Back to Details

research-web

Compare original and translation side by side

🇺🇸

Original

English
🇨🇳

Translation

Chinese
Research request: $ARGUMENTS
研究请求：$ARGUMENTS

Thoroughness Level

详尽程度等级

FIRST: Determine thoroughness before researching. Parse from natural language (e.g., "quick lookup", "thorough research", "comprehensive analysis") or auto-select based on query characteristics.
Auto-selection logic:
  • Single fact/definition/date → quick
  • Focused question about one topic → medium
  • Comparison, evaluation, or "best" questions → thorough
  • "comprehensive"/"all options"/"complete analysis"/"deep dive" → very-thorough
Explicit user preference: If user explicitly specifies a thoroughness level (e.g., "do a quick lookup", "thorough research on X"), honor that request regardless of other triggers in the query.
Trigger conflicts (auto-selection only): When auto-selecting and query contains triggers from multiple levels, use the highest level indicated (very-thorough > thorough > medium > quick).
LevelAgents/WaveWave PolicyBehaviorTriggers
quick1Single waveSingle web-researcher, no orchestration file, direct answer"what is", "when did", factual lookups, definitions
medium1-2Single waveOrchestration file, focused research on 1-2 anglesspecific how-to, single technology, focused question
thorough2-4Continue while critical gaps remainFull logging, parallel agents, cross-reference, follow-up waves for critical gaps"compare", "best options", "evaluate", "pros and cons"
very-thorough4-6Continue until comprehensive OR diminishing returnsMulti-wave research until all significant gaps addressed or new waves stop yielding value"comprehensive", "complete analysis", "all alternatives", "deep dive"
Multi-wave research: For thorough and very-thorough levels, research continues in waves until satisficing criteria are met. Each wave can spawn new investigators to address gaps, conflicts, or newly discovered areas from previous waves. There is no hard maximum - waves continue as long as they're productive and gaps remain at the triggering threshold.
Ambiguous queries: If thoroughness cannot be determined AND the query is complex (involves comparison, evaluation, or multiple facets), ask the user:
I can research this at different depths:
- **medium**: Focused research on core aspects (~3-5 min)
- **thorough**: Multi-angle investigation with cross-referencing (~8-12 min)
- **very-thorough**: Comprehensive analysis covering all facets (~15-20 min)

Which level would you prefer? (Or I can auto-select based on your query)
State: 
**Thoroughness**: [level] — [reason]
then proceed.
第一步：开始研究前先确定详尽程度。可从自然语言中解析（如“快速查询”“深度研究”“全面分析”），或根据查询特征自动选择。
自动选择逻辑：
  • 单一事实/定义/日期 → 快速模式
  • 针对单个主题的聚焦问题 → 中等模式
  • 对比、评估或“最佳”类问题 → 深度模式
  • 包含“全面”“所有方案”“完整分析”“深度挖掘” → 极深度模式
用户明确偏好：若用户明确指定详尽程度（如“做个快速查询”“对X进行深度研究”），无论查询中存在其他触发条件，均优先遵循用户要求。
触发条件冲突（仅自动选择时）：自动选择时若查询包含多个等级的触发条件，选择最高等级（极深度 > 深度 > 中等 > 快速）。
等级每轮Agent数量轮次策略行为表现触发关键词
快速1单轮单个web-researcher，无需orchestration file，直接返回答案“是什么”“何时”、事实查询、定义类问题
中等1-2单轮使用orchestration file，针对1-2个角度进行聚焦研究特定操作指南、单一技术、聚焦类问题
深度2-4存在关键缺口时持续完整日志记录、并行Agent、交叉验证、针对关键缺口启动后续轮次“对比”“最佳方案”“评估”“优缺点”
极深度4-6持续至全面覆盖或收益递减多轮研究，直至所有重要缺口被填补或新轮次无法产生新价值“全面”“完整分析”“所有替代方案”“深度挖掘”
多轮研究：深度和极深度模式下，研究将分多轮进行，直至满足终止条件。每一轮均可生成新的调查Agent，以解决前一轮发现的缺口、冲突或新领域问题。无硬性最大轮次限制——只要能产生有效信息且仍存在缺口，就会持续启动新轮次。
模糊查询处理：若无法确定详尽程度且查询较为复杂（涉及对比、评估或多方面内容），请询问用户：
我可以通过不同深度完成这项研究：
- **中等**：聚焦核心方面的研究（约3-5分钟）
- **深度**：多角度调查并交叉验证（约8-12分钟）
- **极深度**：覆盖所有方面的全面分析（约15-20分钟）

您希望选择哪个等级？（或我可根据您的查询自动选择）
先声明：
**详尽程度**：[等级] — [理由]
，再继续研究。

Deep Web Research Skill

深度网络研究Skill

Orchestrate parallel web researchers to comprehensively investigate a topic through iterative waves, then synthesize findings into actionable intelligence.
Loop: Determine thoroughness → Decompose topic → Launch Wave 1 → Collect findings → Evaluate gaps → [If gaps significant AND waves remaining: Launch next wave → Collect → Evaluate → Repeat] → Synthesize → Output
Orchestration file: 
/tmp/research-orchestration-{topic-slug}-{YYYYMMDD-HHMMSS}.md
- external memory for tracking multi-wave research progress and synthesis.
协调并行web研究Agent，通过迭代轮次全面调查主题，然后将结果整合为可落地的情报。
流程：确定详尽程度 → 分解主题 → 启动第1轮 → 收集结果 → 评估缺口 → [若存在重要缺口且轮次仍有价值：启动下一轮 → 收集 → 评估 → 重复] → 整合结果 → 输出
Orchestration文件：
/tmp/research-orchestration-{topic-slug}-{YYYYMMDD-HHMMSS}.md
- 用于跟踪多轮研究进度和结果整合的外部存储文件。

Satisficing Criteria

满足性终止标准

Research continues in waves until satisficing criteria are met for the given thoroughness level.
研究将分轮次进行，直至达到对应详尽程度的终止标准。

Wave Continuation by Level

各等级轮次延续规则

LevelContinue WhenStop When (Satisficed)
quickN/AAlways single wave
mediumN/AAlways single wave
thoroughCritical gaps remain AND previous wave was productive AND ≤50% source overlap with prior wavesNo critical gaps OR diminishing returns OR >50% source overlap
very-thoroughSignificant gaps remain AND previous wave was productive AND ≤50% source overlap with prior wavesComprehensive coverage (no significant gaps) OR diminishing returns OR >50% source overlap
No hard maximum: For thorough and very-thorough, waves continue based on necessity, not arbitrary limits. The satisficing criteria drive when to stop.
Source overlap: Percentage of sources in current wave that were also cited in any previous wave. >50% overlap indicates research is cycling through same sources.
等级继续轮次的条件终止条件（满足要求）
快速不适用始终为单轮
中等不适用始终为单轮
深度存在关键缺口且上一轮有产出且本轮与之前轮次的来源重叠≤50%无关键缺口或收益递减或来源重叠>50%
极深度存在重要缺口且上一轮有产出且本轮与之前轮次的来源重叠≤50%全面覆盖（无重要缺口）或收益递减或来源重叠>50%
无硬性最大轮次：深度和极深度模式下，轮次延续基于实际需求，而非任意限制。终止判定条件决定何时停止。
来源重叠：本轮中与之前任何轮次引用过的来源占比。超过50%的重叠意味着研究在重复使用相同来源。

