跳到主要内容

工作流清单文件编写

workflow.json 是 workflow 的清单文件（Manifest），定义了 workflow 的全部元数据和行为。Workflow Manager 通过此文件发现和加载 workflow。

基本结构​

{
  "schemaVersion": 2,
  "id": "my-workflow",
  "label": "我的 Workflow",
  "version": "1.0.0",
  "provider": "pass-through",
  "display": {
    "core": false,
    "emoji": "🔧"
  },
  "trigger": { "requiresSelection": true },
  "inputs": {
    "member": { "kind": "parent" },
    "grouping": { "mode": "each" }
  },
  "validateSelection": {
    "select": { "policy": "input-member", "source": "selected" },
    "filters": []
  },
  "parameters": {},
  "execution": {},
  "request": { "kind": "pass-through.run.v1" },
  "hooks": {
    "preflight": "hooks/preflight.mjs",
    "applyResult": "hooks/applyResult.mjs"
  }
}

字段详解​

基本标识​

字段必需类型说明
id✅string唯一标识符，不可重复。推荐使用 kebab-case
label✅string用户可见的显示名称
versionstring语义版本号，如 "1.0.0"
provider✅string后端类型。可选值见下文

provider 可选值​

值说明
"pass-through"纯本地执行，无需后端。适用于文件操作、导出等
"skillrunner"通过 Skill-Runner 后端执行 skill
"acp"通过 ACP 后端执行 skill
"generic-http"通过 Generic HTTP 后端调用 API

provider 决定 workflow 与哪种类型的后端兼容，也决定了 Dashboard 中显示哪些后端可执行。

显示控制​

{
  "display": {
    "core": true,
    "emoji": "📊"
  },
  "taskNameTemplate": "Processing: {query}",
  "debug_only": false
}
字段类型说明
display.coreboolean是否标记为核心 workflow（在 Dashboard 中优先展示、加 core 徽章）
display.emojistring显示名前缀图标，如 "📖"
taskNameTemplatestring任务名称模板，用 {参数名} 占位，执行时将替换为实际值
debug_onlybooleantrue 时仅在调试模式下可见

Input Planning Contracts​

inputs and validateSelection have separate, non-interchangeable roles. inputs is the consumer contract for prepared execution members and grouping; validateSelection is the producer contract for raw-selection validation, candidate selection, ordered filtering, and candidate cardinality.

inputs — Execution Input Contract​

{
  "inputs": {
    "member": {
      "kind": "attachment",
      "accepts": {
        "mime": ["text/markdown", "text/x-markdown", "application/pdf"]
      }
    },
    "grouping": { "mode": "parent" }
  }
}
  • member.kind: selection, parent, child, attachment, note, generated-note, or digest-image-target.
  • member.accepts.mime applies only to attachment execution members.
  • grouping.mode: "each" creates one unit per candidate.
  • grouping.mode: "all" creates one unit containing all candidates.
  • grouping.mode: "parent" creates stable parent groups. Candidates without parent identity are skipped as missing-parent.

validateSelection — Candidate Production Contract​

{
  "validateSelection": {
    "require": {
      "selection": {
        "counts": {
          "parents": { "min": 1 },
          "total": { "min": 1 }
        },
        "allowMixed": false
      },
      "candidates": { "min": 1 }
    },
    "select": {
      "policy": "input-member",
      "source": "related"
    },
    "filters": [
      {
        "kind": "source-file-exists",
        "phase": "availability"
      }
    ]
  }
}

require.selection checks the raw SelectionContext exactly once. select then produces ordered atomic candidates. MIME compatibility and filters run before require.candidates. Count rules use either { "exact": n } or non-negative min/max values.

Supported selectors are input-member (source: selected|related), selection, literature-source, generated-note-candidates, and digest-representative-image. Supported filters are source-file-exists, candidates-per-parent, generated-note-kinds-absent, and artifact-absent. Parameter-dependent artifact checks require phase: "execute"; availability filters run during preview and are reapplied during confirmed planning.

trigger — Empty-selection Gate​

{
  "trigger": {
    "requiresSelection": true
  }
}

trigger.requiresSelection is required in schema v2. It controls only whether an empty selection may enter planning; it does not replace require.selection.

