oh-my-dsh-slim

A port of oh-my-opencode-slim's specialist subagent delegation for DeepSeek Harness (DSH): an orchestrator + 5 specialist roles, each with its own persona, model, tool permissions (toolFilter), reasoning effort, and MCP access. Delivered as a shareable DSH agent preset (with bundled configuration plugins), not a standalone application.

Persona text adapted from oh-my-opencode-slim (MIT © 2025 alvinunreal), attribution retained — see LICENSE. 中文版见 README.zh.md.

⚠️ DSH version requirement (0.5.0): this release targets DSH 0.1.2-rc.1 (latest). DSH changed substantially in 0.1.2, so oh-my-dsh-slim 0.5.0 is NOT compatible with older DSH releases — on DSH 0.1.1 or below, stay on oh-my-dsh-slim 0.4.0. If you upgrade anyway on an older host, nothing breaks: the plugin detects the mismatch, leaves your existing preset directory untouched (it stays fully usable), and shows a notice under Settings → Plugins → oh-my-dsh-slim-compat. Also note: upgrading requires a DSH restart (plugin code is mounted once per host process; new sessions alone do not pick it up).

What it solves

DSH's default orchestration is "one model does everything". This preset splits work into specialist lanes, and the orchestrator plans, dispatches, and integrates:

• oracle — architecture decisions, complex debugging, code review. Read-only.
• designer — UI/UX design and visual polish. Writable.
• fixer — bounded mechanical implementation. Writable.
• explorer — fast codebase reconnaissance. Read-only.
• librarian — external research via official docs / GitHub (context7 + gh_grep MCP). Read-only.

Delegation is background-first (continuable): the orchestrator dispatches lanes and ends its turn; the runtime wakes it with a settlement notice when a lane finishes, so it can integrate results. The orchestrator follows a strict delegation discipline — after dispatching independent lanes it ends its turn with a brief status note (no polling, no re-doing a running lane's scope in the same turn), treats a lane's interim report as "not settled yet", and only finalizes on the settlement notice. The subagent_result tool reads a finished subagent's final message without waking it (zero extra model turns).

Role matrix

Role	Tool	Default model	Effort	Permissions
oracle	subagent_oracle	deepseek-v4-pro	max	read-only
designer	subagent_designer	deepseek-v4-flash	high	writable
fixer	subagent_fixer	deepseek-v4-flash	high	writable
explorer	subagent_explorer	deepseek-v4-flash	low	read-only
librarian	subagent_librarian	deepseek-v4-flash	high	read-only + MCP

• All roles inherit global tools; read-only roles deny edit/write; all roles deny control tools (OMO-style deny-only policy)
• Roles cannot delegate further (maxDepth: 1); librarian exclusively mounts context7/gh_grep in its own child scope
• observer (visual analysis) is reserved but disabled in this release: DSH's send-time gate blocks image attachments based on the main model's vision capability, and delegation prompts are text-only, so pasted images cannot reach a subagent yet. It will be re-enabled when the harness supports forwarding message attachments into subagent contexts.

Install

Requires DSH 0.1.2-rc.1 or newer and a DeepSeek API key (default models route through deepseek-official). Older DSH releases are not supported by 0.5.0 — on DSH 0.1.1 or below use oh-my-dsh-slim 0.4.0 (see the version note at the top).

Option A — plugin marketplace GUI (recommended): open Settings → Plugins in the DSH web GUI, search for oh-my-dsh-slim in the marketplace, and install. It is also listed in the awesome-dsh-plugin catalog.

Option B — CLI:

dsh plugin --profile web add oh-my-dsh-slim

The package ships a seeder that materializes the preset into $DSH_HOME/.agent-presets/oh-my-dsh-slim automatically (updates come with plugin upgrades; your previous copy is backed up).

ℹ️ This uses the default harness home (~/.dsh). If your deployment uses a custom home (e.g. an isolated desktop-app environment), set DSH_HOME to it first — the marketplace GUI (Option A) resolves this automatically.

Option C — git clone:

git clone https://github.com/ninipa/oh-my-dsh-slim "$DSH_HOME/.agent-presets/oh-my-dsh-slim"

Done — the preset appears immediately. Create a new session and pick 极简角色委派 in Settings → Agent Presets.

• Update: cd "$DSH_HOME/.agent-presets/oh-my-dsh-slim" && git pull (or upgrade the plugin), then restart DSH — plugin code (tool schemas, injected reminders) is mounted once per host process, so new sessions alone do not load the new code
• Rollback: git checkout <old-tag> or just delete the directory. Presets are locked per session at creation time; running sessions are unaffected.