Gap Classification

缺口分类

After each wave, classify identified gaps:
Gap TypeDefinitionTriggers New Wave?
CriticalCore question aspect unanswered, major conflicts unresolved, key comparison missingYes (thorough, very-thorough)
SignificantImportant facet unexplored, partial answer needs depth, newly discovered areaYes (very-thorough only)
MinorNice-to-have detail, edge case unclear, tangential infoNo - note in limitations
每轮研究后，对发现的缺口进行分类：
缺口类型定义是否触发新轮次
关键核心问题未解答、重大冲突未解决、关键对比缺失是（深度、极深度模式）
重要重要方面未探索、部分答案需深化、发现新研究领域是（仅极深度模式）
次要锦上添花的细节、边缘情况不明确、无关信息否 - 在局限性中注明

Satisficing Evaluation

终止判定评估

After Phase 4 (Cross-Reference), evaluate whether to continue:
Definitions:
  • Finding: A distinct piece of information answering part of the research question, with at least one source citation. Multiple sources confirming the same fact count as one finding with higher confidence.
  • Substantive finding: A finding that provides new information not already established in previous waves. Variations or restatements of known information do not count.
  • High-authority source: Official documentation, peer-reviewed research, established news outlets (e.g., major tech publications), or sources from recognized domain experts. Company blogs about their own products count as high-authority for factual claims about that product.
  • Independent sources: Sources with different underlying information origins. Two articles citing the same primary source count as one source. Multiple pages from the same domain count as one source unless they represent different authors/teams with distinct research.
  • High confidence: Finding corroborated by ≥3 independent sources OR ≥2 high-authority sources.
  • Medium confidence: Finding corroborated by 2 independent sources OR 1 high-authority source.
  • Low confidence: Finding from a single non-authoritative source with no corroboration.
  • Medium+ confidence: Confidence level of Medium or High (i.e., not Low, Contested, or Inconclusive).
Satisficed when ANY true:
  • All critical gaps addressed (thorough) OR all significant gaps addressed (very-thorough)
  • Diminishing returns detected: new wave revealed <2 new substantive findings AND no finding's confidence increased by at least one level AND no new areas discovered
  • User explicitly requested stopping or a specific wave count
  • Comprehensive coverage achieved: all identified facets addressed with medium+ confidence
Continue when ALL true:
  • Gaps exist at the triggering threshold:
    • thorough: Critical gaps remain (core question unanswered, major conflicts)
    • very-thorough: Significant gaps remain (important facets unexplored, conflicts, newly discovered areas)
  • Previous wave was productive (≥2 new substantive findings OR ≥1 finding's confidence increased by at least one level OR new areas discovered)
  • Research is still yielding value (≤50% of sources in this wave were cited in previous waves)
完成第4阶段（交叉验证）后，评估是否继续研究：
定义说明：
  • 研究结果：可回答部分研究问题的独立信息片段，至少包含一个来源引用。多个来源证实同一事实计为一个高可信度结果。
  • 实质性新结果：提供了之前轮次未涉及的新信息的结果。已知信息的变体或重述不计入。
  • 高权威来源：官方文档、同行评审研究、知名新闻媒体（如主流科技出版物）、或公认领域专家的来源。企业关于自身产品的博客在涉及该产品的事实声明时视为高权威来源。
  • 独立来源：信息来源不同的渠道。两篇引用同一原始来源的文章计为一个来源。同一域名下的多个页面，若来自不同作者/团队且研究内容不同，可计为多个来源。
  • 高可信度：结果得到≥3个独立来源证实 或 ≥2个高权威来源证实。
  • 中等可信度：结果得到2个独立来源证实 或 1个高权威来源证实。
  • 低可信度：仅来自单个非权威来源且无其他佐证。
  • 中等及以上可信度：可信度为中等或高（即非低、有争议或无结论）。
满足以下任一条件则终止：
  • 所有关键缺口已填补（深度模式） 或 所有重要缺口已填补（极深度模式）
  • 检测到收益递减：新轮次产生的实质性新结果<2个 且 没有结果的可信度提升至少一个等级 且 未发现新领域
  • 用户明确要求停止或指定了轮次数量
  • 实现全面覆盖：所有已识别的方面均已得到中等及以上可信度的解答
满足以下所有条件则继续：
  • 存在对应等级的触发缺口：
    • 深度模式：仍存在关键缺口（核心问题未解答、重大冲突未解决）
    • 极深度模式：仍存在重要缺口（重要方面未探索、冲突未解决、发现新领域）
  • 上一轮有产出（≥2个实质性新结果 或 ≥1个结果的可信度提升至少一个等级 或 发现新领域）
  • 研究仍有价值（本轮中≤50%的来源在之前轮次中被引用过）

Wave Planning

轮次规划

When continuing to a new wave:
  1. Identify specific gaps to address (from Cross-Reference Analysis)
  2. Design targeted research prompts for each gap
  3. Assign 1-3 agents per wave (focused investigation)
  4. Update orchestration file with wave number and assignments
  5. Launch agents and collect findings
  6. Return to gap evaluation
Topic-slug format: Extract 2-4 key terms (nouns and adjectives that identify the topic; exclude articles, prepositions, and generic words like "best", "options", "analysis"), lowercase, replace spaces with hyphens. Example: "best real-time database options 2025" → 
real-time-database-options
Timestamp format: 
YYYYMMDD-HHMMSS
. Obtain via 
date +%Y%m%d-%H%M%S
.
启动新轮次时：
  1. 确定需填补的具体缺口（来自交叉验证分析）
  2. 针对每个缺口设计针对性研究提示词
  3. 每轮分配1-3个Agent（聚焦式调查）
  4. 更新orchestration文件，记录轮次编号和任务分配
  5. 启动Agent并收集结果
  6. 返回缺口评估环节
Topic-slug格式：提取2-4个关键词（识别主题的名词和形容词；排除冠词、介词和通用词如“最佳”“方案”“分析”），小写，用连字符替换空格。示例：“best real-time database options 2025” → 
real-time-database-options
时间戳格式：
YYYYMMDD-HHMMSS
。可通过
date +%Y%m%d-%H%M%S
获取。

Phase 1: Initial Setup (skip for quick)

