Three medium advisories land on hono before 4.13.5: a toSSG() path
escape, a query parser that reads parameters past the URL fragment, and
unbounded dot-notation nesting in parseBody(). Only the second one
touches this server -- c.req.query() is read in imageproxy, icsproxy and
app -- and even there safeFetch validates the value it actually fetches
rather than a separate pre-check, so there was nothing to desync. toSSG
and parseBody are never called. The bump is still worth taking on its
own: it is a patch release with no API change.
The declared range moves with it, from ^4.7.4 to ^4.13.7, so the
security floor is recorded in server/package.json and not only in the
lockfile.
The dependabot.yml is the actual fix for how these were found. There was
no config, so nothing opened a PR and the alerts sat on a dashboard
until someone thought to look. Routine updates now group into one PR a
week; majors stay separate, because they are migrations.
Closes#310.
A dark campaign rendered with beige cards inside it. markKeptSurfaces
marks any element whose declared background is below the luminance
threshold, with no area cap, so a 600px layout card is marked exactly
like a button. The CSS then exempted the marked element and its whole
subtree via [data-ihm-keep] *, so a light table nested in that card was
never touched. In the reported specimen 14 of 21 light panels survived.
The rule now is that being inside a painted surface is not inherited past
a sheet. The walk tracks that state and emits a second mark,
data-ihm-in-keep, for elements sitting on paint with no background of
their own; the CSS exempts those explicitly instead of exempting every
descendant. A nested light sheet ends the protection, and paint resumes
below it, so a button inside such a sheet is still kept whole.
The alternatives in the report were not taken. Dropping the descendant
half of the selector outright puts back what #294 fixed: a nested label
on a coloured cell loses its colour. An area threshold is a magic number
that misfires on both a legitimate hero banner and a small dark panel
with a light chip in it.
The tests assert against the neutraliser selector lifted out of
EMAIL_BASE_CSS rather than against the marks. The first draft of them
checked which attributes were set and passed against the unfixed code,
which proved nothing: the bug was in the rule that reads the marks, not
in the marking. All four fail without this change.
The deploy on 2026-09-08 went out at the origin and did not arrive.
Cloudflare went on handing out the previous `sw.js` -- `cf-cache-status:
HIT`, with an edge TTL of four hours, longer than the hour we asked for
-- because the file is neither a hashed asset nor HTML and so fell into
the ordinary `max-age=3600` case.
That is not a freshness preference. The service worker is the app's whole
update mechanism: a browser holding the old one goes on being served the
shell that worker knows and never learns a deploy happened, so the deploy
simply does not land. The manifest matters for a second reason -- the two
have to agree. A fresh manifest advertising a share target, answered by a
worker that has never heard of one, sends the share to the server for a
405. Either being old is survivable; disagreeing is not.
`no-cache` rather than `no-store`: both may still keep a copy, they just
have to revalidate it, which is a 304 and costs nothing. Neither gets to
answer with its own copy without asking.
Narrow on purpose -- two files, named, rather than a policy that quietly
stops the icons and fonts being cached as well.
Both happen in the background. The phone stays where it is.
This was twice described as impossible, here and in FEATURES.md: the
service worker was said to have no session, so anything touching mail had
to open the app. That is wrong, and checking it rather than repeating it
is the whole of this change. ihasmail's session is an httpOnly cookie
against its own origin and the only other thing the API asks for is a
fixed `x-requested-with` header, which is not a secret and is not held
anywhere. A same-origin fetch from the worker carries the cookie like any
other. Confirmed against the mock: logging in with curl and then issuing
`Email/set` with nothing but that cookie and the static headers marked a
message read and moved it to Archive, HTTP 200. Nothing the tab holds in
memory is involved, because the API asks for none of it.
Two actions, because `maxActions` is two on Android and anything past it
is dropped without a word. Archive and Mark as read are the two worth
having: they are what somebody does to a notification they have already
read the whole of. Reply is not among them -- it would have to open the
app, which is what tapping the notification does already.
