dsh-research-autoresearch

The most grueling part of research is rarely the writing—it is the invisible pipeline in between: screening literature, scoring it, tracking progress, detecting stalls, simulating peer review, and iterating. DSH ships tools, but passive tool descriptions are almost never invoked spontaneously. Post-release forensics on 83 session logs revealed zero calls to all six tools, pointing to a missing orchestration layer. dsh-research-autoresearch turns the Deli Chen AutoResearch framework into a working DSH bundle: it registers an autoresearch runtime skill that injects the full protocol into research sessions, and mounts six execution tools—arxiv-search for arXiv recall with dedup, lit-score for LQS five-dimension scoring with emerging-citation imputation, research-state for a seven-operation progress panel, stall-check for pivot / escalate / cap traffic-light signals, peer-review for five-persona median scoring with judgeDisagreement and a MAD-based scoreBand, and research-ui for a runtime visual control panel. Three-level toggles (suite, deployment, session) let you switch visualizations on or off without affecting any tool logic.

Who is it for? If you are using DeepSeek Harness to run literature reviews, draft proposals, polish a survey through self-play iterations, or simply want your agent to act inside the research loop rather than emit a single paragraph of advice, this bundle is built for you. It does not write the paper for you, but it turns the screen → score → track → stall-check → review chain into a state you can observe, replay, and switch off at a moment's notice.

The design choices are grounded in empirical evidence: median scoring over mean (to resist single-persona score inflation), honest score-dropped signaling, structural pivot guidance over tactical tinkering, and robust uncertainty bands drawn from LLM-as-a-judge literature. Every historical defect is locked into regression assertions, and the build is fully self-contained for git-based installation.