[dtt-github]

Why

A non-expert reviewing an AI proposal sees a code diff they can't read, plus "all examples pass". Passing examples don't say what else changed. They should approve a behaviour change, not a code diff.

What

• Proposals open on a behaviour tab instead of the code diff.
• Compare behaviour runs every example on the base commit and with the proposal's files, in the browser, and compares every output cycle by cycle, including the extra cycles after each example ends. Each difference is put in words (lib/compare.ts):
  • a shift of the same behaviour reads "led now changes 1 cycle earlier";
  • anything else reads "busy at cycle 40 was 1, now 0 (different in 3 cycles)";
  • with no difference at all, it says the change only touches cases the examples don't try.
• For editors, how well each version is tested: examples passing, coverage, and each state machine's moves added or removed. This comes from POST /agent/{id}/compare (design::quality runs the gate on both trees).

Checks

• API test: the compare endpoint's shape and member-only access.
• cargo test --workspace (122), clippy, fmt; web typecheck, lint, build.
• In the browser, with two test proposals added to the local database (no model call):
  • Counting to N-2 was described cycle by cycle, with the first example going from passes to fails and examples passing 2 → 1.
  • Starting the counter at 1 was described as "led now changes 1 cycle earlier".

🤖 Generated with Claude Code