A video generation runs a single ComfyUI prompt for minutes, and the media
card only advanced after each prompt finished, so it sat still until the
clip was done.
The ComfyUI provider now follows ComfyUI's standard /ws stream while its
prompts run and maps step progress to the job's overall fraction across
all prompts: never backwards, and short of each prompt's share until its
outputs are downloaded. The history poll stays the source of truth for
completion and errors; a socket that cannot connect, drops, or stays
silent leaves generation exactly as before, and it is closed on finish,
failure, cancel and abort. Each generation uses its own client id, so
concurrent jobs do not take over each other's messages.
The socket goes through a new createOpenAiCompatibleWebSocket next to the
guarded fetch: same URL and key checks, same resolved-address lookup,
ws/wss derived from http/https, the same Authorization header, a message
size cap, and a hard drop when a server never finishes the closing
handshake. It uses the undici package's WebSocket, so no dependency was
added. A provider given a replaced fetch without a matching socket opens
none, so tests and emulators never reach the network.
The ComfyUI emulator now serves the websocket and reports progress for
running prompts; provider, guard and conformance tests cover ordering,
multi-prompt mapping, unavailable and failing sockets, stray messages,
closing on cancel, and the address checks against a loopback server.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
* chore: remove legacy orchestrator, trading and one-off script leftovers
The repository still carried the pre-evOo prototype: a Python orchestrator,
the old HTML "evoo-platform" cockpit (whose test ran in the root unit
suite), an apex-trading package, the hot-reload deploy.sh for it, a
permissions manifest, an architecture note with local file links, agent
handoff notes, a Steve Jobs persona template and about ninety one-off
scripts that probe a developer database with hard-coded local URLs. None
of it is used by the product, CI or the real deploy in ops/evo-deploy.
Delete those files, drop the evoo-platform test glob from the root vitest
config, the Cockpit section from .env.example and the dead biome and
gitignore entries, and update the docs that listed the removed tools.
scripts/start-tmux.sh no longer starts the removed sentinel and resolves
the checkout from its own location instead of a machine-specific path;
scripts/seed-ollama.ts requires DATABASE_URL instead of falling back to a
local URL with a username.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
* fix(web): drop trading and persona prompts from the conversation shell
The web shell still showed leftovers from the trading prototype: composer
chips with German trading and "Steve Jobs" prompts, trading suggestion
chips under the empty conversation line, a German trading hero for the
empty transcript, and a German "data protection" status panel in the
sidebar. They suggested actions the product does not have and added
persistent chrome.
Remove the composer chips and their helper, the empty-state chips and the
hero (the empty conversation keeps its existing one-line start hint), the
sidebar trading indicator and the data trust panel. The new-bot E2E now
asserts the plain empty state; the sidebar spec no longer expects the
trading status, and the placeholder Steve Jobs spec with no assertions is
gone.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
Co-authored-by: cynarAI <cynarAI@users.noreply.github.com>
* refactor: remove the trading bot status integration
With the sidebar trading indicator gone, the only remaining consumer of
the trading bot status was a banner in General settings that reused the
same component. evOo is a personal assistant; a status proxy and poller
for one external trading service, configured through service-specific
TRADING_BOT_URL and VITE_TRADING_BOT_URL variables, does not belong in
the core product.
Remove the settings banner and its prop plumbing, the web poller, the
indicator component, the /api/trading-bot/status proxy, the shared core
contract, the env loading and both variables from .env.example, together
with their tests.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
* chore: remove one-off database scripts and a stale review note
packages/db carried four hand-run scripts outside the package build that
query or seed a developer database with a hard-coded connection string,
one of them for a specific account. docs/VOICE-FIRST-REVIEW.md was a
one-time verification note for a workflow that no longer exists and
named private machines and local paths. Nothing references them.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
* chore: replace machine-specific paths and addresses with placeholders
ORIGIN.md pinned the intended checkout to an absolute path under a
personal home directory, docs/LOCAL-SETUP.md listed folders on one
developer's machine, and two tests used what looks like a real tailnet
address. Use a home-relative path, drop the machine-specific folder list,
and use 100.100.100.100 in the tests; it is still inside 100.64.0.0/10,
so both tests check the same thing as before.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
* test(web): drop the trading status from the shell contrast test
main's contrast test now waits for the sidebar trading status before
measuring; this branch removes that status.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
---------
Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: cynarAI <cynarAI@users.noreply.github.com>
Each project (space) now has a main agent and a main conversation of its
own. The user talks to it, and it delegates to the project's specialists.
Creating a project sets both up and opens the conversation, and deleting
an otherwise empty project removes them with it.
