@sstreichan/oh-my-dsh-slim
English | 中文
A function-form Cordis plugin that ports the oh-my-opencode-slim agent pantheon — Orchestrator + 7 specialists, slash commands, AST-grep + webfetch + task control tools, 9 bundled skills, context7 + gh_grep MCP fallbacks, multiplexer pane adapters, interview dashboard, and a Rust companion window — to DeepSeek Harness.
What it does
Registers the following on the dsh Context:
- The pantheon as real agents — the bundle layer restates the base
agent-loop row so orchestrator + 7 specialists (explorer, librarian, oracle, designer, fixer, council, councillor) exist as selectable agents in every surface. Each gets a stable sessionId (omo-<name>) and the base default model route. observer is omitted to match disabledAgents: ['observer'].
- A web agent preset — on boot the plugin seeds an
orchestrator agent preset (preset.yml + agent.cordis.yml) into $DSH_HOME/.agent-presets/orchestrator/, where dsh web's agent-presets registry discovers it automatically and lists it in the session picker. Seeding is idempotent and never overwrites an existing preset.
- Specialist delegation — the bundle layer routes the base
subagent tool through the plugin's omo subagent provider (ctx.subagents.registerProvider). When the orchestrator delegates with an @handle (e.g. subagent(description: "@explorer: locate route handlers", ...)), the provider swaps the plain-text persona for that specialist's FULL prompt, applies its tool filter (explorer stays read-only, designer may write, ...), and routes its configured model (agents.explorer.model: 'anthropic/claude-3-5-sonnet' → the child runs on anthropic/claude-3-5-sonnet) before spawning the child via the standard in-process driver. Delegations that name no specialist get a neutral fallback identity.
- UI-configurable models — when
@deepseek-ai/dsh-settings is mounted (for example via @deepseek-ai/dsh-settings-file, which the dsh base bundle mounts by default), the plugin registers an oh-my-dsh-slim settings namespace and a dedicated Settings → Oh My DSH Slim menu with one model dropdown per pantheon agent. Options come from dsh's live host-wide model catalog. Edits persist to $DSH_HOME/settings.yaml, apply live to running agent-loop agents, and take effect for future specialist delegations; the Cordis config remains the base layer (see Settings UI: per-agent model selection).
- System-prompt sections —
omo:orchestrator-rules (the delegation prompt with <Role>, <Agents>, <Workflow>, <Communication>) and omo:agent-catalog (display names + roles for each specialist), contributed through ctx.systemPrompt.section({ name, order, text }). Both are orchestrator-scoped: their text provider inspects the assembly's agent and renders nothing for spawned specialists, their descendants, and specialist agent-loop agents — while the main user session (the omo-orchestrator agent-loop session, a headless fresh session, or a preset-less web session) IS the orchestrator lane and gets the full rules.
- Tools — 11 model-facing tools via
ctx.tools.register(defineTool(...)): ast_grep_search, ast_grep_replace, webfetch, acp_run, task_cancel, task_message, task_result, task_revive, task_status, wait_for_user, preset_switch.
- Slash commands —
/deepwork, /reflect, /loop, /preset via ctx.commands.register({ name, description, handler }). Handlers return { kind: 'success', text } and schedule model-visible work through agent.session.append('user/message', ...).
- Skills — 9 bundled
SKILL.md files (simplify, codemap, clonedeps, deepwork, verification-planning, reflect, worktrees, oh-my-opencode-slim, loop-engineering) registered via ctx.skills.registerProvider() when dsh-skill is mounted, or as ctx.systemPrompt.section() fallbacks otherwise.
- MCP fallbacks —
context7_search and gh_grep_search tools that fetch via HTTP when dsh-mcp-client is not mounted in the profile.
- Hooks — 7 event listeners via
ctx.on(event, listener): cache-monitor (agent/request), json-error-recovery (tools/post-execute), phase-reminder (agent/pre-step), image-routing (agent/pre-step), foreground-fallback (agent/request-error), orchestrator-wake (agent/turn-stopping), post-file-tool-nudge (tools/post-execute + agent/pre-step).
