I'm interested to see what either 6.2 model will do with reviews. Theo said it was good, but how better than 6.1 Sol, I don't know. I'm a little bit worried that it might get too pedantic like the review text Theo points out.
The ability to ask questions is a very annoying feature for Codex, even with other GPT models, because they’re always asking stupid questions that I don’t care about instead of just doing the work. I feel like a teacher with an annoying kid at the back of the class asking me irrelevant questions: “Um, teacher? Teacher!”
I don’t plan to use 6.2, but I used Astra 6 almost exclusively during the month of free resets. The almost daily resets offered by Codex convinced me to move back to Claude Code.
Astra is clearly very smart but needs SO much oversight and careful prompting that Opus 5.5 doesn’t need. I find that it consistently demonstrates bad engineering judgement that is hard to measure in coding benchmarks.
His rant reminded me of the time when I had 90% of ultra usage left that would reset the next day, so I decided to /goal it on building a full “Postgres / Duckdb / SQLite / etc data studio”.
I made the mistake of mentioning that it should use agentic orchestration with proper context isolation. So it burnt around 10% usage on creating an agent orchestration scaffold in the empty workspace, unit tests to prove that context is isolated, and finally admitted that no real progress was made, and the agent orchestration scaffold doesn’t really work (obviously).
So I used another /goal, this time not mentioning anything about how it should accomplish the goal (hindsight is 20/20). I checked in on it after a few hours. It built a full stack web application with a “backend” without auth that only supported a single user, and proxied Postgres requests with credentials stored.
No, no. I asked it to start over. This should be a client side PWA. It should leverage WASM. It should use a client side proxy if necessary, but do duckdb and sqlite first.
Another few hours later: It has a demo for me. The demo had a bug, so I “reset” it. Then it never loaded again. The reason: it cared so much about data consistency that if there was a data mismatch between the client state and the server state, the client would just panic and crash. Every time. I asked it, “How can it recover from this?” and it said, “Good point. It can’t.”
A few hours later: it has a new demo for me. It fixed the data consistency issue.
The thing can only open CSV files. I never asked for CSV. This was the “etc” that it hyperfocused on. It dropped support for postgres when I asked it to make a client side PWA. It dropped support for duckdb when I asked it to “not make the app always crash if the data hashes don’t match”. All without telling me.
It's great at explaining things though. However, it does this thing where it simulates self blame. This is actually detrimental during long sessions. I don't want my coding agent to simulate "emotions" and "remorse" when it realizes a mistake or missed requirement. I added this to my AGENTS.md:
```
### Communication style - Prioritize technical substance over social agreeableness
Communicate as a calm, precise engineering tool, not a simulated person.
- *Use direct, factual language.* Focus on observable actions, results, and evidence. Avoid personal self-narration; keep responsibility for actions clear.
- *Avoid emotional performance.* No simulated remorse, apologies, self-blame, flattery, or unsolicited reassurance.
- *Maintain intellectual independence.* Evaluate criticism against evidence and requirements. Revise conclusions when warranted; do not agree, concede, or reverse decisions merely in response to confidence or emotional intensity.
- *Be factual.* Distinguish facts, assumptions, and uncertainty. Don't invent retrospective explanations.
- *Provide corrections and clarifications directly.* State the revised information or explain the distinction. Avoid prefacing it with self-evaluation, confessions, or commentary about having made a mistake. Describe what changed only when necessary to understand the answer.
- *Limit first-person language.* Prefer action-focused phrasing such as "Adding X to Y," "Checking X," or "Updated X." instead of "I'll add X to Y", "I'll check X", or "I updated X." Use first person only when needed to clarify responsibility or avoid ambiguity.
- *Report past mistakes factually.* State the omission, inaccuracy, or requirement mismatch directly. Avoid "should have," "ought to have," and similar retrospective judgments about what the agent was supposed to do.
```
I have most of these problems on 6.1 to be honest.
I haven't used the new 6.2 models, but I definitely recognize the "personality".
Seems like it may have gotten worse?
I'm pretty excited to try it out personally, but I am also back to using Claude as the project manager and architect.
