ipx/.claude/skills/ipx-prod-check/SKILL.md

name, description

name

description

ipx-prod-check

Look for problems in production ipx (the ipx-web, iPX and ipx-directory containers on Tower, Tower itself and the monitoring that watches them) from their logs in Loki, their traces in Tempo and the metrics in Prometheus, and file what is found as Gitea issues. Use when asked to check on production, look for issues or errors in ipx, see why something is slow or failing in production, review the logs, or investigate a report about the live site.

Checking production ipx

Production is three containers in the Arcane project ipx on Tower (/mnt/fast/arcane/projects/ipx/compose.yaml), one request passing through them in this order:

container	what	logs, in Loki	traces, in Tempo	metrics, in Prometheus
ipx-web	Caddy: serves the page, passes /api and /media to ipx (#203). Port 8099, the LAN and the tunnel	{container="ipx-web"}, JSON: a line a request (http.log.access), its own errors	service ipx-web: a span a request, passing traceparent on to ipx (#213)	job="ipx-web": Caddy's own caddy_http_*, from METRICS_BIND, published as 192.168.1.130:9190 (#371)
iPX	the daemon: API, scans, downloads, DeadAir	{container="iPX"}, JSON (IPX_LOG_FORMAT=json)	service ipx; passes traceparent to the directory	job="ipx", its /metrics
ipx-directory	the directory of public podcasts, directory.ipxreader.app (#204, #214)	{container="ipx-directory"}, JSON (DIRECTORY_LOG_FORMAT=json)	service ipx-directory: a span a request, joining ipx's trace when ipx called; a trace a refresher round (#213)	job="ipx-directory", its /metrics

Grafana's alert rules (grafana/alerts.py, #365) watch all of it between checks: start by asking which are firing (the monitoring section). Under them, Tower itself: job="node" (the node exporter: load, CPU, memory, disks). The Grafana dashboard iPX draws all of it in that order (grafana/dashboard.py). A trace without the production tag (resource.deployment.environment.name="production", on all three) is a daemon run by hand, not production. Since #213 a request is one trace across the services it touched: a Directory page is ipx-web, then ipx, then the directory, so start from the trace, not the logs. Alloy ships the logs; Loki, Tempo and Prometheus publish no query port to this machine, so use the script beside this file:

Q=.claude/skills/ipx-prod-check/query.py
$Q logs    '<LogQL>'           [since]   # lines, newest first (500 at most)
$Q metric  '<LogQL metric>'    [since]   # $range becomes `since`
$Q traces  '<TraceQL>'         [since]   # slowest first
$Q spans   '<TraceQL>'          [since]   # the matching spans, each with its own time, slowest first
$Q trace   <trace id>                    # one trace as a tree, with its log lines
$Q prom    '<PromQL>'          [since]   # Prometheus, now; $range becomes `since`
$Q flame   '<TraceQL>'  [since] [out.svg] # where the matching spans' time went, summed over up to
                                         # 100 traces: a flame graph, and out.svg.folded for speedscope

since is 30m, 1h, 24h, 7d. Default to 24h; widen it to see whether something is new. "Since the last deploy" is since that container started (each deploys on its own):

for c in iPX ipx-web ipx-directory; do S=$(docker inspect $c -f '{{.State.StartedAt}}'); echo "$c $(( ($(date +%s) - $(date -d "$S" +%s)) / 60 + 1 ))m"; done

Start every check by seeing that all three run, since Loki cannot see a container that is not:

docker ps -a --filter name=iPX --filter name=ipx-web --filter name=ipx-directory --format '{{.Names}} {{.Status}}'

All three should be Up, with no recent restart; they have no Docker healthcheck (#218). Then work through ipx (the checks below), the page, the directory, Tower and the monitoring itself, in that order.

ipx

What the log carries

Every line has timestamp, level, message and target; lines inside a span have span (the innermost: {"name":"feed","feed":"x"}). Loki's | json flattens it to span_name, span_feed. Lines inside a traced span, requests and scans, also carry trace_id and span_id: give the trace_id to $Q trace to see the whole request or scan. (From 2026-09-29 16:30 UTC; before that, lines had no trace id, the access log's time was ms, and each event was logged twice, words under ipx::scan and fields under ipx::io.)

