@shijunan123/dsh-subagent-effort

dsh-subagent-effort

A DeepSeek Harness plugin, mounted the dsh way — everything is a plugin: a bundle with a cordis.patch.yml, installed/removed through dsh plugin, configured from the GUI (Settings → Plugins → configurable) or the profile patch layer.

It lets the harness's in-process subagents (subagent, subagent-fork-in-process, subagent-spawn-in-process, and workflow children) run with their own model and reasoning effort — chosen by the GUI settings panel, by config (defaults), or by the parent model per delegation (the subagent_select tool).

Dependencies: @deepseek-ai/dsh-tools (tool definition), @deepseek-ai/schemastery (config schema), @shijunan123/dsh-client-ui-subagent-effort (settings card). Pure ESM, no build step.

Lifecycle — one command, clean uninstall

Install / update / uninstall go through manage.ps1 (project root), which keeps the profile's pnpm-workspace.yaml release-age excludes and leftover dirs consistent — dsh plugin alone leaves those behind.

powershell -ExecutionPolicy Bypass -File manage.ps1 install     # install latest
powershell -ExecutionPolicy Bypass -File manage.ps1 update      # upgrade to latest
powershell -ExecutionPolicy Bypass -File manage.ps1 uninstall   # remove + clean leftovers
powershell -ExecutionPolicy Bypass -File manage.ps1 status      # what is installed

After install/update: restart dsh web and hard-refresh the browser.

Config

GUI (recommended)

Settings → Plugins → configurable → Subagent model and reasoning effort card: default model, default reasoning effort (low/medium/high/max or inherit). Writes to settings.yaml; takes effect immediately, no restart.

Profile patch layer (optional composition defaults)

- patch:
    - id: subagent-effort
      config:
        model: deepseek-v4-flash        # subagent default model (optional)
        reasoningEffort: high           # low|medium|high|max (optional)
        provider: spawn                 # delegation provider for subagent_select (default spawn)
        toolName: subagent_select       # model-facing tool name (default)
        enableRunInBackground: true     # expose run_in_background on the tool

Every field is optional; the bundle ships inert (subagents inherit the parent; the tool is registered but unused until the model picks it). Runtime precedence: subagent_select per-call > settings section (settings.yaml) > profile config > inherit.

Behavior

Part 1 — config defaults

Subagent children carry AgentOptions.subagentDepth >= 1. An agent/request waterfall listener rewrites model / reasoningEffort for those agents only; root agents are never touched.

Precedence per subagent request:

1. the child's own AgentOptions.reasoningEffort wins (set by the model through subagent_select, or smuggled through the built-in tool's agentOptions — the tool schema does not advertise the key, but undeclared keys pass through);
2. the settings.yaml section subagent-effort applies otherwise;
3. then the profile patch config;
4. with neither set, the request keeps its inherited effort.

settings.yaml (and then profile config) is authoritative for the subagent model when set.

Part 2 — subagent_select tool

A model-facing delegation tool like the built-in subagent tool, plus two optional parameters:

parameter	meaning
model	model id for the subagent (e.g. deepseek-v4-flash, deepseek-v4-pro); omit to keep the parent's model
reasoningEffort	low | medium | high | max; omit to keep the parent's effort

The parent model picks the model/effort per task (cheap + fast for routine work, strong + max for hard problems). Chosen values travel through agentOptions and are applied by part 1.

To avoid two delegation tools in the prompt, rename the built-in tool first by patching its base rows (tool-subagent, tool-subagent-fork) with a distinct toolName, letting subagent_select be the one tool.

Install

From npm (published):

dsh plugin --profile web add @shijunan123/dsh-subagent-effort

From a local checkout:

dsh plugin --profile web add link:C:\path\to\subagent-effort

Then restart dsh web and hard-refresh the browser. Remove with dsh plugin --profile web remove dsh-subagent-effort.

Verify

dsh --profile web --dump-config   # the subagent-effort row is mounted
node --test test/                 # 25 behavior unit tests, mock ctx, no model needed

Model Experience

Indirectly, through the subagent_select tool schema and its description, which reach the parent model's prompt; the plugin itself adds no other model-visible text.

KV Cache effect

The tool description is a fixed, prefix-stable literal at registration; no per-request prefix changes, so it does not invalidate cache reuse.

Notes

• Works for every in-process subagent path, including workflow children and the built-in subagent tool (its agentOptions.reasoningEffort is honored even though its schema does not advertise it).
• Does not affect external product subagents (Codex / Claude Code / ACP bridges); those translate their own effort flags.
• Both parts use documented extension points (agent/request waterfall, ctx.tools.register) — no core changes.

Publish

Publishing requires an npm account with 2FA; the registry rejects CLI publishing unless a granular token with Bypass 2FA for publishing is used (Classic tokens are no longer creatable). Flow: bump versions in packages/*/package.json, publish the client package first, then the host:

# in the project root, with a bypass-2FA token in ~/.npmrc or a temporary project .npmrc
pnpm publish --filter @shijunan123/dsh-client-ui-subagent-effort --no-git-checks
pnpm publish --filter @shijunan123/dsh-subagent-effort --no-git-checks   # rewrites workspace:* deps

pnpm pack output is verified: the tarball carries lib/, cordis.patch.yml, README.md, package.json (dsh.bundle.patch included, so dsh plugin add recognizes it as a bundle).