PPT Master 生成管线失败恢复治理:跨路由停止/继续规则、恢复矩阵与断点续跑实战指南
PPT Master 的 Default Generate PPTX 是一条多阶段长链路:从源内容处理、两阶段用户确认、图像获取、逐页 SVG 创作,到质量门禁与原生 PPTX 导出。任何一个环节失败,若处理不当都会导致重复规划、丢状态或"静默降级"。本文以仓库治理文档 failure-recovery.md 为骨架,完整展开其三层治理模型——跨路由全局停止/继续规则(§2)、Generate PPTX 专用恢复矩阵(§1)与断点恢复指针(§3),并结合 generate-pptx.md、confirm_ui.md、executor-base.md 等源码与命令文档,给出可直接落地的命令、工件与判断依据。读完本文,你将掌握:什么故障必须阻断管线、什么故障自动降级即可;每类故障的自动恢复动作、何时需要用户介入,以及从哪个阶段入口续跑。
1. 治理模型总览:三层规则如何分工
该治理文档把停止/继续规则组织为三层,作用域严格区分:
| 章节 | 作用域 | 职责 |
|---|---|---|
| §2 全局停止/继续规则 | 跨全部三条路由(Generate PPTX、Create Template、Edit Native PPTX) | 通用兜底:门禁工件缺失、便利面故障、派生工件过期、静默降级禁令等 |
| §1 Generate PPTX 恢复矩阵 | 仅 Generate PPTX | 按故障点给出"是否阻断 / 自动恢复 / 用户介入 / 恢复入口"四元组 |
| §3 Generate PPTX 恢复指针 | 仅 Generate PPTX | 按"最后良好状态"给出精确续跑入口 |
文档还规定:拥有路由与阶段的文档可以添加更窄的处理,但绝不能削弱这些规则或重复矩阵内容——即本文件是停止/继续治理的唯一权威,其他文档只能引用。
两条贯穿全文的硬规则:
硬规则 1:失败的必要工件(required artifact)阻断下一个门禁(gate);失败的便利面(convenience surface,如 Confirm UI、live preview)回退到规范通道(chat 或命令行)且不阻断路由。
硬规则 2:运行时路径必须在每次工具调用内从入口时的
SKILL_DIR锚点展开,绝不改变 CWD 或继承先前 shell 的工作目录(见 generate-pptx.md)。
这与仓库的工件所有权规范一脉相承:每个事实只能从它的属主工件读取,绝不把多个通道合并成第二个事实源(见 artifact-ownership.md 的开篇硬规则)。
2. 全局停止/继续规则(跨路由,§2)
2.1 规则总表
| 条件 | 动作 |
|---|---|
| 必要门禁工件缺失(required gate artifact missing) | 在该门禁处停止并点名该工件 |
| 可选阶段未被显式请求 | 不作为恢复动作运行它 |
| 便利 UI / 服务器失败 | 回退到聊天或无需该表面继续 |
| 派生工件过期(derived artifact stale) | 从其属主源重新生成 |
| 必要人工工件缺失 | 暂停并点名确切工件;只有它们存在后才继续 |
| 校验或导出失败 | 修复属主源工件,然后只重跑失败及其受影响的下游操作 |
| 已确认的执行选择无法被满足 | 保持已确认需求可见;只按属主工作流允许的方式重试已确认的 provider、mode、voice、effect 或 path;仍不可用时停下、请求新决策,或使用属主工作流声明的恢复。Quick 的 AI 生成例外是一个显式的无 AI 重规划:保留沟通任务,绝不换用另一个图片来源 |
2.2 缺失值(Missing values)的边界
对于已有工件中的某个字段,只遵循其属主 schema 或工作流声明的三种规则之一:必填性(requiredness)、推断过程(inference procedure)或固定默认值(fixed default)。一个没有此类规则的活动性省略(active omission),在属主边界处停下。
同时明确禁止"按类比扩展回退"(do not extend a fallback by analogy)——空值、非激活条件字段、整个工件或派生工件、文件格式属性,都不能用类比方式套用回退。仓库中只存在两种回退术语:
- declared-inference / declared-procedure fallback:声明缺失条件 + 有界过程,不需要新的用户决策;
- fixed compatibility default:声明确切值,应用时给出一次警告。
2.3 禁止静默降级(Forbidden — silent downgrade)
绝不因为下游命令"可能容忍"缺失文件而跳过必需门禁;绝不为了保持路由前进而改动已确认的执行值。正确动作只有四种:修复、暂停、请求新决策、在属主边界应用显式的 profile 属主恢复。
这条规则在导出阶段有直接体现:svg_to_pptx.py 导出器要求当前匹配的 final 质量报告,缺少、不可读、不支持、非 final、阻断、过期或不可验证的报告都会导致非零退出(见 generate-pptx.md Step 7.3),这正是"下游可能容忍缺失文件"的倒置——导出器在技术上拒绝容忍,从而把"静默降级"堵死。
2.4 主动生产化解(Proactive production resolution)
对于生产字段(Speaker Notes / Custom Animations / Narration Audio),解决顺序固定为:
最新显式指令 → final Stage 2 主动值 → 工作流默认(enabled / disabled / disabled)
- 启用 Narration Audio 会抬高 Speaker Notes(只要 Notes 未被显式禁用);
- 显式 notes-off / audio-on 冲突时,在 Generate 的单问题依赖门禁处停下询问;
- 结果与溯源(provenance)只记录在 design_spec.md §I,绝不进入 spec_lock.md。
这也与 strategist.md §6.2 的规划工件语法一致:Design Spec 是可读的人类决策,lock 是其执行子集,result.json 只消费一次且不再重开。
3. Generate PPTX 恢复矩阵(仅 Generate PPTX,§1)
这是治理文档的核心资产:一张 43 行、五列的恢复矩阵,每行给出"故障点 → 是否阻断 → 自动恢复 → 用户介入 → 恢复入口"。矩阵必须被完整继承,下面先原样呈现完整矩阵,再按故障族分组给出命令级展开。
| Failure point | Blocking | Automatic recovery | User intervention | Resume entry |
|---|---|---|---|---|