What is routine is logged at debug, which goes to the admin page's Log view and not to Loki: a request that went well (#364), a scan's end (scan_done, #363), a feed not due (feed_skip), a feed starting (feed_start), and a feed that brought nothing (feed_done with every count 0). Count those from the metrics, not the log.

target	fields	what
ipx::http	method, path, route, status, duration_ms	a warning for each request answered 4xx or 5xx; route is the pattern, empty for an unrouted path. Every request, as info, until #364; now those that went well are only in ipx_http_request_duration_seconds
ipx::scan	ev and the event's own: feed, new, downloaded, failed, bytes, msg, url, feeds, reason, from, to; on a failure error.type and, from an HTTP error, http.response.status_code	the daemon's events, one line each, in words; warnings are feed and download failures
ipx::io	ev, feeds, pending, downloaded on the status reply	commands arriving on the socket (-> {...}) and the answer to ipx status
ipx::daemon, ipx::cli, ipx::catalogue, ipx::scan, ...	message, sometimes fields	start-up, shutdown, account and config messages, under the module that logs them; plain ipx before 2026-10-24, when main.rs was split (#356)
ipx::client	ev, user, feed, enclosure, deadair; msg, src, line on an error; ms on a timing	what the page reports from the browser (#190): page_error, rejection and player_error as warnings; play_start (ms from play to sound), stall (ms waiting mid-play) and list_shown (ms from a click to its list drawn, with feed the view and filter, #246) as info. Lines carry the trace of the action in the page: its API requests send traceparent, one trace id per tap or key

Events (ev): feed_start (debug), feed_done (new, downloaded, failed, torrents), feed_skip (not due, debug), feed_error (msg), feed_moved (from, to: a permanent redirect followed, the address updated), download_done (bytes), download_error (msg, url), torrent_deferred, reaped, reap_done, scan_done (feeds; debug), status (feeds, pending, downloaded: the answer to ipx status, only when someone runs it since #218), deadair_done (feed, enclosure, saved_s at the default pause), error (msg). DeadAir also logs DeadAir found the silences (cuts, saved_s, took_ms) under ipx::deadair, and warns DeadAir could not read a file when ffmpeg fails on one, which it then passes over until the next restart.

error.type is the HTTP status (404, 503) or one of dns, redirect_loop, timeout, tls, not_a_feed, site_message, connect, parse, other; Loki's | json names it error_type.

Filter on the text before | json where you can (|= "\"ev\":\"feed_error\""): it is much cheaper than parsing every line. Lines before 2026-09-29 14:00 UTC are text, not JSON, and | json | __error__="" drops them.

What Prometheus carries

The daemon's own state between requests, which no log line or trace holds (#194): ipx serves /metrics on [web] metrics_bind (production: 192.168.1.130:9187, LAN only), on a thread and port of its own, from memory alone, and Prometheus scrapes it every 15 s as job="ipx". Kept from 2026-10-11 15:15 UTC. Gauges, but for the counters (_total) and the one histogram:

