Context Tracker

Product description: Context Tracker is a zero-config prompt recorder, context profiler, and prompt co-pilot for VS Code and Cursor. You click Start capture. Muse Spark inspects the project, probes local model servers, and finds the active endpoint in code, .env, or user config. Each captured call is split into sections with a token-budget share, a WHAT label, and a WHY label, then linked back to the file and line in the project. Dynamic templates are traced by Muse Spark. Click any section and ask the co-pilot to find contradictions, prune bloat, write a summary, or apply a patch straight to the source file. A Meta model stays available to answer questions as you go. Problem it solves: LLM apps build every request from system instructions, templates, chat history, tool schemas, and retrieved documents. When a model hallucinates, ignores a rule, or burns more tokens than expected, the useful question is: what exact prompt was sent, which file each slice came from, and why that slice was included. Answering that today means editing .env files by hand, standing up a proxy, hunting through ~/.config, and reading raw JSON. Dynamic prompts (f-strings, Jinja2, Mustache, RAG injections) do not match the source text, so a captured request still does not point at the line that built it. Windows of 10k to 50k+ tokens are too large to scan for a contradiction between a system rule and a retrieved chunk.