[Xiiiing]

V0.7.1 restarted its Python/Codex reader for individual requests and layered polling and duplicate UI state onto interactive work. V0.8.0 keeps one installed Codex App Server under the Rust Agent, delivers authorized traffic over an outbound persistent connection, and shows cached history immediately while reconciling in the background.

• Replace Python runtime packaging with native Codex discovery/configuration, request IDs, bounded concurrent reads, shared session indexing and turn supervision. Keep Agent durable receipts/outbox, schema 7, single-session execution and unknown-result recovery.
• Add authenticated Agent WebSocket transport with connection replacement, save acknowledgements, ephemeral streaming and explicit service/Hub/Codex readiness. Keep HTTP compatibility for Agents being upgraded.
• Adopt Ant Design X, TanStack Query and Virtual for the Codex workspace. Preserve system navigation, fixed composer, per-session drafts, Markdown/KaTeX, tool groups and scrolling anchors. Clear bounded in-memory content and receipt state on logout.
• Update both deployment languages, package tests and third-party notices. Upgrade Hub before Agent; reuse the existing locally authenticated Codex executable.

Validation: make check, make test (95 Rust and 19 Console tests), make privacy, Console build and 59 browser checks passed; nine redundant browser matrix cases are intentionally skipped. Native protocol checks cover Codex 0.153.4 and 0.147.0. Release installation/rollback tests and a real isolated Hub → Linux Agent → browser loop cover reading, sending, queuing, interruption, notifications and reconnection. A V0.7.1 → V0.8.0 → V0.7.1 → V0.8.0 database roundtrip retains completed receipts and does not replay uncertain work.

Controlled browser benchmarks load all 100 pages of a 2000-turn history: 11 turns remain mounted, cached switching p95 is 92 ms, and 20 sends use 40 HTTP requests (previously about 80). Each simulated network segment has 80 ms RTT; these figures exclude model time. Private machine-specific evidence is intentionally kept outside Git.

The final real-host test found and fixed duplicate history reads when terminal events and command receipts arrived separately. Refreshes now deduplicate by operation/turn identity, with a regression covering both arrival orders. Under two added 80 ms RTT segments, 30 real native reads of the latest 20 turns reached 243 ms p95, 30 save acknowledgements reached 208 ms p95, and 20 sends used 42 total HTTP requests. An active turn also survived Hub reconnection before interruption, and its queued follow-up completed.