The worker still cannot reach a catalogue. It is plain JavaScript copied
into the build, outside the bundle, with no i18n and no idea which
mailbox is the archive. So the app writes both down in the same cache it
already uses for handoffs, and rewrites them whenever the language, the
account or the folder list changes. Where there is no such note -- between
installing this worker and next opening ihasmail -- the notification
appears with no buttons at all, rather than English ones over a mailbox
guessed by name. That also fixes two strings the worker had always shown
in English regardless: "New mail" and "(no subject)".
A session can be gone by the time a button is pressed. That comes back as
a refusal and the notification says so, rather than vanishing as though
it had worked. It does not open the app to recover: being interrupted is
what the button existed to avoid.
The two claims that were wrong are corrected rather than quietly deleted,
including the one about push renewal -- which still needs a tab, but for
a different reason than the one given. The reason is when the worker
runs, not what it may do: it wakes only for a push, and the push stops
when the subscription lapses.
Two new strings, in all nine catalogues.
CI caught this on Node 22 while it passed here on 26. `new File([blob],
…)` only puts the blob's contents in the file where that implementation
recognises a Blob as a part; where it does not, it stringifies it, and
the file contains the thirteen characters "[object Blob]". No error
anywhere -- the name, the type and the attachment are all correct and
the contents are gone.
A browser would not have done this. It is worth not relying on that: an
ArrayBuffer is a part on every implementation, and the whole file is in
memory a moment later regardless, since it is about to be uploaded.
ihasmail could hand a file to the share sheet as of #306, and was still
not in it. Share a photo from the gallery, a link from the browser or a
document from a file manager and ihasmail was not among the places it
could go, which is the one piece of operating-system integration a mail
app is expected to have.
A share is a POST that navigates, and there is nothing on this side that
can answer one: the app is a client-side router with no endpoint at that
address, and the server behind it would need a route that understood the
composer. So the service worker intercepts it, takes the form body, puts
the files and text in its cache, and redirects to the app -- which finds
them on start and opens a draft holding them. The subject is the shared
title, the text and the link become the body, and files are attached and
begin uploading. Nothing is addressed: a share says what to send, never
who to.
The body is pushed in above the signature rather than passed to open(),
because open() only fits a signature when it is given no body at all --
the obvious version drops the signature from every message that started
as a share, and nothing about the draft looks wrong afterwards.
Collected on every start rather than when the launch URL says so. A share
to a signed-out ihasmail lands on the sign-in page, and there is no
account to attach to until it is done, so the payload has to outlive a
redirect and a login -- which the query string does not. What that costs
is a stash nobody came back for, so it carries a timestamp and expires
after ten minutes.
`accept` names wildcard families and explicit types and extensions both.
A mail client attaches anything, but wildcards are not in the
specification and operating systems differ over which form they match on,
so the explicit list is what holds if the families are ignored.
The cache name the worker and the app have to agree on now has one home
on the app side. It was written out twice, and a drift would not fail --
a push verification would simply never complete and a share would arrive
at an empty composer.
One case is deliberately left to fail loudly: an app still installed
whose worker has been cleared away POSTs to the server, which answers
405. A server route would trade a plain error for a silent nothing, and
the payload is gone in both -- it only ever existed in that request body.
Verified by test, not on a device: Android is the only place this exists
at all, and the extension driving Chrome is not connected here. The
handoff is pinned from the tab's side against a cache shaped exactly as
the worker leaves it, since the two files never see each other.
Three things an installed ihasmail did not do that a phone user expects,
and all three are about the app once it is off the browser tab.
The unread count was painted into the tab title and the favicon, neither
of which exists in `display: standalone` -- so putting ihasmail on a home
screen threw the count away entirely. It goes to the Badging API as well
now. Web Push marks the icon while the app is closed, and marks it with a
dot rather than a figure: the service worker has no session to ask how
many messages are unread, and a push carries the new mail rather than a
total, so counting the payload would badge "2" over an inbox holding
forty. The next tab to open writes the real count over it.
Sharing is new. Everything that left ihasmail left as a download, which
on a phone is close to a dead end -- the file lands in Downloads and
whoever meant to send it somewhere goes looking for it in a file manager.
The share sheet is now on the message menu, on each attachment row, and
in the file viewer, which is where an attachment is already open and
where both callers meet. A message shares as text rather than as the
.eml beside it: a share sheet is aimed at everything that is not a mail
client, and an .eml in a chat app is an attachment nobody can open.
