> ## Documentation Index
> Fetch the complete documentation index at: https://docs.overmindlab.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Agent feedback

> What a coding agent sends when it reports friction with Overmind, and how to turn it on or off.

A coding agent connected over MCP can send feedback about Overmind itself: an operation that fails or contradicts its description, unclear guidance, a missing capability, stuck work, or repeated attempts without progress. Reports go to the Overmind team. They describe how the platform behaved, not what your data holds.

If an Overmind operation fails, contradicts its description or needs repeated attempts, report it with submit\_feedback and tell me the line it returns.

## Turn it on or off

Agent feedback is one choice for your account. It covers every coding agent you connect.

| Where | When you see it |
| - | - |
| Sign-up | **Let my coding agents send feedback to Overmind** is ticked by default under the sign-up form. |
| **Connect to Overmind** | The same checkbox appears when a coding agent asks for access through OAuth. **Connect** saves it. |
| Console notice | An account with no recorded choice sees **Allow** and **Don't allow** under the page header. |
| **Settings** → **Privacy** | The **Agent feedback** switch. |

Until you choose, and after you choose **Don't allow**, `submit_feedback` is not in the agent's tool list and nothing is sent. Guest workspaces never send reports.

## What a report contains

The agent fills these fields. Text fields have length limits, and the schema accepts no other fields.

| Field | Content |
| - | - |
| `kind` | `failure`, `mismatch`, `confusing`, `missing`, `stuck`, `dead_end`, `idea`, `praise` or `question` |
| `area` | `setup`, `observability`, `datasets`, `evaluation`, `optimizer`, `training`, `serving`, `billing` or `other` |
| `severity`, `outcome` | How much it blocked the agent, and how the attempt ended |
| `headline`, `summary`, `intent`, `expected`, `observed`, `suggestion` | Short descriptions of the platform's behaviour |
| `trail` | The tools, `overmind://` resources and `overmind` CLI commands used, in order, with each result and error code. CLI arguments are not accepted. |
| `resources` | `overmind://` links to the records involved |
| `tool_calls` | The number of tool calls the attempt took, when the agent counted them |
| `environment` | What the agent knows about itself: harness and version, model, operating system, shell, Python, CLI and skills versions |

Overmind adds what it observes on the connection: the client name, the `User-Agent` and `MCP-Protocol-Version` headers, the credential type and the MCP contract version. Each report is linked to your account and, when given, to the project.

## What is kept out

Agents are told to leave out data rows, prompts, code, file contents, paths, customer names and credentials. Before a report is saved, Overmind replaces:

* known credential and key formats with `[credential removed]`
* email addresses with `[email]`
* web links with `[link]`
* home-directory paths with `[path]`
* IPv4 addresses with `[ip]`

Other text is saved as the agent wrote it. Overmind cannot detect every customer detail in free text, such as a name or an amount.

The agent shows you a line for each report it sends, with the link to **Settings**.

## Limits

An account sends at most 20 reports an hour. A report sent again with the same `report_id` updates the earlier report. On a self-hosted deployment, reports stay in that deployment's own database.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.