Skip to content

v4.0.2 — Post-release patch pass

Choose a tag to compare

View all tags
@elibarnett elibarnett released this 01 Jun 19:57
· 34 commits to main since this release
v4.0.2
6e02918

Post-release patch pass for v4.0.1. No new models — just bug fixes, validation hardening, accurate pricing, Docker hygiene.

Highlights

  • Veo upscale handlers were silently broken (PR #6). veo_upscale_4k could never return a URL — kie returns code: 422 even on success with data.resultUrls (plural), while our handler looked for singular resultUrl and kieRequest threw on non-200. veo_upscale_1080p had the same throw-on-non-200 issue during processing. Both rewritten with a tolerant fetch + dual singular/plural URL extraction.
  • Veo PRICING numbers replaced with empirically-measured values (PR #5). Probed each tier 2026-06-01. Every existing rate was wrong, between -75% and +110%:
    • veo-3/* (Quality): 50 → 31.25 cr/s
    • veo-3-fast/*: 10 → 21 cr/s
    • veo-3-lite/*: 5 → 3.75 cr/s
    • veo/1080p: 20 → 5 cr flat
    • veo/4k: 80 → 120 cr flat
  • Polling now tolerant of kie.ai's transient malformed JSON on /api/v1/jobs/recordInfo (PR #2). Previously a single bad interim poll response would abort a whole 15-minute task with a generic JSON parse error even though kie's task was rendering fine.
  • Cross-model option validation in generate_image / generate_video (PR #1). aspect_ratio checked against each model's aspectRatios; option-level enum/min/max enforced. The low-level MCP SDK doesn't validate per-tool inputSchemas — these were getting through as opaque kie.ai errors.
  • gemini-omni/video hardening: array caps + Array.isArray checks, reference asset quota enforcement (images + videos×2 + character_ids ≤ 7), String()-coerced duration, per-model maxPromptChars: 20000, full 30-voice enum on audio_id, schema maxLength constraints (PRs #1 + #7).
  • Omni IDs now in taskHistory so kieAudioId / characterId survive context compaction (PR #7).
  • Docker hygiene: npm install → npm ci, new .dockerignore (was streaming multi-GB build context per rebuild) (PR #1).

Verification

Live API integration verified with real Gemini Omni and Veo probes; 663 credits (~$3.32) spent on empirical measurements.

Full diff: v4.0.1...v4.0.2

🤖 Generated with Claude Code

Assets 2
Loading