metric	what
ipx_db_connections{state="in_use"|"idle"}, ipx_db_connections_max	the database pool: in use reaching the max means requests and scans wait for a connection
ipx_worker_queue	commands waiting for the sequential worker (scans, downloads, adds)
ipx_event_listeners	open pages and socket clients on the live-update stream
ipx_torrent_slots_free	of 2
ipx_deadair_playing	episodes being played through DeadAir's playlist
ipx_runtime_workers, ipx_runtime_tasks, ipx_runtime_queue	tokio: threads, live tasks, tasks waiting for a thread
ipx_process_resident_bytes, ipx_process_open_fds	memory and open files and sockets; mimalloc gives back what is freed, so memory follows use and a rise that never comes down is a leak (#224)
ipx_feed_fetches_total, ipx_feed_bytes_total	counters: feeds fetched (304s included) and the bytes of their bodies, read whole, so what scans allocate
ipx_uptime_seconds, ipx_build_info{version}	since the daemon started; what runs
ipx_worker_busy, ipx_worker_job_seconds{kind}	the worker's job now (scan, reap, download, status, reload) and for how long; ipx status still answers while it is stuck
ipx_jobs_total{kind}, ipx_job_seconds_total{kind}	counters: jobs finished, and their seconds all told
ipx_last_scan_age_seconds	since a scan last finished; absent until one has since starting
ipx_http_requests_in_flight, ipx_scan_feeds_fetching, ipx_downloads_active	work in progress
ipx_db_query_seconds{kind}, ipx_db_wait_seconds	histograms (#367): each statement's time once it has a connection, by its first word (select, insert, ...); and the wait for a connection, measured for hand-written SQL only, about half the queries. A wait growing while queries stay flat is the pool full
ipx_page_seconds{ev}	histogram: what the page timed, list_shown, play_start, stall, as the ipx::client lines say one by one (#369)
ipx_feed_failures_total{error_type}, ipx_new_items_total, ipx_downloads_total{result="saved"|"failed"}, ipx_deadair_saved_seconds_total	counters of what the events say (#368): feeds failed by kind, as the feed_error lines' error.type; new items; downloads; seconds of silence DeadAir found
ipx_http_request_duration_seconds{method,route,status}	histogram: each request answered and its seconds, buckets 5 ms to 10 s; _count is requests, _sum their time. route is the pattern, "" for an unrouted path; the event stream is left out (#364). From 2026-10-25
ipx_download_bytes_total	counter; rate() is the download speed
ipx_event_queue, ipx_events_missed_total	live updates unread by some listener; counter of those dropped for one that fell behind
ipx_deadair_backlog, ipx_deadair_busy_seconds	files of shows with DeadAir on not yet looked at (episodes everyone heard through stay counted); how long it has been on its file
ipx_feed_retries	feeds waiting to retry a failed name lookup
ipx_art_cache_bytes, ipx_art_cache_limit_bytes	artwork kept, as of the last sweep, and the most kept: kept sits at the limit by design
ipx_art_evicted_total, ipx_art_requests_total{result="hit"|"miss"}	counters: images the sweep threw out to stay under the limit, and artwork served from disk or fetched from the publisher; many evicted at every sweep with a falling hit rate means the limit is too small
ipx_signals_judged_total{user,key}, ipx_signals_hits_total{user,key}	counters: items Jev judged for Signals, and Signals they fit, by person and whose key paid (server or own, #278) (#254)
ipx_jev_requests_total{result="ok"|"key_refused"|"malformed"|"rate_limited"|"error"}, ipx_jev_seconds_total, ipx_jev_paced_seconds_total	counters: calls to TypeSafe's Jev, Signals' and icons', by how they ended; their seconds, the wait for a turn included; and that wait alone. malformed is ipx's request at fault (#257); rate_limited is a 429 from TypeSafe, whose limits are 80 calls and 100,000 input tokens a second, which ipx paces itself under at 60 and 80,000 (#277)
ipx_jev_tokens_total{kind="input"|"output",user,key}	counter: tokens TypeSafe counted, as its answers say, by person and whose key paid: what Signals, icon picks and key checks cost (#275, #278). A call shared on the server's key is split by each person's share of its questions
ipx_signals_last_pass_age_seconds, ipx_signals_keys_refused	since the Signals judge last finished a pass (at most 5 minutes when well; absent until one has); people whose key TypeSafe refused
directory_process_resident_bytes, directory_process_open_fds, directory_feed_fetches_total, directory_feed_bytes_total, directory_uptime_seconds, directory_build_info, directory_http_request_duration_seconds{method,route,status}, directory_feed_checks_total, directory_feed_checks_failed_total, directory_snapshots_total	the directory's (the last four #370), job="ipx-directory", from its metrics_bind (production 192.168.1.130:9188, LAN only)
ipx_feeds, ipx_downloads_pending, ipx_files_downloaded	the library, as the daemon counts it every 30 s

The checks

Run these, then read the lines behind whatever stands out. Most of the time is in the reading: a count says something happened, the lines and traces say why.

1. Warnings and errors, grouped. What went wrong, how often, and since when.

   $Q metric 'sum by (target, message) (count_over_time({container="iPX"} | json | __error__="" | level=~"WARN|ERROR" [$range]))'
   

   Feed and download failures name the feed in the message; group them in the next check instead.

2. Failing feeds and downloads.

   $Q metric 'sum by (feed, error_type) (count_over_time({container="iPX"} |= "\"ev\":\"feed_error\"" | json | __error__="" [$range]))' 7d
   $Q metric 'sum by (feed, error_type) (count_over_time({container="iPX"} |= "\"ev\":\"download_error\"" | json | __error__="" [$range]))' 7d
   

   Tell the publisher's problems from ipx's. A 404, 410, DNS failure or 503 from the feed's own server is the publisher (worth saying, since the feed may have moved; one issue for a feed that has been dead for days, not for a 503 once). A parse error on a feed that loads in a browser, a redirect loop ipx should follow, or the same failure on many feeds at once is ipx.

