Point the extractor at any agent codebase and get back the ACDL specification of how it assembles context — with a line-by-line evidence table.
Runs through the Claude Code CLI, so it bills your Claude subscription rather than API credits. Needs claude on your PATH and node for the run capture.
The standalone runner. Bills your Anthropic API key; add --dry-run to see the assembled prompt without spending anything.
Commit this workflow to the repo and run it from the Actions tab. It runs on your Actions minutes with your ANTHROPIC_API_KEY secret, and leaves the spec as a downloadable artifact.
Can't commit to the repo you want to analyze? Fork the ACDL repo, add the same secret there, and run its ACDL extract workflow with any public repository as the input.
Reverse-engineers the message array your agent builds before every model call
It works from the code that actually assembles the prompt. READMEs describe intent; the message array is the ground truth.
Read-only file access, scoped to the checkout. Nothing in the target is executed or modified — the deliverables land in a separate output directory.
Every line of the emitted spec is traced to a file and line number in the report, so you can check the extraction instead of trusting it.
A real extraction is a long agentic run over the whole codebase — minutes of wall time and a few dollars of tokens, not an instant answer.