Skip to main content
A script card explains one run step’s script, so that a reader can review it without reading the code. The dashboard shows each card beside its script.
A card has:
  • a one-sentence summary, and whether the script changes nothing or changes data;
  • effects found by code analysis, never by a model: the hosts it calls, the declared secrets and other environment variables it reads, the files it reads and writes, and the commands it runs;
  • 3 to 8 numbered steps in plain English, with line ranges, and the edge cases that the code does not handle.

​
How a card is checked

  1. Names and numbers. Every number, quoted name, and host in the card’s text must appear in the code. A failed check gets one retry.
  2. Round trip. A model writes a program from the card’s steps alone. Up to 3 recorded inputs of the step from your local runs go through both programs, with no network access, and the outputs are compared. A script that calls the network is not replayed. --no-round-trip skips this check.
A failed check is shown on the card. It never blocks a publish.

​
When cards are made

A signed-in run that saves a version starts method explain in the background for each script without a card. method publish makes the missing cards before it publishes. A card belongs to one script hash: change the script, and the step gets a new card. method explain needs a saved version of the file’s current content, and uses your hosted model credit. Without sign-in it makes no card. See Observed effects for what a script really did in a run.