Connector Approval Failure — Findings + Pending Notion Edit
Date: 2026-07-26 Owner: Arden (Builder — tools, diagnostics, infrastructure) Status: Text is final. Two Notion pastes pending — blocked from the cloud session, needs a local session.
What Arden needs to do
Replace the section titled "A note on connectors and reliability" — in full — on two Notion pages. The replacement text is below, already finished and voice-checked. No drafting required.
| Page | URL |
|---|---|
| Setting Up Claude for a Google Workspace Professional Services Firm | https://app.notion.com/p/3a4a879ece9d8107865fddfe1ce1ca3d |
| Setting Up Claude for a Microsoft-Based Professional Services Firm | https://app.notion.com/p/3a4a879ece9d81a7991bd788030b9956 |
Two ways to land it:
- Paste manually. Open each page, select the existing section, paste the replacement.
- Let a local session write it.
updatepagewithcommand: updatecontent, using the text below asnew_str. If it fails with-32003, just try again — retries often go through.
Also pending: the KHB Operations action entry for this task. It could not be created from the cloud session. Title: "Paste connector-reliability section into both firm setup guides." Ball With: Arden. Owner: Arden. Workstream: Infrastructure. Type: Task. Priority: Medium. Repo Link: this file.
What was actually observed
All from a single Claude Code cloud session on 2026-07-26, in order:
| Call | Result |
|---|---|
notion-search | ✅ |
notion-fetch (page body, board schema) | ✅ |
notion-update-page ×3 (incl. a nine-word edit) | ❌ -32003 |
notion-create-pages | ❌ -32003 |
Gmail create_draft ×3 | ❌ -32003 |
Calendar list_calendars | ✅ |
Calendar create_event | ✅ |
Calendar delete_event ×2 | ❌ -32003 |
The conclusion is that there is no clean pattern. createevent succeeded and deleteevent failed on the same connector seconds apart, both writes. That rules out every tidy explanation:
- Not reads-work/writes-fail — a write succeeded.
- Not one-broken-connector — Calendar both worked and failed.
- Not session-wide — something succeeded.
What's left: individual calls intermittently fail to complete their approval step. Approval prompts do reach the user — Kathryn saw them arrive both before and after a failed attempt — and clicking them does not reliably rescue the call that spawned them.
Retrying is worth doing. The Calendar write succeeded on its first attempt after six consecutive failures on other connectors. Declaring a connector dead after one failure is wrong.
Two claims in earlier versions of this doc that were wrong
Recorded so the mistake isn't repeated, not for its own sake.
- "Usually affects one connector while the others keep working normally." Asserted before testing. Contradicted when Gmail and Notion failed together.
- "Reads pass, writes fail, across every connector." The correction to (1), also asserted from a small sample. Contradicted when
create_event— a write — succeeded.
Both were generalized from roughly six failures that happened to be writes and four successes that happened to be reads, then written into client-facing text with more confidence than the evidence carried. The client-facing section below now describes only the observed behavior and the practical response, and makes no mechanism claim.
Not verified: the fresh-session fix. It could not be tested from inside the affected session.
GOOGLE VERSION — replacement text
A note on connectors and reliability
Connectors — Gmail, Drive, Calendar, and third-party ones like Otter or Zoom — are reliable most of the time, but they can be finicky. Every so often an action comes back saying it needs approval and then simply fails. Sometimes no prompt appears for you to click. Sometimes one appears, you click it, and the action has already given up by then.
Don't confuse this with the approvals that are supposed to happen. Normally a connector asks, you click Allow, and the work proceeds. Here the approval and the action get out of step, and the action fails regardless of what you click.
There's no clean pattern to it, so don't spend energy looking for one. The same connector can work one minute and fail the next — saving a file can go through while the very next action fails. It isn't a permissions problem on your account and it isn't something you configured wrong.
What to do about it: try the action again. Retries often go through, and that's the cheapest fix by far. If a couple of retries don't clear it, start a fresh session, which reconnects the tools cleanly. If it's still stuck and the work matters, do that piece in the desktop app and move on — it's rarely worth troubleshooting in the moment. What doesn't help is re-approving the same request over and over. If a connector fails this way persistently across days, raise it with whoever administers your Google Workspace, since at that point it may genuinely be a permissions setting on the admin side.
The reason this is worth knowing up front: don't build a routine that depends on a connector running unattended. For anything time-critical — a scheduled send, a recurring calendar action — have a person confirm it actually went through.
MICROSOFT VERSION — replacement text
A note on connectors and reliability
Connectors — Outlook, SharePoint, OneDrive, Calendar, and third-party ones like Otter or Zoom — are reliable most of the time, but they can be finicky. Every so often an action comes back saying it needs approval and then simply fails. Sometimes no prompt appears for you to click. Sometimes one appears, you click it, and the action has already given up by then.
Don't confuse this with the approvals that are supposed to happen. Normally a connector asks, you click Allow, and the work proceeds. Here the approval and the action get out of step, and the action fails regardless of what you click.
There's no clean pattern to it, so don't spend energy looking for one. The same connector can work one minute and fail the next — saving a file can go through while the very next action fails. It isn't a permissions problem on your account and it isn't something you configured wrong.
What to do about it: try the action again. Retries often go through, and that's the cheapest fix by far. If a couple of retries don't clear it, start a fresh session, which reconnects the tools cleanly. If it's still stuck and the work matters, do that piece in the desktop app and move on — it's rarely worth troubleshooting in the moment. What doesn't help is re-approving the same request over and over. If a connector fails this way persistently across days, raise it with whoever administers your Microsoft environment, since at that point it may genuinely be a permissions setting on the admin side.
The reason this is worth knowing up front: don't build a routine that depends on a connector running unattended. For anything time-critical — a scheduled send, a recurring calendar action — have a person confirm it actually went through.
Microsoft deltas from Google
- Opening line: Outlook, SharePoint, OneDrive, Calendar replaces Gmail, Drive, Calendar.
- Routing line: whoever administers your Microsoft environment replaces whoever administers your Google Workspace.
- Everything else identical.
What changed from the section currently live on the pages
The live version says the failure is "a known issue on Anthropic's side" and to start a fresh session. Three changes:
- Retry is now the first move, ahead of starting a fresh session. It's cheaper and it demonstrably works.
- The two ways it presents — no prompt at all, or a prompt that arrives too late — so people recognize both.
- An explicit "there's no clean pattern", which saves a firm from hunting for a rule that isn't there, and an escalation path for when it persists.
No mechanism claim. Earlier drafts had one; it didn't survive testing.