An attachment downloaded and then rendered as a broken image. A blob URL
is decoded according to the blob's type, so handing the webview
application/octet-stream makes img refuse bytes that are a perfectly
good PNG — and octet-stream is exactly what the client substitutes when
the server names no type.
Prefer a concrete image type from the server, otherwise identify the
format from its leading bytes. Every signature checked is a fixed marker
at a fixed offset, so this reads no user content, and a JSON error body
served with HTTP 200 is deliberately not mistaken for an image.
Whether that was the cause is not yet established, so the alternative is
made self-diagnosing rather than guessed at. The download outcome —
status, the type the server gave, the type the bytes indicate, and the
length — is now traced, and the frontend reports why an attachment did
not render instead of leaving a bare broken-image icon: a failed
download shows its error, and a payload the browser cannot decode says
so. Both paths were previously swallowed.
Refs: N-6.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The first capture identified the real cause of a reply appearing twice,
and it was neither candidate. All 524 frames carried a message_id and
each reply used exactly one, so the event_id fallback is not involved
and neither is the history session fallback.
AstrBot interleaves two kinds of plain frame in one reply. Frames with
streaming true are incremental deltas whose total matches the complete
frame exactly. Frames with streaming false carry whole finished blocks —
one reply had five of them totalling 915 characters against 173
streamed. map_plain appends both, so the blocks show up inline during
generation, and map_complete then replaces the accumulation with the
streamed text alone, discarding them. That is the reply appearing twice
and then collapsing, and the collapse loses content.
One thing lengths cannot settle: whether a block repeats streamed text
or carries its own. One reply had 133 block characters against 142
streamed, which looks like a repeat; another had 915 against 173, which
cannot be. The two need opposite fixes, so this adds a per-stream
summary reporting how many blocks appear in the final text and in the
streamed text, as counts and lengths only.
Also records that plugin images were never sent: no media or attachment
frame appeared in any capture, and no plain frame carried a message
chain array. The attachment path has still never received data.
Refs: N-9 (new), N-4 (root cause found), N-6 (not a client defect).
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The first capture answered the question the trace was built for and
eliminated it: all 77 plain frames of one reply carried a message_id, so
a reply splitting into two bubbles is not the event_id fallback in
F-34.
What remains depends on the value rather than its presence. If the id
changes partway through a reply then the client is correctly showing two
messages the server sent; if it stays the same then the split is ours,
in the historySessionId fallback that loads another session's history.
The id is an opaque row identifier, not content.
Also record the length of a string data payload. It separates "the
server streamed many small chunks" from "the server sent the reply in
one piece" without recording any text.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The event set an explicit target of astrbot::frame while the enabled!
guard used the default module-path target. A filter naming either one
therefore failed: naming the crate did not reach the event, and naming
the target left the guard returning false. Either way the trace emitted
nothing, which would have looked like the application producing no logs
at all and wasted a reproduction cycle.
Give the guard the same target, hoist it into a constant, and add a test
that drives the exact RUST_LOG value printed in the acceptance
instructions. The test was checked against the broken guard and fails
with empty output. A second test asserts the trace stays silent at the
default level, and both assert no payload text or attachment URL reaches
the log.
Document the capture procedure in the acceptance checklist.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Two reported defects cannot be diagnosed from the code alone. A reply
sometimes renders as two bubbles during generation, which happens when a
streaming frame carries no message_id — the client then keys it by
event_id, which is unique per event. And images sent by plugins never
appear, which happens when an image frame is not followed by the
attachment_saved frame the mapper waits for. Both depend on what the
server actually sends, and picking the wrong one would produce a fix
that looks right until another conversation path hits the other.