| Confirm UI launch failure | No | Re-check confirm_ui/result.json once, then chat fallback |
No | generate-pptx Step 4 chat confirmation |
| Confirm UI wait timeout | No, if no final result yet | Re-check result.json once; server cleanup stays mandatory |
Only if the user still wants the page | Step 4 same stage or chat fallback |
| User switches from Confirm UI to chat mid-stage | Yes until the current stage is confirmed | Follow the in-run switch in confirm-surface.md: keep persisted confirmed stages, continue the current and remaining stages in chat, never relaunch UI |
Confirm in chat unless delegated | Step 4 current chat stage |
| Stage 1 completed, then interrupted while UI remains selected | Yes until final Stage 2 is confirmed | Read Stage 1 result.json, derive a fresh recommendations.stage2.json without changing Stage 1, then --wait-only |
Usually no | Step 4 final Stage 2 write/wait |
| Missing final confirmation | Yes | None | User confirms or changes the values | Step 4 final confirmation |
Final confirmed value or later explicit override missing, changed, or weakened in design_spec.md |
Yes | Repair from the retained final-confirmation object plus any newer explicit instruction; only a fresh recovery turn with no retained state reads persisted evidence once | Only when the value genuinely cannot be honored | Step 4 Gate 1 |
spec_lock.md changes confirmed identity or omits a required anchor/routing decision |
Yes | Re-author the affected rows from the completed Design Spec and context; do not enumerate page-local literals | No unless the Design Spec is incomplete | Step 4 Gate 2 |
| Execution exposes a missing Strategist-owned role/plan detail | Yes for the page | Repair the Design Spec/lock fragments under executor-base.md §2.1 |
Only if confirmed intent changes | Step 4 Gate 1/2 → Step 6 current page |
| Execution context is fresh, resumed, compacted, external, or unknown | Yes until rebuilt | Read the complete Design Spec, then lock, once; reload triggered inputs and the latest completed SVG when mid-deck | No | Step 6 current page |
apply-template-workspace rejects a legacy or incomplete template |
Yes | Stop template consumption; create a new workspace through Create Template from the original PPTX/reference and return with its exact root | Only when source evidence or template choices are unavailable | Create Template → Step 4 Stage 1 |
| Native formula marker validation or LaTeX compilation failure | Yes for the page | Repair the marker or LaTeX and rerun the checker; no formula-image fallback | Clarify the equation only when the source is ambiguous | Active SVG authoring step |
| AI image generation failure | Default blocks at the recovery decision; Quick does not | Default auto: try A → B, then ask once to repair/retry, generate manually, or cancel/replan; explicit api / host-native: retry that path, then ask. Quick removes exhausted AI/dependent-slice jobs, replans with native text/SVG or prepared non-AI assets, and continues |