3. Server errors and slow requests.

   $Q prom 'sum by (method, route, status) (increase(ipx_http_request_duration_seconds_count{job="ipx",status=~"4..|5.."}[$range])) > 0' 24h
   $Q prom 'topk(10, histogram_quantile(0.95, sum by (method, route, le) (rate(ipx_http_request_duration_seconds_bucket{job="ipx",route!=""}[$range]))))' 24h
   $Q logs '{container="iPX"} |= "\"target\":\"ipx::http\""' 24h
   

   Any 5xx is worth a look; the log has each failed request's own line, with its trace id. 401s are people signing in, not a problem unless one address is hammering. The 95th percentile is read off the buckets, so it says which bucket, not the exact time: for a slow route, find its spans (check 5) and see which step holds the time. The histogram restarts at 0 with the daemon, which increase and rate allow for.

4. Is the worker keeping up? Scans should finish regularly, the queue should drain, and the daemon should not be restarting on its own.

   $Q prom 'sum(increase(ipx_jobs_total{job="ipx",kind="scan"}[$range]))' 6h
   $Q prom 'max_over_time(ipx_downloads_pending{job="ipx"}[$range])' 6h
   $Q logs '{container="iPX"} |= "daemon started"' 7d
   

   A daemon started not matched by a deploy (see git log and the image's build time) is a crash or an OOM kill: check docker inspect iPX -f '{{.State.OOMKilled}} {{.RestartCount}}' and the lines just before it. A pending count that only grows means downloads are not keeping up or not running.

   Since 2026-10-02 the daemon sleeps until the next feed is due (at most 10 minutes) instead of scanning every minute (#114), so expect a scan every few minutes, most of one or two feeds; none at all for over 10 minutes means the worker is stuck. A scan's end (scan_done) is logged at debug since #363, so it is in the page's log view and not in Loki. And pending is the real queue (#113): files a scan will download on its own. Back-catalogue files are held, listed but not counted, so it is usually 0 or a handful.

5. Slow and failed traces. Look at ipx's requests as spans, not traces: a trace started by a tap in the page also holds that page's live updates and any audio it streamed, minutes long, so a list of traces over a few seconds is nothing but those. spans lists each matching span with its own time and route, slowest first:

   R='resource.deployment.environment.name="production" && resource.service.name="ipx"'
   $Q spans "{$R && span.http.route != \"\" && span.http.route != \"/api/events\" && span.http.route !~ \"/media/.*\" && duration > 1s} | select(span.http.route)"
   $Q spans "{$R && name=\"step\" && duration > 500ms} | select(span.of)"
   $Q traces '{resource.deployment.environment.name="production" && status = error}'
   $Q trace <id>
   

   Tempo keeps about two weeks. A trace id from a log line that Tempo answers 404 for, or a search that finds only the last quarter of an hour, is Tempo not reading its stored blocks: see the monitoring section.

   Group the first by hour and by route (awk '{print substr($3,1,2), $NF}' | sort | uniq -c), then lay the hours beside Tower's (the Tower section): load over 100 with processes blocked and memory under 4 GB is the host stalling, and every request of those minutes is slow with it, ipx's queries included. On 2026-10-18 that was 120 of 130 requests over a second, up to 43 s. What is left at quiet hours is ipx's own: open its trace and see which step holds the time. A request far longer than its steps added up waited outside them, for a connection or a thread. The second line finds steps over half a second wherever they ran; a step slow only once, first thing after hours idle, is Postgres reading from disk what it no longer had in memory (All Subscriptions' entries_in, a second where it is 60 ms warm).

   Scans (scan) are long by nature, since they fetch many feeds one after another: look for one feed or fetch span holding most of it ($Q spans '{... && name="feed" && duration > 10s}'), or a download far slower than its size explains.

   Steps are step spans named by of (#187): an item list's entries_in (with the list asked for: filter, sort, dir, offset, limit, search), sanitize, deadair, count_in; the feed list's catalogue, feed_list and its queries; a reap's missing_files, prune_abandoned_failures, reap_candidates, clean_directory (once a day since #193, so a reap is about a second), trim_artwork, with what it removed on the reap span. DeadAir's work on a file is a deadair trace (feed, enclosure, cuts) with probe, silencedetect, frame_bytes, store_plans: seconds each, more is a file ffmpeg struggles with. A trace begun in the page (#190, #192) has the tap or key as its root span, tap or key (#376), from the moment it began to its last API answer, with the requests it made under it (API calls, the playlist and audio of a play, position saves, the page's own report): a tap long with its requests short is the page itself, drawing. The page reports the tap when the next one begins or it is hidden, so the root arrives last; a trace with no name in $Q traces is requests no tap began (a live update's, a timer's), or a tap not reported yet.

