Ask for the account password before minting an app password, and keep
sessions the proxy checks from writing the account's own registry objects,
so a session left open on someone else's machine cannot take a credential
away from it. The password is compared with what the session holds; Stalwart
is asked only when 2FA moved the session onto an app password.
Serve attachments and proxied images with no-store on a device that is not
the person's own. Give files from a winmail.dat only the types the server
would show inline. Strip direction controls from sender and attachment
names and from saved filenames.
On signing out, send what is inside its undo window, then close every
composer, so the next person to sign in does not find the last one's draft.
Group sessions by the account Stalwart names and its server, so "sign out
other sessions" also reaches a session opened as a bare or differently
cased username.
Cap JSON bodies at 64 KB on every API route except JMAP and uploads, which
bound themselves. Sign-in used to read a body of any size before its rate
limits ran; the flood ceiling now also runs before the body is read.
For sessions whose JMAP requests are checked, lower the read cap from 16 MB
to 4 MB, allow four such reads per session at once, and turn requests away
with a 503 once 32 MB is held across everyone.
Count sign-in limits per /64 for IPv6, since one host holds a whole /64.
Bind the compose example to loopback, and run it read-only with no
capabilities and no-new-privileges. Keep .env.* out of git and the image
build context.
server/src/mock/index.ts was 1,545 lines, the largest file in the repo.
It had carried `/* ---------- data ---------- */` style markers for a
while, so the seams were already drawn; this turns six of them into
files.
mock/config.ts 51 env-derived constants, `account`, `state`
mock/data.ts 410 fixtures and the builders that make them
mock/engine.ts 442 the generic JMAP machinery -- get/set, filters,
patches, refs, limits, recurrence plumbing
mock/handlers.ts 455 the `Method/name` dispatch table
mock/events.ts 26 SSE fan-out and the Email/changes ring buffer
mock/auth.ts 11 checkOtp
mock/index.ts 195 HTTP routing, the session document, listen
TWO THINGS THAT COULD NOT JUST MOVE:
`counter` and `vacation` were module-level `let`s written from both the
fixture builders and the handlers. An ES module can export a `let` and
importers see it update, but they cannot assign to it, so both became
containers: `seq.counter` and `vacationBox.current`. Seven call sites.
`recordEmailChange`, `broadcast`, `sseClients` and `checkOtp` lived in
the HTTP section, but the handlers call them -- and index.ts imports the
handlers. Leaving them there is a cycle, so they became events.ts and
auth.ts rather than being dragged into data.ts, which is fixtures.
`account` is still exported from index.ts, because account.test.ts and
login-guard.test.ts reach for `mock.account` and `mock.server`.
Verified by running it, not only by compiling it: `npm run mock` boots
and listens, `/.well-known/jmap` returns a session, and a POST to
`/jmap/` answers Mailbox/get with the nine seeded folders and
Email/query with the seeded messages.
The dashboard's link to Stalwart's own administration needed
STALWART_ADMIN_URL, which an operator had to know to set. Everything it
holds can be read from the server:
- the origin is the host Stalwart advertises in its own session URLs, the
one people reach it at even when ihasmail talks to it on a private
address;
- the prefix is where its web interface application is installed. The
production server's x:Application reads "Stalwart Web Interface",
enabled, urlPrefix {"/admin", "/account"}, and /admin is also what
Stalwart writes at first boot.
So, for an account that administers, the account info fetch now asks the
account's own server for its applications and links to origin + /admin/.
An installation whose web interface is disabled or moved gets no link; an
administrator who may not read applications gets Stalwart's default
/admin. It is cached with the rest of the account info.
STALWART_ADMIN_URL and a servers file entry's adminUrl still win, for an
administration that lives somewhere else. A routed domain without one now
takes what its own server said, never the default server's.
On a server that is not Enterprise the Tenants page is still only the
notice. On Enterprise the notice is gone -- a real installation that has
tenants has the licence -- unless SHOW_ENTERPRISE_NOTICES=1 asks for it above
the list. The public demo will set it: it reports Enterprise so tenants can be
shown, and should not suggest they come without the licence. The setting
reaches the browser as session.ihasmail.server.enterpriseNotices.
A run on the production server with throwaway tenants, a role, lists and a
domain, all removed, found three things the source reading had not:
- Something in a tenant has to be on a domain in that tenant (a list in a
tenant on an unassigned domain is invalidForeignKey), while something in
no tenant may be on a tenant's domain. The account panel's tenant choice
offered every tenant; it offers only the domain's now, and a new account
starts in the tenant of the domain it is made on. The domain list reads
memberTenantId for it.
