Conversation
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
![cursor[bot]](https://avatars.githubusercontent.com/in/1210556?s=60&v=4){kind=small}
This comment was marked as outdated.
This comment was marked as outdated.
Sorry, something went wrong.
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| with sync_playwright() as p: | ||
| browser = p.chromium.launch() | ||
| page = browser.new_page( | ||
| viewport={"width": DISPLAY_WIDTH, "height": DISPLAY_HEIGHT} | ||
| ) | ||
| page.goto(url, wait_until="networkidle") |
There was a problem hiding this comment.
Every call to process_tool_use launches a new Playwright browser and page for the requested action (lines 123‑128) and then disposes them when the function returns. This means each click, key press, or scroll happens on a fresh instance of the game and is immediately discarded, so subsequent screenshots never reflect the effects of earlier tool actions. The agent therefore cannot make observable progress in the Heatseeker game. The tool handler should reuse a persistent browser/page across iterations and only capture new screenshots after actions have been applied.
Useful? React with 👍 / 👎.
| if action == "screenshot": | ||
| result = "Screenshot taken" |
There was a problem hiding this comment.
When the model requests a screenshot action, the handler returns the literal string "Screenshot taken" (lines 135‑136) and this string is forwarded as the tool result. Claude’s computer‑use API expects the tool result for a screenshot to contain the captured image so the model can view the updated screen. Without providing that image payload, the model cannot see the game state after actions and the conversation will stall. The handler should capture the screenshot and return it as an image block in the tool result.
Useful? React with 👍 / 👎.
| browser.close() | ||
| page.wait_for_timeout(1000) |
There was a problem hiding this comment.
process_tool_use closes the browser and then calls page.wait_for_timeout(1000) on the same page object (lines 171‑172). After browser.close() the page’s channel is destroyed and Playwright raises a Target closed error when any method is invoked, so every tool invocation will fail. Delay should be implemented before closing the browser or via time.sleep outside Playwright.
Useful? React with 👍 / 👎.
1 participant
Note
Introduces a Claude 4.5 computer-use agent to autonomously play Heatseeker, with full docs, tests, packaging, and Docker support; removes obsolete
claude.md.ComputerUseAgentinmodels/claude/src/agent.pyusing Anthropic Claude 4.5 + Playwright to screenshot, analyze, and act (screenshot,click,type,key,scroll) on the Heatseeker game.MODEL,COMPUTER_USE_BETA,DISPLAY_WIDTH/HEIGHT, max iterations) and CLI entry viapython -m src.agent.docs/computer_use/claude-4_5-api.mdandmodels/claude/README.mdwith setup, config, API reference, troubleshooting, and Docker usage.models/claude/pyproject.toml,pytest.ini, andDockerfilefor local/dev installs and containerized runs.models/claude/tests/covering initialization, tool actions, loop flow, and config.claude.md.Written by Cursor Bugbot for commit a64a985. This will update automatically on new commits. Configure here.