dsh-second-opinion

The primary model answers, the stream closes, and the reply lands in your transcript. Spotting a shaky assumption, a hallucinated API parameter, or a skipped edge case usually means scrolling back and re-reading at your own pace. dsh-second-opinion fires an independent reviewer model asynchronously after the assistant message is committed. It re-reads the full context, anchors each concern to the exact quoted text, and surfaces a structured verdict — all without delaying, blocking, or overwriting the original turn.

Two modes keep you in control. The default advisory mode stores the review as a suggestion; you choose whether to send one comment or the whole verdict back for reconsideration. In auto mode, a concern or disagree verdict automatically triggers a bounded revision, capped by autoRevisionsPerTurn so the loop can never grow unbounded. The reviewer is handed only a single structured submit_review response tool and none of the primary agents executable tools, eliminating any risk of side effects during review. Reviews run serially within a session for context consistency and concurrently across sessions, with results persisted in a Storage Domain sidecar that supports retry and per-comment replay.

If your daily work leans on Harness multi-turn tool workflows and you want a second pair of eyes on every committed answer without sacrificing first-token speed, this plugin is the quiet quality inspector: it never changes the conversation cadence, never adds perceived latency, and simply leaves a quote-anchored checklist beside the speech bubble when you are ready to look.