- Multiplexer pane adapters — tmux, zellij, kitty, herdr, cmux with auto-detection via environment variables.
- Interview subsystem — per-session HTTP server + optional dashboard with SSE.
- Rust companion — detached egui desktop window that polls a JSON state file.
Specialist delegation requires the subagent spine (dsh-subagent + dsh-subagent-spawn-in-process + dsh-tool-subagent, all mounted by the base bundle). The plugin layers an omo provider on top: the bundle patch points tool-subagent at it with backgroundMode: one-shot, so every delegation (foreground and background) flows through the provider, which resolves each delegation's specialist @handle to the specialist's full prompt + tool set + model (see Specialist delegation).
Export shape
A function/namespace plugin: it exports name / inject / Config / apply and NO default. A stray export default would collapse the module via the Loader's unwrapExports and drop inject (see docs/postmortem/0001).
inject: ['tools', 'agents', 'systemPrompt', 'llm', 'sessions']. Optional services (subagents, skills, commands) are resolved via ctx.get(name).
Configuration
| Field | Default | Meaning |
|---|
preset | — | Active preset name (must exist in presets). |
setDefaultAgent | true | Accepted for backwards compatibility but inert — dsh has no per-plugin default-agent knob (the web surface picks its default via the agent-presets row). The pantheon is registered through agent-loop.agents. |
autoUpdate | true | Periodic npm registry poll; surfaces a notice when an update is available. |
seedWebPreset | true | Seed the orchestrator agent preset into $DSH_HOME/.agent-presets/ so it appears in dsh web's picker. Idempotent; set to false to stop seeding (and delete the dir to remove an existing one). |
presets | — | Named sets of per-agent overrides; selected by preset. |
agents | — | Per-agent overrides applied on top of the active preset. model accepts a string or an ordered fallback chain [{id, variant?}]. |
disabledAgents | ['observer'] | Agents to omit from the prompt catalog. orchestrator and councillor are protected. The bundle layer's agent-loop.agents list is what makes an agent exist as a real selectable agent — remove a row there (or in your profile's cordis.patch.yml) to disable it outright. |
disabledMcps | [] | MCPs to skip (context7, gh_grep). |
disabledTools | [] | Tools to skip. |
disabledSkills | [] | Skills to skip. |
imageRouting | 'auto' | 'auto' routes images through @observer; 'direct' never intercepts. |
multiplexer | |
Install
Install the plugin into a profile with one command:
dsh plugin --profile web add @sstreichan/oh-my-dsh-slim
This does three things:
- Installs the package via pnpm into
$DSH_HOME/profiles/web.
- Detects the
dsh.bundle declaration and appends it to the profile's bundle stack.
- On the next boot, applies the bundle's
cordis.patch.yml layer — no manual config editing.
Then run dsh (restart if already running):
dsh web
Use any profile name (e.g. dsh plugin --profile my-profile add ...; dsh web is an alias for dsh --profile web).
Upgrade
dsh plugin --profile web add @sstreichan/oh-my-dsh-slim
Re-running add is idempotent and pulls the latest version. Pin a version with @sstreichan/oh-my-dsh-slim@0.1.0.
Uninstall
dsh plugin --profile web remove @sstreichan/oh-my-dsh-slim
Customizing config
Override the plugin's config in your profile's $DSH_HOME/profiles/web/cordis.patch.yml:
- id: oh-my-dsh-slim
name: '@sstreichan/oh-my-dsh-slim'
config:
preset: openai
setDefaultAgent: true
agents:
explorer:
model: anthropic/claude-3-5-sonnet
temperature: 0.1
Restart dsh for the change to take effect (hot-reload picks up cordis.patch.yml edits).
