Simon Willison has released version 0.5 of llm-echo, a debug plugin that simulates an LLM by simply repeating the prompts it receives. The update modifies how the tool calculates simulated token usage, ensuring that counts now include tool calls, system prompts, and reasoning tokens alongside standard input and output. A user running the command `uvx –with llm-echo llm -m echo ‘hi there’ -o thinking 1 –usage` receives a response that mimics a thinking process while displaying a JSON object containing the prompt, system instructions, and attachment data. The output explicitly lists token usage as 2 input and 24 output for the example.
This refinement matters for developers testing local models because it provides a realistic baseline for cost estimation and latency analysis without incurring actual API charges. By accounting for hidden overheads like system instructions and reasoning steps, the plugin offers a more accurate proxy for real-world model behaviour during the debugging phase. This allows engineers to validate their prompts and evaluate performance metrics before deploying code to production environments.
* Supports tool call simulation within the token count
* Includes system prompt length in the total usage
* Outputs JSON data for machine-readable verification
