- LICENSE is the plain Apache 2.0 text, without the "Portions of this
software" preamble, so GitHub's About panel names the license. The two BSL
folders keep their own LICENSE, which names each folder as its scope; the
README's License section and the Terms page now name both, and the Terms
point at the LICENSE files rather than the root one alone.
- /plugins: the eight plugins shipped in 2.2.1, with what each reads, how to
start (download, free 30-day licence, Settings › Plugins, @ in a chat),
third-party plugins and the registry, and why Reddit and Instagram are not
offered.
- /pricing: 7 platforms and the web crawler, not 9; the plugin note, the
licence and trial answers and the expiry answer describe the shipped
plugins; third-party plugins need no licence.
- Homepage plugin band: no longer "Coming soon".
- README: a Plugins section linking /downloads for the free licence, in place
of Star history.
A CMake from MinGW or MSYS2, first on PATH on many Windows machines,
defaults to Ninja, which refuses -A x64 and finds no compiler outside a
Visual Studio prompt, so pnpm dev stopped at configuring audio.cpp. The
recipe now names the generator vswhere finds, such as "Visual Studio 17
2022", instead of leaving CMake to choose.
Picking Needs approval for a first-party Read-only group switched blocked
tools on, and they then ran without asking. A group choice now skips any
tool the approval rules keep from it, and the menu closes a choice no tool
in the group can take.
Base UI stacks popups by portal order and covers everything opened before
a modal dialog with its own click blocker, which carries no z-index. Our
wrappers put z-50 on every popup, lifting a parent dialog above its child's
blocker so it stayed clickable. Follow Base UI's setup instead: #root gets
isolation: isolate and the popup wrappers drop z-50, so every nested
dialog, alert dialog and popup stacks by the order it opened. The toaster
portals to <body>, outside #root, to stay above dialogs.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Raw <button> elements become Button; ones with their own layout (cards,
stretched row links, flip surfaces) use a new bare variant that keeps the
focus ring, transition and press without the look. Press feedback lives
only in Button: a 0.95 scale on every button, menu triggers included, in
over 150ms and back over 250ms on a no-overshoot curve. Wide buttons and
rows pass static and press by colour instead. The per-button press nudges
are gone, and ESLint now rejects a raw <button> or press motion in a
Button's className.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Each tool row trades its switch and Ask/Always allow control for three
icons with tooltips that say what each does, or why the approval rules
close it. Descriptions move to a hover on the name, and a blocked tool's
name is muted. Read-only and Can make changes fold away, with a count and
a menu of the state their tools share (Custom while they differ) that
sets every tool at once. A pick is one write of enabled and approval.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The settings dialog grows to 1072x768 and its sidebar to 200px so long
labels fit; both are capped to the window. The add-plugin dialog moves
from max-w-md to max-w-lg. Sizes use Tailwind scale classes.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Raise the secondary token in both themes so secondary buttons no longer
blend into hovered rows. Browse cards move off the token to a muted fill
with a softer hover, and their Connect button is now secondary like the
connected table's.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The Model setup chat test caught httpx errors but not the new
StreamedError, so a model that failed mid-reply answered 500 instead of
the 502 "chat request failed: ..." a refused request gets.
After dev's one-click ChatGPT sign-in, the local-only speech section must
still offer no remote model; a test now proves it with a plan that claims
the slot, and fails without the localOnly gate. The docs drop two stale
Known gaps (no speech screen; no transcription test), name the settings nav
as it reads, count migrations to 0030, mark steps 2 to 7 done in the
proposal, and note the curated recogniser's type in the catalog proposal.
- Google Search: Google's sorry wall is recognised by host and path, not a
"/sorry" substring, so a result linking to a page with "sorry" in its URL
no longer burns the wave's exit (goto and the captcha check share
captcha.is_sorry_url).
- Walmart: a redirect still pending after the last allowed hop is a refused
try, so the next exit is tried instead of reading an empty page.
- Amazon: a hedge whose sticky lease is unavailable falls back to the
gateway instead of spending the call's hedge budget on nothing.
- Fixtures: live tracking and session tokens scrubbed from the captured
pages (Indeed, Amazon, Google, Google Maps, TikTok, YouTube).
- plugin-host.md: shared_access() everywhere, and the Reddit and Instagram
gaps now describe their fail-fast latches.
