Every HTTP client passed an integer timeout to requests, applying the same
value to BOTH connect and read phases. A slow Jellyfin /Items page or qBit
/sync/maindata blew through the 10s read budget → ReadTimeoutError. Split
into a (connect=5s, read=60s default) tuple via shared http_timeout() helper.
The media index build worker uses a 180s read floor. Existing services with
low timeout_seconds benefit from bumping to 60+.
Route all prometheus widget queries through Grafana /api/ds/query instead of
direct Prom HTTP. PrometheusConfig: drop base_url, add grafana_url +
datasource_uid; secret grafana_api_key (required). PrometheusWidgetSource →
MetricSource with _gateway_query POST method. normalize_grafana_frames
recovered from 65bae95 + shared _dedup_label helper. Gateway-path status
check. Startup old-config validation. CHANGELOG migration note. All adapter
tests rewritten for POST /api/ds/query + Grafana frames mock. Backend: 331
pytest pass, ruff clean. Frontend: build green (unchanged in S1).
Register MediaIndex as a harness concern with ALTER TABLE migration to add
service_id column (idempotent). Scope replace_items by service_id (FIXES latent
global-clear bug where building for one Jellyfin wiped another's rows). Scope
query by service_id (empty-string = all rows, backward-compat). Thread
service_id through build_media_index + worker + query_media router. New
regression test proves scoped replace preserves other services' rows.
Backend: 321 pytest pass, ruff clean. Frontend: build green.
The build endpoint had Depends(get_jellyfin_client) and Depends(get_user_id)
which executed BEFORE the function body. If Jellyfin was unreachable, these
raised HTTPException(503), the function never ran, and the worker was never
started. The frontend mutation had no onError handler, so the failure was
completely silent — the button briefly showed 'Building...' then reverted
to 'Build index' with zero feedback.
Backend fix: removed the Jellyfin dependencies from post_build_index.
The worker subprocess resolves its own Jellyfin connection via
_resolve_jellyfin(service_id) — the endpoint just needs to start the
worker process. The libraries count starts at 0 and gets updated by
the worker once it connects.
Frontend fix: added onError to useBuildIndex that invalidates the status
query (so the UI reflects the non-building state). MediaTab now displays
the build error inline: 'Build failed: <message>' next to the button.
283 backend tests pass (updated build test for new no-dependency flow);
128 frontend tests pass; ruff/eslint clean.
Four fixes:
1. Settings > Services tab: master/detail layout. Replaced the vertical
stack of ServiceConfigEditor cards with a SelectionRailCard (list on
left) + SectionCard (details on right) — same pattern as Machines
and SSH Keys tabs. Click a service in the rail to edit it.
2. Service Overview widget restriction. When adding widgets on a
service's Overview, the dialog now only shows built-in widgets +
widgets for THAT service type (not all services). Dashboard/named
dashboards (no serviceId) still see all.
3. Reorder buttons fixed. The swap-sort_order approach was a no-op when
both items had sort_order=0 (the default). Now moveInstance
renumbers ALL items by their new index position (i * 10) after the
swap, guaranteeing values change. References use updateRef, owned
widgets use saveWidget, both sequential.
4. Media index worker resolution. The subprocess worker called
get_jellyfin_client/get_user_id (FastAPI request dependencies) which
don't work outside request context. Now accepts a service_id
parameter (passed from post_build_index) and resolves the Jellyfin
instance directly from the settings store via _resolve_jellyfin.
Raises RuntimeError (not HTTPException) on failure.
283 backend tests pass; 128 frontend tests pass; ruff/eslint clean.
Extract the duplicated _resolve_service_record helper (identical in
routers/monitoring.py and routers/authentik_users.py) into a shared
services/service_resolution.py module. Both routers now import
resolve_service_record from the shared module.
The authentik router previously hardcoded service_type='authentik' in
its local copy; the shared helper takes service_type as a param (same
as monitoring's did).
Tests updated: test_api.py patches now target the correct module paths
(resolve_service_record on the monitoring module where it's imported,
build_service_record on the service_resolution module).
283 backend tests pass; ruff clean.
Three reusable-widget follow-up fixes:
1. Reference sort_order independently reorderable. Reordering a
referenced widget now updates the widget_references.sort_order (per-
dashboard), not the shared widget instance sort_order. New backend
update_widget_reference method + PUT /api/widgets/references/{id}
endpoint. Frontend moveInstance checks _ref_id to choose the right
mutation (updateRef for references, saveWidget for owned).
