Features

Agent loop

Every message runs through an agent loop:

1. Build the request: OpenDocBot assembles the conversation for the model: the system prompt (tool definitions and rules), the prior history, and your message with the document's current state (<doc_state>) prepended as private context.
2. Stream the reply: the model streams back text and/or tool calls.
3. Execute tools: any tool calls are run against the document via the Office.js APIs; each result is appended to the conversation and fed back to the model.
4. Repeat: the loop returns to step 2 with the updated history, until the model sends no more tool calls and gives its final answer.

• The model sees the document outline (<doc_state>) at the start of every turn.
• It calls tools (read, write, format, verify) through the Office.js APIs.
• The loop repeats until the model is done, max 30 iterations per message.
• If the exact same tool call fails 3 times consecutively, the loop aborts with an error instead of burning iterations.

Custom instructions

The Behavior tab has a Custom Instructions box whose contents are injected into the system prompt on every turn. They apply to the whole conversation and take precedence over conflicting built-in rules. Useful for persistent preferences like tone, language, or formatting conventions. See Configuration.

Tool calling per host

The tool registry is filtered per host, so the model only sees relevant tools:

Host	Tools
Word	read/search/edit text, lists, collapse blank paragraphs, verify structure, execute_office_js, ask_user_question
Excel	list sheets, read/write/format ranges, insert/delete rows & columns, merge, clear, sort, set sizes, execute_office_js, ask_user_question
PowerPoint	deck structure, slide read/list, styled text read & edit, shape insert/remove/format, structure ops (create/delete/duplicate/move), masters, verify slides, edit_slide_xml, execute_office_js, ask_user_question

Write tools are tracked separately (see Human in the Loop below).

Human in the Loop

When enabled (Settings → Behavior tab), every document-modifying tool call pauses for your approval:

1. The model proposes an action (e.g. "Change the heading to orange")
2. An ApprovalCard appears with the friendly label and technical details
3. You Approve or Reject

• Approval covers tools that change the document, including: edit_doc_text, edit_doc_list, collapse_blank_paragraphs, write_range, format_range, insert_rows_columns, delete_rows_columns, merge_cells, clear_range, sort_range, set_column_width, set_row_height, modify_presentation_structure, insert_slide_element, remove_slide_element, edit_slide_text, edit_slide_xml, format_shape, and execute_office_js.
• Read-only tools run without prompting.
• Rejecting a tool feeds a "user rejected this tool call" result back to the model so it can adapt rather than repeat the call.

Why

It's your document. The model proposes, you decide; useful for anything you can't easily undo.

Reasoning visibility

While the model thinks, its chain-of-thought streams into a clickable "reasoning" block at the top of the reply:

• ▸ thinking while reasoning is in progress
• ▸ reasoning once content arrives; click to expand the full thinking
• Expanded reasoning shows the model's actual deliberation before it acted

Reasoning is captured per provider: reasoning_content (chat-completions), response.reasoning_text.delta / reasoning_summary_text.delta (Responses), thinking_delta (Anthropic), and thought parts (Gemini).

How much the model thinks is configurable via Settings → Advanced → Reasoning Effort (off by default); see Configuration.

Ask user questions

When the model needs your input before acting, it presents tappable option cards via the ask_user_question tool instead of guessing:

• 1–4 questions per card, each with a header, question, and 2–4 options
• Other option: click it and type your own answer
• Multi-select questions join selections (including an "Other" custom value)
• Answers are sent back as [Header] answer lines so the model knows exactly what you chose

Prompt caching

Long conversations repeat a large system prompt + history prefix. Where the provider supports it, OpenDocBot enables server-side caching automatically to cut cost and latency. Support and settings differ per provider; see Providers for the general explanation and Configuration for the settings.

Debug export

One click copies the whole session for troubleshooting:

• A markdown export (# OpenDocBot Debug Export) with message counts
• The full debug log (timestamps, provider cache info, tool traces)
• Every user/assistant message, with reasoning in a <details> block and tool calls with (truncated) arguments

Paste it into an issue or a chat with the maintainers; it contains everything needed to reproduce a bad turn.

Resilience

• Stream timeouts: the SSE stream aborts after 120 s idle or 10 min total, surfacing a real error instead of hanging.
• Provider errors: response.failed / error events reject the stream with the provider's message.
• Truncation notice: if the model hits the output token limit, the reply is flagged with a note to raise Max Tokens.
• Empty-response guard: a genuinely empty reply (no content, no tool) shows a visible note instead of silent nothing; tool-driven turns (like questions) never trigger it.
• Thrash guard: repeated identical tool failures stop the loop.

Word & Excel specifics

• Selection context: text you highlight before typing is passed as <user_selection> and beats the whole-document state for ambiguous requests ("fix this", "make it bold").
• Paragraph markers: <doc_state> marks headings, lists, page breaks and images so the model understands structure, not just text.

PowerPoint specifics

• No selection API: PowerPoint.js has none, so there's never a <user_selection> block; ambiguous slide references are asked or resolved to the most recently used slide.
• Re-import caveat: OOXML write-backs (edit_slide_text, edit_slide_xml) re-import the slide and reassign slide/shape ids; the tool returns a warning to re-run list_slide_shapes before further edits.
• Verification: verify_slides checks overlaps, out-of-bounds shapes, and WCAG text contrast after styling.