Every control feature-detects, and sharing a file is a separate question
from sharing at all -- desktop Linux and Firefox have neither, and not
every browser with `share` takes files. Anything that fails, including
the transient activation running out while a large attachment is fetched,
falls through to the download the button sits beside, so the worst case
costs a tap rather than the file. `NotAllowedError` is reported as
unsupported for that reason: it cannot be told apart from a refusal, and
a toast about activation is not something a reader can act on.
The share strings are contextual keys rather than the existing "Share…".
That one means granting another account access, and several languages use
a different verb for it -- German had "Freigeben" where the sheet wants
"Teilen". Three new strings, in all nine catalogues.
The manifest gains `launch_handler: navigate-existing`, so a mailto:, a
shortcut or a notification tapped while ihasmail is running arrives in
the copy that is running: two windows on one inbox disagree about what
has been read. `focus-existing` would have been wrong -- it only focuses
and leaves the target URL to launchQueue, which nothing here consumes, so
it would swallow the mailto. There is deliberately still no `id`, and the
manifest now says why: it is the one member resolved against the origin
of start_url rather than against the manifest's own address, so no
relative form can name a subpath mount, and the default id already is
start_url -- writing one now would give every installed copy a new
identity and orphan it as a second app.
Verified by test rather than on a device: the extension driving Chrome
was not connected, and Chrome on Linux has no Web Share to drive anyway.
The preview dialog is covered by a component test that stubs the browser
both ways.
The bullet asserted that ihasmail stays stateless without saying what
that is scoped to, which reads as a claim about the file rather than
about the process. Name both halves: the format is ihasmail's, the file
is the account's.
The plural-key gotcha and the reasons store tests miss visible bugs are
contributor guidance, not a side file: they belong next to the rest of
the pull-request checklist where anybody sending a change will read them.
The file was added without being asked for. It stays, but with its scope
stated at the top so it does not grow into a second contributor guide:
the nine catalogues and what it takes to confirm a visible change works,
and nothing else.
Contacts and calendars disagreed on a re-import: a vCard or LDIF entry
whose identity a book already held overwrote the card there (#242, #274),
while an event whose UID a calendar held was counted and thrown away
(#222). The asymmetry was never decided -- it was where each half stopped.
Decided on #279: calendars update too, with two properties held back.
`participants` carries every attendee's accepted/declined and
`recurrenceOverrides` holds every "just this Wednesday" edit made here.
Both are decisions taken after the file was written, and a file that
mentions them at all describes them as they were at export, so writing
either one over would destroy work silently and return no error. A
corrected export now fixes the time, the title and the location, and
leaves who said yes alone. `uid` is held back with them: it is what the
two were matched on, so it is already equal.
The scan returns uid -> id rather than a set of UIDs, since updating
needs something to address, and creates and updates now share one
`maxObjectsInSet` budget the way contacts' `writeCards` does -- 300 new
and 300 changed batched separately would be two calls of 300, neither
over a ceiling of 500 and both refused. Counts become created/updated,
reported as the contacts import reports them.
Still no scheduling messages, on an update as much as on a create. That
is a real cost -- an event a re-import moves is moved here and nowhere
else -- and it is the lesser one: an import is not the place to start
mailing a room full of people who never asked for it.
Driven against the mock end to end: a second file with the same UID
updated the event in place, took the file's title, start and location,
and left an accepted RSVP and a per-occurrence override untouched even
though the file carried participants of its own.
On a phone the folder list is the drawer, so it is also where a new folder
is started -- and the New folder dialog was stacked at 900 against the
drawer's 950, so it opened behind the folder list with only a sliver
showing past the drawer's right edge. Unusable: the name field and the
Cancel button were both underneath.
The same trigger, the same fault, one layer down: Compose in the drawer
opens a full-screen composer, and at 800 that came up behind the drawer
too.
A modal has to outrank the navigation that raised it. The dialog backdrop
goes to 960 and the composer dock to 955, which keeps every relationship
those two already had -- a dialog still clears a composer, popovers,
tooltips and toasts still clear both -- and adds the one that was missing.
