| Commit message (Collapse) | Author | Age | Files | Lines |
|---|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
says who stopped
Bounding the client's read-ahead changed what a transcode slot is. It used to
be a burst — the browser took segments as fast as it could append them, so a
slot came back within the minute whatever the length of the film. Now it is
held for as long as someone is watching, so the cap counts simultaneous
viewers, and two of them meant the third was refused for the next hour and a
half.
The right number depends on the machine, so it belongs to the operator:
`[node] max_concurrent_streams` in node.toml, or
MESHBAY_MAX_CONCURRENT_STREAMS. Default 8 — one ffmpeg per viewer, remuxing
rather than encoding, idle on a pipe for most of the film. Zero, a negative
number, a non-number and a bool are refused with a warning naming the setting:
`Semaphore(0)` is not "no limit", it is a node where no video ever plays and
nothing says why, and TOML `true` would have become 1 by way of `int()`.
A stream also ends on the peer's silence now rather than on its stinginess. A
viewer buffered well ahead deliberately grants nothing for minutes, and the
old budget accumulated over the whole wait, so a keepalive that granted no
credit could not keep a paused film alive.
The rest is diagnosis, which is what this cost. `client_diag` carries the
player's own view — readyState, refused appends, buffered ranges, the video
element's error — into the node's log at DEBUG, next to the node's view of the
same stream. It is the only window into a phone, and every field is
stringified and cut short because all of it is peer-controlled. The node also
logs the first keepalive, which distinguishes a paced client from an unpaced
one at a glance, and progress every hundred segments, whose last line says
where a stream stopped and which side stopped it.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Reported from a phone: play a video, close the viewer, open another — the second
hangs and the third is refused. Three separate causes, found by instrumenting
rather than guessing, after two fixes that addressed real but different bugs.
A task nobody holds can be collected mid-flight. asyncio keeps only a weak
reference, so `ensure_future` with the result discarded may be garbage-collected
while running — "Task was destroyed but it is pending!" — and `_stream_video`
never reached the exit of its `async with sem`. `_spawn` holds every background
task; all nineteen call sites go through it.
Losing the peer must stop its work. The connectionstatechange handler popped the
session from a dict and nothing else, so a closed tab went on transcoding for
the full 120 s credit timeout. Measured in the log: 91 s of ffmpeg after the
connection closed. `shutdown_tasks()` now runs on the way out, and the credit
wait checks the channel before sleeping and polls in slices instead of once.
And `await proc.wait()` after `kill()` still deadlocks. ffmpeg outruns a
credit-paced viewer and fills the stdout pipe; stop reading it and the transport
cannot finish closing, SIGKILL or not. Measured against the live node with a
169 MB video, closing the viewer after 20 segments and asking for the next one:
15.1 s then "Server busy" before, 0.1 s / 0.0 s / 0.0 s after.
Chunk replies wait for room on the channel. Eight megabyte-sized chunks answered
as they arrived queued 8 MB with nothing watching — measured at 7.3 MB of
bufferedAmount in milliseconds. Fine on a LAN, minutes of head-of-line delay on
a busy link.
Upload names accept any script. The rule was ASCII-only, so `été.txt` was
refused — and so was `rapport (1).pdf`, which is the form `_free_name` produces
itself, meaning the node rejected names it had chosen. Widened to Unicode with
the C5a and H2 protections intact, plus a refusal of names that lie about
themselves: trailing space or dot, and the right-to-left override. Errors now
name the file, so one bad name no longer fails every upload in flight.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
`get_messages` pages forward from the oldest message. That is the right shape
for "what happened since I last looked" and the wrong one for opening a
conversation, and the browser asked it for `since=0, limit=200` — so a group
with more than two hundred messages showed its first two hundred and the
exchange anyone came for was unreachable. Demonstrated on 300 messages: the
newest was simply absent from the answer.
`get_recent` and `get_before` page backwards, cursored on the row id rather than
the timestamp. Nothing makes a `time.time()` float unique, and a cursor on a
value two rows can share eventually skips a message or repeats it.
PING/PONG covers liveness on an already-open channel: a DataChannel whose peer
vanished without closing still reads as connected, and nothing noticed until a
real request hung. It is not a discovery mechanism — opening a connection to
ping costs a full ICE/DTLS handshake, measured at 0.6-7 s across two ISPs — so
presence in the group list comes from the hub's registry instead.
Both additions are backward compatible: an 0.1 peer sends no `before` and is
answered with the newest page, which is what it wanted.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
**Removing a member.** The owner can do it from the Members tab, and it
is two halves in the order that fails safe: the node stops serving the
group key first (an operator-signed request, so a paired browser only),
then the hub drops the membership row. The other order would leave
someone able to reach a node that still serves them.
It is a membership, not an account. The user row is never written: their
other groups, their files and their pinned identity survive, because one
group's owner must not be able to erase someone from the hub. It is also
per group — a node hosting two loses them from one — and it does not take
back the key they already unwrapped, which is what rotating the GEK is
for. The confirmation and the panel both say so.
**Downloads and streaming through the disk, in both browsers.** The audit
this started as found two ways to put gigabytes in a tab.
Firefox and Safari have no File System Access API, so every download
there was collected in memory. A service worker fixes it: the page keeps
the writable half of a transferred stream, the worker answers a made-up
URL with the readable half and a Content-Disposition header, and the
browser writes it to disk as it arrives, with real backpressure. The
worker caches nothing and falls through on every request that is not one
of these downloads. A zip announces no Content-Length, since the archive
is larger than the files in it and a length we miss truncates the file.
