dsh-web-search-ddg
aooyoo
Zero-token DuckDuckGo search provider for the DeepSeek Harness (DSH) web seam — local headless browser, no API key, no m…
PROJECT TOPICS
PROJECT README
Automatic Flash/Pro routing for DeepSeek Harness subagent calls.
Your main conversation keeps running on a strong model (e.g. V4 Pro). When the agent delegates a task to a subagent, this plugin classifies the task and sends light tasks to a cheap model (e.g. V4 Flash) while heavy tasks stay on the strong model — deterministically, without relying on the model's cooperation.
provider/model parameters are added to the
subagent / subagent_fork tools, so you can force a model on any call./delegate command — per-session mode: /delegate auto | off | flash-all.budgetCapTokens, new subagents downgrade to Flash.The plugin wraps the official subagent / subagent_fork tools on each agent's
scoped context (same seam the proven dsh-reasoning-settings uses), classifies
the task inside the wrapper, and rewrites the child's first request in the
agent/request waterfall — the official lifecycle stays untouched. Verified
end-to-end: a "search" delegation ran the child session on
deepseek-v4-flash while the parent session stayed on deepseek-v4-pro.
dsh plugin --profile web add dsh-delegate-router
Set the Flash/Pro routes either in the profile patch config or via environment variables (both are optional; the router only acts when the relevant route is configured):
setx dsh_delegate_router_FLASH_PROVIDER "<provider>"
setx dsh_delegate_router_FLASH_MODEL "<flash-model-id>"
setx dsh_delegate_router_PRO_PROVIDER "<provider>"
setx dsh_delegate_router_PRO_MODEL "<pro-model-id>"
Restart DSH afterwards. Modes: auto (default), off, flash-all; switch at
runtime with /delegate <mode>.
npm install
npm run smoke
node scripts/e2e-delegate.mjs # browser E2E against a test instance
MIT
Built for the DeepSeek Harness plugin ecosystem — thanks to the community on LINUX DO for feedback and testing.
CLASSIFICATION EVIDENCE
系统优先读取 GitHub Topics,再与站内分类词典和词根规则比对。