Omer Ben-Amram

[Omer Ben-Amram @OBenamram]

1. Codex CLI is still nowhere close to CC in terms of general usability, CC env variable hacks to use sol works okaish, but is not the default for most users. 2. Opus 5 is unusable - but still sticking with CC because of fable. It is still better overall is understanding user intent when planning, even if we don't use it to execute. 3. Please allow users in enterprise settings to see their spend!

[Omer Ben-Amram @OBenamram]

@charliermarsh Post sol - we need a better way to see subagent transcripts (rather than /root/cool_agent)

[Omer Ben-Amram @OBenamram]

@jarredsumner No point in hyping it if fable is unavailable 💔

[Omer Ben-Amram @OBenamram]

ultimate AI cli setup is definitely @FactoryAI droid + codex subscription via proxy. missions are insanely powerful with GPT-5.4. can also mix up w/ opus for orchestrating but have the codex subscription foot most of the bill. goated tool!

[Omer Ben-Amram @OBenamram]

@nir_benz שום סיכוי שהמסקנות שיש שם אמיתיות, רואים גם שה״דו״ח״ הזה זה פרומפט מצוקמק בלי הרבה השקעה. לא בלתי סביר שפשוט יש שם גם קוד/ספריות פיתוח פנימיות שלהם שמשמשות להשוואה או סתם קוד ניסיוני. אבל זה נראה כמו סתם engagement bait

[Omer Ben-Amram @OBenamram]

@NousResearch please fix gpt-5.4 from sounding like a deranged bot (if you want, i can bla bla)! the lobster have fixed the tone properly. for other stuff Hermes still cooks!

[Omer Ben-Amram @OBenamram]

. @NousResearch Switched from Opus/glm-5.1 to gpt-5.4, it crushes tasks, code output is amazing. but can't get it to shake off the extremely verbose responses, and ending all responses with "if you want...". even when meddling with it's system prompts. What do?

[Omer Ben-Amram @OBenamram]

@NousResearch loving Hermes so far! How could we response to messages from crons on telegram topics? The only feature I really miss from @openclaw !

[Omer Ben-Amram @OBenamram]

@JoelDeTeves QMD locally with custom memory plugin!

[Omer Ben-Amram @OBenamram]

@nir_benz general agent - replaces openclaw for me. i found it much easier to configure, and the self skill synthesis is incredible

[Omer Ben-Amram @OBenamram]

Spend last 24hrs evaluating replacements for opus in Hermes Agent. GLM-5.1 and nothing else is even close (tried MiniMax, Mipo, Kimi and Qwen 3.6). GLM-5.1 is hands down the best!

[Omer Ben-Amram @OBenamram]

שמות מבצעים צהליים להיות כמו

[Omer Ben-Amram @OBenamram]

@mitsuhiko I think another interesting metric is how much of the code do we still actually read? a lot of code becomes "compiled" artifact and i only read the api/boundaries.

[Omer Ben-Amram @OBenamram]

@aye_aye_kaplan GPT-5.2-XHIGH for basically all tasks. opus if really quick ones or questions/research. id rather have gpt cook on longer tasks and run them in parallel than babysit opus which fails at 30-50% of tasks ("let me simplify this!").

[Omer Ben-Amram @OBenamram]

@aloncarmel המנוי של openai דרך codex מאוד מאוד נדיב (משמעותית יותר מקלוד) גם בחבילה הזולה (לא הpro). בהרבה משימות גם מדובר במודל יותר טוב (אם משתמשים בextra high)

[Omer Ben-Amram @OBenamram]

@jediahkatz @aye_aye_kaplan thanks! another example :) request id 1655ffb4-b752-4479-afda-870bcb6b9265

[Omer Ben-Amram @OBenamram]

@ericzakariasson Subagents. and don't obliterate system resources on longer chats (1 chat could literally eat up all memory on a beefed up mbp until system just dies).

[Omer Ben-Amram @OBenamram]

it was a few days ago, - i couldn't pull up the exact place in the conversation though - but i explicitly remember those complaints (also i added into cursorrules later to explicitly avoid the web search beacuse of consistently poor results). here is claude codes best attempt of grepping cursor's state files! thanks :)

[Omer Ben-Amram @OBenamram]

@ketacode כן לגמרי - אבל ��ראה שבשביל שיהיה לו מספר צריך לארגן לו מספר אמיתי (בשביל לנצל את הheartbeats/ שהוא יוזם שיחה)