websearch-search

name	websearch-search
description	Run the lit-review orchestrator keyless agent-driven web search channel that uses WebSearch and WebFetch outputs normalized through websearch_ingest.py. Use when the user invokes the web search channel, asks for Stage 4d open-web literature discovery, or needs a Claude Code web-search fallback without SearchAPI, Gemini, or Undermind credentials.

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:

   python websearch-search/scripts/websearch_ingest.py --emit-tasks \
       --queries-file OUT/scholar_queries.json --research-question "<rq>" \
       --batch-size 3 -o OUT/websearch_tasks.json
   

   This writes {system_prompt, research_question, tasks:[{batch_id, queries:[...]}]}. With no queries it prints WEBSEARCH_DEFERRED and writes an empty plan.
2. Fan out across Opus subagents — one per tasks[k]. Hand each subagent the system_prompt, the research_question, and its queries, and have it run WebSearch on each query, WebFetch the most promising hits (publisher / SSRN / arXiv / NBER / OpenAlex / Semantic Scholar) to read the real title, authors, year, venue, DOI, and abstract — never inventing a field, leaving unknowns "" — and Write its candidates to OUT/websearch_results_batch_<id>.json:

   [{"title": "...", "authors": "First Last, Second Author", "year": "2021",
     "journal": "...", "doi": "10.xxxx/...", "url": "https://...", "abstract": "..."}]
   

   Only title is required. Do NOT WebFetch scholar.google.com (bot-blocked).
3. Merge the partial files into the stage output:

   python websearch-search/scripts/websearch_ingest.py \
       --results OUT/websearch_results_batch_*.json -o OUT/stage4d_websearch.json
   

   This dedups by title across all batches, does best-effort keyless Crossref DOI fill (--no-enrich to skip), and writes stage4d_websearch.json (+ .ris, source="websearch"). The stage[0-9]*.json dedup glob then picks it up. If no batch yielded a usable candidate it prints WEBSEARCH_DEFERRED and writes an empty file, so the pipeline continues on the other channels.

Inline fallback (a handful of queries)

For a small query set you can skip the fan-out: run WebSearch/WebFetch yourself, collect everything into one OUT/websearch_results.json, and merge the single file (--results accepts one or many):

python websearch-search/scripts/websearch_ingest.py \
    --results OUT/websearch_results.json -o OUT/stage4d_websearch.json

Anti-hallucination

Web hits are real records, so fabrication is far lower than asking the model to recall papers from memory. It is not zero — a snippet can carry a wrong year, or a non-peer-reviewed page can slip in — so keep Stage 5b verification ON: it confirms every paper against OpenAlex / Crossref / Semantic Scholar and drops anything that cannot be confirmed. Never pair this channel with --no-verify.

Notes

• Keyless: the only network call the script makes is the keyless Crossref polite pool (set LITREVIEW_CONTACT_EMAIL to use your own contact). No LLM API calls.
• Google Scholar itself is bot-blocked, so do not WebFetch scholar.google.com directly; rely on WebSearch results and on fetching the underlying source pages.
• Coverage depends on what surfaces in search; this is a strong keyless baseline, not a replacement for Undermind / Deep Research / the SearchAPI Google Scholar channel.