Log the shape of every SSE frame: its type, whether message_id and
session_id are present, the streaming flag, and the key names inside
data. No payload text is recorded, so this is safe to leave enabled in
the field. Off unless RUST_LOG selects debug, and gated on
tracing::enabled! so a disabled level costs one atomic read rather than
a JSON parse.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
send_connection_slot_message wrote the optimistic user row from the
runtime snapshot alone, then called send_slot_message, which can still
refuse the send for an empty message, a busy slot or a disconnected
slot. The rejected question stayed in history with no failure marker,
and because each attempt minted a fresh local- id, retrying appended a
second copy rather than replacing the first.
Add a read-only pre-flight, ensure_slot_can_send, and run it before any
database write. The conditions it checks are extracted into
slot_send_precondition, which send_slot_message now uses inside its own
critical section, so the two cannot drift. Mint the message id once so
the failure path can address the row it wrote.
A send that fails after AstrBot accepted it must keep its question:
cancelled and streamInterrupted both mean the message is already in the
remote conversation and a partial reply may already be stored. Those
two keep the row; every other code marks it failed. server_received
encodes that rule and is covered by a test that walks every
AstrBotClientError variant through astrbot_send_error, so a future
rename of an R2 error code cannot silently change the classification.
Marking rather than deleting keeps chat_sessions.last_message_id valid,
re-uses the idempotent upsert, and leaves unread counts alone since they
already ignore user rows. The queue sanctions either choice.
The R2 retry loop and its error mapping are untouched; all five R2
transport sentinels still pass.
Refs: F-18 (audit P2-3).
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
probe_connection_slot never consulted the slot assignment table, so the
same profile could be activated into both runtime slots at once. That
bypasses the ProfileAlreadyAssigned policy the database enforces, and
both slots then race to overwrite the same sticky_session_id.
Enforce it in three places. The runtime invariant lives inside
activate_slot's claim phase, so it also covers startup auto-connect,
which calls activate_slot directly and would bypass a command-layer
check. The command adds a cross-check against connection_slots. And the
operations switch guard now covers probe and disconnect, not just slot
assignment.
Both rules are deliberately negative — they ask whether the profile is
bound to the *other* slot, never whether it is bound to this one.
auto_connect_configured_slots activates from profile.auto_connect_slot
and never writes connection_slots, so a positive rule would break
reconnect for every auto-connected profile. For the same reason the
probe guard only refuses a change of profile, so reconnecting the
profile the operations slot already holds still works during a task.
The operations guard has no production writer yet (F-74, audit P3-7),
so it is inert at runtime. It ships with a test that flips the guard
directly rather than as an unexercised branch.
Refs: F-19 (audit P2-4).
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
switch_slot_session released the slot lock while set_session_id ran and
then re-acquired it to write active_session_id, with no check that the
slot still belonged to the same activation. A deactivation or
re-activation in that window put the old session id onto the new
occupant's snapshot, and persist_session_identity then copied it into
that profile's sticky session.
Capture the generation alongside the client in the first critical
section and compare before writing. The write-back is extracted into
commit_session_id so the invariant can be tested directly rather than
raced.
Refs: F-17 (audit P2-1).
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
activate_slot held the single mutex that covers both slots across
health() and probe(). Those go through the REST client, whose total
timeout is 120s, so one unreachable address froze every runtime command
on both slots for that long: status polling, sending, cancelling and
disconnecting all blocked on the same lock. The cancellation token was
also only created after a successful probe, so the activation itself
could not be interrupted.
Split the function into three phases. Validation and client
construction happen first with no lock held; a short critical section
then bumps the generation, installs the cancellation token up front and
publishes the Connecting snapshot; the probe runs outside the lock
inside a biased select against that token; and a final critical section
writes the result back only if the slot still belongs to this
activation.
The listener receives the same token that phase one installed. Creating
a second one there would leave deactivate_slot cancelling a token no
listener observes, so the desktop-v1 reconnect loop would outlive the
disconnect.
This also fixes a stuck-slot bug the queue did not name. Validation used
to run after the Connecting snapshot had been written, so a rejected
request tore down a working connection and left the slot in Connecting
forever, with no client and no last_error to explain it.
