Error creating task or Error starting conversation

If you're able to repro this, please use the /feedback slash command to upload your logs and session data and post your thread ID here. We'll need this additional information to help diagnose the problem you're seeing.

Did you try restarting the Codex IDE extension? It sounds like that instance of the extension is in a bad state.

Yes, restarted all. Extension, VS Code, Mac.
Nothing helps.

Thread ID: 019ba4ab-a7f6-7523-a929-47325cf6f6e4

The logs weren't of much help unfortunately.

By any chance, are you trying to start a cloud task? That would explain why I'm not seeing any relevant local logs. In the lower left part of the extension panel, you should see a menu that allows you to select between cloud environments and your local workspace. Unless you intend to start a cloud task, make sure this is set to "Local Workspace".

No, it's a local task. A cloud task is working, but only one time, a followup als gives an error.

Here is a small video showing the problem.
https://github.com/user-attachments/assets/e6742199-d06a-4f14-8dda-8c3936dc822d

I found the problem with the help of Codex CLI 0.80. This version showed me an error a skill file.
Codex in VS Code silently failed.

/.codex/skills/implement-rate-limiting/SKILL.md: invalid YAML: mapping values are not allowed in this context at line 2 column 53

After fixing this Codex in VS Code worked again.

Thanks for the analysis. Codex should be more resilient here — or at least provide a better error message.

@Urkman, thanks for reporting this. When you have a moment, could you try reproducing the issue again by breaking the SKILL.md file? If it reproduces, could you also share the exact line that’s invalid in your skill definition? I have trouble to to reproduce it on my side.

Similar issues with New Agent tab, reported in 019c53e3-1b7c-7f03-9720-d54ef213e315

VSCode 1.109.2
WSL
codex-cli 0.98.0 - 0.101.0
extension 0.4.74 and 0.5.74

Errors:

• Failed to resume task
• Error creating task
• To continue, return to the window where this conversation started.

May also be relevent: At about the same time as these New Agent windows stopped working, I was getting HTTP 403 errors in the GitHub output window. Somehow the repo that I was working on had become un-authorized for use with the Codex Extension. I checked the GH account and the repo wasn't listed as authorized. I re-authed it and the local connection resumed. No one else uses the account. There was no log of the disconnection. It simply lost that relationship.
Maybe the conversations are linked with the repo ID and all of that config info was lost at the same time?

Moving my focus to #11727 and #11721, which are current and seem more relevant.

I had this issue too, i fixed it;

Hopefully this works for you;

Depending on your IDE, mine was windsurf, the ide's openai codex extension fails because codex could not parse the local agent config files in '~/.codex/agents/*.toml'

Those files used TOML multiline basic strings and contained shell patterns like \|, which made the config invalid. That broke conversation startup and led to the localhost:57017 refusal symptoms i saw.

What I changed (I had these installed which caused some failures)
I patched all of these to use TOML literal multiline strings instead:

gsd-codebase-mapper.toml
gsd-debugger.toml
gsd-executor.toml
gsd-integration-checker.toml
gsd-phase-researcher.toml
gsd-plan-checker.toml
gsd-planner.toml
gsd-project-researcher.toml
gsd-research-synthesizer.toml
gsd-roadmapper.toml
gsd-verifier.toml
Backups were created as *.toml.bak.

Verification
I verified Codex can now load config again:

codex features list
codex mcp list
So the parse error is resolved.

Then i restarted my ide completely and codex worked.

Still happening on latest version.

I am currently on v26.513, I simply reinstalled the IDE extension and it worked again. To my surprise when the session resumed it detected something went wrong with the previous thread