Zimeng Xiong's WeblogThursday, 5 February 2026
slight improvement!
really impressed with CODEX plus use limits. having a hard time exhausting the 5 hour limit (2-3 sessions running for 6+ hours today continously). the weekly limit though...
hosted version of excalidash when?? might consider it! convex backend should be good enough. definitely need to create the pointer-image system
codex is so much better than opus its not even close
can confirm, 5.3 still sucks at UI compared to base gemini 3 even with frontend design skills
really liking the codex app, it's a bit laggy, but still very useful, i think i prefer it over CLI tools now...ironic
also done on the convex migration (1hr 45min). also looks promising
5.2 high managed to get the PRs done! let's see how good it is
i have to say i still much prefer CLI tools over GUI/IDEs for agentic workflows? i totally believe that GUIs will be the future for harnesses like CC and Codex, but something about just being able to open a new tab or sigkill the current instance and just be in a terminal just feels like i have more control. theirs nothing i can do in a IDE/GUI than hitting stop, it just feels helpless (yes i can open a terminal, but it doesnt feel the same)
letting it run overnight, lets see if it can close all my PRs haha
(also testing out the codex app)
something you would never see Opus saying
5.2 high migrating everything to convex 🤩 (its been 1.5hrs so far)
i think it will work first try
really proud of the settings page preview animation that shows a sloppy stylization of the drawn path, really cool
just had 5.2 high code up a openclaw alternative.
i just want it to:
- respond to my signal messages
- keep a memory file (might be use implementing semantic search or vector based search)
- heartbeat, i.e. agent can set a alarm for it to wake up, default 30minutes wake up and do stuff
- shell access
- ability to edit itself/it's own config (temp migration of the agent onto another place, and then the agent is free to edit it's config/harness, and then agent gets transferred back to its own harness)
- browser use
will test it out tomorrow!
well, we got there. not bad, only 20 minutes left
been using 5.2 high for a day
1. the CLI is SOOO much better than CC and Gemini its not even close. super fast and responsive. does lack a lot of features like ctrl+o to inspect, etc, etc.
2. as expected, reasons A LOT more than gemini or opus models, average session is ~2-3min for a quick fix and easily 30+ minutes for multi stage changes, not a quick model. having it refactor an entire website from firebase to convex right now, running non stop for nearly an hour now.
3. with 3-4 sessions I'm not hitting my hourly limit at all? i'm sure i could hit it if I tried, but surprising as this is just the $20 tier.
4. context comopression is nicely done and doesnt leave me waiting for 2 minutes while opus tries to compact its tiny context
so far, everything it changed at least builds and is solid so far, definitely prefer over opus.
just bought GPT $20/mo plan (free first month). the 5hr limits are surprisingly fine, already 30% done with the weekly cap. don't have access to 5.3 on the $20/mo.
after using 4.6 for a bit, it feels like opus is finally catching up to GPT level reasoning, definitely is a long reasoning model. never had opus sustain a session for over 20+ minutes before.
restarted my server and forgot to update my bind mount so my auth was down all day 😦