- Add gemini-3.1-flash-image-preview to supported Google multimodal models
- Improve preferences loading with blocking first-time setup flow
- Add first-time-setup.md reference for guided configuration
- Update model references in SKILL.md and preferences schema
- Update x-goog-ext-525001261-jspb header format to match upstream
HanaokaYuzu/Gemini-API (commit 42900f7, 2026-02-03), appending
`,null,null,1` to fix image generation failures
- Replace deprecated gemini-2.5-pro and gemini-2.5-flash with
gemini-3.0-flash and gemini-3.0-flash-thinking
- Add gemini-3.1-pro-preview (empty header, server auto-routed)
- Update SKILL.md and CLI help text to reflect new model options
Closes#50
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Bun's fetch implementation has a known issue where long-lived connections
through HTTP proxies (e.g., Clash, V2Ray) get their sockets closed
unexpectedly, causing Google image generation requests to fail with
"The socket connection was closed unexpectedly".
This change adds automatic proxy detection and falls back to curl as the
HTTP client when a proxy is configured (via https_proxy, http_proxy,
HTTPS_PROXY, HTTP_PROXY, or ALL_PROXY environment variables). When no
proxy is detected, the original fetch-based implementation is used.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Add dense-modules layout for data-rich guides and 3 new styles:
morandi-journal, pop-laboratory, retro-pop-grid. Add keyword shortcuts
for 高密度信息大图 auto-selection.
Prompt credit: AJ (https://waytoagi.feishu.cn/wiki/YG0zwalijihRREkgmPzcWRInnUg)
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Runtime.evaluate reuses the same JS execution context across calls in a
session. The previous expression used `const thumbs = ...` which throws
"Identifier 'thumbs' has already been declared" on the second loop
iteration, causing result.value to be undefined and JSON.parse to throw
"JSON Parse error: Unexpected identifier 'undefined'".
Fix by inlining the querySelector into a single expression with no
variable declaration, eliminating the re-declaration error.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Switch md-to-html from manual line-by-line parsing to marked + front-matter +
highlight.js + remark-cjk-friendly for robust markdown conversion with syntax
highlighting, proper CJK handling, and standard frontmatter parsing.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
When generating multi-image series, use image 1 as --ref for all
subsequent images. This anchors character design, color rendering,
and illustration style across the entire series — critical for styles
with recurring characters or mascots.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Embed file path directly in PowerShell script with single-quote escaping
instead of using param()/−Path which fails with -Command parameter.
- Use fileURLToPath() for getScriptDir() to handle Windows drive-letter paths.
- Add deep analysis template for extracting specific visual elements
- Require MUST INCORPORATE section with concrete, reproducible details
- Add integration approach for spatial layout instructions
- Clarify ref-capable backend requirements (Google/OpenAI)
- Support --ref with OpenAI GPT Image models (gpt-image-1.5)
- Auto-select Google or OpenAI when --ref provided
- Change ref-related warnings to explicit errors with fix hints
- Add reference image validation before generation
- Improve retry logic to skip non-retryable errors
Jim Liu 宝玉
Yangkai.Shen
李野
justnodejs
LyInfi
jeffrey94