Claude-Session: https://claude.ai/code/session_01R8hSsKTB3W3e61wD6kvzur
Co-authored-by: Claude <noreply@anthropic.com>
Newer ffmpeg releases refuse to encode JPEG from limited-range video
frames, so poster frames failed on CI's ffmpeg while older versions
accepted the same command. The poster is converted to full-range
yuvj420p before encoding, which every supported ffmpeg version takes.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
The artifact pane from main loads and previews files through the streamed
artifact route like the file card, so its hook takes only the artifact id.
attach_file keeps media types and gains present, which marks image and
file blocks for the pane.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
The live screen opens wide beside the chat while the bot works on it or asks
for a takeover, a file the bot presents as its result opens next to the
conversation, and a question that scrolled away gets a "Needs you" pill back
to it. Shared core rules derive the focus from signals clients already have
(pointer events, takeover requests, a new `present` flag on attach_file,
pending asks), move the pane only on focus edges, never replace a pane the
user opened, and learn per bot whether the screen opens on its own. Pane
widths persist per device; mobile shows a pill that opens the computer.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01C7VSRHHzyt2KwoCkUkZ2Gc
Bots now announce every click and typing step before it runs, and the
web app draws a cursor in the bot's colour that glides to that point
over the live computer screen, ripples on clicks and shows typing.
- contracts: ComputerPointerPayload on the existing thread.computer event
- executor: computer_act and browser_act announce pointer steps
(best effort); the page-browser helper reports element screen boxes,
which are stripped before results reach the model
- web: agentPointerFeed + AgentCursor over the side-panel preview and
the full-window screen, mapped through the noVNC letterbox
- in-app browser: native WebContentsView on desktop (own session, http(s)
only, clamped bounds, back/forward, reload, DevTools, load errors) and
a sandboxed iframe on the web; replaces the dead panel, the fake
50%/40% cursors and the unimplemented devServer bridge
Claude-Session: https://claude.ai/code/session_01WiGooXtsKrZQmdszepRrNi
Co-authored-by: Claude <noreply@anthropic.com>
spaces.create linked any workspacePath to the new project, and
computer.files and computer.readFile then listed and read files under
that folder for anyone in the space. On a deployment with signups, any
member could link / or the API's own checkout and read whatever the API
process can read, such as its .env file. Linking also ran git in the
chosen folder and stored its remote URL.
A linked folder lives on the host, and the host is the deployment
owner's machine, the rule the onboarding scan already follows:
- spaces.create refuses a folder from anyone but the deployment owner,
before the space is created or the folder is touched.
- Project files are listed and read only for the deployment owner, so a
space a member linked before this change stays closed too. Space
navigation no longer shows a linked folder's path and git remote to
other members.
- A linked folder is stored as its real path, and listing and reading
follow symlinks only while they stay inside it. Symlinks that leave
the folder or lead nowhere are left out of a listing instead of
failing it.
- The web New project dialog offers the folder field only to the
deployment owner.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EyQw3AEaHYXegvgvTPAsi1
AGENTS.md asks for top-level `import type` instead of inline `type`
specifiers; this applies it to the import statements the media work added
or changed.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
The model saw edit_video and export_video with artifact_id next to
generate_image, generate_video, edit_image and transcribe with artifactId,
sourceArtifactId and negativePrompt. Mixed conventions in one tool list invite
wrong calls, so all media inputs now use snake_case like continue_after and the
other builtin tools: artifact_id, source_artifact_id, negative_prompt,
duration_seconds. Handlers, artifact checks, the media card's transcribe
action and the tests follow. Nothing has shipped, so stored job inputs need no
migration.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
Node's default 5-minute requestTimeout cut off a 500 MiB video upload on
any uplink slower than about 14 Mbit/s, so the video cap was unusable over
ordinary connections.
- The API server's requestTimeout is 30 minutes, enough for the largest
upload at about 2.5 Mbit/s. headersTimeout keeps its 60-second default.
- Each streamed upload aborts when the client sends no bytes for 60
seconds: the route answers 408, the partial file is deleted and no row
is written, so a slow-drip client cannot hold a connection for the full
request timeout. Time spent writing to disk does not count.
- The store and the upload guard no longer wait for a body to confirm its
cancellation, which a stalled or hostile client may never do.
- A guard test checks that every file the executor attaches records the
thread's group, which the render_plot chart fallback once dropped.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
Dictation falls back to the owner's transcription server when the voice
provider cannot transcribe, so that server is a voice recipient too: it is
listed for voice use (as "Transcription (<origin>)") whenever it would take
the recording, and not when the voice provider transcribes itself.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
A card that only says "Transcript" does not tell two transcriptions apart,
and it is English whatever the user's language.
- transcribe takes an optional title like the video operations, so a bot
names the card in the user's language; "Transcript" stays the default.
- Transcribe on a media card passes the source file name as the title.
