v1.0.0-beta4

新增 img-gen 独立生图技能

本版本新增 skills/img-gen 技能，提供与 MCP 服务端同等的图像生成与编辑能力，无需 MCP 中间层。

核心特性

• 纯环境变量驱动：直接读取 OS 环境变量，不依赖任何 .env 文件，不绑定特定 profile 路径
• 同等覆盖：支持全部 13 个 preset（GPT Image 2 × 10、Nano Banana 2 × 3），与 MCP 服务端 gpt_image_2_official / nano_banana_2_official 能力对等
• 渠道检测：check_env.py 报告各渠道（GPT / Banana）可用状态与缺省配置，供大模型直接读取判断
• 零依赖：仅 Python 标准库，无 pip 依赖

使用方式

# 检测环境
python3 skills/img-gen/scripts/check_env.py

# 生成图片
python3 skills/img-gen/scripts/img_gen.py generate \
    --preset laozhang_gpt_image_2_default \
    --prompt "A futuristic steel plant" \
    --save output.png --size 2K --ratio 16:9

所需环境变量

变量	用途
IMG_GEN_GPT_IMAGE_2_OFFICIAL_API_KEY	GPT 渠道 API key
IMG_GEN_GPT_IMAGE_2_OFFICIAL_PRESET	GPT 渠道缺省 preset
IMG_GEN_NANO_BANANA_2_OFFICIAL_API_KEY	Banana 渠道 API key
IMG_GEN_NANO_BANANA_2_OFFICIAL_PRESET	Banana 渠道缺省 preset
IMAGE_OUTPUT_DIR	输出目录（可选，缺省 ~/images）

MCP 服务端改进

• 恢复可配置 retry_count（0-3），支持对瞬态 HTTP 错误自动重试
• 依赖 mcp[cli] 收紧为 >=1.12.0,<2.0.0

GitHub: https://github.com/zztdandan/image-generate-mcp-remote/releases/tag/v1.0.0-beta4

v1.0.0-beta3

Changes

• Disable all upstream request retries; every MCP image request is attempted exactly once.
• Keep existing preset timeout values unchanged.
• Return an acknowledgement one second after the upstream response while decoding, URL download, validation, and persistence continue in a background thread.
• Return source_url for URL responses with guidance to prefer the persisted save_path.
• Return response format and estimated file size for base64/inlineData responses without URL or provider extras.
• Apply the same asynchronous persistence behavior to official and temporary image tools.

Verification

• uv run ruff check image_generate_mcp_remote tests
• uv run pytest -q (36 passed)
• wheel and sdist built successfully

v1.0.0-beta2

Changes\n- add LaoZhang Nano Banana Pro preset\n- archive LaoZhang Nano Banana Pro provider guide\n- bump package/service version to 1.0.0-beta2

v1.0.0-beta1

Summary

• reduce preset retry budget to one retry plus one initial attempt
• align preset upstream timeouts by max supported size tier: 120s for 1K, 150s for 2K, 200s for 4K
• update deployment guidance and bump package version to 1.0.0-beta1

v1.0.0

Summary

• Promote the project to v1.0.0 and align package metadata, service version, lockfile, and install docs with the stable release number.
• Keep the preset-driven official tool model as the release baseline, including per-call preset + api_key override behavior and provider-owned timeout/dispatch handling.
• Include the latest verified improvements since v0.9.9: corrected retry timing behavior, appended anti-fragment prompt suffixes for GPT presets, and added the apiyi_nano_banana_2 preset.

Verification

• uv run pytest

Release Notes

Publishing this GitHub release triggers the repository's release workflow, which builds distributions, uploads them to the release, and publishes the package to PyPI.

v0.9.9

Summary

• Formalize recent preset behavior refinements so provider-specific differences stay inside preset definitions instead of leaking into tool callers.
• Block unsupported large-size generation for tested presets: LaoZhang default now rejects all 2K and 4K requests, LaoZhang VIP rejects all 4K requests, and Right Codes gpt-image-2 rejects all 2K and 4K requests.
• Tighten preset request handling by keeping LaoZhang VIP payloads minimal, allowing URL-based image downloads beyond HTTPS-only responses, and updating tests and docs for the preset-mode workflow.

Validation

• Run uv run pytest successfully (30 passed).

v0.9.6

• unify image size inputs across tools to image_size plus aspect_ratio
• replace the shared size tuple with a provider-aware registry that records both GPT and Nano Banana size mappings
• update catalog, tests, and docs for the new size contract

v0.9.5

Summary

• add runtime api_key/base_url/model overrides plus gateway-compatible auth and response parsing for the official GPT and Nano Banana tools
• improve image size contracts to record both GPT request sizes and Nano Banana observed output sizes, with catalog and test coverage
• publish version 0.9.5 docs, lockfile, and compatibility updates

v0.9.4

Full Changelog: v0.9.2...v0.9.4

v0.9.2

Summary

• expose per-tool supported size presets, timeout details, and parsed image dimensions so MCP clients can validate requests and saved outputs more reliably
• tighten gpt-image-2-url size validation to benchmark-backed presets and surface discoverable error messages through the catalog and tool descriptions
• refresh README and deployment docs to describe the PyPI-backed uvx / uv tool install workflow and the automatic GitHub Release → PyPI publishing path