fix: float precision and parameter-level x-enum-varnames - #42
Merged
lightsofapollo merged 2 commits intoJul 27, 2026
Conversation
The remaining findings from live-testing a generated client against the RunPod v2 API. - `format: float` now maps to f64 rather than f32. JSON carries no binary32, so the declared format describes the server's storage, not the transport: a price sent as 0.03 survives in f64 but becomes 0.029999999329447746 through f32. RunPod declares float on catalog prices and double on billing, so the same money took two different precisions depending on endpoint. Opt out with float_precision = "f32"; --types-conservative keeps the literal mapping. - Parameter-level inline enums honor x-enum-varnames. Schema-level enums already did, so an identical enum produced different Rust variant names depending on whether it lived in components.schemas or on a parameter. A varnames array whose length disagrees with enum is ignored rather than applied to a prefix, since the pairing would be a guess. problem_details() is deliberately unchanged. It decodes the RFC 9457 profile, identified by Content-Type: application/problem+json; RunPod returns plain application/json, so None is correct. Loosening the check would misparse arbitrary JSON error bodies as Problem Details. Its doc comment now explains why third-party APIs yield None and points at `typed`/`body` instead. Verified: 369 tests pass, clippy and fmt clean, and the full 54-spec corpus generates and compiles against its own dependency fragments. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
full-spec-compile has been failing with SIGTERM (143) roughly 25 minutes into a 240-minute budget, on main as well as on branches, since well before the current fixes. It is not a timeout and not disk. Measured cause: microsoft-graph generates 2.4M lines (16,153 operations) and peaks at ~14.3 GB RSS in a single rustc process during cargo check. A hosted ubuntu-latest runner has 16 GB, so it is killed. CARGO_BUILD_JOBS=1 changes the peak by 0.1% — the memory is one rustc type-checking one crate, so parallelism and sharding do not help. Adds GENERATE_ONLY_SPECS, currently just microsoft-graph. It is still generated, which is where most generator defects surface; only the compile step is skipped. SPEC_COMPILE_FORCE_CHECK=1 runs it anyway where there is headroom. Generate-only specs are counted separately and never folded into `passed`, and the success line degrades to "N compiled cleanly; M generated but not compiled" — a green run must not read as full verification when one spec was never compiled. Full corpus: 53 passed, 0 gen-failed, 0 check-failed, 1 generate-only, 1 skipped (gitea, Swagger 2.0). Refs: openapi-generator-lnj Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
lightsofapollo
changed the base branch from
fix/live-tested-generator-defects
to
main
July 27, 2026 03:37
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Stacked on #41 — the remaining three findings from live-testing a generated client against the RunPod v2 API. Base this on #41; merge that first.
format: floatnow maps tof64(behavior change)JSON carries no binary32. The declared format describes the server's storage, not the transport, so mapping it literally discards precision the response actually carried:
This is a live observation, not a hypothetical: RunPod declares
floaton catalog prices anddoubleon billing amounts, so the same money took two different precisions depending on which endpoint you called.Filed under Changed rather than Fixed because it alters generated APIs. Opt out with
float_precision = "f32"under[generator.types];--types-conservativekeeps the literal mapping. There's a test asserting the actual round-trip loss so the rationale in the docs can't silently rot.Parameter-level inline enums honor
x-enum-varnamesSchema-level enums already honored the extension through
SchemaAnalysis::enum_extensions, but parameter enums are inline and have no analyzed-schema name to key on, so they silently fell back to the naming heuristic. An identical enum therefore produced different Rust variant names depending on whether it lived incomponents.schemasor on a parameter.Plumbed through a new
ParameterInfo::enum_varnames. A varnames array whose length disagrees withenumis ignored rather than applied to a prefix — the pairing would be a guess. Serde renames still target the wire value, and suffix disambiguation still applies.problem_details()— intentionally unchangedReported as "can never succeed against this API," but that is correct behavior, not a defect.
The helper decodes the RFC 9457 profile, which is identified by
Content-Type: application/problem+json. RunPod returns its errors as plainapplication/json, soNoneis the right answer. Loosening the media-type check would let arbitrary JSON error bodies be misinterpreted as Problem Details — a worse bug than the one reported. Servers generated by this tool always emit the problem media type, so it works against them.The real gap was discoverability, so the generated doc comment now explains why third-party APIs yield
Noneand points attyped/bodyinstead. Happy to add opportunistic parsing behind a flag if that's wanted.Verification
cargo clippy --all-features -- -D warningsclean,cargo fmtcleangitea, Swagger 2.0). Run locally, since the float change touches every spec with a numeric field and the per-PR tier only compile-checks OpenAI and Anthropic.One pre-existing test asserted
format: float→f32for query parameters; updated tof64with the rationale inline, since that is precisely the behavior being corrected.🤖 Generated with Claude Code