Cad sim design orchestrator

CAD Sim Design Orchestrator

Author the first reviewable candidate only after intake has produced a complete intake/project_brief.yaml. Keep this core loaded. Use the conditional contract map below instead of preloading every advanced policy.

Invocation Order

1. Use $cad-sim-project-intake unless the complete brief already exists.
2. Use this skill to create source/CAD, one canonical annotation_view/, handoff.json, current evidence, and the brief-required execution packet.
3. Use $cad-sim-review-revise-loop from the same Direct Meta-Agent when iterative review, maturity promotion, or autonomy-before-user-review is in scope. Review-ready states are internal, not user-visible completion.

Once a current candidate and handoff exist, do not use this skill as the main revision loop.

Core Invariants

• Work only in the active fresh writable root. Treat old projects and any legacy or comparison framework checkout as read-only references.
• Do not start CAD from a vague prompt. Missing user answers, material interfaces, or rubric intent route back to intake or a precise needs_user.
• In an ordinary real project, the calling thread is the user-facing Direct Meta-Agent: the combined Manager and primary author. Do not launch an author subagent, replacement author, emergency builder, or manager-fallback author. The same Direct Meta-Agent launches fresh independent reviewers and owns revision, route promotion, and closeout.
• ready_for_independent_review, review_pending, revision_complete_pending_review, and ready_for_fresh_re_review are internal transitions. An External Controller exists only when a rehearsal or benchmark explicitly tests split topology.
• A split-topology run has one author role and one settle cycle. If that role cannot publish, route blocked with reason_code: blocked_meta_agent_authoring; do not create a rescue author chain or score controller fallback output as normal provenance.
• For bounded goal runs, reserve the final 35 percent of tokens and at least 15 minutes for full-surface audit, independent review, schema repair, route promotion, and full preflight.
• In source-blind or comparator-sensitive work, write source_boundary_preflight.yaml before reading memory, controller work/workflows, prior runs, comparator/raw material, or local framework files. Authoring roles never use controller state, including at closeout.
• Seed missing artifact_contract/README.md and viewer3d/ only from this installed skill's bundled resources. Missing resources route missing_project_support; never reach into the reset source repo.
• Stop before unauthorized heavy solvers, commercial/cloud/GPU work, large installs, physical execution, or safety/fit/fatigue/production claims.
• Viewer validity and native traceability are necessary, not sufficient. Use actual render-mesh/native evidence and a three-view visual read; schematic plots cannot prove a reviewable product concept.
• For master's-level work, keep architecture_and_behavior_integrity open until the configured release state has independent inputs, a continuous applicable path, complete expected-contact/required-clearance/forbidden contact coverage including environment/reference geometry, analysis-ready loads/joints/boundaries/scenarios, and a governing alternative when required. Apply profile semantics for physical products, robotics/electromechanical, simulation/optimization, and software-connected hardware.
• Respect the brief's target_maturity, autonomy_after_grill, autonomy_envelope, stage_frontier, review_contract, analysis/tooling permissions, and stop budgets. These contracts permit bounded work; they do not authorize invented facts or broader objectives.
• Before serious primary-surface polish, use $cad-sim-quality-harness. New harnesses use schema_version: 2, an acceptance_profile, every mandatory profile row, at most three run-specific risks, and generated quality/quality_harness_results.json. Keep quality/quality_harness.yaml authoritative. Use quality/quality_reset_seed.yaml only for the existing clean-restart route.
• A website/report/notebook/dashboard or equivalent front door is the user-facing product. It must contain readiness-critical instructions, evidence, decisions, states, boundaries, and the next action at point of use; Markdown and side files may be audit/download evidence but not missing UI.

