Fetching https://<site>/favicon.ico directly would resolve a DNS name - and
open a connection - to EVERY domain kept in the vault. The list of sites where
someone holds an account is precisely what this vault protects; handing it to
the network for a sixteen-pixel drawing would be a poor trade. The relay sees
only that someone is interested in one domain, which is why the call is asked
for entry by entry and never happens on its own.
Two tries: the full host, then the second-level domain - chat.qwen.ai has no
icon at the relay, qwen.ai does. The type comes from the first bytes, never
from the extension: the relay answers PNG under a .ico address.
A dedicated patch rather than PatchEntryField, because icon_custom is a NUMBER
in the snapshot and writing it as a string would give "0" - which the desktop
reads with GetValue<Boolean> and would choke on. It is set to zero: the icon
comes from an automatic lookup, not from a choice, and that flag is what stops
the desktop from ever replacing a chosen icon during a refresh.
Base64 without line breaks. Delphi's encoder wraps at 76 characters by
default, a wrapped data: URI does not display, and that trap has already been
paid once on the containers.
Three things of different natures lived in collapsible panels on the home
screen - repository credentials, the unlock code, the log - on the one screen
where one mostly wants to open one's vault. They move to a settings sheet,
reachable from the list too.
A sheet, not a fourth view of the rail: each view costs its transform rules
and a place in back-key navigation, and settings is not somewhere one moves
about - it is a parenthesis one opens and closes.
The unlock code can now be REMOVED. It could not before: clearing the field
meant "nothing typed", not "drop the protection", so one could set a code and
never get out. It requires an open vault, and that is the proof being asked
for - the cache only let itself be read under the right code. Without that,
anyone holding the unlocked phone could lift the protection without knowing
it.
The refusal that guards every other path stays: it is what stopped a sync
from silently undoing a code. Removing is the one deliberate way out, and it
says so at the call site.
The old fields stay in the document as the source Delphi knows; the sheet is
only a view of them. Emptying them would lose the configuration at the first
sync.
A probe every ten minutes compares the remote fingerprint to ours and lights
an indicator. It does not pull: adopting on its own while someone is reading
an entry would change it under their eyes. The user decides when.
The body is downloaded anyway - Gitea's contents API cannot answer without it
- but it is thrown away undecrypted: no derivation, no write, just a string
comparison. The cost is network, not time or battery.
It rides the idle timer rather than a second one. That timer already beats
every 15 seconds for auto-lock; counting its beats costs an integer, where
another timer costs an object, a lifecycle, and one more chance to forget it
on shutdown.
Two indicators, two colours, two gestures. Orange: things wait HERE and the
repository does not know. Blue: the repository holds things we do not have.
One dot for both would say "something is wrong" without saying what - and one
is resolved by publishing, the other by fetching.
Two defects in what offline saving had just introduced, both losing data
silently - found by re-reading rather than by testing, since the device was
away.
The background refresh REPLACED the cache with the repository's version. An
edit made offline vanished at the next fetch, without a word, which is
exactly what offline saving promises to prevent. It merges now, and keeps the
flag: merging is not publishing.
And the flag lived only in memory. An application killed between the offline
save and the publish forgot it had anything to send - and the next refresh
then clobbered the cache. It is a file now, read before any refresh can run.
A file rather than a field of the sealed cache: the flag carries no secret,
its whole content is yes or no, and the cache's signature is already nine
parameters long.
The save used to start from the network: without it, nothing was kept. But
what someone has just typed is already theirs - publishing is a consequence,
not a condition. The cache is written first, always, and the repository
follows when it can.
A failed publish no longer loses anything. It raises a flag, and the toast
says both things: that it is kept, and that the repository does not know yet.
A bare "failed" would have people retype what is already there.
On conflict the two snapshots are now MERGED rather than the change replayed.
Replaying one edit onto the fresh snapshot was right while writing required
being online; with offline saving several changes can be waiting here, and
replaying one would drop the others.
Sync publishes what is pending BEFORE fetching. Fetching first would overwrite
our cache with the repository's version and the offline change would vanish
without a word - exactly what offline saving promises to prevent.
A dot on the padlock says changes are waiting: discreet until looked for,
present at every glance.
Two questions, and the code was only the second one.
WHERE THE FILE LANDS. The private directory is useless - no file manager
opens it and it leaves with the uninstall - and direct external storage has
been closed since Android 10. Of the three remaining doors, MediaStore ->
Downloads is the only one needing neither a permission nor an extra screen:
since API 29 an app may write what it inserted itself. The destination picker
would need an activity result, which FMX makes painful; sharing would need a
FileProvider in the manifest. Below API 29 it would take
WRITE_EXTERNAL_STORAGE - one of the ten permissions just removed - so it
refuses and says so.
WHERE THE CONTENT IS. TVault drops it on purpose: a single attachment took
the cache from 13 to 174 Ko. It is re-read from the raw snapshot, which has
been in memory ever since saving stopped fetching before publishing. The same
data now serves two needs that had nothing to do with each other.
