question :
how does the agent harness works ?
Is T3 just a control surface/UI over Claude Code, Codex, , Grok, etc., or do you also run your own harness/tools on top of the provider?
Want to know if the agent loop still belongs 100% to the provider, or if T3 changes how the work actually happens.
Is it allowed to use Claude oAuth subscription or i must use API ?
@NousResearch@Teknium did something happen to hermes desktop in mac ?
suddenly cant update, claude did debug and is blaming that it is your bug :D
so just wanted to make sure , this is what claude says :
Heads-up: commit 6d4e851d8 (Aug 27, "fix(serve): bounded flush-on-SIGTERM") breaks Hermes Desktop boot on macOS — every backend spawn gets killed after 90s with "Timed out waiting for Hermes backend port announcement" even though the backend is healthy. Root cause + fix below 👇
Reply / detail block:
Root cause: tui_gateway/server.py line ~401 does a module-level sys.stdout = sys.stderr (JSON-RPC stdout guard). That was fine while only the gateway imported it. 6d4e851d8 added from tui_gateway.server import install_exit_flush_signal_handlers into hermes_cli/web_server.py, so the swap now runs inside every hermes serve — before the port announce.
Result: HERMES_BACKEND_READY port=N prints to stderr, but the Electron waiter (apps/desktop/electron/backend-ready.ts) only watches stdout → 90s timeout → healthy backend SIGTERMed → infinite boot loop. Deceptive to debug because desktop.log echoes both streams, so the READY line appears in the log while the waiter starves.
Repro: python -m hermes_cli.main serve --host 127.0.0.1 --port 0 >out.txt 2>err.txt — READY lands in err.txt, out.txt is empty.
Fix that works for me locally: print the machine-parsed sentinels (HERMES_BACKEND_READY, BACKEND_PORT_IN_USE) with file=sys.__stdout__ in web_server.py (or capture real stdout before the tui_gateway import). Happy to share logs — macOS 15 / arm64, git install, v0.20.5 @ a588685fb.
מודל מסתורי בשם Ox Alpha הופיע ב-OpenRouter — חינם, קונטקסט מיליון טוקנים, וכמעט 4 טריליון טוקנים בשישה ימים. היום נחשף: ZAI הסינית, והמודל הוא GLM-5.3-Flash. הפרטים בעמוד
אנתרופיק מאחדת את הזיכרון של Claude: אותו זיכרון בצ'אט ובמצב העבודה בענן, נכתב תוך כדי השיחה, וכל מה שנשמר שקוף — אפשר לקרוא, לערוך או למחוק. נושאים רגישים