- When GET /mcp fails, the turn used to go on with only SurfSense's tools
re-registered, so the session kept the last turn's rules and a held plugin's
tool switched off since could still be offered. Without the server list
there is nothing to deny by name, and a "*" deny would also block read and
edit, so the turn now ends with a network error before the model is asked,
as it does when the rules cannot be set. This predates the PR.
- The plugin proxy closes its upstream connection on every path, including
a failed 200 write and cancellation.
dev took ADR 0055 for the plugin registry's Verified rule, so the speech
recognition record is now 0056, Accepted with the change it decides, and
every link to it follows. Its 'Where the code stands' section went: the
code now holds the whole decision.
The history read that follows a failed Agentic message reached
ready_opencode too and touched the configuration again, so one failed
message restarted opencode twice and raised two toasts. A history read
now leaves the configuration alone and fails at once with Electron's
reason.
Also backdates the earlier failure file in its test, since Windows file
times can tie it with the configuration written right after.
The interface half of local speech recognition: Settings > Speech recognition
and its set-up popover, recordings in the sources panel, the microphone in
Electron, and dictation in the composer.
Review of the step 5 backend half (553048c02).
- An Ogg Vorbis recording whose path held a character outside the Windows
ANSI code page (a file named 会議.ogg, or a data folder under an account
named José) was refused for good as "This recording could not be decoded.":
miniaudio's vorbis_get_file_info opens the path with fopen, while its stream
reader opens a wide one. The decode child now reads an Ogg file's length from
its own pages, the last page's granule position over the rate its first
packet names; it matched miniaudio's reading on every fixture.
- A speech model chosen midway restarted the recording under the old language
intent, so the header read "Japanese" and the checkpoint kept "ja" although
the new model, whose settings start empty, detected the language. The job now
asks the status route again and starts over under what it reports.
- A cancel of a recording queued again after giving way left its decoded audio,
115 MB an hour, with no run left to clear it. The cancel route now drops it
and progress.json, keeping the plan and windows for a retry, as a run's own
end does (recordings/ended_run.py, shared with the worker, best-effort for a
file another process holds). A cancel that lands just as a job would give
way now ends it, instead of pause_job refusing quietly with the audio kept.
- build-audiocpp.yml's path filter takes modules/documents/recordings/, which
the end-to-end gate reads too.
Step 5 of the local speech recognition proposal, backend half. A recording
uploaded as a source becomes a timestamped transcript in its content, so
chunking, search, citations and the agent take it unchanged, without one long
file holding every other upload behind it on the one-thread ingest queue.
- Upload: .wav .mp3 .flac .ogg .oga with server-owned audio/* types, checked
by their bytes (RIFF/WAVE, ID3 or an MPEG frame, fLaC, Ogg with a Vorbis
first packet). Ogg Opus is rejected per file: "Opus support comes in a
later update". can_hold_figures is false for audio/*, so Docling never
reads a recording.
- worker/audio_decode/: the worker's own binary, started as
`worker --decode-audio <in> <out>` ahead of the consumer's imports, checks
the disk from the file's header length, streams it through pyminiaudio
(MIT) to 16 kHz mono s16 and exits when its parent does. A malformed file
kills the child, not the consumer. miniaudio joins pyproject, uv.lock and
worker.spec (with its licence metadata).
- worker/ingestion/audio/, reached from parsing before Docling: readiness
through the API's status route, the decode, Silero over the file a minute at
a time (the API's reader, now able to stream), spans padded and joined,
windows of at most 28 s cut at the longest pause or a run's quietest
100 ms, each window sent to /internal/transcription/windows with retries and
waits, windows.jsonl and progress.json as the checkpoint, a progress event
per window, then the header and [hh:mm:ss] paragraphs, segments.json and
transcription {language, duration_s, model} in the row's metadata.
- Yielding: after a window or a 503 waiting, once a minute has passed and a
source that is not a recording is pending, the job raises JobYieldedError;
the pipeline sets the row back to pending (jobs.pause_job) and queues it at
PRIORITY_BULK. Recordings never yield to each other.
- No speech model: the job fails without retry and marks
transcription.awaiting; choosing a speech model, by the route or an install
with select, queues every waiting recording again. Refusals (no room, no
speech, no audio, undecodable, memory short for 10 minutes) are
CancelExecution(retry=False); every other retry resumes from windows.jsonl
when the model and language intent match.
- DocumentRead gains transcription {language, progress, awaiting} for audio
rows, progress read from progress.json; InternalEvent and the SSE frame
gain an optional transient progress {done_s, total_s}.
