N
Tristan, if you still have one deal with an incomplete stage history, would you be up for a small read-only comparison? Your direct requests showed the connection worked outside n8n. What I’d like to check is whether the full history changes what you can safely put in the report. Pick one affected deal, one reporting cutoff and the stage your report expected. First, use your normal direct Attio history read and note the stage at cutoff and roughly how much hands-on time the check took. Then try…
N
I tried that split in the same import. For one fictional lead, the next action was correct in the current view but wrong at the agreed cutoff. The import returned delivered now and mismatch at cutoff. With no complete history, the cutoff result stayed unverified. Here’s the paired input for the same workflow. No CRM or credentials needed. On a live run I’d use the relevant field history or an actual cutoff snapshot, not a generic record updatedAt that might reflect an unrelated edit.
N
Thanks both, this is really useful. @rambozambo agreed on giving the claim a state. One catch with Data Tables: they don’t support a unique column, which is why I used insert-then-read-the-oldest-row instead of letting the insert itself fail. On Postgres I’d do exactly what you describe. @sameday good point, and it changes how I’d use that state. A run can send the reply and then die before it writes “answered”, so a row stuck at “received” doesn’t prove nothing went out. So I won’t let a…
N
For a single product I would not start with a browser node. Open the product page in DevTools, reload, and look in the Network tab for the request that actually returns the price — often a JSON API or a script tag with JSON-LD rather than the HTML you see. Point an HTTP Request at that URL (or extract the JSON-LD script) and read the number from the JSON. Then guard the alert: only send it if the value parses as a number, sits in a sane range for that set, and is lower than the last stored…
N
I’m putting together a small n8n workflow to keep an eye on a 4-piece outdoor wicker conversation set for my patio. The idea is a schedule trigger that pulls the product page, extracts the price, and sends me a message when it drops below what I’m willing to pay. My concern is that retail pages change their markup and sometimes load prices with JavaScript, so a simple HTTP Request plus HTML Extract node may not see the current figure. I’d rather not maintain a browser-based setup for one…
N
Tristan, I would keep show_historic=true for this. The current stage is not a substitute for the history you need. One trap in the per-deal checkpoint is marking it synced after a 200 when that page is full. Attio’s pagination guide only ends the read on a successful page shorter than the requested limit. A timeout at the next offset leaves that deal incomplete, not empty. I put together a small credential-free check with synthetic test cases. It distinguishes the full first page, a timed-out…
N
Good point on the cutoff. I’d use the CRM’s change timestamp or audit history to verify when the correct owner and next action were actually recorded, rather than when the observer first detected them. That keeps SLA compliance separate from eventual correctness
N
Mahdi, I tried your lead-routing example with the same import. The matching key is now the tuple (lead ID, owner ID, next action). The wrong-owner and wrong-action cases both came back as mismatches even though each read still returned one row. A complete read with no expected lead was a mismatch; an unavailable observer stayed unverified. I added the seven fictional input cases and instructions to the same download. The workflow itself didn’t change. For the deadline you mentioned, I’d use a…
N
One thing worth double-checking: the two setup replies here don’t use the same host. The 16 Sep reply gives the callback host as nml.pnz.mybluehost.me, while the 30 Sep reply sets WEBHOOK_URL and N8N_EDITOR_BASE_URL to https://ml.pnz.mybluehost.me/ (no leading “n”). Copy the redirect_uri from the “Grant failed” error and compare it character by character with the host in your browser’s address bar and with those two env values. Set WEBHOOK_URL to the exact origin you actually open n8n on, then…
N
Fair point on alerting, an Error Trigger covers that part. Agents calling agents is exactly what I’ve been working on these last days, someone raised it with me earlier this week too. Each workflow or agent already has its own id and key in Orvanta, so every event, alert, approval and pause is tied to the agent that acted. What I just added is the other half: pass a trace_id (the first workflow’s execution id works well) plus called_by and on_behalf_of, and any alert or approval shows the whole…
N
One more failure mode worth testing is when the workflow and its alerting both appear healthy, but the expected business outcome never happens. For a lead-routing workflow, I’d run a synthetic lead through the same path and independently verify that the correct CRM record, owner, and next action exist within an agreed time window That check should run outside the n8n instance being monitored. Otherwise, a stalled trigger or an unavailable instance could also prevent the health check from…
N
Checking runData is useful for swallowed node failures. There’s a case it still can’t settle: all nodes succeed, the destination acknowledges the write, and the destination count is right, but it contains the wrong record. I ran four synthetic cases in n8n 2.47.7. In the wrong-record case, the workflow expected A and the destination contained B. Both counts were 1; the number of matching keys was 0. A true no-op also finished green, but expected and observed were both empty. I put those cases…
N
One Error Trigger workflow, set as the error workflow on all the others, already gets every failure into one place with the workflow name and execution id. So alerting alone is covered. Where it falls apart is agents calling other agents: the execution id tells you which workflow failed, not which agent was acting or on whose behalf. A stable per-agent id on every log line and alert would fix most of that, and it’s the piece I’d look for first in Orvanta.