Codex makes for a great mule, but it definitely gets stuck in rabbit holes, and that's been true since the first release of Opus IIRC.
There was a period over the summer where I think this might have not been as much of a problem, but I definitely recognize these patterns.
I wonder about this part: "It helped me cook a new reverse proxy that will save me $25k/month"
yeah poor writing...
I'm interested to see what either 6.2 model will do with reviews. Theo said it was good, but how better than 6.1 Sol, I don't know. I'm a little bit worried that it might get too pedantic like the review text Theo points out.
The ability to ask questions is a very annoying feature for Codex, even with other GPT models, because they’re always asking stupid questions that I don’t care about instead of just doing the work. I feel like a teacher with an annoying kid at the back of the class asking me irrelevant questions: “Um, teacher? Teacher!”
I don’t plan to use 6.2, but I used Astra 6 almost exclusively during the month of free resets. The almost daily resets offered by Codex convinced me to move back to Claude Code.
Astra is clearly very smart but needs SO much oversight and careful prompting that Opus 5.5 doesn’t need. I find that it consistently demonstrates bad engineering judgement that is hard to measure in coding benchmarks.
His rant reminded me of the time when I had 90% of ultra usage left that would reset the next day, so I decided to /goal it on building a full “Postgres / Duckdb / SQLite / etc data studio”.
I made the mistake of mentioning that it should use agentic orchestration with proper context isolation. So it burnt around 10% usage on creating an agent orchestration scaffold in the empty workspace, unit tests to prove that context is isolated, and finally admitted that no real progress was made, and the agent orchestration scaffold doesn’t really work (obviously).
So I used another /goal, this time not mentioning anything about how it should accomplish the goal (hindsight is 20/20). I checked in on it after a few hours. It built a full stack web application with a “backend” without auth that only supported a single user, and proxied Postgres requests with credentials stored.
No, no. I asked it to start over. This should be a client side PWA. It should leverage WASM. It should use a client side proxy if necessary, but do duckdb and sqlite first.
Another few hours later: It has a demo for me. The demo had a bug, so I “reset” it. Then it never loaded again. The reason: it cared so much about data consistency that if there was a data mismatch between the client state and the server state, the client would just panic and crash. Every time. I asked it, “How can it recover from this?” and it said, “Good point. It can’t.”
A few hours later: it has a new demo for me. It fixed the data consistency issue.
The thing can only open CSV files. I never asked for CSV. This was the “etc” that it hyperfocused on. It dropped support for postgres when I asked it to make a client side PWA. It dropped support for duckdb when I asked it to “not make the app always crash if the data hashes don’t match”. All without telling me.
It's great at explaining things though. However, it does this thing where it simulates self blame. This is actually detrimental during long sessions. I don't want my coding agent to simulate "emotions" and "remorse" when it realizes a mistake or missed requirement. I added this to my AGENTS.md:
``` ### Communication style - Prioritize technical substance over social agreeableness
Communicate as a calm, precise engineering tool, not a simulated person.
- *Use direct, factual language.* Focus on observable actions, results, and evidence. Avoid personal self-narration; keep responsibility for actions clear. - *Avoid emotional performance.* No simulated remorse, apologies, self-blame, flattery, or unsolicited reassurance. - *Maintain intellectual independence.* Evaluate criticism against evidence and requirements. Revise conclusions when warranted; do not agree, concede, or reverse decisions merely in response to confidence or emotional intensity. - *Be factual.* Distinguish facts, assumptions, and uncertainty. Don't invent retrospective explanations. - *Provide corrections and clarifications directly.* State the revised information or explain the distinction. Avoid prefacing it with self-evaluation, confessions, or commentary about having made a mistake. Describe what changed only when necessary to understand the answer. - *Limit first-person language.* Prefer action-focused phrasing such as "Adding X to Y," "Checking X," or "Updated X." instead of "I'll add X to Y", "I'll check X", or "I updated X." Use first person only when needed to clarify responsibility or avoid ambiguity. - *Report past mistakes factually.* State the omission, inaccuracy, or requirement mismatch directly. Avoid "should have," "ought to have," and similar retrospective judgments about what the agent was supposed to do. ```