mirror of
https://github.com/google-gemini/gemini-cli.git
synced 2026-05-12 12:54:07 -07:00
2.3 KiB
2.3 KiB
Offload Skill
This skill provides a high-performance, parallelized workflow for offloading intensive developer tasks (PR reviews, fixing CI, preparing merges) to a remote workstation. It leverages a Node.js orchestration engine to run complex validation playbooks concurrently in a dedicated terminal window.
Performance Workflow
The offload skill supports the following specialized playbooks:
review(default): Clean build, CI status check, static analysis, and behavioral proofs.fix: Build + Log analysis of CI failures + Iterative Gemini-led fixing and pushing.ready: Final full validation (clean install, full preflight, and conflict checks).open: Provision a worktree and drop the user directly into a remote shell/tmux session.implement: Read an issue → Research → Implement → Verify → Create PR.
Workflow
1. Initializing an Offload Task
When the user asks to offload a task (e.g., "Offload PR 123 fix" or "Make PR 123 ready"), use the run_shell_command tool to execute the orchestrator:
- Command:
npm run offload <PR_NUMBER> [action] - Action: This will sync scripts to the remote host, provision a worktree, and pop a new terminal window for the playbook dashboard.
- Response: Inform the user which playbook has been launched.
2. Monitoring and Synthesis
The remote worker saves all results into .gemini/logs/offload-<PR_NUMBER>/. Once the playbook finishes, the agent should synthesize the results:
- Read logs corresponding to the playbook tasks (e.g.,
build.log,review.md,test-execution.log,diagnostics.log). - Check the
.exitfiles to confirm success of each parallel stage.
3. Final Recommendation
Provide a structured assessment based on the physical proof and logs:
- Status: PASS / FAIL / NEEDS_WORK.
- Findings: Categorized by Critical, Improvements, or Nitpicks.
- Conclusion: A clear next step for the maintainer.
Best Practices
- Isolation First: Always respect the user's isolation choices (
~/.offload/gemini-cli-config). - Be Behavioral: Prioritize results from live execution (behavioral proofs) over static reading.
- Multi-tasking: Remind the user they can continue chatting in the main window while the heavy offloaded task runs in the separate window.