aider

Recovered with lab prompts aider from the trace proxy (newest capture 20260711T110347Z). aider routes its model calls through litellm, which the proxy sees as ordinary chat-completions, so the prompts below are the exact text that reached the model, not a copy from the tool's source. The prompts are stored raw under prompts/aider/; this page reviews what each one does and links to the file. Sizes are in tokens, counted with the o200k_base encoding.

Coder prompt

• 260 tokens
• no tools
• raw: prompts/aider/1-coder.md

aider takes a different path from every other agent here: it exposes no tools at all. Instead of a function schema it opens with "Act as an expert software developer" and teaches the model an edit format in prose, asking for whole-file or search-and-replace blocks it can parse out of the reply and apply itself. That is why the prompt is one of the smallest in the set even though aider is a capable coding tool: the machinery lives in aider's own parser, not in the prompt. One artifact from the Hi! run is worth flagging. aider sends no output cap, so on the ceiling model the completion ran to the proxy's max_tokens floor and stopped at exactly 32,000 tokens, a runaway that reflects the floor rather than anything aider asked for.

Summarizer

• 175 tokens
• no tools
• raw: prompts/aider/2-summarizer.md

The second prompt is aider's chat summarizer, the side call it makes to compress a long conversation back into a short running history. It asks for a terse, second-person summary written as if the user said it, so the digest can be fed back in as context. Like the coder prompt it is tool-free and small, and it never touches the scored task.