Measured here: an 8.6-minute file from pending to ready in 60 s at 5.83% word
error; an hour of MP3 decodes in 3.6 s; miniaudio's linear resampler read the
44.1 kHz copy at 5.91% against scipy's 5.83%, under the proposal's 2% relative,
so it stays. Docs: documents.md gains Recordings; data-model.md, overview.md,
selection.md and packaging.md follow; the proposal records the two step 5
measurements.
Review of the speech recognition API slice.
- A transcription request past its deadline raised httpx's ReadTimeout out of
the client, so the phrase and windows routes answered a bare 500 instead of
503 transcription_unavailable, which the worker retries and the composer
explains. Any HTTP error the client meets, and a 200 with no JSON text, now
become TranscriptionUnavailableError; a timed-out request is not asked again.
- The cleanup's character-level collapse for zh, ja, yue and th counted a
number's digits as characters, so 100000 in a Chinese transcript became 10,
and Hmmmm became Hm. A run of digits or Latin letters now counts as one unit.
- A WAV cut mid-sample passed the check with half a sample, which made Silero
fail (the clip then skipped the speech check) and reached the server as is.
The check keeps whole samples only.
- The keepalive and close routes were sync handlers, so they changed the
dictation sessions from the threadpool while the gate read them on the event
loop. They are async now, like every other use.
Docs: the proposal's Transport table gains the windows route's `busy` wait and
a phrase's 409 transcription_needs_memory after a hold, both of which the step
documented in admission.md only; the cleanup rule names the unit; admission.md
says a timed-out request answers 503.
Each item gets a Hugeicons mark in a tile beside its title: book for the
blog, megaphone for announcements, history for the changelog, envelope
for contact.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Links sit straight on the page ground over a full-width engraved scene,
replacing the closing download pitch and the card on the sky painting.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Reddit refuses logged-out .json and Instagram's profiles sit behind a login.
Reddit now gives up after one rotation instead of about 2m50s and latches the
refusal, so calls answer at once with an error naming the countries tried.
Instagram no longer rotates on its require_login 401 and points to /p/ and
/reel/ URLs, which still work. Both use the shared license cache, progress and
text results.
- Pages are crawled level by level, several at once, with an exact page budget.
- HTML is capped at 1 MB before trafilatura and parsed once: a 7.3 MB page
took 17-25 s to extract, now about 1.5 s. Extractors warm up at startup.
- The tier ladder follows the HTTP status and challenge markers: a 404 fails
in about 1 s instead of trying the browser tiers. The dynamic tier runs on
the shared browser pool with a bounded settle.
- Output is bounded: links and contacts are capped with totals, and every
page gets its share of the agent's text. A 3-page crawl reply went from
1.85 MB to 49 KB.
- Fixes: GitHub and GitLab URLs filed as social, 404s and Cloudflare pages
returned as success, loadedUrl ignoring redirects, links resolved against
the pre-redirect URL, the oldest date on the page used as its date,
permalink anchors in titles, and raw HTML from the stealth tier.
A 3-page docs crawl went from about 28 s to 4-8 s.
- Comments and trending read TikTok's unsigned JSON endpoints, with the
browser as fallback: comments went from 25-96 s to about 2 s.
- Listings run on the shared browser pool, with event-driven waits, a JS
click, no home-page warm-up, media blocked, a hedged second country, and
concurrent targets; profile and video pages go over HTTP.
- Geo-blocked and login-bounced exits end the attempt at once.
- user_search recovers an empty first body inside the page instead of
starting a new browser, which made it intermittent.
- Fixes: textLanguage, isAd, isPinned, isSlideshow, mentions and slideshow
media dropped by parse_video, business accounts marked verified, withheld
comment pages, and the profile item counting toward max_items.
- Review pages after the first are fetched in parallel and deduplicated;
start URLs and search terms run together, in order.
- Fail-fast transport with hedging; 444, 403, 5xx and a /blocked redirect are
blocks that rotate, not pages; raw bytes and orjson instead of the lxml
round trip.
- A failed detail returns the listing card with detailsError instead of an
error item and an extra listing fetch.
- Fixes: variant products with 0 stars and 0 reviews, the sponsored flag and
search provenance lost on details, duplicate reviews at page boundaries,
reviews of other variants unlabelled, specifications and fulfillment always
null, and lone surrogates in review text.
200 reviews went from 34-99 s to about 10-13 s. WALMART_PARALLEL=0 restores
sequential fetches; WALMART_CANONICAL_URL=1 (off) skips the /ip redirect.
