Ask Reddit some twenty more searches on a first sweep, and read deeper where page one found buyers

[kev1n]

The median project's first sweep ran 6 searches, found about 670 posts and ended with 29 buyer leads; 47% of the searches found nothing.

What changes

• sweepSearches (new, src/lib/scan/searches.ts): one model call at the start of a first sweep writes about 18 more Reddit searches: 5 plain asks for a tool, 4 "alternative to ", 3 symptom, 3 acting-for-someone, 2 workaround, 1 moment. They are never stored as problem phrasings, so Google discovery and the Reddit SEO tab are untouched. A failed call adds nothing and the sweep carries on.
• First pass is 1 page (25 posts) of every search instead of 2. Only the 10 walks whose first page found the most buyer leads read on to 6 pages.
• The 8 made-up searches that found the most buyers are kept as plan keywords (source = sweep, active), so recurring scans ask them too. Discovery's plan publish leaves them alone; a person can switch them off on the Sources tab.
• Post budget (2,500) and lead cap (150) are unchanged.

Measured (offline, real Reddit search, real triage and judge; data in .context/widen, not committed)

• Experiment, 24 prod projects, 557 searches: tool asks 16.8 leads per 100 posts and rival searches 12.0, against 8.1 for the page's phrasings; blind Opus check of 312 leads rated 73% / 67% / 51% real. Page-1 leads rank a search's yield at Spearman 0.89; title pass rate at 0.37.
• Gate, 11 projects the experiment never saw, today's sweep against this one:

	Today	This PR
Buyer leads, median project	27	73
Buyer leads, total (capped at 150 each)	375	803
Today's leads still found		337 of 375
Leads per 100 posts	5.8	6.4
Reddit search cost per project	$0.010	$0.025
Posts judged per project	325	801
Blind Opus check: good / weak / bad	37% / 44% / 19% (91)	new-only leads 52% / 32% / 15% (130)

• Adds about 17s (the model call) before the first search. All-in cost per sweep goes from about $0.07 to about $0.12 at list model prices.
• Not helped: products with almost nobody asking (WinstarPack, DONGFU, SSCG stayed near zero or bad).

Tests: two backfill tests updated for the one-page first pass, three added. Full suite passes except tests/jobs.test.ts "seeds a scan", which times out locally against the shared dev database before this change too.

Existing projects are not touched: this only runs on a first sweep.

🤖 Generated with Claude Code