// Wednesday Oct 14: the Friday Oct 9 09:00 fire is days ago, missed.
+	r := e.start(at(2026, 10, 14, 12, 0), p.svc.DraftSchedule())
+	r.runUntil(at(2026, 10, 16, 8, 59))
+	wantDrafts(t, "before Friday 09:00", e.drafts())
+	wantKeys(t, "plans before Friday 09:00", e.weekPlans(), nil)
+
+	r.runUntil(at(2026, 10, 16, 9, 0))
+	wantDrafts(t, "at Friday 09:00", e.drafts(), succeeded)
+	wantKeys(t, "plans at Friday 09:00", e.weekPlans(), []string{"2026-10-19"})
+	wantKeys(t, "plan_ready of W43", e.planReady(), []string{"plan_ready:2026-10-19"})
+	got, err := p.svc.Get(context.Background(), w43)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(got.Days) != 7 || got.Days[0].Dinner.Name != "Beef Chili" || got.Status != plan.StatusDraft {
+		t.Errorf("plan of W43 = %+v, want the drafted week", got)
+	}
+
+	r.runUntil(at(2026, 10, 23, 9, 0))
+	wantDrafts(t, "W44", e.drafts(), succeeded, succeeded)
+	wantKeys(t, "plans", e.weekPlans(), []string{"2026-10-19", "2026-10-26"})
+	wantKeys(t, "plan_ready of W43 and W44", e.planReady(), []string{"plan_ready:2026-10-19", "plan_ready:2026-10-26"})
+	wantKeys(t, "schedule_runs", e.claimed(scheduler.PlanDraft), []string{"2026-W42", "2026-W43", "2026-W44"})
+}
+
+// TestPlanDraftLeavesAnExistingPlan: when next week already has a plan,
+// here one the household drafted by hand, the period drafts nothing,
+// notifies no one and says so.
+func TestPlanDraftLeavesAnExistingPlan(t *testing.T) {
+	e := newEnv(t)
+	p := e.plans(e.pool())
+	p.reply(w43)
+	p.reply(w43) // a second draft would succeed too
+	if _, err := p.svc.StartDraft(context.Background(), w43, false); err != nil {
+		t.Fatal(err)
+	}
+	wantDrafts(t, "manual draft", e.drafts(), succeeded)
+	wantKeys(t, "plan_ready of the manual draft", e.planReady(), []string{"plan_ready:2026-10-19"})
+	before, err := p.svc.Get(context.Background(), w43)
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	r := e.start(at(2026, 10, 16, 8, 0), p.svc.DraftSchedule())
+	r.runUntil(at(2026, 10, 16, 9, 0))
+	wantDrafts(t, "with a plan", e.drafts(), succeeded)
+	wantKeys(t, "plan_ready with a plan", e.planReady(), []string{"plan_ready:2026-10-19"})
+	wantKeys(t, "schedule_runs", e.claimed(scheduler.PlanDraft), []string{"2026-W42", "2026-W43"})
+	after, err := p.svc.Get(context.Background(), w43)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if fmt.Sprint(after) != fmt.Sprint(before) {
+		t.Errorf("plan of W43 = %+v, want it unchanged from %+v", after, before)
+	}
+	if logs := p.logs.String(); !strings.Contains(logs, "the week already has a plan") || !strings.Contains(logs, "period=2026-W43") {
+		t.Errorf("logs = %s, want the skipped period logged", logs)
+	}
+}
+
+// TestPlanDraftNoDuplicateAfterRestart: a restart after the week was
+// drafted, still within the catch-up window, drafts nothing more and sends
+// no second plan_ready: the period is already claimed.
+func TestPlanDraftNoDuplicateAfterRestart(t *testing.T) {
+	e := newEnv(t)
+	ids := e.pool()
+	p := e.plans(ids)
+	p.reply(w43)
+	r := e.start(at(2026, 10, 16, 8, 0), p.svc.DraftSchedule())
+	r.runUntil(at(2026, 10, 16, 9, 0))
+	wantDrafts(t, "first process", e.drafts(), succeeded)
+	r.stop()
+	p.stop()
+
+	// The restarted process has a new job runner, plan service and
+	// scheduler; it would answer a second draft if one were started.
+	p2 := e.plans(ids)
+	p2.reply(w43)
+	r2 := e.start(at(2026, 10, 16, 10, 0), p2.svc.DraftSchedule())
+	r2.runUntil(at(2026, 10, 23, 8, 59))
+	wantDrafts(t, "after restart", e.drafts(), succeeded)
+	wantKeys(t, "plans after restart", e.weekPlans(), []string{"2026-10-19"})
+	wantKeys(t, "plan_ready after restart", e.planReady(), []string{"plan_ready:2026-10-19"})
+	wantKeys(t, "schedule_runs", e.claimed(scheduler.PlanDraft), []string{"2026-W42", "2026-W43"})
+}
+
+// TestPlanDraftAIFailureEmitsNothing: when draft_week fails the draft job
+// fails as ai_failed with the failure recorded in ai_runs, no plan is made
+// up and no one is notified; a restart within the catch-up window does not
+// draft the claimed period again.
+func TestPlanDraftAIFailureEmitsNothing(t *testing.T) {
+	e := newEnv(t)
+	ids := e.pool()
+	p := e.plans(ids)
+	p.model.QueueGenerate(llm.GenerateResult{Err: errors.New("provider down")})
+
+	r := e.start(at(2026, 10, 16, 8, 0), p.svc.DraftSchedule())
+	r.runUntil(at(2026, 10, 16, 9, 0))
+	wantDrafts(t, "after the AI failed", e.drafts(), aiFailed)
+	wantKeys(t, "plans after the AI failed", e.weekPlans(), nil)
+	wantKeys(t, "plan_ready after the AI failed", e.planReady(), nil)
+	wantKeys(t, "schedule_runs", e.claimed(scheduler.PlanDraft), []string{"2026-W42", "2026-W43"})
+
+	var status, msg string
+	if err := e.st.Reader.QueryRow(`SELECT status, COALESCE(error, '') FROM ai_runs WHERE skill = 'draft_week'`).Scan(&status, &msg); err != nil {
+		t.Fatalf("ai_runs row of draft_week: %v", err)
+	}
+	if status != string(harness.StatusProviderError) || msg == "" {
+		t.Errorf("ai_runs draft_week = %s %q, want %s with its message", status, msg, harness.StatusProviderError)
+	}
+	if logs := p.logs.String(); !strings.Contains(logs, "week draft failed") || !strings.Contains(logs, "reason="+plan.ReasonAIFailed) {
+		t.Errorf("logs = %s, want the draft failure with reason %s", logs, plan.ReasonAIFailed)
+	}
+
+	r.stop()
+	p.stop()
+	p2 := e.plans(ids)
+	p2.reply(w43)
+	e.start(at(2026, 10, 16, 10, 0), p2.svc.DraftSchedule())
+	wantDrafts(t, "after restart", e.drafts(), aiFailed)
+	wantKeys(t, "plan_ready after restart", e.planReady(), nil)
+}
+
+// TestPlanDraftFollowsSettingsWithoutRestart: a plan_draft_dow and
+// plan_draft_time change takes effect in the running scheduler; the old
+// Friday 09:00 no longer fires and the new Wednesday 18:30 drafts the next
+// week.
+func TestPlanDraftFollowsSettingsWithoutRestart(t *testing.T) {
+	e := newEnv(t)
+	p := e.plans(e.pool())
+	p.reply(w43)
+
+	r := e.start(at(2026, 10, 12, 12, 0), p.svc.DraftSchedule()) // Monday
+	e.update(household.SettingsPatch{PlanDraftDow: ptr(3), PlanDraftTime: ptr("18:30")})
+	r.changed()
+	r.runUntil(at(2026, 10, 14, 18, 29)) // Wednesday
+	wantDrafts(t, "before Wednesday 18:30", e.drafts())
+	r.runUntil(at(2026, 10, 14, 18, 30))
+	wantDrafts(t, "at Wednesday 18:30", e.drafts(), succeeded)
+	wantKeys(t, "plan_ready", e.planReady(), []string{"plan_ready:2026-10-19"})
+
+	// Friday 09:00, the old fire time, drafts nothing more.
+	r.runUntil(at(2026, 10, 16, 9, 0))
+	wantDrafts(t, "at Friday 09:00", e.drafts(), succeeded)
+	wantKeys(t, "schedule_runs", e.claimed(scheduler.PlanDraft), []string{"2026-W42", "2026-W43"})
+}

