* Open the workflow write-access link in a browser tab by default
Launching a `gr.Workflow` locally now opens the write-access link in a new
browser tab automatically, the way `jupyter notebook` opens its token URL,
so you land in an editable canvas instead of a read-only one.
Only done when a browser on this machine is plausibly the right place to
open it: Spaces, headless Linux, containers, CI/pytest, Colab and SSH
sessions keep the previous print-the-link-only behavior. An explicit
`inbrowser=` still wins, and `GRADIO_WORKFLOW_INBROWSER=false` opts out.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* Call vision-language models through chat completions in gr.Workflow
`image-text-to-text` mapped to the `visual_question_answering` endpoint,
which no Inference Provider serves for any model on the Hub — the router
routes every model carrying that tag as `conversational`. VLM model nodes
therefore always failed with "Task 'visual-question-answering' not
supported for provider ...", and even where served, VQA has no token
budget for emitting a whole file.
Adds a `chat_completion` endpoint schema (image + prompt → text) and
points `image-text-to-text` at it, on both the Python and canvas side.
Chat images are inlined as data URIs because the provider, not this
process, dereferences them, so a local path or a relative
`/gradio_api/file=` URL is unreachable.
Also adds demo/workflow_vlm_html_comparison, comparing two recent VLMs
on the same screenshot → webpage task.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* Don't clobber the installed skill when the agent skills dir is a symlink
`gradio skills add --claude` resolves the agent's skills directory with
`Path.resolve()`, which follows symlinks. When that directory is a link to
the central location — e.g. `.claude/skills -> ../.agents/skills`, a common
way to share one skills dir across agents — the link path collapses onto the
skill that was just installed, so `--force` deleted it and replaced it with
a symlink pointing at itself. Without `--force` it reported the fresh
install as a pre-existing conflict.
Skip linking when the resolved link path is the central skill path: the
skill is already reachable there, and no link is needed.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* Stream chat completions in gr.Workflow so long generations survive the router
The router returns a bare 504 for a non-streaming /v1/chat/completions
request that takes longer than ~120s to produce its response. A
vision-language model writing a whole file routinely exceeds that — a
reasoning model can spend the entire window before its first visible
token — so image-text-to-text nodes failed with an opaque gateway error
on any non-trivial image. Measured: the same request 504s at 121s
buffered, and runs 308s to completion when streamed.
Stream the response instead, so the request is bounded by the model
rather than by an idle proxy, and raise the token cap to 16384 (sized to
the canvas executor's 300s per-node timeout at the ~100 tok/s these
models stream).
Truncation is now also reported accurately: reasoning counts against
max_tokens, so a model can hit the cap while still thinking and return
no content at all. Say so, with the reasoning volume, instead of
guessing that the prompt was too long.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* Cover the auto-opened browser tab and chat-completion dispatch with tests
Both paths were added in this PR without tests, as Copilot noted. The browser
tests assert the parts that are easy to regress: an explicit `inbrowser=True`
opens the write-access URL rather than the read-only one, an explicit `False`
opens nothing even with the env var asking for it, and the auto-open stays
suppressed under pytest, CI and SSH.
The chat-completion tests pin the shape of the request the router receives — a
text part plus an image part, streamed — and the two ways it can come back
empty, since a reasoning model that spends the whole budget thinking needs a
different message than one that was cut off for another reason.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* Silence the optional playwright import for the backend type check
The demo's renderer imports playwright lazily and raises a clear install hint
when it is absent, but `ty` resolves imports statically and playwright is not
in the CI environment. Suppressed the same way demo/spectogram handles scipy.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
Abubakar Abid
d3c70fa5cb