Introduction¶
When processing agent tasks, DeepSeek Harness (DSH) requires users to approve specific operations. The current approval card details area only supports displaying command-type tool parameters (such as bash and pwsh). For tools such as write and edit that use file_path or path parameters, this area remains empty. As a result, users cannot intuitively understand the operation details and potential risks before clicking “Allow” or “Deny.”
The dsh-approval-explain plugin adds an “Explain” button to the approval card and generates an operation summary using a single LLM call. The summary includes the operation details, affected files, and a safety assessment, resolving the issue of the empty details area.
Core Features¶
- Explain button: Adds an explain button to the approval card details area. Clicking it generates an operation summary.
- Full parameter support: Supports the
command,file_path, andpathparameter shapes, without being limited to specific tools. - Visual reuse: Reuses native button styles (outline, 36px capsule shape) and uses the DeepSeek family color scheme.
- Command support: In addition to the button, supports using the
/explain <content>command directly in the editor. - No build authorization required: The plugin ships prebuilt artifacts directly without triggering pnpm build prompts.
- Non-invasive: As an incremental plugin, it does not disable other DSH features.
Installation and Enablement¶
This plugin requires DSH version 0.1.5-rc.1 or higher and only supports the Web GUI.
- Ensure pnpm is installed:
corepack enable pnpm
- Run the installation command:
dsh plugin --profile web add github:Martlet-Tech/dsh-approval-explain
- After installation, restart the DSH process so the configuration file is reloaded.
If the plugin does not take effect after installation, check $DSH_HOME/profiles/web/package.json and ensure that dsh.profile.bundles includes dsh-approval-explain.
Usage¶
After installing and restarting DSH, an “Explain” button appears in the approval card. Clicking the button calls the LLM to analyze the pending operation and returns a three-line summary:
* What: The specific content of the operation
* Touches: The file paths involved
* Safety: Safety assessment
You can also use the following command directly in the message editor:
/explain <content>
Notes¶
- Web GUI limitation: This plugin currently supports only the Web interface.
- Audit trail: Each explanation click leaves
command/runandcommand/donerecords in the session log, which helps with auditing but also means every operation leaves a trace. - Token consumption: The explanation feature generates an LLM call, with output length limited to about 700 tokens.
- Card takeover: This plugin takes over the display of the entire approval card, not just adds a button.
- Internal coupling: The plugin internals depend on DSH’s
StreamChunkfield names (text-delta,block-end, andfinish). If DSH updates these fields, the plugin may need adaptation. - No effect on results: The explanation is for reference only. It does not change the actual approval outcome or cache judgments.