internal/scheduler/scheduler.go

@@ -0,0 +1,587 @@
+// Package scheduler runs the household's timed work: each registered
+// schedule fires once per period at a household-local time read from the
+// household settings.
+//
+// # Periods and the claim
+//
+// Every fire time belongs to one period, named by a period key derived in
+// the household zone from the schedule's cadence (see Rule.PeriodKey): the
+// local date "2026-10-09" for a daily rule, and the ISO week "2026-W41" of
+// the fire date, or of the week after it, for a weekly rule. Before running,
+// the scheduler inserts schedule_runs(schedule, period_key); the insert is
+// the claim, and a period already claimed, by an earlier run, an earlier
+// process or another instance, is skipped. A period therefore runs at most
+// once: a Run that fails is logged and its period is not retried.
+//
+// # Catch-up
+//
+// When the scheduler starts, and whenever it wakes late, it looks only at
+// each schedule's most recent fire time. If that time passed within
+// CatchUpWindow and the period is unclaimed, the period runs once. If it
+// passed longer ago, the period is claimed without running and logged as
+// missed, so it can never run later, not after a restart and not after a
+// settings change moves the fire time. Fire times before the most recent
+// one are never looked at.
+//
+// # Claim failures
+//
+// A claim that fails (the database refused the insert) is never skipped:
+// the scheduler logs it and keeps that period's fire time, and every
+// ClaimRetry claims that same period again until the claim succeeds. Only
+// then does it apply the catch-up rule to it, running it if its fire time
+// is within CatchUpWindow of the successful claim and recording it as
+// missed otherwise, and only then does it go on to later fire times, even
+// when one has passed meanwhile. Each period is claimed by its own fire
+// time, so the later period runs once as well.
+//
+// # Time zone and DST
+//
+// Fire times are wall-clock times in the household zone, never the process
+// zone, turned into instants by household.LocalTime. Its rule is the
+// scheduler's DST rule: a time that does not exist on a spring-forward day
+// fires later by the length of the gap (02:30 fires at 03:30 local), and a
+// time that occurs twice on a fall-back day fires once, at its first
+// occurrence. Either way the period key is that local date's, so the
+// period still runs exactly once.
+//
+// # Settings changes
+//
+// The scheduler subscribes to the household settings and records each change
+// with the instant it was saved. A change takes effect at once without a
+// restart: fire times up to that instant follow the old settings and later
+// ones the new, even when the scheduler gets to the change late. So moving
+// a fire time to earlier the same day does not fire the passed time, and a
+// new fire time that was still ahead when the change was saved fires.
+package scheduler
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"log/slog"
+	"sync"
+	"time"
+
+	"code.cubecraftlabs.com/Maniacs_Mansion/Meal-Quest/internal/db"
+	"code.cubecraftlabs.com/Maniacs_Mansion/Meal-Quest/internal/household"
+	"code.cubecraftlabs.com/Maniacs_Mansion/Meal-Quest/internal/store"
+)
+
+// Schedule names, as the schedule_runs.schedule CHECK constraint allows.
+const (
+	SuggestionsWeekly    = "suggestions_weekly"
+	PlanDraft            = "plan_draft"
+	ShoppingChatAutosend = "shopping_chat_autosend"
+	DayOfReminder        = "day_of_reminder"
+	RotationCheck        = "rotation_check"
+	WeekRollover         = "week_rollover"
+	AIRunsPurge          = "ai_runs_purge"
+)
+
+// names are the schedule names Register accepts.
+var names = map[string]bool{
+	SuggestionsWeekly:    true,
+	PlanDraft:            true,
+	ShoppingChatAutosend: true,
+	DayOfReminder:        true,
+	RotationCheck:        true,
+	WeekRollover:         true,
+	AIRunsPurge:          true,
+}
+
+// CatchUpWindow is how long after its fire time a period that has not run
+// still runs. A period found later than that is recorded as missed.
+const CatchUpWindow = 24 * time.Hour
+
+// ClaimRetry is how long the scheduler waits before repeating a pass in
+// which a claim failed.
+const ClaimRetry = time.Minute
+
+// maxWait bounds one sleep, so the scheduler checks the wall clock at least
+// this often even if the host was suspended or its clock stepped.
+const maxWait = time.Hour
+
+// Rule is when a schedule fires: at household-local Time ("HH:MM") every
+// day, or on Weekday only when Weekly is set. Build one with Daily, Weekly
+// or WeeklyForNextWeek.
+type Rule struct {
+	Weekly  bool
+	Weekday time.Weekday
+	Time    string
+	// NextWeek keys a weekly period by the ISO week after the fire date,
+	// for work done ahead for the coming week.
+	NextWeek bool
+}
+
+// Daily fires every day at hhmm; its period is the local date.
+func Daily(hhmm string) Rule {
+	return Rule{Time: hhmm}
+}
+
+// Weekly fires on day at hhmm; its period is the ISO week of the fire date.
+func Weekly(day time.Weekday, hhmm string) Rule {
+	return Rule{Weekly: true, Weekday: day, Time: hhmm}
+}
+
+// WeeklyForNextWeek fires on day at hhmm; its period is the ISO week after
+// the fire date's.
+func WeeklyForNextWeek(day time.Weekday, hhmm string) Rule {
+	return Rule{Weekly: true, Weekday: day, Time: hhmm, NextWeek: true}
+}
+
+// PeriodKey returns the period key of a fire on household-local date
+// fireDate: the date ("2026-10-09") for a daily rule, the ISO week
+// ("2026-W53") for a weekly one, and the ISO week after that ("2027-W01")
+// when NextWeek is set. It panics if fireDate is the zero LocalDate.
+func (r Rule) PeriodKey(fireDate store.LocalDate) string {
+	if !r.Weekly {
+		return fireDate.String()
+	}
+	if r.NextWeek {
+		return household.PeriodKey(addDays(fireDate, 7))
+	}
+	return household.PeriodKey(fireDate)
+}
+
+// validate reports whether r can fire.
+func (r Rule) validate() error {
+	if r.Weekly && (r.Weekday < time.Sunday || r.Weekday > time.Saturday) {
+		return fmt.Errorf("invalid weekday %d (want 0-6, Sunday = 0)", r.Weekday)
+	}
+	if !r.Weekly && (r.Weekday != time.Sunday || r.NextWeek) {
+		return errors.New("a daily rule has no weekday or next week")
+	}
+	// Any date checks the time's form.
+	_, err := household.LocalTime(time.UTC, store.LocalDateOf(time.Unix(0, 0).UTC()), r.Time)
+	return err
+}
+
+// Period is one period of a schedule, as given to its Run.
+type Period struct {
+	// Key is the period key claimed in schedule_runs.
+	Key string
+	// FireAt is the instant the period fired, in the household zone.
+	FireAt time.Time
+	// Date is the household-local date of FireAt.
+	Date store.LocalDate
+	// Week is the ISO Monday of the week Key names; for a daily rule, of
+	// the week containing Date.
+	Week store.LocalDate
+}
+
+// Schedule is registered work that runs once per period.
+type Schedule struct {
+	// Name is one of the schedule name constants.
+	Name string
+	// Rule returns when the schedule fires under settings. It is called
+	// again after every settings change.
+	Rule func(settings household.Settings) Rule
+	// Run does the period's work. It runs on the scheduler's goroutine,
+	// one schedule at a time, so it must return promptly: long work is
+	// enqueued as a job. ctx is cancelled when the scheduler stops. An
+	// error is logged; the period stays claimed and is not retried.
+	Run func(ctx context.Context, p Period) error
+}
+
+// Settings is the household settings source: *household.SettingsService.
+type Settings interface {
+	Get(ctx context.Context) (household.Settings, error)
+	Subscribe(fn func(household.Settings)) (unsubscribe func())
+}
+
+// Scheduler runs the registered schedules.
+type Scheduler struct {
+	st       *store.Store
+	settings Settings
+	logger   *slog.Logger
+
+	// now is the clock and after the timer; tests replace both.
+	now   func() time.Time
+	after func(d time.Duration) <-chan time.Time
+
+	// wake holds one token when the settings changed since the loop last
+	// looked; pending is the changes to apply, oldest first.
+	wake      chan struct{}
+	pendingMu sync.Mutex
+	pending   []change
+
+	mu        sync.Mutex
+	schedules []Schedule
+	started   bool
+
+	wg sync.WaitGroup
+}
+
+// New returns a scheduler over st reading the household settings from
+// settings.
+func New(st *store.Store, settings Settings, logger *slog.Logger) (*Scheduler, error) {
+	if st == nil {
+		return nil, errors.New("scheduler: nil store")
+	}
+	if settings == nil {
+		return nil, errors.New("scheduler: nil settings")
+	}
+	if logger == nil {
+		return nil, errors.New("scheduler: nil logger")
+	}
+	return &Scheduler{
+		st:       st,
+		settings: settings,
+		logger:   logger,
+		now:      time.Now,
+		after:    time.After,
+		wake:     make(chan struct{}, 1),
+	}, nil
+}
+
+// Register adds sch. It must be called before Start, once per name.
+func (s *Scheduler) Register(sch Schedule) error {
+	if !names[sch.Name] {
+		return fmt.Errorf("scheduler: register: unknown schedule %q", sch.Name)
+	}
+	if sch.Rule == nil {
+		return fmt.Errorf("scheduler: register %s: nil rule", sch.Name)
+	}
+	if sch.Run == nil {
+		return fmt.Errorf("scheduler: register %s: nil run", sch.Name)
+	}
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	if s.started {
+		return fmt.Errorf("scheduler: register %s: already started", sch.Name)
+	}
+	for _, have := range s.schedules {
+		if have.Name == sch.Name {
+			return fmt.Errorf("scheduler: register %s: already registered", sch.Name)
+		}
+	}
+	s.schedules = append(s.schedules, sch)
+	return nil
+}
+
+// Start reads the settings and starts the scheduler, which first catches
+// up on each schedule's most recent period and then fires schedules as
+// their times come. Settings whose zone or rules are invalid are an error
+// and nothing starts. The scheduler stops when ctx is cancelled; Wait waits
+// for it.
+func (s *Scheduler) Start(ctx context.Context) error {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	if s.started {
+		return errors.New("scheduler: already started")
+	}
+	// Subscribing first means a change saved after the read below still
+	// wakes the loop.
+	unsubscribe := s.settings.Subscribe(s.changed)
+	settings, err := s.settings.Get(ctx)
+	if err != nil {
+		unsubscribe()
+		return fmt.Errorf("scheduler: %w", err)
+	}
+	p, err := compile(s.schedules, settings)
+	if err != nil {
+		unsubscribe()
+		return err
+	}
+	s.started = true
+	s.wg.Add(1)
+	go func() {
+		defer s.wg.Done()
+		defer unsubscribe()
+		s.loop(ctx, p)
+	}()
+	return nil
+}
+
+// Wait blocks until the scheduler has stopped.
+func (s *Scheduler) Wait() {
+	s.wg.Wait()
+}
+
+// change is a settings change and the instant it was saved.
+type change struct {
+	at       time.Time
+	settings household.Settings
+}
+
+// changed is the settings subscriber. It runs on the updating goroutine, so
+// it only records the change and wakes the loop.
+func (s *Scheduler) changed(settings household.Settings) {
+	s.pendingMu.Lock()
+	s.pending = append(s.pending, change{at: s.now(), settings: settings})
+	s.pendingMu.Unlock()
+	select {
+	case s.wake <- struct{}{}:
+	default: // the loop already has a token and reads pending then
+	}
+}
+
+// takeChanges returns and clears the pending changes.
+func (s *Scheduler) takeChanges() []change {
+	s.pendingMu.Lock()
+	defer s.pendingMu.Unlock()
+	out := s.pending
+	s.pending = nil
+	return out
+}
+
+// plan is the schedules compiled against one version of the settings.
+type plan struct {
+	loc   *time.Location
+	items []item
+}
+
+type item struct {
+	sch  Schedule
+	rule Rule
+}
+
+// compile builds the plan for settings.
+func compile(schedules []Schedule, settings household.Settings) (*plan, error) {
+	loc, err := household.LoadZone(settings.Timezone)
+	if err != nil {
+		return nil, fmt.Errorf("scheduler: household zone: %w", err)
+	}
+	p := &plan{loc: loc, items: make([]item, len(schedules))}
+	for i, sch := range schedules {
+		rule := sch.Rule(settings)
+		if err := rule.validate(); err != nil {
+			return nil, fmt.Errorf("scheduler: %s: %w", sch.Name, err)
+		}
+		p.items[i] = item{sch: sch, rule: rule}
+	}
+	return p, nil
+}
+
+// loop fires schedules until ctx is cancelled. done is the instant up to
+// which fire times have been handled; it starts at the zero time so the
+// first pass catches up. retry is the periods whose claim failed, each
+// with its own fire time. Each pass first claims retry again; while any
+// claim there still fails, the pass stops and is repeated after
+// ClaimRetry. Otherwise it applies the pending settings changes in order,
+// handling fire times up to a change under the plan it replaces, then
+// handles fire times up to now under the current plan, stopping after the
+// first stretch in which a claim failed.
+func (s *Scheduler) loop(ctx context.Context, p *plan) {
+	var done time.Time
+	var pending []change
+	var retry []target
+	for {
+		now := s.now()
+		pending = append(pending, s.takeChanges()...)
+		retry = s.retry(ctx, retry, now)
+		for len(retry) == 0 && len(pending) > 0 {
+			c := pending[0]
+			if c.at.After(done) {
+				if p != nil {
+					retry = s.fire(ctx, p, done, c.at, now)
+				}
+				done = c.at
+			}
+			p = s.compileChange(ctx, c.settings)
+			pending = pending[1:]
+		}
+		if len(retry) == 0 {
+			if p != nil {
+				retry = s.fire(ctx, p, done, now, now)
+			}
+			done = now
+		}
+		if ctx.Err() != nil {
+			return
+		}
+
+		var timeout <-chan time.Time
+		switch {
+		case len(retry) > 0:
+			timeout = s.after(ClaimRetry)
+		case p != nil:
+			if next, ok := p.next(now); ok {
+				timeout = s.after(min(next.Sub(now), maxWait))
+			}
+		}
+		select {
+		case <-ctx.Done():
+			return
+		case <-timeout:
+		case <-s.wake:
+		}
+	}
+}
+
+// compileChange returns the plan for changed settings, or nil when they
+// are unusable.
+func (s *Scheduler) compileChange(ctx context.Context, settings household.Settings) *plan {
+	p, err := compile(s.schedules, settings)
+	if err != nil {
+		// Settings are validated before they are saved, so this is a
+		// bug; firing at a guessed time would be worse than not firing
+		// until the settings are fixed.
+		s.logger.ErrorContext(ctx, "scheduler: settings unusable; no schedule fires until they change", "error", err)
+		return nil
+	}
+	return p
+}
+
+// target is one schedule's period, named by the fire time it belongs to.
+type target struct {
+	it  item
+	loc *time.Location
+	at  time.Time
+}
+
+// fire handles, for every schedule, its most recent fire time at or before
+// until if it is after done; now is the current instant. It returns the
+// targets whose claim failed, to be claimed again.
+func (s *Scheduler) fire(ctx context.Context, p *plan, done, until, now time.Time) []target {
+	var failed []target
+	for _, it := range p.items {
+		if ctx.Err() != nil {
+			return failed
+		}
+		at, due := fireBefore(it.rule, p.loc, until)
+		if !due || !at.After(done) {
+			continue
+		}
+		if t := (target{it: it, loc: p.loc, at: at}); !s.handle(ctx, t, now) {
+			failed = append(failed, t)
+		}
+	}
+	return failed
+}
+
+// retry handles targets again at now and returns those whose claim still
+// failed.
+func (s *Scheduler) retry(ctx context.Context, targets []target, now time.Time) []target {
+	var failed []target
+	for i, t := range targets {
+		if ctx.Err() != nil {
+			return append(failed, targets[i:]...)
+		}
+		if !s.handle(ctx, t, now) {
+			failed = append(failed, t)
+		}
+	}
+	return failed
+}
+
+// handle claims the period of t and runs it, or records it as missed when
+// its fire time is more than CatchUpWindow before now. It reports false
+// when the claim failed and the period is still unclaimed.
+func (s *Scheduler) handle(ctx context.Context, t target, now time.Time) bool {
+	it, at := t.it, t.at
+	date := household.Today(at, t.loc)
+	period := Period{Key: it.rule.PeriodKey(date), FireAt: at, Date: date}
+	if it.rule.NextWeek {
+		period.Week = household.WeekStart(addDays(date, 7))
+	} else {
+		period.Week = household.WeekStart(date)
+	}
+	log := s.logger.With("schedule", it.sch.Name, "period", period.Key, "fire_at", at.Format(time.RFC3339))
+
+	claimed, err := s.claim(ctx, it.sch.Name, period.Key, now)
+	if err != nil {
+		log.ErrorContext(ctx, "scheduler: claim period failed; retrying", "error", err, "retry_in", ClaimRetry.String())
+		return false
+	}
+	if !claimed {
+		return true
+	}
+	if late := now.Sub(at); late > CatchUpWindow {
+		log.WarnContext(ctx, "scheduler: period missed; recorded without running", "late", late.String())
+		return true
+	}
+	if err := it.sch.Run(ctx, period); err != nil {
+		log.ErrorContext(ctx, "scheduler: run failed", "error", err)
+		return true
+	}
+	log.InfoContext(ctx, "scheduler: ran")
+	return true
+}
+
+// claim inserts schedule_runs(name, key) and reports whether this call
+// inserted it.
+func (s *Scheduler) claim(ctx context.Context, name, key string, now time.Time) (bool, error) {
+	n, err := db.New(s.st.Writer).ClaimScheduleRun(ctx, db.ClaimScheduleRunParams{
+		ID:        store.NewID(),
+		Schedule:  name,
+		PeriodKey: key,
+		CreatedAt: store.Timestamp{Time: now.UTC().Truncate(time.Millisecond)}.String(),
+	})
+	if err != nil {
+		return false, err
+	}
+	return n == 1, nil
+}
+
+// next returns the earliest fire time of any schedule after now, or false
+// when there is no schedule.
+func (p *plan) next(now time.Time) (time.Time, bool) {
+	var next time.Time
+	for _, it := range p.items {
+		at, ok := fireAfter(it.rule, p.loc, now)
+		if ok && (next.IsZero() || at.Before(next)) {
+			next = at
+		}
+	}
+	return next, !next.IsZero()
+}
+
+// window is how many days either side of a date fireBefore and fireAfter
+// search: more than a week, so a weekly rule always has a fire time on
+// each side.
+const window = 8
+
+// fireBefore returns the latest fire time of r in zone loc at or before t.
+func fireBefore(r Rule, loc *time.Location, t time.Time) (time.Time, bool) {
+	var best time.Time
+	for _, at := range fireTimes(r, loc, t) {
+		if !at.After(t) && (best.IsZero() || at.After(best)) {
+			best = at
+		}
+	}
+	return best, !best.IsZero()
+}
+
+// fireAfter returns the earliest fire time of r in zone loc after t.
+func fireAfter(r Rule, loc *time.Location, t time.Time) (time.Time, bool) {
+	var best time.Time
+	for _, at := range fireTimes(r, loc, t) {
+		if at.After(t) && (best.IsZero() || at.Before(best)) {
+			best = at
+		}
+	}
+	return best, !best.IsZero()
+}
+
+// fireTimes returns r's fire times on the local dates within window days
+// of t's. A rule that fails to resolve was validated at compile, so it
+// cannot.
+func fireTimes(r Rule, loc *time.Location, t time.Time) []time.Time {
+	today := household.Today(t, loc)
+	var out []time.Time
+	for i := -window; i <= window; i++ {
+		d := addDays(today, i)
+		if r.Weekly && weekday(d) != r.Weekday {
+			continue
+		}
+		at, err := household.LocalTime(loc, d, r.Time)
+		if err != nil {
+			panic(fmt.Sprintf("scheduler: validated rule %+v: %v", r, err))
+		}
+		out = append(out, at)
+	}
+	return out
+}
+
+// addDays returns d plus n calendar days.
+func addDays(d store.LocalDate, n int) store.LocalDate {
+	y, m, day := d.Date()
+	return store.LocalDateOf(time.Date(y, m, day+n, 0, 0, 0, 0, time.UTC))
+}
+
+// weekday returns the day of the week of d.
+func weekday(d store.LocalDate) time.Weekday {
+	y, m, day := d.Date()
+	return time.Date(y, m, day, 0, 0, 0, 0, time.UTC).Weekday()
+}

