Have a local LLM parse a large stdout from the given command to cut down on massive stack-trace-induced context window explosions

parse-output.sh

#!/bin/bash

# Script: parse-output.sh
# Purpose: Run a command and use Qwen LLM to parse any error output
# Usage: ./parse-output.sh [--show-output] <command>
# Default: Quiet mode (LLM-friendly output only)
# Debug mode: Use --show-output to see full command output and progress

MODEL="mlx-community/Qwen3-30B-A3B-4bit-DWQ"
ENABLE_THINKING=false

# Check for show-output flag
SHOW_OUTPUT=false
if [ "$1" = "--show-output" ]; then
    SHOW_OUTPUT=true
    shift  # Remove the flag from arguments
fi

# Check if a command was provided
if [ $# -eq 0 ]; then
    echo "Usage: $0 [--show-output] <command>"
    echo "Example: $0 ./gradlew test"
    echo "Example: $0 --show-output ./gradlew test"
    exit 1
fi

# Create temporary files
TEMP_OUTPUT=$(mktemp)
TEMP_ERROR=$(mktemp)
TEMP_MLX_OUTPUT=$(mktemp)

# Function to clean up temporary files
cleanup() {
    rm -f "$TEMP_OUTPUT" "$TEMP_ERROR" "${TEMP_OUTPUT}.combined" "$TEMP_MLX_OUTPUT"
}
trap cleanup EXIT

# Run the command
if [ "$SHOW_OUTPUT" = true ]; then
    echo "Running: $@"
    echo "----------------------------------------"
    "$@" > >(tee "$TEMP_OUTPUT") 2> >(tee "$TEMP_ERROR" >&2)
    EXIT_CODE=$?
else
    # Default quiet mode - capture silently
    "$@" > "$TEMP_OUTPUT" 2> "$TEMP_ERROR"
    EXIT_CODE=$?
fi

# Show status in debug mode
if [ "$SHOW_OUTPUT" = true ]; then
    echo ""
    echo "----------------------------------------"
    echo "Command failed with exit code: $EXIT_CODE"
    echo "Analyzing error with Qwen..."
    echo "----------------------------------------"
fi

# Combine outputs
cat "$TEMP_OUTPUT" "$TEMP_ERROR" > "${TEMP_OUTPUT}.combined"

# Create the prompt
# Conditionally add /no_think based on ENABLE_THINKING
if [ "$ENABLE_THINKING" = true ]; then
    NO_THINK=""
else
    NO_THINK="/no_think"
fi

PROMPT="$NO_THINK
You are an expert error analyzer for Java applications and build systems.

<input>
  $(cat "${TEMP_OUTPUT}.combined")
</input>

<instructions>
- Extract only essential information for debugging
- Focus on user code, not framework internals
- Include specific error messages and line numbers
- Remove all JUnit/TestNG/Gradle/Maven internals
- Be extremely concise but include all critical details
- Do NOT make suggestions or explanations, just extract information.
</instructions>

<output_format>
For each failed test, note:
- The name of the suite(s) and test(s) that failed.
- All relevant text from each error output,
  e.g. 'java.lang.NullPointerException: Cannot invoke org.springframework.cache.Cache.invalidate() because the return value of com.netflix.uhcapi.cache.TopNContentCache.getTopNQueryCache() is null'
- Only 3-5 most important lines from user code, not framework internals.
</output_format>
"

# Run mlx_lm and capture output
mlx_lm.generate \
    --model "$MODEL" \
    --prompt "$PROMPT" \
    --max-tokens 2000 \
    --temp 0.1 > "$TEMP_MLX_OUTPUT" 2>&1

# Extract and display the generated text
if [ "$SHOW_OUTPUT" = true ]; then
    # Debug mode - show full mlx output
    cat "$TEMP_MLX_OUTPUT"
else
    # Default quiet mode - extract only the generated content
    # First extract content between ========== markers, then remove <think> blocks
    awk '/^==========/{if(++count==1) next; if(count==2) exit} count==1' "$TEMP_MLX_OUTPUT" | \
    sed '/<think>/,/<\/think>/d' | \
    sed '/^$/d'
fi

# Exit with original error code
exit $EXIT_CODE

Prerequisite installs:

pip install --upgrade mlx-lm transformers tokenizers