Agent Workflow Guardrails

Core Rules

• Always prefer the smallest possible change that fixes the root cause.
• Avoid broad refactors unless strictly required to unblock correctness.
• After each patch, re-run the minimum relevant validations, then full flow when needed.
• For every bug, first do a deep root-cause audit with evidence. Do not ship a patch that only masks symptoms or "seems to help" without identifying the actual source of the bug.

Parallel Execution Rules

• Never wait idly on subagents.
• While a subagent is running, always propose and run useful parallel tasks.
• Parallelize everything that can be safely parallelized via subagents (audit, tests, repo checks).
• If waiting is unavoidable, ask what additional parallel work should be done meanwhile.

Standard Deep-Audit Pipeline

• This flow is executed by Codex + subagents (not by local shell automation scripts).
• Cross-repo checks are diff-scoped only: validate changed files and their direct integration paths, not full-repo audits by default.

1. Identify change scope and impacted modules.
2. Run deep audit on the diff.
3. Add/adjust tests that cover the changed behavior (unit first, then integration/e2e on devnet when relevant).
4. Execute validations.
5. If a step fails, apply a minimal patch and restart the pipeline from audit.
6. Run cross-repo validation only on impacted paths across:
   • 8004-solana-indexer
   • agent0-ts-solana
   • 8004-solana
   • 8004-atom
7. Only finalize when all checks are green.