   Every command the worker runs from the socket, the CLI or the page is a command span (cmd, #373) with its job inside; a scan on the clock is a scan of its own. A list's sync is list_sync (list, listed) with steps upsert_listed, subscribe_readers, drop_unlisted, and an OPML upload opml_import (#372). The daemon's start is a start trace (#374), from open_database and assemble_config to daemon started: how long a deploy took to answer.

6. The daemon's state, from Prometheus. Pressure and leaks that requests and traces only hint at.

   $Q prom 'up{job="ipx"}'
   $Q prom 'max(max_over_time(ipx_db_connections{job="ipx",state="in_use"}[$range]))' 24h
   $Q prom 'max(max_over_time(ipx_db_connections_max{job="ipx"}[$range]))' 24h
   $Q prom 'max(max_over_time(ipx_worker_queue{job="ipx"}[$range]))' 24h
   $Q prom 'max(max_over_time(ipx_runtime_queue{job="ipx"}[$range]))' 24h
   $Q prom 'max(resets(ipx_uptime_seconds{job="ipx"}[$range]))' 24h
   $Q prom 'ipx_process_resident_bytes{job="ipx"} - ipx_process_resident_bytes{job="ipx"} offset $range' 24h
   $Q prom 'ipx_process_open_fds{job="ipx"} - ipx_process_open_fds{job="ipx"} offset $range' 24h
   $Q prom 'ipx_runtime_tasks{job="ipx"} - ipx_runtime_tasks{job="ipx"} offset $range' 24h
   $Q prom 'max(max_over_time(ipx_worker_job_seconds{job="ipx"}[$range])) by (kind)' 24h
   $Q prom 'max(max_over_time(ipx_last_scan_age_seconds{job="ipx"}[$range]))' 24h
   $Q prom 'sum by (kind) (increase(ipx_job_seconds_total{job="ipx"}[$range]))' 24h
   $Q prom 'max(max_over_time(ipx_http_requests_in_flight{job="ipx"}[$range]))' 24h
   $Q prom 'histogram_quantile(0.99, sum by (le) (rate(ipx_db_wait_seconds_bucket{job="ipx"}[$range])))' 24h
   $Q prom 'histogram_quantile(0.95, sum by (le, kind) (rate(ipx_db_query_seconds_bucket{job="ipx"}[$range])))' 24h
   $Q prom 'sum(increase(ipx_events_missed_total{job="ipx"}[$range]))' 24h
   $Q prom 'max(max_over_time(ipx_deadair_busy_seconds{job="ipx"}[$range]))' 24h
   $Q prom 'sum by (result) (increase(ipx_jev_requests_total{job="ipx"}[$range]))' 24h
   $Q prom 'max(max_over_time(ipx_signals_last_pass_age_seconds{job="ipx"}[$range]))' 24h
   

   up 0 means Prometheus cannot reach the port: the container is down or metrics_bind is unset. The daemon keeps all 20 connections open on Postgres (#241), so idle is the pool's 20 less those in use; in use at 20, idle at 0, means something is waiting for one: find the slow requests or scan of that time in check 3 and 5. A worker queue that only grows means downloads are not keeping up, or the worker is stuck on one job. A runtime queue above 0 for more than a scrape or two means every thread is busy: something is blocking a thread or burning one. Memory, open files or tasks only climbing across a day with no restart is a leak (a file or socket not closed, a task never ending); a step after a deploy is not. A reset of the uptime is a restart: match it with a deploy, or it is a crash (check 4).

   The worker: one job over 30 minutes (ipx_worker_job_seconds), or a scan last finished over 15 minutes ago, is the worker stuck, which ipx status does not show; its trace and the log lines of that time say on what. A kind whose share of ipx_job_seconds_total jumps is where the worker's time went (reap was most of it until #193). Web requests in flight near the pool's max mean the pool is what they wait on, and ipx_db_wait_seconds for how long. Live updates missed mean a page showed stale counts until reloaded. DeadAir over a minute on one file is a file ffmpeg chokes on. Any rate_limited Jev call means the pacing is not holding ipx under TypeSafe's limits (#277); each jev span has paced_ms, its wait for a turn. Any malformed Jev call is ipx asking wrongly (#257); key_refused is someone's key; the Signals judge over 10 minutes since a pass is stuck on Jev: its signals trace shows which jev call.

