dsh-cabinet-mode

Cabinet Mode (内阁模式) for DeepSeek Harness

Version compatibility (0.7.0): Targets DSH 0.2.2-rc.1. This release adds keyboard-accessible session navigation and member usage snapshots; it preserves the shorter default prompts while retaining budget, authorization and final-state verification requirements, and retains the per-office model selection fix before the host freezes each step; 0.5.1/0.5.2 only repaired preset loading and display metadata. Verified with official host components and isolated profiles, without real model calls; see the model regression scope below. Earlier rc.3/rc.5 UI and installation results are historical evidence for earlier packages. On DSH 0.2.1-rc.3/rc.5 retain the previously verified package, including 0.5.0: the current preset omits the separate list_agents entry required on rc.5. Select the compatible version in the market.

[中文](README.zh.md) | English

Install: [Use the plugin market (recommended)](#install-from-the-plugin-market-recommended).

Website: [luodaoyi.github.io/dsh-cabinet-mode](https://luodaoyi.github.io/dsh-cabinet-mode/) (bilingual intro page, served from site/ in this repository).

Icon: Ming-dynasty grand-secretary portrait provided by the project owner (shipped as icon.svg).

A small DSH bundle that adds Cabinet Mode (内阁模式) to

[DeepSeek Harness](https://github.com/deepseek-ai/deepseek-harness). Install the

bundle, restart DSH, and pick 内阁模式 when starting a new session. The mode is

a research-team preset modeled on the Great Ming Research Bureau cabinet

(大明科研司·内阁): a Lead (内阁首辅) directs members (阁臣) of four bureaus, defaults

to research before touching code, and must never fabricate sources.

The package ships a host half (index.js, the member schema and Cabinet-only prompt provider) and a browser half

(client.js, the right-Sidebar Cabinet panel) with five role portraits and an unranked fallback portrait

embedded as 128px data URLs (square sources in docs/images/avatars/,

downscaled by src/build_avatars.py); it also keeps the optional Node.js

fallback scripts (scripts/*.mjs). It uses @deepseek-ai/schemastery for the host settings schema. React and friends come from the DSH client module table, with no

bundled duplicate and no CDN.

It declares two rows — preset-cabinet (one @deepseek-ai/dsh-agent-preset

declaration) and cabinet-ui (the Loader row that loads this package's

JavaScript) — and reuses the plugin list shipped with DSH Standard Mode, so

the mode has the same tools as Standard Mode; what differs is the persona, plus

the right-Sidebar Cabinet panel. Team tools (spawn_teammate, team_task_*,

and the rest) come from the profile-level

@deepseek-ai/dsh-experimental-agent-team-profile bundle; **this preset does not

ship them**.

This bundle declares no hard DSH version gate: no engines.dsh, no DSH

peerDependencies, no install-time version check. Earlier package releases were tested on DSH 0.2.1-rc.3; use the compatibility guidance above for 0.7.0 (see [DSH-side constraints](#dsh-side-constraints)).

Member usage snapshots

Cards show host-reported token totals, elapsed time and approximate context composition. Missing values show a dash. The public projection read may return a cached snapshot; these figures are not guaranteed final usage or billing totals. Tests use real React with simulated services, not live-team E2E. See [validation scope](docs/member-usage-validation.md).

Update notes

0.6.3

• Default prompts are shorter than in 0.6.2; the reduction varies by role and custom settings. No fixed per-turn or token saving is claimed.
• Removed repeated instructions and examples.
• Prompt-level task budgets (not runtime-enforced counters): up to 30 tool calls for a single-file task.
• The prompt budget for a multi-file task is up to 100 tool calls.
• The prompt limits retries for the same failure to 2 and requires the task brief to state its budget and stop conditions; this is not a runtime guarantee. These budget limits predate 0.6.3.

0.6.2

• Explains when falling back to the public source.
• Shows the source of installed plugins; unknown when unreadable.
• Keeps state separate per connection.

0.6.1

• Memory plugin can install from a local path.
• The default registry source needs network access. A local source is for first installation only; dependencies may still need the network, so offline success is not guaranteed.
• Updates from local paths are rejected.
• Fallback to the registry is limited to a completed source inspection failure before installation; timeouts or thrown errors do not trigger fallback.
• Tests reject the literal /home/fslong in client.js; a separate check rejects selected absolute-path patterns in cordis.patch.yml. These checks do not detect every hardcoded home path.

0.6.0

• Member config page now has five tabs.
• Plugin page adds a memory plugin install card.
• Task assignment spells out reporting points and deliverable scope.
• Reassign only after the active writer stops.
• Remote install reads the official response structure.

Optional memory plugin

The Plugin tab installs the public pinned @rsrsai/dsh@0.1.13 package. Installing and enabling are confirmed separately; a restart notice must be followed. Newer versions are not silently downgraded. Unknown versions are reported as unknown. Each remote wait has a two-minute deadline; a timeout does not cancel the host job. If the outcome remains unknown, check the native plugin manager before retrying. Closing the card stops stale UI updates and does not automatically enable a still-installing bundle.

Cabinet member settings

After a bundle install, open Settings → Cabinet members (内阁成员). The five

office groups have a display name (40 characters), portrait and prompt (4000

characters). Names remain accompanied by the original member ID in Cabinet cards

and Owner labels. Task ownership and native roster names never change.

Edit, then click **Sa