| name | smart-compact-v6 |
| description | Reduce Codex context replay and visible output with an exact acceptance contract, scoped edits, concise handoffs, and normal safety rigor. |
Smart Compact v6
Optimize total task context without compressing requirements, code, commands, or hidden reasoning.
For a bounded local task with a complete specification, one owner, an exact target root, and a supplied acceptance command:
- Do not create a plan or reread instructions already in context.
- Read only the specification and target state needed to implement. Do not inspect sibling solutions, prior benchmark arms, memory, or the acceptance harness source.
- Before each edit, verify every path starts with the exact target root. Never write beside or above it.
- Implement in one coherent patch when practical. Split only for tool limits or a diagnosed failure.
- Execute the supplied acceptance command verbatim. Ad hoc smoke tests do not replace it.
- If it fails, inspect only the reported target, patch only the cause, and rerun the same acceptance command.
- After acceptance passes, run one scoped status check and stop. Do not add confidence tests or process narration.
Honor active shell wrappers. If AGENTS.md requires RTK, start every exec_command command with literal rtk, use rtk proxy when needed, and retain the prefix in acceptance commands. Otherwise do not assume RTK is installed.
These are workflow defaults, not a hard call budget. Expand when failure, ambiguous requirements, destructive impact, security, permissions, production, or high-stakes risk requires it.
Communicate only the outcome, changed paths, decisive verification, blocker, and material residual risk. Preserve exact identifiers, numbers, values, negation, conditions, and order.