In a November 13 GitHub report, an OpenClaw user said that when policy excluded the configured primary model, requests went to another provider’s model despite an empty fallback list, exhausting subscription usage. An automated review identified a substitution path in the code but said the intended behavior still required a maintainer decision. It did not independently reproduce the reported usage loss by running the service.
A Codex user seeks the reason for a cleanup refusal
A November 12 Windows Codex report, followed by further accounts on November 13, describes user-authorized cleanup of task-created temporary files being rejected before execution in a Full access session. The message said “blocked by policy” without identifying the enforcing rule or a review route; separate Windows permission checks did not establish which Codex policy layer had refused the operation. The report asks for an explanation that would let a user understand the refusal.
Claude Code user says a task answer became a rating
On November 13, a macOS Claude Code user reported typing 1 to answer the first numbered task question and having it treated as a low feedback rating instead. The user said the intended rating would have been 3, but they were answering the task rather than the feedback prompt. As of a November 14 check, the public issue contained no explanation from Anthropic of how the input was routed.
Microsoft 365 connection problems reach two AI services
Anthropic’s November 12–13 incident log describes failed mail, calendar and search requests through its Microsoft 365 connection, while OpenAI recorded problems with new Microsoft connections and tool use. Both developers posted recovery updates on November 13, although Anthropic warned of possible lingering errors while Microsoft recovered. These logs describe the AI services’ connections; Microsoft’s underlying cause and the outage’s full scope remain unverified here.
In a November 12 release filed with the SEC, Fermi announced agreements with an unnamed large cloud customer to develop an initial 600 megawatts of power capacity at Project Matador, with supply planned to begin in 2027. It said the definitive long-term lease was still being finalized. The capacity describes planned power development, not computing capacity already delivered to a tenant.