Deployed, the judge's query still read every entry: an expression index
has no statistics until its table is analysed, and Postgres, guessing a
third of entries were recent, chose a scan (898 ms) over entries_recent.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The judge's query read all of entries (605 MB, 566 ms) on every wake; an
index on each feed's items by date, as the query asks, and the rechecks
in a half of their own. Calls ran one at a time, about 150 ms each, so a
new Signal's month of about 8,000 items took twenty minutes: six in flight
now. On the server's key everyone's Signals on an item share one call.
TypeSafe's token counts go to /metrics and each jev span.
An item's text is cut at 1,000 characters instead of 2,000: on 55 real
items of eight podcast feeds and six Signals, on the development key,
1,000 changed none of 330 verdicts (nor did the same request twice) and
saved 12% of the input tokens, 1,266 to 1,112 a call; 500 changed one.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
With signals_on_server set and TYPESAFE_KEY present, the judge asks Jev
about everyone's Signals on the server's key; otherwise each person turns
Signals on in Settings with a key of their own, which is now checked with
TypeSafe before it is saved. The Signals folder shows only when they are on.
The browser suite's daemon inherited the shell's TYPESAFE_KEY, so fixture
feeds were sent to the real TypeSafe to be categorised; it is blanked now,
and an empty key counts as none.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The judge had spans but nothing in Prometheus: no count of items judged or
of how calls to Jev ended, nor whether the judge was still passing. The jev
span recorded an error field without the span's status, so Tempo's failed
traces left Jev out, and a timeout recorded nothing. Calls now send
traceparent like every other call out of ipx.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signals (#254) in ipx and the page, and the compression and tracing
work of #239-#253 in all three.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
xkcd's posts are an <img> whose alt and title are the comic's words. text_of dropped tags whole,
attributes and all, so Jev was sent "Feed: xkcd.com / Title: Ground Effect", ground effect being
an F1 aerodynamics term, and the comic went into the Formula 1 Signal at 0.60. With the alt text
it scores 0.19. Alt and title are read once each, and once when they say the same.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Jev chooses among options and cannot name one, so the page searches Font Awesome's index for each
word of the Signal's name, words and sentence, the best six each, and offers the likeliest 40.
POST /api/signals/icon asks a choice question on the server's TYPESAFE_KEY and answers with one
of those offered, never another. "Let Jev pick" in the icon picker, and a Signal made without an
icon gets one a moment after. Tried against TypeSafe: for AI Development, laptop-code 0.45 just
ahead of robot 0.40.
Each Signal's row says "12 items · 3 unread", as a feed's row counts its own, in place of its
words; /api/signals gives `items` beside `unread`, by the same rules as its list.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
unjudged() limits rows, one for each item and Signal, to 50, and the pass took the queue for
empty whenever the items numbered fewer than 50. With two Signals every batch is 25 items, so the
re-judge after #259 crawled at 25 items every 5 minutes. A batch is now full by its rows. A pass
TypeSafe stopped as malformed waits for the next wake however full it was, rather than asking again
at once.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Jev was asked "Is this item about any of these: ai, development, programming?", so any one broad
word was enough and the Signal's name, which says what is meant, was never sent. AI Development
held 54 items: Sega pulling a game remaster (0.90), a Surface Laptop review (0.95), a C++ tutorial
(0.99). It is now asked about one subject, called by the Signal's name, described by its words
and, when those are not enough, its owner's sentence. Scored against TypeSafe over those 54 with
"AI tools for writing software: coding agents, AI dev tools, vibe coding, AI in code review.":
5 stay, Linux kernel AI code review highest at 0.94; the remaster 0.06, the C++ tutorial 0.02.
signals.description is new, and its first start deletes every verdict, which answered the old
question, so everything in the window is asked again. Any change to a Signal's name, words or
sentence drops its verdicts, since each is part of the question. The queue query names the
Signal's sentence apart from the item's description, which share a row; the test caught the
item's winning.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The page fetched the Signals' unread counts only with the feed list, so reading an item, here or
in its own feed, left its Signal's badge as it was, and so did Jev adding one. The Apple Signal said
2 unread over a page with none. The page now fetches them again after a read or unread lands, and
on a new signal_hits event, which the judge sends once a pass for each person it found items for.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Every call was refused with 422: a noul question's `criteria` is an object of what true and false
mean, and it was sent a sentence. The error kept only the status, from error_for_status, so the log
and the trace said 422 and not why; TypeSafe's body said "Input should be a valid dictionary"
at questions.3.noul.criteria. The body is now in the error and on the jev span, with the status.
