mirror of
https://github.com/paperclipai/paperclip.git
synced 2026-11-13 07:12:00 +01:00
DottaandClaude Opus 5.5
fa50b8d662
The fixture saved a built-in fake sandbox as its unsupported control. The product has refused to save fake sandboxes since #4449 ("reserved for internal probes"), so setup returned 422 before any agent run, and the test API withheld the reason. No saveable computer is advertised as unable to run Claude Code, so the case now contrasts advertised support with runtime health: the fixture saves an SSH computer on a reserved .invalid host that is advertised for Claude Code but never tested. The yaml context records this premise change (manifest changedFields pinned). The grader no longer reads reply keywords. It requires the agent boundary (403 on capabilities and probe), a rendered reply link to the environment settings page that the person then clicks to see both computers and the checked capabilities, no board-only SSH settings in the reply, unchanged targets and worker, and exactly one chat turn. Cookie-session API failures with 400/409/422 now report the product's own validation message, bounded and with request values and the session redacted. Details and other statuses stay withheld. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>