Repeated hard-stop "API Error: Sonnet 5 can't help with this. Start a new session to continue" errors

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

I'm building an internal team credential vault inside a Google Apps Script project — AES-encrypted passwords for billing staff, decrypt-on-demand via a reveal button that auto-hides after a timeout, with all access logged. This is a standard internal password-manager pattern, not credential harvesting.

The problem: mid-session, I repeatedly hit a hard-stop error — "API Error: Sonnet 5 can't help with this. Start a new session to continue" — with a link to the AUP page and no further explanation. Once a conversation touches the AES encryption and password-reveal logic together (from my observation), the error fires on nearly every subsequent message, making the session unusable. This has forced me to start over approximately 15 times.

This appears to be a false-positive safety block surfacing as an API error rather than a normal refusal. The code is legitimate, the use case is legitimate, and there's no actionable feedback — just a hard stop.

I confirmed this is session-specific, not content-specific: I opened a brand-new session and asked Claude to review the exact same code, and it completed without error. The long-running session continued to fail on later messages — including one that wasn't discussing the code at all, just prose about how to file this report — indicating the block is tied to something that accumulates within a session rather than any specific file or line of code.

What Should Happen?

Claude should be able to read, review, and help fix code for a legitimate internal tool — including encryption/decryption logic for a team credential vault — without hitting a hard-stop error. If there's a genuine concern about the content, it should show up as a normal response/explanation from Claude (something I can read and respond to), not an unexplained "API Error" that ends the session and forces me to start over. This shouldn't need to happen almost every message once the conversation has touched this topic once — and it definitely shouldn't fire on messages that aren't even about the flagged content.

Error Messages/Logs

req_011CdqV3UioH8wtb5EFHZCvU
req_011Cdt5yNyVZDFvdZtqL9x43
req_011CduRJih9c2Yd78TQja2do
req_011CduNj16CdkwRFpM9J5AV2
req_011CduhCh3shs7BUcLNUvV26
req_011CduG7ytbJ3CAjDFKjgsBE
req_011CdtDcR75VA9ytJntQk8oX
req_011CdweFMUPNW22PD1NLmmf4
req_011CdwfyXShgjw2BBvNrNy86
req_011Cdwggg2XYXfA9xXdEuddn
req_011Cdwh6cba8TGDkG29CS9SN
req_011CdwinK4rdfk8dR6cu2ydR
req_011Cdxsrue95B1SKJ9VS5YS5
req_011CdxwS9cYKYKPQSgsyzy5L

Steps to Reproduce

1. In an ongoing session, ask Claude to review AES encryption/decryption code and a "reveal password" UI (auto-hides after ~20s, logs each reveal) across several files, and continue the conversation across many exchanges discussing that code.

1. Observe a hard-stop error: "API Error: Sonnet 5 can't help with this. Start a new session to continue," linking to the AUP page, with a Request ID — no refusal message or explanation from Claude first.

1. Open a separate, brand-new session and ask Claude to review the same underlying code (same file, same functions) with no prior history attached.

1. Observe this completes normally — no error.

1. Return to the original long-running session and continue the conversation, including messages that don't reference the flagged code at all (e.g., plain prose discussing this bug report). Observe the error recurs anyway.

Note: This rules out any single file or code snippet as the trigger — identical content read cleanly in a fresh session. The block appears tied to something that accumulates within a specific session once it's been touched (possibly including the session's own prior AUP-error messages, which then compound). Once triggered, it blocks unrelated follow-up messages too, not just ones referencing the original code. I can't share the actual files publicly since they contain proprietary business logic, but can provide them privately if it would help investigate.

Claude Model

Other

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

2.1.228

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

VS Code integrated terminal

Additional Information

_No response_