Create Aero

Create an aerospace concept

Treat the text after /create-aero as the concept argument.

Resolve the argument

1. Require exactly one concept path. Accept:
   • designs/<name>
   • <name> as shorthand for designs/<name>
   • an existing design.yaml
2. Work from the repository root.
3. Reject .., absolute paths outside this repository, and names that would write outside designs/ or results/.

Scaffold a missing concept

If the resolved concept directory does not exist:

1. Do not copy the template or create the concept directory from this skill. Offer both supported bootstrap paths:

   • For an intent plus requirement documents, sketches, and optionally a reference 3D model (a triangle mesh of the real aircraft), invoke the dedicated initializer so it measures, authors, and runs the bounded geometry checkpoint:

     /initialize-aero <name> "<intent>" @requirements.md @sketches @reference.stl
     

   • For direct visual authoring, start the schema-driven Studio:

     python -m openair.designer new <name> --open
     

2. For the Studio-first path, ask them to write the brief and inference record, trace the sketch photos, and click Create concept. A successful save creates designs/<name>/design.yaml, brief.md with the auto-generated Sketch measurement worksheet, and original/rectified sketch PNGs.

3. For the initializer path, inspect design.yaml .sketch.treatment. requirement mode still requires the reviewed Studio save. inspiration mode may proceed without Studio when the user requests the autonomous/no-manual-review flow; the measured worksheet and hard identity bounds are then the source contract.

4. Ask them to re-invoke /create-aero designs/<name> after saving, or proceed directly in an explicitly autonomous initializer invocation.

5. Stop. Do not run solvers against unconfirmed template requirements, and do not treat an abandoned Studio session as a source concept.

Establish sketch evidence before optimization

Do not start MDO from an image and an undocumented guess. Before the full pipeline:

1. Require a Sketch measurement worksheet in brief.md that records: source image and view; four rectification/control points or an explanation of why none are needed; dimensional scale reference; fuselage length, width, height, and stations; wing span, root/tip chords, sweep, and root location; fin geometry; every inferred value; and measurement tolerances. Use python -m openair.designer <name> --open to revise the brief, four-point graph-paper rectification, tracing, and auto-generated worksheet. The browser's derived values are preflight guidance, not gate evidence.

2. Confirm design.yaml .sketch contains those measured targets and tolerances. A non-reference planform must not inherit the template envelope. When wing.sections is present, require sketch.treatment: reproduction, 3–12 strictly increasing eta stations from 0 to 1, and scalar root/taper/sweep/dihedral values matching WingSpec.equivalent_trapezoid. The measured sections—not the scalar trapezoid—must drive OpenVSP, OAS, and Studio.

3. Run the baseline geometry checkpoint:

   python -m openair.geometry run designs/<name>
   

4. In interactive or requirement mode, open results/<name>/baseline/threeview.png beside every designs/<name>/sketch-*.png. Compare span/length, root and tip chord, sweep direction and magnitude, wing station, body silhouette, and fin placement. If the artifact is visibly outside the worksheet tolerance, update the source design and repeat this checkpoint; do not let MDO hide an input-shape error. In autonomous inspiration mode this is an automated evidence checkpoint, not a request for user approval: require geometry truth and the hard identity bound, then let MDO repair soft-prior mistakes.

5. When the concept carries a reference model (designs/<name>/reference/reference.json, guidebook chapter 13), the baseline checkpoint must also show results/<name>/baseline/geometry.json .reference_fidelity.ok == true (mandatory for reproduction; disclosed for other treatments), and you must open results/<name>/baseline/reference_overlay.png. Departures are acceptable only where the brief lists the feature as unrepresentable. For a sectioned wing, also read geometry.json .wing.planform_mode, .openvsp.readback.wing_sections, and .reference_fidelity.disclosed.p95_model_to_reference_exposed_components_m.wing; this is model-to-reference only. Do not mistake the buried centreline carry-through's component p95 for exposed shape error. When fuselage.fairings is present, also require .openvsp.readback.fairings_match, .openvsp.readback.vtail_root_extensions_match, mesh_checks rows fairing_*_contained and fin_*_attached, and .reference_fidelity.checks.p95_body.basis == "fuselage + measured fairing components". Fairings and fin_*_root are loft-only/non-lifting. The fairing remains in body-union fidelity; verify only buried-root components are marked excluded from component fidelity.

6. Require the concept inputs (brief.md, design.yaml, sketches, and reference/) to be committed before the full run. If they are uncommitted, stop and ask the user to commit them or explicitly authorize a commit.

Run an existing concept

Activate the repository environment before Python starts:

source .venv/bin/activate
export LD_LIBRARY_PATH="$PWD/tools/libs/usr/lib/x86_64-linux-gnu:${LD_LIBRARY_PATH:-}"
export PATH="$PWD/tools/openvsp/opt/OpenVSP:$PWD/tools/su2/bin:${PATH:-}"
export PYTHONPATH="$PWD/src:${PYTHONPATH:-}"
export MAMBA_ROOT_PREFIX="$PWD/tools/mamba"
export OPENMDAO_REPORTS=0
python -m openair run designs/<name>
pytest

The geometry-only call above is the required pre-MDO checkpoint. After it passes (or is manually accepted in interactive mode), do not substitute a sequence of hand-picked stages: the concept orchestrator owns phase ordering, stale-artifact cleanup, optimized-design publication, stretch runs, and presentation generation.

QA review

After a successful run:

1. Read .cursor/skills/aero-qa/SKILL.md.
2. Review results/<name>/baseline/ and results/<name>/optimized/ as different aircraft. Execute every gate in docs/guidebook/00-qa-workflow.md; never infer a verdict from ok: true.
3. Inspect results/<name>/optimized/threeview.png and the interactive mesh in results/<name>/report.html. Confirm the displayed artifact agrees with the configuration and mesh-truth checks.
4. Recompute at least one headline value independently.
5. For inspiration concepts, inspect baseline/mdo.json:
   • all branches, including airfoil selection and any htail fallback;
   • a converged calibration_history with NP error ≤ 0.05 MAC and stable dash;
   • fidelity_sweep weights 2.0/1.0/0.25 with weight 1.0 delivered;
   • every beyond-tolerance row in sketch_departures, including its clamp study reason; and
   • fin volume plus fin_te_overhang_m <= 0.
6. Report gate, evidence path + JSON key + value, and verdict for any failure. Distinguish TACS/SU2 calibration results from flight-worthiness gates.

Finish with a concise concept verdict, headline performance, validation count, the optimized three-view image, and links to:

• results/<name>/report.html
• results/<name>/executive_brief.pdf
• results/<name>/optimized/design.yaml

When publishing a new committed preview, add its curated title, summary, kind, and optional featured status to site/designs.yaml. The Pages build is deliberately fail-closed when that manifest and the results/*/report.html set differ.

To make the delivered geometry the explicit source for another iteration, use python -m openair promote <name> <new-name>; never copy generated YAML over the current source concept.