A request refused as malformed (400, 422) ends the pass, since every other item's would be refused
the same way: the 50 warnings a pass were the same failure 50 times. The state is text now, as
TypeSafe's docs show it, rather than an object that sent null for a feed with no title, and the
questions are keyed signal_<id>. Checked against TypeSafe with the corrected request: 0.96 for an
item about the keywords, 0.02 for one that was not.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The dialog asks what to listen for, the page says what it picked out of the noise of your feeds,
and a new Signal says nothing has come through yet instead of showing an empty list.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A Signal is a name and keywords. A task beside DeadAir's asks TypeSafe's Jev, on the owner's own
key, about each item of their subscribed feeds from the last 30 days and every new one: one call
an item, one noul question for each of their Signals still to hear about it, 0.5 or more a fit.
Every verdict is kept, fitting or not, in signal_verdicts, so nothing is asked twice, and new
keywords drop a Signal's verdicts to be judged afresh. A key TypeSafe refuses (401/402/403) is
passed over until its owner saves another, so a bad key is not sent the whole queue.
A Signal's items come through entries_where with an EXISTS on its hits, so filters, search, sort
and read state are an ordinary list's, and reading an item in a Signal reads it in its feed.
The page shows Signals under the places, with New signal, and Settings takes the TypeSafe key,
which is never sent back. The browser suite stubs Jev in the fixture server (IPX_TYPESAFE_URL).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Measured 2026-10-16, uncompressed until now: the directory's answers (the feed list ipx reads,
603 KB of JSON, ~134 ms of a 159 ms Directory request), ipx's snapshot upload (~100 KB each half
hour), both /metrics (4.7 and 0.9 KB every 15 s) and every trace export to Tempo. Already
compressed: ipx-web's answers (zstd, #239), the home page and the directory's public pages
through Cloudflare (brotli, zstd).
- The directory compresses with brotli, zstd or gzip as asked, and unpacks a gzipped request.
- ipx sends its snapshot gzipped; the directory is deployed first.
- Both /metrics answer gzipped, as Prometheus asks.
- opentelemetry-otlp's gzip-http: OTEL_EXPORTER_OTLP_COMPRESSION=gzip compresses the export, set
for all three in production's compose file; Caddy's exporter reads the same variable.
- web/build.mjs writes .br (quality 11) and .gz (level 9) beside each text file, and Caddy serves
them in place (file_server precompressed): app.js 30 KB where zstd on the fly gave 37, app.css
11 where 14. A .br or .gz asked for by name is refused, or /admin.js.br would be the admin
page's script without the check that keeps it to admins; in a handle of its own, since a bare
respond runs after the handles.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A Directory page took 983 ms during a scan with 19 ms under spans: the directory's own span ends
when it has answered, and ipx's reading of the 458 KB answer, its subscriptions and catalogue,
and the matching of 1,471 listings had none. call() is a span now, directory_call, with the
answer's size; rows_for has subscriptions, catalogue and match steps.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
With apple_country set, the directory fetches Apple's top 100 podcasts and the top 50 of each of
its 19 categories for that country (itunes.apple.com/{cc}/rss/toppodcasts), looks up each
show's RSS address (/lookup, 100 ids a call, three seconds apart, as Apple asks for no more than
about twenty a minute), and stores them as the snapshot apple-podcasts with no subscribers. The
refresher fetches them like any feed; the most-subscribed chart is unchanged. Apple's primary
genre goes in as the snapshot's category, used only while a feed names none.
The same weekly run reads Apple's genre tree (ws/genres?id=26) and warns when it differs from
ipx_feed's CATEGORIES; on 2026-10-15 they were the same, all 19 and their subcategories.
A syncs table keeps when it last ran: at start if a week has passed, again weekly, an hour after
a failure. keep_instances spares apple-podcasts, which no token holds, and no token may take
the name. A trace a sync, a span each answer.