Video was worse and affected both browsers. The node pushed ffmpeg's
whole output as fast as it was produced while the player consumed a
segment at a time, so the queue held the film — and appending all of it
hit the SourceBuffer's cap, where the handler logged the error and
dropped the segment, leaving a hole in the middle of the film with
nothing to show for it. Streaming is credit-based now, 24 segments of
256 KB in flight, verified against the live node: three credits, three
segments, then silence until more are granted. The player evicts what is
more than a minute behind the playhead and retries a refused segment
rather than dropping it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Two things a Files panel needs and did not have.
**Removing a directory** is privileged, where creating one is not: it
acts on a name other members are using, on the operator's disk. It is
refused unless the directory is empty, and that rule is the safety
property — whatever the browser sends, this cannot destroy content. The
check runs twice, once before the challenge and once after the signature
comes back, because a file can land during the round trip. A file also
accepts its uploader's key; a directory has no uploader, so only the
operator's key will do.
**Downloading a folder** produces a zip built in the browser, written
straight to disk as the chunks arrive. An archive of a group folder is
routinely tens of gigabytes, so nothing is held: peak memory is one chunk
plus a small record per file. The node is not involved at all — it serves
the same encrypted chunks as any other download, holds no temporary
files, and cannot be asked to compress anything.
zipstream.js is store-only. Group content is video and images, already
compressed, so deflate would spend CPU on every byte to save nothing, in
the thread that is also decrypting. Sizes and CRCs go in a data
descriptor after each file because a stream cannot seek back to patch a
header, and zip64 kicks in per entry past 4 GiB and for the archive
itself. Because none of that can be checked from the Python side of the
house, test_zipstream.py runs the real module under Node and reads what
it produces with zipfile — CRCs, UTF-8 names, zip64 records and all. The
archives also pass `unzip -t`.
Firefox and Safari have no File System Access API, so there is nowhere to
stream to: the fallback builds the archive in memory and says so, with
the size, before starting rather than after failing.
One mistake worth recording: the first version of deleteDirectory passed
the node's own answer as the value to check the challenge against, which
turns the comparison into a tautology. It checks the path we asked for.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Attaching a group to a node meant hand-editing node.toml with a UUID
copied from a browser URL, restarting, and knowing that gek-init exists.
Nothing in the CLI said so, and on a node reached over SSH there is no
paste buffer to carry a UUID across in the first place.
meshbay-node group add grenet --dir ~/grenet-share
The name is resolved against the operator's groups on the hub by the
daemon, which is the process holding the session. The [[groups]] block is
appended to node.toml as text rather than round-tripped through a TOML
writer: the file is hand-written and its comments explain decisions worth
keeping. The directory is created, and the command says what remains —
restart, then gek-init for that group.
It refuses a name it cannot find by printing the groups it can, with
their ids. That listing is the useful half of the answer and it was
missing everywhere: _daemon_api now renders an `available` list from any
endpoint that offers one.
The key is per group and pairing is not, which is the part that reads as
a gap until it is written down: one paired browser covers every group the
node hosts, while each group's key admits only its own members. §4 of the
user guide now says all three of those in one place.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
A description could only be set the moment a group was created, so every
group made before anyone thought of one stayed blank for good. The owner
can now edit it from the group's page, and PATCH /v1/groups/{id} takes it.
That endpoint takes the description and nothing else, deliberately. The
name, the visibility and the join policy are the terms members joined on;
a private group that can quietly become public is not the group they
agreed to be in. Changing those needs a decision about who gets told, not
a field on a form — there is a test saying so.
Separately, the legacy operator key is gone. `admin_pk_ed25519` in
node.toml named the operator before the roster existed and was kept so
that an existing deployment would keep working; nothing uses it, and a
second source of node authority is not something to carry around out of
politeness. Authority is the roster, read fresh on every check.
It is removed rather than ignored: a config that still names the key gets
a warning at startup pointing at the file. Dropping it in silence would
refuse invites and file deletion with a signature error that looks like a
bug somewhere else — which is exactly how finding M3 presented.
Two tests were verifying admin operations by naming a key in the context,
which was the legacy path. They now pair an operator into a roster, the
way an operator does. The authority test anchored on the deleted function
and passed vacuously once it disappeared; it states the invariant against
the verifier and the daemon instead.
Also defined .btn-secondary, used in four places and styled in none.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Correction to the previous commit. Uploads went wherever the member happened to
be looking, which spreads chat attachments through the tree and makes the
destination a client-supplied path — surface that had to be defended. Everything
a member sends now lands in `uploads/` at the root of the shared directory:
visible, one place, easy for the operator to look into or empty.
Chat attachments go there too, so the separate out-of-tree thumbs directory is
not needed and is not built. They were already ordinary uploads; now they are
ordinary uploads that land somewhere sensible.
The destination is chosen by the node, so a client naming somewhere else changes
nothing — the traversal surface simply is not there on this path. safe_subdir()
remains for dir_create, where the path genuinely does come from the client, and
keeps its tests.
One shared directory means name collisions are ordinary rather than adversarial:
every camera produces IMG_1234.jpg. The node finds a free name — "IMG_1234 (2).jpg"
— and reports it in the ack, because a chat message has to point at the file that
was actually written and not at someone else's. Nothing is ever replaced, which
is the property the per-user quarantine existed for (C5a) and the one the tests
assert; they fail if the free-name search is removed.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The per-user quarantine is gone. `.uploads/{user_id}/` was the fix for C5a, and
it worked, but it made the shared directory something nobody could organise:
every file landed under a uuid nobody recognises. Files now go where the member
is looking, most often the root.