Desktop is untouched: the drawer's z-index only exists below 768px, and
nothing sat between 800 and 960 anywhere else.
The stack is now written down beside `.dialog-backdrop`, and guarded by a
test on the stylesheet rather than a component test: jsdom has no paint
order, so nothing in a rendered tree can tell that a dialog is behind the
drawer that opened it.
No user-visible strings change; the nine catalogues are untouched, and the
fallback count holds at 16 in each.
The README badge still read 0.16.20, in both the label and the shield it
links to. It is the first version number a reader sees and it was the one
place the prose update missed, because it is HTML rather than Markdown.
Two entries had gone further than stale and were wrong. FEATURES said
occurrence ids are not stable across a write, and KNOWN-ISSUES carried
that as a live hazard with the five-week series that proved it. 0.16.21
fixed exactly that: an occurrence is identified by its recurrence id now,
and holding an id across a write keeps it on its own date. Both entries
say so, keep the old behaviour and the evidence for it because the client
still supports 0.16 as a whole, and record what replaced it.
The defence in the client stays either way, and the reason is written
down: re-resolving by recurrenceId costs one lookup, a date can still
leave a series, and 0.16.20 is still a server someone may be running.
The KNOWN-ISSUES header now says the live instance runs 0.16.21 and,
unlike the upgrades before it, that this one was re-run rather than read
against the diff — with what was exercised by hand.
Themes were not in "What's in it" at all, which is odd for something a
reader sees before anything else. There is now a bullet for the twelve,
saying that palette and light-or-dark are separate choices and that a
palette which would not meet the contrast this app claims is not written.
FEATURES lists the six new palettes and what each borrows for its light
half, and records the rule that changed with them: body text used to be
checked and then accepted or rejected, which would have turned away five
of the six over a bar their designers never aimed at, so it is now lifted
along its own hue like every other text tone. Twenty-one of the twenty-two
borrowed halves need at least one lift.
The message-theming entry gained the second switch, including why the
first one alone did nothing for most real mail.
The Stalwart section records what the release is validated against rather
than only what it requires: 0.16.21, run against a real instance, with the
four client-visible JMAP changes named. The mock section gains its third
switch and says it tracks the current release, confirms each behaviour
against a real server first, and rewrites rather than deletes the test
that pinned an old behaviour.
Every catalogue was at 1,255 of 1,279 with 24 strings rendering English.
Eight of those are real UI text and are now translated in all nine
languages: the five sort options that had no entry while their opposites
did (Read first beside Unread first, Unstarred first beside Starred
first, Smallest first beside Largest first, and the two alphabetical
directions), and the three sentences behind the link and external-sender
warnings. Each follows the phrasing its own catalogue already used for
the sibling it sits next to.
The remaining sixteen are left in English deliberately, because
translating them would be wrong: product and project names, the sample
addresses in placeholder text, bare URL prefixes, the ellipsis used as a
masked value, and two mail header names.
Per locale: 1,263 of 1,279, up from 1,255.
**The stale list is not touched, and should not be cleaned blindly.** The
checker reports 41 keys as translated-but-never-looked-up, and some of
them are live. "Classic" is the clearest: the palette picker renders it
through translate(p.name) from a constant, so the extractor sees no
literal, while the German "Klassisch" it would delete is the exact fix
issue #247 asked for. "Add star" and "Remove star" are the same shape,
reached through a ternary in a JSX label. Teaching the extractor those
two call sites is the prerequisite for trusting that list.
Catppuccin, Solarized, Ayu, Kanagawa, Everforest and Primer, each with the
light and dark variant its own project publishes: Latte and Mocha, Lotus
and Wave, and so on. Values were fetched from each project's own repository
and recorded in .palette-sources/palettes-upstream.md, with the two tiers
no project publishes marked derived rather than passed off as upstream.
Four candidates were rejected rather than adapted. Nord and Synthwave '84
publish no light variant, and inventing one is not porting a theme.
Monokai is proprietary and its licence forbids redistribution. Material
Theme has become a commercial product whose repository no longer publishes
a palette at all.