Decoding and writing happen off the main thread, under the waiting veil: a
few megabytes would freeze the screen, and the veil would not even paint.
What leaves the application leaves its protection: the file lands in the
clear in Downloads. The confirmation names the place rather than saying
"saved".
Release configuration compiles too - checked, since that is what section 5
caught out.
The home screen's buttons already dimmed during an operation, but from the
list nothing moved: deleting or duplicating takes a second or two - the
container's round trip - and the screen stayed identical, as if the tap had
not registered. A veil with a turning arc now covers any screen, and it also
prevents the second tap, which would not wait any better.
Its safety net is the existing 25 s deadline: even if an operation never
hands back, the veil lifts.
The plan records where deleted entries actually are: every publish is a Gitea
commit, so the snapshot from before the deletion is still in the repository's
history, readable in its web interface and recoverable by import. Not a
screen, but a real net - and one that catches what a trash would not: an
unlucky edit, a bad import.
The floating menu was wrong for a list. It closed on the next tap, so opening
an entry afterwards took two gestures and felt like a button had got stuck.
Android's contextual bar instead: a long press selects, the top bar takes the
place of search, says how many rows are held, and offers duplicate and
delete. It goes away when nothing is held. One does not search while choosing
what to delete, and two rows of controls would fight over the same thumb.
Selection is multiple, because the cost is in the bar, not in the work: the
patch closure loops, and ten deletions become ONE publish. The container goes
up whole on every write - deleting ten entries one by one would pay that trip
ten times, and give nine chances to fail halfway.
A tap during selection picks rather than opens, which is also the way out:
unpicking the last row closes the bar. The hardware back cancels the
selection before locking the vault - locking because someone wanted to cancel
a choice is the opposite of what the gesture means.
Both menus stay, with distinct jobs: the bar acts on rows chosen in the list,
the detail's three dots act on the entry in front of you. Duplicate copies
everything, attachments included, as the desktop does - a copy that silently
lost the attached document would be a surprise discovered too late.
Two things, and the second was me misreading the question: the affordance was
asked for in the LIST, not in the opened entry.
A long press on a row opens the same menu. Not a cross per row - an
irreversible action a thumb's width away on a list one scrolls - and not a
third button per row, when those were just cleared away. A long press is
invisible until used, it is Android's convention for acting on a list item,
and it is where multi-select would go if it ever comes.
And the deleted entry reopened itself. editDone hands control back after a
refusal, but it redrew the LAST detail - after a delete the view has already
returned to the list, so the entry one had just erased came back on screen.
It now hands back to what is on screen, not to what was.
Delphi deletes by rank when the list asks, by the open entry otherwise -
exactly like open and copy. The page never sends a uuid: it has one, but
letting it name a row Delphi never displayed is a door with no reason to be
open.
Copying a password or a username opened the entry instead - about half the
time, which is the worst kind of failure to diagnose. Since those buttons
carry an icon, the tap target is the <svg> or one of its strokes, so reading
the attribute off the target returned null and the handler fell through to
"open". A finger landing sometimes on the glyph and sometimes on the margin
made it look random. Both handlers now resolve to the button first.
The new-entry form gets a folder, like the editor: an entry filed from the
start saves having to move it.
And a folder can now be created, not only chosen. Filing an entry under a
name is not enough - it is `folders` that fills the desktop's column, and a
name missing from it leaves the entry in a folder nothing displays, looking
lost. The two now go together, in one place, for both editing and creation.
'All' is never registered: it is the default, not a folder.
The folder had no row of its own - it shared "Folder / tags" with the tags,
so neither could be touched. It is now CHOSEN from the known folders rather
than typed: a typo would file the entry under a folder nothing displays,
here or on the desktop, and it would look lost. Tags stay read-only because
they are an array in the snapshot, and writing one as a string would break
its shape for the desktop.
The 2FA secret gets its own masked row. The code has had a panel for a long
time, but it is computed and renewed every thirty seconds; the secret is what
one sets, changes and removes, and nothing allowed that from the phone.
User, Pass, Copy, Show and Edit become drawings. Four words on a list row left
little room for the entry's name and had to be re-read every time. Every icon
carries an aria-label: replacing a word with a drawing must not take the
information away from someone listening to a screen reader.
Decided: it is outside the repository, it no longer enters a release build,
and the token it carries opens a test repository only. Keeping it saves
retyping forty characters at every reinstall of the bench.
What would warrant revocation is a token granting access to anything beyond
that repository - not this one. Recording the decision so it stops coming
back as an open item.
Until now the two situations never met: a local vault had no repository, a
pulled vault had no local container. The day both exist, the rule "the local
container is the authority" means saves never reach the repository - and the
two diverge in silence, which is the worst outcome available.
