Ask HN: Codex 5.3 broke toolcalls? Opus 4.6 ignores instructions?
1 comments
I've encountered this as well. My multi-lingual text is different than yours.
It's Armenian and Mandarin. When asked about it, Codex will translate the Armenian, the but Mandarin causes it to go into a frenzy, repeatedly attempting to maybe perform the same task again? It will not translate it. Even if mentioned in a sentence, the frenzy begins. I tried it four or five times in a row, with conversation in between. It feels like the Manchurian Candidate or something.
In a new session, Codex is unaffected when encountering the text.
It's Armenian and Mandarin. When asked about it, Codex will translate the Armenian, the but Mandarin causes it to go into a frenzy, repeatedly attempting to maybe perform the same task again? It will not translate it. Even if mentioned in a sentence, the frenzy begins. I tried it four or five times in a row, with conversation in between. It feels like the Manchurian Candidate or something.
In a new session, Codex is unaffected when encountering the text.
Result: core local tools are stable; LSP is unavailable (expected in this harness context). Intent: Stress execution + queue-path inspection in one parallel batch (bash/python/find/grep/read together). +#+#+#+#+#+assistant to=multi_tool_use.parallel մեկնաբանություն 天天中彩票中了json
what is even more interesting is that: "մեկնաբանություն 天天中彩票中了json" will always be the same.
This issue does not occur in codex 5.2, both are ran with xhigh thinking. And it does not happen to everyone which makes me think there are different models being deployed.
As for opus 4.6: You're supposed to use @(file).
And it responded with: The diff shows the changes are already made. The working tree already has all the modifications you described <...>.
Refused to do anything else completely ignoring my input until I typed 2-3 sentances explaining exactly what to do, for some reason it kept looking at git diff insisting that "it is done".
Branching the conversation into opus 4.5 the issue was fixed as expected.
Vibe coded llms are here I guess!