Dumb Dumb Code Checker

Slow and simple but somewhat working code analysis agent that can navigate through code base.

Quick start

Installation

git clone https://github.com/knilink/dumb-dumb-code-checker.git
cd dumb-dumb-code-checker
npm install

Then edit src/config.ts for preferred models and options.

Run

Here using this project as an example. For different project, specify a different --workspace.

# please let me know immedediately if you come across with a model smart enough to notice this project is actually itself
OLLAMA_HOST='http://localhost:11434' npm run start -- --workspace . --query "what does this stupid project do?" --max-iterations 16 --init-ctx=none

# lol
npm run start -- -w . -q 'find evidences in this project to prove YOU ARE "dumb-dumb-code-checker"'

# this one is interesting as "model" could mean data model, llm model or else
npm run start -- -w . -q "how do i use a different model?" --init-ctx=none

npm run start -- -w . -q "getting connect ECONNREFUSED 127.0.0.1:11434"

# should know it's unrelevant and ask for clarification straight away
npm start -- -w . -q "is calling kakapo dumb offensive?" --init-ctx=relevant-files

Suggested models

This is base my experience testing with open source models.

Thinking

Chain of thought model is highly recommended.

• qwq:32b-preview-q4_K_M: creative but heavily hallucinating, suggested to use with a decent summarizer to filter out noises, recommended temperature 0.7 ~ 1.2
• deepseek-r1:32b-qwen-distill-q4_K_M: concise and stable but less creative, being path dependence and not able to think out of box. qwq is preferred over this model

Summarizing

Models with decent reasoning capabilities will be more likely to capture key information and filter out noise.

• mistral-large:123b-instruct-2407-q2_K: top open source reasoning model
• qwen2.5:32b-instruct-q4_K_M: could be sufficient for simple, closed ended query

Tool Calling

Reasoning is not mandatory but a good reasoning model will be more likely to capture the correct information.

• qwen2.5:32b-instruct-q4_K_M: sufficient, but rarely it will simply copy/pasting content although the arguments make no sense
• mistral-large:123b-instruct-2407-q2_K: tool calling structure may deformed due to heavily quant, and it too good at reasoning thus sometime it may refuse to pick any tool if the query is too vague.

Tips

Although, technically, any query would work, however, dumb dumb is dumb and doesn't known when to stop investigating. Thus, a clear, simple, single purpose, closed ended question and explicitly stating the end condition could help ending investigation earlier.

• ✗ "i'm conducting an security audit and wondering where foo in file bar is from": security audit is not directly related to the query and dumb dumb will eventually try to access the source code of node.js looking for vulnerabilities
• ✓ "locate the source of 'foo' in file 'bar'": simple and explicitly state it's a navigation task

Also try a higher temperature if observing copy/pasting entire file or other content when summarizing.