Sync now merges them once and then deletes the local container: the
repository becomes the authority and the device behaves like any other. That
makes divergence impossible rather than unlikely. Order matters - publish
first, delete after; the reverse would lose the whole vault if the publish
failed.
The remote is the base, never the local. It carries what the phone cannot
rebuild: attachment bodies, the avatar, folder colours, and whatever the
desktop adds next. Starting from the local would amputate all of it, and the
desktop would adopt the amputated version on its next pull.
Arbitration is the desktop's, term for term: last write wins on updated_at, a
tombstone beats an older entry but not one edited after it - that one was
revived on purpose. Timestamps are ISO 8601 UTC on both sides, milliseconds
included, so lexicographic order IS chronological order.
MergeSelfCheck asserts the four rules at startup in debug. The merge never
runs against a real vault before it has run there.
The manifest asked for fine and coarse location, camera, phone, phone state
and external storage - Delphi's defaults, none of them used. On a password
vault those are exactly the wrong signals: whoever reads the app's listing saw
a request for position and camera with no possible explanation. The list is
written by hand now.
A trap on the way: naming the substitution marker inside the comment that
explains its removal is enough to bring it back. The substitution is textual
and does not care that it lands in a comment.
The bench credentials are no longer compiled into a release - the local
include is debug-only. Verified in the binary in UTF-8 AND UTF-16: Delphi
stores strings as UTF-16, so an ASCII grep returned zero on both sides and
proved nothing.
The log leaves the home screen in release without leaving the application: a
1.5 s press on the status line brings it back. Nothing secret is in it, but a
permanent Log button is development noise, and diagnostics must stay reachable
for the day something goes wrong remotely.
And the Release configuration did not compile at all, which nobody knew: the
two self-checks had their declarations under IFDEF DEBUG and their bodies
outside. Invisible as long as one only ever builds Debug.
Signing is what remains, and the keystore password is the user's to hold.
The retry path finally ran, and did not recognise what it was handed. A
desktop publishing between our read and our write earns a 422, not a 409, so
the save failed instead of retrying - which is exactly what a never-exercised
path hides: it was written, it compiled, it was even commented, and it did
not fire.
Gitea answers 409 when the sha no longer matches and 422 when it is missing
or the request is judged unprocessable, "file already exists" among them.
Both say the same thing to us: the sha is stale. Re-reading and retrying is
right for both and cannot clobber anything - the retry publishes against the
fresh sha. A genuinely malformed body would fail again and give up after
three tries, carrying Gitea's own words.
The log now prints which code it was, rather than assuming.
It went from 0000 to 00000000 with nobody deciding it. The typed code and the
code the cache is sealed under were the same variable, so every write of the
cache re-sealed under whatever sat in the field - a sync, which asks for no
code at all, was enough to change it, silently.
They are two things now. Only opening the vault touches the seal, because
that is the one moment we know the current code is known: the decryption just
proved it. Everything else - sync, background refresh, saving an entry -
rewrites the cache under the code already in place.
Hence a second field, shown only once a code exists: the current one must open
the vault for the new one to take. Without it this fix would have locked the
device under the code set by accident, which is exactly the harm it is meant
to prevent.
Both, or nothing: the row leaves the snapshot AND a tombstone takes its
place. Removing the row alone does not delete - the desktop still holds its
own copy, and the next pull would bring it back as an entry we never had.
The time matters too: the desktop's arbitration keeps an entry edited AFTER
the deletion rather than killing it a second time.
No trash on the phone, deliberately. The desktop's is a `deleted` flag that
does not travel in sync, so a deletion stopping there would propagate
nowhere. Hence a button at the very bottom, in two taps, that disarms itself
after four seconds - a button left armed gets touched by accident much
later, and this is the one action with no way back.
The confirmation is in the page because the engine cannot offer one: a
JavaScript prompt needs a WebChromeClient FMX does not provide.
Not exercised on the device: it would have destroyed a real entry.
PatchEntryField already took a field name; what was missing was carrying it
from the row. It travels in the detail row itself rather than being guessed
from the label - a caption is prose, and the day it changes the page would
send a field name that does not exist, or quietly stop offering the edit.
Delphi does not take it on trust either: four names accepted, the rest
refused. The snapshot also carries tombstones, attachments and timestamps
that PatchEntryField would overwrite without understanding them.
An empty row now shows IF it can be filled - a missing username you cannot
add is a dead end. Emptying is allowed, except on a password, where it would
almost always be a slip; the refusal says so rather than doing nothing.
Show only appears on a masked field: on a site in clear it has nothing to
reveal, and would in fact hide it. A note keeps only Name and Content - a
site and a username mean nothing there.
The crypto is down to about 135 ms, so the wait was never the crypto: the
log says HTTP 200 in 3244 ms for 266299 bytes. Every save fetched the whole
vault, decrypted it, patched it, and sent it back.
The fetch existed to be sure of modifying the latest published version. The
sha proves that better and for free: publish against the snapshot in hand,
and if another device published in between, Gitea refuses with a 409 - then,
and only then, pull for real. The retry loop was already written for exactly
that.