Body text is now lifted for contrast like every other text tone rather
than exempted and merely checked. Most of these palettes target their own
~4.5:1 for body text where ihasmail asks 7:1, so the old rule would have
rejected five of the six on a bar their designers never aimed at. Nudging
the published colour along its own hue is what the script already does for
muted text, links and accents, and every shift is printed in the generated
CSS: Solarized light moves 4.13 to 7.07, Primer needed nothing at all.
Primer is named for the design system, not for GitHub. The colour values
are MIT; the name and the logo are trademarks, and NOTICE says plainly
that nothing here is endorsed.
The picker grid already wrapped on its own, so twelve cards needed no
layout change.
The hint under the theme picker listed the third-party palettes by name.
That sentence is translated into nine languages, so every palette added
meant rewriting it, retranslating it nine times, and leaving the previous
version behind as a stale key nothing looks up.
It now describes the rule instead of enumerating the cases: a palette
named after another project is that project's work, used under its own
licence. True of the four here, true of the next one, and true without
saying "MIT" for a palette that might not be. The names are already in
Settings beside each swatch and in NOTICE with their copyright lines,
which is where a credit belongs.
Swapped rather than added in all nine catalogues, so the old key is gone
rather than left stale: 1,255 of 1,279 translated per locale, unchanged,
and the 41 pre-existing stale keys are neither added to nor cleaned up
here.
Appearance gained "Apply the theme to messages too" some time ago, and it
themes an HTML message only when the message brings no colours of its own.
That predicate is the right default and it almost never passes: one
`color:#FFFFFF` on one button label opts a whole message out, so in real
mail — receipts, shipping notices, anything from a template — the switch
did nothing at all and the reader kept a bright white card on a dark UI.
A second switch, off by default and only meaningful with the first on,
forces the palette over the sender's colours. It cannot be done perfectly,
which is why it is a separate, explicit choice: the same bargain a
dark-reader extension makes.
What it does is tell two kinds of colour apart. A *sheet* the design sits
on — the white 600px wrapper — is neutralised, and a *painted surface* —
a call to action, a footer banner — is kept whole so its label stays
legible on it. Relative luminance decides, at 0.5: white wrappers sit at
1.0, a blue button near 0.09. Only the painted ones are marked, with
data-ihm-keep, and one rule in EMAIL_BASE_CSS neutralises everything else.
Nothing the sender wrote is removed, so the switch is reversible, colours
arriving from a <style> block are covered as well as inline ones, and
print still pins the tokens to ink on white.
The mock grew the message this is about: an outer wrapper on
bgcolor="#ffffff", a <style> block, a coloured button, a grey footer.
Without one, neither the bug nor the fix could be seen.
Verified in a browser against the mock: with only the first switch on the
card is still white; with both, the wrapper computes to transparent, body
text follows the theme, and the button keeps white-on-blue. Two surfaces
marked, which are the two the message paints.
Closes#290
Four changes, each confirmed against a real 0.16.21 rather than read from
the changelog.
Synthetic recurrence ids are now built from an occurrence's recurrenceId
instead of its position, so they survive a write. This reverses a hazard
the mock reproduced on purpose: up to 0.16.20 writing one override
renumbered the series and a held id silently named a different date. A
five-week series was expanded live, its third occurrence retitled through
its synthetic id, and all five original ids re-read; every one still
resolved to its own date. The test that pinned the instability now pins
the stability, with two more around it.
Calendar/get and AddressBook/get return every property when properties is
omitted or null, shareWith included. Mailbox/get on the same server still
omits it, so that stripping stays and now applies to mailboxes alone.
EventSource ping events advertise the interval in seconds, not
milliseconds. The mock parses the parameter it used to ignore: a 30 s
floor, larger values honoured, 0 disables pings, a non-numeric value is a
400. The first ping now arrives one interval in rather than on connect,
which is what the server does.
CalendarEvent/set rejects create, update and destroy with forbidden when
the request asks for scheduling messages and the account may not send
them. MOCK_NO_SCHEDULING_SEND=1 develops against that account.
A signed-in tab held two sockets: the browser's, and one from ihasmail to
Stalwart carrying that tab's push stream. The upstream one was most of what a
tab cost, and the only reason Stalwart's connection limit applied to ihasmail
at all.