What the quarantine actually bought is kept, and is now what the tests assert
rather than the location:
- an existing file is never replaced. That was the real defect — overwriting a
file also made the attacker its recorded uploader, and therefore able to
delete it through the uploader path
- the name allowlist is unchanged
- the destination is confined under the shared root
That last one is new surface: the directory arrives from the client. safe_subdir()
is the single place that decides, with two independent guards — every segment
against the name allowlist, and the resolved result under the root — because one
of them will eventually be refactored by someone who does not know why it is
there. Ten traversal cases are covered, and they fail if both guards go.
Also adds `dir_create` (any member may organise a shared directory; audited like
anything that writes to the operator's disk) and makes the node report its real
directory list in index_sync — folders were inferred from file paths, so a new
empty one, or one that had been emptied, simply did not exist as far as the UI
was concerned.
Two C5a tests changed their assertions deliberately, as C5b's did before: they
encoded the quarantine path, which is the thing being removed. The property they
existed for is asserted more directly than before.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
One keypair was copied to every node its owner joined, so cracking the bundle on
any single node yielded the identity used on all of them: their content on other
operators' machines, and the ability to sign as them anywhere. That lateral reach
was the part of C4 worth attacking.
Each node now gets its own keypair, generated the first time its owner joins it
and left with that node alone. An operator who cracks what sits on their own disk
holds a key that is a stranger to every other node — and on their own node, one
that unlocks nothing they did not already hold: they serve the content, the index
and every byte of it by design.
Nothing changes for the user. A first contact with a node already needed that
operator's code, and the key is created in the same step; a second browser still
recovers it from the node with the passphrase alone. Two operators can also no
longer tell they host the same person by comparing keys.
BREAKING, and deliberately without a compatibility path — the deployment is wiped
for the next demo:
- users.pk_ed25519 / pk_x25519 dropped (migration a7c31f9e40b2)
- registration no longer sends or stores a key
- PUT /v1/users/me/keys and regenerateKeys() gone; rotation is now
`member unpin` plus a fresh code, decided on the machine that pinned it
- /pubkeys returns an account id and a node's linking key. It was the directory
H3 read, and nothing wraps for it any more
- the pk_user JWT claim is gone
That last one closed a live defect the inventory turned up: the node recorded
pk_user as the uploader's identity and authorized deletion against it, so a hub
issuing a token naming its own key could delete anyone's uploads on any node.
Attribution now uses the key the node itself pinned.
A simplification falls out. Registration generates nothing, so a scripted signup
is a real account: `demo.py bootstrap` takes a wiped hub and node to a working
demo with no browser, which was impossible while keys were born in one.
Also fixes, found by running it on a wiped deployment: the key handed back on a
join now belongs to the group the connection is for, not the group named in the
invitation — an operator pairs node-wide but redeems the code while opening a
group, and expects to read it.
Tests: 343, including the two that state the property — a key pinned by one node
is refused at another, and someone else's code does not admit it. Verified end to
end against a wiped hub and node: bootstrap, pair, invite, join, download,
stream, second browser, revoke.
Design: docs/per-node-identity-v1.md
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Both found by deploying the thing and running the workflow end to end. Neither
was reachable from the test suite, for the same reason in each case: the tests
knew something a real client cannot.
1. A first-time joiner had no way to learn node_pk.
join_request signs a transcript naming the node, and the node key was only
sent in handshake_ack — which an invited member cannot reach, having no GEK to
prove. joinGroup() therefore threw "handshake incomplete" and the browser path
for an invited member was broken. Every test built the transcript from a node
key it already had, so nothing noticed.
The challenge now carries node_pk. It is unverified at that point and never a
substitute for the ack: the ack still proves possession and signs the
transcript, the client checks the two values match and refuses a peer that
changed identity mid-handshake, and TOFU pinning is unchanged. A wrong value
only makes our own verification fail.
test_invite_then_join_delivers_the_gek now takes the key from the challenge
instead of from sk_node, so it proves a real client can learn it.
2. The roster pinned everyone without a name.
`_do_join_request` took the username from the session, which takes it from the
JWT — and the hub puts no username claim in a token. So identities were pinned
with an empty name and `member revoke <name>` could never match: the live node
answered "known: , ,". Invitations now carry the name (new invites.username
column, with a migration for the roster DBs already out there), and the CLI
resolves a name through the daemon: its own roster first, the hub as fallback
for identities pinned before this.
The harness that found them is QE/deploy/e2e.py — gitignored with the rest of
QE/, so it is not in this commit. It does the SPA's job in Python against the
live deployment: hub login, WebRTC via hub signaling, the unified handshake,
joining with a code, index, chunk download and MSE segments.
Verified against meshbay.org and the local node: an account registered from
scratch is invited by code, receives the group key wrapped for a key it proved it
holds, downloads and decrypts a file, streams 5 encrypted fMP4 segments,
reconnects with no code, and is refused after `member revoke`. The node audit log
shows invite_create → join_pinned(via=code) → gek_wrapped → handshake, then
join_no_gek once revoked.