The snapshot kept is the one RECEIVED, never a re-serialisation of TVault:
that would drop everything the phone does not read - attachment bodies, the
avatar, tombstones - which is the invariant PMA.VaultPatch exists for. It
goes at lock, with the rest of the plaintext.
Note this makes the 409 path load-bearing, and it has still never run. It is
now what guarantees a save cannot clobber another device.
The PUT still carries the whole container; that is inherent to the format.
Opening from the cache derives nothing, so the key learned last session was
gone and the FIRST save of every session paid the 4.5 seconds again. It now
travels in the sealed blob, next to the plaintext vault and to the password
that opens the container - that neighbourhood is what makes it acceptable:
it exposes nothing the file does not already hold.
The three parts go together or not at all. A key without its salt does not
say what it fits, which makes it useless, or worse, used where it does not
belong. A cache written by an earlier version carries none: LoadVault
returns nil, one derivation happens, and the next write stores it.
A freshly registered vault still derives once - EncryptContainer does not
hand back the key it just derived.
The plan records what the desktop confirmed: the two entries created on the
phone imported with nothing to correct.
Adding an entry took five to seven seconds. The log now says where they
went: key derived in 4395 ms. Every save re-derived the master password -
600000 iterations - to arrive at exactly the key that had just been used.
The remedy was already half written: EncryptContainerWithKey exists for
this reason, but only within a single save. The key now lives for the
session, and is reused only when the container carries the same salt and
the same iteration count - the desktop may have re-encrypted its own with
a fresh salt, and a key that does not fit fails on GCM's tag while the
message blames the password, which would be right.
Nothing new is exposed: the plaintext vault is already in memory, and the
cache already holds the password that produces this key. It goes with
them, at lock - salt included, or the next open would think it holds a
usable key and decrypt against nothing.
The envelope is read in one place now rather than copied per entry point:
at three copies, a drift between them would surface as a wrong-password
error.
Measured against the real repository: HTTP 200, decrypted, 8 entries - so
the envelope refactor holds. The reuse line itself is not yet measured on
the device.
The form disables its Save button on send, so a refusal that only speaks
leaves it dead with the typed text behind it - which is what happened to
the one guard added with entry creation: name it, press Save, nothing.
The rule had already been forgotten twice before, at the editor. It is a
procedure now rather than two calls side by side, because two calls side
by side is exactly what gets written as one.
Measured on the device: refuse, let the message fade, refuse again - the
second press produces the message, so the button had come back.
A vault you cannot add to is not a vault, so an account created here was
useless without this. A plus button next to the padlock, a form, and a
kind - login or note, which the vault distinguishes and whose body is the
password field either way.
One save path now, two transports, and the rule that decides between them
is: if a local container exists it is the authority. A vault born here has
no repository; a vault pulled from one has no local container. The case
where both exist - a local vault that later gains a repository - needs a
merge, and that is section 7 step 4.
The patch arrives as a function rather than an already-patched snapshot:
the remote path must modify WHAT IT HAS JUST PULLED, or it would overwrite
whatever another device published in between. That constraint is why the
two transports could not simply share a finished payload.
Two traps closed while passing: the form left the typed password sitting
in the document after the save, and a refusal left its Save button dead
with the text behind it.
Verified on the device that the button and the form appear and read
correctly. Saving was not exercised - the vault on this phone syncs to a
real repository, and a test entry would have been published to it with no
way to delete it from here yet.
Until now the phone could only read a vault that came from elsewhere:
without a URL, a token and a sync password it showed nothing. It now
signs in like the desktop does - username and master password - and can
create its own vault, with no network at all.
The local container carries exactly what the repository would carry: same
format, same encryption under the password. Two reasons, not one. It is
the desktop's model, where the master password protects the vault at rest
rather than the hardware alone; and the format staying the repository's
means sync keeps working with nothing to convert. No Keystore on that
file - it is already encrypted under the password, and binding it to the
device would make it unrecoverable after a reinstall. vault.bin keeps
that role: it is the fast-open cache, and only it needs sealing.
The password opens the container - the decryption itself is the check, no
secret is ever compared - and the username, read from inside once open,
says whose vault it is.
The home screen now knows which of three situations it is in: nothing
here yet, a container without a cache, or a cache ready to open.
Verified on the device: with the cache set aside, the screen offers
username, master password and Create vault, alongside the download path.
Creating an account could not be exercised to the end - the device locked
its screen mid-run - and the vault that was set aside is back in place,
same size, still sealed under its code.
Two defects, one feature made useless. The cache did seal under the code,
but nothing on the home screen asked for it: opening without one failed,
the failure was read as corruption, and the device offered to download a
vault it held perfectly well. And the next refresh rewrote the cache in
the old shape, so the code was undone seconds after being set, silently.
