v5.1.0

Hardening pass from a multi-session friction analysis of real production runs: documented gotchas
that were never enforced, and a noisy feedback metric.

Added

• build-workflow-v2 design-quality gate. workflow-reviewer now runs as a hard gate between
  Validate and Deploy (≤2 fix cycles) in both greenfield + edit flows. Design anti-patterns the n8n
  validator cannot catch — raw HTTP in Code nodes, continueOnFail/onError masking real errors, AI
  sub-nodes miswired via .out().to() — now block the push instead of shipping. Warnings are
  surfaced (reviewWarnings) without blocking.
• Orchestrator StructuredOutput resilience. Schema'd subagent calls in build-workflow-v2 +
  build-stack-v2 now route through a safe() wrapper: a subagent that ends without calling
  StructuredOutput (or dies terminally) gets one retry, then a graceful fallback into the existing
  gate-failure path — instead of crashing the whole run with an opaque error after burning the
  tokens (observed: a heavy agent burned ~786k tokens, then the workflow aborted opaquely).
• build-stack-v2 greenfield auto-detect. Before decomposing, a greenfield stack build checks
  docs/*.architecture.md for a stack that already covers the use-case and returns
  needs-decision (re-run as extend, or pass mode: 'greenfield' to force) instead of silently
  rebuilding a working stack from scratch.
• Feedback sync → public repo. Target moved from the private n8n-autopilot-internal to the
  public neurawork-git/n8n-autopilot. repoLabel (a customer basename = PII on a public repo) is
  stripped from the issue title, summary, and raw NDJSON before push; records keep it locally.
• SessionStart plugin-staleness probe (check-plugin-version.sh) — INFO nudge to
  claude plugin update when the installed version is behind the latest GitHub release. Never
  auto-runs (env-changing). Catches the case where a stale install silently lacks newer gotcha hooks.
• Cheat-sheet rows: credential-schema-first (n8nac credential schema <type> before create) and
  sub-workflow-publish-before-parent.

Fixed

• Wrong execution-list flag in the cheat-sheet/reference: --workflow → --workflow-id (n8nac
  rejects --workflow; the documented form caused the same CLI error across sessions). 5 files.
• REST-guard dodge. The PreToolUse curl-block only matched curl/wget, so urllib /
  Invoke-RestMethod / requests. slipped past it. Guard now catches those against /api/v1 too;
  the /api/v1/data-tables carve-out is preserved, and the skill blesses looping its curl for
  polling (instead of reading n8nac's internal secret store).
• env-blind workspace status trap. CLAUDE.md recommended workspace status to verify the env,
  but it reports the GLOBAL active env — this misled the model into the forbidden env use across 3
  sessions. Docs now route env verification to env list --json.
• Noisy feedback metric. capture-feedback.sh matched skill-listing + SessionStart-hook
  injections (false positives) and double-counted on resume. Now scans only real conversation turns
  and keeps one last-write-wins event per session.