RFC 8620 section 7.2 defines the other push transport: a PushSubscription,
where the server POSTs StateChange objects to a URL the client registers.
Stalwart 0.16.20 implements it. ihasmail now registers one subscription per
account at sign-in, and when Stalwart POSTs a change, fans it out to that
account's open tabs over the browser-facing streams it already holds. A tab
opens on the relay as before and is moved to fan-out the moment its account
verifies -- the upstream request is ended, the browser stream is untouched,
and nothing keeps a reference to what was torn down. After that there is no
upstream connection at all. The shapes are the RFC's; nothing here is taken
from any other client.
Measured at a 256 MiB cap over a private plain-HTTP route, against a real
Stalwart with 6,144 accounts verifying during the ramp and no failures:
tabs client Stalwart system KiB/tab
raw relay (before) 5,000 48.2 46.4 94.6
push by subscription 6,144 33.3 4.8 38.0
a direct-to-server client 12,389 4.8 53.8 58.6
Descriptors per tab: one, the browser's. Stalwart pays 4.8 KiB per tab and
holds no connection for it, so its per-listener connection limit no longer
applies to ihasmail. What remains per tab on the client is Node's cost for a
held HTTP/1.1 connection.
PUSH_URL is the https origin Stalwart can reach ihasmail at. The RFC requires
https and Stalwart enforces it, so Stalwart must trust that certificate: a
public TLS front already does; a private segment needs an internal CA in
Stalwart's trust store. An account whose subscription cannot be verified
stays on the relay, so nothing breaks -- only the saving needs the
certificate. PUSH_MODE=relay disables the subscription path entirely.
/api/push/:token accepts only a JSON body under 64 KiB for a known 32-byte
token, answers 200 or 404, and echoes nothing. /api/health reports how many
accounts are verified, pending or failed and how many tabs are on each path.
Listing latency at one user went from 1.95 ms on the previous release to
3.25 ms on main, and a bisect put the whole of it on the compression commit.
Not on compressing: the harness never sent Accept-Encoding, so nothing was
ever gzipped. Hono's middleware still inspects every compressible response it
declines and sets Vary on it, and setting a header on a streamed passthrough
rebuilds the Response off its fast path -- about 1.2 ms per JMAP call, on a
request that had asked for nothing.
The middleware now runs only when the request names gzip or deflate. Measured
at one user against the same Stalwart:
compressor touches but declines, no Accept-Encoding 3.25 ms
skipped entirely, no Accept-Encoding 2.02 ms
compressor applied, Accept-Encoding: gzip 2.27 ms
previous release, either 1.95 ms
Applying gzip to a JMAP response costs about a quarter of a millisecond and
saves three to five times the bytes on every listing and body, so JMAP
responses stay compressed by default; COMPRESS_JMAP=0 turns that off for a
deployment that would rather not.
The raw push relay is also made safe to tear down from outside -- the
browser stream keeps its headers and is not ended when the upstream request
goes -- which the next change relies on.
Only sign-in and the account endpoints were rate limited. JMAP, blob
downloads and the image and calendar proxies had no budget at all, and the
proxy is one Node process that saturates a core at roughly 2,000 operations a
second -- measured at 110% CPU under 150 concurrent users. One signed-in
account looping requests could slow every other user on the instance.
Each session now gets API_RATE_LIMIT requests a minute on those routes, 1,200
by default: twenty a second sustained, well above what a busy tab does and an
order of magnitude below where one tab starts to hurt the rest. Over budget
returns 429 with Retry-After. Sign-in keeps its own, separate limiter.
Checked in situ: one session driven flat out was cut off after exactly 1,200
requests, and with API_RATE_LIMIT=0 throughput at 50 users is unchanged.
639 MB unpacked and 119 MB compressed, against 239 MB and 59 MB now. Two
causes, both in the runtime stage.
The build stage's node_modules was copied across whole: 132 MB of vite,
TypeScript, esbuild, jsdom and React that the server never loads, since it
needs hono and its Node adapter and nothing else -- about 4 MB. The runtime
stage now installs the server workspace's production dependencies on its own.
Then `chown -R node:node /data /app` rewrote every one of those files, which
on overlayfs copies the whole tree into a second layer of the same size. Only
/data is written to at runtime; /app stays root-owned and read-only to the
process, which is what an immutable container wants anyway.