- A domain created in a tenant puts its DKIM keys there too, and they keep
the tenant from being deleted. They are counted with the rest, so Delete
is not offered while any remain.
- Stalwart lets a domain leave a tenant while the tenant still has accounts
on it, stranding them. The panel asks first and refuses while any are
there.
The refusal to delete a tenant that holds anything was confirmed, as were
tenant create, quota pointers, logo and rename. The mock follows the domain
rule, filters DKIM keys by tenant, and KNOWN-ISSUES records the run.
The non-Enterprise notice is now just "Tenants are a Stalwart Enterprise
feature." Two sentences were reworded and one plural added, in all nine
catalogues, and the old sentences are gone.
On a server that does not report Enterprise -- or reports no edition --
tenants hold nobody to anything beyond an ordinary user's permissions, so
the page is the notice alone: no New tenant, no search, no list, and no
tenant query is made. The mock's edition is MOCK_EDITION now (default oss,
as before), so MOCK_EDITION=enterprise brings the section back to work on.
A tenant is a separate organisation on one server: its own people,
domains and limits, and an administrator who manages only what is in it.
It gets a section under Access, gated by sysTenantQuery and sysTenantGet,
with a notice on a server that does not report Enterprise, where anyone
inside a tenant is held to an ordinary user's permissions.
The panel edits the tenant's name, logo, role and limits. The logo is an
https address, drawn through the image proxy the strict image policy
requires, or an image data URL. Limits change one quotas/<name> pointer
each, so the four ihasmail does not offer keep their values, and an empty
field is no limit. The role is the most anyone inside can be allowed.
Stalwart keeps no list on a tenant -- each account, group, domain, list and
role names its own -- so what a tenant holds is counted with memberTenantId
queries and shown against its limits. Domains are added and taken out from
the tenant's panel, one memberTenantId change each; only a domain in no
tenant can be added, and its accounts stay where they are. Delete is offered
once every count reads zero.
A tenant does nothing until someone administers it, so the account panel
gains a Tenant choice for an administrator who can read tenants: an
Administrator inside a tenant administers that tenant. Nobody moves their
own account.
The mock has a tenant holding a domain and an administrator, a spare domain
to assign, memberTenantId filters on every query, and Stalwart's rule that
only an administrator outside every tenant may move things into one. A test
of taking a domain back out found that the mock's pointer handling dropped a
top-level null instead of storing it, so nothing had ever been cleared that
way; it stores null now, as the server reads it back.
Nothing about tenants has been written on a live server: production has
none. KNOWN-ISSUES says what was read from source.
Thirty-nine new strings and one plural, in all nine catalogues.
A throwaway role on the production server confirmed the create shape,
one-pointer changes to permissions, bases and name together, the grant
refusal, and the in-use refusal when another role builds on it. It also
showed that a permission name Stalwart does not know fails the whole
update -- which is how jmapEmailSet, carried by the mock since Accounts
was built, turned out not to exist. The mock uses jmapEmailUpdate and now
refuses unknown names against the 0.16.22 snapshot.
Some permissions an administrator holds are never listed by /api/account
(sysLogCreate was granted without complaint), so the picker locks their
Allow; KNOWN-ISSUES says so.
A role is a named set of permissions given to accounts, groups and
tenants. It gets its own section under a new Access heading: every role
listed with the permissions it grants once its bases are followed, and a
panel to create, edit and delete one.
A role builds on others and has everything they grant; a denial anywhere in
the tree wins, which is how Stalwart resolves it (permissions.rs unions
enabled and disabled across the tree, then subtracts). The picker is
Stalwart's own list of permissions, under its headings, searchable and
filterable to what is granted or set here. Each permission is not set,
allowed or denied, and one that is inherited says which role it comes from.
Only permissions the viewer holds can be allowed, because Stalwart refuses
the rest, and a role carrying anything the viewer lacks opens read-only with
no delete, because Stalwart checks a grant but not a delete. Saving sends a
pointer for each permission and base role that changed.
The roles Stalwart hands out by default, read from x:Authentication, say so
before they are changed and cannot be deleted here; a role still in use is
kept by the server, and the refusal names what uses it.
The permission list is Stalwart's schema. A new route, GET
/api/admin/permissions, fetches /api/schema as the signed-in account and
returns only names and labels, behind the same two gates as the registry
methods and held in memory for an hour. Its labels are English only, so
every one of the 661 has a translation in each of the eight other
languages, in its own file keyed by permission name and loaded only when
Roles opens. A permission a later Stalwart adds shows its English label. A
test holds every language to the 0.16.22 snapshot: nothing missing, nothing
stale.
The mock answers x:Role/set with the grant check, loops and in-use
refusals, reads the defaults from x:Authentication, and serves the schema
gzipped as the real one is.