阶段1：初始设置（快速模式跳过）

1.1 Get timestamp & create todo list

1.1 获取时间戳并创建待办事项列表

Run two commands:
  • date +%Y%m%d-%H%M%S
    → for filename timestamp (e.g., 
    20260112-060615
    )
  • date '+%Y-%m-%d %H:%M:%S'
    → for human-readable "Started" field (e.g., 
    2026-01-12 06:06:15
    )
Todos = research areas to investigate + write-to-log operations, not fixed steps. Each research todo represents a distinct angle or facet. List expands as decomposition reveals new areas. Write-to-log todos ensure external memory stays current.
Starter todos (seeds - list grows during decomposition):
- [ ] Create orchestration file; done when file created
- [ ] Topic decomposition→log; done when all facets identified
- [ ] (expand: research facets as decomposition reveals)
- [ ] Launch Wave 1 agents; done when all agents spawned
- [ ] Collect Wave 1 findings→log; done when all agents returned
- [ ] Cross-reference findings→log; done when agreements/conflicts mapped
- [ ] Evaluate gaps→log; done when gaps classified
- [ ] (expand: Wave 2+ if continuing)
- [ ] Refresh: read full orchestration file
- [ ] Synthesize→final output; done when all findings integrated + sourced
Critical todos (never skip):
  • →log
    after EACH phase/agent completion
  • Refresh:
    ALWAYS before synthesis
Expansion pattern: As decomposition reveals facets, add research todos:
- [x] Create orchestration file; file created
- [x] Topic decomposition→log; 4 facets identified
- [ ] Research: real-time database landscape 2025; done when options cataloged
- [ ] Research: performance benchmarks; done when latency/throughput data found
- [ ] Research: conflict resolution strategies; done when CRDT/OT patterns documented
- [ ] Research: production case studies; done when 3+ cases collected
- [ ] Launch Wave 1 agents (4 parallel); done when all agents spawned
- [ ] Collect Agent 1→log; done when findings written
- [ ] Collect Agent 2→log; done when findings written
- [ ] Collect Agent 3→log; done when findings written
- [ ] Collect Agent 4→log; done when findings written
- [ ] Cross-reference→log; done when agreements/conflicts mapped
- [ ] Evaluate gaps→log; done when gaps classified
- [ ] (expand: Wave 2 if continuing)
- [ ] Refresh: read full orchestration file
- [ ] Synthesize→final output; done when all findings integrated + sourced
执行两个命令：
  • date +%Y%m%d-%H%M%S
    → 用于文件名时间戳（如
    20260112-060615
    ）
  • date '+%Y-%m-%d %H:%M:%S'
    → 用于人类可读的“开始时间”字段（如
    2026-01-12 06:06:15
    ）
待办事项 = 需调查的研究领域 + 写入日志操作，而非固定步骤。每个研究待办事项代表一个独特的角度或方面。随着主题分解，列表会不断扩展。写入日志的待办事项确保外部存储保持最新。
初始待办事项（基础项 - 分解过程中会扩展）：
- [ ] 创建orchestration文件；文件创建完成后标记为已完成
- [ ] 主题分解→写入日志；确定所有研究方面后标记为已完成
- [ ] （扩展：分解过程中新增研究方面）
- [ ] 启动第1轮Agent；所有Agent启动完成后标记为已完成
- [ ] 收集第1轮结果→写入日志；所有Agent返回结果后标记为已完成
- [ ] 交叉验证结果→写入日志；完成一致性/冲突映射后标记为已完成
- [ ] 评估缺口→写入日志；完成缺口分类后标记为已完成
- [ ] （扩展：若继续则启动第2轮及以后）
- [ ] 刷新：读取完整orchestration文件
- [ ] 整合结果→最终输出；所有结果整合并添加来源后标记为已完成
关键待办事项（不可跳过）：
  • 每个阶段/Agent完成后必须执行
    →写入日志
  • 整合结果前必须执行
    刷新：
    操作
扩展示例：分解主题发现新方面后，添加研究待办事项：
- [x] 创建orchestration文件；已创建
- [x] 主题分解→写入日志；已识别4个研究方面
- [ ] 研究：2025年实时数据库格局；完成方案分类后标记为已完成
- [ ] 研究：性能基准测试；找到延迟/吞吐量数据后标记为已完成
- [ ] 研究：冲突解决策略；记录CRDT/OT模式后标记为已完成
- [ ] 研究：生产案例；收集3个以上案例后标记为已完成
- [ ] 启动第1轮Agent（4个并行）；所有Agent启动完成后标记为已完成
- [ ] 收集Agent 1结果→写入日志；结果写入后标记为已完成
- [ ] 收集Agent 2结果→写入日志；结果写入后标记为已完成
- [ ] 收集Agent 3结果→写入日志；结果写入后标记为已完成
- [ ] 收集Agent 4结果→写入日志；结果写入后标记为已完成
- [ ] 交叉验证→写入日志；完成一致性/冲突映射后标记为已完成
- [ ] 评估缺口→写入日志；完成缺口分类后标记为已完成
- [ ] （扩展：若继续则启动第2轮）
- [ ] 刷新：读取完整orchestration文件
- [ ] 整合结果→最终输出；所有结果整合并添加来源后标记为已完成

1.2 Create orchestration file (skip for quick)

1.2 创建orchestration文件（快速模式跳过）

Path: 
/tmp/research-orchestration-{topic-slug}-{YYYYMMDD-HHMMSS}.md
markdown
undefined
路径：
/tmp/research-orchestration-{topic-slug}-{YYYYMMDD-HHMMSS}.md
markdown
undefined

Web Research Orchestration: {topic}

Web研究协调：{主题}

Timestamp: {YYYYMMDD-HHMMSS} Started: {YYYY-MM-DD HH:MM:SS} Thoroughness: {level} Wave Policy: {single wave | continue while critical gaps | continue until comprehensive}
时间戳：{YYYYMMDD-HHMMSS} 开始时间：{YYYY-MM-DD HH:MM:SS} 详尽程度：{等级} 轮次策略：{单轮 | 存在关键缺口时持续 | 持续至全面覆盖}

Research Question

研究问题

{Clear statement of what needs to be researched}
{清晰的研究需求说明}

Topic Decomposition

主题分解

  • Core question: {main thing to answer}
  • Facets to investigate: (populated in Phase 2)
  • Expected researcher count: {based on thoroughness level}
  • 核心问题：{需解答的核心内容}
  • 待研究方面：（阶段2中填充）
  • 预计Agent数量：（基于详尽程度等级）

Wave Tracking

轮次跟踪

WaveAgentsFocusStatusNew FindingsDecision
1{count}Initial investigationPending--
轮次Agent数量研究重点状态新结果数量决策
1{数量}初始调查待启动--