Tests: 232 node+common.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
A node admits people from its own roster, and until now a headless operator had
no way to put anyone on it: pairing worked from the CLI, everything else needed a
browser on a machine that does not have one. Absorbs milestones 14.3/14.4.
member list who is admitted, role, status, when and how pinned
member invite <username> one-time code; the node wraps the key when they
connect, so nobody has to be online then
member revoke <username> stop serving them the key
member unpin <username> forget the pin so they can pair again after a reset
All of it goes through the daemon's loopback API with the per-run session token
(11.5.3) — _daemon_api() in daemon.py, which also replaced three hand-rolled
urllib blocks. `status` deliberately still reads the keystore, config and roster
directly, so it works while the daemon is stopped.
Two things the commands say out loud, because getting them wrong is silent:
- revoke ends by telling the operator to rotate the key. The ex-member stops
receiving it on their next connection, but they hold the current one, and
"revoked" reads like it took the key back.
- revoke/unpin refuse a username the roster does not know instead of acting on
nobody. A typo must not look like success.
Code lifetimes now differ by what the act is: 7 days for an invitation, which
crosses a human conversation and gets answered whenever someone reads their
messages, and 24 h for operator pairing, which is typed during the SSH session
that printed it. Both configurable ([node] invite_ttl_hours, pair_ttl_hours). A
day was long enough for the second and not for the first — a code that dies over
a weekend means finding a browser to issue another one.
The roster is also in the local admin UI, escaped: usernames come from the hub
and land on the page that can re-key groups and read the audit log, so H2's rule
covers them exactly as it covers filenames.
Verified by driving the real CLI against a stub daemon over a socket, which is
how the "known: <nothing>" bug in the not-found path turned up.
Tests: 89 node here (roster, endpoints, CLI routing, TTL config).
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The invite flow fetched the invitee's pk_x25519 from the hub and wrapped the
GEK for whatever came back (app.js:1466, and gek-init did the same server-side).
The hub is the key directory, so a hub answering with its own key was handed the
group key by an honest member following the protocol exactly. No forgery, no
injection, nothing for the client to notice. That was H3.
The fix is not safety numbers. Nobody reads the directory any more:
- the node holds the GEK and wraps it itself, on every connection, for the
X25519 key the joiner signed with their Ed25519 identity in one transcript
(meshbay:join:v1), so the identity key vouches for the encryption key;
- identities are bound to accounts by a one-time code the hub never sees —
40 bits, single use, one account, bounded per connection AND node-wide;
- the node's own roster decides who may receive the key. Hub membership lets
someone reach a node; it no longer gets them anything. A hub that invents an
account and mints it a token is answered not_authorized_for_group.
Safety numbers would have made substitution detectable by a human who checks, at
the moment there is nothing to check against — first contact. Removing the lookup
makes it impossible, and costs the user one code to pass along.
M3 falls out of the same work. The daemon auto-pinned its own keystore key as
admin_pk_ed25519 while the browser signs with the user identity key, so every
privileged operation failed closed with a signature error that looked like a bug
somewhere else; the demo only worked because a deploy script overwrote the value.
Authority now comes from the roster, established locally by `operator pair`.
Asking the hub for the operator's key — the obvious-looking fix — would have let
the hub install itself as node administrator.
BREAKING: gek_bundle_store is deleted, not gated. No member hands the node key
material at all, so C5b becomes structural rather than an authorization to check.
Existing stored bundles are still served, so current deployments keep working.
Also:
- join_policy (invite|open) is read from node.toml, never from the hub — a hub
able to declare a group open would be handed its key. Unknown group ⇒ invite.
- admin signatures are verified against the roster on every check, so unpinning
takes effect without a restart. admin_pk_ed25519 stays readable as legacy.
- two C5b tests were rewritten, deliberately: they asserted that
gek_bundle_store demanded an operator signature, and the message is gone. They
now assert the stronger property. The file says not to fix these tests, so
this is the record of why they changed.
- a slice-1 bug found while writing slice 2: connect() never passed skEdB64, so
pairing would have failed at runtime with no test able to catch it.
Tests: 152 node+common here, including an end-to-end DataChannel run where a
member who has never held the group key redeems a code in the pre-proof window
and receives the key wrapped for a key only they can open.
Design: docs/invite-pairing-v1.md
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
| |
|
|
|
|
|
|
|
|
|
|
| |
test_daemon.py started the real admin UI on a fixed port, so every test file
that also brought up a node collided with it. Each file passed on its own and
the full node suite failed with EADDRINUSE on test_daemon_creates_chat_store —
which reads as a flaky regression rather than a test-isolation bug.
Confirmed against a clean worktree at HEAD before touching anything: the
failure predates the invite work.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Phase 11.5.4/5/6 — finding C6, the last open critical finding.
QUIC ran a JWT-only handshake: a forged or stolen token reached the node and
could inject chat without ever holding the group key. It now runs the same
challenge/response as WebRTC through meshbay_common.handshake — client nonce,
role-bound length-prefixed transcript, GEK proof, and the node proving itself
with a GEK proof plus an Ed25519 signature over the transcript (C3).
11.5.6 channel binding, resolved by spike and then by two findings the spike
could not predict:
* aioquic 1.3.0 exposes no RFC 5705 exporter, and the peer certificate only
via a private attribute. The server reads its own certificate from disk, so
no internals are touched on that side; the client's access is guarded and
fails loudly if an upgrade moves it.
* A RESUMED TLS session carries no certificate — aioquic does not re-send it,
so there is nothing live to bind to. The anchor therefore travels with the
session ticket, which is sound because the ticket is cryptographically
derived from the handshake where that certificate was presented.