Configuration

Zero configuration required (sensible defaults ship with the preset). User intent is read in descending priority:

1. The file pointed to by the OH_MY_DSH_SLIM_CONFIG env var (test/CI channel)
2. The host settings namespace oh-my-dsh-slim (recommended): the bundled seeder registers this namespace, so configuration lives under the oh-my-dsh-slim: section of the host's settings.yaml. Effort/temperature are re-read on every delegation (edits apply immediately); model/maxTokens apply to new sessions
3. The legacy $DSH_HOME/oh-my-dsh-slim.json file (fallback for hosts without a settings service). On hosts with the npm package installed, the file is imported into the settings namespace on first boot and archived as oh-my-dsh-slim.json.imported-<timestamp>

All three channels share one document shape (schema: oh-my-dsh-slim.schema.json):

{
  "preset": "my-dsh-normal",
  "presets": {
    "my-dsh-normal": {
      "fixer": { "model": "kimi-k3", "effort": "high" },
      "librarian": { "mcps": ["context7", "gh_grep"] }
    }
  }
}

• Per-role overrides: enabled / model / effort / deny / mcps; temperature / maxTokens are advanced keys (advanced.roles.<roleId>)
• Effort vocabulary: none omits the reasoningEffort parameter entirely — for models that do not support effort control (e.g. local LLMs without a reasoning-effort field); off explicitly disables reasoning on models that support the parameter. Other levels (low/medium/high/max …) are model-scoped: the adapter accepts only the levels the selected model declares (the DeepSeek adapter, for example, accepts off/low/high/max and rejects medium), and an unsupported level fails loudly at request time. The GUI card's effort dropdown is built from each model's declared set and blocks out-of-set values with an inline warning
• Model validation: at delegation time the configured model id is checked against the providers you imported in Settings → Models. An unknown model fails loud on the first delegation, listing every imported model (including the vision-capable subset) — no silent breakage
• observer is locked: observer.enabled: true is ignored with a warning (see above)
• Changes take effect in new sessions; running sessions are unaffected

GUI card (ships with the npm package): after install, a card appears under Settings → Plugins → Plugin configuration — per-role toggle/model/effort editable inline, advanced maxTokens/temperature behind a warning sub-section, and a model dropdown sourced from the same catalog as the composer's picker. The effort dropdown is scoped to the selected model's declared reasoning efforts (unsupported levels are hidden; an explicit mismatch warns inline and blocks save). The orchestrator row is informational only: it is the session's main model, changed in the composer's picker (defaults under Settings → Models). Saving reports which changes apply immediately (effort/temperature) and which start with new sessions.

Conversational configuration (no JSON editing needed): just say e.g. "change fixer's model to kimi-k3" or "disable the oracle role" — the orchestrator edits the JSON per the schema.

Multiple configurations (multi-preset): the settings card has a Delegation configuration dropdown on top (it appears when the seeder is installed — its /omds RPC feeds the roster). It manages named configurations, each backed by its own native agent preset:

• The dropdown always lists 极简角色委派 / Minimal Role Delegation (the bundled profile — the new-session default until you change it) plus ＋ New configuration. Choosing "＋ New configuration" edits an in-place draft copied from the configuration you were just editing: nothing is written until you hit Save, which then asks only for a display name (the internal id is generated from the name and never changes afterwards).
• Restore defaults resets only what you are currently editing; it never deletes a configuration or clears the roster.
• Selecting a configuration only chooses what is edited — the current session never switches. Which configuration a NEW session uses is decided by the native Agent preset picker and its default; the card's Set as default for new sessions button writes exactly that native setting, so the card and the picker always agree (clicking a preset card in the picker is the same write).
• Saved configurations become real agent presets: directories under $DSH_HOME/.agent-presets/profile-<prefix>-<hash>/, selectable in the Agent preset picker like any other preset. Each profile's per-role settings are stored as its own snapshot (profile.json beside the preset composition), which is how two profiles never leak into each other.

web_fetch (follows the host)

Since host DSH 0.1.2 the harness itself ships web_fetch with built-in SSRF protection, enabled by default for every session and every delegated child. This preset no longer wires its own fetch provider or exposes a webFetch switch: the previous "advanced configuration" section and the web-fetch-gate plugin were retired. web_search remains available from the host web service.

Roadmap

• observer re-enable — waiting on upstream DSH support for forwarding message attachments into subagent contexts (see the role matrix note above).

Self-tests & probes (all zero-cost)

# Static validation (structure / keys / persona dead-references / soft-disable assertions)
node scripts/t0-validate.mjs .

# Unit tests (config merge / effort injection / delegation contract / subagent_result /
# settings schema / sandbox strip / early-close ledger / preset seeder / profile RPC / client card)
node scripts/test-config-loader.mjs && node scripts/test-effort-plugin.mjs
node scripts/test-role-subagent.mjs && node scripts/test-subagent-result.mjs
node scripts/test-settings-schema.mjs && node scripts/test-sandbox-strip.mjs
node scripts/test-early-close-context.mjs && node scripts/test-preset-seeder.mjs
node scripts/test-profile-rpc.mjs && node scripts/test-client-card.mjs

# Host-contract probe battery (8 probes / 9 phases, zero-model) — run after every DSH upgrade.
# Bootstraps a scratch DSH_HOME automatically (no credentials needed); see scripts/TEST-INVENTORY.md
node scripts/run-host-probes.mjs          # --list / --only <name> / --keep for options

Acceptance checklist

GUI-TEST-TASKS.md provides 7 non-explicit-dispatch scenarios (prompts + expected behavior) for verifying a fresh deployment. T3 uses the baseline project under examples/omo-probe-baseline.

