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

# Traces

> Read a trace: the conversation, the grading, what the environment showed, the files the agent changed, QA results, logs, and cost.

A **trace** is one attempt by one model at one task. It records the whole conversation, every tool
call and its result, what the environment showed, the grade, and the cost. Open a trace from a
job's **Traces** tab, a task's **Traces** tab, or a result in the taskset grid.

## Read the conversation

The conversation shows the prompt, the model's messages, and each tool call with its result. It ends
with the **Evaluation result**: the reward and, for rubric grading, the judge model, a
**Judge summary**, and every criterion with its weight, whether it was met, and the judge's reason.
**Copy JSON** copies the raw result. The score at the top of the conversation jumps to it.

The reward is the verdict, not the explanation. Read the steps and the grading together to tell
whether the agent solved the task or found a shortcut; [QA agents](/platform/qa-agents) do this
reading for you at scale.

The header shows who ran the trace, its job, its reward, whether it is public, and **LIVE** while
it runs. When a trace stopped early, a badge says why:

| Badge | Meaning |
| - | - |
| **Step limit** | The agent reached the job's maximum steps. |
| **Output limit** | The model reached its output token limit. |
| **Timed out** | The trace ran out of time. |
| **Bad tool call** | The model produced a tool call that could not be parsed. |

An execution badge says where the trace ran, which decides what HUD bills:

| Badge | Where it ran | Billed by HUD |
| - | - | - |
| **Hosted** | HUD ran the agent and the environment. | Model usage and environment runtime |
| **HUD Sandbox** | Your agent ran outside HUD; the environment ran on HUD. | Model usage and environment runtime |
| **External** | The agent and the environment both ran outside HUD. | Model usage only |

## Inspect a trace

**Open inspector** shows the trace's views beside the conversation. Which views appear depends on
what the environment recorded:

| View | Shows |
| - | - |
| **Screen** | Screenshots from computer-use and browser steps. |
| **Robotics** | Robot observations, policy inference, and video. |
| **Workspace** | The workspace files at the step you select, and what changed. See [file tracking](/platform/file-tracking). |
| **Source** | The environment source the trace ran on. |
| **QA** | QA results for the trace, and **Run QA**. See [read QA results](/platform/qa-results). |
| **Logs** | Environment and worker logs. |
| **Term** | The shell commands the agent ran and their output. |
| **Debug** | Runtime diagnostics for failed or stuck traces. |
| **Info** | **Manage trace**: title, description, and **Invalidate** or **Revalidate**. **Execution**: mode, agent, environment, and stop reason. **Cost**: total, model usage, and environment runtime. |

Invalidating a trace keeps it out of leaderboards and reward summaries. See
[keep bad traces out of the score](/platform/evaluations#keep-bad-traces-out-of-the-score).

## Share a trace

The **More trace actions** menu has **Compact view**, which collapses sections of the conversation,
**Copy conversation**, and, for editors, **Share trace**. Share invites people or teams as
**Viewer** or **Editor**, or makes the trace public so anyone with the link can view it.

Public viewers see the conversation and the **Screen**, **Robotics**, and **Workspace** views. They
also see your team's QA verdicts, summaries, and findings, without the evidence behind them.
**Source**, **Logs**, **Term**, **Debug**, and **Info** stay private.


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