* The anchor had to travel with the ticket rather than live on the client
object: resumption constructs a fresh client, so an instance-level cache
was silently useless. Caught by the resumption test, not by inspection.
Both paths refuse rather than degrade. No certificate and no cached anchor
means the handshake fails; it never falls back to an unbound proof, which would
silently drop MitM detection (L4).
QuicChunkClient gains a peer_cert_der property and constructor argument,
mirroring how session_ticket is already carried by the caller.
Tests: 9 quic/multi-group, full node+common suite green.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Phase 11.5.4 — QUIC half. Findings M1, M9 on this transport.
quic_server._do_handshake_sync was a second, weaker copy of the WebRTC logic:
group_id was optional, so omitting it skipped the membership check entirely and
fell back to the node's first group (M1); node-scoped daemon tokens were
accepted as client tokens (M9); and the checks could drift from the WebRTC path
independently, which is how they diverged in the first place. Authorization now
comes from meshbay_common.handshake, shared with WebRTC.
C6 IS STILL OPEN ON THIS TRANSPORT. There is no GEK proof here yet: a forged or
stolen token still reaches the node over QUIC and can inject chat without
holding the group key. What remains is the challenge/response and the mutual
node proof — quic_binding() is written and unit-tested for exactly this, and
11.5.6 (whether a certificate hash is the right anchor, or an RFC 5705 exporter
is reachable from aioquic) is still unproven. This commit narrows the gap to
the proof itself; it does not close the finding.
QUIC tests updated: default tokens are members of the test group, and clients
pass group_id, since it is mandatory now.
Tests: 9 quic/multi-group, full node+common suite green.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Phase 11.5.4/5/7/8 — findings C6 (WebRTC half), C3, L4, M1, M9.
New meshbay_common/handshake.py is the single implementation of authorization
and proof: JWT verify, scope, denylist, mandatory group_id, membership, hosting.
The handshake previously existed three times over and only the newest copy
enforced the GEK proof.
C3 — mutual authentication. Authentication ran one way: the client proved
itself, the node proved nothing. handshake_ack.node_pk was never verified
against anything and per-chunk signatures had been dropped in Phase 9.15, so a
peer that had hijacked signaling (C2) or been substituted by the hub could
accept the client's proof, ignore it, and serve a forged index, forged chat
history and a forged is_node_admin flag. The client now sends a nonce; the node
answers with its own GEK proof over that nonce AND an Ed25519 signature over
the transcript; the browser verifies both and refuses otherwise. It also
refuses an unchallenged handshake_ack, which previously let a peer skip proving
anything at all.
L4 — the proof was nonce ‖ offer_fp ‖ answer_fp: bare concatenation, and a
missing fingerprint silently degraded it to nonce-only, dropping MitM detection
(NS5). Every field is now length-prefixed and domain-separated, the role is
bound so a client proof cannot be replayed as a node proof, and an absent
channel binding is refused rather than tolerated.
M1 — group_id was optional; omitting it skipped the membership check entirely
and fell back to the node's first group. Now mandatory.
M9 — node-scoped daemon tokens are refused on the client path.
NOT DONE: quic_server.py still runs its own JWT-only handshake, so C6 remains
open — a forged or stolen token reaches a node over QUIC and can inject chat
without holding the GEK. quic_binding() is written and unit-tested but unwired.
11.5.6 (whether the certificate-hash anchor works with aioquic, or an RFC 5705
exporter is reachable) is unproven. 11.5.8 TOFU pinning of pk_node is not done:
the client verifies the node's signature but does not yet remember which key it
saw last.
Adds packages/meshbay-common/tests/test_handshake.py (18 tests) covering the
properties every transport must inherit. WebRTC test helpers rewritten around
the shared module; _make_jwt now defaults to the test group, since group_id is
mandatory.
Tests: 24 webrtc, 176+ node+common.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Phase 11.5 — findings H6, C4 (partial), and milestone 11.5.3.
H6 — resource exhaustion. Several paths let one peer degrade or stall a node:
* the DataChannel frame limit was a flat 64 MB applied BEFORE authentication,
so an unauthenticated peer could announce a huge frame and dribble bytes
into it. Unauthenticated peers now get 64 KB; the large budget is granted
only after the GEK proof, where it is needed for uploads.
* _do_stream_segment ran subprocess.run(..., timeout=30) directly in the event
loop, stalling the entire daemon — every peer, every group — for up to
thirty seconds per request. Now async, with a timeout and process kill.
* ffmpeg was spawned per stream request with no cap. Both streaming paths now
share a transport-wide semaphore.
* POST /v1/nodes/{id}/webrtc/offer was reachable by any authenticated user for
any node, with no membership check and no rate limit, making the target node
allocate an aiortc PeerConnection and gather ICE on demand — remote resource
exhaustion against a third party's machine. Now rate limited, capped per
user, SDP size bounded, and the caller must share an active group with the
node. That also closes the H4 gap where signaling ignored group status.
* POST /v1/nodes/{id}/incoming took peer_ip verbatim, so any user could make an
arbitrary node emit UDP packets to an address of their choosing. The probe
target must now match the caller's own source address.
C4 (partial) — the pre-proof bundle window. GEK and keypair bundle fetches are
served before the GEK proof by necessity: the client needs its wrapped bundle in
order to compute the proof. That window is a disclosure surface a hub can reach
by forging a JWT. Bounded to 4 fetches per session and audited as
"pre_proof_fetch". The real fix is removing remote keypair bundles entirely,
which belongs to the native client (Phase 13.3).
