dsh-subagent-setting
Configure the default provider / model / reasoning effort for subagents from the DeepSeek Harness Web settings page. Newly created subagents pick the values up immediately; an optional switch also lets later settings changes reach subagents that were created earlier.
Requires DSH 0.2.x.
image-20260821012945376
Features
- Settings page (Settings → Subagent Model): pick provider, model and reasoning effort from the live model catalog — no manual YAML editing.
- A subagent can run a different model than its parent — the reason this plugin exists. The parent session keeps its own model; children go where you point them here.
- New subagents use the new settings immediately — the route is snapshotted when the child is created and applies to that child alone.
- Optional live propagation (Apply setting changes to subagents created earlier): when enabled, a settings change also updates every existing idle subagent — the new values apply on its next request (a running subagent picks them up from the next step). When disabled, only subagents created after the change are affected.
- Empty provider / model / effort means inherit the parent. With all three empty the plugin does not interfere at all.
- WebView-friendly custom dropdowns: the settings form uses fully controlled custom dropdowns instead of native
<select> popups, so it behaves the same in the browser and inside a Tauri (WebView2) shell.
How it works
- Host (
lib/index.js) is a plugin whose Config is the settings document: all five fields are marked volatile(), and the 0.2 settings surface projects only volatile fields — so the entry shows up on its own, with no namespace to register.
- On
agent/created, for every agent whose session.header.origin === 'subagent', it installs an agent/request waterfall listener on that agent's own scope and rewrites the final LlmCallConfig to the configured route. Scope filtering is what keeps that listener to this one child.
- The listener reads the child's live route holder, so
applyToIdle only needs to replace that holder to re-point an existing child on its next request. Config changes arrive through the loader's loader/volatile-update event.
- The reasoning effort is validated against the route's real capability with
llm.resolveModelInfo() before it is injected. The settings dropdown reads the same call (remote.session.modelCatalog → buildModelCatalog), so the two can never disagree; when the model genuinely does not support the level, the field is dropped rather than sent and rejected.
- Changing the route without naming an effort drops the inherited effort so the new model resolves its own default — the same rule the Host's own child-option resolution applies.
- Client (
client/client.js) reads and writes that entry through ctx.configForms, reads the catalogue through ctx.remote.session.modelCatalog(), and uses ctx.configForms.whileServed() so the page appears only while the Host actually serves the plugin.
Install
dsh plugin --profile desktop add <spec>
You can also install a local directory from the sidebar Plugins page. The package declares dsh.bundle.patch, so it joins the profile's bundle list on install.
The Host half depends on @deepseek-ai/schemastery to declare its Config schema. A third-party plugin's module resolution cannot see the DSH-bundled @deepseek-ai/* packages, so it is declared as an ordinary dependency and installed into the profile from npm, pinned to the runtime's own version (3.18.4).
Open Settings → Subagent Model to configure it.
Known limitations
- The reasoning-effort dropdown is only populated when the model advertises reasoning. A model hand-declared under
llm-pi-ai without reasoningEfforts is treated by pi-ai as carrying no reasoning metadata, and only inherit is offered. Declare reasoningEfforts on that provider's model entry to expose levels for it.
- Editable from the Web settings page only; the auto-generated plugin configuration page does not list this entry (that surface projects volatile fields too, but only for bundle rows).
License
MIT
