narrative-prompt-polish

Anyone who works with AI agents regularly hits the same wall: you have the idea, maybe even a rough draft, but handing it straight to the agent often falls short—vague phrasing, missing context, unclear intent. narrative-prompt-polish exists to close that last mile from draft to executable prompt. Hit the ✨ button beside the main input, and the plugin spins up a sidebar conversation where the agent inherits the full session context (system prompt plus history), rewrites your draft into a specific, action-ready prompt, and stays available for multi-round refinement. When you are happy, one click copies the result back into the main input—without ever touching the session history.

Under the hood the plugin follows a three-tier fallback chain. In full mode it replicates the [system]+[history] prefix to hit the prompt cache, so only the draft and polish instruction are computed incrementally. If context is limited it degrades gracefully to partial (recent conversation window) or none (bare draft), logging every step into the response and a JSONL trace. Local regex classifies the input into one of four intents—debug, implement, explain, or chat—and shapes the prompt skeleton accordingly. A CAS fingerprint check guards the write-back: if the draft changed during polishing, the plugin refuses to overwrite rather than silently losing edits.

This plugin is built for writers and engineers who drive AI agents daily inside DSH Web, for automation pipelines that can hit the single-shot polish endpoint without a sidebar, and for teams that want every polish run auditable via structured traces (intent, context mode, cache hit, token counts). The multi-round sidebar flow depends on better-sidebar; without it the single-shot path still works, so the plugin is ready to use out of the box either way.