The code now lives on the home screen, not among the server settings: it
seals this device's cache, while the other three fields describe the
shared remote. Always visible - hiding it when none is set would remove
the only place to set one - the hint saying which case you are in. An
absent code is no longer attempted, a wrong one is no longer corruption,
and a sealed cache is never rewritten without its code.
The two panels each got an arrow, and only one opens at a time.
Measured end to end on the device: version byte 1, code typed, open, byte
4; cold start says "Unlock code required" and open without the code
leaves the byte at 4; correct code opens the vault, and the refresh that
follows leaves it at 4.
It was our own handler. The web view is not a field the system knows, so
it does not intercept the key for us - it arrives here, and we consumed
it in every case: the vault locked, or the list came back, while the
keyboard stayed up with no way left to close it.
hideSoftInputFromWindow answers whether there was anything to fold, which
is the synchronous signal needed to decide. Something folded, key eaten;
nothing folded, the key keeps its usual meaning.
Measured on the device, before and after: mInputShown stayed true, it now
goes false and the application stays in front. A second press locks, as
before.
Same origin as the sideways slide. The keyboard's validate key asks for
the next field, and the next one in the document sits on a view parked
outside the frame. Focus landed there unseen: the keyboard changed shape
- numeric for the code, alphabetic for the search - and would not close,
because the field holding it open was off screen.
No field of a resting view keeps the focus now, and the validate key
closes the keyboard instead of jumping. In the editor it saves, which is
what the key is for after typing a password.
The write test goes with it: it proved the write path before the first
real PUT, that path now carries the vault itself, and the button, its
handler, the probe and the debug hook that revealed it were only ever
scaffolding.
Setting a code and syncing did nothing, and the log said why: "304: unchanged,
vault served from cache". A sync only writes when the remote actually changed,
and it is the write that re-seals - so vault.bin kept its old format and its
old date, and the code looked inert. My instruction to type the code and sync
was wrong; I gave it without following the path.
Re-sealing belongs where the plaintext is already in hand: opening the vault.
That spot has a precedent - the existing path that re-seals under
authentication when a screen lock reappears - and the code now takes priority
over it, since the two protections do not stack.
The log also names the format it opened, so "sealed under the unlock code" can
be read rather than inferred from the absence of a warning.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The reported pain: changing the pattern destroys the local cache every time.
That is Android doing exactly what it promises - a key created with
setUserAuthenticationRequired is invalidated the moment the screen lock
changes, so a new lock cannot open data sealed under the old one. No storage
format changes that, which is why SQLite would not have helped: the problem is
what the key is tied to, not what the file looks like.
So the cache gains a third sealing format, VER_PIN, with two layers:
inner - AES-GCM under PBKDF2(code, 100k) : it takes the code
outer - the Keystore AES key WITHOUT the auth constraint : it takes the device
The outer key is never invalidated, so the file survives any change to the
screen lock; the code restores the protection that removing the constraint
would otherwise cost. 100k iterations rather than 600k because a four to six
digit code carries little entropy - more rounds mostly slow the honest user,
and the real barrier is the outer layer requiring this phone.
The code lives in a fourth settings field and is NEVER stored: empty at every
launch, typed to open, and used to re-seal on every write. That is what
separates it from the token, whose persistence is a convenience with no
consequence since it only opens ciphertext.
Built but NOT installed - the phone went off the network.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The log named it exactly: "Cle ou sel absent pour le rechiffrement". The edit
that swapped DecryptContainer for DecryptContainerEx had never been applied -
the timing line from the same change was missing too, which is what should have
told me. So the key was never collected and re-encryption refused.
Worth keeping in mind: the guard in EncryptContainerWithKey did its job. It
refused rather than encrypting under an empty key and writing a container
nobody could read into the vault. A failed save is an inconvenience; a
published unreadable snapshot would have been the kind of loss this whole
section is built to avoid.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Reported together, and the second explains the first: Cancel did nothing
because Save disables it and nothing re-enabled it - a disabled button fires no
click. And the screen never came back because the entry guards (busy, no uuid,
no sync password) returned with a toast and nothing else, leaving the editor
open with every control dead. A cul-de-sac with no way out but killing the app.
Each guard now calls PM.editDone, which redraws the row from the last detail
received. The page can leave edit mode without asking Delphi for anything,
which is what it needed to recover from a refusal that never reached the
network.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The except only covered decryption and the patch. An exception from
re-encryption or from the PUT skipped the block that restores the screen - and
that block also held the message, so the failure was invisible as well as
permanent: the row stayed on "Saving..." with nothing to explain it.
Restoring the screen now lives in the finally, so it happens whatever the code
above did. A catch-all around the body turns any unforeseen exception into a
message and a log line, because a failure that says nothing is indistinguishable
from slowness - which is exactly how this one was reported.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Two defects, both reported from the device.