Research Assignments

研究任务分配

(populated in Phase 2)
（阶段2中填充）

Agent Status

Agent状态

(updated as agents complete)
（Agent完成后更新）

Collected Findings

收集到的结果

(populated as agents return)
（Agent返回后填充）

Cross-Reference Analysis

交叉验证分析

(populated after each wave)
（每轮后填充）

Gap Evaluation

缺口评估

(populated after each wave - drives continuation decisions)
（每轮后填充 - 决定是否继续的依据）

Synthesis Notes

整合笔记

(populated in final phase)
undefined
（最终阶段填充）
undefined

Phase 2: Topic Decomposition & Agent Assignment

阶段2：主题分解与Agent任务分配

2.1 Decompose the research topic into ORTHOGONAL facets

2.1 将研究主题分解为独立的研究方面

Before launching agents, analyze the query to identify non-overlapping research angles. Each agent should have a distinct domain with clear boundaries.
  1. Core question: What is the fundamental thing being asked?
  2. Facets: What distinct aspects need investigation? Ensure minimal overlap:
    • Technical aspects (how it works, implementation details)
    • Comparison aspects (alternatives, competitors, trade-offs)
    • Practical aspects (real-world usage, adoption, case studies)
    • Current state (recent developments, 2025 updates)
    • Limitations/concerns (drawbacks, issues, criticisms)
  3. Orthogonality check: Before assigning agents, verify:
    • Each facet covers a distinct domain
    • No two facets would naturally search the same queries
    • Boundaries are clear enough to state explicitly
Bad decomposition (overlapping):
  • Agent 1: "Research Firebase"
  • Agent 2: "Research real-time databases" ← Firebase is a real-time database, overlap!
Good decomposition (orthogonal):
  • Agent 1: "Research Firebase specifically - features, pricing, limits"
  • Agent 2: "Research non-Firebase alternatives: Supabase, Convex, PlanetScale"
启动Agent前，分析查询以确定无重叠的研究角度。每个Agent应负责一个边界清晰的独立领域。
  1. 核心问题：用户本质上要问什么？
  2. 研究方面：需要调查哪些独立的方面？确保最小化重叠：
    • 技术方面（工作原理、实现细节）
    • 对比方面（替代方案、竞品、权衡）
    • 实践方面（实际应用、采用情况、案例研究）
    • 当前状态（最新进展、2025年更新）
    • 局限性/关注点（缺点、问题、批评）
  3. 独立性检查：分配Agent前，确认：
    • 每个方面覆盖一个独立领域
    • 任意两个方面不会导致重复搜索相同内容
    • 边界足够清晰可明确说明
错误分解示例（重叠）：
  • Agent 1：“研究Firebase”
  • Agent 2：“研究实时数据库” ← Firebase属于实时数据库，存在重叠！
正确分解示例（独立）：
  • Agent 1：“专门研究Firebase - 功能、定价、限制”
  • Agent 2：“研究非Firebase替代方案：Supabase、Convex、PlanetScale”

2.2 Plan agent assignments with explicit boundaries

2.2 规划Agent任务分配并明确边界

FacetResearch FocusExplicitly EXCLUDE
{facet 1}"{what to research}""{what other agents cover}"
{facet 2}"{what to research}""{what other agents cover}"
Agent count by level:
  • medium: 1-2 agents (core + one related angle)
  • thorough: 2-4 agents (core + alternatives + practical + concerns)
  • very-thorough: 4-6 agents (comprehensive coverage of all facets)
If decomposition reveals more facets than agent count allows:
  • Prioritize facets by: (1) directly answers core question, (2) enables comparison if requested, (3) addresses user-specified concerns
  • Combine related facets into single agent assignments where orthogonality allows
  • Schedule remaining facets for Wave 2 if initial wave is productive
Orthogonality strategies:
  • By entity: Agent 1 = Product A, Agent 2 = Product B (not both "products")
  • By dimension: Agent 1 = Performance, Agent 2 = Pricing, Agent 3 = Security
  • By time: Agent 1 = Current state, Agent 2 = Historical evolution
  • By perspective: Agent 1 = Official docs, Agent 2 = Community experience
研究方面研究重点明确排除范围
{方面1}“{研究内容}”“{其他Agent负责的方面}”
{方面2}“{研究内容}”“{其他Agent负责的方面}”
各等级Agent数量：
  • 中等：1-2个Agent（核心方面 + 1个相关角度）
  • 深度：2-4个Agent（核心方面 + 替代方案 + 实践 + 关注点）
  • 极深度：4-6个Agent（全面覆盖所有方面）
若分解出的方面超过可用Agent数量：
  • 按优先级排序：(1) 直接解答核心问题，(2) 若需要对比则支持对比，(3) 解决用户明确提出的关注点
  • 在保持独立性的前提下，将相关方面合并为单个Agent的任务
  • 若初始轮次有产出，剩余方面安排到第2轮
独立性策略：
  • 按实体划分：Agent 1 = 产品A，Agent 2 = 产品B（而非都负责“产品”）
  • 按维度划分：Agent 1 = 性能，Agent 2 = 定价，Agent 3 = 安全性
  • 按时间划分：Agent 1 = 当前状态，Agent 2 = 历史演变
  • 按视角划分：Agent 1 = 官方文档，Agent 2 = 社区反馈

2.3 Expand todos for each research area

2.3 为每个研究方面扩展待办事项

Add a todo for each planned agent assignment:
- [x] Topic decomposition & research planning
- [ ] Research: {facet 1 description}
- [ ] Research: {facet 2 description}
- [ ] Research: {facet 3 description}
- [ ] ...
- [ ] Collect and cross-reference findings
- [ ] Synthesize final output
为每个规划的Agent任务添加待办事项：
- [x] 主题分解与研究规划
- [ ] 研究：{方面1描述}
- [ ] 研究：{方面2描述}
- [ ] 研究：{方面3描述}
- [ ] ...
- [ ] 收集并交叉验证结果
- [ ] 整合最终输出

2.4 Update orchestration file

2.4 更新orchestration文件

After decomposition, update the file:
markdown
undefined
分解完成后，更新文件：
markdown
undefined

Topic Decomposition

主题分解

  • Core question: {main question}
  • Facets identified:
    1. {facet 1}: {why this angle matters}
    2. {facet 2}: {why this angle matters} ...
  • 核心问题：{核心问题}
  • 已识别的研究方面：
    1. {方面1}：{该角度的重要性}
    2. {方面2}：{该角度的重要性} ...

Research Assignments

研究任务分配

AgentFacetPromptStatus
1{facet}"{prompt}"Pending
2{facet}"{prompt}"Pending
...
undefined
Agent研究方面提示词状态
1{方面}“{提示词}”待启动
2{方面}“{提示词}”待启动
...
undefined

Phase 3: Launch Parallel Researchers

阶段3：启动并行研究Agent

3.1 Launch web-researcher agents

3.1 启动web-researcher Agent

Launch 
vibe-extras:web-researcher
agents for each research angle. Launch agents in parallel (single message with multiple agent invocations) to maximize efficiency.
Wave 1 prompt template (broad exploration with boundaries):
{Specific research question for this facet}

YOUR ASSIGNED SCOPE:
- Focus areas: {specific aspect 1}, {specific aspect 2}, {specific aspect 3}
- This is YOUR domain - go deep on these topics

DO NOT RESEARCH (other agents cover these):
- {facet assigned to Agent 2}
- {facet assigned to Agent 3}
- {etc.}

Current date context: {YYYY-MM-DD} - prioritize recent sources.