2. Detach preserves service_id. detach_widget_reference now copies the
original widget's service_id into the clone, so service-bound widgets
(Grafana chart, Jellyfin activity) continue to render after detach.
3. Named dashboards support widget references. NamedDashboardPage
fetches useWidgetReferences('named:<slug>') and renders them via
WidgetInstanceCard alongside pinned links. 'Edit widgets' button
opens WidgetConfigDialog with dashboardScope='named:<slug>'.
Also: removed useMemo on combinedWidgets in WidgetConfigDialog to fix
a react-hooks/preserve-manual-memoization lint error (the React Compiler
ESLint plugin couldn't verify the spread+sort memoization).
283 backend tests pass (+1 update_reference test); 128 frontend tests
pass; ruff clean; 0 lint errors.
Widgets configured on one dashboard (e.g., a Grafana service's Overview)
can now be live-referenced on other dashboards. Editing the widget config
updates it everywhere it's referenced. References can be detached into
independent clones.
Backend: new widget_references table (dashboard_scope, widget_id,
sort_order) with ON DELETE CASCADE. CRUD methods + 4 endpoints:
GET/POST /api/widgets/references, DELETE /api/widgets/references/{id},
POST /api/widgets/references/{id}/detach (clones the widget into a
standalone instance, then removes the reference).
Frontend: WidgetConfigDialog gains a dashboardScope prop. When set
(the main Dashboard passes 'main'), the dialog shows:
- Owned + referenced widgets in a combined list, with a link badge on
references.
- 'Add existing widget' picker: searchable list of ALL widget instances
not already on this dashboard. Click to create a reference.
- Detach button on references: clones the widget (service_id=NULL) and
removes the reference.
- Delete on a reference removes the REFERENCE (not the original widget).
Dashboard renders referenced widgets alongside owned widgets.
Detaching a service-bound widget clones it with service_id=NULL — the
clone may need re-binding to a service to render correctly. Named
dashboards don't pass dashboardScope yet (pinned-links-only); when they
gain widget support, the backend already handles any scope string.
282 backend tests pass (+2 reference lifecycle); 127 frontend tests
pass; ruff/eslint/tsc/vite all green.
Four coupled changes to the services-as-hub IA:
1. Nav entries use service TYPE names (Jellyfin, SSH Tasks, Alertmanager,
Grafana, Prometheus, Backups, Authentik) instead of conceptual names
(Media, Files, Actions, Alerts, Users). ssh_tasks collapses to one
entry ('SSH Tasks') instead of two. The content tabs inside each
service page surface the concepts (Files, Actions).
2. Service page gains a two-level tab structure when multiple enabled
instances of the same type exist: instance tabs on top ([Main Jellyfin]
[Backup Jellyfin]), content tabs below ([Overview] [Media] [Requests]
[Widgets]). Clicking an instance tab navigates to the sibling's route.
Single instance: no instance tabs. Replaces the dropdown switcher.
3. Config tab (connection fields, secrets, enable/disable, delete) moves
from the service page to Settings > Services tab. The service page
becomes a PURE operational view (Overview + content tabs + Widgets) --
no save/delete/config state. Settings gains a 4th tab 'Services' with
ServiceConfigEditor per instance (schema-driven config fields, secrets
with leave-blank-to-keep semantics, ConfirmDialog on delete).
4. Overview tab is now a configurable widget grid per service instance.
Each instance manages its own set of widgets on its Overview. Backend
widget list endpoints gain ?service_id= and ?scope= (dashboard|service)
filter params; the main Dashboard uses scope=dashboard to exclude
service-scoped widgets. The OverviewTab reuses WidgetInstanceCard +
WidgetConfigDialog. Empty state CTA for instances with no widgets.
All service-tab stubs are replaced; stubs.tsx deleted.
272 backend tests pass (+1 widget filter); 121 frontend tests pass (+3
instance-tabs + OverviewTab); lint/build green both sides.
Slice 1 of jellyfin-service-registry. Jellyfin is configured exclusively
via the service registry now; the machine-level media_root/path_prefix
fields were dead duplicates of the global config.
- services/settings_store.py: DEFAULT_SERVICES no longer includes
"jellyfin" (now ["monitoring", "files"]). Removed machine-level
media_root/path_prefix from _default_local_machine, _row_to_machine,
_normalize_machine_payload, _seed_local_machine, get_machine_config,
and upsert_machine. _default_local_machine no longer reads global
config, so the get_settings import is dropped.
