A real card, and a real run

This card is from a run of 19 September 2026, shortened for the page but not rewritten. A person writes the top half. The loop writes the bottom.

What a person writes

files:
  - app/src/main/java/dk/cocode/xpost/PostLink.java
  - app/src/test/java/dk/cocode/xpost/PostLinkTest.java
gate_files_are_the_work: true
requirement:
  goal: A pure function that returns a normalized https fetch URL for a valid
    x.com or twitter.com single-post link, or says the input is not one.
  done_when: PostLinkTest compiles and exits 0, proving the function normalizes
    valid status links (with or without a query string, http or https, with or
    without www) and refuses anything else.

The gate

A command, nothing more. Its exit code is the verdict, and the loop proves it fails before anyone is allowed to make it pass.

set -e -o pipefail
TMPD=$(mktemp -d)
javac -d "$TMPD" app/src/main/java/dk/cocode/xpost/PostLink.java \
                 app/src/test/java/dk/cocode/xpost/PostLinkTest.java
java -cp "$TMPD" dk.cocode.xpost.PostLinkTest

What the loop writes back

Only front matter, never the words. The note stays yours.

status: done
kept_at: '2026-09-19T14:06:00Z'

What that run cost

Four cards planned, three of them buildable. All three were built, gated, reviewed and kept, none parked, none dropped, with no restart by hand.

• 31 minutes from start to stand-down.
• $3.66 in total: $1.68 building, $1.30 reviewing, $0.68 planning.
• One thing was refused as needing a person, with the reason recorded: turning a fetched page into the post's text could not be proved by any gate the loop could write.

The code all three cards produced compiles and passes its tests from a clean clone. That check was run by a person, outside the loop, because the thing that builds evidence never grades it.

How the loop runs a card Back to the start