internal/scheduler/scheduler_test.go

@@ -0,0 +1,757 @@
+package scheduler_test
+
+import (
+	"context"
+	"errors"
+	"io"
+	"log/slog"
+	"slices"
+	"sync"
+	"testing"
+	"time"
+
+	"code.cubecraftlabs.com/Maniacs_Mansion/Meal-Quest/internal/household"
+	"code.cubecraftlabs.com/Maniacs_Mansion/Meal-Quest/internal/scheduler"
+	"code.cubecraftlabs.com/Maniacs_Mansion/Meal-Quest/internal/store"
+	"code.cubecraftlabs.com/Maniacs_Mansion/Meal-Quest/migrations"
+)
+
+const zone = "America/Chicago"
+
+// chicago is the household zone of every test.
+var chicago = mustZone(zone)
+
+func mustZone(name string) *time.Location {
+	loc, err := household.LoadZone(name)
+	if err != nil {
+		panic(err)
+	}
+	return loc
+}
+
+// at returns the instant clocks in Chicago read the given time.
+func at(year int, month time.Month, day, hour, minute int) time.Time {
+	return time.Date(year, month, day, hour, minute, 0, 0, chicago)
+}
+
+// fakeClock is a clock whose time moves only when the test moves it.
+type fakeClock struct {
+	mu     sync.Mutex
+	now    time.Time
+	timers []fakeTimer
+	// waits receives one value per After call, so the test knows the
+	// scheduler finished a pass and is waiting.
+	waits chan struct{}
+}
+
+type fakeTimer struct {
+	deadline time.Time
+	ch       chan time.Time
+}
+
+func newFakeClock(now time.Time) *fakeClock {
+	return &fakeClock{now: now, waits: make(chan struct{}, 100)}
+}
+
+func (c *fakeClock) Now() time.Time {
+	c.mu.Lock()
+	defer c.mu.Unlock()
+	return c.now
+}
+
+func (c *fakeClock) After(d time.Duration) <-chan time.Time {
+	c.mu.Lock()
+	ch := make(chan time.Time, 1)
+	c.timers = append(c.timers, fakeTimer{deadline: c.now.Add(d), ch: ch})
+	c.mu.Unlock()
+	c.waits <- struct{}{}
+	return ch
+}
+
+// set moves the clock to t and fires the timers due by then, reporting
+// whether any fired.
+func (c *fakeClock) set(t time.Time) bool {
+	c.mu.Lock()
+	defer c.mu.Unlock()
+	c.now = t
+	fired := false
+	kept := c.timers[:0]
+	for _, tm := range c.timers {
+		if tm.deadline.After(t) {
+			kept = append(kept, tm)
+			continue
+		}
+		tm.ch <- t
+		fired = true
+	}
+	c.timers = kept
+	return fired
+}
+
+// nextDeadline returns the earliest pending timer deadline.
+func (c *fakeClock) nextDeadline() (time.Time, bool) {
+	c.mu.Lock()
+	defer c.mu.Unlock()
+	var next time.Time
+	for _, tm := range c.timers {
+		if next.IsZero() || tm.deadline.Before(next) {
+			next = tm.deadline
+		}
+	}
+	return next, !next.IsZero()
+}
+
+// recorder records the periods a schedule ran.
+type recorder struct {
+	mu   sync.Mutex
+	runs []scheduler.Period
+	err  error
+}
+
+func (r *recorder) run(_ context.Context, p scheduler.Period) error {
+	r.mu.Lock()
+	defer r.mu.Unlock()
+	r.runs = append(r.runs, p)
+	return r.err
+}
+
+func (r *recorder) periods() []scheduler.Period {
+	r.mu.Lock()
+	defer r.mu.Unlock()
+	return slices.Clone(r.runs)
+}
+
+func (r *recorder) keys() []string {
+	var keys []string
+	for _, p := range r.periods() {
+		keys = append(keys, p.Key)
+	}
+	return keys
+}
+
+// env is a migrated database with household settings in Chicago.
+type env struct {
+	t        *testing.T
+	st       *store.Store
+	settings *household.SettingsService
+}
+
+func newEnv(t *testing.T) *env {
+	t.Helper()
+	ctx := context.Background()
+	st, err := store.Open(ctx, t.TempDir())
+	if err != nil {
+		t.Fatal(err)
+	}
+	t.Cleanup(func() {
+		if err := st.Close(); err != nil {
+			t.Error(err)
+		}
+	})
+	if err := store.Migrate(ctx, st.Writer, migrations.FS); err != nil {
+		t.Fatal(err)
+	}
+	if err := household.SeedSettings(ctx, st, zone); err != nil {
+		t.Fatal(err)
+	}
+	return &env{t: t, st: st, settings: household.NewSettingsService(st)}
+}
+
+// update changes the household settings.
+func (e *env) update(patch household.SettingsPatch) {
+	e.t.Helper()
+	if _, err := e.settings.Update(context.Background(), patch, "member"); err != nil {
+		e.t.Fatal(err)
+	}
+}
+
+// claimed returns the period keys schedule_runs holds for name.
+func (e *env) claimed(name string) []string {
+	e.t.Helper()
+	rows, err := e.st.Reader.Query(`SELECT period_key FROM schedule_runs WHERE schedule = ? ORDER BY period_key`, name)
+	if err != nil {
+		e.t.Fatal(err)
+	}
+	defer rows.Close()
+	var keys []string
+	for rows.Next() {
+		var k string
+		if err := rows.Scan(&k); err != nil {
+			e.t.Fatal(err)
+		}
+		keys = append(keys, k)
+	}
+	if err := rows.Err(); err != nil {
+		e.t.Fatal(err)
+	}
+	return keys
+}
+
+// running is a started scheduler on a fake clock.
+type running struct {
+	t     *testing.T
+	clock *fakeClock
+	stop  func()
+}
+
+// start starts a scheduler on e at now with the schedules, and stops it
+// when the test ends. It returns once the first pass, the catch-up, is done.
+func (e *env) start(now time.Time, schedules ...scheduler.Schedule) *running {
+	e.t.Helper()
+	s, err := scheduler.New(e.st, e.settings, slog.New(slog.NewTextHandler(io.Discard, nil)))
+	if err != nil {
+		e.t.Fatal(err)
+	}
+	clock := newFakeClock(now)
+	scheduler.SetClock(s, clock.Now, clock.After)
+	for _, sch := range schedules {
+		if err := s.Register(sch); err != nil {
+			e.t.Fatal(err)
+		}
+	}
+	ctx, cancel := context.WithCancel(context.Background())
+	if err := s.Start(ctx); err != nil {
+		cancel()
+		e.t.Fatal(err)
+	}
+	var once sync.Once
+	stop := func() {
+		once.Do(func() {
+			cancel()
+			s.Wait()
+		})
+	}
+	e.t.Cleanup(stop)
+	r := &running{t: e.t, clock: clock, stop: stop}
+	r.idle()
+	return r
+}
+
+// idle waits until the scheduler finished a pass and is waiting.
+func (r *running) idle() {
+	r.t.Helper()
+	select {
+	case <-r.clock.waits:
+	case <-time.After(10 * time.Second):
+		r.t.Fatal("scheduler did not finish its pass")
+	}
+}
+
+// runUntil moves the clock to to, stopping at every instant the scheduler
+// asked to wake at on the way, as a real clock would.
+func (r *running) runUntil(to time.Time) {
+	r.t.Helper()
+	for {
+		next, ok := r.clock.nextDeadline()
+		if !ok || next.After(to) {
+			if r.clock.set(to) {
+				r.idle()
+			}
+			return
+		}
+		if r.clock.set(next) {
+			r.idle()
+		}
+	}
+}
+
+// changed waits for the scheduler to apply a settings change.
+func (r *running) changed() {
+	r.t.Helper()
+	r.idle()
+}
+
+func reminder(rec *recorder) scheduler.Schedule {
+	return scheduler.Schedule{
+		Name: scheduler.DayOfReminder,
+		Rule: func(s household.Settings) scheduler.Rule { return scheduler.Daily(s.ReminderTime) },
+		Run:  rec.run,
+	}
+}
+
+func suggestions(rec *recorder) scheduler.Schedule {
+	return scheduler.Schedule{
+		Name: scheduler.SuggestionsWeekly,
+		Rule: func(s household.Settings) scheduler.Rule {
+			return scheduler.Weekly(time.Weekday(s.SuggestionsDow), s.SuggestionsTime)
+		},
+		Run: rec.run,
+	}
+}
+
+func planDraft(rec *recorder) scheduler.Schedule {
+	return scheduler.Schedule{
+		Name: scheduler.PlanDraft,
+		Rule: func(s household.Settings) scheduler.Rule {
+			return scheduler.WeeklyForNextWeek(time.Weekday(s.PlanDraftDow), s.PlanDraftTime)
+		},
+		Run: rec.run,
+	}
+}
+
+func ptr[T any](v T) *T { return &v }
+
+func wantKeys(t *testing.T, what string, got, want []string) {
+	t.Helper()
+	if !slices.Equal(got, want) {
+		t.Fatalf("%s = %q, want %q", what, got, want)
+	}
+}
+
+func TestFiresOncePerPeriod(t *testing.T) {
+	e := newEnv(t)
+	var rec recorder
+	// Seeded reminder time is 16:30; boot just after it on Oct 6 runs it.
+	r := e.start(at(2026, 10, 6, 16, 31), reminder(&rec))
+	wantKeys(t, "after boot", rec.keys(), []string{"2026-10-06"})
+
+	r.runUntil(at(2026, 10, 7, 16, 29))
+	wantKeys(t, "before the fire time", rec.keys(), []string{"2026-10-06"})
+	r.runUntil(at(2026, 10, 9, 23, 59))
+	wantKeys(t, "after three days", rec.keys(), []string{"2026-10-06", "2026-10-07", "2026-10-08", "2026-10-09"})
+
+	for i, p := range rec.periods()[1:] {
+		want := at(2026, 10, 7+i, 16, 30)
+		if !p.FireAt.Equal(want) {
+			t.Errorf("run %s fired at %s, want %s", p.Key, p.FireAt, want)
+		}
+		if p.FireAt.Location().String() != zone {
+			t.Errorf("run %s fired in zone %s, want %s", p.Key, p.FireAt.Location(), zone)
+		}
+	}
+	wantKeys(t, "schedule_runs", e.claimed(scheduler.DayOfReminder), rec.keys())
+}
+
+func TestRestartMidPeriodDoesNotRerun(t *testing.T) {
+	e := newEnv(t)
+	var first recorder
+	r := e.start(at(2026, 10, 7, 12, 0), suggestions(&first)) // Wednesday
+	// The seeded Sunday 08:00 of Oct 4 is days ago: missed, not run.
+	wantKeys(t, "first boot", first.keys(), nil)
+	r.runUntil(at(2026, 10, 11, 9, 0))
+	wantKeys(t, "first process", first.keys(), []string{"2026-W41"})
+	r.stop()
+
+	// The restart is an hour after the Sunday 08:00 fire, mid-period.
+	var second recorder
+	r2 := e.start(at(2026, 10, 11, 9, 0), suggestions(&second))
+	wantKeys(t, "after restart", second.keys(), nil)
+	r2.runUntil(at(2026, 10, 18, 7, 59))
+	wantKeys(t, "rest of the period", second.keys(), nil)
+	r2.runUntil(at(2026, 10, 18, 8, 0))
+	wantKeys(t, "next period", second.keys(), []string{"2026-W42"})
+	wantKeys(t, "schedule_runs", e.claimed(scheduler.SuggestionsWeekly), []string{"2026-W40", "2026-W41", "2026-W42"})
+}
+
+func TestCatchUpWithin24Hours(t *testing.T) {
+	e := newEnv(t)
+	var rec recorder
+	// Sunday Oct 11 08:00 passed 23h59m before boot.
+	e.start(at(2026, 10, 12, 7, 59), suggestions(&rec))
+	wantKeys(t, "catch-up", rec.keys(), []string{"2026-W41"})
+	if p := rec.periods()[0]; !p.FireAt.Equal(at(2026, 10, 11, 8, 0)) {
+		t.Errorf("caught-up period fired at %s, want its fire time", p.FireAt)
+	}
+}
+
+func TestNoCatchUpBeyond24Hours(t *testing.T) {
+	e := newEnv(t)
+	var rec recorder
+	// Sunday Oct 11 08:00 passed 24h01m before boot.
+	r := e.start(at(2026, 10, 12, 8, 1), suggestions(&rec))
+	wantKeys(t, "boot", rec.keys(), nil)
+	// The missed period is recorded, so it never runs later.
+	wantKeys(t, "schedule_runs", e.claimed(scheduler.SuggestionsWeekly), []string{"2026-W41"})
+
+	// Moving the fire time to Sunday 23:00 puts that week's fire time
+	// within 24 h of now. The change itself never fires a passed time...
+	e.update(household.SettingsPatch{SuggestionsTime: ptr("23:00")})
+	r.changed()
+	wantKeys(t, "after a settings change", rec.keys(), nil)
+	r.stop()
+
+	// ...and the restart's catch-up finds the period already recorded.
+	var again recorder
+	r2 := e.start(at(2026, 10, 12, 8, 2), suggestions(&again))
+	wantKeys(t, "after restart", again.keys(), nil)
+	r2.runUntil(at(2026, 10, 18, 23, 0))
+	wantKeys(t, "next period", again.keys(), []string{"2026-W42"})
+}
+
+func TestCatchUpOnlyMostRecentPeriod(t *testing.T) {
+	e := newEnv(t)
+	var rec recorder
+	// Days of downtime: only Oct 9 16:30, 30 minutes ago, runs.
+	e.start(at(2026, 10, 9, 17, 0), reminder(&rec))
+	wantKeys(t, "catch-up", rec.keys(), []string{"2026-10-09"})
+	wantKeys(t, "schedule_runs", e.claimed(scheduler.DayOfReminder), []string{"2026-10-09"})
+}
+
+func TestSpringForward(t *testing.T) {
+	// Clocks in Chicago jump 02:00 CST → 03:00 CDT on Sunday 2026-03-08.
+	e := newEnv(t)
+	e.update(household.SettingsPatch{ReminderTime: ptr("08:00"), ReviewTime: ptr("02:30")})
+	var daily, skipped recorder
+	review := scheduler.Schedule{
+		Name: scheduler.RotationCheck,
+		Rule: func(s household.Settings) scheduler.Rule { return scheduler.Daily(s.ReviewTime) },
+		Run:  skipped.run,
+	}
+	r := e.start(at(2026, 3, 7, 9, 0), reminder(&daily), review)
+	r.runUntil(at(2026, 3, 9, 12, 0))
+
+	wantKeys(t, "08:00 runs", daily.keys(), []string{"2026-03-07", "2026-03-08", "2026-03-09"})
+	wantFired(t, daily.periods(),
+		time.Date(2026, 3, 7, 14, 0, 0, 0, time.UTC), // 08:00 CST
+		time.Date(2026, 3, 8, 13, 0, 0, 0, time.UTC), // 08:00 CDT
+		time.Date(2026, 3, 9, 13, 0, 0, 0, time.UTC),
+	)
+
+	// 02:30 does not exist on Mar 8: it fires once, an hour later at
+	// 03:30 CDT, keyed by Mar 8.
+	wantKeys(t, "02:30 runs", skipped.keys(), []string{"2026-03-07", "2026-03-08", "2026-03-09"})
+	wantFired(t, skipped.periods(),
+		time.Date(2026, 3, 7, 8, 30, 0, 0, time.UTC), // 02:30 CST
+		time.Date(2026, 3, 8, 8, 30, 0, 0, time.UTC), // 03:30 CDT
+		time.Date(2026, 3, 9, 7, 30, 0, 0, time.UTC), // 02:30 CDT
+	)
+	if got := skipped.periods()[1].FireAt.Format("15:04 MST"); got != "03:30 CDT" {
+		t.Errorf("skipped 02:30 fired at %s local, want 03:30 CDT", got)
+	}
+}
+
+func TestFallBack(t *testing.T) {
+	// Clocks in Chicago go back 02:00 CDT → 01:00 CST on Sunday 2026-11-01.
+	e := newEnv(t)
+	e.update(household.SettingsPatch{ReminderTime: ptr("08:00"), ReviewTime: ptr("01:30")})