7. What the page saw. The browser's own errors and timings, which no server line shows.

   $Q metric 'sum by (ev, msg) (count_over_time({container="iPX"} |= "\"target\":\"ipx::client\"" | json | __error__="" | level="WARN" [$range]))' 7d
   $Q logs '{container="iPX"} |= "\"target\":\"ipx::client\"" | json | ev=~"play_start|stall" | line_format "{{.timestamp}} {{.user}} {{.ev}} {{.ms}}ms deadair={{.deadair}} {{.feed}}"' 24h
   $Q metric 'quantile_over_time(0.95, {container="iPX"} |= "\"ev\":\"list_shown\"" | json | __error__="" | unwrap ms [$range]) by (feed, filter)' 24h
   

   A player_error is playback that failed in front of someone: hls.js: ... is DeadAir's playlist (its trace has the playlist and audio requests), media error 4 a file the browser cannot play. A page_error or rejection is a bug in the page; src and line point into the minified app.js, so match the message to web/src. A list_shown over 1000 ms is a list someone waited for: its trace says whether the API or the network held it. play_start over 2000 ms, or a stall, is worth a trace: was it the server, Cloudflare, or the network.

docker logs --since 10m iPX 2>&1 | tail -5 shows the newest lines when Loki is behind.

ipx-web, the page

Caddy logs one JSON line a request (#210), logger http.log.access.log0: Loki's | json gives request_uri, request_method, request_remote_ip, status, duration (seconds, not ms), size. The request's headers and a ?token= are filtered out before it is written. Requests from 127.0.0.1 are local checks; leave them out. /api and /media pass on to ipx, which counts them in ipx_http_request_duration_seconds with its own time; Caddy's duration for them adds the hop. This is the only log with a line for every request. Caddy's own metrics, job="ipx-web" (caddy_http_request_duration_seconds{code,handler,method} and the like), count every request too, 502s included, which ipx never sees.

W='{container="ipx-web"} |= "http.log.access" | json | __error__="" | request_remote_ip != "127.0.0.1"'
$Q metric "sum by (status) (count_over_time($W [\$range]))"
$Q metric "sum by (request_uri) (count_over_time($W | status >= 500 [\$range]))"
$Q metric "topk(10, quantile_over_time(0.95, $W | request_uri !~ \"/(api|media)/.*\" | unwrap duration [\$range]) by (request_uri))"
$Q logs '{container="ipx-web"} | json | __error__="" | level=~"warn|error" | msg !~ "admin endpoint disabled|automatic HTTPS is completely disabled|HTTP/[23] skipped.*|exiting; byeee.*" | request_uri != "/api/events"'

A 502 is ipx not answering Caddy: a restart (match the deploy) or ipx down; Caddy's error line names it (dial tcp ...:8099: connect: connection refused). The page itself (/, /admin) waits on /api/page-check before it is sent, so a slow page is usually a slow page-check: its route in the histogram (check 3), or ipx's span inside Caddy's in the trace. 401 and 303 on / are people signed out being sent to /login, not a problem. Caddy's own notices at each start (no admin endpoint, no HTTPS or HTTP/3, its goodbye on SIGTERM) are by design and filtered out above, and so is aborting with incomplete response on /api/events: a page closing or reconnecting its live-update stream, about 600 a week.

Each access line carries the request's traceID (| json gives traceID): give it to $Q trace to see where the time went, Caddy's span with ipx's inside it. A slow page with ipx's span short is Caddy or the network; a long one is ipx.

$Q spans '{resource.service.name="ipx-web" && span.url.path != "/api/events" && span.url.path !~ "/media/.*" && duration > 2s} | select(span.url.path)'
$Q traces '{resource.service.name="ipx-web" && status = error}'

Who reaches it: request_remote_ip is 192.168.64.1 for the tunnel and anything else on Tower (Docker's gateway), 192.168.1.130 for Tower itself, a LAN address for the LAN. Caddy passes Cloudflare Access's headers on only from 192.168.64.1 (TUNNEL_PEER), and ipx trusts them only from Caddy, with a valid signed token. Testing sign-in from Tower proves nothing about the tunnel (#209): send it from a container on another network.

ipx-directory

It logs a line a request (target ipx_directory::http: method, path, status, duration_ms, token, the token's name and never the token), a line a snapshot (message snapshot: instance, accepted, refused_private, refused_blocked, refused_invalid), a line a refresher round (feeds checked: checked), a line a failed feed check (feed check failed: feed, the directory's id, and error), the admin's actions (feed removed and blocked, feed unblocked, snapshot dropped, each with by), and config read again when its token file changes. Its lines carry their own fields and no span (#363). Its metrics, as job="ipx-directory" (production 192.168.1.130:9188): directory_process_resident_bytes, directory_process_open_fds, directory_feed_fetches_total, directory_feed_bytes_total, directory_uptime_seconds, directory_build_info. Its traces (#213): a span a request (http), a child of ipx's when ipx called with a traceparent, and a trace a refresher round (refresh, feeds) with a feed span for each check. Its lines inside a span carry trace_id and span_id, as ipx's do (#375): a feed check failed leads to its check's span.

