dsh-research-loop

Auto-research loop plugin for DeepSeek Harness. Two halves:

• Tool research_log — durable research memory: an append-only JSONL log per topic (.dsh-research/.jsonl in the workspace) with append / read actions. Entries carry a phase (plan | search | read | synthesize | conclude), a self-contained note, and source URLs — so multi-step research survives context compaction and session restarts.
• Skill auto-research-loop (skills/auto-research-loop/SKILL.md) — the loop discipline: bounded plan → search → read → synthesize iterations, source-quality ordering, conflict recording, explicit stop conditions (evidence coverage / diminishing returns / iteration cap), resume-from-log.

Status: example / reference skeleton from awesome-deepseek-harness. Structure follows observed community plugin conventions (see plugins/README.md); not yet verified against a live DSH instance. Search/fetch tools themselves are assumed to come from the harness or other plugins — this plugin deliberately only contributes the durable-state + methodology layer.

Install

pnpm install && pnpm build
dsh plugin --profile  add file:/absolute/path/to/plugins/dsh-research-loop

Then restart that profile.

Tool: research_log

Parameter	Type	Notes
action	string	append | read
topic	string	stable topic label; slugged into one log file per topic
phase	string	append only: plan | search | | |