Conditional Contract Map

Detailed behavior remains authoritative in references/full_authoring_contract.md. Load only the activated headings or search clusters from its Search Map:

• Source-blind, delegated, comparison, recovery, or goal-budget work: load the matching role/source-boundary/settle rules.
• Master's/A+, interface-heavy, robotics, simulation, or hybrid work: load the architecture, maturity, Quality Harness, decision, alternative, and verifier rules.
• Website, prototype, build/use, interactive, or evidence-promotion work: load the Primary Outcome Surface, action-grounding, runtime, and handoff rules.
• Tool-sensitive, external-capability, optimization, or new-stage work: load the tooling, provenance, analysis, and stage-frontier rules.
• For exact files, field expectations, commands, and final reporting: load the reference's Inputs, Minimal Study Files, Workflow, and Output Standard headings as needed.

Do not copy the reference into the run. It is procedural guidance, not project evidence.

Core Workflow

1. Confirm the brief and, when required, the source-boundary preflight. Restate the goal, maturity, interfaces, assumptions, approved stage, autonomy, analysis/tooling envelope, review contract, and stop budgets.
2. Create authoring_status.yaml before heavy work. Record route, role/thread identity, last_progress_at, next_file, expected outputs, published files, and any blocker; advance it whenever a declared intermediate appears.
3. Choose the smallest useful tools and evidence. Honor project_brief.yaml.tooling_strategy before the first affected design or simulation decision. Use a separate $cad-sim-tooling-research-scout when its policy is required_before_first_design; use it at the first material trigger when allowed_when_triggered; and do not launch it when forbidden. Keep lightweight decisions inline only when the named decision owner allows that topology. Record tooling/tooling_scout.md even when the decision is use_default_stack, no_extra_tooling, or no simulation, including the agent-operability assessment, simulation decision/fidelity boundary, and evidence-export plan required by the brief.
4. Build one coherent baseline before variants or optimization. Preserve reconstructive source, units, parameters, imported-reference provenance, and a compact claim boundary.
5. Run the cheapest configured-state checks capable of falsifying the governing interfaces, behavior, clearances, contacts, connections, action states, and release representation. When the selected Quality Harness profile requires it, publish and check quality/configured_project_integrity.yaml: bind each selected identity across at least two current surfaces, and either verify every route-controlling transition against current implementation/receipt evidence or mark a passive artifact not_applicable with a rationale. This applies to replacement parts, bags/racks, configurable tools and printers, robotics, and software-connected hardware without pretending the check is physical validation. For robotics/electromechanical work, also publish and check quality/system_interface_manifest.yaml.
6. Create one canonical annotation_view/. Use the artifact contract and viewer3d enrich-topology only when approximate overlays materially improve review; never present them as kernel-accurate proof.
7. For build/use or prototype maturity, include bought/standard parts, tools, manufacturing/print assumptions, assembly/install/remove/service actions, inspection, first tests, reject conditions, and evidence limits. Map each readiness-critical action to visible geometry, hardware, access, sequence, and exact output files.
8. Create one obvious Primary Outcome Surface and its annotation_view/primary_outcome_surface_brief.yaml. Embed or summarize readiness evidence at point of use, consolidate before expanding, and keep only true real-world action inputs in the required Use list.
9. Write handoff.json with exactly one canonical_annotation_view. Declare other bundles noncanonical or remove them. Refresh iter_NNN labels.
10. Validate with PYTHONPATH=. python -m viewer3d handoff <root> --json; add --require-native-model for comparison replay. Record the exact command and result. Use viewer3d serve for canonical interactive CAD inspection.
11. Perform the full-surface and three-view self-audits, run all declared local verifiers, and distinguish nominal/source checks from downstream release evidence and physical/external boundaries.
12. When iterative refinement is in scope, enter $cad-sim-review-revise-loop in the same Direct Meta-Agent, launch a fresh reviewer, and continue to a validated terminal route. Do not return an internal review-ready state to the user.

Output

Report the changed files, brief and unresolved assumptions, canonical handoff and viewer validation, freshness warnings, visual/mechanical or system feasibility, Quality Harness and interface status, Primary Outcome Surface and launch path, action/build-use/prototype closure, analysis/tooling provenance and fidelity limits, autonomy/stage/review route, whether old projects were modified, and the exact next bounded action. Never make the final claim stronger than the current digital, physical, or external evidence.