- The E2E seeds the card through bootstrap as well as threads/get, since the
bot thread's first snapshot arrives with bootstrap.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
The image E2E now runs in the main conversation, which is a group thread,
so the browser test covers group uploads and reads from the artifact route
as well as bot threads. Bot file attachments keep their run's operation id
for tracing, and the non-streaming attach cap is the inline cap directly.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
Generated clips (nothing binary committed) go through extractAudio: a video's
audio reaches the server as 16 kHz mono WAV, a 12-second recording is
uploaded in two ranges under a small limit, and a video without sound is
refused before anything is uploaded. Skipped where ffmpeg is missing, which
may include CI runners.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
Edit cards showed English words such as "crop · flip · extend" to every user.
edit_image now takes an optional title written by the bot in the user's
language, used for the card and the file name like edit_video. Without one,
the card falls back to the repaint prompt or to language-neutral tokens only:
sizes, ratios, degrees, arrows and format names.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
Brings in ffmpeg (probing, extractAudio) and the video operations. The
transcribe registration now gets the ffmpeg audio tools, so video is
transcribed through extractAudio with the demuxer forced from the MIME type,
and long recordings are converted and uploaded in ranges.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
- A handler that ignores its abort is left behind after the grace period;
a late rejection from it (or from the grace timer after the job settled)
was unobserved, and the worker exits on unhandled rejections. Both sides of
the race are now observed and the grace timer is cleared with the job.
- Ending a job removes its temp directory when no runner in this process
holds it (leftovers of a runner that died here), and reconciliation sweeps
directories of jobs that no longer run.
- An executor test runs a media tool end to end through applyTool: shown
only for operations the space can run, started with the run's scope and
effect key, continue_after split off, refused when not exposed.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
A Postgres suite starts edit_video and export_video through the real
runtime and local artifact store: outputs become linked versions with
posters on the card, exports store one version per preset, appended clips
and subtitle files of another user are refused at start, and an
out-of-range cut fails the job with a readable error. It skips without
Postgres or ffmpeg.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
- The GPU guard waited for free VRAM at the operation's floor even on a card
whose total memory is below it (an 8 GB card and a 12 GB video floor), so
such jobs deferred until their budget ran out. The requirement is now capped
at 90% of the card's total; a card with less than half the floor fails at
once with "The GPU does not have enough memory for this job."
- Tests pin that media tools reach a run only through the space's
capabilities and within the bot's allowlist now that the image tool
definitions exist.
- Type-only imports are top-level `import type` in the media runtime, tools,
RPCs and tests (AGENTS.md).
- The ComfyUI settings test gives saveMediaSettings a store with nothing saved,
since the save reads the previous settings before validating.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
The card's follow-ups now reach the media job RPCs, and it only offers what
the space can run.
- media.capabilities answers which operations the actor's space can run
(booleans only). The card asks once per minute per space, and only when it
shows finished video or audio, so other cards cost nothing.
- Export and Transcribe call media.jobs.start with the card's message, so the
same bot posts the new card; Cancel calls media.jobs.cancel. A refused
start shows the server's message under the card.
- New files use top-level import type.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
Picks up ComfyUI generation, media tool dispatch and the streaming runner.
The video tools now share the image tools' schema helper, and both sets of
operations join the runtime's operation list.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
Brings in the media job RPCs, generic media tool dispatch and the image
operations; transcribe joins the operation list and the tool catalog next to
them.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
Agents only received the documents of a turn on their computer and could
attach workspace files up to 10 MiB, so they could neither cut a clip a user
sent nor post the video they rendered.
- Current-turn images, video and audio are copied into the bot's workspace
(Team Computers under bots/<botId>/) like documents. Remote computers move
files as base64 JSON, so anything above their 12 MiB transfer cap stays in
storage and the agent is told it by artifact id instead of failing the
turn; local computers copy up to the kind cap.
- attach_file accepts video and audio by extension and attaches up to the
kind cap when the store streams (still bounded by the transfer cap of a
remote computer), stored through the shared artifact helper with a real
SHA-256.
- The render_plot fallback that attaches a chart PNG now records the group,
so the chart opens in group conversations.
- Agent history names image artifacts by id, like video and audio, so a
later turn can refer to them.
- artifacts.get refuses files above 25 MiB: base64 reads hold the whole
file in memory, and larger media is served by the artifact route.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
A transcript is whatever was said in the recording, so the summary the bot
is woken with quotes it and says it is not instructions.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
Binary paths, the ffmpeg version, the probe's container name and a
configurable stall timeout had no caller; detection is cached per process,
so every helper looks the binaries up itself.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
Media tools and card actions send prompts, images, video and audio to the
deployment owner's ComfyUI and transcription servers. Like the embeddings
server behind built-in memory, those self-hosted endpoints are recipients in
the AI data-sharing disclosure: they appear with the model uses ("ComfyUI
(<origin>)", "Transcription (<origin>)", keyed by the full endpoint so a new
server asks again), and starting a card job asks for model consent on clients
that run the foreground check.
