"This model does not support assistant message prefill." in /agent mode (but prefill not configured)

[ekr-nttd-emeal]

Issue

When I work with /agent, I get this output:

architect> /agent
Context management enabled for large files
cecli v0.98.0
Main model: github_copilot/claude-opus-4.6 with agent edit format
Weak model: github_copilot/claude-sonnet-4.5
Git repo: .git with 2,214 files
Warning: For large repos, consider using --subtree-only and .cecli_ignore
See: https://cecli.dev/docs/faq.html#can-i-use-cecli-in-a-large-mono-repo
Repo-map: using 1024 tokens, auto refresh
MCP servers configured: Local

agent> Find the location where XYZ and change it to do ABC.
...
# some agentic back-and forth with local "rg" calls and such - seems to work ...
# eventually:
...
Tool Call: Local • ShowContext
Arguments: {"show": 
[{"file_path":"somefile","start_text":"bla","end_text":"blabla","padding":2}]}

Successfully retrieved context for 1 file(s)
LiteLLM API Error: litellm.BadRequestError: Github_copilotException - This model
does not support assistant message prefill. The conversation must end with a 
user message.
litellm.APIError: Error building chunks for logging/streaming usage calculation
Retrying in 0.2 seconds...
...

Note that it does not say anything about "infinite output" in the model info, which I think would be the case if the "prefill" mode was enabled for these models.

It is unclear to me how to proceed. I have the feeling that the error message is misleading. Searching the Cecli documentation brings this for "prefill": https://cecli.dev/docs/more/infinite-output.html - and this suggests that Cecli would mention "infinite output" in the model lines if this feature were enabled.

Is there something I can do as a user to fix this issue? Some configuration to set, or prompt to give?

Version and model info

cecli v0.98.0
Main model: github_copilot/claude-opus-4.6 with agent edit format
Weak model: github_copilot/claude-sonnet-4.5
Git repo: .git with 2,214 files
Warning: For large repos, consider using --subtree-only and .cecli_ignore
See: https://cecli.dev/docs/faq.html#can-i-use-cecli-in-a-large-mono-repo
Repo-map: using 1024 tokens, auto refresh
MCP servers configured: Local

On MacOS, if it matters.

[dwash96]

So it's caused by this generally:
https://www.reddit.com/r/LLMDevs/comments/1r0ih5p/opus_removes_lastassistantturn_prefill_you_can_no/

[ekr-nttd-emeal]

Thanks @dwash96!

Is there some way to configure Cecli to know about this? Or would this be a code change?

[dwash96]

It's a code change, and I'm still thinking about what makes sense to do in this case in agent mode since usually, a tool call response is the last message in the stream and its alternating assistant messages with those tool call responses

[dwash96]

Alright, so if the model doesn't support prefilling and the conversation system would send a message chain that does not end in a user response, I am going to inject a simple dummy user message prompting the model to proceeed. I'll add that to 0.98.1

[ekr-nttd-emeal]

Awesome, thank you @dwash96!