Add image, file, and dataset attachments for web and CLI

README.md

@@ -67,6 +67,40 @@ ml-intern --max-iterations 100 "your prompt"
 ml-intern --no-stream "your prompt"
 ```
 
+**Attach local files and datasets:**
+
+Use `--file` or `--image` when a local file should be visible to the next
+agent turn only:
+
+```bash
+ml-intern "summarize this CSV" --file ./data.csv
+ml-intern "what does this screenshot show?" --image ./screenshot.png
+```
+
+In interactive mode, queue local files for the next submitted message:
+
+```text
+/attach ./data.csv ./notes.txt
+/attach ./screenshot.png
+```
+
+Use `--dataset` or `/dataset` when the file should be imported to a private
+Hugging Face dataset repo for training jobs or later reuse:
+
+```bash
+ml-intern "fine-tune on this data" --dataset ./train.jsonl
+```
+
+```text
+/dataset ./train.jsonl
+```
+
+Dataset imports are stored under a private repo named
+`{username}/ml-intern-user-datasets`. Plain `--file`, `--image`, and `/attach`
+do not upload to the Hub; they are per-turn context only. In the web UI, use
+the paperclip button or drag and drop files into the composer, then choose
+between **Attach to turn** and **Import as dataset** before sending.
+
 Run `ml-intern` then `/model` to see the full list of suggested model ids
 (Claude, GPT, HF-router models like MiniMax, Kimi, GLM, DeepSeek, and local
 model prefixes).

agent/core/agent_loop.py

@@ -25,6 +25,7 @@
 from agent.core.cost_estimation import CostEstimate, estimate_tool_cost
 from agent.messaging.gateway import NotificationGateway
 from agent.core import telemetry
+from agent.core.attachments import build_user_content
 from agent.core.doom_loop import check_for_doom_loop
 from agent.core.hub_artifacts import start_session_artifact_collection_task
 from agent.core.llm_params import _resolve_llm_params
@@ -1125,6 +1126,7 @@ async def _abandon_pending_approval(session: Session) -> None:
     async def run_agent(
         session: Session,
         text: str,
+        attachments: list[dict[str, Any]] | None = None,
     ) -> str | None:
         """
         Handle user input (like user_input_or_turn in codex.rs:1291)
@@ -1138,10 +1140,35 @@ async def run_agent(
         if text and session.pending_approval:
             await Handlers._abandon_pending_approval(session)
 
-        # Add user message to history only if there's actual content
-        if text:
-            user_msg = Message(role="user", content=text)
-            session.context_manager.add_message(user_msg)
+        redacted_user_msg: Message | None = None
+        raw_user_msg: Message | None = None
+
+        def _redact_transient_user_content() -> None:
+            if raw_user_msg is not None and redacted_user_msg is not None:
+                raw_user_msg.content = redacted_user_msg.content
+
+        # Add user message to history only if there's actual content. Image
+        # bytes are sent transiently for the current LLM turn, then replaced
+        # with text-only placeholders before the history is persisted/reused.
+        if text or attachments:
+            content = build_user_content(text, attachments or [])
+            redacted_text = build_user_content(text, [
+                m for m in (attachments or []) if m.get("type") == "dataset_import"
+            ])
+            if attachments:
+                # Preserve manifest/preview context in the persisted text while
+                # avoiding raw image data URLs.
+                from agent.core.attachments import attachment_note
+
+                redacted_text = (text or "").rstrip() + attachment_note(attachments or [])
+            raw_user_msg = Message(role="user", content=content)
+            redacted_user_msg = Message(role="user", content=redacted_text)
+            if content != redacted_text:
+                session.context_manager.items.append(raw_user_msg)
+                if session.context_manager.on_message_added:
+                    session.context_manager.on_message_added(redacted_user_msg)
+            else:
+                session.context_manager.add_message(raw_user_msg)
 
         # Send event that we're processing
         await session.send_event(
@@ -1596,6 +1623,7 @@ async def _exec_tool(
                     }
 
                     # Return early - wait for EXEC_APPROVAL operation
+                    _redact_transient_user_content()
                     return None
 
                 iteration += 1
@@ -1637,6 +1665,8 @@ async def _exec_tool(
                 errored = True
                 break
 
+        _redact_transient_user_content()
+
         if session.is_cancelled:
             await _cleanup_on_cancel(session)
             await session.send_event(Event(event_type="interrupted"))
@@ -1942,7 +1972,8 @@ async def process_submission(session: Session, submission) -> bool:
 
     if op.op_type == OpType.USER_INPUT:
         text = op.data.get("text", "") if op.data else ""
-        await Handlers.run_agent(session, text)
+        attachments = op.data.get("attachments", []) if op.data else []
+        await Handlers.run_agent(session, text, attachments)
         return True
 
     if op.op_type == OpType.COMPACT: