dsh-socrates

The hardest part of deep research is rarely finding information — it is producing a report you actually dare to cite. dsh-socrates engineers the Socratic principle of clarifying before digging: a Planner generates up to three disambiguation questions on the topic, and the research loop only starts after you confirm the direction, cutting wasted runs at the source.

The plugin closes the loop end to end. In adaptive multi-round research, parallel researchers search, close-read, and verbatim-extract evidence each round; at round end, high-priority gaps are collected for marginal-gain verification, and the loop halts on consecutive zero gain, backed by a hard round cap and a token budget. In cross-validation, a Verifier issues a three-way verdict (entailed / contradicted / baseless) on every claim, with single-hop backfill and conditional revision rounds, leaving unverified claims explicitly flagged. Citations may only come from verbatim evidence snippets capped at 100 characters, audited programmatically for numbering, format, and exact-match consistency. Built-in access to arXiv, PubMed, and Semantic Scholar (no API key) plus automatic model tiering — Pro for planning and synthesis, Flash for research — keep costs in check.

It suits any workflow that demands a trustworthy, auditable, numbered-citation report: technology stack comparisons, ecosystem surveys, literature reviews, competitive analyses. Simply state your research intent and purpose in plain language; the plugin decides how many sub-questions to spawn, how many rounds to run, and whether to invoke adversarial review. Pass depth 3 for exhaustive coverage, or enable review for citation correction and coverage auditing — the rest converges on its own.