Also adds unit tests for the media settings store: URL normalization and
refusals, encryption, the key-free view, one decryption per saved version, key
retention per server and removing empty settings.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
Covers what the unit tests fake: the job starts from validated input with
the "Transcript" card, runs against the transcription emulator, stores the
.srt and .vtt as artifacts next to the source and lists them on the card,
and an artifact of another user is refused before a job exists.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
With the media runtime in place, transcription becomes an operation like the
others: bots call the transcribe tool, the job runs in the background and
the card shows the subtitle files.
- createTranscriptionOperations registers transcribe: available when a
transcription server is configured, refuses artifacts the user cannot
read before the job exists, and writes <name>.srt and <name>.vtt next to
the source (a .txt when the server returns no timings). With ffmpeg the
audio is converted to 16 kHz mono WAV and long media is transcribed in
upload-sized ranges on one timeline; without it audio is sent as stored
and video is refused with a clear message. A server that is down or busy
(unreachable, 429, 502, 503) defers the job instead of failing it, since
nothing has been stored yet. The bot gets the transcript, cut at a word.
- The transcribe tool definition uses the shared media schema helper and
continue_after; dispatch stays generic.
- The main bot's run creates Studio once a media server is configured,
right before the specialist directory is built.
- Dictation: when the user's voice provider cannot transcribe (or there is
none), /api/voice/transcribe uses the self-hosted transcription server,
and the voice status reports server dictation so the web app uses it
instead of the browser's speech recognition. A voice provider that
transcribes keeps being used as before.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
Bots and users need to start and follow media jobs, and the deployment owner
needs to point the deployment at their GPU and transcription servers.
- Media tools are named like their operations. A run sees one only when the
space can run that operation (MediaRuntime.capabilities); the per-bot
allowlist still applies. applyTool hands every media tool to one
dispatcher (media-tools.ts) that splits off continue_after, starts the job
under the tool call's effect key (a replay returns the same job) and returns
at once; refusals (invalid input, unknown artifact, unavailable) go back to
the model as tool errors.
- media.jobs.start starts an export, transcription or image edit from a card:
the conversation is resolved like every thread procedure, the card's bot
(or the group's lead) posts the new card, and the runtime refuses source
artifacts outside the actor's space. media.jobs.get and media.jobs.cancel
only see the actor's own jobs.
- media.settings.get/update are deployment-owner only, never return keys and
answer invalid URLs with a 400 that says what is wrong.
- createApp accepts media operations as a test seam.
- Tests: a Postgres suite over the RPCs (owner-only settings, encryption,
key retention per server, URL policy, start/get/cancel isolation between
users, card jobs end to end) and unit tests for tool exposure and dispatch.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
The image server is the deployment's, not necessarily the user's, and
repainting an edit only works when generation is available.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
edit_video and export_video now run as media jobs: both registrations join
the shared operation list the API and the worker build their runtime from,
name every artifact they read (source, appended clips, subtitle files) so
starting a job refuses artifacts the user cannot read, and their tool
schemas use the runtime's continue_after helper.
Container deployments get ffmpeg, ffprobe and the subtitle font in the
API/worker image. Tests cover the handlers against a fake job context:
cuts, blurred 9:16 reframes with subtitles, appended silent clips, audio
extraction, single frames, batch exports with posters and progress,
rotated phone video, clear input errors and cancellation. They skip where
ffmpeg is not installed; the planner and parser tests always run.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
Queue, interrupt and prompt requests now go through one JSON helper, so the
cancel path reads like the rest of the client.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
Bots can now make images and short clips on the user's own GPU server and
rework images locally, through the media job runtime.
- generate_image: prompt, aspect preset or exact size, up to 4 images, optional
source image for image-to-image. Sizes target what the default models are
tuned for, and sources are fitted to that size before upload, so large
photos cannot exhaust GPU memory.
- generate_video: prompt, aspect, duration up to 10 s, optional start frame.
- edit_image: crop, rotate, flip, resize to a size or format (crop, pad or
blurred background), extend, convert, in that order, with sharp; an optional
prompt repaints the result on ComfyUI. Edits keep the source as parent.
- Registrations declare the input artifacts so jobs refuse ones the user
cannot read, and generation declares a fixed GPU memory floor (6 GB images,
12 GB video) that a custom workflow cannot lower.
- Tool definitions reuse the operation input schemas plus continue_after.
- Saving media settings rejects custom workflows the provider could not fill.
- Tests: sharp edits on generated fixtures, handlers against the ComfyUI
emulator, and a PostgreSQL suite running the operations through the media
runtime (card outputs, artifact versions, foreign artifacts, GPU wait).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MRtGG9ARfqqhma3t85Ro2u
