[pullfrog]

Fixes #28.

_harmonizeMelody filtered each token's start/duration for finiteness, then summed them (t.start + t.duration) to pick the last window. Two finite values can sum to Infinity — the issue's repro {start: 1e308, duration: 1e308, midi: 60} — turning windowCount = Math.ceil(Infinity / windowSeconds) into an infinite main-thread loop. Large finite timestamps were equally unhappy: a 1e10s span means ~5 billion synchronous window iterations.

Changes:

• _harmonizeMelody now validates the summed endpoint for finiteness and against a supported song horizon (MAX_HARMONIZED_MELODY_SECONDS = 900.0, sharing routes.py's 15-minute MAX_AUDIO_DURATION_SECONDS bound; unlike the audio path's truncation, an out-of-horizon melody is rejected outright) and returns no chords, which the caller turns into the existing graceful null — the same result a rest-only or lyrics-only track already produces.
• The window count is hard-capped (MAX_HARMONY_WINDOWS = 3000) independently of timestamps, so the loop can never iterate an impractical number of times even if the horizon or minimum windowSeconds ever change.
• Regression tests cover the overflow repro (with a real token alongside it, so the null can't be waved off as an empty window grid), a very large finite span past the horizon, and an in-horizon boundary (>=890s melody still harmonizes).
• Docs/version: README generateAccompanimentFromLyrics null conditions, CHANGELOG Fixed entry, plugin.json 0.6.1 → 0.6.2, CLAUDE.md prose + test count.

Verified: JS suite 110/110 pass (was 107), Python 20/20 pass, node --check clean. Reintroducing the old reduce-based endpoint hangs the new overflow test (killed by a 12s external timeout), confirming the regression coverage pins the fix.

  ^{｜ View workflow run ｜ via Pullfrog ｜ Using big-pickle (free) ｜ 𝕏}

[kilo-code-bot]

Code Review Summary

Status: No Issues Found | Recommendation: Merge

Files Reviewed (6 files)

• chordr/screen.js - _harmonizeMelody endpoint validation + window cap; verified against routes.py MAX_AUDIO_DURATION_SECONDS (900.0) and test expectations
• tests/generate_accompaniment_from_lyrics.test.js - 3 regression tests; verified t=[0,888] math for the in-horizon case
• CHANGELOG.md - Fixed entry consistent with code
• README.md - null-condition docs match implementation
• CLAUDE.md - 0.6.2 / 110 pass prose matches plugin.json and tests
• chordr/plugin.json - version bump 0.6.1 → 0.6.2

---

_{Reviewed by free · Input: 357.8K · Output: 26.9K · Cached: 223.7K}