deep-researchlisted

<!-- prettier-ignore-start --> **Document Version:** 2.0 **Last Updated:** 2026-04-05 **Status:** ACTIVE <!-- prettier-ignore-end --> # Deep Research Multi-agent research engine that does what a single conversation cannot. Decomposes questions, dispatches parallel searcher agents, synthesizes with citations and confidence levels, runs mandatory contrarian/OTB challenges, and produces structured output with downstream routing. ## Critical Rules (MUST follow) 1. **Always show research plan before executing** -- present decomposition, agent allocation, cost estimate. Wait for approval. `--auto` skips plan approval only (all other checkpoints still require interaction). Recommend `--auto` only when invoked by another skill or for topics with prior research. 2. **Contrarian + OTB mandatory at ALL levels** -- no exceptions. 3. **Floor depth is L1 (Exhaustive)** -- no shallow modes. If someone wants quick, they ask Claude directly. 4. **Write to disk first** -- findings must survive crashes. **Output capture fallback:** On Windows, agent output files may be 0 bytes (anthropics/claude-code#39791). After each background agent completes, check the output file size. If 0 bytes, capture the task-notification result text and write it to the expected output file using the Write tool. Empty results must never be silently accepted -- if no output is available via any channel, report the failure to the user. 5. **State file updated after every state-changi