11.5.3 — the node admin UI was unauthenticated because it binds loopback. But
any local process can reach it, and so can a page in the operator's browser via
DNS rebinding — and this API re-initialises group keys and reads the audit log.
H2 showed script execution there equals full control. Now gated by a per-run
token, printed at startup, accepted as ?t= or X-MeshBay-Token.
One test needed rewriting rather than adding: the first version asserted
"subprocess.run(" was absent from the source, which also matched the comment
documenting the old behaviour. It now parses the AST and checks the property.
Tests: 121 node, 142 hub+common. Regression suite 47 node + 10 hub.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Phase 11.5 hardening batch — H7, H4, M2, M6, M7, L1, L3, L6.
H7 — private content hashes leaked to the hub. The daemon registered blake3
hashes for every group it hosted, private ones included, giving the hub a
content fingerprint of every private file and letting anyone confirm whether a
known file exists in the network. The leak was dormant only because the routes
were declared on the groups router with a full path and mounted at
/v1/groups/v1/swarm/* — the node's calls 404'd into a swallowed exception.
Fixing the path alone would have activated the leak, so both land together:
registration is gated on group visibility, the routes moved to a real
/v1/swarm router, and the lookup now requires authentication.
H4 — revocation was advisory. Group revocations were signed and broadcast by
the hub and then dropped by the node, whose handler understood only "user" and
"jti", so "suspend a group" enforced nothing. The denylist was also in-memory
only, so a restart silently un-revoked everyone. Now persisted to
data_dir/denylist.json, group targets honoured on both transports, and live
sessions for a revoked group are closed.
M2 — the node keystore, which protects the node's Ed25519 and X25519 private
keys, was still deriving at 64 MB long after the hub's password verifier moved
to 256 MB; the docs recorded the bump as done, true for the hub only. Raising
the constant alone would have made every existing keystore permanently
undecryptable, so envelopes now record the parameters they were written with
and pre-M2 files continue to open under the legacy profile.
M6 — registration inserted its audit row with a NULL user_id and then ran
UPDATE ip_logs SET user_id=<new> WHERE user_id IS NULL, claiming every
unattributed row in the table: failed logins for other usernames, concurrent
registrations. In logs retained a year for legal requests, that attributed
other people's connections to the wrong account.
M7 — X-Forwarded-For was trusted unconditionally at four call sites, so anyone
could forge the IP written to the compliance log and evade per-IP rate limits.
New netutil.client_ip honours the header only from a trusted proxy and takes
the rightmost hop (the one our proxy appended); no direct header reads remain.
L1 dead GEK_REQUEST/GEK_RESPONSE constants removed; L3 peer errors no longer
echo exception text (paths, internal state); L6 email sanity-checked instead of
accepting any string — deliberately not RFC 5322, to avoid a new dependency.
test_daemon_index_change_pushes_to_peers asserted that a PRIVATE group's hashes
are registered with the hub. Split: private asserts not-called (index push to
members still asserted), and a new test proves public groups still register.
That is the fourth pre-existing test found asserting a vulnerability as
intended behaviour, after gek auto-activation, the transport-wide chat_store
and the blind admin challenge.
Tests: 116 node, 132 hub+common. Regression suite now 43.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Phase 11.5 — findings H1, C5a, H2, C5b, H5 (see second-review.md).
Batched together because the node-side changes share webrtc_server.py and
cannot be separated into working commits.
H1 — cross-group chat leak. chat_store, the peer registry and the display-name
cache were read from the shared transport context, and daemon.py hoisted the
FIRST group's chat store onto it. On a node hosting several groups every
group's messages went to one database, chat_history served them back to members
of every other group, and chat broadcast reached all peers regardless of group.
All three now resolve through _group_ctx().
C5a — upload confinement. Uploads landed in the shared root under a
client-chosen name and overwrote whatever was there. Any member could destroy
the operator's files, and by becoming the recorded uploader of the replaced
file could then delete it through the uploader path, bypassing the Ed25519
admin challenge. Uploads now go to a per-user quarantine (.uploads/{user_id}/),
refuse to overwrite, and enforce chunk ordering, a filename allowlist and a
size cap.
H2 — stored XSS in the node admin UI. Filenames chosen by any group member were
interpolated raw into the localhost UI, which has no authentication, so script
execution there equals control of the node admin API. Now html.escape()
throughout, textContent in the audit table, plus CSP/nosniff/no-referrer. The
CSP contains exfiltration but cannot stop injected inline script — escaping is
the fix.
C5b — group key seizure. gek_bundle_store wrote whatever any member sent and
auto-activated bundles addressed to the node operator. The operator's X25519
public key is public (the node publishes it in handshake_ack), so any member
could wrap a key of their choosing for it and take over the group, locking
every legitimate member out. Storing now requires an operator signature and
_try_activate_gek is removed: nothing arriving over MNP can set a live GEK.
H5 — unbound signing oracle. The node challenged with 32 raw random bytes and
the client signed them blind, so a signature named no operation, subject, node
or time. New meshbay_common/adminop.py defines a length-prefixed,
domain-separated transcript; both sides build it independently and the client
refuses to sign when the announced op/subject do not match its request.
BREAKING: a group admin who does not operate the node can no longer store GEK
bundles on it. Invites must be performed by the node operator.
