Comparing changes

Packages the existing CLI behind a single POST /convert endpoint so callers can import .ork files without carrying a JDK, a JVM or a 70 MB OpenRocket jar in their own image. Additive only: nothing under rocketserializer/ changes. - One subprocess per request. OpenRocketSession.__exit__ never calls shutdownJVM() (jpype1<1.5 cannot restart one), and __enter__ silently reuses an existing classpath when a JVM is already live, so a warm worker would eventually parse against the wrong OpenRocket version. A fresh process also contains JVM segfaults and isolates cli.py's import-time serializer.log write. - Inlines thrust_source.csv and drag_curve.csv into the response. Raw parameters.json references them by host-relative path (backslash-separated on Windows), so it is not self-contained; callers get one object instead of a file lifecycle. - Reports serializer_version (package version + pinned jar) so a bad parse can be traced to the parser that produced it. - Maps the known user-fixable failures (no simulation data, non-English file) to 422 with stable codes rather than leaking tracebacks. - Concurrency capped: every job boots a JVM and loads OpenRocket's full preset database. Dockerfile pins OpenRocket-23.09 on Temurin 17 and forces java.awt.headless, since OpenRocket's startup path constructs a Swing GuiModule.

Expands service/test_app.py from 12 to 64 unit tests across _read_pairs, _classify, _inline_csv, _safe_name, the subprocess contract, the envelope, the error map, the HTTP layer and the concurrency cap. Subprocess is stubbed throughout, so the suite needs no JVM and no jar and runs in ~0.7s. Writing them surfaced four real defects, each fixed and each verified by reverting the fix and confirming the suite goes red: - A filename of ".." escaped the job directory. Path("..").name returns ".." unchanged, so job_dir / ".." resolved to the temp root and write_bytes raised an unhandled IsADirectoryError. Path().name also does not split on backslashes under POSIX, so a Windows client's "C:\a\b.ork" arrived whole. Extracted _safe_name() to normalise separators and reject dot-only names. - _inline_csv crashed on a null section. ork_extractor wraps lookups in _safe_search and can emit null rather than a dict, which made parameters.get(section, {}).get(key) raise AttributeError. Same fix applied to the id.filepath strip. - Malformed ork2json output escaped as a traceback. A non-numeric CSV cell or truncated parameters.json now returns 500 parse_failed. - The upload size limit was enforced only after the body was read, so an oversized request was pulled into memory just to be discarded. Now checked against the declared size first, with the post-read check kept as a fallback for clients that send none.

Verified end to end for the first time: image builds, container serves, and examples/Anonymous--Mu/rocket.ork converts in 11s (EPFL--BellaLui in 6.5s) with both CSVs inlined and no temp directories left behind. Three problems only a real build could surface: - jpype1 is pinned <1.5, which predates aarch64 wheels, so pip compiles it from source on arm64 and the image had no C++ toolchain. build-essential and python3-dev are now installed and purged inside a single layer. - jammy ships pip 22.0.2, whose setuptools predates the PEP 660 build_editable hook, so `pip install -e .` failed outright. pip/setuptools are upgraded first, and the package is installed non-editable — an image has no reason to hot-reload it. - A corrupt or non-OpenRocket upload returned 500 with a raw Python traceback. parse_ork_file raises "The .ork file is not in UTF-8" for both a genuine encoding problem and any file that is not an .ork archive, so it now maps to 422 invalid_encoding with copy covering both readings. README documents the arm64/amd64 split: a build on Apple Silicon produces an image that will not run on an amd64 host.

Navigation Menu

Uh oh!

Comparing changes

Open a pull request

Uh oh!

Commits on Aug 6, 2026

This comparison is taking too long to generate.

Uh oh!