The code review re-does the judge's playtest #307

Closed
opened 2026-10-29 15:32:32 -04:00 by cmoriarty · 1 comment
Owner

What happens
Run 139's agent.code-review spent most of its hour on a desktop and phone playtest it handed to the browser subagent through window.__app; its own scripted player could not reach a win or a loss, it was cut at its 1 h limit (#303), and the judge then played the game again for 29 min. After apply, the review was the largest cost in the item.

Fix
The code review reads the code and runs the lanes; playing the app is the judge's, and the review brief says so.

Done when
A code review on a web game finishes without a browser playtest, and the judge still plays it.

From run 139.

**What happens** Run 139's `agent.code-review` spent most of its hour on a desktop and phone playtest it handed to the `browser` subagent through `window.__app`; its own scripted player could not reach a win or a loss, it was cut at its 1 h limit (#303), and the judge then played the game again for 29 min. After apply, the review was the largest cost in the item. **Fix** The code review reads the code and runs the lanes; playing the app is the judge's, and the review brief says so. **Done when** A code review on a web game finishes without a browser playtest, and the judge still plays it. _From run 139._
Author
Owner

Confirmed on production (6f440e5, PR #310); closing.

Run 147 (thorough, the Breakout project's item, vLLM) ran its agent.code-review as the new review agent, which has no browser:

  • 10.8 minutes, against run 139's code review, which ran to its 60-minute cut and then 5.1 more.
  • No browser or app calls and no subagents. Nothing served or played the game.
  • What it did: 15 reads, and 6 shell commands (git status, openspec validate --strict, a grep for leftover logging, a lockfile check, a speed calculation in node -e, and two unit-test files). It made 2 small fixes, the starting keypress in src/main.js and a test's speed cap, then wrote its report.

The judge plays the game after it, in a session of its own.

**Confirmed on production (6f440e5, PR #310); closing.** Run 147 (thorough, the Breakout project's item, vLLM) ran its `agent.code-review` as the new `review` agent, which has no browser: - **10.8 minutes**, against run 139's code review, which ran to its 60-minute cut and then 5.1 more. - **No browser or `app` calls and no subagents.** Nothing served or played the game. - **What it did:** 15 reads, and 6 shell commands (git status, `openspec validate --strict`, a grep for leftover logging, a lockfile check, a speed calculation in `node -e`, and two unit-test files). It made 2 small fixes, the starting keypress in `src/main.js` and a test's speed cap, then wrote its report. The judge plays the game after it, in a session of its own.
Sign in to join this conversation.
No labels
No milestone
No project
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference
cmoriarty/braid#307
No description provided.