Nothing came back after Save. AdoptVault rebuilds the LIST, not the detail, so
the row stayed in edit mode: the input remained on screen and nothing said
whether the write had gone through. The detail is now redrawn from the vault in
both cases - after a success it carries the new value, after a failure the old
one, which reads without needing a sentence. It is found by UUID rather than by
index: the vault has just been replaced by the server's, and nothing guarantees
the same entry sits at the same rank.
The button also stayed inert-looking for the several seconds a save takes, so it
now says "Saving..." and disables itself while the work runs.
And the typed password showed in clear. My reasoning was that you cannot verify
it anywhere else - but that argues for a reveal button, not for leaving it
exposed the whole time. It is masked, with Show next to it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Reported as "Save takes 6 to 8 seconds". A save decrypts the snapshot it just
pulled and re-encrypts the patched one - with the same password. That was two
full PBKDF2 runs at 600k iterations, four to six seconds each on this phone,
for two keys that protect the same file.
DecryptContainerEx now hands back the key, its salt and the iteration count,
and EncryptContainerWithKey reuses them. The IV is still drawn fresh every
time: reusing an IV with the same key breaks GCM entirely, whereas reusing the
salt merely yields the same key - which is the whole point. A salt slows an
attack on the password; it has no reason to differ between two versions of the
same file.
EncryptContainer keeps its old shape and now delegates, so there is a single
place that composes a container.
The log times both halves, so the next measurement says where the remaining
seconds are rather than inviting another guess.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Pull, patch, re-encrypt, publish - in that order, and against the SERVER's
snapshot rather than the local cache. That is what replaces the merge the plan
asked for: anything untouched is already everyone's newest version, so there is
nothing to reconcile.
The only conflict left is another device publishing between our read and our
write. Gitea refuses it on the sha with a 409, and the whole cycle repeats,
decryption included, since the content changed. Three attempts: past that the
other device is publishing in a loop and insisting would not help.
The editable field is marked in Delphi rather than recognised by its caption in
the page. A caption is prose; the day it changes, matching on it would silently
stop offering the edit.
The editor is inline rather than window.prompt: a JavaScript prompt needs a
WebChromeClient that FMX does not provide, and prompt() then returns null
without a word - the button would appear to do nothing. The value shows in
clear while typing, deliberately: you are entering a password you cannot read
back anywhere else, and masking it would mean confirming a keystroke blind.
This is the one command that carries a secret over the bridge, agreed
beforehand. It stays in the process, the iframe carrying it is destroyed, and
the handler logs the action only - never the query.
Only the password is editable. Creating or deleting an entry would need
tombstones and uuid generation, and changing a password covers the need
without either.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Section 4 step 4. TVault reads a subset of the fields, so re-serialising the
vault from what it understood would drop attachment contents, created_at, the
account avatar, tombstones, and whatever the desktop adds next - and the
desktop would adopt that amputated version on its next pull, the merge being
last-write-wins on updated_at. Silent, and final.
So nothing is rebuilt: the received JSON is edited in place and the rest is
handed back untouched. What the phone does not understand travels intact by
construction rather than by vigilance.
A corollary worth naming, because it removes the step the plan feared most:
since the patch applies to the snapshot just pulled, there is no merge to
write. Anything untouched is already the server newest version. The only
possible conflict is another device publishing between our read and our write,
which Gitea refuses on the sha.
The self-check tests what is PRESERVED, not what changes: a field unknown to
TVault, an attachment content_b64, the untouched sibling entry, the
tombstones. It reports "patch keeps what it does not understand" on the device
at every debug launch.
Also confirmed by this log: the write token now survives a restart, so the 304
settings fix holds.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Proven on device, against the real repository:
GET first: HTTP 200, sha 8feb3b7d...
PUT .../write-test.json
OK 200 - new sha 67d6a4b2...
Read the sha, write with it, receive the next one. Section 4 step 3 is done.
The probe reads before writing because that is the real sequence: an empty sha
means "create", and Gitea answers 422 when the file is already there - which is
what it did, and rightly.
The second fix comes from an interaction between two changes made hours apart
today, and only visible once both were in place. Settings are stored on the
line after "received N bytes"; but now that the sha comparison works, a sync
answers 304 and returns before ever reaching it. So a freshly typed token was
never saved and vanished on close - which is exactly what the user reported.
They are now kept as soon as the transport answered at all, 200 or 304: both
prove the address and the token, only the body differs.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Reported as "the token field saves nothing", and it was exact: LockVault
called PushConfig, which pushes the STORED server settings back into the page.
Every lock therefore overwrote a freshly typed token with the old one -
including the three-minute auto-lock, so it happened with no gesture from the
user at all, and the field looked like it forgot on its own.
The user's own sequence isolated it: deleting one character gave a 401, which
proves the typed value does reach the wire; then open, lock, sync gave a 200
again, the lock having restored the old token in between.
PushConfig now runs at startup only. It was pointless in the other four places
anyway: PM.wipe clears the list and the detail, never the configuration fields.