Fifty-two new strings and two plurals in all nine catalogues, and 661
permission labels with 59 headings in each of the eight translations.
A mailing list is an address that passes mail on to everyone on it. To
Stalwart it is its own object, x:MailingList, behind sysMailingList*, so
it gets its own section under Directory after Groups: search, fifty to a
page with each list's recipient count, and a panel to create, edit and
delete one.
Recipients are a property of the list, so unlike a group's members they
save with the rest of the panel. What Save sends for them is only what was
added and removed, one recipients/<address> pointer each -- the patch the
live server accepted -- so a recipient added elsewhere while the panel was
open is not taken out. They can be pasted several at a time, from a
spreadsheet column, a comma-separated line or Name <address>; anything with
an @ that is not an address stays in the box with a note. Past a dozen, a
filter narrows them.
That is all a list is in Stalwart -- no owners, moderation or posting
rules -- so that is all the panel offers.
The mock answers x:MailingList with two lists, the recipient set's live
shape, and the refusals a wrong address, a clash with an account and a
missing permission get.
Twenty-five new strings and one plural, in all nine catalogues.
A throwaway group on the production server confirmed what the code was
built on: a Group account with Default roles, membership as a pointer on
the member, the members query, and a delete refused while a member still
names the group. Its objectId is an {object, id} pair rather than a bare
id; the mock answers the same way now.
A group is a shared address and mailbox and the people who share it. To
Stalwart it is an x:Account of type Group, behind the same sysAccount*
permissions as a person, so it sits under Directory beside Accounts:
search, a page of fifty with each group's member count, and a panel to
create, edit and delete one.
Membership lives on the member, not the group. Members are the users whose
memberGroupIds name it, and adding or removing one is a single
memberGroupIds/<group> pointer on that user's account -- true or null --
which leaves their other groups alone. Changes apply straight away rather
than riding on Save, so the list is always what the server has. Nobody can
add or remove themselves, the same line the account panel draws at one's
own role.
A group's role is Default or Custom, not a person's User or Admin, and it
is what the group may do: in 0.16 a user's permissions come from their own
roles only, and a group gives its members what is shared with it. Only
roles the viewer could grant are offered.
Delete takes the members out first and then deletes the group, the order
a domain's keys go before the domain, because the registry keeps anything
another object names. A role that cannot change the members' accounts is
not offered a delete it could only half finish.
The mock's groups had a person's roles, accepted a memberGroupIds filter
without applying it, and answered a linked delete with the wrong shape;
all three follow the source now, and it refuses nested groups and
memberships of things that are not groups.
Nothing about groups has been run against a live server yet: production
has none, and every operation is a write. KNOWN-ISSUES says what was read
from source.
Thirty-five new strings, two of them plurals, in all nine catalogues.
A line under the cards says where the rest is: detailed metrics, the
delivery queue, logs and server settings are in Stalwart's own
administration. It links there when the operator sets STALWART_ADMIN_URL,
and stays plain text otherwise, because STALWART_URL is how this server
reaches Stalwart and is often an address no browser can open.
Several servers: a servers file entry may now be an object,
{"url": ..., "adminUrl": ...}, and a session routed to that server gets its
adminUrl. A routed domain without one gets no link rather than the default
server's, for the same reason routing never falls back. The URL is sent
only to a session that may administer.
The shipped example file stopped the server at startup: its "_comment"
key was read as a domain and refused as not a URL, while the test that
checks the example skipped it. Keys starting with an underscore are notes
now -- no mail domain starts with one -- and the example is also loaded
through the real parser in a test, so the two cannot disagree again.
Two new strings, in all nine catalogues.
Administration used to open on its first section. It opens on a grid of
cards now: users, domains, messages waiting in the delivery queue, server
memory, and the last 24 hours' received and sent. Each card is there only
when the role holds what its number needs -- a count is a query, the
metric history a query and a get -- so a helpdesk role that reads accounts
and domains sees those two cards and nothing about the server.
What the cards count is whatever Stalwart answers for the signed-in
account, which scopes a tenant administrator's accounts, domains and queue
to the tenancy. The metric history has no tenant in it, and Stalwart's
Tenant Administrator role does not hold it, so a tenant's dashboard is
users, domains and pending.
The history is Enterprise-only and switched off by default. A server that
refuses it leaves those cards off; one that records nothing says so rather
than showing zeroes. Received and sent add up the queue counters Stalwart's
own dashboard uses, filtered with the comparison names the live server
accepts (a bare timestamp is unsupportedFilter). The column count follows
the number of cards so rows stay even, and falls back by the grid's own
width rather than the window's.
