---
url: /handbook/advisors/repo-copilot.md
description: >-
  The AI panel beside the Company Repo that answers, reviews, proposes edits and
  drafts asks, and only ever proposes.
---

# DutyGraph Agent

**DutyGraph Agent™** is a panel beside the files of the [Company Repo](../discovery/company-repo.md). It answers questions about the company, finds gaps, writes edit proposals and drafts questions for the right people. It only proposes. The advisor accepts or rejects each proposal. Until November 2026 the app called it the Copilot.

## One message box

| What you write | What DutyGraph Agent does |
| --- | --- |
| A question ("What is missing from this role?") | Answers it. The answer names the files it uses. With Claude, the answer streams as it is written. |
| A check ("Find gaps and conflicts", "Review these files") | Finds gaps, disputes and weak sources in the context files. |
| A change, with a file open ("Rewrite this file", "Fill in the goals") | Writes a proposal for the open file. Read it, then accept or reject it. |
| An ask to a person ("Ask the CFO about the close", "Draft an ask") | Drafts clear questions for the people responsible. Nothing is sent until you send it. |

There are no mode buttons. DutyGraph Agent reads your request and does one of these. When it does more than answer, a line under your question says how it took the request, with **Just answer it** to get a plain answer. A change is saved only when you select **Accept**. An ask goes out only when you send it.

Open **Repo** in the company sidebar. DutyGraph Agent is the right-hand pane. On a phone it opens as a sheet, and the Overview's **Ask DutyGraph Agent** quick action opens it too.

## Which AI it uses

DutyGraph Agent calls an AI model only through the company's own approved connection:

1. the company's approved **Claude** (Anthropic) connection;
2. else the company's approved **OpenAI** connection;
3. else none. DutyGraph Agent says no AI connection is approved and makes no call.

"Approved" means an enabled key for this company, or the firm's key approved for this company in **Settings › Connections**. A company never uses a key from the server's own environment. The sample companies never use AI connections.

## Edit: what Accept writes

The advisor does not pick fields. They write what the file must say, for example "Fill this with your best guess". DutyGraph Agent maps the request onto the file's fields and puts the rest in the file's **Notes**.

The panel shows field changes as a diff and Notes as a preview. **Accept** then writes, depending on the file:

| File | What Accept writes |
| --- | --- |
| A record file: Task Card, duty, workflow, handoff or person | A new version of the record, plus Notes when there are any |
| A research file, for one plain field | An advisor correction to that finding, or Notes |
| Any other file, such as a charter or the operating rhythm | A new version of the file's Notes |
| A generated file, such as an index or the company brief | Nothing. DutyGraph Agent says which file to open instead. |

* Code, not the model, decides how a proposal is applied.
* Accepted text is labelled "Advisor draft — based on common practice, not confirmed by `<Company>`" or "Advisor draft — from your words, not confirmed by `<Company>`".
* The file stays **Inferred** until a person at the company confirms it.
* If the file changed after DutyGraph Agent read it, Accept stops with "This file changed. Ask again."
* Accept needs the edit permission. Without it the button is disabled and says why.

## Notes

Notes are the advisor's own content on a Repo file. Each save is a new version; earlier versions are kept and cannot be changed or deleted by the application. Today Notes are written through DutyGraph Agent's **Accept**.

## Draft an ask

1. Write who to ask and about what ("Ask the controller about the close steps"), or write "Draft an ask". Under an answer, **Draft an ask about this →** does the same with no typing.
2. Read the draft. Each person shows why they are asked.
3. Select **Add to Discovery queue**. The asks wait as unsent drafts in **Discovery › Asks**. Or select **Edit in the composer first**.

DutyGraph Agent asks only about missing, unconfirmed or disputed items, one question per item and at most 10 per person. It addresses each question to a person responsible for that file. **DutyGraph Agent never sends an ask.**

From **Repo › To review**, **Review in file →** runs a focused review of one research finding. It ends with one verdict: looks right, needs a correction, or ask a named person.

## What it reads

* **The context files.** The open file by default. With no file open, the whole company, up to a size limit; files that do not fit are counted and left out. The advisor can narrow it to up to 100 chosen files.
* **The company brief** (`company/brief.md`): status, initiatives, work in motion, recent activity, key people and top gaps.
* **The chain**: the strategy, initiative, department mandate, role and duty above the selected person, file or department, and the work below and across it.
* **The conversation**: the last 6 turns. **Clear** starts again.

## What it does not read or keep

* **Accepted values only.** Proposed readings and files waiting for review do not reach the model. A waiting file sends a placeholder.
* **No quotes and no email addresses.** Email addresses are masked in every file and message.
* **A person's own words** reach the model only in Ask mode, when the advisor asks about one returned answer from **Repo › To review**. They are never written to a Repo file, and they never flow into Edit or Draft an ask.
* **Route files**, which hold captured numbers, are not read.
* **Holdings** never reach DutyGraph Agent. See [Holdings](../governance/holdings.md).
* **No stored text.** Each call writes a usage row with counts and sizes, never the message, the files, the answer or the key. OpenAI calls set `store: false`.

## Limits

* Calls per company are capped per day, with a short pause between one user's calls.
* A replayed request returns an error and never makes a second model call.
* With OpenAI, answers arrive in one piece; only Claude streams.

An advisor's own AI tool can also read the Repo over MCP. See [MCP for advisors](../connect/mcp.md).

::: info
**Status.** **Built**: Ask, Review, Edit and Draft an ask through the company's approved Claude or OpenAI connection; Accept for records, research corrections and Notes; the Discovery queue; focused review; the company brief and chain context; the privacy rules above. **Simulated**: none; the sample companies make no AI calls. **Planned**: a typed Notes editor, a Notes history view in the file, and streaming for OpenAI.
:::