$Q traces '{resource.service.name="ipx-directory" && duration > 2s}'
$Q traces '{resource.service.name="ipx-directory" && name="feed" && duration > 10s}'
$Q traces '{resource.service.name="ipx" && name="directory_share"}'

A slow Directory page in ipx is usually its http span in the directory: the trace says whether. A refresher round far over a minute is one publisher's feed holding it, the slow feed span.

D='{container="ipx-directory"} | json | __error__=""'
$Q metric "sum by (status, token) (count_over_time({container=\"ipx-directory\"} |= \"ipx_directory::http\" | json | __error__=\"\" [\$range]))"
$Q metric "topk(5, max_over_time({container=\"ipx-directory\"} |= \"ipx_directory::http\" | json | __error__=\"\" | unwrap duration_ms [\$range]) by (path))"
$Q logs "$D | level=~\"WARN|ERROR\""
$Q logs "$D | message=\"snapshot\"" 7d
$Q metric "sum(count_over_time($D | message=\"snapshot\" [\$range])) by (instance)"
$Q metric "sum by (feed, error) (count_over_time($D | message=\"feed check failed\" [\$range]))" 7d
$Q metric "sum(sum_over_time($D | message=\"feeds checked\" | unwrap checked [\$range]))" 24h

Expect a snapshot from each server when what it shares changes and every half hour, so about 48 a day each; one every half minute is ipx resending what has not changed (#211, fixed). A burst of 401s with an empty token is someone without a token trying the public address, expected now and then; a steady stream is worth an issue, since Cloudflare can rate-limit it. A snapshot whose refused_private is not 0 means an ipx server tried to send a feed that looks private: find which in that server's log, since the directory refuses it but the sender should not have sent it. Many feed check failed lines for one feed are the publisher's, not the directory's. Each feed is checked every twelve hours (#243), 25 at a time, so feeds checked over a day should be about twice the feeds listed; none for over twelve hours means the refresher is stuck. could not store a feed check or could not read which feeds are due is its database (ipx_directory on the databases project's Postgres). A config that does not load (config changed but does not load) keeps the old tokens: worth fixing before someone's new token is needed. Weekly, Apple Podcasts synced (shows, accepted) brings Apple's charts in as the apple-podcasts snapshot (#252); Apple Podcasts' sync failed retries in an hour, worth an issue when it keeps failing, and Apple's podcast categories are not ipx's (added, gone) is an issue to update CATEGORIES in crates/feed.

Tower

The three share Tower with everything else, so a slow or failing ipx can be the host. The node exporter (job="node") has the history; twelve threads.

$Q prom 'max_over_time(node_load1{job="node"}[$range])' 24h
$Q prom 'min_over_time(node_memory_MemAvailable_bytes{job="node"}[$range]) / 1e9' 24h
$Q prom 'sum by (mode) (rate(node_cpu_seconds_total{job="node", mode=~"user|system|iowait"}[$range]))' 1h
$Q prom 'max by (device) (rate(node_disk_io_time_seconds_total{job="node", device=~"nvme.*|md.*"}[$range]))' 1h

For when, step back with offset: max_over_time(node_load1[10m] offset 60m). Load far above twelve with little CPU busy is processes waiting, on a disk or on the user shares: on 2026-10-12 debug builds linking through /mnt/user (shfs, FUSE) took load to 248 with mostly system time and stalled everything on the shares, ipx's downloads included. Available memory near nothing makes it worse. Disk busy near 100% on nvme1n1 is /mnt/fast, appdata and Postgres's data; on md* the array behind /mnt/user, the downloads. Say what was running at the time if you can tell (docker stats --no-stream); a host problem is the user's, not an ipx issue, unless ipx caused it.

