dsh-capability-toggle-plugin

Control agent capabilities from the DSH WebUI — with real runtime enforcement.

English · 简体中文

_{Session · Project · Global — blue check = on, red cross = off, dashed dash = unset.}

What it is

A DeepSeek Harness (DSH) WebUI plugin for controlling skills, MCP servers, tools, prompt injections, approval escalation, and safety guards at session, project, or global scope. Depending on the family, disabling removes, suppresses, rejects, or intercepts the capability on the agent's next step.

Compatible DSH versions

This release targets DSH 0.2.0-rc.1. Every DSH peer dependency is pinned to that release candidate. The Host uses the agent/created lifecycle, persistent configEditor edits, and Loader volatile-config updates; each committed change reconciles every live agent binding.

This release makes no compatibility claim for other DSH versions.

Quick start

Requires Node.js ^22.19.0 or ≥ 24.0.0 (same as DSH 0.2.0-rc.1).

dsh plugin --profile web add github:lifeopsgo/dsh-capability-toggle-plugin#v2.0.0

Restart the existing DSH Web GUI process, then refresh the page. Start it with the command below when it is stopped:

dsh --profile web web

Open the control beside the ➕ button while the agent is idle. Replace web with another profile name when needed.

# Upgrade or downgrade: use any tag listed on the releases page
dsh plugin --profile web add github:lifeopsgo/dsh-capability-toggle-plugin#v2.0.0

# Remove
dsh plugin --profile web remove dsh-capability-toggle-plugin

Features

Three-level resolution

Each capability has three independent levels:

session  ›  project  ›  global  ›  default (enabled)

The nearest explicit value wins. Unset defers to the next level; with every level unset, the capability remains enabled. The row badge always shows the resolved result.

The button displays only its current state: click to toggle on ↔ off, or use its small clear badge to return to unset.

Capability families

Enforcement

Every mechanism is scoped to the current agent; global registrations are not mutated.

Security controls

Turning off Approval escalation rejects every approval request from that agent without changing the system /permission setting.

Safety guards are opt-in:

Usage stats

Skills, MCP servers, and tools carry a small badge showing how many times the model called them this session (called 7). Counts update when a turn ends while the panel is open, live for one agent's lifetime, and are never persisted — the same retention as a safety guard's matched N badge.

The tally counts requests, not successful runs: a call a guard blocked or sent to confirmation still counts, because the model asking for a capability is the signal worth seeing. Guards are matched against rather than called, so they keep their own badge and show no usage count; prompt and approval rows show none either.

Panel preferences

A disclosure arrow beside the panel title opens three display preferences, stored in localStorage so they survive page reloads and restarts:

The fraction counts a guard as enabled only while it is active. A guard row reuses the same disabled field to mean ACTIVE — the inverse of every default-on family — so a Security tab with the approval gate open and all five guards inactive reads 1/6, not 6/6.

Narrowing the level columns is display-only: the three-level resolution keeps running exactly as before, so a hidden project or global override still applies. Each row's badge and level switches always reflect the resolved state — a default-on family reads Active/Disabled, a guard reads Guarding/Inactive — which is why hiding a column cannot hide an effect. The name column absorbs the freed width, and the layout is driven by CSS variables so it stays aligned with the narrow-screen adaptation.

Additional behavior: while the agent runs, the controls that write a stance lock — each row's level switches, their clear badges, and the bulk menus — but browsing stays open, so the search box, the level-column selector, the tabs, and a row's expand still work. State survives popup close and turn boundaries, and the UI follows the WebUI language.

Roadmap

Planned, not yet implemented:

• Cross-project config sync — copy or link project-level settings from another project instead of configuring each project from scratch.
• Capability invocation stats — shipped in v1.2.0: skills, MCP servers, and tools badge how many times the model called them this session.
• Fraction-format tab counts — shipped in v1.3.0: each tab badge renders as enabled / total, and the panel's disclosure arrow holds three display preferences. A guard counts as enabled only while active, so its inverted disabled flag never inflates the Security numerator.
• Settings menu for customizable defaults — expose the plugin's own options in a settings menu, such as the default stance for newly discovered capabilities (today a capability with all three levels unset resolves to enabled, except the opt-in safety guards, which stay inactive).
• Show only enabled / only disabled — add a state filter next to the search box, which today matches names and descriptions only. Guards need the same care the fraction counts need: a guard reuses disabled to mean ACTIVE, so "only disabled" must not list a guard that is actually enforcing. The filter would also narrow what bulk actions apply to, since they act on every currently visible row.
• Filter and select-all — shipped in v1.1.0: the toolbar's search box filters rows, and each level's bulk menu applies enable/disable/clear to every currently visible row.
• Bulk actions on a filtered selection — shipped in v1.1.0 together with the filter (search narrows, bulk acts on what's shown).