The pantheon agents
The bundle layer registers the orchestrator + specialists as real agents by overriding the base agent-loop row. Each row carries a stable sessionId (so remounts resume the same history) and the base default route (deepseek-official / deepseek-v4-flash). To customize, restate the whole list in your profile's cordis.patch.yml — a patch replaces a row's entire config, so include every agent you want to keep:
- id: agent-loop
config:
agents:
- id: orchestrator
sessionId: omo-orchestrator
provider: deepseek-official
model: deepseek-v4-flash
- id: explorer
sessionId: omo-explorer
provider: deepseek-official
model: deepseek-v4-flash
# ...
- Disable an agent — remove its row from the list.
- Re-enable
observer (disabled by default) — add its row.
- Pick different models — set
provider/model per agent; they must name a route with a registered adapter (the llm-deepseek row is the default).
The web agent preset
dsh web composes every session from an agent preset — that is what its session picker lists. On boot the plugin seeds an orchestrator preset into the harness-home user preset root ($DSH_HOME/.agent-presets/orchestrator/), which the agent-presets registry scans by default, so the orchestrator shows up in the web picker with no manual steps. The preset is minimal on purpose: the orchestrator's delegation rules, specialist catalog, and tools come from the plugin's registrations — the omo prompt sections render only for the orchestrator lane, so a session created from this preset sees the full orchestrator surface while spawned specialists never inherit it.
- Set it as the web default — the
agent-presets settings namespace holds the default ($DSH_HOME/settings.yaml → agent-presets: { default: orchestrator }, or the web UI's settings).
- Customize it — edit
$DSH_HOME/.agent-presets/orchestrator/agent.cordis.yml; the plugin never overwrites it.
- Remove it — delete the
$DSH_HOME/.agent-presets/orchestrator/ directory and set seedWebPreset: false in the plugin's config (otherwise the next boot re-seeds it).
Settings UI: per-agent model selection
When the settings service is mounted, the plugin publishes an oh-my-dsh-slim namespace to dsh's settings system and ships a web client bundle that renders a dedicated Settings → Oh My DSH Slim menu in the dsh web UI. The page lists one model dropdown per pantheon agent (orchestrator, explorer, librarian, oracle, designer, fixer, observer, council, councillor) in pantheon order.
Each dropdown is populated from dsh's live host-wide llm.models catalog and stores the selected value as provider/model (for example anthropic/claude-sonnet-4-5). The catalog refreshes when dsh reports adapter or settings changes. Previously saved values that are temporarily absent from the catalog remain available as a current compatibility option until replaced.
How edits behave:
- Persistence — saving writes the user layer into
$DSH_HOME/settings.yaml under oh-my-dsh-slim: (one flat key per agent). Fields left empty inherit the base layer.
- Reset — the
· marker next to a field label means the value is a user override; clearing the field (and saving) removes the override and restores the configured base value.
- Live application — saved values are pushed into the running agent-loop agents' option objects, so the next request from that agent uses the new model without restarting dsh or losing session history.
- Delegation routing — future specialist delegations resolve the model through the same value (UI override first, then
agents.<name>.model config).
Precedence for an agent's effective model, highest first:
- The settings UI value (
oh-my-dsh-slim user layer in $DSH_HOME/settings.yaml).
- The Cordis config's
agents.<name>.model (plugin config / preset / profile patch).
- The base agent-loop row's route (default:
deepseek-official / deepseek-v4-flash).
Requirements and notes:
- The settings provider row (
@deepseek-ai/dsh-settings-file, mounted by the dsh base bundle) must be active; without it the plugin logs a debug line and runs without UI settings — models then come from config only.
- The section schema is flat (one top-level field per agent). That shape is required by the web settings scope, which writes top-level path edits; do not restate it as a nested
agents object.
- Staged edits that do not parse are rejected at save time (the Save button stays disabled) rather than silently dropped; a failed save keeps the drafts so they can be corrected.
