dsh-subagent-effort
A DeepSeek Harness plugin that gives you per-call control over a subagent's thinking intensity.
The stock subagent / subagent_fork tools make the child inherit the session's reasoningEffort. This plugin adds one tool — subagent_effort(prompt, effort, cwd?, model?) — that runs a one-shot child with effort = low | high | max for that call only. The child otherwise joins your preset like a stock subagent: same persona, prompt sections, skill catalog and tools.
subagent_effort(prompt="List the files under src/", effort="low") # mechanical work, cheap
subagent_effort(prompt="Refactor this module and explain the risks", effort="max")
Why it works
reasoningEffort is a first-class per-request field of the subagent seam, not a profile or preset setting:
SubagentStartRequest.agentOptions documents provider / model / reasoning-effort / output-token overrides and requires the provider's agentOptions capability.- The host's
resolveChildAgentOptions(parent, requested, depth) spreads the requested options over the parent's route; an explicit effort survives both inheritance and a model switch.
- The LLM layer validates the effort against the model's advertised
reasoning.efforts before any request and rejects unsupported values (UNSUPPORTED_REASONING_EFFORT) — no clamping, no silent fallback.
The plugin therefore registers one small provider implementing the 0.2.1 local activation contract (prepareContinuable) and advertises agentOptions: true. The host materializes the child, mounts the parent's preset composition and applies the requested effort. The plugin also pre-checks the model/effort pair through the llm service so a bad combination fails with a readable message instead of an opaque child error.
Requirements
- DeepSeek Harness 0.2.1 or newer (the activation API
ctx.subagents.startActivation is mandatory).
- Node.js 22+ (the host ships its own runtime; this is only for
node --check).
- A model that advertises reasoning efforts. With the DeepSeek account provider,
deepseek-flash advertises low / high / max.
Install
From your DSH profile directory (for the desktop app: %USERPROFILE%\.dsh\profiles\desktop):
-
Add the dependency to package.json:
"dependencies": {
"dsh-subagent-effort": "file:C:/path/to/dsh-subagent-effort"
}
A published tarball or git URL works the same way (npm: / git+https: specifier).
-
Append the bundle so the plugin is mounted with the profile:
"dsh": { "profile": { "bundles": [ "...", "dsh-subagent-effort" ] } }
-
Install and restart the app:
pnpm install
Plugins are loaded at host startup — a restart is required after installing or updating. If you edit the plugin source in place, remember that a file: dependency is copied into node_modules, so re-run pnpm install (or copy the files) before restarting.
Configuration
The plugin reads its settings from the bundle row (all optional):
| option | default | meaning |
|---|
providerName | subagent-effort | provider name registered on ctx.subagents |
toolName | subagent_effort | name of the model-facing tool |
- id: dsh-subagent-effort
name: dsh-subagent-effort
config:
toolName: subagent_effort
Tool
| parameter | required | description |
|---|
prompt | yes | Self-contained task for the child. It does not see this conversation. |
effort | yes | low | high | max, validated against the model's advertised efforts. |
model | no | Model id for the child (same provider as this session). Omit to inherit. |
cwd | no | Absolute working directory for the child. Omit to inherit the session's. |
The call is foreground: it waits for the child and returns its final text. Use low for mechanical, deterministic work (run a command, read files, apply a template), high for judgment without long reasoning chains, and max for cross-file reasoning, source-level diagnosis or design trade-offs.
How it works (implementation notes)
apply() registers a provider implementing prepareContinuable — the host's local backend. No child agent is created by the plugin; the host materializes it and applies request.agentOptions.- The tool builds
{ provider, label, signal, delivery: 'caller', request: { maxDepth, parent, prompt, agentOptions, cwd? } } and awaits ctx.subagents.startActivation(spec).
delivery: 'caller' keeps the host from pushing a settlement notice into the parent's inbox — the tool result is the only delivery channel for a one-shot call.maxDepth is parentDepth + 1, so the child may not delegate further while a nested caller still works.- Result first, dispose second: detaching the child before settling the result would cancel the turn.
Verified vs. not verified
Verified on DSH 0.2.1-rc.1 / 0.2.1-rc.2 (Windows desktop, 2026-10-17/18):
- the tool registers and appears in the tool surface;
subagent_effort(effort="low") returns completed with the child's text;- the child session records
config.reasoningEffort: "low" (effort="max" likewise); the ids are 48d6d4fe-… (low) and 02d17f6e-… (max);
- an unsupported effort or an unknown model fails before delegation with the supported list.
Not verified:
high was not separately recorded in a child session log (the value is accepted and goes through the same validation path).- Only the Windows desktop host was exercised; other shells/hosts are untested.
- The tool is foreground-only; there is no background/job mode (use your host's background delegation for that).
License
MIT © 2026 tietoumiao
