Coding And Data

Data Tasks

• If the data source is clear, include it in the background-session prompt.
• If the data source is unclear, resolve that first instead of delegating blindly.
• For warehouse or SQL work, include connector names, schema hints, date ranges, and the exact question.
• For file-based analysis, include file paths and desired outputs such as charts, CSVs, or summaries.

Parallel Work

If the user explicitly wants multiple reviewers or multiple technical angles:

• start multiple background sessions in parallel
• give each one a distinct goal
• read the results back and synthesize agreements, disagreements, and unique findings

Code Review Pattern

For substantial PR review:

• extract owner, repo, and PR number or use the full PR URL
• start parallel background sessions if multiple reviewers are desired
• give each session the PR context and review objective
• after completion, synthesize findings into:
  • summary
  • agreements
  • disagreements
  • unique findings
  • final recommendation

If the user wants comments posted back to the PR, use the real GitHub CLI flow only after synthesizing the review.

Post-Completion

After a background coding or data session completes:

• read the session output carefully
• read any generated files or artifacts it produced
• summarize what was done, what was verified, and any important decisions
• present deliverable files with show

Do not blindly rerun the whole task unless the result itself indicates a failure or missing verification.