Default chooses one outcome; manual files still required before export | Default: Step 5 decision, Step 4 plan repair, or Step 7 image readiness gate; Quick: current resource plan |
| Web image search/download failure | No | Adjust query/source per image-searcher rules, then Needs-Manual if unresolved |
Only if required with no substitute | Step 5 |
| Slice sheet missing | Yes for derived slice rows | Wait for the parent sheet; run slice_images.py; rerun image analysis |
Yes when the sheet was manual/offline | Step 5 slice handling / Step 7 image readiness gate |
| Strict-alpha slice failure | Yes for every named output | Return the parent to preparation; correct an evidenced key/tolerance mismatch, then enlarge cells or split incompatible shape families and regenerate; a parent never substitutes for its slices | Only after automated recovery is exhausted | Quick §2 resource closure / Default Step 5 |
Residual Pending or Failed image row before Executor |
Yes | Re-run the owning path; Default AI follows its three-outcome decision and reaches Needs-Manual only after manual confirmation; Quick follows its no-AI replan |
Only when the owning rule requires a new choice or file | Step 5 terminal-state check |
| User replaces/adds images after analysis | No | Re-run analyze_images.py before reading image facts |
No | Step 4/5/6 image-fact read |
| Live preview fails to start / closed by user | No | Continue generation; report unavailability | Only if the user requires browser preview | Step 6, or live-preview Step 1 on request |
| Browser annotations submitted during generation | No | Defer application until after Step 7 | User asks to apply | live-preview Step 2 |
svg_quality_checker.py error |
Yes | Review the complete issue set from one unfiltered run; fix all errors and selected warnings in one consolidated pass, then one verification rerun; a remaining failure is the next batch — never check between individual fixes | No unless a required asset is missing | Step 6 Visual Construction |
svg_quality_checker.py warning |
No | Continue without mandatory modification; preserve compatible user syntax, report material advice when useful | No | Step 6 advisory handling |
Missing notes/total.md while Speaker Notes is enabled |
Yes | Generate notes before Step 7 | No | Step 6 Logic Construction |
| Step 7 image readiness missing manual files | Yes | None; list required filenames and prompts | Yes | Step 7 image readiness gate |
total_md_split.py failure |
Yes | Fix notes format/path, rerun Step 7.1 | Usually no | Step 7.1 |
finalize_svg.py failure |
Yes | Fix SVG/assets, rerun Step 7.2 | Only if a source asset is missing | Step 7.2 |
svg_to_pptx.py failure |
Yes | Without a current passing final report: get the complete blocking set from the final checker, fix, rerun the checker, proceed only after passed; otherwise fix the conversion issue and rerun Step 7.3 |
Only if a required artifact is missing | Step 6 final quality gate or Step 7.3 |
| Export succeeds but the user wants browser edits re-exported | No | Rerun Step 7.2 and 7.3 after the edits | No | Post-export live-preview handling |
