Introduction¶

When running long tasks with DeepSeek Harness (DSH), the native trace records messages, tool calls, and events, but a single round of tasks often contains hundreds or thousands of steps. Answering a few basic questions isn’t easy: What strategy did the Agent actually adopt? Which steps are advancing the task, and which are repetitive trials? Did failures come from model judgment, tool usage, or environmental conditions? Is there enough evidence in an “already completed” response?

Going through the trace step-by-step is inefficient. What dsh-trace-insight introduced below does exactly that: organize dense trace events into a continuously updated analysis timeline, allowing every conclusion to be traced back to the corresponding original event.

What is it¶

dsh-trace-insight (DSH Trace Insight) is a read-only execution review plugin for DSH maintained by Liu-Bot24, licensed under MIT. DSH’s philosophy is “everything is a plugin.” Trace Insight integrates into the right sidebar of the DSH web interface as a plugin; the left side preserves the conversation or trace, and the right side displays the interpretation results, with adjustable width.

It directly reads DSH’s structured Session Event Log, first using local rule-based analysis to establish a factual foundation, and then uses independently configured models on demand to explain decisions, risks, and improvement directions.

Boundaries need to be clarified first: the plugin only reads data and does not modify the checked Session, Workspace, Skill, or Global Memory, nor does it intervene, pause, or block the execution of the development Agent.

Core Features¶

Capabilities on the analysis side:

• Continuous Review: Organizes rule-based and model-based analyses by Turn and Seq, allowing you to see progress at different stages of a long task without waiting for the entire round to end.
• Rule-based Analysis: Identifies tool failures, repeated failures, no-progress loops, path guessing, tool misuse, completion signals, and evidence gaps locally, without calling the model.
• Model Interpretation: Uses independently configured DSH models to analyze strategy, root cause, risks, next steps, and reusable experiences, without affecting the main model of the development Agent.
• Controlled Re-analysis: Temporarily switches models to re-analyze a specified interval without changing the default model used for future automatic analyses.

Capabilities for consuming analysis results:

• Evidence Location: Traces conclusions back to the corresponding Seq, Turn, Step, Tool, excerpt, and original context.
• Task Overview: Summarizes the entire task based on development stage, tool usage, and problem clues, organizing only existing analyses without generating new model calls.
• Result Comparison: Compares two successful model analyses of the same trace, displaying conclusions, configurations, resource usage, and original evidence separately, allowing observation of differences from different models or configurations.
• History & Export: Analysis records are saved locally and can be exported separately as analysis history, original Session history, or complete analysis packages.

The right sidebar contains four pages: Review, Overview, Comparison, and Settings.

Installation and Enablement¶

First, confirm the environment: Node.js 22.19.0 or higher is required (the engines requirement in package.json is node >=22.19.0). The plugin declares compatibility with DSH versions 0.1.0-rc.7, 0.1.0-rc.8, 0.1.1-rc.1, and 0.1.1-rc.2 via dshCompatibility, with surface set to web.

You must close the running DSH before installation, then download or clone the repository and run the installer in the repository root directory.

macOS or Linux:

bash ./install.sh

On Windows, double-click 安装到DSH.cmd, or run in PowerShell:

powershell -ExecutionPolicy Bypass -File .\install.ps1

The installer will connect the right sidebar for each supported DSH found; the plugin package is saved in <DSH_HOME>/trace-insight/packages. After installation, DSH continues to function without relying on the source code directory or download directory.

Start DSH:

dsh web

If there is no global dsh command, use:

npx --yes --package=@deepseek-ai/dsh dsh web

After opening any session, click “Interpret” to open the right sidebar.

A note on versions: The README suggests that for the current version, use “Standard Unpatched Sidebar 1.5.0” under packages/standard; the install/uninstall scripts in the repository root belong to the old patch version. See packages/standard/README.md in the repository for the specific installation method for the standard version. The specific relationship between the two versions is not detailed in the documentation; it is recommended to confirm which version to use before installing.

