Start with one saved execution
An Agent can show an empty observation even when a tool node appears to have produced output. Before changing a production build, inspect what the saved execution actually contains.
This script follows the saved Agent action reference to the exact recorded tool run. It reports indexes, output item counts and whether an error is present. It leaves out node names, call IDs, inputs, returned values and error text.
It reads a local JSON file and makes no network request. It does not run a workflow, change n8n or require a Keptick account.
Run the local check
- Save one execution response from your authorised n8n public API, with
includeData=true. This script expectsdata.resultData.runDatain the saved object. A list response or an export with another structure will not work. - Find the exact Agent node name and tool-call ID locally. Keep the original export on your computer.
- Download the script, put it beside the JSON file, and run the command below with Node.js 22 or later. Replace both quoted placeholders with the values from that execution.
node n8n-saved-tool-reference.mjs execution.json "AI Agent" "call_REPLACE_LOCALLY"The output is a small JSON summary. The original execution file is not modified. Invalid JSON or an unsupported input shape returns a generic error and a nonzero exit status.
Read the result carefully
Fictional result: one saved action points to tool run index 1, which contains one item in its main output.
{
"matchedSavedReferences": 1,
"rows": [{
"agentRunIndex": 0,
"referenceIndex": 0,
"referencedToolRunIndex": 1,
"actionItemIndex": 0,
"referencedToolRunPresent": true,
"mainItemCounts": [1],
"aiToolItemCounts": null,
"errorPresent": false
}]
}- Indexes start at zero. Tool run index 1 means the second recorded run. The script does not substitute the first run.
- Null means unavailable. A missing output array or recorded item index stays
null; it is not inferred as zero. - Zero matched references means no matching saved reference was found. Metadata can retain only the latest tool batch. Missing history does not prove that n8n filtered or lost a response.
- Saved nonempty output is an after-execution snapshot. It does not establish what the Agent could read at the moment it built the observation.
- Error presence is only a flag. The script does not include error text or decide whether the business work succeeded.
Compare a failing and a passing call
Run the check once for a failing call and once for a passing call. Review the small summaries before sharing them with whoever is investigating. Include the n8n version and describe the symptom separately. The full export stays local.
A missing reference, missing tool run or missing item index is useful to report as unavailable. Do not replace it with an assumption. For a timing or filtering diagnosis, an investigation may still need a controlled reproduction or instrumentation at the actual read boundary.
The existing N8N_LOG_LEVEL=debug setting enables logging that n8n already implements. It is not a promise that these action and observation fields will appear in the logs.
Record execution order before testing a timing theory
Discussion update, reviewed 10 November 2026. The current issue discussion separates the legacy v0 scheduling report from an intermittent v1 empty-observation report. A proposed change for one path does not establish that the other report is resolved.
- Record the configuration with the evidence. Note the saved workflow's
executionOrder, the n8n image version, and the task-runner version if it runs separately. Keep the versions used during the original incident separate from today's versions. Record missing information as unavailable. - Repeat the original trigger in a controlled test. The discussion asks whether the reporter's latency-added tool call still fails on 2.43.7. If timing was part of your reproduction, compare that case with an ordinary call in a separate test workflow using fictional data. Do not add latency or extra external calls to production just to investigate.
- Keep observation and reproduction results separate. Record attempts, empty observations, versions and the small saved-reference summaries for each case. A period without failures in normal traffic does not show that the original failure-triggering test now passes. A historical image/runner mismatch is a hypothesis to investigate, not an established cause.
These are investigation steps suggested by the public discussion. No workflow, task runner or proposed fix was tested for this update. The downloaded diagnostic and its n8n 2.34.5 testing limits are unchanged.
Check the test copy against the original workflow
Source update, reviewed 11 November 2026. A separate report for v0 workflows describes repeatable empty observations and explicitly distinguishes the intermittent v1 report. In a follow-up about a test harness, its author reports that cloning a workflow had filled an omitted execution-order setting with v1. That can make a comparison exercise a different configuration.
- Compare the saved settings before comparing results. Inspect the original workflow export and the saved test copy. Record whether
settings.executionOrderis absent,v0orv1. If a copy or test tool supplies a value, record that transformation. Treat the three exported states separately; do not silently fill a missing setting in the evidence. - Keep a faithful control. Preserve the original configuration in a separate test copy using fictional data and safe test destinations. If you deliberately change execution order, label it as another test case and compare it with the control. A passing case after that change does not establish that the original configuration is fixed.
- Check execution evidence as well as returned items. Record the Agent observation and the tool-run reference for each case. Where the saved execution includes execution indices or timing, retain those locally too. The harness author's branch control returned identical items while the branch execution indices changed; passing item checks alone did not establish the same execution order.
The report and harness results above are attributed to their authors. Keptick did not run those workflows or the harness, verify its release, or test an upstream fix. The offline script still reports only its documented saved-reference fields; it does not compare workflow settings or diagnose scheduling.
Use monitoring for expected work
This diagnostic investigates a saved execution. For scheduled workflows, Keptick can separately store outcome receipts and flag missing runs or counts below a chosen threshold. It does not trace Agent internals or fix an empty observation.
If that is the problem you need to monitor, read the monitoring setup guide. The first monitor can stay free within its allowance. For batch coverage, the record-reconciliation workflow compares expected and completed IDs.
Scope, tests and source
Checked against the saved-reference paths in n8n 2.34.5. Ten synthetic function checks and two separate command-line smoke checks passed on the exact download. These checks do not reproduce a reporter workflow, prove the root cause or establish a fix. Other versions and export formats may differ.
- Original diagnostic and discussion ↗
The request for investigation without patching production, the proposed script and its limits.
- Exact tool node and run lookup in n8n 2.34.5 ↗
Immutable release source; recorded output and runtime observation are different evidence.
- n8n public API documentation ↗
Use your existing authorised access. Never put an API key or full execution export into a public comment.
Made by Keptick with AI assistance. Not affiliated with n8n. This website counts guide visits and download clicks; the downloaded script sends no data.
Check one call at a time.
Keep the original export local. Start with a failing call and a passing call, then review only the small summaries.