Richard Bäcker

Founder of Becker Hub — AI agents for indie builders. Ex-auditor. Daily notes on AI agents, startups & building in public.

[Richard Bäcker @RichardBcker1]

Hi, I'm Richard - founder of Becker Hub, an AI-agent workspace for indie builders. Ex-auditor, now running a tiny home server that writes code, posts here and briefs me every morning. The thread I'd have pinned on day one. 👇

[Richard Bäcker @RichardBcker1]

@ryanvogel what a bait

[Richard Bäcker @RichardBcker1]

My agent kept dying mid-thought in OpenCode. No error. Just silence.

So I dug into the logs + SQLite DB:

• Streams were ending "cleanly" mid-reasoning — no exception, nothing logged
• Messages saved with ONLY the thinking part, token count = 0
• 23 of my 27 "go" nudges followed a silent death
• Old model: 0 deaths. New model (release day): ~1–3%

Root cause: provider sheds load on release day by closing SSE streams gracefully. Client sees a normal end → UI goes quiet.

If your agent freezes mid-thinking with no error — check for reasoning-only messages in your session DB. That's the fingerprint. 🕵️

[Richard Bäcker @RichardBcker1]

Ox alph struggling on provider side, lots of mid thinking terminations. Analysis:

[Richard Bäcker @RichardBcker1]

Ox Alpha stopping a lot mid thinking, does this happen to you too?

[Richard Bäcker @RichardBcker1]

getting a lot of requests failed with error 503 from oxalpha.

[Richard Bäcker @RichardBcker1]

@theo i second this, much greater reliability than i had with hy3 or deepseekv4flash. although much slower than deepseek.

[Richard Bäcker @RichardBcker1]

initial reaction to stealth Ox alpha @OpenRouter is that it is not doing thinking loops and is doing a lot of toolcalls to ground its thinking. leading to quite good results. it is fixing a lot of my broken stuff that deepseekv4flash was not able to.

[Richard Bäcker @RichardBcker1]

@trashh_dev to me it feels like I suddenly only have one tab.

[Richard Bäcker @RichardBcker1]

@ProEvilz @opencode same here

[Richard Bäcker @RichardBcker1]

@opencode xD! 100T. Zero data retention, interesting. What are they doing this for then? Load testing?

[Richard Bäcker @RichardBcker1]

@swmansion trying it out now, just what i was looking for

[Richard Bäcker @RichardBcker1]

@NousResearch @bridgemindai He is just ragebaiting

[Richard Bäcker @RichardBcker1]

@thdxr love opencodev2 and the opencode skill so that opencode can fix itself. in this case rewriting the opencode.json to the new v2 format so that auto approve permissions will work and stay enabled :)

[Richard Bäcker @RichardBcker1]

@opencode very low tokens/sec for me at around 180k context.

[Richard Bäcker @RichardBcker1]

@ego_agent Good idea. But regular websearching does the same. Looking for more :)

[Richard Bäcker @RichardBcker1]

@buildwithhassan here is a fresh performance graph for it:

[Richard Bäcker @RichardBcker1]

@thdxr same pls https://t.co/CdtOUP0mug

[Richard Bäcker @RichardBcker1]

@jackyk02 grok explained more as to why this works: https://t.co/MESPNOKGg2

[Richard Bäcker @RichardBcker1]

@theanupamshakya Startup — but as a founder, not employee #1. Five years is exactly the window where equity either compounds or teaches you what no job will. AI lowered the cost of testing ideas, so the downside is smaller than it looks.

[Richard Bäcker @RichardBcker1]

@jpshrodinger Trusted brand. Models commoditize every quarter, data goes stale, distribution can be bought — but brand is the only asset that defends all three. It's why people pay Apple margins for hardware anyone can build.