The server's test for whether an account is offered Administration matches
the client's again, now that a count is enough. The mock answers the queue
and an hourly history ending in the current hour; MOCK_METRICS=off refuses
the history as Community does, a tenant administrator gets the queue, and
helpdesk reads domains, as the demo's does.
ROADMAP and FEATURES said reporting and queues were out of scope; they say
the dashboard reads a handful of numbers and that managing queues, logs
and settings stays out. KNOWN-ISSUES records what was settled on the live
server and what was only read from source.
Fourteen new strings, in all nine catalogues.
0.16.22 changed four things a client sees from CalendarEvent/get and
ContactCard/get. Read from its source and the tests that came with it:
- baseEventId is the master's id on a synthetic id and null otherwise; an
event read by its stored id used to report its own id. A one-off from an
expanded query still has a synthetic id, so it still carries a base.
- recurrenceRule and recurrenceOverrides named on a synthetic id come back
null rather than absent.
- useDefaultAlerts is the reader's own and reads false until set.
- an empty properties list returns id alone, for both methods. pick already
did that, so only a comment changes for contacts.
With properties omitted the stored object comes back as before.
The README said a one-off now carries a null base, which only holds for one
read by its stored id; it now says that, and that the mock follows.
Stalwart explains a refused change in English, and several of its words
reached the page as they were: "Invalid domain name" for a reserved TLD,
"Invalid email address" for a catch-all, a grant refusal, and ihasmail's own
proxy messages. Every registry error type now has its own message, and a
value one of the registry's string validators refused is recognised by the
validator's wording and explained again. A domain clash or a missing domain
is worded for a domain rather than an account.
The one exception is kept on purpose: a password policy's reason follows a
translated sentence, because the rule is the server's and dropping it would
leave no way to find out why.
The mock now refuses a reserved TLD and a catch-all without a domain the way
the live server did. KNOWN-ISSUES records the fix, and that the last two live
cases -- an administrator-set password and the outranking guard -- held.
15 new strings in all nine catalogues, 3 retired; strings falling back to
English stay at 16.
A session signed in without "This is my own device" can no longer
administer. The server withholds the account's permissions from it and the
JMAP proxy refuses registry methods beyond the account's own, the same gate
ADMINISTRATION=0 uses. A borrowed or shared machine is where nobody should
be able to reset a password or remove a domain.
An administrator in such a session still sees Administration in the account
menu, greyed out, with the reason and the fix: sign in again with the box
ticked. The server tells that session only that the account administers.
The gate now reads the body only when it could name a registry method --
"x: in the text, or a \u escape that could spell one -- so ordinary mail
traffic from an untrusted session is forwarded untouched.
1 new string, translated in all nine catalogues, quoting each language's own
label for the tickbox; strings falling back to English stay at 16.
The Accounts list sent x:Account/query with {"type": "User"}, and a live
0.16 server refuses it: "unsupportedFilter - type". The registry keys a
filter by the property's name on the object, which for the discriminator is
@type, so the whole list failed to load. {"@type": "User"} is accepted.
The mock took the wrong name without complaint, which is how it shipped. It
now refuses any filter name the real server does not index for that object,
answering the way Stalwart does.
A role that can read domains now finds a Domains section beside Accounts:
list and search with each domain's account count and whether its DNS, DKIM
and certificate are managed automatically; add a domain; edit its
description, other names, catch-all address and plus addressing; copy its
DNS records one at a time or as a zone file; see its DKIM keys and their
stage; and remove it once no accounts use it.
The records come from the zone file Stalwart computes per domain. A long
DKIM record, which the BIND serialiser splits into quoted chunks, is joined
back into the single value a DNS provider's form wants.
Removing a domain takes its DKIM keys first, in the same request, because the
server will not remove a domain its keys still name. Removal is not offered
while accounts use the domain, or when the role cannot remove the keys.
The Administration nav is now built from the sections the role can read, and
the menu appears when there is at least one. The mock gains domains, DKIM
keys and zone files.
61 new strings, translated in all nine catalogues; strings falling back to
English stay at 16.
ADMINISTRATION=0 at launch removes in-app administration for everyone. The
account's permissions are no longer sent to the browser, so the menu never
appears, and the JMAP proxy refuses Stalwart registry methods other than the
account's own (settings, password, app passwords, API keys, public keys,
masked addresses). Hiding the menu alone would have left an administrator's
browser console able to make every call the menu made.
With administration on, the request body streams through untouched as before;
only an installation that turns it off reads and checks the body, forwarding
the parsed form so the server receives exactly what was inspected.
An account whose Stalwart role manages accounts now finds Administration in
the account menu. It lists, searches, creates and edits accounts -- display
name, other addresses, role, storage limit -- sets a new password, and
deletes, each offered only when the role holds the matching permission.
