af workspace failure with codex

[juanmed]

Hi,

Thanks for sharing your project. I am trying to get started with codev. The documentation is useful but I guess it could be improved to get up and running faster; will try to put a PR to improve this. However I am having issues with af workspace start using codex:

error: unexpected argument '--append-system-prompt' found

  tip: to pass '--append-system-prompt' as a value, use '-- --append-system-prompt'

Usage: codex [OPTIONS] [PROMPT]
       codex [OPTIONS] <COMMAND> [ARGS]

For more information, try '--help'.

My af-config.json file looks like this:

{
  "shell": {
    "architect": "codex",
    "builder": "codex",
    "shell": "bash"
  }
}

Any ideas? Thanks!

[juanmed]

According to chatgpt:

Agent Farm always appends --append-system-prompt in shared role handling, which is valid for Claude but not Codex CLI.

I’ve confirmed the failure path affects both af workspace start architect sessions and spawned builder/worktree sessions when role prompts are enabled.

So I guess it is not my configuration but a problem in codev.

[baogiadoan]

I’m running into the same problem. Codev seems too tightly coupled to Claude. It would be better if it natively supported switching between agent backends, instead of depending heavily on Claude—for example, allowing the Architect agent’s underlying model to be changed easily (between Claude, Codex, Gemini), possibly even during an active session.

[waleedkadous]

I'm investigating implementing this and will come back with a spec shortly. But I will need your help to ensure that evrything works since I don't use those agent harnesses regularly. If I build it, can you help with that?

…

On Wed, Apr 1, 2026, 8:33 PM Bao Doan ***@***.***> wrote: *baogiadoan* left a comment (cluesmith/codev#591) <#591 (comment)> I’m running into the same problem. Codev seems too tightly coupled to Claude. It would be better if it natively supported switching between agent backends, instead of depending heavily on Claude—for example, allowing the Architect agent’s underlying model to be changed easily (between Claude, Codex, Gemini), possibly even during an active session. — Reply to this email directly, view it on GitHub <#591?email_source=notifications&email_token=AAUXGUHRFIKRUXP5AYBK6SD4TXNRPA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTIMJXGQ2DAMRQGQ2KM4TFMFZW63VKON2WE43DOJUWEZLEUVSXMZLOOSWGM33PORSXEX3DNRUWG2Y#issuecomment-4174402044>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAUXGUEKWC7EHFCYBJGSTZD4TXNRPAVCNFSM6AAAAACWLKHUJOVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHM2DCNZUGQYDEMBUGQ> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[baogiadoan]

Sure I'm happy to help with that. In fact, I am still using your initial version of Codev which is based on Markdown files, which is quite flexible. Great work @waleedkadous.

[waleedkadous]

Can you have a look at this spec? #654 -- it follows a similar model to the approach @nharward used to support different forges (gitlab vs github vs gitea etc). Please leave comments on the spec file if you have any feedback.

[baogiadoan]

@waleedkadous I think the spec #654 is good enough for now.

[waleedkadous]

OK I'll get started. Curious. What do we need to do to go from "good enough for now" to "this is a really great implementation"?

…

On Thu, Apr 2, 2026 at 5:09 PM Bao Doan ***@***.***> wrote: *baogiadoan* left a comment (cluesmith/codev#591) <#591 (comment)> @waleedkadous <https://github.com/waleedkadous> I think the spec #654 <#654> is good enough for now. — Reply to this email directly, view it on GitHub <#591?email_source=notifications&email_token=AAUXGUCFA4N34OPDYQO7MYL4T36LJA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTIMJYGEYDSNBSHE4KM4TFMFZW63VHNVSW45DJN5XKKZLWMVXHJLDGN5XXIZLSL5RWY2LDNM#issuecomment-4181094298>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAUXGUGX3SGDSM7CQ4YMMCT4T36LJAVCNFSM6AAAAACWLKHUJOVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHM2DCOBRGA4TIMRZHA> . You are receiving this because you were mentioned.Message ID: ***@***.***>

[baogiadoan]

Thanks,

It would be great if:

1. In the UI or command line we can choose which model to use ["Claude", "Codex", "Gemini"] (or any other models), and assign each to each role (Architect, Builders, etc).
2. We can have the option to switch between models during active sessions. This would be helpful if one has different subscriptions (ChatGPT, Claude, Gemini) and one subscription hits the rate limit (like Claude), and we can easily switch to another subscription/agent to keep working.
3. If it can supports 3rd party API like OpenRouter etc. Using something similar to Claude Code Router, but I think this might be another Spec, not this one.