Trajectory Viewer

Roller includes a read-only Web Viewer that requires no additional dependencies. It groups a run by task and displays model rollouts, reasoning, tool calls, tool results, and final replies. The Viewer reads existing Roller artifacts directly. It does not generate a new trajectory copy, and it does not modify trace_SFT.json or sft.jsonl.

1. Start

Pass the path of a Roller run root that contains info.json or samples/:

Bash

roller view /path/to/runs/my-run

The Viewer listens on 127.0.0.1:17865 by default. The terminal prints the actual URL and the locked run path:

Text

Roller Viewer: http://127.0.0.1:17865
Run directory: /path/to/runs/my-run

Without --port, if 17865 is already in use (for example, by a server that an agent started on the shared host network during collection), the Viewer prints a notice and switches to a free port; use the URL printed in the terminal.

The port can also be set explicitly. If an explicitly set port is in use, the Viewer reports an error and does not switch to another port:

Bash

roller view /path/to/runs/my-run --port 9000

The Viewer process only reads files and serves the page. It does not start Docker, OpenClaw, or a model service. Stop it with Ctrl-C.

2. Access from a remote machine

For remote access, keep the Viewer on the loopback address and open an SSH tunnel from the local machine:

Bash

ssh -L 17865:127.0.0.1:17865 user@remote-host

If the Viewer switched to another port, use the printed port in both places.

If the service must be exposed directly, run:

Bash

roller view /path/to/run --host 0.0.0.0 --port 17865

The Viewer currently has no authentication. Listening on a public interface exposes system prompts, user input, reasoning, tool arguments, tool results, and run metadata. Use it only on a trusted network that already has access control.

3. Page layout

The page has three columns:

1. The left column lists tasks by task_id. Multiple model rollouts of the same task appear in the same group. The list can be searched by task/model/error text and filtered by status or model.
2. The middle column shows the ACTF timeline of the selected rollout. The step title initially shows tokens, elapsed time, stop reason, and reasoning/tool counts. The full step is loaded only after the step is clicked.
3. The Inspector in the right column shows sample information, the first system prompt saved at the ACTF top level, the full tools schema, and the original info.json. A User Simulator task also shows conversation.json.

Drag the dividers between the columns to resize the side columns; the timeline keeps at least 360px. A focused divider also responds to ←/→ (larger steps with Shift), and double-clicking it restores the default proportions. The widths are saved in the current browser and shrink proportionally when the window becomes too narrow for them. At a window width of 1180px or less, the Inspector becomes an overlay and only the left column can be resized.

A divider is also located below the Run summary cards at the top. Dragging it up reduces the height of the summary and increases the height of the three columns; dragging it to the top hides the summary. Drag it back down or double-click it to restore the full height. A focused divider also responds to ↑/↓. The height is also saved in the browser.

Reasoning uses its own region and is collapsed by default. Tool calls are paired with their results by tool_call_id. Results with isError=true use an error style. Everything on the page is rendered as text. HTML or scripts inside the trajectory are not executed.

The page hash stores the rollout ID; the current URL can therefore be copied and used as a deep link to that task in the same Viewer process.

4. Data sources

Content	File	When loaded
Run name, state, processed/total, total usage, collection config	Run info.json	Home page load/refresh
Task/model/status/error/path	Run and sample info.json	Home page load
Step summaries and system/tools	trace_ACTF.json	Rollout selected
Step reasoning/content/tool/result	trace_ACTF.json	Step expanded
Dialogue both sides can see	conversation.json	User Simulator rollout selected

The Viewer does not use trace_SFT.json as its display source, because SFT does not contain step ID, timing, metric, stop reason, and observation metadata. trace.jsonl, cache_trace.jsonl, and wire_trace.jsonl are also not sent to the browser by default.

The Viewer shows the ACTF analysis projection of session events. It does not show the full context of every model request. When OpenClaw applies compaction, pruning, runtime/hook injection, or protocol adaptation between the session and the provider call, the middle timeline cannot prove which history the model saw verbatim. For that kind of audit, enable --wire-trace full at collection time and inspect wire_trace.jsonl and cache_trace.jsonl in the sample offline. See the limits in Trajectory output file guide.

During collection, the Run card on the home page shows state and processed / total from the latest checkpoint. Roller updates on the first result, then after every additional 25 results, and also immediately when completion events are at least 30 seconds apart. It forces an update on interrupt and during finalization. After Resume, the sample's on-disk info.json still contains the original status=ok. The Viewer overlays status=skipped from this run's index so that the Completed card shows the reuse count accurately. The Thinking card shows the retention policy (kept/stripped) and configured native thinking level.

Until the run has finalized info.json, the Viewer also scans samples/<model>/<task>/info.json, so samples finalized after a checkpoint still appear. In-progress temporary directories without a sample info.json are not listed. Click Refresh in the upper-right corner to reload state and samples.

5. Lazy loading and large content

The home page does not read every ACTF. The server parses the matching trace_ACTF.json only when the user selects a rollout, and keeps a limited number of recent trajectories in memory.

When a step is expanded, each user/reasoning/content/arguments/tool result field returns at most 40,000 characters by default. A truncated region shows the original length and a "Load more" button. A further load is capped at 1,000,000 characters. This limit applies only to the browser response and does not modify files on disk.

6. Missing and failed traces

• Samples with status=error/timeout still appear in the list. If ACTF already exists, the partial trajectory up to the failure can be browsed.
• Without ACTF, the page still shows sample status, error, and metadata, and the middle column shows NO ACTF.
• When no cache trace was captured, the system prompt is shown as empty.
• When no wire tools were captured, the tools panel is shown as empty. This does not affect the display of tool call/result in existing steps.
• ACTF reasoning produced with --no-keep-thinking is empty. The page does not fabricate content or reconstruct it from SFT.

7. Read-only and path safety

roller view pins one run directory as the only root it is allowed to read. The API uses an internal rollout ID and does not accept an arbitrary host directory from the browser. Sample/ACTF paths in run info.json that escape the root are rejected. The static server does not provide directory listing, file writes, command execution, or an arbitrary artifact-download API.

Browser responses use security headers that block cross-origin reads, MIME sniffing, and execution of inline/external scripts. These limits are not a substitute for network authentication; therefore a sensitive run should still be accessed only through loopback or an SSH tunnel.

8. Common questions

(1) Newly finished tasks do not appear on the page

Click Refresh in the upper-right corner. The Viewer does not continuously poll the disk, so that a large run does not generate additional I/O.

(2) Port already in use

This error occurs only when the port is set with --port; without this option, the Viewer switches to a free port automatically. Choose another port, or remove --port:

Bash

roller view /path/to/run --port 17866

(3) The Viewer reports that the path is not a Roller run

The input path must be the run root, not samples/, a single sample, or trace_ACTF.json. A valid root contains run info.json or samples/.