The rollout of the GPT-5.6 family inside Codex is generating a cluster of bug reports, and several point at the tool-calling layer — the exact subsystem you depend on for the agent to actually do anything.
The sharpest is a high-priority report of a ToolName namespace collision. On gpt-5.6-sol
, the agent emits a tool call that collides with an existing one and surfaces as unsupported custom tool call: execexec
. The label says it is a CLI tool-calls bug. The practical effect: a capability the model reaches for gets rejected, so the task stalls or degrades to a weaker path. Namespace collisions like this are easy to introduce when a new model version adds tools whose names overlap with custom ones you registered — and they are annoying precisely because the model "thinks" it called something valid.
A second report: on the GPT-5.6 Codex Desktop, new tasks can start without workspace tools. If the file and shell tools are not attached at task start, the agent is effectively blind and handcuffed until something reconnects them — and the user may not realize why the first few turns went nowhere. The label ties it to Windows and the app surface.
The third is simpler but just as real: a user reports not having GPT-5.6 available on a personal cloud workstation despite expecting it. That is a rollout/entitlement gap — the model the marketing talks about is not the model your account can reach, which makes any "try the new build" instruction a dead end.
Three bugs, one theme: the new model family is shipping faster than the tool-wiring and entitlement plumbing around it. If you are on Codex and planning to lean on GPT-5.6, verify your workspace tools attach and your account actually has the model before you trust a session. The Claude Code data-loss writeup is a reminder that model-version bumps are exactly when agent reliability regresses — watch the tool layer first.