Quick overview
This workflow receives an agent’s final reply and tool execution log via webhook, checks whether the reply claims actions that aren’t supported by completed tool results, optionally posts gaps to Slack via an Incoming Webhook, and logs every check to Google Sheets.
How it works
- Receives a POST request on the
/webhook/say-do-check webhook containing an agent reply plus either a tool-call log or an n8n execution JSON with a node map.
- Applies your configured failure markers and node mapping to normalize tool-call results and infer whether each action actually finished.
- Extracts completed-action claims from the reply text (and optionally from a structured
actions array) and grades each claim against the tool-call evidence.
- Produces a verdict of
pass, review, or stop_ship and returns it to the caller along with findings, duplicates, gap counts, and a text report.
- If configured to notify (gaps, stop-ship only, or always), posts a formatted review message to Slack using the Slack Incoming Webhook URL.
- Appends a row to Google Sheets with the verdict, gap metrics, reply text, and report for audit and tracking.
Setup
- Set the Slack Incoming Webhook URL, Google Sheets spreadsheet ID, and target sheet tab name in the Config step.
- Connect Google Sheets OAuth credentials for the Google Sheets append action.
- Create the header row in your sheet tab:
timestamp, receivedAt, source, verdict, stopShip, claims, gaps, gapRate, reply, report.
- If you plan to send n8n execution JSON instead of a raw tool log, update the default node map in Config to match your workflow’s reply node and action nodes.
- Activate the workflow, copy the production webhook URL for
/say-do-check, and configure your agent/calling workflow to POST the reply plus tool log (or execution + node map) to it.
Requirements
- Slack workspace (Incoming Webhook) and Google Sheets. No AI key. No community nodes.
Customization
- Edit
notifyOn in Config: gaps (any unsupported claim), stop_ship (only pay or delete gaps), or always. Change defaultNodeMap to match your agent workflow (reply node name and field, plus each action node with action, id path, and target path or target_value). Edit failureMarkers to match your APIs; an empty list turns off body checks, while HTTP 4xx or 5xx and a failed status string still count. Swap Slack for email if you prefer.
Additional info
A buyer can put this in front of any send step and hold a message when the agent claimed "I refunded the order" or "all set" but the refund node timed out or never ran. The check is deterministic (verb, target, amount and evidence grade, including a hard stop-ship on pay or delete), needs no AI key, and returns a verdict the calling workflow can branch on.
Limitations:
- This checks a finished reply. It does not recall a message that already went out. Run it before you send.
- Targets are only as good as the tool log or node map. n8n often does not store a resolved recipient, so
target_value is sometimes a fixed string.
- Only the first output item of each node run is read.
- Replies with no completed action claims pass, including future tense such as "I will send it", unless a structured
actions list is present.
- The verb list still misses some phrasings. Version 1.2.0 covers the passive and noun cases listed above; unusual wording can still pass quietly unless you send a structured actions array.
- Anyone who can POST to the webhook can run a check. Keep the URL private or put the workflow behind authentication.
- An invalid payload (missing reply, or no calls and no execution) stops the run with an error before a webhook response is sent.
- If Slack or Sheets fails after the verdict has already been returned, the caller still has the JSON result but the notify or log step may be missing.
- No duplicate check across requests. The same payload posted twice is checked twice.
- Unusual API bodies may need the
failureMarkers list in Config edited. An empty list turns off body checks.