Add important warning about not pushing/creating PRs unless explicitly asked

[xingyaoww]

Summary

This PR adds an important warning to the <PULL_REQUESTS> section of the system prompt to address the issue where Claude-4 has been too eager to start PRs without being explicitly requested to do so by users.

Changes

• Added **Important**: Do not push to the remote branch and/or start pull request unless explicitly asked to do so. as the first line of the <PULL_REQUESTS> section in openhands/agenthub/codeact_agent/prompts/system_prompt.j2

Context

Based on team discussion, Claude-4 has been ignoring existing prompt guidance about not opening PRs until asked, causing user confusion. This change makes the instruction more prominent with bold formatting and explicit placement at the beginning of the section.

Testing

• Pre-commit hooks pass
• Change is minimal and focused
• Existing functionality preserved

@xingyaoww can click here to continue refining the PR

---

To run this PR locally, use the following command:

docker run -it --rm   -p 3000:3000   -v /var/run/docker.sock:/var/run/docker.sock   --add-host host.docker.internal:host-gateway   -e SANDBOX_RUNTIME_CONTAINER_IMAGE=docker.all-hands.dev/all-hands-ai/runtime:9694cc9-nikolaik   --name openhands-app-9694cc9   docker.all-hands.dev/all-hands-ai/openhands:9694cc9

[xingyaoww]

I'm running an eval locally for this 🙏

[xingyaoww]

This PR:

Instances submitted: 200
Instances completed: 199
Instances incomplete: 300
Instances resolved: 136
Instances unresolved: 63
Instances with empty patches: 0
Instances with errors: 1
Unstopped containers: 0
Unremoved images: 500
Report written to claude-sonnet-4-20250514_maxiter_100_N_v0.46.0-no-hint-pr9357-run_1.20250625_215337.json
MODEL_NAME_OR_PATH: claude-sonnet-4-20250514_maxiter_100_N_v0.46.0-no-hint-pr9357-run_1
RESULT_OUTPUT_DIR: /home/xingyaow/OpenHands-dev/evaluation/evaluation_outputs/outputs/princeton-nlp__SWE-bench_Verified-test/CodeActAgent/claude-sonnet-4-20250514_maxiter_100_N_v0.46.0-no-hint-pr9357-run_1
Checking for changes: 0it [00:00, ?it/s]
Updating output file: 200it [00:01, 101.96it/s]

On main (from #9140 (comment)):

Total instances: 500
Instances submitted: 200
Instances completed: 198
Instances incomplete: 300
Instances resolved: 135
Instances unresolved: 63
Instances with empty patches: 0
Instances with errors: 2
Unstopped containers: 0
Unremoved images: 500

Should be ready for review