---
Research context:
- Wave: 1 (initial investigation)
- Mode: Broad exploration within your assigned scope
- Stay within your boundaries - other agents handle the excluded areas
- Report any gaps or conflicts you discover for potential follow-up waves
Wave 2+ prompt template (gap-filling):
{Specific gap or conflict to resolve}

Context from previous waves:
- Previous findings: {summary of relevant findings from earlier waves}
- Gap being addressed: {specific gap - e.g., "Sources conflict on X" or "Y aspect unexplored"}
- What we already know: {established facts from Wave 1}

YOUR ASSIGNED SCOPE:
- Focus narrowly on: {targeted aspect 1}, {targeted aspect 2}
- This gap was identified because: {why previous research was insufficient}

DO NOT RESEARCH:
- Topics already well-covered in Wave 1 (don't repeat)
- {areas other Wave 2 agents are handling}

Current date context: {YYYY-MM-DD} - prioritize recent sources.

---
Research context:
- Wave: {N} (gap-filling)
- Mode: Targeted investigation - focus narrowly on the gap above
- Build on previous findings, don't repeat broad exploration
- Flag if this gap cannot be resolved (conflicting authoritative sources, no data available, etc.)
Batching rules:
  • thorough: Launch all 2-4 agents in a single parallel batch
  • very-thorough: Launch in batches of 3-4 agents; for 5 agents use 3+2, for 6 agents use 3+3 (avoid overwhelming context)
  • Wave 2+: Launch 1-3 focused agents per wave
为每个研究角度启动
vibe-extras:web-researcher
Agent。并行启动Agent（单条消息中调用多个Agent）以最大化效率。
第1轮提示词模板（带边界的广泛探索）：
{该方面的具体研究问题}

你的任务范围：
- 重点领域：{具体方面1}, {具体方面2}, {具体方面3}
- 这是你的专属领域 - 深入研究这些主题

禁止研究（由其他Agent负责）：
- {Agent 2负责的方面}
- {Agent 3负责的方面}
- （其他）

当前日期：{YYYY-MM-DD} - 优先选择近期来源。

---
研究背景：
- 轮次：1（初始调查）
- 模式：在分配范围内广泛探索
- 严格遵守边界 - 其他Agent负责禁止研究的领域
- 报告发现的任何缺口或冲突，以便后续轮次跟进
第2轮及以后提示词模板（填补缺口）：
{需解决的具体缺口或冲突}

前序轮次背景：
- 前序结果：{前序轮次的相关结果摘要}
- 待填补的缺口：{具体缺口 - 如“来源对X的结论存在冲突”或“Y方面未探索”}
- 已知信息：{第1轮已确定的事实}

你的任务范围：
- 重点聚焦：{针对性方面1}, {针对性方面2}
- 该缺口的产生原因：{前序研究的不足}

禁止研究：
- 第1轮已充分覆盖的主题（避免重复）
- {其他第2轮Agent负责的领域}

当前日期：{YYYY-MM-DD} - 优先选择近期来源。

---
研究背景：
- 轮次：{N}（填补缺口）
- 模式：针对性调查 - 重点聚焦上述缺口
- 基于前序结果展开，避免重复广泛探索
- 若无法解决该缺口（权威来源冲突、无可用数据等），请标记说明
批量启动规则：
  • 深度模式：将所有2-4个Agent作为一个并行批次启动
  • 极深度模式：按3-4个Agent为一批次启动；5个Agent则分3+2，6个Agent分3+3（避免上下文过载）
  • 第2轮及以后：每轮启动1-3个聚焦式Agent

3.2 Update orchestration file after each agent completes

3.2 每个Agent完成后更新orchestration文件

After EACH agent returns, immediately update:
markdown
undefined
每个Agent返回结果后，立即更新文件：
markdown
undefined

Agent Status

Agent状态

AgentFacetStatusKey Finding
1{facet}Complete{1-sentence summary}
2{facet}Complete{1-sentence summary}
...
Agent研究方面状态关键结果
1{方面}已完成{1句话摘要}
2{方面}已完成{1句话摘要}
...

Collected Findings

收集到的结果

Agent 1: {facet}

Agent 1：{研究方面}

Confidence: {High/Medium/Low/Contested/Inconclusive} Sources: {count}
{Paste key findings from agent - preserve source citations} {If Contested: note the conflicting positions} {If Inconclusive: note what couldn't be determined}
可信度：{高/中/低/有争议/无结论} 来源数量：{数量}
{粘贴Agent返回的关键结果 - 保留来源引用} {若为有争议：注明冲突的不同观点} {若无结论：注明无法确定的原因}

Agent 2: {facet}

Agent 2：{研究方面}

...
undefined
...
undefined

3.3 Handle agent failures

3.3 处理Agent失败

If an agent times out or returns incomplete results:
  1. Note the gap in orchestration file
  2. Decide based on facet criticality:
    • Retry (narrower prompt) if: facet covers a Critical gap for the research question, OR facet is explicitly required by the research question for comparison/evaluation (e.g., query asks to compare X and Y, and facet covers X or Y), OR user explicitly requested this facet
    • Mark as gap (don't retry) if: facet covers a Significant or Minor gap, OR other agents partially covered the topic, OR research can synthesize without this facet
  3. Never block synthesis for a single failed agent - proceed with available findings and note the limitation
  4. If ALL agents in a wave fail:
    • For Wave 1: Retry with simpler decomposition (fewer agents, broader prompts)
    • For Wave 2+: Mark gaps as unresolvable, proceed to synthesis with prior wave findings
    • Always note the systemic failure in Gaps & Limitations
若Agent超时或返回不完整结果：
  1. 在orchestration文件中记录该缺口
  2. 根据方面的重要性决定：
    • 重试（缩小提示词范围）：若该方面对应研究问题的关键缺口，或研究问题明确要求对比/评估该方面（如查询要求对比X和Y，而该方面负责X或Y），或用户明确要求研究该方面
    • 标记为缺口（不重试）：若该方面对应重要或次要缺口，或其他Agent已部分覆盖该主题，或无需该方面也能完成结果整合
  3. 不要因单个Agent失败而阻碍结果整合 - 基于现有结果继续，并在局限性中注明
  4. 若一轮中所有Agent均失败：
    • 第1轮：简化分解（减少Agent数量、扩大提示词范围）后重试
    • 第2轮及以后：标记缺口为无法解决，基于前序轮次结果进行整合
    • 始终在“缺口与局限性”中记录系统性失败