Along the way, and part of the same confusion: the three settings fields had no
labels, only placeholders that vanish once filled - and since the token was
masked this morning it became indistinguishable from the sync password sitting
right below it. They are labelled now, and the log reports both fingerprints so
a value landing in the neighbouring field shows up immediately.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
TSyncClient.Put publishes a container to the Gitea contents API: token header,
JSON envelope with the file base64 inside, sha omitted rather than empty when
creating, branch omitted so the repo default is used. 409 becomes its own
status, ssConflict, because it means re-read and re-apply rather than retry.
The first PUT is where a mistake costs data, so the probe writes to
write-test.json rather than the vault. Same repository, same token, same
envelope, same concurrency check - and the vault is never at stake. Reachable
from a debug-only button; the page asks Delphi whether to show it.
Two mistakes of mine on the way, both about replacing information with a guess:
- the probe did not push the config first, unlike Sync, so a token that had
just been pasted had not reached Delphi and the 403 blamed the wrong thing.
- 401 and 403 were rewritten as "token lacks write access", discarding Gitea's
body. That is an interpretation, not an observation: the same code covers a
read-only repository, a protected branch, a token from another account. The
body says which - here, plainly, "token scope=read:repository".
Also recorded: the sha comparison now short-circuits (304 where it used to
decrypt), so the difference was case or whitespace, absorbed by SameText on
trimmed values. The offending character itself was never seen.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Section 4 step 2. EncryptContainer produces the exact container the desktop
writes: PBKDF2-SHA256, AES-256-GCM with the tag appended as WebCrypto does, a
32-byte salt and a 12-byte IV drawn per call from SecureRandom on Android and
BCryptGenRandom on Windows - never Random, since a predictable salt or IV ruins
the encryption that uses it and nothing in the output would show it.
The check matters as much as the code. A round trip with itself only proves
internal consistency, and it passed while the container carried two defects.
Node - the same primitives as WebCrypto - found both by reading a container the
phone had produced:
- Delphi's base64 wraps at 76 characters; WebCrypto's never does. Lenient
decoders accept it, but depending on their leniency is not a guarantee.
- the accented literal in the check came out of the COMPILER already broken:
this source has no byte-order mark, so Delphi reads it as ANSI. The defect
was in the test rather than the encryption, and only the cross-check could
tell those apart.
The check runs at launch in DEBUG and leaves selfcheck.json in the private
directory so it can be re-read from the PC.
Also fixed on the way: TPath resolves to FMX.Objects.TPath - a drawing shape -
because FMX.Objects comes later in the uses clause, and the error then talks
about an unknown Combine rather than a name collision.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Asked plainly by the user: why does open trigger a sync? Because
OpenLocalNow displays the cache and then refreshes in the background, so a
change made on the desktop shows up without anyone thinking about it. Sound
intent, applied on EVERY open - including a lock followed by a re-open three
seconds later, which cannot learn anything new. Measured cost of one such
pointless refresh: 1 s of download and 5.5 s of key derivation.
A minimum interval between two background refreshes fixes that, whatever else
is wrong: the desktop does not publish several times a minute, and looking at
most once a quarter of an hour is enough for a change to arrive unnoticed.
Pressing Sync goes through another path and never waits.
The sha comparison that should already have made an unchanged refresh cheap is
still not firing - two fingerprints print identically yet compare unequal. It
is now compared with SameText on trimmed values, which cannot confuse two
distinct snapshots since a hex digest never differs by case or whitespace
alone, and the log reports both lengths so the invisible character shows up on
the next real refresh.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Measured on the phone, opening a cached vault: 121 ms to read the sealed file,
14 ms to display it - and then 5718 ms decrypting a snapshot it already had.
The cause is upstream. A sync pushes unconditionally at the end, and the push
is never a no-op: the container is re-encrypted with a fresh salt and IV, so an
identical vault produces completely different bytes. On Gitea that is one more
commit and a new blob sha, so every other device sees a "new" snapshot and pays
a full key derivation to arrive back where it started. It also grew the
repository on every sync that changed nothing, which is the history cost the
plan warns about.
The merged state is now compared against what was pulled, as plaintext -
the only place equality is observable - ignoring snapshot_at, which changes by
construction. Equal means the push is skipped and the sync still counts as done,
timestamp included: having nothing to write is a success, not an absence.
The comparison is deliberately strict. A false negative costs one commit; a
false positive would silence a push that carried a change and lose it for the
other devices. Four tests fix where that line sits, including a modified
updated_at (which arbitrates the merge elsewhere) and a tombstone (a deletion
that has to propagate).
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Against Zaki/pm_sync on Gitea 1.26.1, container uploaded through the web UI:
HTTP 200 in 1079 ms, decrypted, cache written with the sha; the next Sync finds
the same sha and skips the derivation entirely. That was the point of the ETag
on the PHP endpoint and it survives the change of transport.