CLAUDE.md: builds one cargo command at a time, the load looked at before and after, after a
build chained with a test compile took Tower past 200.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
coalesce(s.position, 0) > 5 and coalesce(s.flagged, false) hid from Postgres that a match needs
a row of entry_state, so it kept the outer join and scanned every item: Currently Listening read
all 574,000 (74,733 blocks from disk), 970 ms of its 1.3 s. s.position > 5 and s.flagged mean the
same, a missing row being neither, and let it start from the person's own rows: 2.2 ms, and
Pinned 32 ms to 2.3 (EXPLAIN ANALYZE, BUFFERS on production, read only). Unread keeps its
coalesce: no row is unread.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
#241: on production, All Subscriptions' query runs in 31 ms, unread oldest-first in 30, the
feed list's count in 5, generic plans included (EXPLAIN ANALYZE, BUFFERS, read only), against
0.3-1.6 s in ipx. A new Postgres connection takes 80-330 ms to open, and a page load asks for
more than are idle: four of six slow loads came as the pool grew (3 to 6, 2 to 6, 3 to 5, 6 to
8), and 20 requests at once took a median 414 ms while the pool grew from 8 to 20. The daemon
and the directory's server now keep their whole pool open (min = max); a command such as `ipx
status`, and the tests, which share the server's 100 connections, open as they need.
#246: the page reports list_shown, the ms from asking for a list to drawing it, with the view
and filter, under the click's trace; background refreshes are not timed. Timings skip report()'s
once-a-minute limit, which is for errors in a loop. A dashboard panel and table, and the prod
check, read it.
#245: the load test's browse scenario subscribes by OPML and leaves the Directory, which is
ipx-directory's since #204, to it.
CLAUDE.md records the owner's leave to EXPLAIN (ANALYZE, BUFFERS) any query on production,
read only.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Measured on 2026-10-15: the feed list and item lists went out uncompressed at over 100 KB; ipx
read 2.6 GB of feed bodies a day and the directory 3.9 GB.
- ipx-web compresses with zstd or gzip, leaving out /api/events, which a compressor would hold
back, and /media (#239).
- A body identical to the last one (SHA-256 in feeds.body_hash) is taken as a 304, unless
something is missing a read would fill or the check is forced; and a feed without a schedule
of its own whose body has been the same waits a twelfth of that, up to four hours, beside the
failure backoff (#242). 35 Patreon show feeds sent ~300 KB each hourly, about 250 MB a day.
- The directory checks each feed every twelve hours, not six (#243).
- reqwest accepts brotli and zstd too (#244).
#240 (cache the feed list's counts) is left open: at production's size on the test Postgres
that count takes 12 ms and All Subscriptions 32 ms, against 366 and 300 ms in production, so
the cause is production's database, not the query (#241).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A fetch that failed logged reqwest's words, which name the whole address: Patreon's auth key
went to Loki on 2026-10-15, and, since scan events go to every open page, to everyone's
browser too. #225 had taken queries out of the traces' address fields only.
Every event is redacted where it is emitted, for the log, the live stream and a terminal
alike: address fields lose their query and password, and so does any address in an error's
words (feed::redact_text). The artwork proxy's warnings and the directory's failed checks
the same. The error kept on the feed, shown to its own subscribers, keeps the detail.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The DeadAir tests named for each person's clock used one account. Now three play one file at
once (DeadAir at two pauses, and the original) and each keeps their own place; and two fetch
the same /media/{id}/deadair.m3u8 at once and each gets their own playlist, no-store.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The buttons followed the feeds being checked, so they stopped between one feed and the next;
7ecbe82 held them for 600 ms over each gap. Now pressing one starts it, and scan_done stops
it only when it is the end of the scan asked for: scan_done carries `only` and `forced`, so a
scheduled scan finishing first does not stop a button whose scan is queued behind it. An
error event, a job that failed without a scan_done, stops them all.
A row's spinner stood in for its unread count from the moment its feed was checked; most
checks take milliseconds, so each count blinked as the scan went by. It now waits 300 ms.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
From the slow traces of 2026-10-13: the directory checked a round's 25 feeds one after
another, so two libsyn feeds timing out at 30 s made a round 70 s; a forced scan looked up
62 site icons in turn, 22 s of 62; mjg59.dreamwidth.org takes the connection and never
answers, so it ran the full 30 s in every scan it was in; and since #225 strips queries,
every YouTube fetch span read the same address, outside any span naming its feed.
