Attachment limits and formats
Authoritative count, size, format, extraction, and model-capability limits for composer attachments.
Counts are independent by attachment kind.
| Kind | Count | Per-item raw size | Processing |
|---|---|---|---|
| Extracted file | 5 | 8 MB | Main process extracts text; retained text caps at 120,000 characters |
| Native PDF/file | Included in 5-file cap | 8 MB | Base64 file part when the model advertises native file support |
| Image | 4 | 12 MB | Image data URL; requires compatible model input |
| Audio | 2 | 16 MB | Inline audio part; requires compatible provider/model handling |
File picker formats
The picker accepts images, PDF, plain text, Markdown/MDX, JSON/JSONC, YAML, TOML, INI, CSV/TSV, logs, SQL, HTML/XML, CSS/SCSS, JavaScript/TypeScript, Python, Ruby, Go, Rust, Java, Kotlin, Swift, C/C++, C#, PHP, shell scripts, PowerShell, patch/diff, WAV, MP3, and M4A. Text MIME types are also accepted.
Acceptance by the picker does not guarantee successful extraction. The main process returns an actionable error for unsupported or unreadable content.
Audio formats
Audio detection accepts WAV, MP3/MPEG, M4A/MP4, WebM, and Ogg by supported MIME or extension.
Overflow behavior
When the cap has already been reached, the composer reports the maximum. If a multi-select exceeds remaining slots, only available items are considered and a limit message is shown. Oversized or unreadable items are skipped; accepted items remain attached.
Persistence
Attachments can persist across composer remounts for the selected workspace. They are still part of the next prompt, so remove stale attachments before switching tasks.
Provider boundary
The selected model’s catalog metadata determines whether images, native files, audio, tools, and thinking are supported. Extracted text can work when native-file support is absent, but PDFs still have the same 8 MB raw cap.