AEDT integration¶
aster does not talk to AEDT directly. Real runs go through the companion guard system
ansys-aedt-mcp, which owns the AEDT session, the
solve budget and the evidence trail.
What the guard provides¶
- Explicit session targeting: every call takes exactly one AEDT PID or gRPC port; no implicit sessions and no fallback launch.
- A preflight/audit layer, solver budget accounting, archive and export gates.
- Touchstone and far-field export with the artifacts kept on disk.
The guard pins PyAEDT 1.10.0 and is validated on AEDT 2026 R1. From guard v0.3.2 it also
enforces managed-only sessions (unmanaged AEDT is refused and reported), denies arbitrary broker
code by default, and binds run tokens to project/design/setup when AEDT_GUARD_ENFORCE=1.
See runtime safety for the boundary.
aster's own trusted bridge enables the guard's arbitrary-code opt-in for its own broker call
(InProcessBridge sets it only while calling the broker and restores it; StdioBridge injects it
into the child environment) because measured runs execute modelling snippets. The DSH/Codex MCP
server processes are separate and keep the guard's default-deny gate; do not co-host them with
aster in one process, and run aster only from a trusted operator account. PyAEDT documentation:
https://aedt.docs.pyansys.com/version/stable/index.html.
Configuration¶
aster discovers the guard checkout and its interpreter through Environment (see
aster/config.py). Values can be set with ASTER_* environment variables or a config file:
| field | meaning |
|---|---|
aedt_exe |
path to ansysedt.exe |
guard_repo |
checkout of ansys-aedt-mcp |
venv_python |
<guard_repo>\.venv\Scripts\python.exe |
reference_project |
the .aedt project copied for each candidate |
campaign_root |
writable campaign directory |
aedt_version |
default 2026.1 |
python -m aster reports missing fields before a campaign starts; validation is explicit rather
than a stack trace halfway through a solve.
Run lifecycle¶
- Preflight — verify the reference project and target design exist and the session is the one the caller named.
- Model — copy the reference project, apply the candidate's parameter values to the AEDT design.
- Solve — start the analysis with the guard's budget accounting.
- Post-process — export S11 (Touchstone) and the far-field grid; store both under the run directory.
- Score — evaluate the current acceptance rules from the exported artifacts.
- Record — append the run to the campaign SQLite archive with status, metrics and paths.
Failed solves leave a failed run record; they are never silently dropped. A stale solver process is handled by the guard, not by the design layer.
Evidence¶
Each real run directory keeps:
<run-id>/
├── artifacts/result.s1p
├── artifacts/farfield_views.json
├── artifacts/native_reports/*.png
└── run.json
The archive stores the metric values that were computed from these files, so an acceptance change can be applied later without a new solve (provided the recorded data is still compatible with the task).