+	var daily, repeated recorder
+	review := scheduler.Schedule{
+		Name: scheduler.RotationCheck,
+		Rule: func(s household.Settings) scheduler.Rule { return scheduler.Daily(s.ReviewTime) },
+		Run:  repeated.run,
+	}
+	r := e.start(at(2026, 10, 31, 9, 0), reminder(&daily), review)
+	r.runUntil(at(2026, 11, 2, 12, 0))
+
+	wantKeys(t, "08:00 runs", daily.keys(), []string{"2026-10-31", "2026-11-01", "2026-11-02"})
+	wantFired(t, daily.periods(),
+		time.Date(2026, 10, 31, 13, 0, 0, 0, time.UTC), // 08:00 CDT
+		time.Date(2026, 11, 1, 14, 0, 0, 0, time.UTC),  // 08:00 CST
+		time.Date(2026, 11, 2, 14, 0, 0, 0, time.UTC),
+	)
+
+	// 01:30 happens twice on Nov 1: it fires once, at the first (CDT).
+	wantKeys(t, "01:30 runs", repeated.keys(), []string{"2026-10-31", "2026-11-01", "2026-11-02"})
+	wantFired(t, repeated.periods(),
+		time.Date(2026, 10, 31, 6, 30, 0, 0, time.UTC), // 01:30 CDT
+		time.Date(2026, 11, 1, 6, 30, 0, 0, time.UTC),  // first 01:30, CDT
+		time.Date(2026, 11, 2, 7, 30, 0, 0, time.UTC),  // 01:30 CST
+	)
+}
+
+func TestFallBackSecondOccurrenceDoesNotRerun(t *testing.T) {
+	e := newEnv(t)
+	e.update(household.SettingsPatch{ReviewTime: ptr("01:30")})
+	var rec recorder
+	review := scheduler.Schedule{
+		Name: scheduler.RotationCheck,
+		Rule: func(s household.Settings) scheduler.Rule { return scheduler.Daily(s.ReviewTime) },
+		Run:  rec.run,
+	}
+	// Boot between the two 01:30s of Nov 1: the first ran in an earlier
+	// process, so the second is not a new period.
+	first := time.Date(2026, 11, 1, 6, 30, 0, 0, time.UTC)
+	r := e.start(first.Add(-time.Minute), review)
+	r.runUntil(first)
+	wantKeys(t, "first 01:30", rec.keys(), []string{"2026-10-31", "2026-11-01"})
+	r.stop()
+
+	var again recorder
+	review.Run = again.run
+	r2 := e.start(first.Add(30*time.Minute), review)
+	r2.runUntil(first.Add(2 * time.Hour)) // past the second 01:30, CST
+	wantKeys(t, "second 01:30", again.keys(), nil)
+}
+
+func wantFired(t *testing.T, got []scheduler.Period, want ...time.Time) {
+	t.Helper()
+	if len(got) != len(want) {
+		t.Fatalf("%d runs, want %d", len(got), len(want))
+	}
+	for i := range want {
+		if !got[i].FireAt.Equal(want[i]) {
+			t.Errorf("run %s fired at %s (%s), want %s", got[i].Key, got[i].FireAt.UTC(), got[i].FireAt, want[i])
+		}
+	}
+}
+
+func TestISOWeek53Boundary(t *testing.T) {
+	e := newEnv(t)
+	var sugg, draft recorder
+	// Suggestions Sunday 08:00 keys the ISO week ending that Sunday; the
+	// plan draft Friday 09:00 keys the week after.
+	r := e.start(at(2026, 12, 20, 9, 0), suggestions(&sugg), planDraft(&draft))
+	r.runUntil(at(2027, 1, 11, 0, 0))
+
+	wantKeys(t, "suggestions", sugg.keys(), []string{"2026-W51", "2026-W52", "2026-W53", "2027-W01"})
+	// Sunday 2027-01-03 is the last day of 2026-W53.
+	if p := sugg.periods()[2]; p.Date.String() != "2027-01-03" || p.Week.String() != "2026-12-28" {
+		t.Errorf("2026-W53 suggestions ran on %s for week %s, want 2027-01-03 for 2026-12-28", p.Date, p.Week)
+	}
+	// Friday Dec 18, two days before boot, was missed.
+	wantKeys(t, "plan drafts", draft.keys(), []string{"2026-W53", "2027-W01", "2027-W02"})
+	// Friday 2026-12-25 drafts 2026-W53; Friday 2027-01-01 drafts 2027-W01.
+	if p := draft.periods()[0]; p.Date.String() != "2026-12-25" || p.Week.String() != "2026-12-28" {
+		t.Errorf("2026-W53 draft ran on %s for week %s, want 2026-12-25 for 2026-12-28", p.Date, p.Week)
+	}
+	if p := draft.periods()[1]; p.Date.String() != "2027-01-01" || p.Week.String() != "2027-01-04" {
+		t.Errorf("2027-W01 draft ran on %s for week %s, want 2027-01-01 for 2027-01-04", p.Date, p.Week)
+	}
+}
+
+func TestPeriodKey(t *testing.T) {
+	date := func(s string) store.LocalDate {
+		d, err := store.ParseLocalDate(s)
+		if err != nil {
+			t.Fatal(err)
+		}
+		return d
+	}
+	tests := []struct {
+		rule scheduler.Rule
+		date string
+		want string
+	}{
+		{scheduler.Daily("16:30"), "2026-12-31", "2026-12-31"},
+		{scheduler.Weekly(time.Monday, "00:00"), "2026-12-28", "2026-W53"},
+		{scheduler.Weekly(time.Sunday, "08:00"), "2027-01-03", "2026-W53"},
+		{scheduler.Weekly(time.Monday, "00:00"), "2027-01-04", "2027-W01"},
+		{scheduler.WeeklyForNextWeek(time.Friday, "09:00"), "2026-12-25", "2026-W53"},
+		{scheduler.WeeklyForNextWeek(time.Friday, "09:00"), "2027-01-01", "2027-W01"},
+	}
+	for _, tt := range tests {
+		if got := tt.rule.PeriodKey(date(tt.date)); got != tt.want {
+			t.Errorf("%+v.PeriodKey(%s) = %s, want %s", tt.rule, tt.date, got, tt.want)
+		}
+	}
+}
+
+func TestSettingsChangeMovesNextFire(t *testing.T) {
+	e := newEnv(t)
+	var rec recorder
+	r := e.start(at(2026, 10, 7, 12, 0), reminder(&rec))
+	wantKeys(t, "boot", rec.keys(), []string{"2026-10-06"})
+
+	// Earlier but still ahead: fires at the new time, not the old.
+	e.update(household.SettingsPatch{ReminderTime: ptr("13:00")})
+	r.changed()
+	if next, _ := r.clock.nextDeadline(); !next.Equal(at(2026, 10, 7, 13, 0)) {
+		t.Errorf("next wake %s, want the new 13:00", next)
+	}
+	r.runUntil(at(2026, 10, 7, 13, 0))
+	wantKeys(t, "at the new time", rec.keys(), []string{"2026-10-06", "2026-10-07"})
+	r.runUntil(at(2026, 10, 7, 23, 0))
+	wantKeys(t, "at the old time", rec.keys(), []string{"2026-10-06", "2026-10-07"})
+
+	// Moved to a time already passed today: nothing fires now, and
+	// tomorrow fires at the new time.
+	r.runUntil(at(2026, 10, 8, 12, 0))
+	e.update(household.SettingsPatch{ReminderTime: ptr("11:00")})
+	r.changed()
+	wantKeys(t, "moved into the past", rec.keys(), []string{"2026-10-06", "2026-10-07"})
+	r.runUntil(at(2026, 10, 9, 10, 59))
+	wantKeys(t, "before the new time", rec.keys(), []string{"2026-10-06", "2026-10-07"})
+	r.runUntil(at(2026, 10, 9, 11, 0))
+	wantKeys(t, "the new time", rec.keys(), []string{"2026-10-06", "2026-10-07", "2026-10-09"})
+}
+
+// A change saved at 12:00 moving the 16:30 reminder to 12:01 must fire it
+// even when the scheduler only gets to the change at 12:02, after the new
+// fire time has passed.
+func TestSettingsChangeAppliedLateStillFires(t *testing.T) {
+	e := newEnv(t)
+	var rec recorder
+	// blocker holds the scheduler in a Run at 12:00 on 2026-10-07 while
+	// the settings change and the clock moves on.
+	started := make(chan struct{})
+	release := make(chan struct{})
+	blocker := scheduler.Schedule{
+		Name: scheduler.RotationCheck,
+		Rule: func(household.Settings) scheduler.Rule { return scheduler.Daily("12:00") },
+		Run: func(_ context.Context, p scheduler.Period) error {
+			if p.Key == "2026-10-07" {
+				close(started)
+				<-release
+			}
+			return nil
+		},
+	}
+	r := e.start(at(2026, 10, 7, 11, 0), reminder(&rec), blocker)
+	wantKeys(t, "boot", rec.keys(), []string{"2026-10-06"})
+
+	r.clock.set(at(2026, 10, 7, 12, 0))
+	<-started
+	e.update(household.SettingsPatch{ReminderTime: ptr("12:01")})
+	r.clock.set(at(2026, 10, 7, 12, 2))
+	close(release)
+	r.idle() // the pass the blocker held
+	r.idle() // the pass that applies the change
+	wantKeys(t, "after the late change", rec.keys(), []string{"2026-10-06", "2026-10-07"})
+
+	r.runUntil(at(2026, 10, 7, 23, 0))
+	wantKeys(t, "end of day", rec.keys(), []string{"2026-10-06", "2026-10-07"})
+	wantKeys(t, "claims", e.claimed(scheduler.DayOfReminder), []string{"2026-10-06", "2026-10-07"})
+}
+
+// failClaims makes every claim fail until recover is called.
+func (e *env) failClaims() (recover func()) {
+	e.t.Helper()
+	for _, q := range []string{
+		`CREATE TABLE fail_claims (on_ INTEGER NOT NULL)`,
+		`INSERT INTO fail_claims VALUES (1)`,
+		`CREATE TRIGGER fail_claim BEFORE INSERT ON schedule_runs
+			WHEN (SELECT on_ FROM fail_claims) = 1
+			BEGIN SELECT RAISE(ABORT, 'injected claim failure'); END`,
+	} {
+		if _, err := e.st.Writer.Exec(q); err != nil {
+			e.t.Fatal(err)
+		}
+	}
+	return func() {
+		e.t.Helper()
+		if _, err := e.st.Writer.Exec(`UPDATE fail_claims SET on_ = 0`); err != nil {
+			e.t.Fatal(err)
+		}
+	}
+}
+
+// A claim that fails is retried, not skipped, and the period then runs once.
+func TestClaimFailureIsRetried(t *testing.T) {
+	e := newEnv(t)
+	recoverClaims := e.failClaims()
+	var rec recorder
+	r := e.start(at(2026, 10, 7, 12, 0), reminder(&rec))
+	wantKeys(t, "boot with failing claims", rec.keys(), nil)
+	wantKeys(t, "claims while failing", e.claimed(scheduler.DayOfReminder), nil)
+	if next, _ := r.clock.nextDeadline(); !next.Equal(at(2026, 10, 7, 12, 0).Add(scheduler.ClaimRetry)) {
+		t.Errorf("next wake %s, want a retry after %s", next, scheduler.ClaimRetry)
+	}
+
+	recoverClaims()
+	r.runUntil(at(2026, 10, 7, 12, 0).Add(scheduler.ClaimRetry))
+	wantKeys(t, "after the retry", rec.keys(), []string{"2026-10-06"})
+	r.runUntil(at(2026, 10, 7, 23, 0))
+	wantKeys(t, "end of day", rec.keys(), []string{"2026-10-06", "2026-10-07"})
+	wantKeys(t, "claims", e.claimed(scheduler.DayOfReminder), []string{"2026-10-06", "2026-10-07"})
+}
+
+// A claim that keeps failing past the next daily fire is retried for its
+// own period, not the most recent one: that period, now more than
+// CatchUpWindow old, is recorded as missed, and the later period runs once.
+func TestClaimFailureRecoversAfterNextFire(t *testing.T) {
+	e := newEnv(t)
+	recoverClaims := e.failClaims()
+	var rec recorder
+	// The 7 October 16:30 reminder is due and its claim fails, past the
+	// 8 October 16:30 fire.
+	r := e.start(at(2026, 10, 7, 16, 45), reminder(&rec))
+	r.runUntil(at(2026, 10, 8, 17, 0))
+	wantKeys(t, "runs while failing", rec.keys(), nil)
+	wantKeys(t, "claims while failing", e.claimed(scheduler.DayOfReminder), nil)
+
+	// Claims recover 24h31m after the 7 October fire.
+	recoverClaims()
+	r.runUntil(at(2026, 10, 8, 17, 1))
+	wantKeys(t, "after recovery", rec.keys(), []string{"2026-10-08"})
+	wantKeys(t, "claims after recovery", e.claimed(scheduler.DayOfReminder), []string{"2026-10-07", "2026-10-08"})
+
+	r.runUntil(at(2026, 10, 9, 16, 0))
+	wantKeys(t, "next day", rec.keys(), []string{"2026-10-08"})
+}
+
+func TestSettingsChangeTimezone(t *testing.T) {
+	e := newEnv(t)
+	var rec recorder
+	r := e.start(at(2026, 10, 7, 12, 0), reminder(&rec))
+	e.update(household.SettingsPatch{Timezone: ptr("America/New_York")})
+	r.changed()
+	// 16:30 in New York is 15:30 in Chicago.
+	r.runUntil(at(2026, 10, 7, 15, 30))
+	wantKeys(t, "after the zone change", rec.keys(), []string{"2026-10-06", "2026-10-07"})
+	if got := rec.periods()[1].FireAt.Location().String(); got != "America/New_York" {
+		t.Errorf("fired in zone %s, want America/New_York", got)
+	}
+}
+
+func TestRunErrorKeepsPeriodClaimed(t *testing.T) {
+	e := newEnv(t)
+	rec := recorder{err: errors.New("boom")}
+	r := e.start(at(2026, 10, 7, 16, 0), reminder(&rec))
+	r.runUntil(at(2026, 10, 7, 23, 0))
+	wantKeys(t, "runs", rec.keys(), []string{"2026-10-06", "2026-10-07"})
+	r.stop()
+	var again recorder
+	e.start(at(2026, 10, 7, 23, 0), reminder(&again))
+	wantKeys(t, "after restart", again.keys(), nil)
+}
+
+func TestRegister(t *testing.T) {
+	e := newEnv(t)
+	s, err := scheduler.New(e.st, e.settings, slog.New(slog.NewTextHandler(io.Discard, nil)))
+	if err != nil {
+		t.Fatal(err)
+	}
+	var rec recorder
+	bad := []scheduler.Schedule{
+		{Name: "hourly", Rule: reminder(&rec).Rule, Run: rec.run},
+		{Name: scheduler.DayOfReminder, Run: rec.run},
+		{Name: scheduler.DayOfReminder, Rule: reminder(&rec).Rule},
+	}
+	for _, sch := range bad {
+		if err := s.Register(sch); err == nil {
+			t.Errorf("Register(%q) accepted an invalid schedule", sch.Name)
+		}
+	}
+	if err := s.Register(reminder(&rec)); err != nil {
+		t.Fatal(err)
+	}
+	if err := s.Register(reminder(&rec)); err == nil {
+		t.Error("Register accepted a second schedule with the same name")
+	}
+}
+
+func TestStartRejectsInvalidRule(t *testing.T) {
+	e := newEnv(t)
+	s, err := scheduler.New(e.st, e.settings, slog.New(slog.NewTextHandler(io.Discard, nil)))
+	if err != nil {
+		t.Fatal(err)
+	}
+	var rec recorder
+	if err := s.Register(scheduler.Schedule{
+		Name: scheduler.WeekRollover,
+		Rule: func(household.Settings) scheduler.Rule { return scheduler.Weekly(7, "00:00") },
+		Run:  rec.run,
+	}); err != nil {
+		t.Fatal(err)
+	}
+	if err := s.Start(context.Background()); err == nil {
+		s.Wait()
+		t.Fatal("Start accepted weekday 7")
+	}
+}
+
+func TestStopEndsTheLoop(t *testing.T) {
+	e := newEnv(t)
+	var rec recorder
+	r := e.start(at(2026, 10, 7, 12, 0), reminder(&rec))
+	done := make(chan struct{})
+	go func() {
+		r.stop()
+		close(done)
+	}()
+	select {
+	case <-done:
+	case <-time.After(10 * time.Second):
+		t.Fatal("scheduler did not stop")
+	}
+}