Adds tests/test_security_regressions.py. Verified against pre-fix source via
git stash. Three pre-existing tests asserted the vulnerable behaviour as a
feature and were inverted: gek auto-activation, and the transport-wide
chat_store in test_daemon.
Tests: 109 node, 132 hub+common.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Phase 11.5.A — findings C1 and C6 (see second-review.md).
C1: the per-group HTTP file API bound 0.0.0.0 for every configured group,
private ones included, and served two endpoints with no authentication at all:
GET /index (full Mesh Group Index) and GET /file/{id} (raw plaintext file via
FileResponse). Anyone able to reach the port — LAN, forwarded port, permissive
IPv6 — read every private file. This bypassed the entire GEK-proof and node
sovereignty layer. Deleted rather than patched: it duplicated MNP without any
of its controls.
C6: the TCP+TLS chunk server accepted a bare JWT with no GEK proof, leaving a
second non-compliant handshake path. Deleted; QUIC remains and will be brought
to parity with WebRTC by the unified handshake in 11.5.4.
Transport decision recorded in transport/__init__.py: WebRTC/ICE is primary for
browser and native clients (the only NAT traversal validated here — 2 ISPs,
IPv4 STUN + IPv6, 4G CGNAT); QUIC is kept for LAN, port-forwarded and hub-less
group:// access. punch_nat() is a direct-connection helper, not a traversal
stack.
Also removed server_ssl_context()/client_ssl_context() from tls_cert.py (no
remaining callers) and a dead import of the former in quic_server.py.
generate_self_signed_cert() stays: QUIC uses it, and the certificate hash is
the intended channel-binding anchor for 11.5.6, since QUIC has no DTLS
fingerprint to bind the GEK proof to.
BREAKING CHANGE: node.toml keys `port` and `http_port` are gone. Regenerate
config with `meshbay-node init`. Env var MESHBAY_PORT -> MESHBAY_QUIC_PORT.
Tests: 198 passed (209 - 7 test_http_server - 4 test_transport). No other test
changed status. Net -1300 lines.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Baseline commit capturing in-progress Phase 12 work that was already present
in the working tree (uncommitted) before the Phase 11.5 security remediation
begins. Committed as-is, without review or modification, so that remediation
changes arrive as a separable diff.
Contents: BundleStore (P2P GEK + keypair bundles), password split
(auth_key / bundle_key), node Ed25519 auth (POST /v1/nodes/auth, node-scoped
JWT), GEK-HMAC handshake proof with DTLS channel binding, Ed25519 admin
challenge-response, node local admin UI rewrite, browser key persistence.
Not authored in this session — captured to establish a baseline.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Add SQLite audit store for legal compliance (LCEN/DSA): logs user IP,
actions (handshake, file download/upload/delete, stream, chat), and
timestamps. Retention: 1 year, with cleanup method.
WebRTC transport now logs all user actions to the audit store with
remote IP extraction from the ICE transport.
Local web UI rewritten as a proper admin dashboard:
- Stats cards (groups, files, peers)
- Connected peers table with IP, username, group, state
- Group cards with file listings and shared directory info
- Audit log page with event/user filtering
- Dark theme, responsive, auto-refresh
- JSON API: /api/status, /api/groups, /api/peers, /api/audit, /api/config
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
|
| |
|
|
|
|
|
|
|
|
|
| |
When watchdog detects file changes, the daemon now:
- Pushes INDEX_SYNC to all connected WebRTC peers in that group
- Registers file hashes with hub /v1/swarm/register endpoint
Also registers all file hashes on startup for initial discovery.
hub_client: add register_swarm() method for bulk hash registration.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The node daemon was previously a skeleton that only started QUIC/TCP
servers and the local web UI. All browser-facing functionality (WebRTC,
hub WebSocket, chat store, HTTP file API) lived in QE demo scripts.
This rewrites daemon.py to be fully self-contained:
- WebRTC transport for browser clients (aiortc DataChannel)
- Hub WebSocket task (signaling, revocations, WebRTC offers)
- ChatStore per group (SQLite in ~/.local/share/meshbay/)
- HTTP file API per group (create_http_app on configured port)
- Graceful shutdown (all transports, stores, tasks)
- hub_client: _ws tracking + send_ws() for chat notifications
- config: data_dir field for persistent state
- systemd: security hardening (ProtectSystem, StateDirectory)
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
|
| |
|
|
|
|
|
|
|
| |
Replace download-then-play VideoPlayer with MSE (MediaSource Extensions)
streaming. Node remuxes to fMP4 via ffmpeg, probes codecs with ffprobe,
and sends encrypted segments over DataChannel. Browser decrypts and
appends to SourceBuffer — playback starts within seconds.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Complete browser-based client: Preact SPA with login, group file browser,
encrypted download, video playback, group chat, i18n, and dark/light theme.
Browser connects P2P to nodes behind residential NAT via WebRTC DataChannel
(aiortc). Hub handles signaling only — all data flows E2E.
Performance: pipelined downloads (8-chunk sliding window), binary msgpack
wire format (no base64), redundant I/O elimination. Large file downloads
stream to disk via File System Access API (showSaveFilePicker).
Validated on SFR + Orange residential NATs, Chrome + Firefox, IPv4/IPv6.
132 tests passing. Deployed to meshbay.org + Orange node.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Browser clients can now connect P2P to nodes behind residential NAT via
WebRTC DataChannel with ICE/STUN. Validated on SFR Port-Restricted Cone
NAT + 4G CGNAT across three scenarios (WiFi LAN, 4G IPv6, 4G IPv4 STUN).