Phase 4: Collect, Cross-Reference & Evaluate Gaps

阶段4：收集、交叉验证与缺口评估

4.1 Mark collection todo in_progress

4.1 将收集待办事项标记为进行中

4.2 Analyze findings across agents

4.2 跨Agent分析结果

Look for:
  • Agreements: Where do multiple agents reach similar conclusions?
  • Conflicts: Where do findings contradict? (includes agent-reported "Contested" findings)
  • Inconclusive: Areas where agents couldn't determine answers
  • Gaps: What wasn't covered by any agent?
  • Surprises: Unexpected findings that warrant highlighting
Handling agent confidence levels:
  • High/Medium/Low: Standard confidence - use for cross-referencing
  • Contested: Agent found high-authority sources that directly contradict each other - treat as a conflict requiring resolution or presentation of both positions
  • Inconclusive: Agent couldn't find agreement among sources - may warrant follow-up wave with different search angles
重点关注：
  • 一致性：多个Agent得出相似结论的地方？
  • 冲突：结果相互矛盾的地方？（包括Agent报告的“有争议”结果）
  • 无结论：Agent无法确定答案的领域
  • 缺口：所有Agent均未覆盖的内容？
  • 意外发现：值得突出的意外结果
Agent可信度处理规则：
  • 高/中/低：标准可信度 - 用于交叉验证
  • 有争议：Agent发现高权威来源直接矛盾 - 视为冲突，需解决或呈现双方观点
  • 无结论：Agent无法在来源中找到共识 - 可能需要后续轮次尝试不同搜索角度

4.3 Update orchestration file with cross-reference

4.3 更新orchestration文件的交叉验证内容

markdown
undefined
markdown
undefined

Cross-Reference Analysis

交叉验证分析

Agreements (High Confidence)

一致性结论（高可信度）

  • {Finding}: Supported by agents {1, 3, 4}
  • {Finding}: Confirmed across {count} sources
  • {结果}：得到Agent {1, 3, 4}的支持
  • {结果}：得到{数量}个来源的确认

Conflicts (Requires Judgment)

冲突（需判断）

  • {Topic}: Agent 1 says X, Agent 3 says Y
    • Resolution: {which to trust and why, or present both}
  • {Topic}: Agent 2 reported as Contested - {Position A} vs {Position B}
    • Resolution: {present both with supporting sources, or identify which is more authoritative}
  • {主题}：Agent 1认为X，Agent 3认为Y
    • 处理方式：{信任哪一方及原因，或呈现双方观点}
  • {主题}：Agent 2报告为有争议 - {观点A} vs {观点B}
    • 处理方式：{呈现双方观点及支持来源，或确定哪一方更权威}

Inconclusive Areas

无结论领域

  • {Topic}: Agent {N} couldn't determine - {reason}
    • Action: {follow-up wave with different angles, or note as limitation}
  • {主题}：Agent {N}无法确定 - {原因}
    • 行动：{后续轮次尝试不同角度，或在局限性中注明}

Gaps Identified

已识别的缺口

  • {What wasn't answered}
  • {Areas needing more research}
  • {未解答的内容}
  • {需进一步研究的领域}

Key Insights

关键洞察

  • {Synthesis observation 1}
  • {Synthesis observation 2}
undefined
  • {整合观察1}
  • {整合观察2}
undefined

4.4 Evaluate gaps and decide next wave (skip for quick/medium)

4.4 评估缺口并决定是否启动下一轮（快速/中等模式跳过）

For thorough and very-thorough levels, classify each gap:
markdown
undefined
深度和极深度模式下，对每个缺口进行分类：
markdown
undefined

Gap Evaluation (Wave {N})

缺口评估（第{N}轮）

Critical Gaps (triggers thorough/very-thorough continuation)

关键缺口（触发深度/极深度模式继续）

  • {Gap}: {Why critical - core question aspect unanswered}
  • {Gap}: {Why critical - major conflict unresolved}
  • {缺口}：{关键原因 - 核心问题未解答}
  • {缺口}：{关键原因 - 重大冲突未解决}

Significant Gaps (triggers very-thorough continuation)

重要缺口（触发极深度模式继续）

  • {Gap}: {Why significant - important facet unexplored}
  • {Gap}: {Why significant - partial answer needs depth}
  • {Gap}: {Why significant - newly discovered area worth exploring}
  • {缺口}：{重要原因 - 重要方面未探索}
  • {缺口}：{重要原因 - 部分答案需深化}
  • {缺口}：{重要原因 - 发现值得探索的新领域}

Minor Gaps (note in limitations, don't pursue)

次要缺口（在局限性中注明，不继续研究）

  • {Gap}: {Why minor - nice-to-have detail}
  • {缺口}：{次要原因 - 锦上添花的细节}

Wave Productivity Assessment

轮次产出评估

  • New substantive findings this wave: {count}
  • Confidence improvements: {which areas improved}
  • New areas discovered: {list or "none"}
  • Diminishing returns signals: {yes/no - explain}
  • 本轮产生的实质性新结果数量：{数量}
  • 可信度提升：{哪些领域的可信度提升}
  • 发现的新领域：{列表或“无”}
  • 收益递减信号：{是/否 - 说明}

Wave Decision

轮次决策

  • Current wave: {N}
  • Thoroughness level: {level}
  • Wave policy: {single wave | continue while critical gaps | continue until comprehensive}
  • Critical gaps remaining: {count}
  • Significant gaps remaining: {count}
  • Was this wave productive? {yes/no - ≥2 findings OR confidence improved OR new areas}
  • Decision: {CONTINUE to Wave N+1 | SATISFICED - proceed to synthesis}
  • Reason: {explain based on satisficing criteria - what gaps remain or why comprehensive}
undefined
  • 当前轮次：{N}
  • 详尽程度等级：{等级}
  • 轮次策略：{单轮 | 存在关键缺口时持续 | 持续至全面覆盖}
  • 剩余关键缺口数量：{数量}
  • 剩余重要缺口数量：{数量}
  • 本轮是否有产出：{是/否 - ≥2个新结果 或 可信度提升 或 发现新领域}
  • 决策：{继续第N+1轮 | 满足终止条件 - 进入整合阶段}
  • 理由：{基于终止判定条件的解释 - 剩余缺口情况或全面覆盖的依据}
undefined

4.5 Wave Decision Logic

4.5 轮次决策逻辑

If SATISFICED (any of these true):
  • Level is quick or medium → Proceed to Phase 5
  • No critical gaps (thorough) or no significant gaps (very-thorough) → Proceed to Phase 5
  • Diminishing returns: previous wave yielded <2 new substantive findings AND no finding's confidence increased by at least one level AND no new areas discovered → Proceed to Phase 5
  • Comprehensive coverage achieved: all identified facets addressed with medium+ confidence → Proceed to Phase 5
  • User explicitly requested stopping
If CONTINUE (all of these true):
  • Gaps exist at triggering threshold:
    • thorough: Critical gaps remain
    • very-thorough: Significant gaps remain
  • Previous wave was productive (≥2 new substantive findings OR ≥1 finding's confidence increased by at least one level OR new areas discovered)
  • Not cycling through same sources (≤50% of sources in this wave were cited in previous waves)
满足以下任一条件则终止：
  • 等级为快速或中等 → 进入阶段5
  • 无关键缺口（深度模式）或无重要缺口（极深度模式） → 进入阶段5
  • 收益递减：前一轮产生的实质性新结果<2个 且 没有结果的可信度提升至少一个等级 且 未发现新领域 → 进入阶段5
  • 实现全面覆盖：所有已识别的方面均已得到中等及以上可信度的解答 → 进入阶段5
  • 用户明确要求停止
满足以下所有条件则继续：
  • 存在对应等级的触发缺口：
    • 深度模式：仍存在关键缺口
    • 极深度模式：仍存在重要缺口
  • 前一轮有产出（≥2个实质性新结果 或 ≥1个结果的可信度提升至少一个等级 或 发现新领域）
  • 未重复使用相同来源（本轮与之前轮次的来源重叠≤50%）

