Failures

Overstated sales. Early digests used the connector’s net and ran high until the owner caught it. Every digest before the fix was overstated. The tax rule now lives in a shared note every agent reads.

Unverified tax rates. A store’s tax rate is labeled “unverified” until it is proven from order-level data.

Order counts differ by method. A business-date pull and a modified-time pull disagreed, so some figures are labeled approximate.

Redundant recaps. Toast already emails sales totals, so the daily note now covers only reasons and exceptions.

No voids or comps in the connector. It needs a signed-in web session. When the session expired, voids went out as a same-day follow-up.

Noise. The first versions reported daily. Escalate-only-on-threshold made the brief worth reading.

Hidden gaps. The connector has no item mix, covers or close report. They are marked unavailable rather than guessed.

The connector only gives store totals. Live pour cost needs item-level sales, which the read-only connector doesn’t expose. A small read-only “product mix” tool was planned for it; that change is waiting on the owner’s OK, so daily runs give store-level net sales only.

Pour specs are a mapping, not house specs. The recipe pours came from an agent’s mapping (standard pours, cans, draft sizes), not from confirmed house specs. Signature cocktails have no recipe yet.

Pack prices costed as single cans. A unit-cost check found canned drinks in the inventory system priced per pack instead of per can, and liquor buttons depleting the wrong pour size. Both would show false over-pour. The fixes were drafted, not applied, because they need an owner’s OK.

Two inventory systems. A separate weekly inventory-ledger routine did overlapping work and stalled when several browsers were connected and it couldn’t pick one. It was merged into PourCost on 10/23 so there is one system.

Second location has no cost source. The other concept has no recipe or bottle-cost data, so it can’t be costed yet.

Missing is not zero. Early on it was easy to let a missing cost read as $0. The app now shows ”—” with the reason instead.

Portal sign-ins were the whole fight. Sessions expired and every distributor landed on a login wall. The agent correctly refused to guess passwords, bypass verification, or retry more than twice, so zero invoices came down on those runs.

The routine was paused after repeated auth failures; the last attempt was 2026-09-23 and downloaded nothing. Resuming would only create a failed-portal alert every morning.

The fix needs a human. Sign in once with “remember me” checked, then the agent checks the sessions read-only before resuming. The agent can’t do this itself, by design.

Reporting was rerouted mid-build. It first emailed the owner directly, then moved to one front-door agent that bundles everything into a single brief.

Login ownership. One portal uses a legacy login on a separate address, and one distributor login is tied to an individual’s personal email rather than a shared company address. Check who owns each login before you automate.

Empty price log. The first successful run only seeds baselines. Price moves appear from the second pull.

Order guide published with no login (Oct 29). The order guide was reachable without signing in. Fix: switched to invite-only the same day. Staff now need an invite.

Three starts in one night (Jul 2). The same marketplace was started three times in one night on two platforms. Fix: kept the Base44 version and parked the rest.

Duplicate checklist apps (Jun 24). Two duplicate checklist apps were started in the same half hour, three shift apps in all. Fix: kept one, a launcher.

An abandoned rebuild (Jun 26). A from-scratch rebuild was started 3.5 hours after the first version, abandoned after 1 h 58 m, and renamed “bad app”. Fix: went back to the original.

No local image upload in the connector. Postiz’s connector creates text-only drafts. Images were attached afterwards by hand in the Postiz web UI, one post at a time.

No public file host route. The connector’s URL upload accepts only public URLs. Uploading to anonymous public file hosts was blocked by the safety review, correctly, so that route was dropped.

UTC and daylight-saving error. Post times are stored in UTC. One draft was created at 22:00 UTC instead of 17:00 UTC (noon local) after the clocks changed. It was deleted and recreated. Rule: always compute UTC from local time with the DST date in mind.

Text-only list previews. The list view shows only text, so attached images could not be confirmed through the connector. The report from whoever attached them was relied on.

One caption typo still open. A hashtag typo is in one draft. Fixing it means recreating the draft.

Missing assets. Many rows name an image the script can’t find (26 of 48 Ready rows by a filename check). The script skips these, which is safe, but those posts wait until someone supplies the file.

Story captions. Story rows have no caption and are all In Progress. The story path is untested end to end, and the story duplicate check looks only at the date, so two different stories on one day would count as one.

Flagged rows marked Ready. Three rows are Ready and flagged. The script skips them correctly.

Loose filename matching. Images are found by file name alone, ignoring folders and case.

Same-day cutoff. A row for today is skipped if its draft time has passed.

Channel reconnect: unconfirmed. The project’s Instagram account was added to Postiz as its own channel by the owner, separate from other projects’ channels. Whether a social account split needed a reconnect is not known, so nothing is claimed either way.

Archived versions still published (Oct 29). Two archived versions were still published, one with no login. Fix: both switched to invite-only the same day.

Carrier registration blocked every reply. A real test text reached Twilio, the webhook answered correctly, and Twilio sent the reply. The phone never got it: error 30034, meaning the number’s US A2P 10DLC campaign was not registered. Everything on our side worked; the carrier refused delivery. Lesson: register the messaging campaign before you build the bot.

The live bot was answering for a different property. The concierge that was actually running sat on another machine behind a temporary tunnel, and it still had demo data for a different unit. It gave that unit’s details, not the pilot unit’s. It could not be patched from where the new facts lived. Lesson: know exactly which machine serves your live endpoint, and keep demo data out of anything a real guest can reach.

Temporary tunnels as production. Both the old and new webhooks ran behind quick tunnels whose URLs change on restart. A more durable setup was staged, but the cutover waits on the carrier fix.

Fact files are mostly blank. Most operational fields (entry details, vendors, pet policy) are still [FILL]. That is safe, because the bot escalates instead of inventing, but it means few questions get answered automatically yet.

Escalations are held in memory. Pending human approvals live in memory and would be lost on a restart. A persistent store is on the fix list.

Two channels, one brittle path. Guest messages (channel manager) and text messages (Twilio) share code that assumes one channel’s conversation IDs. They need separate end-to-end tests before cutover.