[dinwwwh]

🤖 Generated with Claude Code

https://claude.ai/code/session_01VVaw7RYWCsn8myBE4Rncuq

Problem

Extensions pass through unconverted. When a $ref points into one, the target was left as written and never converted. This happens at the document root (#/x-...), in components (#/components/x-...), or under a key that is already x- right under a component schema. So 3.1 syntax such as type lists, const, type: "null" and examples reached the 3.0 output.

Real 3.1 documents do this:

• Stoplight's bundler puts every bundled file under #/x-ext/<hash>. Examples are GoDaddy domains-v3 (119 refs) and cloudflare/nimbus fixtures.
• CMSIS csolution-rpc uses #/x-jsonrpc-envelope, whose jsonrpc is a const.
• datamodel-code-generator has fixtures with #/x-shared/... and #/x-responses/... unions of const and null.

This also breaks the loosen-only rule. For example, take oneOf: [{ $ref: '#/x-ext/tag' }, { type: 'null' }] with x-ext.tag: { const: 'a' }. A 3.0 reader ignores const, so null matches both branches, and the 3.0 output rejects a null that the input accepts. In a fuzz of grouped and extension refs checked by the Go santhosh-tekuri oracle, this lost 9 accepted values out of 11.9k. With the fix it loses 0.

Change

isInlined now treats these $refs as pointing at unconverted content and replaces them with their converted targets, the same way $refs into $defs or removed parts are handled. The NocoDB check no longer exempts keys that are already x-. Recursion is cut to {} at the back reference, as for $defs. The extension itself stays as written.

The README states the rule. It also extends the existing size note: each such $ref repeats its target once serialized. GoDaddy domains-v3's 3.0 output grows from 97 KB to 277 KB this way.

Checks

• Tests: 312 pass, and coverage of packages/downgrader/src is 100% for statements, branches, functions and lines.
• Stryker on isInlined: 69 of 70 mutants are killed. The 1 survivor is the known equivalent name === 0 || end === -1 guard.
• Outputs on all 72 fuzz dumps (1.12M docs) are identical to main.

---

Generated by Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[codspeed]

Merging this PR will not alter performance

✅ 29 untouched benchmarks
⏩ 2 skipped benchmarks¹

---

_{Comparing claude/project-thread-rixcs1 (92d20db) with main (55d11ff)}

Footnotes

1. 2 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩

[pullfrog]

✅ No new issues found.

Reviewed changes
This run reviewed the net diff of PR #376 against main: the isInlined change in packages/downgrader/src/v3.1-to-v3.0.ts, the new and updated tests in packages/downgrader/tests/v3.1-to-v3.0.test.ts, and the README note.

• Inlining refs into extensions — isInlined now treats #/x-, #/components/x-, and x--prefixed keys right under a component schema as unconverted, so a $ref into them is replaced by its converted target. The old renamed flag's !key.startsWith('x-') exemption is removed.
• Cycle handling — the inline helper cuts back-references to {} via ctx.inlining, so a self-referential extension target terminates (covered by the new Loop case).
• Tests and docs — the new extensions suite covers root #/x-ext, #/components/x-schemas, parameter refs, and cycles; the NocoDB-style test now pins that an x-note extension stays as written while a $ref into it inlines to the converted schema. The README states the rule and the serialized-size cost.

  ^{｜ View workflow run ｜ Using deepseek-v4.1-flash-1017 (free via Pullfrog for OSS) ｜ 𝕏}