Agent测试
- 作者仓库星标 57
- 作者更新于 实时读取
- 作者仓库 joelclaw
- 领域
- AI 智能 · joelclaw · memory · recall
- 兼容 Agent
-
- Claude Code
- Cursor
- Cline
- Codex
- Windsurf
- Gemini CLI
- +20
- 信任分
- 98 / 100 · 已通过审计
- 作者 / 版本 / 许可
- @joelhooks · v1.0.0 · 未声明 license
- Token 消耗评级
- 低消耗
- 接入复杂程度
- 即装即用
- 是否需要外部 API Key
- 不需要
- 兼容的系统
- 未声明(默认跨平台)
- 底层运行要求
- 无特殊要求
- 文件与系统权限
-
- 只读
- 允许写入 / 修改
- 网络行为
- 仅限本地
- 安装命令数
- 26 条
档案由构建时根据 SKILL.md 与安装命令自动衍生,可能与作者实际意图存在差异。
---
name: recall
description: Fan-out search across all memory sources when context is unclear or vaguely referenced. Triggers…
category: AI 智能
runtime: 无特殊运行时
---
# recall 输出预览
## PART A: 任务判断
- 适用问题:提示词、Agent 工作流、模型评估或自动化推理。
- 输入要求:目标材料、限制条件、期望输出和验收方式。
- 证据边界:围绕“Trigger Detection / Fan-Out Search Pattern / 1. Today's Daily Log (fastest, most likely)”读取原文规则,不把推断写成作者承诺。
## PART B: 执行结果
- **01** 任务判断:确认你的需求是否属于提示词、Agent 工作流、模型评估或自动化推理,并标出输入、限制和预期结果。
- **02** 执行计划:优先按“Trigger Detection / Fan-Out Search Pattern / 1. Today's Daily Log (fastest, most likely)”拆成步骤,说明每一步会读取什么、修改什么、产出什么。
- **03** 交付结果:给出可复制的命令、文件改动、检查清单或内容草稿,并说明如何继续迭代。
- **04** 风险边界:结合 读取文件、写入/修改文件、主要在本地完成、通常不需要额外 API Key 给出执行前确认项。
## Running Rules
- 读取文件、写入/修改文件;主要在本地完成;通常不需要额外 API Key。
- 先小样例验证,再放大到真实任务。
- 交付时同时给结果、检查口径和下一步迭代建议。 原文出现了 `/tmp` 这类斜杠命令;如果你的 Agent 支持命令触发,优先用命令开场,再补充目标和边界。
告诉 Agent 目标文件或材料、期望结果、不可改范围、是否允许联网或执行命令。本 Skill 的权限画像是:读取文件、写入/修改文件。
先用一个小任务确认它会围绕“Trigger Detection / Fan-Out Search Pattern / 1. Today's Daily Log (fastest, most likely)”工作;涉及文件或命令时,先看 diff、日志、预览或测试结果。
检查最终产物是否包含明确结果、必要证据和下一步动作;如果输出泛泛而谈,就补充输入、边界和验收标准后重跑。
---
name: recall
description: Fan-out search across all memory sources when context is unclear or vaguely referenced. Triggers…
category: AI 智能
source: joelhooks/joelclaw
---
# recall
## 什么时候使用
- 把 AI / Agent方向的常用动作沉淀成 Agent 可调用的技能 适合处理AI Agent、提示词、模型评估与自动化推理,核心价值是把输入、判断、执行、验证和交付边界固定下来,避免 Agent 泛泛回答。 围绕 joelclaw…
- 面向提示词、Agent 工作流、模型评估或自动化推理,优先处理能明确输入、步骤和验收标准的工作。
## 需要提供什么
- 目标材料、目录范围、期望结果和不可改动内容。
- 是否允许联网、执行命令、读写文件或调用外部服务。
## 执行规则
- 围绕「Trigger Detection / Fan-Out Search Pattern / 1. Today's Daily Log (fastest, most likely)」组织步骤,不把推断写成作者事实。
- 读取文件、写入/修改文件;主要在本地完成;通常不需要额外 API Key。
- 先跑小样例,确认结果可检查后再扩大任务范围。
## 输出要求
- 给出最终产物、关键证据、验证方式和下一步动作。
- 信息不足时标记 unknown,不编造命令、平台或依赖。 作者原文负责流程事实;仓库文件负责来源和命令;流狐只补充适用场景、限制和质量判断。
skill "recall" {
输入层 -> 用户目标 + 目标文件 + 禁止范围 + 验收标准
上下文层 -> Trigger Detection / Fan-Out Search Pattern / 1. Today's Daily Log (fastest, most likely)
规则层 -> SKILL.md 触发条件 / 执行顺序 / 输出格式
运行层 -> 无特殊运行时 | 读取文件、写入/修改文件 | 主要在本地完成
安全层 -> 通常不需要额外 API Key + 小任务验证 + diff / 日志复核
输出层 -> 可复制结果 + 检查清单 + 下一步迭代
} Recall — Find What Was Said
When Joel references something vaguely, don't guess — fan out across all memory sources and find it.