Uninstallation also requires closing DSH first, then running the corresponding uninstall script in the repository root (uninstall.sh / uninstall.ps1 / double-click 卸载插件.cmd). Uninstallation will remove Trace Insight, the right sidebar, and the plugin package managed by the installer; saved analysis history remains in the data directory.

Typical Usage¶

Complete workflow for first-time use:

1. Open any DSH session, click “Interpret”, and enter “Review”. Rule-based analysis reads the existing trace directly and does not require model configuration.
2. If model interpretation is needed, go to “Settings → Models & Auto Strategy”, select a registered DSH provider and model, and save.
3. After saving, you can wait for automatic analysis to trigger, or manually analyze a segment of the trace in the Review.
4. Click on the evidence entry in the analysis conclusion to view the corresponding original event and context.

After the steps above, every model conclusion and rule finding can open an independent evidence drawer; each citation marks the Seq, Turn, and summary, and the original context can be read when needed.

Several things will not trigger model calls: opening or refreshing Trace Insight, filtering the timeline, viewing evidence, entering Overview, or comparing existing analyses. Model requests are only generated during automatic analysis or when the user initiates a model analysis.

Model configuration is divided into three layers: global default model, session-specific model, and a model used temporarily only for this analysis. Only the first two are saved; switching the model temporarily does not modify the default configuration used for future automatic analyses. Automatic analysis only runs when the Session is being monitored in real-time execution, a default model is configured, and trigger conditions are met; if the model fails, is cancelled, or returns invalid results, the analysis progress does not skip that interval.

If the “Interpret” entry is not visible, close DSH completely and run the installer again. If it still does not appear, check if the plugin has entered the web profile:

dsh --profile web --dump-config | grep "trace-insight"

For Windows PowerShell, use Select-String:

dsh --profile web --dump-config | Select-String "trace-insight"

When using DSH only via npx, replace dsh in the command with npx --yes --package=@deepseek-ai/dsh dsh.

Data Directory and Export¶

The default data directory is as follows; when DSH_HOME is set, it is <DSH_HOME>/trace-insight:

The directory mainly contains settings.json and sessions/<session-id-hash>.json.

Exports are divided into three categories: Analysis History (rule analysis, model analysis, runtime status, and analysis progress), Original Session History (DSH raw events, surface, and session lineage), and Complete Analysis Package (both of the above). Confirmation is required when exporting original history or a complete analysis package; analysis history may contain evidence excerpts and model raw text, so check the content before sharing.

Privacy and Cost Boundaries¶

• Rule-based analysis runs completely locally and does not call the model.
• Model inputs use trace evidence that has been trimmed and desensitized for common credentials, and complete raw logs are not sent; when using external models, the corresponding evidence will be sent to the model provider you choose.
• Original Session data is not included in standard analysis exports by default.
• The Trace Insight Host RPC only allows loopback page access and must be used via 127.0.0.1 or localhost on the same machine running DSH; data cannot be read when opening DSH via a LAN address.

Applicable Scenarios and Notes¶

Suitable target audience and scenarios: Developers who frequently run long tasks with DSH and need to review Agent behavior; those who want structured analysis while preserving the original trace; scenarios requiring comparison of interpretation results from different models or configurations on the same trace.

Note a few things before use:

1. The plugin runs with the permissions of the current DSH process; before installation, you should check the repository source code and license yourself (currently MIT).
2. You must close the running DSH before installing and uninstalling.
3. When “Waiting for default model” is displayed, rule-based analysis will continue to run; model analysis will only start after configuring the default model. If model analysis fails, the failure record is retained, progress does not skip the failure interval, and you can retry from the failure interval after checking provider credentials, model routing, and rate limit status.

Conclusion¶

dsh-trace-insight decomposes trace review into two layers: local rule-based analysis for the factual foundation, and independently configured models for the interpretation layer; all conclusions can be traced back to original events, and the process is read-only without interfering with the development Agent. If you are using DSH to run long tasks and struggle with going through the trace step-by-step, give it a try.

• Community Plugin Directory: https://www.skillhub.cn/plugins/Liu-Bot24/dsh-trace-insight
• GitHub Repository: https://github.com/Liu-Bot24/dsh-trace-insight