The server keeps the permissions list from GET /api/account, which it
already called for the edition and threw the rest away. Everything else is
JMAP x:Account, x:Domain and x:Role calls through the existing /api/jmap
proxy, so nothing new is stored and Stalwart decides every call.
Stalwart checks a grant against the caller's permissions but not a password
change or a delete, so an account that outranks the viewer is shown
read-only. Your own password is changed in Settings, which re-seals the
session; changing it here would strand it.
The mock server gains a directory behind the same permission names, with
MOCK_ROLE choosing admin, tenant-admin, helpdesk or user.
68 new strings, translated in all nine catalogues; strings falling back to
English stay at 16.
Two requests ignored the domain mapping from #238 and went to STALWART_URL:
- /api/account/* re-fetched the upstream session without upstreamFor(), so
once the five-minute session cache expired, password, app-password and
2FA calls for a mapped domain reached the default server.
- The locale lookup resolved Stalwart's apiUrl against the default server
rather than the one that issued the session.
Both now use the session's own server, with a test pinning the second.
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.
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.
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.
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.
One ihasmail in front of several Stalwarts, from #238. STALWART_URL stays
required and stays the default, so an installation that sets nothing behaves
exactly as it always has -- the mapping only adds domains that go elsewhere.
An unlisted domain goes to the default. So does a bare username, which
Stalwart accepts and which has no domain to map at all.
A listed domain never falls back. If its server is unreachable that sign-in
fails rather than retrying against the default, because falling back would
authenticate somebody against a server their domain was deliberately routed
away from -- and if the same account name existed there, they would land in
another tenant's mailbox. The fallback is a decision about unmapped domains,
taken before any network call, not a recovery path.
Smaller than it sounds because only four places read config.stalwartUrl, all
in upstream.ts. The upstream session now records which server issued it, since
the relative URLs inside it only mean anything against that server, and every
route already holding a session gets the right upstream without a second
lookup. The client is untouched: it talks to one proxy and never learns there
is more than one server behind it, which is exactly why this is small and
several-servers-at-once is not.
The upstream is derived from the username rather than stored on the session,
so a mapping change takes effect on restart instead of being frozen into
sessions that outlive it.
Validated at boot the way the settings policy is: malformed JSON, a duplicate
domain once normalised, a missing file or a value that is not an http(s) URL
all stop the server. Domains are lower-cased and stripped of a trailing dot,
because that is how one arrives off a username and comparing them any other
way means a mapping that silently never matches. The servers themselves are
not contacted -- a mapping is a routing table, not a health check, and one
customer's outage must not stop ihasmail starting for the other four.
Eight tests on the routing, two on the shipped example, and the four refusals
checked by hand against a real config load.
ihasmail runs in its own container, usually on its own host, so Stalwart being
briefly unreachable is an ordinary Tuesday. Sign-in handled it almost right:
a 401 is invalid_credentials, a timeout is 504 and anything else is 502, none
of which reads as a rejected password.
What it got wrong was the counting. RateLimiter.check() consumes an attempt
when it is called, and it is called before the upstream is contacted; reset()
only runs on success. So every try against an unreachable server burned a
credential attempt, and after ten of them the person was locked out for the
rest of the fifteen-minute window -- including after the server came back. A
thirty-second blip became a quarter-hour lockout, and the second failure was
entirely ihasmail's own doing.
A 401 is a judgement about the password and stays counted. A 502 or 504 is the
upstream failing to answer, says nothing about the credentials, and is now
refunded -- one attempt back, not the key cleared, so a run of real failures
with an outage in the middle still adds up. The old-server refusal refunds too:
those credentials were accepted.
Both guessing keys are refunded, not just the username one. Refunding only
that would not have fixed it -- ten retries still spend the per-address budget,
and behind one office NAT that budget belongs to the whole building, so a
company-wide outage would lock out the company.
Which needs a backstop, because "not counted" must not mean "unlimited": each
attempt still costs an outbound connection that may sit there until
UPSTREAM_TIMEOUT, and an outage is the one moment the endpoint is cheapest to
abuse. So there is a second ceiling per address, twenty times looser and never
refunded. A person retrying will not come near it; something hammering will.
Both messages now say the quiet part -- "This is not a problem with your
password" -- for somebody already worried they have forgotten it.
Closes#239.
#231 added the policy but nothing to copy. The repo already answers this the
same way four times over -- Caddyfile.example, deploy.example.sh,
nginx.example.conf, .env.example -- and the new feature was the one thing
configurable here with no example beside it.
settings-policy.example.json carries all three sections with the reasoning in
it, including the part worth being deliberate about: a `changes` entry
overrides a decision a reader has already made, and if you want it to stay put
regardless that is `enforced` instead. JSON has no comments, so the commentary
is in `_`-prefixed keys, which is safe because the server reads three names and
ignores everything else.
