Summary

Based on #157 and #158.

This PR adds user-selected image, file, and dataset attachments across the web UI and CLI.

The main design choice is to split attachments into two explicit modes:

• Attach to turn: files/images are used as context for the next agent turn only.
• Import as dataset: files are uploaded to a private Hugging Face dataset repo and injected into the agent turn as durable dataset manifest context.

What changed

Shared attachment layer

Adds agent/core/attachments.py, which centralizes:

• filename sanitization
• file validation
• MIME/type detection
• size metadata
• text previews for readable files
• local per-turn manifests
• private HF dataset imports
• dataset manifest generation
• model-context notes for attached files

Web UI

The chat composer now supports:

• paperclip file picker
• drag/drop files and images
• removable attachment chips
• upload error display
• Attach to turn vs Import as dataset toggle

When importing as a dataset, files are uploaded to a private dataset repo before the chat turn is submitted.

CLI

Headless mode now supports:

ml-intern "summarize this" --file ./data.csv
ml-intern "inspect this screenshot" --image ./screenshot.png
ml-intern "train on this" --dataset ./train.jsonl

Interactive mode now supports:

/attach ./data.csv ./notes.txt
/dataset ./train.jsonl

--file, --image, and /attach are per-turn context only.
--dataset and /dataset upload to the private HF dataset repo.

Agent context and privacy

Attached text-like files include metadata and bounded previews in the agent turn.

Images are sent transiently as multimodal image parts so the LLM can inspect them, then redacted back to placeholders/manifest notes before persisted history or approval continuations can reuse the turn.

Dataset imports are stored under:

{username}/ml-intern-user-datasets
sessions/{session_id_or_cli_run_id}/{upload_id}/...

The agent receives the repo id, path prefix, manifest path, file list, sizes, MIME types, and guidance to use the HF dataset path for training/jobs.

Docs

Updated README with CLI and web usage examples for:

• --file
• --image
• --dataset
• /attach
• /dataset

CLI help now also lists /attach and /dataset.

Safety/correctness details

• Plain attachments do not upload to the Hub.
• Dataset uploads only happen through explicit dataset commands or the web Import as dataset toggle.
• Raw image bytes are not persisted in session history.
• Upload refs are preserved for Claude quota retries.
• Duplicate multipart filenames are written to unique temporary paths to avoid accidental overwrite.

Test plan

• python3 -m py_compile agent/core/attachments.py agent/core/agent_loop.py agent/main.py backend/routes/agent.py backend/session_manager.py
• uv run pytest tests/unit/test_cli_rendering.py tests/unit/test_session_manager_persistence.py tests/unit/test_attachments.py
• npm run build

All targeted tests pass. npm run build passes with the existing large chunk warning.

Future improvements

• Add a richer Codex-style CLI/TUI composer with visible [Image #1] / [File #1] draft chips before submit.
• Support easier terminal drag/drop flows where available, instead of requiring explicit /attach PATH or --image PATH.
• Add recursive folder import and larger/resumable uploads if users need bigger dataset ingestion flows.
• Add direct image-specific UX polish in web, such as thumbnails or preview-on-hover.