No TURN relay needed. Hub serves only as signaling relay (<1 KB).
New files:
- webrtc_server.py: aiortc-based WebRTC transport (node side)
- signaling.py: SDP/ICE relay endpoint (hub side)
- transport.js: browser WebRTC client with msgpack framing
- webrtc-test.html: spike test page for browser→NAT→node validation
- test_webrtc_transport.py: 4 tests (handshake, file transfer, auth, guard)
- meshbay-draft-v4.md: architecture spec updated for web client
Modified:
- hub_client.py: WebRTC offer handling via hub WebSocket
- revocation.py: node_id from WS auth + webrtc_answer routing
- pyproject.toml: aiortc>=1.9 dependency
123 tests passing (117 existing + 6 new).
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Implements all 8 milestones (7.0-7.7):
- 7.0: JWT carries `groups` claim; node verifies group membership at
MNP handshake (QUIC + TCP+TLS). Resolves security review C2.
- 7.1: QUIC 0-RTT session resumption via stored session tickets
(17-21ms reconnect vs 47ms cold).
- 7.2: Hub→node WebSocket signaling for NAT punch coordination
(`client_incoming`/`punch_ready`) + jti denylist push. Denylist
class blocks revoked users/jtis at handshake.
- 7.3: Multi-group daemon — one QUIC port serves N groups with
per-group GEK, shared_root, and index routing.
- 7.4: HLS streaming via QUIC (STREAM_SEGMENT message type, ffmpeg
segment extraction).
- 7.5: Sender Keys protocol for group chat (Signal Groups approach).
Each member has own sending chain key, HKDF chain ratchet, AES-256-GCM
encryption, Ed25519 signing. Resolves security review C1.
- 7.6: Chat store (SQLite via aiosqlite), CHAT_MESSAGE MNP wire type
with peer broadcast, web UI with WebSocket push.
- 7.7: Argon2id calibration CLI.
First security review included (first-review.md). 109 tests, demo-v3
validated against meshbay.org production hub.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
| |
QuicChunkServer/QuicChunkClient: same MNP protocol over QUIC/UDP.
Enables hole-punching (Spike 4 Cone NAT validated). Uses aioquic 1.3.0.
Bug found+fixed: asyncio.Event race condition in client recv loop
(quic_event_received overwrote _stream_events[0] after _recv
created it). Fixed with asyncio.Queue (no shared mutable state).
Server uses synchronous handlers in quic_event_received (avoids
ensure_future transmit timing issue). 3/3 tests. Full suite: 50/50.
Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>
|
| |
|
|
|
|
|
|
|
|
|
|
|
| |
FastAPI app on port 19001: GET / (node info), GET /index (public
group index JSON), GET /file/{id} (full download), GET /file/{id}/{n}
(encrypted or plaintext chunk), GET /hls/{id}/playlist.m3u8 +
GET /hls/{id}/{n}.ts (HLS streaming via ffmpeg).
Public groups: index browsable without auth, files downloadable.
Private groups: chunks encrypted with GEK, auth required.
7/7 tests passing. Full suite: 47/47.
Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>
|
| |
|
|
|
|
|
|
|
|
|
|
|
| |
config.py: TOML + env var priority. auth.py: Argon2id passwords,
JWT EdDSA with jti, refresh token hashed (blake3). Routers:
hub (info/pubkey), users (register/login/refresh/pubkeys),
nodes (announce/get), groups (create/gek-bundle/gek-retrieve).
Rate limiting via slowapi. app.py factory with lifespan.
All 40 tests pass (SQLite in-memory, no PostgreSQL required).
Fix: remove tests/__init__.py to resolve namespace conflicts.
Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>
|
| |
|
|
|
|
|
|
|
|
| |
Self-signed TLS cert (RSA-2048, TLS 1.3 min). Server: JWT offline
verify, index sync, file_request → encrypt+sign chunk pipeline.
Client: handshake, fetch_index, fetch_chunk with Ed25519 verify +
blake3 hash check + GEK decrypt. Integration test: 2MB file served
in 2 chunks, reassembled == original. 3/3 tests. Full suite: 29/29.
Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>
|
| |
|
|
|
|
|
|
|
| |
GroupIndex: msgpack→zstd→GEK-encrypt→sign for private groups,
plaintext+sign for public groups. DirectoryIndexer: watchdog-based
watcher, async initial scan via thread pool, on_change callback.
Delta support (diff between versions). 10/10 tests passing.
Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>
|
| |
|
|
|
|
|
|
| |
Register/login/announce/token-refresh/GEK-unwrap. JWT jti and
pk_user verified at login. Hub PK cached after first fetch.
6/6 tests passing with httpx.MockTransport (no network).
Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>
|
| |
|
|
|
|
|
|
| |
Argon2id + AES-256-GCM encryption at rest. Unlock via env var
(MESHBAY_UNLOCK_KEY), unlock.key file (chmod 600), or interactive
getpass. 10/10 tests passing.
Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>
|
|
|
3-package layout: meshbay-common (shared crypto/protocol),
meshbay-hub (FastAPI server), meshbay-node (local daemon).
Includes validated POC spikes 1-6 in poc/, architecture drafts
v1/v2 in docs/, and CLAUDE.md project conventions.
All cryptographic primitives extracted from POC into
meshbay_common/crypto.py (GEK wrap/unwrap, chunk key derivation,
keystore encryption, chunk signing).
Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>
|