armpro24-blip/cad-cae-copilot/aieng-closed-loop-copilot

aieng-closed-loop-copilot

MCP-first closed-loop discipline for recommend → edit → re-simulate → compare. Use this only when a baseline project has CAD, CAE setup/results, and explicit design targets or user-stated metrics.

Purpose

• Bound design iteration before changing geometry.
• Prefer safe parameter edits over full regeneration when editable parameters exist.
• Re-run CAE before claiming improvement.
• Stop on target met, no evidence, approval denial, or budget exhaustion.
• For repeatable loops, follow the relevant entry in

../engineering_skill_contracts.json: cad-mod-propose-verify, solver-run-orchestrate, design-target-review, or evidence-report-synthesize.

Required workflow

1. Define the target, metric, and iteration budget (default: 3).
2. Inspect baseline with aieng.agent_context, aieng.inspect_package, and result/design-target summaries when present.
3. Discover editable dimensions with cad.list_editable_parameters and inspect local/global scope.
4. Choose one change per iteration. Prefer cad.edit_parameter; use cad.execute_build123d, cad.replace_part, or cad.remove_part only when geometry must change structurally.
5. After CAD mutation, inspect regression_diff (topology drift) and critique_diff (a fail/warn verdict means the edit worsened manufacturability) and refresh semantics if needed.
6. Run cad.critique for engineering parts before CAE claims.
7. Prepare CAE with cae.prepare_solver_run, generate deck with cae.generate_solver_input if needed, then call cae.run_solver only through approval.
8. Extract metrics with cae.extract_solver_results, optionally cae.extract_field_regions, and refresh summaries before comparing.

Hard rules

• Respect [APPROVAL REQUIRED] tools; if AIENG_MCP_BLOCK_APPROVAL_TOOLS=1 blocks mutation/solver execution, report the block and stop.
• Do not chain multiple CAD changes before re-simulation unless the user explicitly accepts the loss of attribution.
• Do not claim improvement until post-change solver/result evidence exists.
• Read the credibility tier on every result: executed_solver_result outranks proxy_assembly_result outranks surrogate_prediction outranks critique_finding. Compare like-for-like and never present a lower tier as solver-verified.
• Surrogate proposals are advisory, not evidence: never report a predicted number without its envelope — each prediction carries uncertainty_std / predicted_score_band, and the proposal set carries a leave-one-out validation error band. They guide search only and never substitute for a solver rerun.
• A critique pass or parameter range check is not production certification.
• Stop if target is already met, no safe editable variable exists, no solver evidence can be produced, approval is denied, or budget is exhausted.

Response contract

Report: baseline evidence, chosen change and reason, approval status, post-change CAD/CAE evidence, metric comparison, termination reason, and limitations.