注:为避免表内换行破坏可读性,上表保留了原始英文术语(Blocking / Resume entry 等),与治理文档原文一一对应;本文其余部分使用中文叙述展开。
3.1 确认面故障族:便利面回退,不阻断路由
矩阵前四行全部围绕 Confirm UI 与聊天切换。核心事实:
- 启动失败与等待超时都是"不阻断"——先重新检查一次
confirm_ui/result.json,然后走聊天回退;但服务器清理(--shutdown)是强制性的。 - 阶段中途从 UI 切到聊天是"阻断"的(直到当前阶段被确认):遵循 confirm-surface.md 的 in-run switch——保留已持久化的已确认阶段,当前及剩余阶段在聊天中继续,绝不重新拉起 UI。
命令层证据来自 confirm_ui.md 与 generate-pptx.md Step 4:
python3 ${SKILL_DIR}/scripts/confirm_ui/server.py <project_path> --daemon # 启动组合 Stage 1
python3 ${SKILL_DIR}/scripts/confirm_ui/server.py <project_path> --wait-only --wait-stage stage1 # 等待 Stage 1 回执
python3 ${SKILL_DIR}/scripts/confirm_ui/server.py <project_path> --complete-template-selection # 安装/自由设计闭环后的握手
python3 ${SKILL_DIR}/scripts/confirm_ui/server.py <project_path> --wait-only # 等待 final Stage 2
python3 ${SKILL_DIR}/scripts/confirm_ui/server.py <project_path> --shutdown # Step 4 清理(幂等)
生命周期细节决定了"恢复"怎么写:
- 默认绑定
127.0.0.1:5050起的第一个空闲端口,--port N精确绑定;等待预算默认 590 s(--wait-timeout),超时后服务器仍存活,调用方在聊天回退前重新检查两个 Stage-1 回执; - 空闲 900 s 自动关闭,
--shutdown幂等且优先走/api/shutdown优雅退出; - 项目级锁位于
<project_path>/.confirm_ui.lock,重复启动被拒绝,陈旧锁(死 pid)会被覆盖; --reset-template-selection只清除template_options.json、template_selection.json、template_handoff.json三个侧车文件,绝不触碰 Strategist 文件、已安装模板内容或result.json。
3.2 规划工件族:Gate 1 / Gate 2 与"证据只读一次"
矩阵中"Final confirmed value … missing, changed, or weakened in design_spec.md"与"spec_lock.md changes confirmed identity"两行分别对应用户确认 → Design Spec → lock 的两道门禁:
- Gate 1(active-decision fidelity):用保留的最终确认对象修复
design_spec.md;只有无保留状态的全新恢复轮次才被允许把持久化最终证据读一次。result.json的最终对象(stage: final+status: confirmed)在 Step 4 被恰好读取一次并保留到 Design Spec 写作完成,之后不再重开(见 confirm_ui.md 的 final-result consumption contract)。 - Gate 2(lock context fidelity):保持 Design Spec 不动,只重写不匹配的 lock 行;lock 可以规范化语法、补充有依据的重复角色,但绝不能改变身份、丢弃细化、引入方向或变成字段拷贝/允许清单(strategist.md §6.2)。
"执行上下文是 fresh / resumed / compacted / external / unknown"一行则给出了完整重载顺序:读完整 Design Spec,再读一次完整 lock;若处于 deck 中途,重载触发输入与最新一张已完成的 SVG——这正是 resume-execute.md 的核心机制:新会话里"继续生成 projects/"进入 Step 6 前,先做 design_spec.md + spec_lock.md 的存在性 sanity check,二者任一缺失即按 §3 恢复,绝不进入 Step 6 或把孤儿 lock 当作权威。
3.3 图像与切片故障族:三层恢复哲学
- AI 图像生成失败(Default vs Quick 分叉):Default 的
auto路径先 Path A 再 Path B,两者耗尽后问一次——修复重试同路径、手动生成、或取消相关 AI 图并修复计划;只有确认的manual才产生Needs-Manual行(见 image-base.md §3 的"自动耗尽先于阻断"硬规则)。Quick 则完全不同:耗尽自动化后进入显式 no-AI replan,移除耗尽的 AI/依赖切片任务,用原生文本/SVG 或已备好的非 AI 资产重规划并继续——绝不让另一个图片来源顶替(quick-generate.md §2)。 - 切片(slice)故障:父 sheet 缺失 → 等待父表、运行
slice_images.py、重跑图像分析;严格 alpha 切片失败 → 父图退回准备态,修正有证据的 key/tolerance 不匹配,或扩大单元格 / 拆分不兼容形状族后重新生成——父图绝不顶替其切片输出。命令形如:python3 ${SKILL_DIR}/scripts/slice_images.py <project_path>/images/<sheet>.png --grid RxC --names ... --trim --alpha --bg KEY_HEX_FROM_PROMPT --strict-alpha - 残留
Pending/Failed图像行:Executor 开始前必须终结;Default AI 走其三结果决策,只有手动确认后才到达Needs-Manual。 - 用户在分析后增删图片:不阻断,但读取图像事实前必须重跑
python3 ${SKILL_DIR}/scripts/analyze_images.py <project_path>/images——analysis/image_analysis.csv是images/的再生视图而非持久缓存(artifact-ownership.md §3 再生规则)。
3.4 质量检查器故障族:合并修复纪律