- routers/settings.py: removed media_root/path_prefix from
MonitoringMachineInput (dead API input; store already ignored them).
The global config remote_media_root/path_prefix properties + path_utils.py
are unchanged (files.py and media_index still use them for Jellyfin->SSH
path resolution). ruff clean; 239 backend tests pass.
- Archive the completed observability-service-registry SDD change into
openspec/changes/archive/ (delivered across 5 slices; only
jellyfin-service-registry remains active).
- Stop ignoring .pi-map.md / .pi-map.index.md so the navigation maps are
versioned alongside the code, and add the regenerated map pairs repo-wide.
Slice 4 of observability-service-registry. Removes the shared-file
Prometheus bridge; external Prometheus now consumes node-exporter targets
via http_sd_configs against GET /api/monitoring/prometheus-targets.
- services/targets.py: removed write_prometheus_targets() (the file
writer) and its json/Path/get_settings imports; updated module docstring.
build_node_exporter_targets() is unchanged and still powers the HTTP
endpoint.
- main.py: removed the startup write_prometheus_targets call.
- routers/settings.py: removed the _write_prometheus_targets helper and
its three post machine create/update/delete call sites + the now-unused
targets import.
- config.py: removed the prometheus_file_sd_dir field.
- docker-compose.yml / docker-compose.dev.yml: removed the
PROMETHEUS_FILE_SD_DIR backend env var.
- tests: removed TestWritePrometheusTargets + the write_prometheus_targets
import in test_targets.py; rewrote the two TestSettingsMachines tests to
assert machines appear/disappear from /api/monitoring/prometheus-targets
(the surviving HTTP path) instead of the removed file-writer side effect.
ruff clean; 239 backend tests pass.
Slice 2 of observability-service-registry. The monitoring router resolves
observability components from the service registry instead of env vars.
- routers/monitoring.py: removed _alertmanager_client/_webhook_client env
readers + the get_settings import. Added _resolve_service_record(store,
service_type, service_id?) -> ServiceRecord|None (requested instance with
type+enabled checks, else first enabled instance), plus _base_url/_timeout/
_auth_headers (Bearer from api_key)/_status_response helpers.
- /alerts + /alertmanager-status now take service_id? + Depends(store),
resolve an alertmanager service, return graceful not-configured/
unreachable payloads including service_id/name; status down-branches now
include peers:[] + error (fixes prior type drift).
- NEW /grafana-status (probes /api/health) and /prometheus-status (probes
/-/healthy then /api/v1/status/buildinfo) returning
{up,version,service_id,name,error}.
- Webhook receiver is now log-only (dropped the outbound
ALERTMANAGER_WEBHOOK_URL forward).
- tests: rewrote TestAlertmanager + TestAlertmanagerWebhook to mock
_resolve_service_record/requests.get (not-configured via empty registry);
added TestGrafanaStatus/TestPrometheusStatus and a TestResolveServiceRecord
unit class covering service_id match/type-mismatch/disabled and first-
enabled/none-enabled paths.
Orphaned config fields alertmanager_url/alertmanager_webhook_url and the
env-var removal land in Slice 5. ruff clean; 240 backend tests pass.
Reviewed fresh-context (read-only): no blockers.
Manage now connects to existing Grafana/Prometheus/Alertmanager instances
and never deploys its own stack.
- docker-compose.yml / docker-compose.dev.yml: removed prometheus, loki,
alloy, grafana, alertmanager, node-exporter services, the monitoring
network, and observability named volumes; they now ship only backend +
frontend. Dev frontend now joins the web network so the Vite dev proxy
can reach the backend.
- backend: alertmanager_url default is now empty; /api/monitoring/alerts
and /alertmanager-status return graceful "not configured" responses
when ALERTMANAGER_URL is unset. Added not-configured tests.
- docker-compose.observability.yml: kept as the optional standalone
example; header clarifies Manage does not deploy it.
- Removed orphaned combined monitoring/prometheus/prometheus.yml
(standalone stack uses prometheus.standalone.yml).
- Docs (README, REQUIREMENTS decision log, monitoring-logging-design,
observability-runbooks, context.md, MIGRATION_PLAN, frontend/README,
CHANGELOG) updated to the connect-to-existing model.
