Auto-Compose think

[hellocodelinux]

I have Ollama locally with Qwen3.
It has "think".
This gets included in the response.
Is there a way to avoid that?
Right now I get the answer but it also shows everything inside .

[hellocodelinux]

for the moment I did this

auto_compose_command=/home/eduardo/ia.sh %1

#!/bin/sh

file="$1"
txt=$(cat "$file")
prompt="en esta conversacion tu eres lu9dce, continua y responde, agrega emojis"
resp=$(printf "%s\n%s" "$prompt" "$txt" | ollama run qwen3:0.6b)
clean=$(echo "$resp" \
    | sed '/Thinking\.\.\./,/\.done thinking\./d' \
    | sed '1d')
echo "$clean"

[d99kris]

Ok, good to hear you found a workaround 👍

I'm not sure yet if there is another more generic approach to solving this in the nchat compose script. Hopefully someone else finds this thread and can chip in.

[hellocodelinux]

Let’s look at this.

Ollama offers a cloud service with powerful models.
It’s completely free — you just need to create an account and generate an API key.

These cloud models don’t use local disk space and don’t consume your CPU, since everything runs on Ollama’s servers.

The small models that run locally can’t really follow a full conversation thread.

To get good answers, the AI needs proper context in the prompt.

It’s also useful to remove the “thinking” section included in some responses.

A more elegant solution would be possible if the cloud version supported --json output, but I couldn’t find a way to enable it.

In the nchat configuration, I changed the auto compose command so it points to a bash script:

auto_compose_command=/home/eduardo/ia.sh %1

%1 is the temporary .txt path provided by nchat.

#!/bin/sh

# Set the OLLAMA_API_KEY environment variable for authentication
export OLLAMA_API_KEY="YOU API HERE"

# The first argument ($1) is the input file containing the conversation
file="$1"

# Read the entire content of the input file into the 'txt' variable
txt=$(cat "$file")

# Define the full prompt to be sent to the AI model
# This prompt instructs the model to analyze the conversation, understand the roles,
# and respond briefly to the last unanswered message using a specific persona (lu9dce's style).
prompt="read this conversation and analyze it. lu9dce is me, shoa is my wife, angeles and juan are my children, toto is my dog, pan is angeles's baby cat. respond to the last message that I have NOT responded to in a single, brief context. answer with my writing style and my words, in the language of the messages, and add an emoji"

# Combine the prompt and the conversation text, separated by two newlines, and pass it to ollama
# The 'deepseek-v3.1:671b-cloud' model is used for the generation
resp=$(printf "%s\n\n%s" "$prompt" "$txt" | ollama run deepseek-v3.1:671b-cloud)

# Filter the response: remove the 'Thinking...' block that some ollama models output
resp=$(echo "$resp" | sed '/^Thinking\.\.\./,/^...done thinking\./d')

# Remove the first line of the output (often a blank line or a greeting from the model)
resp=$(echo "$resp" | sed '1d')

# Print the final, cleaned response to the standard output
echo "$resp"