geniusrun/docs/superpowers/specs/2026-07-27-workout-not-found-design.md

Stop retrying a permanently-missing Garmin workout forever — Design

Status: Approved, ready for implementation planning. Origin: user report -- /api/sync/status's workouts_pending count stuck at 1 forever, no matter how many times "Sync now" is clicked.

Root cause

fillPendingWorkouts/fillActivityWorkout (backend/internal/sync/service.go:339-371) treats every get_workout_by_id failure identically: log it, leave workout_raw_json NULL, and let ActivitiesMissingWorkout naturally retry it on the next sync. This is the right behavior for a transient failure (network blip, rate limiting), but wrong for a genuine HTTP 404 -- the activity's workout_id points at a workout that was deleted on Garmin's side after the activity was recorded (confirmed via the user's own backend log: a 404 for a specific workout_id). A 404 is definitive and will never succeed on retry, yet the current code retries it every single sync, forever, with the failure only ever logged server-side (log.Printf) -- never surfaced to the user beyond a perpetual, unexplained "1 more workout pending" nudge.

Fix

An activity can legitimately end up with workout_id set but workout_raw_json permanently NULL when the workout can't be found on Garmin -- workout_raw_json stays NULL forever (no fabricated data), but a new, separate marker records "confirmed missing" distinctly from "not yet fetched," so ActivitiesMissingWorkout/CountActivitiesMissingWorkout stop counting it.

internal/garmin/pyscript/wrapper.py

garminconnect already defines GarminConnectNotFoundError (a subclass of GarminConnectConnectionError) specifically for this case -- its own docstring says "so callers can now catch a missing resource specifically (e.g. deleting an already-deleted workout)." It's raised by connectapi() (used by get_workout_by_id and other lookups) whenever the underlying HTTP response is a 404.

dispatch()'s generic exception handler (used by every cmd, including the generic call dispatcher that get_workout_by_id goes through) gains a check: if the caught exception is a GarminConnectNotFoundError, the returned JSON error response gets an additional "not_found": true field alongside the existing "error" string. This is deliberately generic (any Garmin API call that 404s gets this marker, not just workouts) -- the decision about what to do with a not-found error stays method-specific in the Go/sync layer.

internal/garmin/client.go

• wireResponse gains NotFound bool json:"not_found,omitempty"``.
• A new sentinel: var ErrNotFound = errors.New("garmin: resource not found").
• roundTrip: when resp.Error != "" && resp.NotFound, the returned error wraps ErrNotFound (fmt.Errorf("%s: %s: %w", cmdName, resp.Error, ErrNotFound)), so callers can errors.Is(err, garmin.ErrNotFound) regardless of which method was called.

internal/store (schema + queries)

• schema.sql: activities gains workout_not_found_at TEXT (nullable), placed next to workout_raw_json, following the exact same style as the existing details_fetched_at/ splits_fetched_at timestamp markers.
• New method SetActivityWorkoutNotFound(ctx, userID, activityID) error, mirroring SetActivitySplitsFetched's exact shape (UPDATE activities SET workout_not_found_at = datetime('now'), updated_at = datetime('now') WHERE id = ? AND user_id = ?).
• ActivitiesMissingWorkout/CountActivitiesMissingWorkout both gain AND workout_not_found_at IS NULL in their WHERE clause.
• UpsertActivity's ON CONFLICT DO UPDATE already never touches details_fetched_at/ details_raw_json/splits_fetched_at/workout_raw_json -- workout_not_found_at follows the same pattern, so it's untouched by a later re-sync and the "confirmed missing" marker persists indefinitely once set.
• docs/DATABASE.md regenerated via go run ./cmd/dumpschema after the schema edit.

internal/sync/service.go

In fillPendingWorkouts's loop, when fillActivityWorkout returns an error:

• If errors.Is(err, garmin.ErrNotFound): call db.SetActivityWorkoutNotFound(ctx, s.userID, a.ID) and log at Info level ("workout not found on Garmin, likely deleted -- marking as such, will not retry"). No further retry.
• Otherwise: keep the existing behavior exactly as it is today (log a warning, leave workout_raw_json NULL, naturally retried on the next sync).

internal/garmin/mock

mock.Client's existing WorkoutErrByID map[int64]error field (added in a previous plan) needs no shape change -- a test simulating a 404 just sets the mapped error to fmt.Errorf("...: %w", garmin.ErrNotFound) (or garmin.ErrNotFound directly), which errors.Is picks up the same way a real wrapped error would.

Non-goals

• No change to fillActivityDetails/fillPendingActivityDetails's error handling (a details/splits fetch failure still aborts the batch and surfaces as a sync error, as it does today) -- this fix is scoped to workouts specifically, matching the confirmed root cause. If a details/splits 404 ever turns out to need the same treatment, that's a separate, unconfirmed problem to investigate on its own evidence.
• No frontend change -- workouts_pending simply stops counting a confirmed-missing workout, so the existing SyncModal/banner UI needs nothing new.
• No handling for the (very unlikely) edge case of Garmin re-linking a different workout_id to the same activity after workout_not_found_at was set for an earlier one -- accepted as an edge case not worth the added complexity.

Testing

• internal/garmin/pyscript/tests/test_wrapper.py: a new test mirroring the existing test_call_propagates_garminconnect_exception_as_error, but raising GarminConnectNotFoundError and asserting the response includes "not_found": true.
• internal/garmin (Go): a test asserting roundTrip wraps the error with ErrNotFound when the wire response sets not_found: true.
• internal/store: extend the existing ActivitiesMissingWorkout test (or add a new one) with a fixture that has workout_not_found_at set, asserting it's excluded from both ActivitiesMissingWorkout and CountActivitiesMissingWorkout.
• internal/sync: a test using mock.Client.WorkoutErrByID set to a garmin.ErrNotFound-wrapped error, asserting fillPendingWorkouts/FillPendingDetails completes without error, the activity's workout_not_found_at gets set, and it's excluded from a subsequent CountActivitiesMissingWorkout call (i.e., confirming it does NOT get retried on a second call, unlike today's forever-retry behavior for this exact scenario).