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

# Command harness

> Run your own agent as a process inside the sandbox

The `command` harness runs a process you own instead of a coding agent Anpord installs. Anpord starts it in the sandbox, hands it the case through environment variables, reads events from its stdout, and scores the trial with the case's verifier.

```ts theme={null}
tasks: [
  {
    harness: { base: "command", profile: { name: "sample", dir: "./profile" } },
    model: "openai/gpt-5.5",
    provider: "e2b",
  },
]
```

The profile's `run` is the command line; its `install` step runs once while the sandbox is prepared. The profile directory ships everything else the process needs.

## Environment

The process receives these variables:

| Variable                    | Value                                                                    |
| --------------------------- | ------------------------------------------------------------------------ |
| `ANPORD_PROMPT`             | The case prompt                                                          |
| `ANPORD_MODEL`              | The task's `model`, verbatim                                             |
| `ANPORD_HOME`               | The sandbox home directory                                               |
| `ANPORD_WORKSPACE`          | The checked-out workspace; the working directory when the process starts |
| `ANPORD_SYSTEM_PROMPT_FILE` | Path to the profile's system prompt, when it has one                     |
| `ANPORD_TRACE_LOG`          | Where the shell recorder appends what it sees                            |

Anything in the profile's `env` is present too. Stdin is closed.

## Events

Print one JSON object per line on stdout. A line with a known `_tag` is recorded; every other line is ignored, so ordinary logging is harmless. `at` is epoch milliseconds and optional; a line without it takes the moment it was read.

```json theme={null}
{"_tag":"Started","sessionId":"run-42","model":"openai/gpt-5.5"}
{"_tag":"Message","role":"assistant","text":"Reading the failing test first."}
{"_tag":"Command","command":"bun test","exitCode":1,"output":"expect(5).toBe(6)\n"}
{"_tag":"FileChange","paths":["/workspace/src/total.ts"]}
{"_tag":"ToolCall","callId":"call_7","name":"search","input":"{\"query\":\"total\"}","status":"completed"}
{"_tag":"Usage","inputTokens":9120,"outputTokens":312,"cacheReadTokens":8000}
{"_tag":"Finished","reason":"done","at":1788300005000}
```

| Line         | Recorded as                                                                                                                                       |
| ------------ | ------------------------------------------------------------------------------------------------------------------------------------------------- |
| `Started`    | Opens the session; `sessionId` and `model` are taken from it. One is synthesised if you never print it.                                           |
| `Message`    | A turn of the conversation; `role` is `assistant` or `user`                                                                                       |
| `Command`    | A shell command with its exit code and output. `exitCode` may be `null` when you did not capture it.                                              |
| `FileChange` | Paths written, absolute                                                                                                                           |
| `ToolCall`   | A tool invoked by name; `callId` and `status` may be `null`                                                                                       |
| `Usage`      | Tokens spent by one turn, added to the run's total. `cacheReadTokens` and `cacheWriteTokens` default to 0 and `totalTokens` to input plus output. |
| `Finished`   | Ends the journal with a reason                                                                                                                    |

The JSON Schema for these lines is generated from the same definition Anpord decodes with, so the examples above are checked against it.

## Exit

A non-zero exit does not fail the trial and does not discard what was printed before it. When the process exits without printing `Finished`, one is recorded with the reason `exit N`. The verifier scores the trial either way: a process that prints `Finished` and writes nothing fails its verifier exactly as one that crashes does.

The process has fifteen minutes. A run that is still going at the cap is ended and recorded as unavailable.

## The recorder

Anpord sources a small script into every non-interactive shell the process starts, through `BASH_ENV`. Its `DEBUG` trap appends one line per command to `ANPORD_TRACE_LOG`, and each becomes a `Command` in the journal with a `null` exit code, unless you already printed a `Command` whose text matches byte for byte.

What it sees is limited to bash and zsh, the two shells with a `DEBUG` trap:

* `bash -c` and nested bash scripts are traced, line by line.
* `sh -c` on Debian is dash, which has no `DEBUG` trap. Only the `sh -c …` invocation is recorded, not what ran inside it.
* Shells spawned by Node, Python or any other runtime are invisible. A Python agent leaves one line: its own invocation.
* Exit codes are never captured. Bash runs the trap before the command, and zsh reports the command it is about to run, so neither knows how it ended.

If you want inner commands and exit codes in the journal, print `Command` yourself.