The base image's npm, npx, yarn and corepack are removed from the runtime
stage as well. The server is started with `node` directly and never calls
them; anyone who gains code execution should not find a package manager
waiting.
Checked that the image starts --read-only, serves the gzipped bundle, signs
in against Stalwart, holds a push stream, and that `hono` loads from the
3.1 MB that remains.
Two changes on the push path, both measured against a real Stalwart 0.16.20
with the container capped at 256 MiB and tabs added in steps of 200 until the
kernel killed it:
tabs held per tab of which native
before 1,665 133 KiB 81 KiB
pin upstream calls to STALWART_URL 3,400 58 KiB 8 KiB
+ raw push relay 4,979 37 KiB 10 KiB
Stalwart advertises absolute https URLs in every session, and the proxy
followed them -- so even with STALWART_URL naming a private plain-HTTP hop on
the same Docker network, every held push stream went out through TLS. That leg
is about 80 KiB of OpenSSL state per tab: native memory Node cannot see, which
is why neither the heap ceiling nor the stream buffer size ever moved the
number. absoluteUpstream() now keeps the path and query from the advertised
URL and the scheme, host and port from the configured one. A setup that must
reach Stalwart at an origin other than the one it was given sets
STALWART_FOLLOW_ADVERTISED_URLS=1.
With the transport out of the way, the fetch()-based relay was the next cost:
an undici Response, a web ReadableStream, a reader and Hono's stream bridge
held alive per tab, about 44 KiB of heap for a session that otherwise costs
4 KiB. relayPushRaw() pipes the upstream socket into the Node response and
tells the adapter the response is already sent. RAW_PUSH_RELAY=0 restores the
fetch path for comparison.
JMAP throughput is unchanged (2,383/s against 2,484/s at 50 users, inside
run-to-run noise); the relay does not touch that path. Verified that a push
stream through the raw relay delivers a StateChange while mail is written.
The install page's advice to set --max-old-space-size was measured in the same
runs and made no difference at all -- 3,400 tabs with it and without -- and
is withdrawn in the docs alongside this change.
* Compress our own responses
The bundle went out uncompressed unless a proxy in front did the work: 933 KB
on the wire where 311 KB does, on every first load. Both example proxy configs
compress, but that only helps deployments that copied them, and the default
should not depend on reading the examples.
Hono's middleware, with the proxy routes held back. `/api/blob`, `/api/image`,
`/api/ics` and `/api/upload` forward somebody else's bytes under a
content-length copied from upstream, and issue #76 was a silent truncation
caused by exactly that header disagreeing with its body. Re-encoding them
would be safe in principle -- the length is dropped and the response goes out
chunked -- but they carry attachments and images that are already compressed,
so there is nothing to win and a scar to respect.
`/api/events` is listed with them even though Hono already skips
text/event-stream by content type, so that changing the push route's type
cannot quietly start buffering the stream.
`/api/health` is excluded for the opposite reason: at 47 bytes gzip made it 73.
Hono's size threshold cannot catch that on its own, because it only applies
when a response carries a content-length and `c.json()` does not set one. The
other JSON routes stay compressed -- a JMAP response has just as unknown a
length and can run to hundreds of kilobytes.
Verified against the built image: assets come back gzipped with Vary set,
662 KB to 209 KB; /api/events still returns text/event-stream with no
content-encoding and delivered a StateChange while mail was being written;
health is 47 bytes either way. No user-visible strings, so no catalogue work.
* Word the comment for either side compressing
The app compresses its own responses as of the follow-on change, so a note
saying the bundle ships uncompressed would be wrong as soon as that lands.
nginx passes through what the upstream already encoded rather than re-encoding
it -- verified single-encoded with both layers active -- so the directives are
correct either way and the comment now says so without asserting which side
does the work.
* Test compression against a fixture, not the web build
The compression tests asked for `/` and asserted a gzipped 200. That passes
locally, where `web/dist` is lying around from an earlier build, and fails in
CI, which runs `npm test` before `npm run build`: with no bundle the shell
route serves the "web build not found" fallback, which is short, plain text and
correctly uncompressed. The failure read as compression being broken when the
tests were simply depending on a build step that had not run.