Manual profile (no dsh plugin)
If you cannot or prefer not to use dsh plugin, use the standalone profile template at examples/oh-my-dsh-slim-profile/cordis.yml. Mount it in $DSH_HOME/profiles/oh-my-dsh-slim/cordis.yml and run:
npm i -g @sstreichan/oh-my-dsh-slim
dsh --profile oh-my-dsh-slim web
Migrating from a manual insert? If your profile's cordis.patch.yml already inserts an oh-my-dsh-slim row by hand, remove it before running dsh plugin add. Otherwise the bundle layer and your manual insert both contribute the same row id, which crashes dsh on boot with duplicate loader entry id.
MCP servers (context7, gh_grep), drop the sample mounting configs from examples/mcp-context7.cordis.yml and examples/mcp-gh-grep.cordis.yml into your profile. These mount @deepseek-ai/dsh-mcp-client rows that provide the full MCP tool surface. Without them, this plugin registers degraded HTTP-fetch fallback tools.
Composition
The sample profile composes the dsh base (settings, credentials, llm-deepseek, session, agent-loop, tools, system-prompt, sandbox, bash, fs, todo, subagent spine, skill subsystem, commands adapter) plus this plugin:
| Plugin | Role |
|---|
@deepseek-ai/dsh-session + persistence + projection | Event-sourced session log with JSONL persistence and SQLite query. |
@deepseek-ai/dsh-agent + dsh-agent-loop | Agent factory + agent loop. |
@deepseek-ai/dsh-tools + dsh-system-prompt | Tool registry + system-prompt assembly. |
@deepseek-ai/dsh-subagent + dsh-subagent-spawn-in-process + dsh-tool-subagent | Specialist delegation spine. |
@deepseek-ai/dsh-skill + dsh-skill-filesystem + dsh-tool-skill | Skill registry + filesystem provider + skill tool. |
@deepseek-ai/dsh-commands | Slash command adapter. |
@deepseek-ai/dsh-tool-bash + dsh-tool-fs + dsh-tool-todo | Bash, filesystem, and todo tools. |
@deepseek-ai/dsh-sandbox + dsh-sandbox-policy + dsh-approval | Sandbox + permission gate. |
@sstreichan/oh-my-dsh-slim | This plugin: orchestrator + 7 specialists + tools + commands + skills + hooks + multiplexer + interview + companion. |
Model Experience
Orchestrator system prompt
What the model sees
The orchestrator sees a ~3k-token system prompt assembled from omo:orchestrator-rules (the delegation prompt with <Role>, <Agents>, <Workflow>, <Communication>) and omo:agent-catalog (specialist display names + descriptions). Specialist prompts add ~1k token each but are only visible when the orchestrator dispatches via the subagent tool.
Token effect
Fixed cost on every orchestrator turn. Specialist prompts are per-delegation, not per-turn.
KV Cache effect
Prefix-stable while the orchestrator prompt text and per-agent overrides are unchanged. Switching presets, disabling an agent, or hot-reloading a prompt invalidates reuse from the first affected system-prompt token.
Specialist delegation
The orchestrator delegates with the subagent tool (from dsh-tool-subagent), naming the specialist lane with its @handle in the description argument:
subagent(description: "@explorer: locate route handlers", prompt: "...", run_in_background: true)
The plugin's omo provider intercepts the call and:
- resolves the
@handle (canonical name, display name, or alias — in persona, description, or the prompt) against the configured pantheon;
- replaces the persona with the specialist's full prompt (e.g. the read-only
@explorer prompt, the write-capable @designer prompt) and applies the specialist's tool filter;
- routes the specialist's configured model into the child's
agentOptions — agents.<name>.model: 'anthropic/claude-3-5-sonnet' becomes provider: 'anthropic', model: 'claude-3-5-sonnet'; a bare model id ('deepseek-v4-flash') overrides just the model and keeps the parent's provider route;
- hands off to the same in-process driver the standard
spawn provider uses.