The directory now checks five at a time (SQLite stays on one connection, so its writes
queue). A forced scan looks up the icon in the fetch-ahead task, which already runs six
deep, and hands it to scan_one with the body. Both feed clients give up after 15 s of
silence (no successful fetch took over 8 s in a day of traces); FEED_TIMEOUT stays 30 s
for a feed still sending. The directory's client had no connect timeout; it has ipx's 10 s.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The fetch span and the download span recorded the whole address, so a private feed's
credential (Patreon's ?auth=) went to Tempo with every scan, readable by anyone with
Grafana. feed::redact drops the query, fragment and user:password; a token in the path
would still go, noted where it is.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
ipx's resident memory climbed from about 40 MB to 500-850 MB each night while scans ran five
times as often, and stayed there until a restart; the directory sat at 1.25 GiB idle. glibc's
allocator keeps what each thread's arena freed. mimalloc gives freed pages back, so resident
memory now follows what is in use. jemalloc was the first choice for its in-use figures, but its
upstream was archived in 2025.
Declared once in the directory crate, which ipx builds in, so both binaries get it.
What was missing to see it: the directory served no /metrics (now on metrics_bind, a port apart
from the tunnel's), nothing counted the feed bytes fetched, and the fetch span had no body size.
The feed crate's spans never reached Tempo either: both trace exporters filtered to their own
crate, so ipx_feed's fetch and parse spans were dropped.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Docker runs a healthcheck with exec, and containerd puts each exec's state in XDG_RUNTIME_DIR,
/run/user/0 on Tower, which elogind removes when root's last session ends. Every check then
failed and every container showed unhealthy while all of them answered. The checks are gone from
the sample compose file and production's; ipx's library counts for /metrics, which only the
healthcheck's status call refreshed, are counted every 30 s by the daemon.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Each release is now a Gitea release as well as an image: scripts/release.sh takes the binary (or
the page's files) out of the deployed image, checks its version, and attaches it as a tarball with
the changelog section as the notes.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The directory renders its own pages from the store: the most subscribed, categories, A to Z,
search, and a page per feed with its address and a Copy button. No token, nothing that writes; /v1
keeps its token check. The publisher's HTML is shown as text, item links only when http(s), with
a strict Content-Security-Policy; covers load from their publishers with no referrer.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The zone is in Cloudflare and the rays-unraid tunnel routes the name to ipx-directory, as it
does ipx-dir.sdf1.net, which keeps answering until every server using it has moved: a redirect
to another host would drop the Authorization header. Production ipx shares with the new name.
CLAUDE.md, docs/directory.md, the prod-check skill and the dashboard name it; the admin page's
placeholder and the settings' hint use directory.example.net.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A Directory feed's page offers Check now and Remove to an admin whose token is rwa (the page
learns it from can_admin on the feed's page, asked of the directory only for admins). Admin,
Directory lists the servers sharing, from a new GET /v1/instances, and the blocked feeds by
title: a feed is often blocked because its address carries a key, so the address stays in the
directory. Unblocking or dropping a snapshot sends this server's snapshot again at once; since
directory's 403 reaches the page as 403, not 502.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The workspace stops holding one version. ipx keeps 0.12.1-dev and its vX.Y.Z tags; the directory
starts at 0.1.0-dev, tagged directory-vX.Y.Z, and the page at 0.1.0-dev in package.json, tagged
web-vX.Y.Z, which web/build.mjs gives the script for the logo's tooltip beside ipx's. crates/feed
is a library, not released. CLAUDE.md's release steps are for one service at a time.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
CHANGELOG.md moves to crates/ipx/CHANGELOG.md with its history, which until 0.12.0 was all of
ipx's; web/CHANGELOG.md and crates/directory/CHANGELOG.md start with this release's lines for the
page and the directory. The dashboard's lines go: monitoring is not a service anyone uses.
CLAUDE.md says which file a change goes in, and that a release covers each.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The exporter setup and the traceparent parser move from ipx into ipx_directory::otel, which ipx
already depends on, so both binaries share them. The directory makes a span a request, a child of
the caller's traceparent, and a trace a refresher round with a span a feed; ipx sends the current
span's traceparent on every call to the directory and wraps each share in a span. The dashboard's
Traces row and the prod-check skill cover all three services.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Caddy logs one JSON line a request, filtered: no ?token= in the address and no request
headers, which hold the session cookie and Access's token. Its default log, errors included,
has the same filter, since an error line carries the request too. The directory logs a line a
refresher round, so the dashboard can show its progress.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A scan wakes the sharer, and a scan runs every half minute or so since each checks only the
feeds that are due: Tower sent the same 1,471 feeds every 33 seconds and the directory rewrote
its rows each time. The snapshot, with the address and token it goes to, is compared with what
last went; the same again waits for SHARE_EVERY.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
web/build.mjs builds a static site, web/dist, with the icons and typeface
beside the pages; Dockerfile.web serves it with Caddy, which passes /api
and /media to ipx on the same origin, so cookies, live updates and
Cloudflare Access work as before. ipx serves no page any more: build.rs
and every include_str! of the web files are gone, and its image is built
without node.
