360 Secure SkillHub
360 安全技能中心

Osop Log

该 Skill 指导模型在任务完成后,把本次执行过程整理成两个 YAML 文件:描述步骤、类型与依赖边的 .osop 工作流定义,以及记录运行状态、耗时、工具调用次数与推理依据的 .osoplog.yaml 执行记录,并给出节点类型映射和子代理追踪写法。

archie0125
archie0125
热度 45.59
最近更新 2026-05-11 17:39:05
分类:知识依赖:读写文件应用场景:笔记整理 / 工作流 / 模板

安装技能

复制下面的话术发给你的 Agent,它会通过 secure-skills 检索并安装该技能。

执行命令安装此 skill: npx secure-skills-cli add osop-log
Metadata
name
osop-log
description
Generate OSOP session log — creates .osop workflow and .osoplog.yaml execution record
version
1.2.0
emoji
📝
homepage
https://osop.ai
argument-hint
["short description of what was done"]
metadata
{"openclaw":{"requires":{"bins":["bash"],"config":["~/.osop/config.yaml"]},"install":[],"always":false}}
user-invocable
true
disable-model-invocation
false

OSOP Session Logger

You just completed a task. Now produce a structured session log.

What to create

  1. sessions/YYYY-MM-DD-<short-desc>.osop — workflow definition
  2. sessions/YYYY-MM-DD-<short-desc>.osoplog.yaml — execution record

Create the sessions/ directory if it doesn't exist.

Task description

$ARGUMENTS

.osop format

osop_version: "1.0"
id: "session-<short-description>"
name: "<What you did>"
description: "<1-2 sentence summary>"
version: "1.0.0"
tags: [claude-code, <relevant-tags>]

nodes:
  - id: "<step-id>"
    type: "<node-type>"   # human, agent, mcp, cli, api, cicd, git, db, docker, infra, system, event, gateway, data, company, department
    subtype: "<subtype>"  # Optional: llm, explore, plan, worker, tool, test, commit, rest, etc.
    name: "<Step Name>"
    description: "<What this step does>"
    security:
      risk_level: "<low|medium|high|critical>"

edges:
  - from: "<step-a>"
    to: "<step-b>"
    mode: "sequential"    # sequential, parallel, conditional, fallback, error, spawn, etc.

.osoplog.yaml format

osoplog_version: "1.0"
run_id: "<generate-uuid>"
workflow_id: "<matches .osop id>"
mode: "live"
status: "COMPLETED"

trigger:
  type: "manual"
  actor: "user"
  timestamp: "<ISO timestamp>"

started_at: "<ISO timestamp>"
ended_at: "<ISO timestamp>"
duration_ms: <total ms>

runtime:
  agent: "claude-code"
  model: "<current model>"

node_records:
  - node_id: "<step-id>"
    node_type: "<type>"
    attempt: 1
    status: "COMPLETED"
    started_at: "<ISO>"
    ended_at: "<ISO>"
    duration_ms: <ms>
    outputs:
      <what you produced>
    tools_used:
      - { tool: "<tool-name>", calls: <n> }
    reasoning:
      question: "<what was decided>"
      selected: "<chosen approach>"
      confidence: <0.0-1.0>

result_summary: "<1-2 sentence summary>"

Node type mapping

Action type subtype
Read/explore files mcp tool
Edit/write files mcp tool
Shell commands cli script
Run tests cicd test
Git operations git commit / branch / pr
Analyze/reason agent llm
Search codebase mcp tool
Ask user human input
User reviews human review
Spawn sub-agent agent explore / plan / worker
API calls api rest

Sub-agent tracking

If you spawned sub-agents, use parent on child nodes and spawn edge:

nodes:
  - id: "coordinator"
    type: "agent"
    subtype: "coordinator"
  - id: "explore_1"
    type: "agent"
    subtype: "explore"
    parent: "coordinator"
edges:
  - from: "coordinator"
    to: "explore_1"
    mode: "spawn"

Important

  • Be accurate about what tools were used and how many calls
  • Include reasoning for non-obvious decisions
  • Estimate durations based on tool call timing
  • If the task failed, set status to FAILED and include error details
  • View logs at https://osop-editor.vercel.app