One log line was lying: it announced If-None-Match, which is never sent to
Gitea - the API does not know that header, and the comparison happens on the
sha after reception. Announcing it sent you looking for a 304 from the server
that will never come. It now says "known sha" on that transport.
The plan records the two frictions met on the way, both already fixed: settings
kept only after a successful decryption (so a wrong password lost the URL and
the token every time), and the launch refresh being silent and using the
password stored with the vault rather than the field.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The saved settings win over the first-run constants, so a typed URL that never
survived a successful transport kept losing to the old PHP endpoint. The
device's settings.bin was removed (the vault cache was NOT - it is the only
offline copy on that phone, and wiping it is not needed to change an address).
The Android constant now points at the contents API rather than the repo URL:
it is the one that returns the file with its sha, and the client recognises
"/api/v1/repos/" to pick that transport. Windows keeps the local endpoint - the
machine hosts it, and nothing requires it to go out on the network to read
itself.
The token stays out of the repository, as does the sync password: both live in
the gitignored PMA.LocalConfig.inc, or get typed on the device.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
They were saved only after a successful decryption, in the name of storing
proven settings rather than typed ones. The consequence showed up the moment
the password did not match: the URL and the token were lost on the next
launch, and both had to be typed again - an API URL and forty characters of
token, on a phone, on every attempt.
A 200 with a container in hand already proves the address and the token. The
password is a separate matter, and this call never stored it anyway. So the
settings are kept there, before decryption, and the vault plus its fingerprint
still wait for a decryption that actually worked.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The log printed "HTTP 0 in 2013 ms" when no HTTP response arrived at all.
Zero is not a status code, and written that way it reads like one the server
returned - which sends you looking for the cause on the wrong side of the
connection. It now says "no HTTP reply after ... ms".
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Reported symptom, with the local server stopped: Sync leaves the app inert -
neither Open vault nor Sync responds, and nothing is displayed. Log and
Server settings still work, because they are local to the page.
The log proves where it stops: the GET line is there, the "HTTP ... in ... ms"
line that immediately follows the call is not. So the request never returned,
SetBusy(False) never ran, and the two actions stayed disabled - a disabled
button fires no click, hence the silence. THTTPClient is given 15 s timeouts
and the Android request does inherit them, so this should not happen; rather
than keep chasing why it did, the situation is now impossible.
A deadline timer, armed while a request is in flight and longer than the HTTP
timeouts so the precise network error still wins when it works. On expiry the
interface comes back and says so. The thread is abandoned, not cancelled: it
dies when its socket fails, and if it somehow succeeds its result still shows.
The page's bridge queue had the same shape of bug: a throw while removing the
iframe left "sending" true forever, and every later command was dropped in
silence. The removal is now in a try/finally.
Also fixed while in there: SetBusy hardcoded "Sync" as the button label, so a
vault-less device lost "Download vault" on every busy transition.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The caret in an Android TEdit does not land where you touch, which makes a long bearer token impossible to correct - the reason the last FMX screen had to go. A field in the engine inherits the system's text handling (caret, selection, handles, paste) for free.
With no FMX control left, two things became dead weight and are gone: the vendored Dark.style, which had nothing left to style, and Skia - 25 MB carried solely so FMX labels would shape Arabic. The engine does its own shaping; checked on the device, an Arabic filename still reads correctly. The APK drops from 22.0 to 12.7 MB.
Delphi stays the owner of the server settings; the page only edits them and hands them back when a field loses focus, one navigation per changed field rather than one per keystroke.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
A swipe with nothing to scroll dragged the whole document, showing the body gradient sliding under the screen gradient - two gradients stacked and no content at all. overflow:hidden does not prevent that in an Android WebView; position:fixed does.
The lock button went missing when the list moved to HTML: it lived in the old FMX bottom bar and had no counterpart in the page. It returns as a padlock next to the search field, no label - it is the one icon universal enough to stand alone.
Icons on the unlock screen too, as code points since this source is read as ANSI. All four are basic-plane and common, so a font without them would show an empty box rather than a wrong glyph; checked on the device, they render.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The FMX controls lost this one on merit. Every screen re-taught the same lessons: a template missing from the loaded style, an alignment order left to chance, a HitTest swallowing a tap, a native view painting over everything. The engine draws the same interface in forty lines of CSS and shapes Arabic on its own.
The bridge needs no local server, so no port is open on the device: evaluateJavascript carries the data down, an intercepted cmd:// navigation carries the actions up. Through a throwaway iframe - FMX always tells the engine it did not handle the navigation, which on an unknown scheme replaces the page with an error screen.
Secrets do not travel over that bridge. The list gets names, sites and usernames, which are already on screen; copying from a row sends an index and Delphi reads the vault itself. Only the open entry has its values in the JS heap, and PM.wipe clears them on lock.
Folders become a row of chips with counts instead of a full-screen panel: one tap, and the count says what to expect.
Also gone: about eight hundred lines of the screen this replaces.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
r-zakarya