4.6 Launch Next Wave (if continuing)

4.6 启动下一轮（若继续）

When continuing to a new wave:
  1. Update Wave Tracking table in orchestration file:
markdown
undefined
决定继续后：
  1. 更新orchestration文件中的轮次跟踪表：
markdown
undefined

Wave Tracking

轮次跟踪

WaveAgentsFocusStatusNew Findings
14Initial investigationComplete12 findings
22Gap-filling: {focus areas}In Progress-

2. **Add wave-specific todos**:
  • Wave 2: Investigate {critical gap 1}
  • Wave 2: Resolve conflict on {topic}
  • Wave 2: Deep-dive {significant gap}

3. **Design targeted prompts** for gaps:
   - Be specific: "Resolve conflict between X and Y regarding Z"
   - Include context: "Previous research found A, but need clarification on B"
   - Narrower scope than Wave 1 agents

4. **Launch 1-3 agents** for this wave (focused investigation)
   - Launch `vibe-extras:web-researcher` agents
   - Prompts reference specific gaps, not broad topics

5. **Collect findings** and return to 4.2 (cross-reference including new findings)
轮次Agent数量研究重点状态新结果数量
14初始调查已完成12个结果
22填补缺口：{重点领域}进行中-

2. **添加轮次专属待办事项**：
  • 第2轮：调查{关键缺口1}
  • 第2轮：解决{主题}的冲突
  • 第2轮：深入研究{重要缺口}

3. **针对缺口设计针对性提示词**：
   - 具体明确：“解决X和Y在Z方面的冲突”
   - 包含上下文：“前序研究发现A，但需明确B”
   - 范围比第1轮Agent更窄

4. **启动1-3个Agent**进行本轮研究（聚焦式调查）
   - 启动`vibe-extras:web-researcher` Agent
   - 提示词需指向具体缺口，而非宽泛主题

5. **收集结果**并返回至4.2步骤（交叉验证，包含新结果）

4.7 Mark collection todo complete (when proceeding to synthesis)

4.7 收集完成后标记待办事项为已完成（进入整合阶段时）

Phase 5: Synthesize & Output

阶段5：整合结果与输出

5.1 Refresh context (MANDATORY - never skip)

5.1 刷新上下文（必须执行 - 不可跳过）

CRITICAL: Read the FULL orchestration file to restore ALL findings, cross-references, gap evaluations, and wave tracking into context.
Why this matters: By this point, findings from multiple agents across potentially multiple waves have been written to the orchestration file. Context degradation means these details may have faded. Reading the full file immediately before synthesis brings all findings into recent context where attention is strongest.
Todo must show:
- [x] Refresh context: read full orchestration file  ← Must be marked complete before synthesis
- [ ] Synthesize final output
Verification: After reading, you should have access to:
  • All collected findings from every agent
  • Cross-reference analysis (agreements, conflicts, inconclusive)
  • Gap evaluations from each wave
  • Wave tracking with decisions
  • All source citations
关键：读取完整的orchestration文件，恢复所有结果、交叉验证、缺口评估和轮次跟踪信息至上下文。
原因：此时，多个Agent在多轮研究中产生的结果已写入orchestration文件。上下文衰减会导致这些细节被遗忘。整合前读取完整文件可将所有结果置于当前注意力集中的上下文中。
待办事项必须显示：
- [x] 刷新上下文：读取完整orchestration文件  ← 整合前必须标记为已完成
- [ ] 整合最终输出
验证：读取完成后，你应能获取：
  • 所有Agent的所有结果
  • 交叉验证分析（一致性、冲突、无结论）
  • 每轮的缺口评估
  • 带决策的轮次跟踪
  • 所有来源引用

5.2 Mark synthesis todo in_progress

5.2 将整合待办事项标记为进行中

5.3 Generate comprehensive output

5.3 生成全面输出

Only after completing 5.1 - synthesize ALL agent findings into a cohesive answer. Include:
markdown
undefined
仅在完成5.1后执行 - 将所有Agent的结果整合为连贯的答案。输出需包含：
markdown
undefined

Research Findings: {Topic}

研究结果：{主题}

Thoroughness: {level} | Waves: {count} | Researchers: {total across waves} | Total Sources: {aggregate} Overall Confidence: High/Medium/Low (based on agreement and source quality) Satisficing: {reason research concluded - e.g., "All significant gaps addressed" or "Diminishing returns after Wave 3"}
详尽程度：{等级} | 轮次数量：{数量} | 研究Agent总数：{所有轮次的总数} | 总来源数量：{汇总数} 整体可信度：高/中/低（基于一致性和来源质量） 终止原因：{研究结束的理由 - 如“所有重要缺口已填补”或“第3轮后出现收益递减”}

Executive Summary

执行摘要

{4-8 sentences synthesizing the key takeaway. What does the user need to know?}
{4-8句话的核心结论总结。用户需要知道什么？}

Detailed Findings

详细结果

{Major Finding Area 1}

{主要结果领域1}

{Synthesized insights with inline source citations from multiple agents}
{整合后的洞察，包含来自多个Agent的内联来源引用}

{Major Finding Area 2}

{主要结果领域2}

{...}
{...}

Comparison/Evaluation (if applicable)

对比/评估（若适用）

OptionProsConsBest For
{opt 1}{from agents}{from agents}{synthesis}
{opt 2}{from agents}{from agents}{synthesis}
方案优点缺点适用场景
{方案1}{来自Agent的结果}{来自Agent的结果}{整合结论}
{方案2}{来自Agent的结果}{来自Agent的结果}{整合结论}

Recommendations

建议

{Based on synthesized evidence - what should the user consider/do?}
{基于整合后的证据 - 用户应考虑/执行什么？}

Confidence Notes

可信度说明

  • High confidence: {findings with strong multi-source agreement}
  • Medium confidence: {findings with some support}
  • Contested: {where high-authority sources directly contradicted - present both positions}
  • Inconclusive: {where agents couldn't determine answers despite searching}
  • Low confidence: {single source or weak agreement}
  • 高可信度：{得到多来源强支持的结果}
  • 中等可信度：{有一定支持的结果}
  • 有争议：{高权威来源直接矛盾的地方 - 呈现双方观点}
  • 无结论：{Agent搜索后仍无法确定答案的领域}
  • 低可信度：{仅单个来源或支持薄弱的结果}

Research Progression (for multi-wave)

研究进展（多轮研究时）

WaveFocusAgentsKey Contribution
1Initial investigation{N}{what this wave established}
2{Gap focus}{N}{what this wave resolved}
............
轮次研究重点Agent数量核心贡献
1初始调查{N}{本轮确定的内容}
2{缺口重点}{N}{本轮解决的问题}
............