What ipx did to the page as it served it is kept another way. Caddy asks
ipx before serving a page (forward_auth): /api/page-check sends a browser
not signed in to /login and a non-admin away from /admin, and refuses them
/admin.js; /api/login-check sends someone Access vouches for past the
form; a page asked for with ?token= goes to /api/signin, which sets the
cookie. The theme and the admin link, once written into the page per
person, are set by web/boot.js in <head> from the theme cookie and the
last visit, before the first frame, and /api/me brings the version and
the account's theme from before.
Caddy strips the Cf-Access-* headers from every peer but TUNNEL_PEER, the
address cloudflared's requests come from, which is the boundary
trusted_proxies drew; ipx then trusts only Caddy's address.
The browser suite runs Caddy in front of its daemon, so it tests what
production would run. Not deployed: production's cutover is on #203.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
ipx asks ipx-directory on its people's behalf, with the token an admin set
in the new Admin, Directory tab, so the page never talks to it and never
sees the token or a feed's address. /api/directory, /api/popular and a
feed's Directory page answer from it; subscribing takes the address from
the directory and goes through Add feed's steps. With no directory set
up, the Directory says so. The local directory code is gone; its privacy
checks are the snapshot builder's now, and their tests moved with them.
With sharing on, ipx sends its snapshot after every scan, whenever the
config changes (set_cfg) and every half hour: each public feed and how
many here subscribe, never an OPML or Patreon creator itself or anything
private. A feed can be marked private in its settings, for a paid feed
whose address does not give it away; anyone subscribed may mark it, only
an admin unmark it. The admin's category for a feed that names none goes
with it, which the directory uses only while the feed names none, so the
category an admin gives a blog still reaches the Directory.
The directory lists only feeds it has read as a feed: the first snapshot
after a start went before the first scan, when an OPML was not yet known
to be one, and the directory listed the OPML. It now fetches what a
snapshot brings in at once instead of within the minute.
The browser suite runs its own ipx-directory on SQLite; the Directory's
tests match feeds by title, since the directory names them by address.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A new crate and service, crates/directory, live at ipx-dir.sdf1.net. ipx
servers send the public feeds their people subscribe to (POST /v1/snapshot,
replacing what that token sent before); the directory fetches each feed's
title, artwork, category, description and latest twenty items every six
hours with ipx-feed, and lists them A to Z and by subscribers summed over
every server. It holds metadata only, never a file or a file's address.
Tokens are in its config file as SHA-256 hashes, each ro, rw or rwa, and
the file is read again within a minute of a change, so adding and revoking
one is an edit and nothing else; a revoked token takes its server's counts.
At first they were to come from environment variables, which would have
meant recreating the container to add one. A file that does not load keeps
the tokens as they were.
Every address sent is checked with ipx_feed::url_looks_private, the check
ipx makes before sending, so one server's miss does not publish a paid feed
to every other; an rwa token can remove and block what both miss.
Deployed: ipx-directory in the ipx Arcane project, database ipx_directory
(its own login) on the databases project's Postgres, which db-backup
already covers, and a tunnel route with no Access in front. The store's
tests run on Postgres as well as SQLite when IPX_TEST_DATABASE_URL is set.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The first step of the split (#207). The daemon moves to crates/ipx unchanged;
feed.rs becomes crates/feed (ipx-feed), which the directory service (#204)
will share for its metadata refresher and its privacy check.
To stand alone, fetch takes the feed's address and its credentials instead
of ipx's config::Feed (config::Feed::auth gives them), collect_outlines
moves in with the OPML parsing, and the address half of looks_private is
feed::url_looks_private. ipx imports the crate as `feed` at its root, so
every crate::feed path is unchanged. The version is the workspace's, in the
root Cargo.toml.
Checked: cargo test on Postgres (118 + 29), tsc, page smoke, native bridge,
contrast and Playwright (72) pass; the image built from the new Dockerfile
is deployed and healthy.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