执行控制​

{
  "execution": {
    "timeout_ms": 600000,
    "poll_interval_ms": 2000,
    "mcp": {
      "requiredTools": ["search_items", "get_item_detail"]
    },
    "zoteroHostAccess": {
      "required": false,
      "allowWriteApprovalBypass": false
    },
    "feedback": {
      "showNotifications": true
    }
  }
}
字段说明
timeout_ms超时时间（毫秒，仅对 Generic HTTP 后端有效）
poll_interval_ms轮询间隔（毫秒），控制进度检查频率
mcp.requiredTools此 workflow 所需的 MCP 工具列表（工具名称字符串数组）
zoteroHostAccess.required是否需要 Zotero 主机访问权限（读写库数据）
zoteroHostAccess.allowWriteApprovalBypass是否允许绕过写操作审批
feedback.showNotifications是否显示执行通知。默认 true，设为 false 可静默执行

执行模式（auto / interactive）已移至 request.create.mode——参见 请求种类。

结果获取​

{
  "result": {
    "fetch": { "type": "bundle" },
    "final_step_id": "finalize",
    "expects": {
      "result_json": "result/result.json",
      "artifacts": [
        "result/artifact1",
        "result/artifact2"
      ]
    }
  }
}
字段说明
fetch.type获取方式。"bundle"（下载 zip 包）、"result"（仅获取结果 JSON）
final_step_id对于 sequence 工作流，指定最后一步的 id，用于确定最终结果
expects.result_json期望的结果 JSON 文件路径（相对于运行时工作区）
expects.artifacts期望的产物文件路径列表

请求定义​

声明式请求定义，与 hooks.buildRequest 互斥（如果同时存在，hooks.buildRequest 优先）。

{
  "request": {
    "kind": "skillrunner.job.v1",
    "create": {
      "skill_id": "my-skill",
      "skill_source": "local-package"
    },
    "input": {
      "upload": {
        "files": [
          { "key": "source", "from": "selected.markdown" }
        ]
      }
    },
    "poll": {
      "interval_ms": 2000,
      "timeout_ms": 600000
    }
  }
}

有关各 kind 的详细说明，参见请求种类。

Hook 声明​

{
  "hooks": {
    "preflight": "hooks/preflight.mjs",
    "buildRequest": "hooks/buildRequest.mjs",
    "normalizeSettings": "hooks/normalizeSettings.mjs",
    "applyResult": "hooks/applyResult.mjs"
  }
}
字段必需说明
applyResult✅必需。执行后处理结果的脚本路径
preflight可选。在选择解析之后、请求构造之前运行。可继续、跳过、短路到 applyResult，或将一个输入单元替换为多个虚拟请求单元
buildRequest可选。构建发送给后端的请求。与 request 字段互斥
normalizeSettings可选。规范化用户设置参数

输入过滤已被声明式 validateSelection 机制替代——参见 选择验证 章节。

preflight 不参与菜单启用判断、debug-probe 选择分类或 Host Bridge 就绪检查。请将选择约束放在 validateSelection 中，将后端请求构造放在 buildRequest 或 request 中，将 Zotero 写入放在 applyResult 中。

路径是相对于 workflow.json 所在目录的。

本地化​

{
  "i18n": {
    "defaultLocale": "en-US",
    "messages": {
      "zh-CN": {
        "label": "我的工作流",
        "parameters.language.title": "语言"
      }
    }
  }
}

参见本地化页面获取详细说明。

完整示例：一个带参数的文献分析 workflow​

{
  "schemaVersion": 2,
  "id": "my-literature-analysis",
  "label": "我的文献分析",
  "version": "1.0.0",
  "provider": "skillrunner",
  "display": { "emoji": "📄" },
  "trigger": { "requiresSelection": true },
  "inputs": {
    "member": {
      "kind": "attachment",
      "accepts": { "mime": ["application/pdf"] }
    },
    "grouping": { "mode": "each" }
  },
  "validateSelection": {
    "require": {
      "selection": {
        "counts": { "attachments": { "min": 1 } },
        "allowMixed": false
      }
    },
    "select": { "policy": "input-member", "source": "selected" },
    "filters": [
      { "kind": "source-file-exists", "phase": "availability" }
    ]
  },
  "parameters": {
    "language": {
      "type": "string",
      "title": "输出语言",
      "default": "zh-CN",
      "enum": ["zh-CN", "en-US"],
      "allowCustom": true
    }
  },
  "execution": {
    "timeout_ms": 600000
  },
  "request": {
    "kind": "skillrunner.job.v1",
    "create": { "skill_id": "literature-analysis" }
  },
  "result": {
    "fetch": { "type": "bundle" },
    "expects": {
      "result_json": "result/result.json"
    }
  },
  "hooks": {
    "applyResult": "hooks/applyResult.mjs"
  }
}

下一步​