max_tokens in requests somehow set to 128

[egorian3-coder]

Hello, guys!
Please help me understand what I’m doing wrong.
I choose agent mode and ask it to do something, but it always cuts the answer off at 128 tokens.

I’m running llama-server, and in verbose mode I see this request from Odysseus:
"74.25.704.164 D srv log_server_r: request: {"model":"Qwen3.6-35B-A3-Uncensored-HauhauCS-Aggressive-Q2_K_P.gguf","messages":[{"role":"user","content":"execute command pwd"}],"temperature":1.0,"stream":true,"stream_options":{"include_usage":true},"max_tokens":128,"session_id":"42a349f5-3517-48d1-b2a2-46f2d1909367","cache_prompt":true}"

I don’t understand where this "max_tokens":128 is coming from in request or how to remove it.

[DevSidd2006]

The Issue

In agent mode, you might notice that responses are sometimes abruptly cut off at exactly 128 tokens, even if you set a higher max_tokens limit or if the model isn't finished generating its response. This typically happens when you send a very short or simple prompt (like "execute command pwd" or "hello").

Why it happens

Odysseus has an internal intent classifier that reads your prompt before sending it to the agent loop. If a prompt is very short or doesn't explicitly look like it requires complex tool usage, the classifier labels it as a "low-signal turn" (usually meant for casual conversational chatter).

To reply faster and save compute resources for these casual chats, Odysseus routes low-signal turns through a "fast-path". However, this fast-path had a hardcoded limit that forced max_tokens to cap at 128, overriding your actual settings:

max_tokens=min(max_tokens or 128, 128)

The Fix

To resolve this, you need to remove the hardcoded 128 limit so the fast-path respects your configured token limits.

Open src/agent_loop.py.
Locate the _direct_low_signal execution block (around line 2674).
Change this line:

max_tokens=min(max_tokens or 128, 128),

To this:

max_tokens=max_tokens,

Restart the Odysseus backend, and your agent will now be able to return full-length responses for short prompts!

[DevSidd2006]

I hope it helps !!