Introduction¶
The “Trajectory” feature in DeepSeek Harness (DSH) describes what happened in the past, while runtime state changes must be inspected through “X-Ray.” dsh-runtime-xray is a read-only runtime inspection plugin for DeepSeek Harness. It does not modify configuration, reload plugins, execute commands, or add tools to the model. Instead, it inspects the currently effective runtime base, session capabilities, model inputs, and lifecycle resources, addressing developers’ need to track agent state in real time.
Plugin Overview¶
dsh-runtime-xray is DeepSeek Harness’s read-only runtime X-Ray, maintained by hollis-openlab. It follows the MIT license.
The X-Ray page organizes information along a clear runtime chain: runtime base → Skill / tools → model context. Services belong to the internal runtime base of the Host and are not mistaken for model inputs.
Core Features¶
The plugin provides the following capabilities:
- Layered navigation: Browse Overview, runtime base, session capabilities, and model inputs by runtime layer.
- Current session: View the services, Skills, tools, model context, and model routing actually available in the current session.
- Entire application: Inspect the plugins, services, and lifecycle resources shared by all sessions.
- Scope inheritance: Show the override order “entire application → preset → current session.”
- Model input: Count system instruction segments, runtime context, and variables, grouped by Harness, tools, application, deployment, Plan, and UI.
- Plugin—service network: Show ownership relationships (solid lines) and inferred relationships from explicit
ctx.providetags (dashed lines). - Lifecycle resources: Aggregate service registrations, child plugin mounts, timers, listeners, and other resources; expand to view source and depth.
- Service dictionary: Show user-friendly names and Chinese/English purpose bubbles for native DSH services; unknown services retain the original key and use a safe fallback.
- Evidence levels: Distinguish exact, inferred, and unavailable; does not disguise guesses as facts.
- Fault isolation: If one inspection domain fails, other domains remain available.
- Reread: Manually capture the current state once; no background polling occurs when the page is closed.
- Download diagnostic file: Download a redacted JSON file for the current view; file size limit is 1 MiB.
- Bilingual Chinese/English: Follow the language setting in DSH Web and switch in real time without reloading the plugin.
Installation and Activation¶
Use the following commands to install the plugin and start the Web profile:
dsh plugin --profile web add "github:hollis-openlab/dsh-runtime-xray#main"
dsh --profile web
After restarting the Web profile, “X-Ray” appears next to “Chat” and “Trajectory.”
Typical Usage¶
- Open a session and click “X-Ray.”
- Select “Current session,” then inspect the current session along the runtime base, session capabilities, and model inputs.
- Open “Skill” to view the effective Skills directory; then open “Tools” to view model-invokable actions.
- Open “Model context” to inspect model inputs grouped by data type and source.
- Select “Entire application” to inspect the plugin—service network and aggregated lifecycle resources.
- To confirm the state after a hot update, click “Reread.”
- To submit issue evidence, click “Download diagnostic file” to download redacted JSON.
Use Cases and Notes¶
The plugin only reads runtime information publicly provided by the current DSH process. It does not connect to any external services, nor does it require its own API key, environment variables, or writable data directory.
The default snapshot and diagnostic files do not include:
* credentials and environment variable values;
* prompt text and runtime context text;
* variable values, approval payloads, and tool results;
* callable objects or arbitrary plugin configuration;
* raw session and runtime entity identifiers.
Exported identifiers are consistently anonymized within a single file. The diagnostic file size limit is 1 MiB.
Conclusion¶
dsh-runtime-xray provides a complete view from the runtime base to model input, helping developers quickly pinpoint state during debugging and troubleshooting. For more information or access to the source code, visit the DSH ecosystem directory or the GitHub repository.