The monitoring itself

A quiet dashboard can be a dead exporter. Check what Prometheus scrapes and that each container's lines are reaching Loki:

curl -s http://192.168.1.130:9090/api/v1/targets | python3 -c 'import json,sys
for t in json.load(sys.stdin)["data"]["activeTargets"]: print(t["health"], t["labels"]["job"], t["scrapeUrl"], t.get("lastError","")[:80])'
for c in iPX ipx-web ipx-directory; do echo "$c $($Q metric "sum(count_over_time({container=\"$c\"} [24h]))" 24h | tail -1)"; done
docker ps -a --filter label=com.docker.compose.project=monitoring --format '{{.Names}} {{.Status}}'
docker logs --since 1h tempo 2>&1 | grep -E 'failed to poll|blocklist poll complete' | tail -2

ipx, ipx-directory, ipx-web, node and the monitoring's own, prometheus, tempo, loki and alloy (#366), should be up. Alloy is scraped at alloy:12345 because its command sets --server.http.listen-addr; it listens on 127.0.0.1 without it.

The alert rules (#365) are grafana/alerts.py, written to /mnt/fast/arcane/projects/monitoring/grafana-provisioning/alerting/ipx.json and loaded by POST /api/admin/provisioning/alerting/reload. There is no contact point yet, so one firing is only in Grafana: ask it which are, at the start of every check. The password is in the monitoring project's compose file; never print it.

PW=$(grep -oP 'GF_SECURITY_ADMIN_PASSWORD=\K\S+' /mnt/fast/arcane/projects/monitoring/compose.yaml)
curl -s -u "admin:$PW" http://192.168.1.130:3444/api/prometheus/grafana/api/v1/rules | python3 -c 'import json,sys
for g in json.load(sys.stdin)["data"]["groups"]:
    for r in g["rules"]: print(r["state"], r["name"])' | sort

Tempo looks over its stored blocks every 5 minutes. failed to poll or create index for tenant means it could not, and then searches and trace lookups see only what it still holds in memory, about the last 15 minutes, while its 2,000 blocks sit unread on disk (its search log says total_blocks=0). On 2026-10-25 that was one block left empty, every file 0 bytes, by a write cut short the night before: unexpected end of JSON input. Find it and move it out of /mnt/user/logs/tempo/blocks/single-tenant/; the next look over, within 5 minutes, logs writing tenant index and finds the rest. Tell the user, since it is the monitoring's, not ipx's:

for d in /mnt/user/logs/tempo/blocks/single-tenant/*/; do python3 -c 'import json,sys; json.load(open(sys.argv[1]))' "$d/meta.json" 2>/dev/null || echo "$d"; done

Every monitoring container should be running; after Docker restarts, one can be left Exited (255) though its policy is unless-stopped (Prometheus, on 2026-10-12, when Docker lost its exit during an overload): docker start it and say so. blocky and nut on 192.168.2.21 are another machine, not ipx's; leave them. A container with no lines in an hour while it runs means Alloy is not shipping it; Alloy can take a few minutes to pick up a container after a deploy.

Rules

docker logs and docker exec iPX ipx ... are fine. On production's Postgres, EXPLAIN (ANALYZE, BUFFERS) of any query is allowed, read only, as CLAUDE.md's "Production's database" shows; any other query, ask the user first.

What to do with what you find

Follow the repository's rules in CLAUDE.md: every problem found gets a Gitea issue, with a closed stdin and a timeout on tea. Before filing, list the open issues and do not file one twice; comment on the existing issue with the new evidence instead.

R="--login git.sdf1.net --repo rays/ipx"
t() { timeout 30 /src/tea "$@" < /dev/null; }
t issues list $R --state open
t issues create $R -t "<what is wrong, as the user would notice it>" -L bug -d "<what, where, since when, how often, the query or trace id that shows it>"

Put in each issue what would let someone pick it up cold: the LogQL or TraceQL that shows it, a trace id, the first time it was seen and how often. A publisher's dead feed is worth one issue saying so (the user may want to unsubscribe or find its new address); a single 503 is not.

Finish with a short report to the user, a line for each of ipx-web, ipx, the directory, Tower and the monitoring: what is healthy, what is wrong (with the issue numbers), and anything you could not tell from logs, traces and metrics alone. Do not fix things unless asked; the check is for finding them.

When the checks come back empty

Check that there is data before concluding all is well (the monitoring section's per-container count over a day): iPX a few hundred (new items, Signals, shares with the directory, the page's timings; more on a busy day), ipx-web two thousand or so while the page is in use (a line a request), ipx-directory several hundred. Since #363 and #364 iPX can go an hour at night with no line at all, so judge it by the day. Nothing in a day means Alloy is not shipping it, or the container is down.