Repository navigation
Description
Before submitting
- I searched existing issues and did not find an open issue covering this.
- I included enough detail to reproduce or investigate the problem.
Why this is not a duplicate of #25292 or #38504
Earlier reports describing this exact capability loss (#24969 and #36955) were closed as duplicates of PR #25292. However, PR #25292 fixed a different problem: it only made client_view_file return an error on binary files instead of decoding raw bytes as garbled UTF-8 text. It did not restore or enable native multimodal reading of local audio or video files from disk.
The only related open issue (#38504) also tracks only the remaining view_file binary-text bug on MP4 files, not delivering local audio or video files natively to Gemini models.
Area
apps/server
Problem description
When a user gives an Antigravity thread a file path on disk to a local audio or video file (such as .mp3, .wav, or .mp4), the Gemini model cannot actually listen to the audio or watch the video:
client_view_fileandview_fileonly handle UTF-8 text files (and are also restricted to workspace directories). On binary media files they either fail as non-text (after fix(antigravity): binary file reads return an error, not garbled text #25292) or return garbled bytes ([Bug]: Antigravity view_file still returns binary MP4 as text after the client_view_file fix #38504).- Shell tools (
run_command) only return text stdout/stderr, so the model still cannot perceive the audio or video natively. - As noted in [Bug] Antigravity (ACP) integration: Gemini cannot receive video/image files as media, only raw bytes or a failing attach #36955, Google's standalone
agyCLI can read local media files (frames and audio track) natively from disk, and T3 Code already delivers composer audio attachments (and macOS/Linux MP4 attachments up to 20 MiB) natively to Antigravity. It is specifically the disk-path / tool read path in T3 Code's Antigravity ACP integration that loses native audio and video understanding. - In addition,
docs/user/providers-antigravity.mdnotes that native MP4 video composer attachments are supported only on macOS and Linux, not Windows.
Why this matters more than ever with Gemini 4 Neon
With Gemini 4 Neon now available in T3 Code's Antigravity provider, native audio and video understanding is one of the strongest reasons to pick Gemini models over text/image-only coding models. Real workflows frequently involve audio or video files that already live on disk (for example long lecture recordings, meeting audio, screen recordings, or large media folders that should not be copied into a git repository or manually re-uploaded through the composer). Having to fall back to external Gemini API scripts, Whisper-only text, or FFmpeg frame extraction defeats the main advantage of running Gemini 4 Neon in T3 Code.
Steps to reproduce
- Open T3 Code with the
antigravityprovider andgemini-4-neon-medium(or another multimodal Gemini model). - Point the agent to a local
.mp3,.wav, or.mp4file by file path and ask it to listen to the audio or inspect the video natively (without external ASR or frame-extraction scripts). - Observe the tools available to the agent (
view_file,client_view_file,run_command).
Expected behavior
Antigravity ACP exposes a native media-reading path (or handles audio/video in its file-reading tool) so Gemini models such as Gemini 4 Neon can natively ingest local audio and video files referenced by path, matching the standalone agy CLI capability. Native MP4 handling should also work on Windows, not only macOS and Linux.
Actual behavior
The agent has no tool that passes a local audio or video file on disk into Gemini as native multimodal input, so it cannot listen to audio or watch video referenced by path.
Impact
Major degradation or frequent failure
Version or commit
T3 Code (Beta) 0.1.10
Environment
Windows 11 (10.0.26200.0), T3 Code Desktop 0.1.10, provider antigravity, model gemini-4-neon-medium.
Workaround
Manually splitting/uploading audio files as composer attachments (up to 50 MiB per file), or calling the Gemini API directly via an external Python script with a separate API key. Neither works seamlessly for local media workflows or Windows native MP4 video.
Activity
kuelshammer commentedon Nov 21, 2026
Adding a data point for images — the title says audio/video, but the same gap covers reading image files from the workspace.
Setup: T3 Code server from 2026-10-16 (pre-#25292), Antigravity managed runtime 1.3.0,
oauth-personal. Workflow: ask the agent to look at a PNG/JPG inside the project ("open diagrams/foo.png and tell me what the diagram shows"). The agent calls its file-read tool, T3 answersfs/read_text_filewith the raw bytes decoded as text, and Gemini reports it perceives nothing. On a post-#25292 build the same read should at least fail with "File 'x' is not a UTF-8 text file." instead of returning garbled bytes — better, but the workflow is still impossible; the image never reaches the model either way.Two observations that might help scoping:
fs/read_text_fileandfs/write_text_file(packages/effect-acp/src/_generated/schema-v1.gen.ts), and the Antigravity CLI mapsview_file/client_view_fileonto those. As long as that holds, no client-side change in T3 can deliver the bytes as media — this needs a runtime-side (Google) media-read capability, plus T3-side plumbing to advertise and handle it.Until then the workarounds are composer attachments or falling back to the Antigravity CLI directly, which handles image paths natively.
Comment written with AI assistance, posted by the account owner.