websearch-searchlisted

# Web Search channel (Stage 4d) — keyless, agent-driven A zero-dependency discovery channel for users who have **only Claude Code** and no search accounts (no SearchAPI / Gemini / Undermind). It uses the agent's own **WebSearch / WebFetch** tools to find real literature on the open web, then funnels the hits through the same dedup -> verify -> screen pipeline as every other channel. There is **no subprocess driver** here: a Python script cannot run WebSearch. The agent (you, in Claude Code) does the searching; `scripts/websearch_ingest.py` only normalizes what you gather into the pipeline schema. ## When to use - The user has no `SEARCHAPI_API_KEY`, `GEMINI_API_KEY`, or Undermind login, OR - You want a broad open-web sweep (working papers, very recent work, SSRN / arXiv / NBER / OpenReview / publisher pages) alongside the keyed channels. Run it **in parallel** with whatever other channels are available; its output merges with theirs at dedup. ## Recipe (agent-driven, subagent fan-out) This is the default. The orchestrator emits a batched task plan, fans the batches out across parallel Opus subagents — so the raw WebSearch/WebFetch text stays inside the subagent contexts — and then merges the distilled candidates. It runs the same way whether web search is the sole channel (no keys) or an add-on alongside the keyed channels. 1. **Emit the task plan** from the Stage-0 queries: ```bash python websearch-search/scripts/websearch_ingest.py --emit-tasks \ --qu