Known limits

• Upgrading requires a DSH restart: the agent-plane composition mounts once per host process — config rows (persona text, model routes) re-resolve per session, but plugin code (tool schemas, tool descriptions, injected reminder strings) is frozen in-process. After any plugin/preset upgrade, restart DSH; new sessions alone still run the old code
• Non-vision main models cannot receive pasted images: DSH blocks image attachments at send time based on the main model's capability (MODEL_DOES_NOT_SUPPORT_IMAGES). For image analysis, use a vision-capable main model (e.g. deepseek-v4-flash-vision-exp) directly — or wait for upstream attachment forwarding. If your model actually supports images but is still blocked, check whether its provider configuration declares the image input modality (input: ["text", "image"]) — a common gap for third-party GPT-class models
• web_search is billed separately: librarian prefers MCP (free). web_search runs through the host search service, which issues an independent auxiliary model request per query. For open-ended research, give the task a search budget in the prompt
• Delegated children cannot escalate sandbox permissions — the preset strips stray escalation fields (sandbox-strip plugin, a workaround): DSH fixes a delegated child's file policy and approval state at startup, but the bash/edit/write tool schemas still expose optional sandbox_permissions / justification fields. Some models fill those fields unprompted; a child cannot escalate anyway, so the extra arguments only trigger parameter-validation errors (invalid justification, not strictly wider). The bundled sandbox-strip plugin removes the two fields from role-subagent child tool calls at the tools/pre-execute waterfall and appends a [sandbox: stripped ...] note to the result so the model sees the correction. In this preset's own top-level sessions it additionally strips only the shapes DSH would always reject before any approval prompt (empty justification, single-field pairs, non-widening modes — judged with the host's own WIDER_MODES table); legitimate escalation requests (strictly wider mode + non-empty justification) are kept and still prompt for approval. Sessions that do not use this preset never load the plugin, so their behavior is unchanged. This is a preset-level workaround, not a fix: the real fix is upstream — DSH should stop exposing escalation fields to children whose permission scope is fixed
• Background subagents and "early close" (early-close-context plugin): DSH is turn-based — a model either outputs or ends its turn; there is no mechanism-level way to force it to wait for a background subagent. Some models therefore close with a final conclusion while a delegated child is still running, claiming "done" without the child's result. The bundled early-close-context plugin mitigates this by supplying the model with facts: a live "currently running background subagents" block in the system prompt (re-rendered every turn, same mechanism as the host's sandbox:policy), a "Decision point" reminder attached to every successful delegation result, and a persona clause against claiming completion while a child is unsettled. The ledger is three-state (running → reported → settled): a child's interim report (host frame "Agent sent a message:") is shown as "已回报内容，等待正式完成通知（reported ≠ 完成）" — a report neither concludes the child's turn nor changes its lifetime, and only the finish notice (unconditional for every established child, incl. failure/cancel/token-ceiling) settles it, so the orchestrator no longer announces "the subagent is done" tens of seconds early. In practice the delegation turn reports "still running; cannot output a final conclusion yet", defers dependent work until the finish notice, and wakes to integrate the result. The model may still end its turn before the child settles (no force-wait), but it no longer misreports completion
• Custom profile presets keep the plugin versions they were copied with: a profile is a full copy of the bundled preset directory at creation time; upgrading the npm package re-seeds only the bundled preset, so old profile directories keep their copied plugins until you recreate or copy them again (their configuration snapshots survive — only the plugins age)

FAQ

Which API key do I need? A DeepSeek API key — the default role models route through deepseek-official. Roles can be pointed at any provider you imported in Settings → Models.

Can I use other models per role? Yes — every role's model and reasoning effort is configurable via the user JSON (or conversational config). Unknown model ids are rejected at delegation time with the full list of imported models.

How do I uninstall? Remove $DSH_HOME/.agent-presets/oh-my-dsh-slim (or disable the preset in Settings → Agent Presets). If you installed the marketplace seeder plugin, uninstalling the plugin does not remove the preset directory.

Image analysis? Use a vision-capable main model (e.g. deepseek-v4-flash-vision-exp) and paste directly. The observer role is reserved until the harness can forward attachments into subagent contexts.

Changelog

See CHANGELOG.md.

Acknowledgments

• oh-my-opencode-slim (MIT © 2025 alvinunreal) — role system and persona source
• DeepSeek Harness — host platform

License

MIT