File Reader

File Reader Skill

Use this skill to list visible file context or read content from a file that has already been resolved for the current assistant turn.

Workflow

1. Use this skill only when the user is asking about a specific file, uploaded attachment, historical file reference, or current console file context.
2. Do not invent file IDs. Use a file ID supplied by resolved page context, attachment context, or governed asset resolution.
3. For listing requests such as "what files do I have", "which files are visible", "current files", or "selected files", use the current backend-backed page context when it is present and fresh. Call list_visible_files when that context is missing, ambiguous, stale, or needs an authoritative refresh. Do not use read_file for a list-only request unless the user asks for file contents.
4. For read, summary, translation, extraction, comparison, or question answering, call read_file with file_id. Set max_chars only when you need more or less returned content.
5. Inspect content_status before answering:
   • If content_status is extracted, answer from content.
   • If content_status is empty, say the file has no extractable text content.
   • If content_status is error, explain that the file could not be read and include the short error reason when useful.
6. The raw content returned by read_file is not durable across history reloads, approvals, navigation, refresh, or later tool phases. If later steps will use the file body, a summary, theme, topic, quote, title, prompt, config value, generated asset, or final answer derived from this file, call submit_turn_state before leaving the file-reading phase. Record a concise reusable fact with kind=working_fact, visibility=model_only, source=file-reader/read_file, and a meaningful key such as source_file_summary, source_file_theme, or exact_short_text. When the summary is useful for the user to verify, use kind=user_deliverable, visibility=user_visible, and a short title/content instead.
7. If content_truncated is true and the missing tail matters, ask for a narrower question or retry with a higher max_chars up to the tool limit.
8. If the user asks to delete, rename, move, or otherwise manage a File Management asset, use file-manager when it is available in the current files-page context. If file-manager is not available, explain that the current chat can read files but cannot perform that file-management operation from this surface.
9. Mention the file name when listing, summarizing, or quoting. Do not expose internal file IDs unless the user explicitly asks for them.

Tool Usage

read_file accepts:

• file_id: required file ID from resolved context.
• max_chars: optional maximum returned content characters. Defaults to 4000 and is capped at 12000.
• Success evidence: the tool result must identify the resolved file and include content_status. Use content only when content_status=extracted. If content_status=empty, report that no extractable text was found. If content_status=error, report the actual read failure instead of summarizing from filename or page metadata.
• Handoff evidence: when content_lifetime=current_tool_result_only and content_redacted_in_history=true, do not assume the raw content will be available after continuation boundaries. If handoff_recommended=true and later steps depend on this content or a derived summary, call submit_turn_state before continuing to those steps.

list_visible_files accepts the current page query when available: page, page_size, keyword, sort, extension, processing_status, category, folder_id, workspace_id, and optional selected_ids. It returns:

• source=backend_api: confirms the list came from the authoritative backend query.
• count: number of files in the returned page.
• selected_count: number of selected files.
• files: ordered files with ID, name, extension, MIME type, optional workspace ID, and verified selection state.
• Success evidence: use the returned files list and count as the current visible-file state. Do not infer file contents from this list-only result.

Truthfulness Contract

• Treat file-reader results as authoritative. Do not answer from stale visible page metadata when a read_file result is required.
• If the target cannot be resolved or read, say so directly and ask for a narrower target only when the available context cannot identify the file.
• Retry at most once with corrected file_id or max_chars when the tool result indicates a recoverable issue.