v2.1.5 — Multi-Scene Prompt Stability

Strengthen Seedream 5 prompt workflows with dedicated landscape, portrait, and flowchart recipes; deterministic routing for exact text and topology; compact runtime documentation; official Lite/Pro evidence scope; and a new optional landscape depth reference.

v2.1.3 - Official prompt and runtime alignment

Fixed

• Added official Seedream 5.0 Pro 1.5K tier support across CLI validation, documentation, and tests.
• Reject streamed duplicate image content assigned to distinct explicit image_index values, preserving an ambiguous billing lock instead of reporting a false multi-image success.
• Retry cleanup of stale completed request state before the next POST and stop safely when the exclusive state lock cannot be removed.
• Stop treating cancelled, empty, or unrecognized model-selection answers as permission to use Lite.

Prompt and capability alignment

• Added the official normalized 0-999 coordinate contract and <point> / <bbox> syntax for Seedream 5.0 Pro interactive editing.
• Kept classification slugs as agent-side routing metadata instead of sending them to the model; only Core request remains mandatory, while purpose and visual fields are added when relevant.
• Aligned prompt guidance with official concise natural-language, sketch/UI text-following, multi-reference role, and edit-preservation recommendations without truncating or blocking prompts by length.
• Replaced the outdated public-evidence gap with direct official support-matrix, Pro tutorial, interactive-editing, and prompt-guide references.

Documentation and validation

• Updated the skill workflow, Lite/Pro references, CLI reference, execution flow, bilingual READMEs, capability manifest, prompt templates, and changelog.
• Local validation: 168 passed, 1 skipped, 100 subtests passed; skill validation, Pro 1.5K dry-run, compileall, pip check, and git diff --check all passed.
• Main and tag CI passed on Ubuntu, macOS, and Windows with Python 3.10-3.13.
• No real Ark request was issued.

v2.1.2 - Streaming safety fixes

Fixed

• Preserve the payload-fingerprinted ambiguous request lock when a streamed generation completes with only partial_failed events, preventing an automatic retry without billing certainty.
• Track occupied stream target indices and reject conflicting reuse of the same image_index, even when --force is enabled.
• Assign indexless stream results to the first unoccupied target and validate usage.generated_images against unique saved targets.

Documentation and validation

• Updated the Lite contract, execution-flow documentation, and bilingual README wording.
• Added regressions for mixed indexed/indexless events, conflicting reused indices, force-overwrite protection, and all-failed stream billing locks.
• Local validation: 165 passed, 1 skipped, 98 subtests passed; compileall, pip check, release-tag verification under UTF-8 and cp1252, and git diff --check all passed.
• No real Ark request was issued.

v2.1.1

Documentation

• Add verified npx skills update imagegen instructions for global and project installs.
• Document manual Git and ZIP update workflows.
• Preserve the skill-local .env, reinstall pinned runtime dependencies, and refresh Claude Code sessions after updates.
• Remove the completed historical improvement report.

Validation

• 162 tests passed, 1 skipped, 98 subtests passed.
• Main and tag CI passed on Ubuntu, macOS, and Windows with Python 3.10–3.13.
• Release tag validation passed under Windows cp1252 output.
• No real Ark request was issued.

v2.1.0

Highlights

• Stop agents from adding --dry-run automatically; explicit local preflight remains available.
• Preserve long prompts without the former 300 Chinese character / 600 English word warning or length-driven rewrites.
• Establish prompting.md as the single source of truth for grounded prompt structure and field ordering.
• Align all scenario templates and visual examples with the unified prompt contract.
• Keep pre-POST validation, no-clobber output protection, and pending / ambiguous billing locks unchanged.

Validation

• 162 tests passed, 1 skipped, 98 subtests passed.
• Release tag validation passed under Windows cp1252 output.
• No real Ark request was issued during validation.

v2.0.7

v2.0.7 — 2026-07-31

修复

• 组图 HTTP 200 下 per-item 错误或部分成功但未允许 partial 时，统一标 ambiguous 保留计费状态锁，避免自动重试造成重复计费。
• _normalize_size 超大位数自定义尺寸不再触发 OverflowError，改为先校验像素范围再计算宽高比。
• 已知拒绝路径删除状态文件失败时，从裸 traceback 崩溃改为告警并保留状态文件。
• 流式事件去重 token 统一基于图片内容，修复乱序/缺 image_index 时的错位。

安全与健壮性

• 输入图片 URL 仅允许 HTTPS；明文 HTTP 仅对本地回环地址放行。
• SSE 连续 data: 行内嵌多个 JSON 时报错提示"缺少空行分隔"。
• 信号处理器加入重入保护，防止二次写入计费状态文件。
• state 子命令输出 reason 再次脱敏并截断。

工程与文档

• CI 增加 fetch-depth: 0，修复浅克隆下空白检查恒为空操作。
• README 分类 slug 数量 20 → 19；execution-flow.md 组图路径与测试文件名修正。
• 抽取两脚本共用的原子写入模块 scripts/_atomic.py。
• pyproject.toml license 改 PEP 639 SPDX；.gitignore 补充常见条目；SKILL.md description 精简。

验证

• python -m pytest -q：161 passed, 1 skipped, 95 subtests passed。
• 未发起真实 Ark 请求。

v2.0.6

修复

• 加固输出 URL 下载：校验全部 A/AAAA 地址、拒绝非公网目标并固定到已验证 IP，防止 DNS rebinding。
• 修复大写 HTTP:// / HTTPS:// 被误判为本地路径的问题。
• 区分 completed 状态写入失败与清理失败，避免错误恢复或重复请求。
• 移除任务内重复计费确认；pending / ambiguous 状态锁仍保持强制停止。

工程与文档

• 拆分并精确固定运行依赖与开发依赖。
• CI 升级至 actions/checkout@v6、actions/setup-python@v6，新增 tag/version 一致性检查。
• tag/version 校验使用 ASCII 机器输出，兼容 Windows runner 的 cp1252 stdout。
• 补充高风险能力常量说明以及 DNS、URL、状态恢复与依赖契约测试。

验证

• 153 passed, 1 skipped, 88 subtests passed
• 编译、依赖完整性、tag/version 一致性和空白检查通过
• 未发起真实 Ark 请求

完整变更见 CHANGELOG.md。