Image Factory v0.8.0

P1 evidence release: opt-in versioned aHash/dHash/pHash distances, sample-sufficient Wilson calibration and reviewer drift alerts, verified-anchor mobile review workspace, and an evidence-gated runtime acceptance matrix. Local 529-test suite, merged CI matrix, skills-check, distribution and offline vendor checks passed. The checked-in 4/8/12 paid-story, cross-host/model and fault-injection cases are all NOT_RUN. Browser rendering at 390/768/1280 remains NOT_RUN because local-page browser access was blocked. Synthetic tests do not establish live character or prop continuity.

Image Factory v0.7.0

P0 visual-continuity foundation: fixed synthetic 4/8/12-shot benchmark packs and offline aggregation, image_batch 1.5.0 structured story state with cross-round inheritance, and scores 1.3.0 versioned advisory reviewer provenance. Real paid continuity, cross-host/model acceptance, and fault-injection evidence remain NOT_RUN and tracked in the active OpenSpec change.

Image Factory v0.6.0 — production quality evidence

Image Factory v0.6.0 completes the third incremental release with production-quality evidence for multi-image work.

Highlights:

• image_batch 1.4.0 adds declared aspect-ratio ranges and optional fixed-algorithm perceptual-hash near-duplicate gates.
• Series review uses closed character, wardrobe, prop, style, scene-state, text-absence, and aspect-ratio dimensions with observable evidence.
• evaluate emits calibration evidence plus rebuildable HTML contact sheets and JSON storyboards; summarize rebuilds them without generation calls.
• Artifact receipts add honest provenance with nulls for values that were not observed.
• Legacy image_batch, receipt, and scores documents migrate in memory.

Verification:

• 489 local tests passed.
• CI passed on Linux, macOS, and Windows across Python 3.11 and 3.13.
• Distribution and offline skill-vendor checks passed.
• OpenSpec production-quality-evidence was validated, synchronized, and archived.

Boundary: no paid image generation or fresh live character-continuity acceptance was run for this release; deterministic behavior is covered with synthetic PNG evidence.

Image Factory v0.5.0

Highlights

• Streams Codex JSONL into durable per-attempt evidence and exposes read-only status --watch.
• Attributes artifacts to the active session/call and recovers late artifacts without another generation call.
• Rejects approved runs before external execution when conservative free-space capacity is insufficient.
• Validates attempt progress atomically and reconciles durable JSONL after interruption.

Verification

• 479 local tests passed.
• Linux, macOS, and Windows CI passed on Python 3.11 and 3.13.
• Distribution and managed-skill supply-chain checks passed.

Runtime boundary

No paid generation was invoked for this release. Fresh-host installation and visual continuity remain separate runtime acceptance gates.

Image Factory v0.4.0 — Series Consistency

Highlights

• Adds a schema-backed consistency profile for characters, props, wardrobe, style, and negative constraints.
• Compiles a deterministic effective prompt and binds it to idempotency and receipts.
• Adds typed reference roles and a strict five-reference ceiling across profile and item inputs.
• Ships migration support for image-batch 1.0.0 through 1.2.0 plans.

Verification

• 465 offline tests passed locally.
• Linux, macOS, and Windows CI passed on Python 3.11 and 3.13.
• Distribution and managed-skill vendor checks passed.
• Marketplace metadata was published at 0.4.0.

Evidence boundary

Live runtime visual acceptance for multi-image character and element continuity was not run in this release environment; this release provides the enforceable contract and deterministic evidence layer.

Image Factory v0.3.0

新增

• Codex 会话优先使用内置 imagegen / image_gen，无需 Provider API Key。
• 依据当前会话元数据或真实工具能力识别 Codex、ZCode、Kimi 与 unknown。
• 仅在明确图片额度耗尽后，才向 Codex 用户提供 Baoyu 降级。
• 捆绑并锁定 image-factory-skills v1.1.0 的逐字节验证 imagegen 快照。

工程化

• 完整测试 454 项通过。
• Codex/ZCode/Kimi manifest 同步到 0.3.0。
• 修复版本提升脚本的 repository marketplace release pin 与单插件安全同步。

已知边界

• Baoyu 产物不会伪造 Factory approval、receipt、evaluation 或 recovery 证据。
• 技能源 release dispatch 因 token 权限返回 HTTP 403，本次由人工执行等价的不可变锁同步并通过在线校验。

v0.2.0 — 循环收敛证据与声明式像素检查

新增

• 循环收敛证据：台账逐轮记录数字评测（numeric_history，schema 1.2.0）——"哪一轮最好、最后一轮是否退步"仅凭台账即可回答，外部评分文档被覆写不再丢失历史
• 命名维度批评：advisory 支持对象式分维度评分 + 可观察证据；无证据维度记为不完整、永不计为差距；新增插件本地静默技能 image-factory-review（独立上下文、反棘轮）
• 回归与两级停滞检测：回归带两轮证据报告；接近停滞拒绝 retry-unchanged（强制结构性改写）；已确认停滞拒绝出下一轮、交还给人——任何评测信号都不能自行开启一轮
• 声明式像素检查：计划条目可声明角点颜色 / 最小边距 / 墨色均值检查，标准库解码器惰性度量，失败并入确定性层（failed_pixel_check）；检查跨轮不可削弱、不入幂等键
• 消费 image-factory-skills v1.0.2：judge 要求独立评分上下文，文档化维度 advisory 与反棘轮规则

实证

• 真实付费多轮作业（3 次调用）：round 1 fail（黑底违反声明白底，实测角点 (0,0,0)）→ round 2 Accepted（白底 254 + 边距 23%），台账 rev 22
• 收敛信号如实记录回归（0.92→0.88，条目集不同）；声明式像素检查在同一组产物上复现人工拒绝依据（runtime.md 两节新证据）

Full Changelog

v0.1.6...v0.2.0

Image Factory v0.1.6

Full Changelog: v0.1.5...v0.1.6

Image Factory 0.1.5

Cross-host maintenance release for Codex, ZCode, and Kimi.\n\n- Uses the canonical public repository identity.\n- Pins repository-local marketplace installation to v0.1.5.\n- Preserves host-specific Codex contracts without using Codex as the public product prefix.\n- Includes validated OpenSpec documentation and current distribution checks.

Image Factory v0.1.4

Highlights

• Synchronizes the checksummed external skill snapshot to image-factory-skills v1.0.1.
• Locks the source release to peeled commit 65fb738da4fc23668299165681c0e3f52e9962ba.
• Keeps plugin-local skills outside the managed lock boundary.
• Preserves immutable tag + digest verification for future upgrades.

Verification

• Local online and offline vendor checks passed.
• Repository test suite passed.
• GitHub Actions succeeded at commit 0eec7352423c1a61a7427033f1ebf8c3290552fd.