subagent-handbacklisted

# Subagent Research Reliability ## Problem Dispatching subagents for research has three silent failure modes. Each produces a confident result that is worth nothing, and they fail in dispatch order: 1. **The findings are a dead letter.** The agent researches correctly, then finishes with nothing the caller can use. Platform docs (verified 2026-09-06, code.claude.com/docs/en/tools-reference.md, sub-agents.md): intermediate tool calls and outputs stay in the subagent context; only the final text result is the return path — inline for foreground, as a completion notification for background. That path is real, and it is not enough: observed 2026-08-04 and 2026-08-24, background agents complete with an empty handback, and work that lived only in the subagent transcript is gone. Name an explicit payload channel in the dispatch, every time. 2. **The agent can't actually search.** Its `tools:` grant lacks WebSearch/WebFetch even though its description says "performs web research." It returns a no-op or fabricates citations from training data. 3. **The agent returns claims that do not hold.** Fabricated citations, CVE IDs, arXiv IDs and cross-references are the web case. The general case is wider: a misquoted rule, a line number dressed as a section, and — the dangerous one — **a checked negative that is false**, where the agent reports "zero hits across all five files" and one of the five hits. All three are invisible if you trust the returned tex