They now build their own static root in a temp directory and point STATIC_DIR
at it, in a separate file so the environment is set before the app module is
imported. Checked by moving web/dist aside and running the suite the way CI
does.
The Caddy example has `encode zstd gzip`; the nginx one had nothing, so a
deployment following it shipped every asset uncompressed. Measured against the
built app that is 915 KB on the wire where 307 KB would do -- the difference
falls entirely on first load, and silently, since nothing about it is visible
without inspecting response headers.
`text/javascript` is listed explicitly. The server sends scripts with that
type rather than `application/javascript`, so a conventional gzip_types list
compresses the stylesheet and leaves the 647 KB script alone -- which is what
happened on the first attempt at this change.
text/event-stream is deliberately not listed. Compressing or buffering the
push stream would break it; proxy_buffering is already off below for the same
reason. Verified that /api/events still delivers a StateChange event through
the proxy, as plain text, while assets come back gzipped with Vary set.
A signed message now says whether that holds up, as it is read. This is
verification only: nothing here signs, encrypts or decrypts, and the
private-key question that blocks those is untouched. Verifying needed
none of it, because the certificate travels inside the message -- which
is why this is the half that could be built.
What it checks. For multipart/signed carrying PKCS#7, the exact bytes of
the signed part -- headers included, canonicalised to CRLF -- are hashed
against the messageDigest attribute, and the signature over the signed
attributes is verified with WebCrypto against the certificate inside the
message. RSA PKCS#1 v1.5 and ECDSA over P-256/384/521, with SHA-256, 384
or 512.
The trust model is the design, and it is deliberately small. A browser
has no system trust store, and the certificate arrives inside the
message, so anyone can self-sign as anyone: on its own a good signature
shows only that the sender held the key they attached. So the word
"verified" is never rendered, and the reassuring case is not the loud
one. What carries the weight is remembering -- the first signed message
from an address pins its fingerprint, later ones are compared, and a
signer that changed is reported with both names and told to check by
another route. Trust on first use, no certificate authority anywhere.
The pins live in the account's settings rather than the browser: one
that only a single device knew would greet the same correspondent as new
everywhere else, which is how people are trained to click past the one
warning that matters. A pin records the message that created it, so the
message that established a signer keeps saying so instead of appearing
to be corroborated by itself -- without that, the very first signed
message anybody receives reads as "the same signer as before", where
before is itself. A changed, mismatched or expired signer is never
pinned, since writing the anomaly into the baseline makes every later
message agree with it.
Three things are declined rather than attempted, and all three say
"could not check" rather than "does not check out", because ignorance
and an accusation are different claims:
- OpenPGP, by name. The signature carries no key and there is nowhere
to get the sender's: x:PublicKey is the account's OWN registry, and
a keyserver or WKD lookup would tell a third party who you
correspond with -- the leak the image proxy exists to close.
- SHA-1. Not forgeable in practice today, still not something to put a
tick beside.
- RSA-PSS, whose salt length lives in parameters this does not read.
Guessing wrong would report a good signature as bad.
Nothing validates a chain: no CA bundle is shipped and revocation is not
checked. "Issued by" reports what the certificate claims, and a
self-signed one claims itself.
The DER, CMS, X.509 and MIME readers are hand-written and deliberately
narrow -- no new dependency, and the whole verifier is a lazily imported
8.6 kB chunk that a reader of unsigned mail never downloads. The one
place this is easy to get quietly wrong has its own function and its own
test: signed attributes are signed as a SET OF, not as the [0] IMPLICIT
they arrive as, and hashing the message instead would make every
signature "pass".
Tested against real `openssl smime -sign` output rather than hand-built
fixtures -- RSA, ECDSA, a tampered copy, and a valid signature by a
certificate for somebody else -- because a signed message written by
hand only agrees with whatever its author believed the format to be.
Also driven in a browser against the mock, which now serves three real
signed messages so every branch of the banner is reachable.
Translations: 34 new strings in all nine catalogues, 306 entries.
Falling back to English is unchanged at 24 per language.
dependabot[bot]
jcoffey-dev