flatten-workflow-parallel-arrays-before-collecting

Workflow pipeline(): a stage returning parallel() is a bare array the collector silently drops

Problem

In the Claude Code Workflow tool, pipeline(items, stage1, stage2, …) runs each item through the stages. A common fan-out-then-verify shape makes stage-2 return different shapes per item:

const per = await pipeline(DIMENSIONS,
  d => agent(findPrompt, {schema: FINDINGS}),          // stage 1
  (findings, d) => {
    if (!findings.confirmed.length)
      return { dimension: d.key, verified: [] };        // OBJECT shape (no findings)
    return parallel(findings.confirmed.map(f => () =>   // BARE ARRAY shape (has findings)
      agent(verifyPrompt(f), {schema: VERDICT}).then(v => ({...f, verdict: v}))));
  });

per[i] is now an object {verified: […]} for dimensions that found nothing, but a bare array [{…verdict…}, …] for dimensions that DID find candidates (because parallel() resolves to an array). A collector written for only the object shape:

const all = per.flatMap(p => Array.isArray(p?.verified) ? p.verified : []);  // BUG

evaluates Array.isArray(undefined) → false for every array-shaped element → returns [] → silently drops all verdicts from exactly the dimensions that found bypasses. The workflow returns confirmed: 0, refuted: 0 — a false "all clear." The real tell: agentCount is high and the journal shows verify/v2: agents both started AND result-ed, yet confirmed_count + refuted_count == 0 (verify agents only spawn for candidates, so candidates existed).

Context / Trigger Conditions

• A Workflow script using pipeline() where a later stage conditionally returns parallel(...) (adversarial-verify / review / loop-until-dry patterns).
• The workflow result is "0 confirmed / 0 refuted" or "0 findings" — suspiciously clean.
• The completion summary's agentCount (or the journal under …/subagents/workflows/wf_*/journal.jsonl) shows verify agents ran, contradicting "0".