[dinwwwh]

🤖 Generated with Claude Code

https://claude.ai/code/session_01VVaw7RYWCsn8myBE4Rncuq

What

#356 keeps the null of an anyOf pair whose $ref leads back to it through other such pairs, because the bare $refs that would result match nothing. Its walk missed one case: a pair nested in the target's own pair, such as

T:
  anyOf:
    - anyOf: [{ $ref: '#/components/schemas/T' }, { type: 'null' }]
    - type: 'null'

It also missed the same shape with oneOf inside, or with annotations or keywords 3.0 drops beside the inner pair (examples, $comment, if, else, propertyNames, ...). In 3.1 these admit null. The 3.0 output was T: { $ref: '#/components/schemas/T' }, which drops the null and which Ajv cannot compile (Maximum call stack size exceeded), though it compiles the input.

The walk now also looks inside a branch that is itself a union holding a null branch. So T keeps its pairs, as #356 does for the flat loop. Returning "doesn't admit null" only keeps a null branch, so a wrong answer could cost a redundant branch but never a value.

Scope

This only touches non-consuming $ref loops, which are degenerate. Re-converting today's fuzz dumps (about 140k docs from several generators) changed 0 to 4 docs per 20k, and every one of them was T: { $ref: T } before. On the loop-focused dumps (12k docs), jsonschema-rs still reports 0 lost.

Tests

The #356 test now adds a nested anyOf pair, a nested oneOf pair, a nested union without null (left as is), and a boolean branch. The new asserts fail on 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 (96fe14d) with main (25774db)}

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]

Important

nonNullBranches recurses into nested unions with no visited guard, so a programmatic input whose object graph has an identity cycle (or shared nested arrays) through a null-bearing union makes refAdmitsNull overflow the stack. The same input converts on main, and object-identity cycles are otherwise supported by this converter — details inline.

Reviewed changes

• Follow a null pair nested in a null pair — refAdmitsNull now flattens a branch that is itself a union holding a null branch via the new nonNullBranches, so the loop walk reaches the inner $ref and keeps the pair instead of collapsing it to a bare $ref (packages/downgrader/src/v3.1-to-v3.0.ts).
• Tests — the #356 loop test gains nested anyOf/oneOf cases, a null-free nested union (left whole), and a boolean branch (packages/downgrader/tests/v3.1-to-v3.0.test.ts). All 137 tests in the file pass; uncheck on both files is green.

  ^{｜ Fix all ➔ ｜ Fix 👍s ➔ ｜ View workflow run ｜ Using deepseek-v4.1-flash (free via Pullfrog for OSS) ｜ 𝕏}

[pullfrog]

✅ No new issues found. The cycle guard resolves the prior review's concern and does not affect parsed (acyclic) input.

Reviewed changes

• Guarded the nested-union walk against cycles — nonNullBranches now threads a seen Set of the unions already flattened and returns [] on a revisit, so an identity cycle through null-bearing nested unions terminates instead of overflowing the stack (packages/downgrader/src/v3.1-to-v3.0.ts:986).
• Added a cyclic-input test — a hand-built B = { anyOf: [{ anyOf: [B, null] }, null] } behind a $ref now converts instead of throwing (packages/downgrader/tests/v3.1-to-v3.0.test.ts:1394).

The guard only fires on identity repeats, which parsed JSON (a tree) never produces, so behavior for real documents is unchanged — it matches the repo's existing cycle-guard convention. Verified locally: pnpm vitest run packages/downgrader/tests/v3.1-to-v3.0.test.ts passes (137 tests) and uncheck is green on both files.

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