dsh-plugin-omoslim

DeepSeek Harness (dsh) bundle that installs an Orchestrator agent preset in the style of oh-my-opencode-slim: the main agent plans and dispatches, and model-pinned specialist subagents do the work.

What you get

One agent preset, orchestrator:

• Main agent = Orchestrator — a workflow manager persona (plan → dispatch → reconcile → verify), delegating instead of implementing.
• 11 subagent tools, each with its own persona and pinned model (routed through the opencode-go pi-ai provider):

Tool	Role	Model
subagent	generic worker (background, continuable)	inherit
subagent_fork	forks parent context	inherit
subagent_explorer	fast codebase navigation (read-only)	deepseek-v4-flash
subagent_oracle	architecture / review advisor (read-only)	glm-5.2
subagent_librarian	docs / external research	minimax-m2.7
subagent_designer	frontend UI/UX	kimi-k2.6
subagent_fixer	bounded implementation	deepseek-v4-flash
subagent_councillor_alpha/beta/gamma	independent multi-model reviews	glm-5.2 / kimi-k2.7-code / qwen3.7-max
subagent_council	multi-model consensus synthesis	kimi-k3

The models above mirror the OMO opencode preset found in ~/.config/opencode/oh-my-opencode-slim.json. The OMO council model gpt-5.6-luna is not in the opencode-go catalog, so kimi-k3 is used.

How it works

A dsh bundle is an npm package declaring "dsh": { "bundle": { "patch": "./cordis.patch.yml" } }. This bundle's patch layer inserts one row: the preset-installer plugin. On boot the plugin installs/updates config/presets/* in the harness-home user preset root (~/.dsh/.agent-presets), which the agent-presets roster always scans.

Why install into the user root instead of registering a new root? The dsh launcher's composeProfile forcibly overwrites agent-presets.config.roots with the shipped root, so a bundle cannot contribute its own root. The user root is the supported extension path.

Model definitions are separated from the composition

Each subagent's model and persona live in different files on purpose — you are far more likely to tweak a model than a persona, and plugin updates should not clobber your model choices:

File (in the bundle)	File (in ~/.dsh/.agent-presets/orchestrator/)	What you edit
config/models.json	models.json	model mapping (copied on install; user-owned afterwards)
config/presets/orchestrator/agent.cordis.yml.tmpl	agent.cordis.yml (rendered)	personas / tool wiring (generated from template)
—	.generated	render stamp { renderedHash } (managed)

Every boot the plugin re-renders agent.cordis.yml from the current template

• the current models.json. Rules:

• Model change → edit models.json, restart dsh. Re-render happens.
• Plugin update (new personas/wiring) → re-render happens on next boot; your models.json is never overwritten.
• Hand-editing the composition → if you edit agent.cordis.yml directly, the plugin detects it (hash differs from the stamp) and leaves your file alone. You own it then; delete the preset directory to take plugin updates.

Install

# from the profile directory (web is the default profile)
dsh plugin --profile web add dsh-plugin-omoslim

dsh plugin forwards to pnpm and reconciles dsh.profile.bundles, so the package is both installed and activated as a bundle layer. Then restart the web app:

systemctl --user restart dsh-web
# or: restart your `dsh web` process

For local development instead of a published package:

# in ~/.dsh/profiles/web/
pnpm add file:/home/<you>/Coding/dsh-plugin-omoslim
# then add the package name to the "dsh.profile.bundles" array in package.json

Use

Open the web UI, start a new session and pick the orchestrator preset in Settings → General (it becomes the default only if you set it, or if ~/.dsh/settings.yaml has agent-presets.default: orchestrator). The main agent will plan and dispatch; the subagent tools appear in its tool catalog.

Uninstall / rollback

1. Remove the bundle: dsh plugin --profile web remove dsh-plugin-omoslim (or drop the dependency + dsh.profile.bundles entry manually), restart.
2. The preset files stay in ~/.dsh/.agent-presets/orchestrator/ — delete that directory to remove them, or keep it to keep using the preset without the plugin.

Customizing

• Change a subagent's model → edit ~/.dsh/.agent-presets/orchestrator/models.json, restart dsh (or the dsh-web service). The plugin re-renders agent.cordis.yml on the next boot and your model mapping is never overwritten by plugin updates.

  { "oracle": "glm-5.2", "explorer": "deepseek-v4-flash", "...": "..." }
  

• Change personas / tool wiring → edit the template in the bundle (config/presets/orchestrator/agent.cordis.yml.tmpl) and bump the plugin, or hand-edit the rendered ~/.dsh/.agent-presets/orchestrator/agent.cordis.yml — hand edits are detected (stamp mismatch) and left alone. To take plugin updates again, delete the preset directory first.

• Reinstall the pristine preset → rm -rf ~/.dsh/.agent-presets/orchestrator and restart dsh.

Layout

dsh-plugin-omoslim/
├── package.json            # dsh.bundle.patch declaration
├── cordis.patch.yml        # bundle patch layer (inserts the installer row)
├── config/
│   ├── models.json         # DEFAULT model mapping (independent file)
│   └── presets/orchestrator/
│       ├── preset.yml
│       └── agent.cordis.yml.tmpl   # composition template (@@models.<key>@@ slots)
└── lib/index.js            # Cordis plugin: idempotent installer + template renderer