internal/scheduler/suggestions_test.go

@@ -0,0 +1,370 @@
+package scheduler_test
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"log/slog"
+	"strings"
+	"sync"
+	"testing"
+	"time"
+
+	"code.cubecraftlabs.com/Maniacs_Mansion/Meal-Quest/internal/ai/harness"
+	"code.cubecraftlabs.com/Maniacs_Mansion/Meal-Quest/internal/ai/llm"
+	"code.cubecraftlabs.com/Maniacs_Mansion/Meal-Quest/internal/ai/skills"
+	"code.cubecraftlabs.com/Maniacs_Mansion/Meal-Quest/internal/db"
+	"code.cubecraftlabs.com/Maniacs_Mansion/Meal-Quest/internal/household"
+	"code.cubecraftlabs.com/Maniacs_Mansion/Meal-Quest/internal/jobs"
+	"code.cubecraftlabs.com/Maniacs_Mansion/Meal-Quest/internal/notify"
+	"code.cubecraftlabs.com/Maniacs_Mansion/Meal-Quest/internal/notify/notifytest"
+	"code.cubecraftlabs.com/Maniacs_Mansion/Meal-Quest/internal/recipes"
+	"code.cubecraftlabs.com/Maniacs_Mansion/Meal-Quest/internal/recipes/recipestest"
+	"code.cubecraftlabs.com/Maniacs_Mansion/Meal-Quest/internal/scheduler"
+	"code.cubecraftlabs.com/Maniacs_Mansion/Meal-Quest/internal/store"
+	"code.cubecraftlabs.com/Maniacs_Mansion/Meal-Quest/internal/suggest"
+)
+
+// The tests in this file run suggest's suggestions_weekly schedule on the
+// fake clock against the real suggestion service, job runner, harness and
+// notification outbox, with the AI and recipe sources faked.
+
+// suggestEnv is a suggestion service over a scheduler env's store, the
+// "process" a scheduler runs it in. side answers suggest_side_quests and
+// model normalize_recipe while grounding.
+type suggestEnv struct {
+	e      *env
+	svc    *suggest.Service
+	side   *llm.Fake
+	model  *llm.Fake
+	search *recipestest.Search
+	pages  *recipestest.Pages
+	images *recipestest.Images
+	logs   *syncBuffer
+	stop   func()
+}
+
+// syncBuffer collects log output written by concurrent goroutines.
+type syncBuffer struct {
+	mu  sync.Mutex
+	buf strings.Builder
+}
+
+func (b *syncBuffer) Write(p []byte) (int, error) {
+	b.mu.Lock()
+	defer b.mu.Unlock()
+	return b.buf.Write(p)
+}
+
+func (b *syncBuffer) String() string {
+	b.mu.Lock()
+	defer b.mu.Unlock()
+	return b.buf.String()
+}
+
+// regulars stores the Regulars Beef Chili and Chicken Fajitas, so a batch
+// has Regulars to suggest from.
+func (e *env) regulars() {
+	e.t.Helper()
+	for _, name := range []string{"Beef Chili", "Chicken Fajitas"} {
+		now := store.Now().String()
+		if _, err := db.New(e.st.Writer).CreateDinner(context.Background(), db.CreateDinnerParams{
+			ID:              store.NewID(),
+			Name:            name,
+			NormalizedName:  harness.Normalize(name),
+			Status:          "regular",
+			Origin:          "regular",
+			GroundingStatus: "manual",
+			CreatedAt:       now,
+			UpdatedAt:       now,
+		}); err != nil {
+			e.t.Fatal(err)
+		}
+	}
+}
+
+// suggestions builds a suggestion service over e with a started job
+// runner, which stop stops (and the test's end, if not before).
+func (e *env) suggestions() *suggestEnv {
+	e.t.Helper()
+	s := &suggestEnv{
+		e:      e,
+		side:   &llm.Fake{},
+		model:  &llm.Fake{},
+		search: &recipestest.Search{Results: map[string][]recipes.Candidate{}},
+		pages:  &recipestest.Pages{Bodies: map[string][]byte{}},
+		images: &recipestest.Images{JPEGs: map[string][]byte{}},
+		logs:   &syncBuffer{},
+	}
+	g := recipestest.NewGrounder(e.t, e.st, e.t.TempDir(), s.search, s.pages, s.images, s.model)
+	guard, err := harness.NewGuard(harness.DefaultMaxCalls)
+	if err != nil {
+		e.t.Fatal(err)
+	}
+	cfg := harness.Config{
+		Settings:    llm.Settings{Provider: "anthropic", Model: "claude-test", Temperature: 0.7},
+		NewProvider: func(llm.Settings, string) (llm.Provider, error) { return s.side, nil },
+		Store:       e.st,
+		Guard:       guard,
+	}
+	suggester := func(ctx context.Context, scope string, in skills.SideQuestsRequest) (skills.SideQuests, error) {
+		return skills.SuggestSideQuests(ctx, cfg, scope, in)
+	}
+	logger := slog.New(slog.NewTextHandler(s.logs, nil))
+	runner, err := jobs.NewRunner(e.st, logger, 1)
+	if err != nil {
+		e.t.Fatal(err)
+	}
+	s.svc, err = suggest.NewService(e.st, g, suggester, runner, notifytest.New(e.t, e.st), logger)
+	if err != nil {
+		e.t.Fatal(err)
+	}
+	ctx, cancel := context.WithCancel(context.Background())
+	if err := runner.Start(ctx); err != nil {
+		cancel()
+		e.t.Fatal(err)
+	}
+	var once sync.Once
+	s.stop = func() {
+		once.Do(func() {
+			cancel()
+			runner.Wait()
+		})
+	}
+	e.t.Cleanup(s.stop)
+	return s
+}
+
+// week scripts the next batch: suggest_side_quests answers with the
+// Side quests named names, each grounded by its own recipe.
+func (s *suggestEnv) week(names ...string) {
+	s.e.t.Helper()
+	var quests []string
+	for _, name := range names {
+		slug := strings.ReplaceAll(strings.ToLower(name), " ", "-")
+		page, photo := "https://r.example/"+slug, "https://r.example/"+slug+".jpg"
+		s.search.Results[name] = []recipes.Candidate{{URL: page}}
+		s.pages.Bodies[page] = recipestest.RecipePage(name, "R Kitchen", photo, 35)
+		s.images.JPEGs[photo] = recipestest.JPEG(s.e.t, 400, 320, uint8(len(s.images.JPEGs)+1))
+		s.model.QueueGenerate(recipestest.NormalizedReply("stovetop", 35))
+		quests = append(quests, fmt.Sprintf(`{"name":%q,"cuisine":"Fusion","method":"stovetop","est_minutes":35,"difficulty":"easy","reason":"Like your Beef Chili.","cited_regulars":["Beef Chili"],"search_query":%q}`, name, name))
+	}
+	js := `{"suggestions":[` + strings.Join(quests, ",") + `]}`
+	s.side.QueueGenerate(llm.GenerateResult{Response: llm.Response{Text: js, JSON: json.RawMessage(js)}})
+}
+
+// batch is a suggestion_batches row.
+type batch struct {
+	ID, Trigger, Status string
+}
+
+// batches returns every suggestion batch, oldest first.
+func (e *env) batches() []batch {
+	e.t.Helper()
+	rows, err := e.st.Reader.Query(`SELECT id, trigger, status FROM suggestion_batches ORDER BY created_at, id`)
+	if err != nil {
+		e.t.Fatal(err)
+	}
+	defer rows.Close()
+	var out []batch
+	for rows.Next() {
+		var b batch
+		if err := rows.Scan(&b.ID, &b.Trigger, &b.Status); err != nil {
+			e.t.Fatal(err)
+		}
+		out = append(out, b)
+	}
+	if err := rows.Err(); err != nil {
+		e.t.Fatal(err)
+	}
+	return out
+}
+
+// settled waits until no batch is pending or running and returns the
+// batches.
+func (e *env) settled() []batch {
+	e.t.Helper()
+	deadline := time.Now().Add(10 * time.Second)
+	for {
+		bs := e.batches()
+		busy := false
+		for _, b := range bs {
+			busy = busy || b.Status == suggest.BatchPending || b.Status == suggest.BatchRunning
+		}
+		if !busy {
+			return bs
+		}
+		if time.Now().After(deadline) {
+			e.t.Fatalf("batches still under way: %+v", bs)
+		}
+		time.Sleep(5 * time.Millisecond)
+	}
+}
+
+// wantBatches fails unless the batches are scheduled with statuses want.
+func wantBatches(t *testing.T, what string, got []batch, want ...string) {
+	t.Helper()
+	ok := len(got) == len(want)
+	for i := 0; ok && i < len(got); i++ {
+		ok = got[i].Trigger == suggest.TriggerScheduled && got[i].Status == want[i]
+	}
+	if !ok {
+		t.Fatalf("%s: batches = %+v, want scheduled %q", what, got, want)
+	}
+}
+
+// notified returns the new_suggestions keys stored, oldest first, and
+// reports an error for any whose title is not the one titles gives it.
+func (e *env) notified(titles ...string) []string {
+	e.t.Helper()
+	var keys []string
+	rows := notifytest.Rows(e.t, e.st, notify.NewSuggestions)
+	for i, r := range rows {
+		keys = append(keys, r.Key)
+		if i < len(titles) && r.Payload.Title != titles[i] {
+			e.t.Errorf("new_suggestions %s title = %q, want %q", r.Key, r.Payload.Title, titles[i])
+		}
+	}
+	return keys
+}
+
+// TestSuggestionsWeeklyOneBatchAndNotificationPerWeek: at the seeded
+// Sunday 08:00 the schedule starts one scheduled batch a week, period
+// suggestions_weekly:YYYY-Www, and each tells the household once, with
+// new_suggestions:YYYY-Www counting its new Side quests.
+func TestSuggestionsWeeklyOneBatchAndNotificationPerWeek(t *testing.T) {
+	e := newEnv(t)
+	e.regulars()
+	s := e.suggestions()
+	s.week("Chili Verde", "Chicken Shawarma Bowls")
+	s.week("Lamb Kofta", "Black Bean Tostadas", "Pork Carnitas")
+
+	// Wednesday Oct 7: the Sunday Oct 4 08:00 fire is days ago, missed.
+	r := e.start(at(2026, 10, 7, 12, 0), s.svc.WeeklySchedule())
+	r.runUntil(at(2026, 10, 11, 7, 59))
+	wantBatches(t, "before Sunday 08:00", e.settled())
+	wantKeys(t, "new_suggestions before Sunday 08:00", e.notified(), nil)
+
+	r.runUntil(at(2026, 10, 11, 8, 0))
+	wantBatches(t, "at Sunday 08:00", e.settled(), suggest.BatchDone)
+	wantKeys(t, "new_suggestions of W41", e.notified("2 new Side quests are in"), []string{"new_suggestions:2026-W41"})
+
+	r.runUntil(at(2026, 10, 18, 7, 59))
+	wantBatches(t, "rest of W41", e.settled(), suggest.BatchDone)
+	r.runUntil(at(2026, 10, 18, 9, 0))
+	wantBatches(t, "W42", e.settled(), suggest.BatchDone, suggest.BatchDone)
