dsh-research-first

LLM coding agents suffer from three compounding failure modes: writing code from memory instead of checking the actual API, looping through trial and error without searching for known community workarounds, and acting on stale versions or endpoints. Each skipped cheap lookup becomes an expensive rework cycle downstream. dsh-research-first targets exactly this gap.

Rather than hard-blocking the workflow, it injects a soft, traceable reminder whenever a mutation tool fires before any investigation tool has run in the current step. If the investigation path itself is unavailable (air-gapped machines, missing docs), the plugin detects the failure and nudges the agent to ask the user instead of stalling. After a configurable number of consecutive failures it prompts a quick GitHub Issues check before another blind attempt. Every nudge lands in the session log as a plugin-sourced message, fully replayable and auditable. Need stronger enforcement? Set intensity to warn or block.

Designed for teams that want a lightweight, zero-dependency research habit by default. It composes with dsh-doublecheck (strict discipline gates) and dsh-pain-point-check (failure vetoes), serving as the everyday background nudge that keeps the agent honest without ever freezing the pipeline.