VITE_GRAFANA_URL / VITE_PROMETHEUS_URL remain as optional frontend
deep-link overrides. .env.example still needs a manual update (safety
policy blocks assistant edits): set ALERTMANAGER_URL empty/optional and
move standalone-only vars out of the root file.
- Add shared task_runner.run_saved_task helper used by routers/tasks.py and
widgets/sources.py SshTaskWidgetSource.
- Saved tasks now target ssh_tasks service instances via default_service_id;
the legacy default_machine_id and saved_task_runs are removed.
- Actions page lists ssh_tasks services for default and run-time selection.
- Update types, API client, hooks, tests, docs, and changelog.
Backend tests: 222 passed. Frontend lint/build/test: clean (71 passed).
Follow-up #1 to the service-registry change. Jellyfin/Jellyseerr now resolve
from the service registry, so the machine-level app fields are dead config.
- dependencies.py: drop dead _jellyseerr_client_for; simplify _resolve_machine
to SSH-only.
- settings_store.py + routers/settings.py: remove jellyfin_*/jellyseerr_* from
machine default config, get_machine_config, normalization, row mappers, and
MachineInput.
- frontend types + Settings.tsx: drop the fields and the Jellyfin/Jellyseerr
form sections + service options.
- Update frontend test fixtures.
Existing DB rows may still carry these keys in config_json; they are inert and
drop on the next machine save. Verification: backend ruff clean, pytest 222;
frontend lint 0 errors, build success, 70 tests.
PR 2 of 4 for the runtime service registry change.
- dashboard_widgets gains service_id + widget_kind columns (legacy
addon_id/widget_type kept but unused).
- Source adapters take (service: ServiceRecord | None, widget_kind, config).
SERVICE_ADAPTERS keyed by service_type; BUILTIN_ADAPTERS for backups/static.
- Backups and static stay as service-less built-ins (service_id nullable),
exposed via GET /api/widgets/builtin.
- SSH task adapter resolves the task + instance, runs over SSH, and appends a
service_task_runs history row on success/failure/timeout/error.
- Retire widgets/registry.py; widget metadata now comes from the integrations
registry + widgets/builtin. Remove /api/widgets/types and /api/widgets/sources.
- Stop default widget seeding (fresh install = empty dashboard).
- Rewrite widget tests around the service-bound + built-in model (26 tests).
Backend-only breaking change; frontend is reconciled in Slice 3. Build/lint
stay green; pytest 222 passed.
The 2026-06-16/17 observability update externalised metrics to
Prometheus + node_exporter + Grafana, but the legacy Manage-side
SSH-scraping monitor was never removed. It duplicated the new stack,
ran SSH df on every machine every 300s, and fed nothing (its UI was
deleted in e2ad731). This slice decommissions the duplication.
Removed (backend):
- services/monitoring_poller.py (MonitoringPoller) — entire file
- services/monitoring_actions.py (disk_space, run_machine_operation,
poll_machine_snapshot, build_machine_client) — entire file;
run_machine_operation had only 2 callers (the poller + /disk), both gone
- tests/test_monitoring_actions.py
- endpoints: POST /api/monitoring/poller, GET /machines/{id}/actions,
GET /disk (and the now-dead _resolve_machine helper)
- lifespan wiring (main.py), dependency wrapper (dependencies.py),
poller.start()/kick() from machine save (routers/settings.py)
- SettingsStore: monitoring_machine_actions table CREATE + 2 indexes +
record/list/prune_machine_actions methods; DROP TABLE IF EXISTS on
startup cleans existing DBs (user-approved)
- config knobs: monitoring_poll_interval_seconds,
monitoring_poll_initial_delay_seconds, monitoring_action_retention_days
- test_api.py: TestMonitoring._ensure_machine + test_disk
Kept (fits the new model): /machines, /prometheus-targets, /alerts,
/alertmanager-status, /alertmanager-webhook; the disk_usage JOB template
(manual on-demand, not monitoring); node_exporter_* machine fields
(they point Prometheus at the right host).
Gate: backend pytest 173 passed; ruff clean.
- backend/ uses proper Python src layout (src/media_library_viewer_api/)
with pyproject.toml, hatchling build, and PYTHONPATH=src convention
- frontend/ is a Vite + React + TypeScript SPA
- archive/ preserves the original Streamlit prototype for reference
- Cleaned up root to only contain docs, license, and subproject dirs
- Updated README for the new dual-subproject architecture
Developer
alex