Preface¶

In the DSH Web UI, when a session reaches an operation requiring approval, users typically need to confirm each item manually. For repetitive or low-risk operations, manual confirmation can become an additional step in the workflow. simon300000/dsh-auto provides an Auto Approve permission preset for this purpose: before each operation requiring approval is allowed or denied by the plugin, a new restricted DSH sub-agent reviews it first.

The current release only supports the Web UI.

What This Is¶

dsh-auto is a DSH plugin whose repository owner is simon300000, licensed under MIT. It adds a permission preset to the DeepSeek Harness Web UI, so that approvals no longer rely solely on manual clicks but instead go through a restricted Reviewer sub-session review, after which the host makes the final allow or deny decision.

Core Features¶

• Handles approval/request only when the session selects Auto Approve.
• Each approval spawns a restricted DSH sub-Reviewer session.
• The Reviewer runs in a read-only sandbox with approval/policy = never.
• An execution guard allows only read, glob, grep, and scoped structured-output tools, rejects other tools, disallows further subagents, and permits at most four investigation steps plus a final response step.
• The Reviewer performs only bounded read-only investigation; sensitive files may be checked only when a minimal read-only check could change the decision.
• The structured result requires only outcome and may include risk_level, user_authorization, and rationale.
• The host always rejects critical risk; it also rejects high risk without at least medium user authorization.
• Invalid output, missing action data, timeouts, infrastructure failures unrelated to cancellation, and tool failures are all treated as fail-closed.
• Model refusals are not retried and do not fall back to user prompts.
• Three consecutive refusals in the same parent turn interrupt that turn; any allow action resets the counter.
• A default 90-second deadline covers sub-session creation, all model steps, local read-only investigation, and the final structured output.

Installation and Enablement¶

First, install the plugin:

dsh plugin --profile web add github:simon300000/dsh-auto

After restarting the Web UI, select Auto Approve in the session Permissions selector or as the default permission preset in General Settings.

Typical Usage¶

• A compact result {"outcome":"allow"} defaults to low risk and unknown authorization.
• reviewerProvider and reviewerModel must be set together; if both are omitted, the Reviewer uses the parent session’s current provider and model.
• The bundle defaults to deepseek-official/deepseek-v4-flash with high reasoning.
• Restart DSH after modifying configuration, policy, or plugin code.

Applicable Scenarios and Notes¶

Suitable for scenarios where you use plugin approval in the DSH Web UI and want to reduce manual confirmation while retaining restricted review. Notes:

• The current release only supports the Web UI.
• The Reviewer’s permissions are restricted: read-only sandbox, approval/policy = never, tool whitelist, no further subagents, and step limits.
• High/critical risk operations are not passed through just because the Reviewer is lenient; the host rejects critical risk and high risk that does not meet authorization conditions.
• The plugin changes the actual behavior of the DSH approval chain and runs with the current dsh process permissions. Inspect the source code, configuration, and MIT license before installation.

Conclusion¶

The value of dsh-auto lies in transforming Auto Approve from “direct approval” into “restricted Reviewer reviews first, then the host rules according to policy.” The current documentation does not provide a table-of-contents link; the repository address is:

https://github.com/simon300000/dsh-auto