The delegation lane is backgroundMode: one-shot (not the base bundle's continuable), so every delegation flows through the provider, where the harness honors the specialist's persona, tool filter, AND model. One-shot background runs return a job id — collect with job_output, stop with job_kill.
If you switch the lane back to continuable (for durable child conversations and send_message follow-ups), the omo provider is bypassed — but the plugin's registerContinuableSetup contribution then installs the specialist's full prompt + tool filter into each continuable child from its durable descriptor, so only the specialist's MODEL is lost on that lane (continuable children materialize with the tool request's agent options).
The specialist child's own prompt assembly never includes the orchestrator's delegation rules: the omo prompt sections render only for the orchestrator agent. Delegations that name no specialist keep an explicitly configured persona or get a neutral fallback identity. To bypass resolution, point tool-subagent back at provider: spawn in your profile's cordis.patch.yml.
Token effect
Fixed schema cost. Each delegation adds the specialist's full prompt + result to the turn's token budget (a specialist prompt is ~1k tokens; the child keeps it in its own context, not the parent's).
KV Cache effect
Append-only; delegation results follow the reusable request prefix and do not invalidate existing KV-cache entries.
Skill catalog
What the model sees
When dsh-tool-skill is mounted, the model sees the skill tool schema plus a <system-reminder> catalog listing available skills. Without dsh-tool-skill, skills are injected as ctx.systemPrompt.section() entries visible only to the orchestrator.
Token effect
Catalog cost scales with the number of enabled skills (9 by default, minus disabledSkills).
KV Cache effect
Prefix-stable while the skill set is unchanged. Adding or removing a skill invalidates reuse from the catalog section.
Phase reminder and post-file-tool nudge
What the model sees
The agent/pre-step waterfall appends a <system-reminder> to the orchestrator's latest user message: the phase reminder (workflow discipline) and the post-file-tool nudge (todo update reminder).
Token effect
~50 tokens per injection, only on eligible orchestrator turns.
KV Cache effect
These are appended to the user message, not the system prompt, so they do not invalidate the system-prompt KV cache. They may invalidate the user-message cache if the message was previously cached.
Known Limitations and Deferred Work
- Subagent delegation requires the subagent spine —
dsh-subagent + dsh-subagent-spawn-in-process + dsh-tool-subagent (all part of the base bundle) must be mounted; the plugin's omo provider sits on top of them. Without the spine the provider is not registered and the subagent tool does not exist.
- Specialist resolution is lexical — the
omo provider matches @handle mentions (or bare specialist names) in the delegation request; it is not a registry of named personas. Delegations that phrase the specialist differently (e.g. "the frontend engineer") fall back to the neutral identity. The orchestrator prompt teaches the @handle convention to keep matches reliable.
- Omo-section identity is session-based — the orchestrator-scoped rendering matches the
omo-orchestrator agent-loop session id, the orchestrator agent preset, or any root (non-subagent, non-omo--specialist) session. A profile that renames the orchestrator's sessionId (or uses a differently-named orchestrator preset) must update OMO_ORCHESTRATOR_SESSION_ID / the preset match accordingly, or the omo--prefix specialist exclusion and the agent-preset check stop applying.
- Specialist model format is
provider/model — a specialist's agents.<name>.model string is split on the first / into dsh provider/model for delegated children. A bare model id keeps the parent's provider route. The route must name a registered adapter (the base default is deepseek-official / deepseek-v4-flash).
- Specialist tool filters name the base-bundle surface — the read-only filter's
allow list names tools that ship in the base bundle plus this plugin's own (read, glob, grep, ast_grep_search, webfetch, gh_grep_search, context7_search, job_output, job_list, list_agents). tools.restrict() rejects unknown names, so a profile that disables one of those tools (or mounts a minimal bundle without it) must trim the filter, or delegated specialists fail to start. The OpenCode-era names from oh-my-opencode-slim (lsp, list, codesearch, apply_patch, task, ) are gone.
