Self Healing Ops — AI skill for Claude Code

Self-Healing Ops

A LangGraph agent that repairs broken n8n workflows behind an independent cross-vendor review gate.

When an n8n workflow fails, the agent diagnoses the root cause, decides whether the failure is fixable in the workflow definition or needs a human, drafts a corrected workflow JSON, has a model from a different vendor independently review that fix, and then either patches the workflow through the n8n API (snapshot-first, with rollback) or escalates to a human in Slack.

The point of the cross-vendor step: the model that proposes a fix never approves its own work. A reviewer from a different vendor has different training and different blind spots, which makes the second opinion genuinely independent. And the reviewer has no write path, so even if it is wrong or manipulated, it cannot apply anything.

Why it is safe

Two properties carry the safety, not the model choice:

1. A deterministic diff gate. Before any write, code (not a model) hard-rejects a patch that touches a node outside the diagnosed set, adds a node that makes a network call, references a new credential, or changes a trigger. This does not depend on the reviewer noticing anything.
2. Snapshot-first remediation. Every write is preceded by a durable snapshot and applied as a transaction with per-step rollback. Nothing is overwritten without a way back.

On top of that: the reviewer is read-only, untrusted execution data is fenced and never reaches the reviewer in raw form, the revise loop is bounded, and `DRY_RUN` is on by default.

The pipeline

error fires -> triage -> route --auto-fixable--> propose -> review --approve--> gate -> remediate
                              |                                |                    |
                              |                          --reject (bounded)-->     |
                              |                                revise -> review     |
                              +--human-only--> escalate <--gate blocked /