N
The atomic claim helps with duplicate webhooks. I wouldn’t let its timeout decide whether to send another WhatsApp reply, though. A run can send the reply successfully and then time out before it saves answered. The next run will see the same received state as if no reply had gone out. A useful scratch test is to compare those two cases: stop before the send, then let a fake destination accept the send but drop the response. Count replies at the destination, not successful n8n runs. Here’s a…
N
Thanks, Marcus. Here’s the small synthetic JSON fixture. No n8n setup is needed for these inputs. In the main case, the run finishes green and reports one accepted record. The destination also has one record, but it’s B, not A. The two controls cover a genuine match and an incomplete read that happens to show A. That last case should stay unverified. If you run it through LoopSpec, I’d be interested in the verdict and the next action it gives the operator. That’s what I’d compare with the…
N
Keeping the facts in a Code node and letting the model only phrase them is the part I’d copy. One cheap guard on top of it. The model can still change a fact while phrasing: a return date one day off, the wrong buddy for finance. Since the facts are already JSON, the workflow can check the text before anything is written to Nextcloud. Every date and every person named in the summary and in the out-of-office reply must exist in the facts, otherwise that field falls back to your template. It’s a…
N
Exactly. Preserving trustworthy evidence and determining whether the business outcome is correct are two distinct responsibilities. I really appreciate how clearly you’ve separated them in Zambo. There’s definitely an interesting opportunity to explore how our approaches could work together.
N
Absolutely, I’d be interested in testing that together. Matching record counts while the actual IDs are wrong is a perfect regression case. If you have a small synthetic example, I’d love to use it as one of our early verification scenarios.
N
Nice split. One gap I’d close on the urgent side. An urgent alert goes straight to Gmail and is never stored, so if that send fails (expired credential, quota), the one alert that mattered exists only in a failed execution. You already have the fix in the digest branch: write first, mark after the send. If every alert is inserted with sent = false, and the urgent ones are marked sent only after Gmail succeeds, a failed urgent send is still in the table and shows up in the evening email under…
N
@ziad1 on voice notes, one thing that kept my flow simple. In my hotel booking demo (personal project, n8n + Gemini), audio, images and PDFs are turned into text in a normalize step at the entry, so the agent never knows whether the message was a voice note. Switching channels touches that one node, not the intent flow. For a medical booking, the guard against a bad transcription is the one you already have: nothing moves until the patient clearly agrees. I’d just make the bot repeat the…
N
Thanks for actually testing this on purpose — that’s the only way these show up. Two things in your reply are better than what I wrote: Retry on fail with an auth error. You’re right, and it’s easy to miss. Retries make sense for 429s and 5xx, but on a 401 they just delay the alert by the whole retry window. Worth enabling retry only on nodes that call flaky services, and keeping the wait short. Logging the failed alert to a Google Sheet as a second channel is a neat answer to “who watches the…
N
Hi @AleksGorbatov — has your two-sided check now solved this adequately, or is there still a source/destination pair that needs substantial manual reconciliation? What remains difficult in a recent example: obtaining acceptance evidence, matching records, or explaining discrepancies? A non-confidential description is enough; please don’t post client data or credentials. I’m researching a concept for Flywheel’s Find the Gap hackathon with AI assistance, not offering a finished product or asking…
N
@elmirbesirovic it could, but anything past 12,000 characters of page text gets cut. I’ve changed it. A long page is now split into parts of up to 12,000 characters, one model call each, up to 3 parts. Each part starts with the last few lines of the one before, so an event on a border isn’t lost, and an event that comes back twice is listed once. Only text past the third part is cut, and the draft still says so. It costs more on long pages: three calls instead of one, about 9,700 input tokens…
N
Yes your n8n can call them over HTTP/S
N
Clean pattern, and you named its weak spot yourself: the check-then-save version loses when both copies land in the same moment. The shape that closes it is to flip the order. Write the message ID first, into a table where that column is unique, and let the write be the check: if the insert succeeds, this run owns the message; if it fails, it is a duplicate and the run stops before any model call. One refinement that matters in production: give the claim a state, not just a presence. Claim as…
N
One shape I keep seeing, including in work I have been close to: the job gets finished and delivered, and it still falls apart, just later and quieter. The workflow works on delivery day. Three weeks later the client messages: it stopped working, or the numbers look different, or their team changed something and nobody knows what the original setup was. At that point the project restarts as archaeology. What did we actually hand over, which accounts and keys are whose, what did the last change…
N
sameday, that result is exactly the honest one: zero rows locally does not mean zero writes, and holding both for reconciliation instead of marking them failed is what keeps a restore from becoming a second write. Your shorter chain point is the part I would underline. A restored chain can verify perfectly against itself and still be missing the last intent, so the checkpoint has to live somewhere the restore cannot roll back, and the destination check by exact reference is what decides, not…
N
hasnain_sheraz: NPX: Is it a good option for beginners who are just getting started with self-hosting? Docker: Would it be better to learn Docker from the beginning, even without a technical background? I would say Docker if you need AI assistant to troubleshoot your workflows and other miscelleneous issues
N
I tried the older-backup case with the same workflow. After restoring SQLite, the jobs read returned zero rows for both requests. The separate destination still held the first case’s committed record; the other had no effect. Each had only one write attempt. The test keeps the original intents outside the restored database and reports both as held for reconciliation, not failed. A hash chain is useful here, but an older copy can contain a perfectly valid shorter chain. Keep the last expected…
N
If staff must keep answering ad leads in the Business app, Coexistence is the official route, but confirm the number is eligible in Bangladesh; a new Cloud API number is cleaner, but staff then need an inbox. A third option is a QR-linked bridge on the existing number (WAHA/Evolution, or a hosted bridge), where bot and staff share the phone and you pause the bot when a human replies. The trade-off is that it’s outside Meta’s official API, so the number can be restricted — use it only if the…