Release v7.96.1

A pool no longer fails with an unexplained "Claude Code's model request failed.": the vendor's failure code and HTTP status are kept, and a member whose token, plan or model was refused no longer blocks the rest of the pool. (Cratis/StudioIssues#944)

Changed

• A provider failure that reaches a pooled chat call names the provider, its vendor and the model it was asked for, so "No AI provider could serve the Modeling agent" says which member failed and why.
• A Claude Code member that answers with an authentication, billing or invalid-request failure no longer ends the whole pool: the call moves on to the next member and, when none can serve it, the first refusal is the answer.

Fixed

• The Claude Code transport reported every API failure that was not a rate limit as "Claude Code's model request failed." It now reads the CLI's failure code (authentication_failed, billing_error, invalid_request, server_error, or an unrecognized identifier-shaped code) and the HTTP status of the retries before it gave up, and says what to do. Only the code and the status are used - never the vendor's own message text.
• A Claude Code server_error is treated as transient, so another pool member can serve the call.