[Daniel-rmc]

ManiLoop can now obtain its own ChatGPT plan authorization and call the official Responses API directly, while keeping prompts, images, history, action validation and the robot loop in ManiLoop. Existing API-key and Codex CLI connections remain available as separate routes.

• Add independent PKCE/OIDC sign-in, protected local account storage and refresh, account/model selection, and workbench/CLI integration.
• Share canonical policy inputs with the Responses API adapter; record input hashes, effective request settings and incomplete usage without claiming identical server-side token budgets.
• Handle observed SSE responses without a Content-Type header and explicit keepalive events while retaining strict completion validation and cancellation protections.
• Add LIBERO-only tcp_target_servo_v3: broad pose-goal guards of 2.5 m / π rad and up to 200 control steps (10 s) for movement, with early completion. Gripper commands retain 20 steps. The pinned OSC scale, 20 Hz control clock, 2 ms physics timestep, VLA behavior and original v2 mode remain unchanged. The broad guard is not a reachable-workspace guarantee.
• Document both connections and control protocols, and keep their configurations in separate comparison groups.

Validation: 545 Python tests and 136 subtests passed locally (62 optional checks skipped); 25 Node tests passed. Real ChatGPT-plan text, vision, structured-action and native Panda movement checks passed. Deterministic LIBERO checks reached a 12 cm translation goal in 39 steps (1.42 mm position residual) and a 0.8 rad rotation goal in 21 steps (0.00498 rad rotation residual). Closed-loop LIBERO transport was exercised, but these changes do not establish autonomous task success or a benchmark success rate.

No credentials, local runtime assets, private development notes or generated experiment records are included.