Rebuild Vibeflow as an n8n safety gate

plugins/vibeflow/skills/vibeflow/SKILL.md

@@ -0,0 +1,40 @@
+---
+name: vibeflow
+description: Audit and repair exported n8n workflow JSON with deterministic safety and contract checks. Use when reviewing AI-generated or customer-facing n8n workflows before deployment, investigating Vibeflow VF000-VF009 findings, adding kill switches or human handoffs, checking secrets, retries, and webhook exposure, or preparing workflow changes for CI and pull requests.
+---
+
+# Vibeflow
+
+Use the CLI as the source of truth. Do not infer that a workflow is safe from its README, prompt, or visual layout.
+
+## Workflow
+
+1. Locate the exported workflow JSON. Never inspect or mutate a live production workflow unless the user explicitly requests it.
+2. Run Vibeflow from a repository checkout:
+
+   ```bash
+   node bin/vibeflow.mjs check path/to/workflow.json
+   ```
+
+   For a released version without a checkout:
+
+   ```bash
+   npx --yes github:domfelipe/vibeflow-n8n#v0.8.0 check path/to/workflow.json
+   ```
+
+3. Read [references/policies.md](references/policies.md) when interpreting or repairing a finding.
+4. Fix errors before warnings. Preserve credential references, expressions, node IDs, and unrelated workflow behavior.
+5. Re-run the check after every repair. Stop only when blocking findings are gone or explicitly waived in `.vibeflow.json` with a documented reason.
+6. Report changed nodes, remaining warnings, and what still needs runtime validation in n8n.
+
+## Safety boundaries
+
+- Never copy literal credentials into a workflow to silence `VF001`.
+- Never disable `VF006` for customer-facing agents without explicit user approval; an off switch must block inference and all AI responses.
+- Treat static analysis as a preflight, not proof of runtime correctness.
+- Prefer fixing a shared upstream node over duplicating guards across branches.
+- Keep fixes local to the exported workflow until the user authorizes deployment.
+
+## Output
+
+Return the CLI result, the smallest safe repair, and any unresolved runtime check. Use SARIF for GitHub code scanning and JSON for automation.

plugins/vibeflow/skills/vibeflow/agents/openai.yaml

@@ -0,0 +1,4 @@
+interface:
+  display_name: "Vibeflow"
+  short_description: "Safety checks for AI-generated n8n workflows"
+  default_prompt: "Use $vibeflow to audit this n8n workflow and fix blocking findings."

plugins/vibeflow/skills/vibeflow/references/policies.md

@@ -0,0 +1,16 @@
+# Vibeflow policy reference
+
+| ID | Default | Meaning | Minimum repair |
+|---|---|---|---|
+| VF000 | error | Invalid workflow export | Export valid n8n JSON with a `nodes` array. |
+| VF001 | error | Literal secret in parameters | Use n8n credentials, a vault, or an environment expression. |
+| VF002 | error | Host-level node | Remove it or explicitly approve the node type in policy. |
+| VF003 | warning | Webhook without supported auth and a credential reference | Add header/basic/JWT auth backed by an n8n credential. |
+| VF004 | warning | External action without a connected failure path | Connect a `continueErrorOutput` branch or configure `settings.errorWorkflow`. |
+| VF005 | warning | Inbound side effects without an atomic deduplication gate | Use a claim that emits no item for duplicates, such as `INSERT ... ON CONFLICT DO NOTHING ... RETURNING`, on every path. |
+| VF006 | error | An AI entry path bypasses the structural kill switch | Put a real IF/Switch/Code gate that reads agent status on every entry path. |
+| VF007 | warning | AI Agent without a reachable external handoff action | Add a downstream ticket, chat, email, or equivalent human escalation action. |
+| VF008 | warning | Missing or excessive execution timeout | Set the workflow timeout between 1 and 3600 seconds. |
+| VF009 | warning | Unsafe retry policy | Add idempotency, bound attempts, and configure backoff. |
+
+Configuration changes severity or domain vocabulary; it does not prove the suppressed risk is safe. Keep waivers visible in `.vibeflow.json` and explain them in the pull request.