why? for n8n

Finds n8n runs that report success but produced nothing, drifted out of shape, or invented details that were never in the input.

n8n tells you when a workflow crashes. It says nothing when a workflow quietly stops working. A step returns nothing. An expression points at a field that got renamed last month. An API replies "200 OK" with an error buried in the body. Every step after that silently doesn't happen, and the run still shows up green, marked Succeeded. You find out three days later, when a client asks why they stopped getting leads. why? reads the execution history n8n already keeps and finds those runs. WHAT IT FINDS A field that wasn't there. {{ $json.customer.email }} where the input has no "customer". It resolves to nothing, the node runs, the field goes out blank, and the run is green. Probably the most common silent failure in n8n, and nothing else surfaces it. An error inside a successful response. The HTTP node is happy, it got a reply. The body says "rate limit exceeded". Expired tokens, throttling and validation failures all arrive this way. Items lost partway through. 50 rows in, 47 out. Three records dropped and nobody noticed. A step that produced nothing, and which exact upstream field was empty, including keys buried inside a JSON string. Output that changed shape. A score that was a number and is now the text "high". A summary that ran 140 characters and now runs 30. Faults inside sub-workflows. It says so, and drills in, instead of pointing at the parent. How long it has been happening, across scans and across sessions. For workflows with AI nodes, it also separates the three cases that need different fixes: bad context (a tool errored, returned nothing, or came back about the wrong record), a wrong decision (the answer contradicts the data it was given), or an invented detail (an id, phone number or email domain that appears in nothing it was handed). And it lays out what the agent actually had to work with, which tools ran, and what they returned. WHILE YOU ARE BUILDING Open it on a workflow, press "Compare each run", and leave it open. Every time you hit Execute it tells you what changed since the run before: run #418 vs #417 tier changed "enterprise" to "mid-market" confidence type number to string "high" summary shorter 144 chars to 34 chars: "I'm sorry, I can't help..." Or, just as usefully: "Output identical to the previous run", meaning your change did nothing. n8n has no memory of your last run, so that comparison normally happens in your head. IT TRIES HARD NOT TO CRY WOLF A tool that flags healthy runs gets uninstalled on day two, so it is built around not doing that. A Filter that legitimately kept nothing, a branch that was not taken, a poll that found no new rows, a Code node that mints an order reference, an expression whose author already handled the field being missing. All reported as the workflow working, not failing. It refuses to judge output shape until a node has at least 8 runs of history. And if it does get something wrong, every finding has a dismiss button that silences it permanently in one click, plus a "certain only" switch that hides everything inferred from history and leaves only plain facts about the run. SETUP There isn't any. No account, no API key, no SDK, no self-hosting. It reads the n8n you are already signed into. Works with both n8n Cloud and self-hosted n8n. PRIVACY No server. No analytics. No account. Nothing is ever transmitted anywhere. Execution data is read into memory, analysed, and discarded. What is stored locally is your settings, your ignore list, a record of which faults were seen and when, and statistics describing what your nodes' output normally looks like. That is it, and it is deleted when you uninstall. Open source under the MIT licence, so you can read exactly what it does before you run it. HONEST LIMITS Reads the most recent 40 executions per scan by default. "Look back further" goes to 100 or 250. Output shape checks need 8 or more prior runs of a node, or a learned baseline for that node type. The three way AI diagnosis and invented detail checks only apply to nodes with a model attached. It can find an answer that contradicts its own source. It cannot tell you an answer is wrong when the data was right. That needs to know the correct answer, and nothing does.