svg_quality_checker.py 的 error/warning 区分在矩阵中对应两行:error 阻断,warning 不阻断。error 的修复纪律极其严格(executor-base.md §3 Checkpoints):
运行一次不过滤的 checker,审阅完整问题集,在一个合并修复轮里修掉全部 error 与选中的 warning,然后做一次验证重跑;仍失败的问题成为下一个批次——绝不在单个修复之间反复检查。
门禁命令(Generate Step 6,generate-pptx.md):
python3 ${SKILL_DIR}/scripts/svg_quality_checker.py <project_path> --canonical-authoring --stage early --json # P05 后、P06 前;≤6 页跳过
python3 ${SKILL_DIR}/scripts/svg_quality_checker.py <project_path> --canonical-authoring --stage final --json # 全部页面存在后
--json 写报告文件(validation/svg_quality_early_report.json / validation/svg_quality_report.json),stdout 保持人类可读摘要且绝不被当作 JSON 解析;final 报告被导出器按指纹绑定到 svg_output/,这正是"报告过期 → 导出拒绝"的实现机制。checker 每次调用必须跟随"门禁点"或"一个合并修复轮"之一——"验证一个早期写作模式"不是理由(pacing violation)。
3.5 备注、后处理与导出故障族:严格串行子步骤
Step 7 是严格串行的三个子步骤,一条命令一个调用,前一个成功且满足成功判据后才进入下一个(generate-pptx.md Step 7):
| 子步骤 | 命令 | 成功判据 |
|---|---|---|
| 7.1 演讲备注切分(仅启用时) | python3 ${SKILL_DIR}/scripts/total_md_split.py <project_path> |
notes/ 下每页 Markdown 覆盖全部发布页 |
| 7.2 自包含 SVG 预览 | python3 ${SKILL_DIR}/scripts/finalize_svg.py <project_path> |
svg_final/ 每页一份自包含预览;其缺失不阻断 7.3 |
| 7.3 原生 PPTX 导出 | python3 ${SKILL_DIR}/scripts/svg_to_pptx.py <project_path>(Notes disabled 时追加 --no-notes) |
退出 0,产出 exports/<project_name>_<timestamp>.pptx 与 validation/<...>.report.json(passed / passed-with-warnings) |
失败恢复遵循同一原则:修复属主源工件,从失败的子步骤续跑;绝不因源未变就重启规划。finalize_svg.py 失败 → 修 SVG/assets 重跑 7.2;total_md_split.py 失败 → 修备注格式/路径重跑 7.1;svg_to_pptx.py 失败 → 若无当前通过的报告,先取 final checker 的完整阻断集修复并重跑到 passed,否则修转换问题重跑 7.3。
"Step 7 图像就绪门禁"也是一个典型阻断点:任何必需行处于 Needs-Manual 时,images/ 下必须已存在每个预期文件与切片输出,否则暂停并列出确切文件名,绝不运行 finalize_svg.py 或 svg_to_pptx.py,绝不发布虚线占位符。
4. Generate PPTX 恢复指针(仅 Generate PPTX,§3)
4.1 "最终确认证据"的定义(关键前提)
Final confirmation evidence = 当前聊天中的显式最终确认,或
<project>/confirm_ui/result.json中status: confirmed+stage: final的对象;仅靠规划工件不能证明任何东西。
在该门禁之后,一条更新的显式指令只更新其在 Design Spec 中的有效生产结果与溯源,从属主步骤续跑,不重开 Confirm UI。UI 等待条目只在 UI 仍是所选表面时适用;更新的聊天表面指令则走 in-run switch,在聊天中续跑。
4.2 断点续跑总表
| Last good state | Resume from |
|---|---|
| Stage 1 confirmed, final Stage 2 missing or unconfirmed, UI selected | Derive a fresh recommendations.stage2.json from confirmed Stage 1 and current inputs, then confirm_ui/server.py <project> --wait-only |
Final evidence exists; design_spec.md missing (with or without a surviving lock) |
Step 4 and strategist.md §6.2: read final evidence once, read design_spec_reference.md, author the complete Design Spec from scratch, pass Gate 1; then read spec_lock_reference.md and re-author the complete lock, replacing any orphan. Never reconstruct the Design Spec from an orphan lock |
| Final evidence exists; Design Spec exists, lock missing | Step 4: read final evidence once to audit the Design Spec, then author the complete lock from it plus context |
| Both planning artifacts exist but Gate 1 fails | Read final evidence once, repair design_spec.md, re-author every affected lock row; do not reopen recommendations or infer from the lock |