+	wantKeys(t, "new_suggestions of W41 and W42", e.notified("2 new Side quests are in", "3 new Side quests are in"),
+		[]string{"new_suggestions:2026-W41", "new_suggestions:2026-W42"})
+	wantKeys(t, "schedule_runs", e.claimed(scheduler.SuggestionsWeekly), []string{"2026-W40", "2026-W41", "2026-W42"})
+}
+
+// TestSuggestionsWeeklyFollowsSettings: the schedule fires at the
+// household's suggestions_dow and suggestions_time.
+func TestSuggestionsWeeklyFollowsSettings(t *testing.T) {
+	e := newEnv(t)
+	e.regulars()
+	e.update(household.SettingsPatch{SuggestionsDow: ptr(3), SuggestionsTime: ptr("18:30")})
+	s := e.suggestions()
+	s.week("Chili Verde")
+
+	r := e.start(at(2026, 10, 12, 12, 0), s.svc.WeeklySchedule()) // Monday
+	r.runUntil(at(2026, 10, 14, 18, 29))                          // Wednesday
+	wantBatches(t, "before Wednesday 18:30", e.settled())
+	r.runUntil(at(2026, 10, 14, 18, 30))
+	wantBatches(t, "at Wednesday 18:30", e.settled(), suggest.BatchDone)
+	wantKeys(t, "new_suggestions", e.notified("1 new Side quest is in"), []string{"new_suggestions:2026-W42"})
+}
+
+// TestSuggestionsWeeklyAIFailureSendsNothing: when suggest_side_quests
+// fails the batch fails with the failure logged and recorded in ai_runs,
+// and the household is not notified.
+func TestSuggestionsWeeklyAIFailureSendsNothing(t *testing.T) {
+	e := newEnv(t)
+	e.regulars()
+	s := e.suggestions()
+	s.side.QueueGenerate(llm.GenerateResult{Err: errors.New("provider down")})
+
+	r := e.start(at(2026, 10, 11, 7, 0), s.svc.WeeklySchedule())
+	r.runUntil(at(2026, 10, 11, 8, 0))
+	wantBatches(t, "after the AI failed", e.settled(), suggest.BatchFailed)
+	wantKeys(t, "new_suggestions after the AI failed", e.notified(), nil)
+	// W40 is the boot's missed period, recorded without running.
+	wantKeys(t, "schedule_runs", e.claimed(scheduler.SuggestionsWeekly), []string{"2026-W40", "2026-W41"})
+
+	var status, msg string
+	if err := e.st.Reader.QueryRow(`SELECT status, COALESCE(error, '') FROM ai_runs WHERE skill = 'suggest_side_quests'`).Scan(&status, &msg); err != nil {
+		t.Fatalf("ai_runs row of suggest_side_quests: %v", err)
+	}
+	if status != string(harness.StatusProviderError) || msg == "" {
+		t.Errorf("ai_runs suggest_side_quests = %s %q, want %s with its message", status, msg, harness.StatusProviderError)
+	}
+	if logs := s.logs.String(); !strings.Contains(logs, "suggestion batch failed") || !strings.Contains(logs, "reason="+suggest.ReasonAIFailed) {
+		t.Errorf("logs = %s, want the batch failure with reason %s", logs, suggest.ReasonAIFailed)
+	}
+}
+
+// TestSuggestionsWeeklyNoDuplicateAfterRestart: a restart after the week's
+// batch ran, still within the catch-up window, starts no second batch and
+// sends no second notification: the period is already claimed.
+func TestSuggestionsWeeklyNoDuplicateAfterRestart(t *testing.T) {
+	e := newEnv(t)
+	e.regulars()
+	s := e.suggestions()
+	s.week("Chili Verde", "Chicken Shawarma Bowls")
+	r := e.start(at(2026, 10, 11, 7, 0), s.svc.WeeklySchedule())
+	r.runUntil(at(2026, 10, 11, 8, 0))
+	wantBatches(t, "first process", e.settled(), suggest.BatchDone)
+	r.stop()
+	s.stop()
+
+	// The restarted process has a new job runner, suggestion service and
+	// scheduler; it would answer a second batch if one were started.
+	s2 := e.suggestions()
+	s2.week("Lamb Kofta")
+	r2 := e.start(at(2026, 10, 11, 9, 0), s2.svc.WeeklySchedule())
+	r2.runUntil(at(2026, 10, 18, 7, 59))
+	wantBatches(t, "after restart", e.settled(), suggest.BatchDone)
+	wantKeys(t, "new_suggestions after restart", e.notified(), []string{"new_suggestions:2026-W41"})
+	wantKeys(t, "schedule_runs", e.claimed(scheduler.SuggestionsWeekly), []string{"2026-W40", "2026-W41"})
+}
+
+// TestSuggestionsWeeklyCatchUpAfterTheWeekEnds: a Sunday fire caught up on
+// Monday, in the next ISO week, still runs W41's batch keyed by W41, and
+// the next Sunday's batch notifies for W42.
+func TestSuggestionsWeeklyCatchUpAfterTheWeekEnds(t *testing.T) {
+	e := newEnv(t)
+	e.regulars()
+	s := e.suggestions()
+	s.week("Chili Verde")
+	s.week("Lamb Kofta")
+
+	r := e.start(at(2026, 10, 12, 7, 0), s.svc.WeeklySchedule()) // Monday, 23h late
+	wantBatches(t, "after catch-up", e.settled(), suggest.BatchDone)
+	wantKeys(t, "new_suggestions after catch-up", e.notified(), []string{"new_suggestions:2026-W41"})
+	r.runUntil(at(2026, 10, 18, 8, 0))
+	wantBatches(t, "W42", e.settled(), suggest.BatchDone, suggest.BatchDone)
+	wantKeys(t, "new_suggestions", e.notified(), []string{"new_suggestions:2026-W41", "new_suggestions:2026-W42"})
+}
+
+// TestSuggestionsWeeklyStartsNoBatchWhileOneIsUnderWay: a period that
+// fires while a batch is pending starts no second one and says so.
+func TestSuggestionsWeeklyStartsNoBatchWhileOneIsUnderWay(t *testing.T) {
+	e := newEnv(t)
+	e.regulars()
+	s := e.suggestions()
+	s.stop() // the manual batch stays pending
+	manual, err := s.svc.StartBatch(context.Background(), suggest.TriggerManual)
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	r := e.start(at(2026, 10, 11, 7, 0), s.svc.WeeklySchedule())
+	r.runUntil(at(2026, 10, 11, 8, 0))
+	if bs := e.batches(); len(bs) != 1 || bs[0].ID != manual.ID {
+		t.Fatalf("batches = %+v, want only the manual batch %s", bs, manual.ID)
+	}
+	if logs := s.logs.String(); !strings.Contains(logs, "scheduled suggestion batch not started") || !strings.Contains(logs, "period=2026-W41") {
+		t.Errorf("logs = %s, want the skipped period logged", logs)
+	}
+}

internal/suggest/batch.go

@@ -22,6 +22,9 @@ import (
 // BatchJob is the KindSuggestionBatch payload: the batch to run.
 type BatchJob struct {
 	BatchID string `json:"batch_id"`
+	// Week is the YYYY-MM-DD Monday of the ISO week a scheduled batch is
+	// for, which keys its new_suggestions; empty for any other trigger.
+	Week string `json:"week,omitempty"`
 }
 
 // Batch stages, as the job's result reports them.
@@ -50,19 +53,22 @@ func (r batchResult) PublicJSON() ([]byte, error) { return json.Marshal(r) }
 // wanted at a time, until BatchSize are kept or none are left. A Side quest
 // no recipe grounds is dropped; any other grounding failure fails the
 // batch, as does suggest_side_quests failing, with no Side quest made up
-// instead. Each round's grounded Side quests are stored before the next
-// round starts, so the deck fills as the batch goes. The batch ends done
-// or failed; Side quests stored before a failure stay in the deck. When
-// recording the batch's end fails, so does the job; when it times out at
-// shutdown, the job is left running and the batch with it, for the next
-// start to fail.
+// instead. The batch ends done or failed. When recording the batch's end
+// fails, so does the job; when it times out at shutdown, the job is left
+// running and the batch with it, for the next start to fail.
 //
-// A scheduled or manual batch emits new_suggestions, keyed by the batch,
-// in the transaction that first adds Side quests to the deck, so the
-// household is told of every Side quest stored, whether or not the batch
-// then ends done: a later round, or recording the batch's end, may fail.
-// A replenish batch, which tops up the deck as the household swipes, never
-// notifies. See store.
+// When a batch stores its Side quests depends on whether it notifies (see
+// noticeOf). A scheduled or manual batch notifies: it holds every round's
+// grounded Side quests and, once the last round is done, stores them all
+// in one transaction with one new_suggestions counting and naming every
+// Side quest stored. So the household is told of exactly what the batch
+// added, once; and a batch that fails at any round, including a later
+// round's normalize_recipe, stores nothing and sends nothing, leaving no
+// Side quest in the deck that was never announced. A replenish batch,
+// which tops up the deck as the household swipes and never notifies,
+// stores each round's Side quests before the next round starts, so the
+// deck fills as it goes; those stored before a failure stay in the deck.
+// See store.
 func (s *Service) runBatch(ctx context.Context, p BatchJob, publish func(batchResult) error) (batchResult, error) {
 	batch, err := db.New(s.st.Reader).GetSuggestionBatch(ctx, p.BatchID)
 	if err != nil {
@@ -78,7 +84,7 @@ func (s *Service) runBatch(ctx context.Context, p BatchJob, publish func(batchRe
 	if n == 0 {
 		return batchResult{}, fmt.Errorf("suggest: batch %s is not pending", p.BatchID)
 	}
-	res, err := s.fill(ctx, p.BatchID, batch.Trigger != TriggerReplenish, publish)
+	res, err := s.fill(ctx, batch.Trigger, p, publish)
 	if err == nil {
 		if err := s.finishBatch(ctx, p.BatchID, BatchDone); err != nil {
 			return batchResult{}, err
@@ -98,10 +104,12 @@ func (s *Service) runBatch(ctx context.Context, p BatchJob, publish func(batchRe
 	return batchResult{}, errors.Join(err, finishErr)
 }
 
