outnever/dsh-research-first

dsh-research-first nudges a DeepSeek Harness agent to do a cheap, deterministic investigation before editing code - without hard-blocking: it injects a model-visible reminder when an edit or command runs uninvestigated, detects blocked research (web_search or GitHub unreachable) and hands the decision back to the user, with intensity levels remind/warn/block.

dsh-research-first nudges a DeepSeek Harness agent to do a cheap, deterministic investigation before editing code, targeting three compounding failure modes: guessing before acting, trial-and-error loops and stale knowledge. Rather than hard-blocking - which would stall work in air-gapped or documentation-free environments - it defaults to a soft nudge that allows the edit but injects a model-visible, logged reminder. It also detects blocked research (for example web_search failing or GitHub unreachable) and reminds the agent to ask the user whether to continue, adds 'golden research habits' to the system prompt (search with versions, prefer official tools, explain when blocked), reminds after repeated failures to check community feedback first, and can escalate to warn or block.

Workflow & Automation ★ 1
View on GitHub ↗

Install

dsh plugin --profile web add github:outnever/dsh-research-first

Copied verbatim from the README's install block ('Option 1: official'); a local form is also documented.

Compatibility

A DeepSeek Harness plugin with three intensity levels - remind (default soft nudge), warn (hold for approval) and block (deny). Every reminder is a plugin-sourced message in the session log, so behaviour is replayable and auditable. It composes with engineering-discipline plugins such as dsh-doublecheck rather than replacing them.

Details

Alternatives

940842546/dsh-permissions · adithya-hmt/fullstack-expert · alib8b8/dsh-plugin-aflare

More plugins in Workflow & Automation

Browse more in Workflow & Automation

Guides for Workflow & Automation plugins