The three new tests were checked against a mutation that re-acquires
the lock across the probe; the two concurrency tests fail under it.
Refs: F-14 (audit P1-4).
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
`should_retry_astrbot` treated every `reqwest::Error` as retryable, but that
class covers both "the POST never left" and "the response body broke". `POST
/api/v1/chat` carries no idempotency key, so retrying the second case appended
the user's message to the remote conversation a second time, produced a second
answer, and interleaved the two attempts' deltas into one garbled bubble.
Retries are now limited to connect-phase failures, the one case where the
request provably never reached the server. Anything after that surfaces a
`streamInterrupted` code so the UI can offer a manual retry instead.
Also stops a send from checking the client out of its slot. The other five slot
commands read the client and report "not connected" when it is absent, so a
send disabled sessions, capabilities, upload, download, and session switching
for its whole duration. That window used to be bounded by the client's 120s
total timeout; now that streaming has no total deadline it would last as long
as the reply, so an ordinary long answer would disable the rest of the slot.
The client is cloned instead and concurrent sends are rejected with `slotBusy`.
(P2-13 / F-49, pulled forward from batch R7 because this batch is what made it
user-visible.)
Refs P1-2 / F-12.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Combines the P0 event-id fix with the three transport fixes that rewrite the
same stream loop; they are not separable in these two files.
fix(astrbot): scope stream event ids per request (P0-1 / F-01)
`AstrBotStreamMapper` restarts its sequence at zero for every stream, but
`seen_event_ids` lives on the client and outlives a request. Event ids were
built from `server_instance_id + sequence` alone, so the second message on a
connection replayed the first message's ids and every one of its events was
discarded as a duplicate: the assistant's reply never reached the UI or the
database. Ids now carry a per-request `stream_id`.
fix(protocol): serialize model options in camelCase (P1-7b / F-07)
`AstrBotModelOption` had no `rename_all`, so the frontend's `isDefault` was
always undefined and no model was preselected.
fix(net): stop total timeouts from killing streams (P1-1 / F-10)
`Client::timeout` runs from the start of the request until the body is
finished and keeps counting while `bytes_stream()` is consumed. The 30s
desktop-v1 budget therefore tore the long-lived SSE connection down every 30
seconds no matter how healthy it was — heartbeats could not keep it alive —
and the 120s AstrBot budget cut off any generation longer than two minutes.
Streaming now uses a separate client with no total deadline, detecting a
stalled peer with `read_timeout` (which resets on every successful read) plus
a TCP keepalive. REST endpoints keep their total deadline.
fix(astrbot): decode sse bytes on line boundaries (P1-3 / F-11)
Decoding each network chunk with `String::from_utf8_lossy` corrupted any
character straddling a chunk boundary. For CJK text that is one replacement
character per unlucky chunk, persisted into the message. The stream is now
split into lines at the byte level and only complete lines are decoded.
fix(astrbot): finalize streams on cancel and error (P2-2 / F-13)
Cancellation and every error path returned without a `MessageCompleted`
event, leaving the row `streaming = 1` so it rendered as "generating" again
after a restart. All exit paths now finalize with `Cancelled` or `Error` and
keep the text received so far. Cancelling also no longer flips the connection
to Disconnected and back, which was a visible flicker (P3-4 / F-56).
Hardening found while reviewing the above: a 64 MiB per-response ceiling
(removing the total timeout removed the only bound on a response, and
`read_timeout` cannot replace it), and a cursor-based line buffer instead of
draining from the front, which was quadratic in the chunk size.
Tests drive a real `AstrBotClient` against a hand-rolled HTTP/1.1 chunked-SSE
server on a raw `TcpListener`: byte-exact control over chunk boundaries and
dropping a socket mid-body without the terminal chunk are not expressible
through an HTTP framework, and both are needed here.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>