A test asserts the shipped example stays valid against the rules the parser
enforces -- unique versions, settings objects, no comment key colliding with a
real section. An example that has drifted is worse than none: somebody copies
it, the server refuses to start, and the first experience of the feature is a
crash loop.
.env.example gains the four variables, commented out, with the file form and
the inline form and the note that the file wins over the variables.
Confirmed against the real image on the deploy host rather than reasoned
about: an immutable container -- --read-only, IMMUTABLE=1, SESSION_FILE=
empty -- starts and serves the policy both with a read-only file mount and
with the environment variables alone. The feature costs nothing in
immutability, because the only thing it writes is the applied-changes stamp,
and that goes in the reader's own settings file on Stalwart like every other
setting.
The last third of #207, and the only part that remembers anything.
An admin turns a setting on for people who are already here -- which a default
cannot do, since a default only seeds an account that has none -- and readers
may still turn it back off afterwards, which enforcement does not allow. The
difference between the two is entirely in the remembering.
Each change carries its own version, and an account stores the ones it has had
in its own settings file. Ids rather than a high-water mark, so a change dated
earlier than one already applied is not silently skipped -- the reporter's
analogy is a schema migration, and this is that shape.
Per account rather than per device, because ihasmail's settings are not
browser-local: they live in a file in the reader's own JMAP Files, with the
browser holding a cache. Signing in on a phone does not apply everything a
second time.
A change reaches somebody who had already decided otherwise. That is intended
and confirmed on the issue: the point is to reach everybody who is already
here. It is applied once, and their next decision sticks.
One `update` for however many are pending, since each would otherwise push a
settings file of its own. Enforced values still win, being applied after. A
change whose settings this build does not have at all is dropped rather than
recorded, or it would never run on the ihasmail that does have them.
The reader is told. A setting moving under somebody without a word is the part
of this worth being uneasy about, so the count is toasted with a way into
Settings.
README gains the Docker half the user asked for: a mounted policy file, the
same thing as environment variables for a deployment with no volume, a compose
fragment, and the fact that a policy is read once at startup so editing it
means a restart.
Closes#207.
The first two thirds of #207. A school wanting "warn about outside senders"
on for three thousand pupils cannot ask three thousand pupils, and the
reporter is right that this is a company policy rather than a preference.
Two powers, and the difference between them is the whole request. `defaults`
seed an account that has never had settings of its own and can be changed
afterwards like anything else -- a starting point, not a rule. `enforced` are
reapplied on every load and cannot be changed at all.
Enforced controls stay visible and go dead, with a line saying why. The issue
asked for that by name: a control that is simply missing reads as a bug to
somebody who has used ihasmail without a policy.
The lock is in the settings store rather than only on the controls. There is
one door -- `update` -- and putting it there means an imported settings file,
a settings file synced from a device that predates the policy, and a control
somebody adds later and forgets to check are all covered by construction.
Reset goes back to the installation's answer rather than to ihasmail's, so it
cannot be a way around a policy either.
Configured by environment variable or by a file, because ihasmail's own
production runs read-only with no volume: an installation that cannot mount a
file can still set a variable. Keys this build does not have are dropped, the
same rule an imported settings file already gets -- a policy written against a
newer ihasmail must not put a setting nothing reads into everybody's synced
settings file. Malformed JSON stops the server rather than quietly doing
nothing, since a policy that silently did not apply is indistinguishable from
the feature not working.
Tier three -- enforcing a setting once while still letting readers change it
afterwards -- is not here. It needs a decision the reporter and I have not
made yet, and it is the only part that stores anything new.
Refs #207.
Dragging an event across the month grid wrote the date of the cell it
landed on into the event's stored start. Those are the same date only
while the event's time zone is the reader's.
An event kept in Asia/Tokyo at 15:00 is drawn to a reader in Phoenix at
23:00 the previous evening. Dropped on the 11th, it was written as the
11th in Tokyo -- which is the 10th on screen. It went where its own
calendar said rather than where the pointer did, one day short, every
time.
Moving by the difference between the two local days instead moves it
exactly as far as the hand did, and adding whole days to a stored wall
clock leaves the time of day alone without touching the zone -- so the
frame the rest of this path is careful about is still not crossed.
Found by giving the mock an event in a zone that is not the machine's.
Every other fixture used the machine's own, which cannot tell a correct
conversion from no conversion at all: the case that works is the one the
fixtures were all testing.