-// fill does runBatch's work, returning a *jobs.Failure for a failure the
-// household may see. When notifies is true, store emits new_suggestions
-// with the Side quests it adds.
-func (s *Service) fill(ctx context.Context, batchID string, notifies bool, publish func(batchResult) error) (batchResult, error) {
+// fill does runBatch's work for job of a batch with trigger, returning a
+// *jobs.Failure for a failure the household may see. A batch with a notice
+// is stored once, when its last round is done; any other batch round by
+// round.
+func (s *Service) fill(ctx context.Context, trigger string, job BatchJob, publish func(batchResult) error) (batchResult, error) {
+	batchID := job.BatchID
 	res := batchResult{BatchID: batchID, Stage: StageSuggesting, SuggestionIDs: []string{}}
 	if err := publish(res); err != nil {
 		return batchResult{}, err
@@ -110,6 +118,10 @@ func (s *Service) fill(ctx context.Context, batchID string, notifies bool, publi
 	if err != nil {
 		return batchResult{}, err
 	}
+	note, err := noticeOf(trigger, job, in.Context.Today)
+	if err != nil {
+		return batchResult{}, err
+	}
 	out, err := s.suggest(ctx, batchID, in)
 	if err != nil {
 		if ctx.Err() != nil {
@@ -122,34 +134,81 @@ func (s *Service) fill(ctx context.Context, batchID string, notifies bool, publi
 	if err := publish(res); err != nil {
 		return batchResult{}, err
 	}
-	var week *store.LocalDate
-	if notifies {
-		w := household.WeekStart(in.Context.Today)
-		week = &w
-	}
 	left := out.Suggestions
-	for len(res.SuggestionIDs) < BatchSize && len(left) > 0 {
-		round := left[:min(BatchSize-len(res.SuggestionIDs), len(left))]
+	// held are a notifying batch's grounded Side quests, not yet stored.
+	var held []groundedQuest
+	for len(res.SuggestionIDs)+len(held) < BatchSize && len(left) > 0 {
+		round := left[:min(BatchSize-len(res.SuggestionIDs)-len(held), len(left))]
 		left = left[len(round):]
 		grounded, err := s.ground(ctx, round)
 		if err != nil {
 			return batchResult{}, err
 		}
 		res.Dropped += len(round) - len(grounded)
-		ids, err := s.store(ctx, batchID, week, grounded)
-		if err != nil {
+		if note != nil {
+			held = append(held, grounded...)
+		} else if err := s.storeInto(ctx, &res, nil, grounded); err != nil {
 			return batchResult{}, err
 		}
-		res.Dropped += len(grounded) - len(ids)
-		res.SuggestionIDs = append(res.SuggestionIDs, ids...)
 		if err := publish(res); err != nil {
 			return batchResult{}, err
 		}
 	}
+	if note != nil {
+		if err := s.storeInto(ctx, &res, note, held); err != nil {
+			return batchResult{}, err
+		}
+	}
 	res.Stage = StageDone
 	return res, nil
 }
 
+// storeInto stores quests for res's batch with note (see store) and adds
+// the suggestions stored, and the Side quests left out, to res.
+func (s *Service) storeInto(ctx context.Context, res *batchResult, note *notice, quests []groundedQuest) error {
+	ids, err := s.store(ctx, res.BatchID, note, quests)
+	if err != nil {
+		return err
+	}
+	res.Dropped += len(quests) - len(ids)
+	res.SuggestionIDs = append(res.SuggestionIDs, ids...)
+	return nil
+}
+
+// notice is how a batch's new_suggestions is keyed: by id, as
+// new_suggestions:{id}, for the week starting on week.
+type notice struct {
+	id   string
+	week store.LocalDate
+}
+
+// noticeOf returns the notice of job, run for a batch with trigger on
+// household-local date today, or nil for a replenish batch, which never
+// notifies. A scheduled batch is keyed by the ISO week it was scheduled
+// for ("2026-W41"), not the week it happens to run in, so each week's
+// schedule tells the household once even when its batch runs late; a
+// manual batch is keyed by its batch id, for the week of today. A job
+// whose week does not fit its trigger is an error.
+func noticeOf(trigger string, job BatchJob, today store.LocalDate) (*notice, error) {
+	if trigger != TriggerScheduled {
+		if job.Week != "" {
+			return nil, fmt.Errorf("suggest: %s batch %s has a week", trigger, job.BatchID)
+		}
+		if trigger == TriggerReplenish {
+			return nil, nil
+		}
+		return &notice{id: job.BatchID, week: household.WeekStart(today)}, nil
+	}
+	week, err := store.ParseLocalDate(job.Week)
+	if err != nil {
+		return nil, fmt.Errorf("suggest: scheduled batch %s week: %w", job.BatchID, err)
+	}
+	if household.WeekStart(week) != week {
+		return nil, fmt.Errorf("suggest: scheduled batch %s week %s is not a Monday", job.BatchID, week)
+	}
+	return &notice{id: household.PeriodKey(week), week: week}, nil
+}
+
 // request reads what suggest_side_quests is given: the household context,
 // the deck's names and every passed name (ListPassedNames). A household with no Regulars
 // fails with ReasonNoRegulars.
@@ -224,13 +283,14 @@ func (s *Service) ground(ctx context.Context, quests []skills.SideQuest) ([]grou
 // or archived. Passed names are read in the same transaction as the store,
 // so a Pass or drop committed before it is never missed.
 //
-// When week is not nil and a Side quest is stored, new_suggestions:{batchID}
-// naming those stored is emitted in the same transaction, for the week
-// starting week, so the Side quests and their notification commit together
+// When note is not nil and a Side quest is stored, new_suggestions keyed by
+// note, counting and naming those stored, is emitted in the same
+// transaction, so the Side quests and their notification commit together
 // or not at all: a notification that cannot be stored fails the store, and
-// the batch with it. The event's key makes the first round to store Side
-// quests the one that notifies; a later round's emit stores nothing.
-func (s *Service) store(ctx context.Context, batchID string, week *store.LocalDate, quests []groundedQuest) ([]string, error) {
+// the batch with it. A notifying batch calls store once, with all its
+// Side quests, so the count is the batch's (see runBatch). This is the
+// only place new_suggestions is emitted.
+func (s *Service) store(ctx context.Context, batchID string, note *notice, quests []groundedQuest) ([]string, error) {
 	ids := []string{}
 	var stored []string
 	err := s.st.WithTx(ctx, func(tx *sql.Tx) error {
@@ -298,15 +358,15 @@ func (s *Service) store(ctx context.Context, batchID string, week *store.LocalDa
 			ids = append(ids, id)
 			stored = append(stored, d.Name)
 		}
-		if week == nil || len(stored) == 0 {
+		if note == nil || len(stored) == 0 {
 			return nil
 		}
-		return s.notifier.EmitTx(ctx, tx, chat.NewSuggestions(batchID, *week, stored))
+		return s.notifier.EmitTx(ctx, tx, chat.NewSuggestions(note.id, note.week, stored))
 	})
 	if err != nil {
 		return nil, err
 	}
-	if week != nil && len(stored) > 0 {
+	if note != nil && len(stored) > 0 {
 		s.notifier.Wake()
 	}
 	return ids, nil

internal/suggest/schedule.go

@@ -0,0 +1,36 @@
+package suggest
+
+import (
+	"context"
+	"time"
+
+	"code.cubecraftlabs.com/Maniacs_Mansion/Meal-Quest/internal/household"
+	"code.cubecraftlabs.com/Maniacs_Mansion/Meal-Quest/internal/scheduler"
+)
+
+// WeeklySchedule is the suggestions_weekly schedule: at the household's
+// suggestions_dow and suggestions_time it starts a scheduled batch for the
+// ISO week of the fire date, its period scheduler.Weekly keys
+// suggestions_weekly:YYYY-Www. The batch tells the household of the Side
+// quests it adds with new_suggestions:YYYY-Www for that week, once; a batch
+// that fails before adding any, as when the AI fails, sends nothing. When a
+// batch is already pending or running the period starts none, and says so.
+func (s *Service) WeeklySchedule() scheduler.Schedule {
+	return scheduler.Schedule{
+		Name: scheduler.SuggestionsWeekly,
+		Rule: func(settings household.Settings) scheduler.Rule {
+			return scheduler.Weekly(time.Weekday(settings.SuggestionsDow), settings.SuggestionsTime)
+		},
+		Run: func(ctx context.Context, p scheduler.Period) error {
+			b, err := s.StartScheduledBatch(ctx, p.Week)
+			if err != nil {
+				return err
+			}
+			if !b.Created {
+				s.logger.WarnContext(ctx, "scheduled suggestion batch not started: a batch is already under way",
+					"period", p.Key, "batch_id", b.ID, "trigger", b.Trigger)
+			}
+			return nil
+		},
+	}
+}

internal/suggest/suggest.go

@@ -19,6 +19,7 @@ import (
 	"code.cubecraftlabs.com/Maniacs_Mansion/Meal-Quest/internal/ai/skills"
 	"code.cubecraftlabs.com/Maniacs_Mansion/Meal-Quest/internal/db"
 	"code.cubecraftlabs.com/Maniacs_Mansion/Meal-Quest/internal/dinners"
+	"code.cubecraftlabs.com/Maniacs_Mansion/Meal-Quest/internal/household"
 	"code.cubecraftlabs.com/Maniacs_Mansion/Meal-Quest/internal/jobs"
 	"code.cubecraftlabs.com/Maniacs_Mansion/Meal-Quest/internal/notify"
 	"code.cubecraftlabs.com/Maniacs_Mansion/Meal-Quest/internal/recipes"
@@ -154,17 +155,35 @@ type Batch struct {
 	Created bool `json:"-"`
 }
 
-// StartBatch creates a batch with trigger and enqueues the
+// StartBatch creates a manual or replenish batch and enqueues the
 // KindSuggestionBatch job that runs it, unless a batch is already pending
-// or running, which it returns instead with Created false.
+// or running, which it returns instead with Created false. A scheduled
+// batch is started by StartScheduledBatch.
 func (s *Service) StartBatch(ctx context.Context, trigger string) (Batch, error) {
+	if trigger == TriggerScheduled {
+		return Batch{}, errors.New("suggest: a scheduled batch is started by StartScheduledBatch")
+	}
 	if !slices.Contains(triggers, trigger) {
 		return Batch{}, fmt.Errorf("suggest: unknown batch trigger %q", trigger)
 	}
+	return s.start(ctx, trigger, "")
+}
+
+// StartScheduledBatch is StartBatch for a scheduled batch for the ISO week
+// starting on Monday week, which keys its new_suggestions (see noticeOf).
+func (s *Service) StartScheduledBatch(ctx context.Context, week store.LocalDate) (Batch, error) {
+	if week.IsZero() || household.WeekStart(week) != week {
+		return Batch{}, fmt.Errorf("suggest: scheduled batch week %s is not a Monday", week)
+	}
+	return s.start(ctx, TriggerScheduled, week.String())
+}
+
+// start does StartBatch's work for a batch with trigger and the job week.
+func (s *Service) start(ctx context.Context, trigger, week string) (Batch, error) {
 	var b Batch
 	err := s.st.WithTx(ctx, func(tx *sql.Tx) error {
 		var err error
-		b, err = s.startBatch(ctx, tx, trigger)
+		b, err = s.startBatch(ctx, tx, trigger, week)
 		return err
 	})
 	if err != nil {
@@ -228,10 +247,11 @@ func failedBatch(ctx context.Context, q *db.Queries) (*Batch, error) {
 	return &Batch{ID: cur.ID, JobID: jobID, Trigger: cur.Trigger, Status: cur.Status}, nil
 }
 
-// startBatch does StartBatch's work in tx: the batch and the job that runs
-// it are committed together, so a batch is never seen without its job. The
-// caller wakes the job runner once tx has committed and Created is true.
-func (s *Service) startBatch(ctx context.Context, tx *sql.Tx, trigger string) (Batch, error) {
+// startBatch does StartBatch's work in tx, for a batch with trigger whose
+// job has week (BatchJob.Week): the batch and the job that runs it are
+// committed together, so a batch is never seen without its job. The caller
+// wakes the job runner once tx has committed and Created is true.
+func (s *Service) startBatch(ctx context.Context, tx *sql.Tx, trigger, week string) (Batch, error) {
 	q := db.New(tx)
 	cur, err := s.activeBatch(ctx, q)
 	if err != nil {
@@ -250,7 +270,7 @@ func (s *Service) startBatch(ctx context.Context, tx *sql.Tx, trigger string) (B
 	if err != nil {
 		return Batch{}, fmt.Errorf("suggest: create batch: %w", err)
 	}
-	jobID, err := s.batchJob.EnqueueTx(ctx, tx, BatchJob{BatchID: row.ID})
+	jobID, err := s.batchJob.EnqueueTx(ctx, tx, BatchJob{BatchID: row.ID, Week: week})
 	if err != nil {
 		return Batch{}, fmt.Errorf("suggest: enqueue batch %s: %w", row.ID, err)
 	}
@@ -428,7 +448,7 @@ func (s *Service) decide(ctx context.Context, id, decision, dinnerStatus, update
 		if left >= ReplenishBelow {
 			return nil
 		}
-		b, err := s.startBatch(ctx, tx, TriggerReplenish)
+		b, err := s.startBatch(ctx, tx, TriggerReplenish, "")
 		if err != nil {
 			return err
 		}

internal/suggest/suggest_test.go

@@ -300,6 +300,45 @@ func (e *env) runBatch(t *testing.T) batchResult {
 	return res
 }
 
+// scheduledWeek is the week the tests' scheduled batches are for: the
+// Monday of 2026-W41.
+var scheduledWeek = mustDate("2026-10-05")
+
+func mustDate(s string) store.LocalDate {
+	d, err := store.ParseLocalDate(s)
+	if err != nil {
+		panic(err)
+	}
+	return d
+}
+
+func mustZone(t *testing.T, name string) *time.Location {
+	t.Helper()
+	loc, err := household.LoadZone(name)
+	if err != nil {
+		t.Fatal(err)
+	}
+	return loc
+}
+
+// startBatch starts a batch with trigger, a scheduled one for
+// scheduledWeek, and returns it with the key its new_suggestions has.
+func (e *env) startBatch(t *testing.T, trigger string) (suggest.Batch, string) {
+	t.Helper()
+	if trigger == suggest.TriggerScheduled {
+		b, err := e.svc.StartScheduledBatch(t.Context(), scheduledWeek)
+		if err != nil {
+			t.Fatal(err)
+		}
+		return b, "new_suggestions:2026-W41"
+	}
+	b, err := e.svc.StartBatch(t.Context(), trigger)
+	if err != nil {
+		t.Fatal(err)
+	}
+	return b, "new_suggestions:" + b.ID
+}
+
 func (e *env) batchStatus(t *testing.T, id string) string {
 	t.Helper()
 	var status string
@@ -404,8 +443,9 @@ func TestBatchStoresGroundedSideQuests(t *testing.T) {
 }
 