| Gate 1 passes, Gate 2 fails | Keep the Design Spec; re-author only the mismatched lock rows |
| No final evidence | If result.json proves Stage 1, resume at final Stage 2; otherwise restart Step 4 at Stage 1. Never infer confirmed choices from partial artifacts |
| Both planning artifacts complete, split mode selected | resume-execute |
| Images acquired, SVGs not started | generate-pptx Step 6 |
| SVGs complete and checker passed; Speaker Notes enabled, notes missing | Step 6 Logic Construction |
| SVGs complete; Speaker Notes disabled | Conditional motion handling, then Step 7.2 and export with --no-notes |
| SVGs and notes complete | Step 7.1 |
| Step 7.1 complete | Step 7.2 |
| Step 7.2 complete, PPTX missing | Step 7.3 |
| Browser annotations saved after export | live-preview Step 2 |
4.3 默认原则
默认 — 从失败的属主步骤续跑:除非属主源发生变化,否则不重启规划、不重新生成先前的工件。
这条默认原则与 Step 7 的"严格串行"硬规则共同构成整条管线的恢复基石:状态以磁盘工件为准,能续跑就绝不重来。也正是这个原则,让 resume-execute.md 能够在全新聊天中仅凭 design_spec.md + spec_lock.md 的存在性检查就安全地接管 Step 6–7——上下文无关(context-independent)的恢复正是这套治理的最终形态。
5. 治理如何落地:源码级印证与相关文档导航
- 恢复矩阵的"属主工件"语义:每个工件由谁读写、何时可再生,见 artifact-ownership.md 的所有权矩阵(§1)与再生规则(§3)——
analysis/image_analysis.csv由analyze_images.py再生、notes/slide_*.md由total_md_split.py从notes/total.md再生、svg_final/由finalize_svg.py从svg_output/再生、validation/svg_quality_report.json由 checker 从svg_output/再生。 - 恢复入口对应的完整流程:Step 1–7 的门禁、角色切换与强制命令,见 generate-pptx.md;Quick 单遍简档的恢复差异(无 AI 重规划、无 spec/lock、lockless final checker),见 quick-generate.md。
- 确认面语义:surface 决策、in-run switch、
recommendations.stage1.json/.stage2.json/result.json的载荷形状,见 confirm-surface.md;服务器生命周期、端口/锁、模板选择侧车与进度守卫,见 confirm_ui.md。 - 执行上下文与修复纪律:Executor 的 §2.1 绑定规则(binding vs realization)与 §3 合并修复纪律,见 executor-base.md;上下文失效后的完整重载顺序,见 resume-execute.md。
- 模板工作区被拒的恢复路径:旧版或不完整模板包绝不原地迁移——通过 Create Template 新建工作区后带着确切 root 返回 Stage 1(apply-template-workspace.md §1 的 raw source boundary)。
- 路由归属:Generate PPTX / Create Template / Edit Native PPTX 三条顶级路由的划分与触发条件,见 routing.md——治理文档的 §2 正是这三条路由共享的停止/继续底线。
一句话总结:PPT Master 的失败恢复不是"出错了重来",而是一张可执行的决策表——阻断与否由门禁工件决定,自动恢复遵循属主源再生,续跑入口永远指向磁盘工件对应的属主步骤;唯一必须停下请人拍板的,只有那些无法自动、无法推断、也无法降级的真实缺口。
atomcodeClaude Code 的开源替代方案。连接任意大模型,编辑代码,运行命令,自动验证 — 全自动执行。用 Rust 构建,极致性能。 | An open-source alternative to Claude Code. Connect any LLM, edit code, run commands, and verify changes — autonomously. Built in Rust for speed. Get StartedRust0631
MiniCPM5-2BMiniCPM5-2B 是一款面向端侧、本地部署和资源受限场景的 2B 稠密 Transformer,能够达到同尺寸开源模型 SOTA 水平。Markdown00
video-shotcraftAI宣传片skill,使用 Remotion 制作电影级产品视频:提供106 张镜头配方卡和可复用的视频魔板。适用于 Claude Code 与 Codex以及所有其他智能体Markdown00
HivisionIDPhotos⚡️HivisionIDPhotos: a lightweight and efficient AI ID photos tools. 一个轻量级的AI证件照制作算法。Python09
DragonOSDragonOS is an operating system developed from scratch using Rust, with Linux compatibility. It is designed for **Serverless** scenarios. 使用Rust从0自研内核,具有Linux兼容性的操作系统,面向云计算Serverless场景而设计。Rust00
Spark-X2.5-1.7BSpark-X2.5-1.7B 旨在让强大的 AI 更加实用、高效且易于获取。这些模型在广泛的日常任务中表现出色,涵盖对话、写作、翻译、推理、编程、工具调用和智能体工作流,并在同等规模的开源模型中取得领先结果。Spark-X2.5 将面向效率的架构与最高 1M tokens 的原生上下文窗口相结合,并支持 200 多种语言。Python00