A timetable, a rota, a public holiday list: the calendars people are given
as a link, which ihasmail could not show at all.
Nothing is stored. The document is fetched when the calendar is opened and
parsed in the browser; the server keeps no copy, no cache and no schedule,
which is what lets an immutable container serve this. There is no timer
either -- there is nowhere to run one -- so the guarantee is that a
subscription is as current as the last time somebody looked, which is also
when it matters. That is said plainly rather than implied.
The fetch has to happen on the server: a calendar URL belongs to whoever
published it and almost none of them send CORS headers. That makes it the
second place this app knocks on a door somebody else chose, so the guard
the image proxy has always had was lifted out and both now call it. A
second SSRF implementation is how one of them ends up missing a case; this
way there is one, and the extraction is covered by the image proxy's own
tests still passing unchanged.
webcal: is understood, because that is how these are published, and it is
read as https: rather than waved past the checks -- a webcal URL pointing
at loopback is refused exactly like an http one.
Recurrence is deliberately not expanded. RRULE is a small language with a
lot of edge cases, and a subscription quietly showing the wrong dates would
be worse than one showing the first occurrence and saying so.
The parser is a subscription parser rather than an importer: a subscribed
calendar is read-only and redrawn from scratch each refresh, so nothing has
to round-trip or survive an edit, which is most of what makes a full
iCalendar implementation large. What it does have to do is never mis-state
a time -- a DATE is built in local time rather than at UTC midnight, which
would land on the day before for anyone west of Greenwich -- and never hang
on a document somebody else wrote.
Events go through instancesIn like the birthdays, so no view has to know
they are not real calendars, and the calendar they hang off reports no
write rights, so everything that asks before offering an edit declines on
its own. A subscription that cannot be read says so in the sidebar rather
than drawing an empty calendar, which looks like a calendar with nothing
in it.
Newest-first was the only order, so the mail you had not read yet was
wherever it happened to fall.
Seven presets and up to three levels of your own. It covers the Inbox
alone by default: unread-first is what people want in the folder they
triage and confusing in Sent, where everything is read and the order that
matters is when it went. Search keeps newest-first whatever the setting
says, since a result list is already ordered by the question that was
asked.
The server does the sorting, over the whole folder, for the same reason
search runs there: a list sorted in the browser is sorted only as far as
the browser has loaded, which on a folder of ten thousand is the first
fifty and a lie about the rest.
Two details that are easy to get wrong and were worth pinning in tests.
hasKeyword sorts a boolean and false comes before true, so "unread first"
is $seen ASCENDING while "starred first" is $flagged DESCENDING -- the
other way round. Getting either backwards puts exactly the mail you were
looking for at the bottom. And every order ends with newest-first as a
tiebreak, because a sort whose last level is a keyword or a subject leaves
every tie undefined, and an undefined order changes between two looks at
the same folder for no reason the reader can see.
Sorting on a keyword is optional in RFC 8621, and a server that will not
do it fails the whole query rather than degrading it -- so this setting
could turn a folder into one that does not open. The refusal is caught
once, the keyword levels dropped and the query retried, and nothing is
said: the reader asked for an order and got the closest the server can
give, and a toast on every folder change would be the app complaining
about its own request.
The mock now honours the sort instead of always answering newest-first,
which had it reproducing a server that silently returns a different order
from the one asked for -- the one shape of wrongness a client cannot
detect. MOCK_NO_KEYWORD_SORT=1 reproduces a server that refuses the
keyword sorts, so the fallback can be developed against.
The dates were already on the contact cards and nothing ever showed them,
so the one thing a birthday is for -- noticing it in time -- was the one
thing the app could not do with it.
Derived, not stored. The dates stay on the cards: a second copy of the
same fact drifts the first time somebody corrects one, and keeping a
calendar of its own is exactly what ihasmail does not do. Entries are
generated when a view asks for a range and vanish when the contact does.
They go through instancesIn like everything else, so no view has to know
they are different.
Off until switched on. It is derived data, and a calendar that fills
itself with dates nobody put there is a surprise rather than a feature. It
can also be hidden from the calendar's own sidebar without being turned
off, which is the same distinction the shared calendars already draw.
They cannot be edited or deleted, and that falls out of the design rather
than being special-cased: the virtual calendar reports no write rights, so
every control that already asks before offering Edit or Delete declines on
its own. updateEvent and destroyEvent refuse a synthesised id as well, so
the store is safe whatever calls it -- including anything added later.
Two things about the dates themselves. A card that records only a day and
month is the common case rather than the exceptional one, and gets a
birthday with no age rather than no birthday. And 29 February falls on the
28th in a year that has no 29th: somebody born in February has a birthday
in February, and moving it into March is the arithmetic winning over the
fact. Both are conventions; these are the ones that keep the fact intact.