 // TestBatchEmitsNewSuggestions: a scheduled or manual batch that adds Side
-// quests stores new_suggestions keyed by the batch, naming them; a
-// replenish batch, or one that adds none, stores nothing.
+// quests stores new_suggestions counting and naming them, keyed by the
+// scheduled ISO week or the manual batch, for that week or the current
+// one; a replenish batch, or one that adds none, stores nothing.
 func TestBatchEmitsNewSuggestions(t *testing.T) {
 	for _, tc := range []struct {
 		trigger string
@@ -428,10 +468,7 @@ func TestBatchEmitsNewSuggestions(t *testing.T) {
 				sideQuest("Chicken Shawarma Bowls", "Spiced chicken, like your Chicken Fajitas.", "Chicken Fajitas"),
 				sideQuest("Chili Verde", "Your Beef Chili gone green.", "Beef Chili"),
 			)
-			b, err := e.svc.StartBatch(t.Context(), tc.trigger)
-			if err != nil {
-				t.Fatal(err)
-			}
+			b, key := e.startBatch(t, tc.trigger)
 			if job := e.waitJob(t, b.JobID); job.Status != jobs.StatusSucceeded {
 				t.Fatalf("batch job %s", job.Status)
 			}
@@ -446,64 +483,70 @@ func TestBatchEmitsNewSuggestions(t *testing.T) {
 				t.Fatalf("new_suggestions rows = %+v, want one on push", rows)
 			}
 			r := rows[0]
-			if r.Key != "new_suggestions:"+b.ID || r.Channel != notify.ChannelPush || r.Payload.URL != "/discover" ||
-				r.Payload.Title != "New Side quests are in" || r.Payload.Body != "Chicken Shawarma Bowls, Chili Verde" || r.Payload.Week == "" {
-				t.Errorf("new_suggestions row = %+v, want key new_suggestions:%s naming the Side quests, /discover", r, b.ID)
+			week := household.WeekStart(store.LocalDateOf(time.Now().In(mustZone(t, "America/New_York")))).String()
+			if tc.trigger == suggest.TriggerScheduled {
+				week = scheduledWeek.String()
+			}
+			if r.Key != key || r.Channel != notify.ChannelPush || r.Payload.URL != "/discover" ||
+				r.Payload.Title != "2 new Side quests are in" || r.Payload.Body != "Chicken Shawarma Bowls, Chili Verde" || r.Payload.Week != week {
+				t.Errorf("new_suggestions row = %+v, want key %s counting and naming the Side quests, week %s, /discover", r, key, week)
 			}
 		})
 	}
 }
 
-// TestBatchNotifiesAsSideQuestsAreStored: new_suggestions is stored with
-// the first round to add Side quests, so a batch whose later round fails
-// still tells the household of those in the deck, once; a later round that
-// succeeds adds no second notification; a replenish batch notifies nothing.
-func TestBatchNotifiesAsSideQuestsAreStored(t *testing.T) {
+// twoRoundBatch scripts a suggest_side_quests reply of BatchSize+1 Side
+// quests, Fajita Dish A to F, whose first round grounds all but B, so a
+// second round grounds F. Unless lastGrounds, normalize_recipe fails for
+// F.
+func (e *env) twoRoundBatch(t *testing.T, lastGrounds bool) {
+	t.Helper()
+	var quests []string
+	for i := range suggest.BatchSize + 1 {
+		name := fmt.Sprintf("Fajita Dish %c", 'A'+i)
+		e.search.Results[name] = nil
+		switch {
+		case i == suggest.BatchSize && !lastGrounds:
+			// A page is found, but no normalize_recipe reply is scripted
+			// for it, so the AI call grounding it fails.
+			slug := strings.ReplaceAll(strings.ToLower(name), " ", "-")
+			page, photo := "https://r.example/"+slug, "https://r.example/"+slug+".jpg"
+			e.search.Results[name] = []recipes.Candidate{{URL: page}}
+			e.pages.Bodies[page] = recipestest.RecipePage(name, "R Kitchen", photo, 35)
+			e.images.JPEGs[photo] = recipestest.JPEG(t, 400, 320, uint8(len(e.images.JPEGs)+1))
+		case i != 1:
+			e.recipe(t, name)
+		}
+		quests = append(quests, sideQuest(name, "Like your Chicken Fajitas.", "Chicken Fajitas"))
+	}
+	e.reply(quests...)
+}
+
+// TestBatchNotifiesOnceWithEverySideQuestStored: a batch that fills across
+// two grounding rounds stores new_suggestions once, when it completes,
+// counting and naming every Side quest it stored, not just the first
+// round's; a replenish batch notifies nothing.
+func TestBatchNotifiesOnceWithEverySideQuestStored(t *testing.T) {
 	for _, tc := range []struct {
-		trigger   string
-		lastFails bool
-		notify    bool
+		trigger string
+		notify  bool
 	}{
-		{suggest.TriggerScheduled, true, true},
-		{suggest.TriggerManual, true, true},
-		{suggest.TriggerManual, false, true},
-		{suggest.TriggerReplenish, true, false},
+		{suggest.TriggerScheduled, true},
+		{suggest.TriggerManual, true},
+		{suggest.TriggerReplenish, false},
 	} {
-		t.Run(fmt.Sprintf("%s last round fails %v", tc.trigger, tc.lastFails), func(t *testing.T) {
+		t.Run(tc.trigger, func(t *testing.T) {
 			e := newEnv(t)
 			e.start(t)
-			// The first round grounds BatchSize Side quests but B; the
-			// second round grounds F.
-			var quests []string
-			for i := range suggest.BatchSize + 1 {
-				name := fmt.Sprintf("Fajita Dish %c", 'A'+i)
-				e.search.Results[name] = nil
-				if i != 1 {
-					e.recipe(t, name)
-				}
-				quests = append(quests, sideQuest(name, "Like your Chicken Fajitas.", "Chicken Fajitas"))
-			}
-			e.reply(quests...)
-			if tc.lastFails {
-				if _, err := e.st.Writer.ExecContext(t.Context(), `CREATE TRIGGER refuse_f BEFORE INSERT ON dinners WHEN NEW.name = 'Fajita Dish F' BEGIN SELECT RAISE(ABORT, 'F refused'); END`); err != nil {
-					t.Fatal(err)
-				}
-			}
+			e.twoRoundBatch(t, true)
 
-			b, err := e.svc.StartBatch(t.Context(), tc.trigger)
-			if err != nil {
-				t.Fatal(err)
+			b, key := e.startBatch(t, tc.trigger)
+			if job := e.waitJob(t, b.JobID); job.Status != jobs.StatusSucceeded {
+				t.Fatalf("batch job %s, want succeeded", job.Status)
 			}
-			job := e.waitJob(t, b.JobID)
-			want, deck := jobs.StatusSucceeded, suggest.BatchSize
-			if tc.lastFails {
-				want, deck = jobs.StatusFailed, suggest.BatchSize-1
-			}
-			if job.Status != want {
-				t.Fatalf("batch job %s, want %s", job.Status, want)
-			}
-			if got := e.deckNames(t); len(got) != deck {
-				t.Fatalf("deck = %v, want %d Side quests", got, deck)
+			deck := e.deckNames(t)
+			if want := []string{"Fajita Dish A", "Fajita Dish C", "Fajita Dish D", "Fajita Dish E", "Fajita Dish F"}; !slices.Equal(deck, want) {
+				t.Fatalf("deck = %v, want %v", deck, want)
 			}
 			rows := notifytest.Rows(t, e.st, notify.NewSuggestions)
 			if !tc.notify {
@@ -512,9 +555,47 @@ func TestBatchNotifiesAsSideQuestsAreStored(t *testing.T) {
 				}
 				return
 			}
-			if len(rows) != 1 || rows[0].Key != "new_suggestions:"+b.ID ||
-				rows[0].Payload.Body != "Fajita Dish A, Fajita Dish C, Fajita Dish D + 1 more" {
-				t.Errorf("new_suggestions rows = %+v, want one keyed by batch %s naming the first round", rows, b.ID)
+			want := fmt.Sprintf("%d new Side quests are in", len(deck))
+			if len(rows) != 1 || rows[0].Key != key || rows[0].Payload.Title != want ||
+				rows[0].Payload.Body != "Fajita Dish A, Fajita Dish C, Fajita Dish D + 2 more" {
+				t.Errorf("new_suggestions rows = %+v, want one keyed %s titled %q", rows, key, want)
+			}
+		})
+	}
+}
+
+// TestBatchFailingInALaterRoundStoresNothing: when normalize_recipe fails
+// in a batch's second grounding round, a scheduled or manual batch stores
+// none of the first round's Side quests and sends no new_suggestions, so
+// no Side quest is left in the deck unannounced; a replenish batch, which
+// never notifies, keeps the first round's in the deck.
+func TestBatchFailingInALaterRoundStoresNothing(t *testing.T) {
+	for _, tc := range []struct {
+		trigger string
+		deck    int
+	}{
+		{suggest.TriggerScheduled, 0},
+		{suggest.TriggerManual, 0},
+		{suggest.TriggerReplenish, suggest.BatchSize - 1},
+	} {
+		t.Run(tc.trigger, func(t *testing.T) {
+			e := newEnv(t)
+			e.start(t)
+			e.twoRoundBatch(t, false)
+
+			b, _ := e.startBatch(t, tc.trigger)
+			job := e.waitJob(t, b.JobID)
+			if job.Status != jobs.StatusFailed || job.Error == nil || *job.Error != suggest.ReasonAIFailed {
+				t.Fatalf("batch job %s, error %+v, want failed %s", job.Status, job.Error, suggest.ReasonAIFailed)
+			}
+			if got := e.batchStatus(t, b.ID); got != suggest.BatchFailed {
+				t.Errorf("batch status = %s, want failed", got)
+			}
+			if got := e.deckNames(t); len(got) != tc.deck {
+				t.Errorf("deck = %v, want %d Side quests", got, tc.deck)
+			}
+			if rows := notifytest.Rows(t, e.st, notify.NewSuggestions); len(rows) != 0 {
+				t.Errorf("new_suggestions rows = %+v, want none", rows)
 			}
 		})
 	}
@@ -1299,7 +1380,7 @@ func TestBatchJobFailsWhenRecordingItsEndFails(t *testing.T) {
 		t.Errorf("batch status = %s with its job failed, want failed", got)
 	}
 	// Its Side quests stay in the deck, and the household was told of them.
-	if rows := notifytest.Rows(t, e.st, notify.NewSuggestions); len(rows) != 1 || rows[0].Key != "new_suggestions:"+b.ID || rows[0].Payload.Body != "Chili Verde" {
+	if rows := notifytest.Rows(t, e.st, notify.NewSuggestions); len(rows) != 1 || rows[0].Key != "new_suggestions:"+b.ID || rows[0].Payload.Title != "1 new Side quest is in" || rows[0].Payload.Body != "Chili Verde" {
 		t.Errorf("new_suggestions rows of a batch whose end was not recorded = %+v, want one naming Chili Verde", rows)
 	}
 }
@@ -1474,6 +1555,47 @@ func TestStartBatchRefusesUnknownTrigger(t *testing.T) {
 	}
 }
 
+// TestScheduledBatchNeedsItsWeek: a scheduled batch is started only by
+// StartScheduledBatch, for a Monday.
+func TestScheduledBatchNeedsItsWeek(t *testing.T) {
+	e := newEnv(t)
+	if _, err := e.svc.StartBatch(t.Context(), suggest.TriggerScheduled); err == nil {
+		t.Error("StartBatch started a scheduled batch with no week")
+	}
+	for _, week := range []store.LocalDate{{}, mustDate("2026-10-06")} {
+		if _, err := e.svc.StartScheduledBatch(t.Context(), week); err == nil {
+			t.Errorf("StartScheduledBatch(%q) accepted a week that is not a Monday", week)
+		}
+	}
+	if n := e.activeBatches(t); n != 0 {
+		t.Errorf("%d active batches, want none", n)
+	}
+}
+
+// TestScheduledBatchKeyedByItsWeekWhenRunLate: a scheduled batch's
+// new_suggestions is keyed by the week it was scheduled for, not the week
+// it runs in, so a batch caught up after its week ended still tells the
+// household of that week's Side quests and leaves the next week's key free.
+func TestScheduledBatchKeyedByItsWeekWhenRunLate(t *testing.T) {
+	e := newEnv(t)
+	e.start(t)
+	e.recipe(t, "Chili Verde")
+	e.reply(sideQuest("Chili Verde", "Your Beef Chili gone green.", "Beef Chili"))
+	// A week long past: the batch runs today, weeks later.
+	b, err := e.svc.StartScheduledBatch(t.Context(), mustDate("2026-01-05"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	if job := e.waitJob(t, b.JobID); job.Status != jobs.StatusSucceeded {
+		t.Fatalf("batch job %s", job.Status)
+	}
+	rows := notifytest.Rows(t, e.st, notify.NewSuggestions)
+	if len(rows) != 1 || rows[0].Key != "new_suggestions:2026-W02" || rows[0].Payload.Week != "2026-01-05" ||
+		rows[0].Payload.Title != "1 new Side quest is in" {
+		t.Errorf("new_suggestions rows = %+v, want one keyed new_suggestions:2026-W02 for 2026-01-05", rows)
+	}
+}
+
 // noCalendar is a plan.Calendar for tests that never push.
 type noCalendar struct{}
 

migrations/queries/schedule_runs.sql

@@ -0,0 +1,6 @@
+-- name: ClaimScheduleRun :execrows
+-- Claims one period of a schedule. The insert is the claim: 1 row means
+-- this caller owns the period, 0 rows means it was already claimed.
+INSERT INTO schedule_runs (id, schedule, period_key, created_at)
+VALUES (?, ?, ?, ?)
+ON CONFLICT (schedule, period_key) DO NOTHING;