research-workflow

research-workflow (Dynamic Workflow asset)

A Dynamic Workflow asset (see CLAUDE.md -> "Dynamic Workflow Assets"; review-workflow is the reference example). research.workflow.js is executed by the Workflow tool, invoked from /lets:research's Workflow Mode via:

Workflow({ scriptPath: "${CLAUDE_PLUGIN_ROOT}/skills/research-workflow/research.workflow.js", args })

${CLAUDE_PLUGIN_ROOT} is substituted at command-load time, so research.md carries the literal absolute path. Treat research.workflow.js as a template the command points at, not a script to reproduce inline.

Why this is a transparent performance lever (not autonomous)

/lets:research --workflow runs the SAME stages as the standard Task path - only the per-sub-question search dumps and per-claim verdicts stay off-context. The decompose (break the question into 3-6 sub-questions) stays IN-CONTEXT in research.md so the user can steer the angles; the per-sub-question web research and the per-claim cross-check move into the workflow (off-context); the final synthesis happens in-context after the aggregate returns. The standard path runs the equivalent stages in-context: per-sub-question research fanned out across multiple DEFAULT web Task subagents, per-claim lets:skeptic via Task. Crucially: the verify stage is a cross-check (single-source / contradicted / low-confidence flags), never a claim that the fact is confirmed correct - the skeptic has no web tools and cannot re-fetch URLs.

What it does (off-context)

1. Research - per-sub-question fan-out via the DEFAULT web subagent (no agentType; the default workflow subagent CAN WebSearch/WebFetch - lets:* agents have tools: Read, Grep, Glob, Bash and CANNOT). Each subagent returns its 2-5 strongest claims (FINDING_SCHEMA: claim, evidence, sources, confidence) with evidence carrying QUOTED/paraphrased source material (the downstream skeptic judges "unsupported" against this string and has no web tool), and self-reports used_web_search=false rather than fabricate. mergeClaims dedupes by normalized claim text, unioning sources and keeping the highest confidence.
2. Verify - per-claim lets:skeptic RESEARCH-VERIFY pass over the merged claims. The skeptic is handed each claim's evidence + its siblings (same sub-question) and flags STRUCTURAL weakness: unsupported (evidence does not back the claim) or contradicted (conflicts with a sibling). Single-source / low-confidence are computed deterministically by the script (those claims skip the skeptic - the deterministic flag already fires). applyVerdicts attaches a flagged[] array per claim - ADDITIVE, never drops a claim (research facts accumulate; departure from review's drop rule). Contradiction flagging is the skeptic's job comparing siblings - there is deliberately NO deterministic contradiction pass.