Trigger Detection
Phrases that indicate a recall is needed (case-insensitive):
- "from earlier", "earlier today", "earlier this week"
- "remember when", "remember that", "you mentioned"
- "what we discussed", "what we talked about", "the conversation about"
- "that thing with", "the thing about", "what was that"
- "did we ever", "have we", "wasn't there"
- "last session", "the other day", "yesterday"
- "correlate with", "connect to what we"
- Any vague pronoun reference to past context ("those photos", "that idea", "the notes")
Key principle: If you'd have to guess what "earlier" or "that" refers to, you need recall.
Fan-Out Search Pattern
Search these sources in parallel where possible, with timeouts on each:
1. Today's Daily Log (fastest, most likely)
# Always check first — most "from earlier" references are same-day
cat ~/.joelclaw/workspace/memory/$(date +%Y-%m-%d).md
2. Recent Daily Logs (if today's doesn't have it)
# Yesterday and day before
cat ~/.joelclaw/workspace/memory/$(date -v-1d +%Y-%m-%d).md
cat ~/.joelclaw/workspace/memory/$(date -v-2d +%Y-%m-%d).md
3. Curated Memory
cat ~/.joelclaw/workspace/MEMORY.md
Search for keywords from the vague reference.
4. Session Transcripts
Use the session_context tool to search recent sessions:
sessions(limit: 10) # find recent session IDs
session_context(session_id: "...", query: "what was discussed about <topic>")
5. Vault Notes
# Keyword search across Vault
grep -ri "<keywords>" ~/Vault/ --include="*.md" -l | head -10
6. System Log
slog tail --count 20 # recent infrastructure changes
7. Processed Media
# Check for images/audio that were processed
ls /tmp/joelclaw-media/ 2>/dev/null
8. Redis State
# Memory proposals, loop state, etc.
redis-cli LRANGE memory:review:pending 0 -1 2>/dev/null
Workflow
- Extract keywords from the vague reference. "Those photos from earlier" → keywords: photos, images, media, telegram.
- Fan out across sources 1-8 above. Use
timeout 5on any command that might hang. - Synthesize — combine findings into a coherent summary of what was found.
- Present context — show Joel what you found, then continue with the original task.
- If nothing found — say so honestly. Don't fabricate. Ask Joel to clarify.
Timeouts Are Mandatory
Every external call (Redis, grep over large dirs, session reads) MUST have a timeout. The gateway session cannot hang on a recall operation.
# Good
timeout 5 grep -ri "keyword" ~/Vault/ --include="*.md" -l | head -10
# Bad — can hang indefinitely
grep -ri "keyword" ~/Vault/ --include="*.md"
Anti-Patterns
- Don't grep one file and call it done. The whole point is fan-out.
- Don't guess when recall fails. Say "I couldn't find it" and ask.
- Don't read entire session transcripts. Use
session_contextwith a focused query. - Don't skip media. Photos, audio, processed images in
/tmp/joelclaw-media/are often what "from earlier" refers to.
先判断是否适合
作者设计意图
作者的方法与取舍
边界和复核