Resetting per-session tool budgets by spawning headless claude -p from inside a session

Built-in WebSearch is capped per SESSION (~200 calls) and shared across all subagents, so a wide fan-out exhausts it and every later agent silently degrades to worse sources. Workaround: spawn fresh headless sessions with claude -p "..." --allowedTools "Read,Write,WebSearch,WebFetch,Bash" --permission-mode acceptEdits, each of which gets its own budget. Verify it rather than assuming by having the test session report SEARCH_WORKED/SEARCH_FAILED. Critical caveat learned the hard way: a fresh session has NO context, and a context-free session will confidently conclude an entity does not exist when the underlying data simply is not public. Pass context by writing brief and input files to disk and naming those paths in the prompt, and explicitly tell the session which facts are ground truth so absence of search results is not read as disproof. Operational details: launch with nohup and background them so the launcher returns immediately, which means the harness completion notification refers to the LAUNCHER, not the work; confirm with ps and by polling for output files. Stagger starts (sleep offsets) when the sessions share a paid third-party scraping key, since concurrency, not quota, is usually what trips its rate limit. Unrelated bug from the same session worth remembering: splitting an interleaved list of two categories into N batches with stride slicing (items[b::N]) clusters by parity, so with alternating categories and N=4 two batches got 100% of one category. Round-robin each category separately instead.