The mock now carries birthdays on most of its contacts, including one with
no year and one on 29 February, so both cases are visible without a real
address book.
Outlook sending in Rich Text packs every attachment into one TNEF blob.
Every other client shows a single unopenable winmail.dat, and the files
inside it are gone as far as the reader is concerned -- which is a
decoding problem rather than a mail one.
Written from the published format: a signature, a key, then a flat run of
attributes, each one a level byte, a 32-bit id carrying its own type, a
length, the data and a checksum. Attachments are delimited by
attAttachRenddata rather than named, which is why the parse is a small
state machine.
The MAPI property stream inside attAttachment is read for two properties:
the long filename and the MIME type. attAttachTitle carries an 8.3 name,
so a file that arrived as "Quarterly Report Final.docx" is QUARTE~1.DOC
there and correct here. The stream stops at a named property (id >=
0x8000) rather than guessing past it, since those carry a GUID before
their value and nothing after one can be trusted to stay aligned.
Decoded in the browser, on request. The server never sees the contents and
has nowhere to keep a decoded copy; doing the work on sight would spend
the bandwidth whether or not anybody wanted what is inside.
A blob that goes wrong part-way through keeps what was read before that
point, whether it ran out or the checksum stopped matching. Half the
attachments beats none: the alternative is a reader who can see the file
is there and cannot have it. The original stays attached either way.
The message body is deliberately not decoded. TNEF can also carry it as
compressed RTF, which is a second format again for a body the reader
already has in plain text or HTML nine times in ten.
The mock now sends one, built by its own encoder rather than by the
parser's fixtures, so the two are independent implementations of the same
description.
`BASE_PATH=/mail` mounts the whole app under a prefix, for a host that is not
ihasmail's alone. Unset -- every deployment that exists -- is the domain root
and is byte-for-byte what it was: the canonical form of the setting is the
empty string, and `""` concatenated onto `/api/health` is `/api/health`.
That choice of canonical form is the whole design. A trailing slash would have
been the obvious alternative, and it fails quietly in exactly one place: at the
root it makes `//api/health`, which is not a path on this host but a
protocol-relative URL to a host called `api`. One call site forgetting to
branch is a request leaving the origin. So the empty string, one leading slash,
no trailing one, worked out once in `scripts/basePath.mjs` -- plain JS, next to
`version.mjs`, because the web build and the server both have to reach the same
answer and two implementations of "what does /mail/ mean" is precisely the bug
where the server serves an app whose script tags point somewhere else.
`/mail`, `mail`, `/mail/` and `//mail//` all mean the same mount; a deployment
should not fail over a trailing slash.
Unlike everything else ihasmail is told, this one cannot wait for the process
to start. The bundle writes its own asset URLs into index.html, so `BASE_PATH`
is read at build time for Vite's `base` as well as at run time for the routes,
and the Dockerfile carries one value into both. Get them out of step and the
page comes up blank with a 404 in a console nobody has open -- so the static
handler, which is reading index.html anyway, checks what it asks for and says
so in the log once per build.
Everything moves together. The API mounts at `${base}/api`; the router is
given the base once, so every `<Route path>` and `<Link href>` stays written
root-absolute and wouter does the rest; `apiFetch` adds the prefix in one place
rather than at forty call sites; the session cookie's Path narrows to the mount
so two instances on one host cannot sign each other out.
Two things need no prefix at all, and it is worth saying why they were not
given one. A manifest's members resolve against the manifest's own address, so
relative URLs there follow the mount with nothing substituted at build time --
which is also why `public/` needed no template step. The service worker is the
same trick: it is served from the mount, so `new URL("./", self.location)`
tells it where that is, and a worker that derives the value cannot disagree
with the page that registered it.
Anything outside the mount is a 404 rather than the app shell, and
`stripBasePath` does not use `startsWith` -- under `/mail` this process shares
a hostname, and answering `/mailbox` with our index would shadow a neighbour
instead of letting it 404 honestly. For the same reason the notification-click
handler now checks the path as well as the origin: `includeUncontrolled` widens
`matchAll` to the whole origin, which off the root would have navigated a
stranger's tab to our inbox.
Inline images in a draft were the one silent trap. They are matched by their
blob URL on the way out, once unanchored and once anchored, and a bare
`/api/blob/` still appears inside `/mail/api/blob/...` -- so one pattern would
have replaced the tail and left `/mail` in front of a `cid:`, and the other
would have missed and sent the message linking to the sender's own webmail.
Both patterns are built from the base now.
jcoffey-dev
jcoffey