Gaps & Limitations

缺口与局限性

  • {What couldn't be definitively answered despite multi-wave investigation}
  • {Areas where more research would help}
  • {Potential biases in available sources}
  • {Gaps intentionally not pursued (minor priority)}
  • {多轮研究后仍无法明确回答的内容}
  • {需要更多研究的领域}
  • {可用来源的潜在偏见}
  • {故意未研究的缺口（优先级低）}

Source Summary

来源汇总

SourceAuthorityDateUsed ForWave
{url}High/Med{date}{finding}1
...
Orchestration file: {path} Research completed: {timestamp}
undefined
来源权威性日期用于支持的结论轮次
{链接}高/中{日期}{结论}1
...
Orchestration文件路径：{路径} 研究完成时间：{时间戳}
undefined

5.4 Mark all todos complete

5.4 标记所有待办事项为已完成

Quick Mode Flow

快速模式流程

For quick (single-fact) queries, skip orchestration:
  1. State: 
    **Thoroughness**: quick — [reason]
  2. Launch a 
    vibe-extras:web-researcher
    agent with: "{query}"
  3. Return agent's findings directly (no synthesis overhead)
针对快速模式（单一事实查询），跳过orchestration步骤：
  1. 声明：
    **详尽程度**：快速 — [理由]
  2. 启动一个
    vibe-extras:web-researcher
    Agent，提示词为："{查询内容}"
  3. 直接返回Agent的结果（无需整合开销）

Key Principles

核心原则

PrincipleRule
Thoroughness firstDetermine level before any research
Todos with write-to-logEach collection gets a todo, followed by a write-to-orchestration-file todo
Write after each phaseWrite to orchestration file after EACH phase/agent
Parallel executionLaunch multiple agents simultaneously when possible
Cross-referenceCompare findings across agents before synthesizing
Gap evaluationClassify gaps after each wave (critical/significant/minor)
Wave iterationContinue waves until satisficed OR diminishing returns
Context refreshRead full orchestration file BEFORE synthesis - non-negotiable
Source preservationMaintain citations through synthesis
Gap honestyExplicitly state what couldn't be answered despite multi-wave effort
Log Pattern Summary:
  1. Create orchestration file at start
  2. Add write-to-log todos after each collection phase
  3. Write to it after EVERY step (decomposition, agent findings, cross-reference, gap evaluation)
  4. "Refresh context: read full orchestration file" todo before synthesis
  5. Read FULL file before synthesis (restores all context)
原则规则
先确定详尽程度开始任何研究前先确定等级
待办事项需包含写入日志每次收集后添加待办事项，随后执行写入orchestration文件的操作
每个阶段后写入日志每个阶段/Agent完成后必须写入orchestration文件
并行执行尽可能同时启动多个Agent
交叉验证整合前跨Agent对比结果
缺口评估每轮后对缺口进行分类（关键/重要/次要）
轮次迭代持续启动轮次直至满足终止条件或出现收益递减
上下文刷新整合前必须读取完整orchestration文件 - 不可协商
保留来源整合过程中保留来源引用
坦诚说明缺口明确说明多轮研究后仍无法回答的内容
日志模式总结：
  1. 开始时创建orchestration文件
  2. 每次收集阶段后添加写入日志的待办事项
  3. 每个步骤（分解、Agent结果、交叉验证、缺口评估）完成后均写入文件
  4. 整合前添加“刷新上下文：读取完整orchestration文件”待办事项
  5. 整合前必须读取完整文件（恢复所有上下文）

Never Do

禁止操作

  • Launch agents without determining thoroughness level
  • Skip write-to-log todos (every collection must be followed by a write todo)
  • Proceed to next phase without writing findings to orchestration file
  • Synthesize without completing "Refresh context: read full orchestration file" todo first
  • Skip orchestration file updates after agent completions
  • Present synthesized findings without source citations
  • Ignore conflicts between agent findings (especially "Contested" findings)
  • Skip gap evaluation for thorough/very-thorough levels
  • Continue waves when diminishing returns detected (wasted effort)
  • Stop prematurely when critical gaps remain (thorough) or significant gaps remain (very-thorough) and waves are still productive
  • 未确定详尽程度就启动Agent
  • 跳过写入日志的待办事项（每次收集后必须执行写入操作）
  • 未将结果写入orchestration文件就进入下一阶段
  • 未完成“刷新上下文：读取完整orchestration文件”待办事项就进行整合
  • Agent完成后跳过更新orchestration文件
  • 呈现整合结果时不包含来源引用
  • 忽略Agent结果之间的冲突（尤其是“有争议”结果）
  • 深度/极深度模式下跳过缺口评估
  • 检测到收益递减时仍继续启动轮次（浪费资源）
  • 仍存在关键缺口（深度模式）或重要缺口（极深度模式）且轮次仍有产出时提前停止

Example: Technology Comparison

示例：技术对比

Query: "Compare the best real-time databases for a collaborative app in 2025"
Thoroughness: thorough — comparison query requiring multi-angle investigation
Decomposition:
  • Facet 1: Real-time database landscape 2025 (what options exist)
  • Facet 2: Performance and scalability comparisons
  • Facet 3: Collaborative app requirements (conflict resolution, sync)
  • Facet 4: Production experiences and case studies
Agents launched (parallel):
  1. "Real-time database options 2025: Firebase, Supabase, Convex, others. Current market landscape."
  2. "Real-time database performance benchmarks and scalability. Latency, throughput, concurrent users."
  3. "Conflict resolution and sync strategies for collaborative apps. CRDTs, OT, last-write-wins."
  4. "Production case studies using real-time databases. Companies, scale, lessons learned."
Output: Synthesized comparison table with recommendations based on use case, backed by cross-referenced sources from all four agents.
查询：“对比2025年适用于协作应用的最佳实时数据库”
详尽程度：深度 — 对比类查询需要多角度调查
分解结果：
  • 方面1：2025年实时数据库格局（现有方案）
  • 方面2：性能与可扩展性对比
  • 方面3：协作应用需求（冲突解决、同步）
  • 方面4：生产实践与案例研究
启动的Agent（并行）：
  1. “2025年实时数据库方案：Firebase、Supabase、Convex及其他。当前市场格局。”
  2. “实时数据库性能基准与可扩展性。延迟、吞吐量、并发用户数。”
  3. “协作应用的冲突解决与同步策略。CRDT、OT、最后写入获胜机制。”
  4. “使用实时数据库的生产案例研究。企业、规模、经验总结。”
输出：基于所有4个Agent的交叉验证来源，生成包含建议的综合对比表格。

Example: Multi-Wave Comprehensive Research

示例：多轮全面研究

Query: "Give me a comprehensive analysis of all the AI coding assistant options in 2025"
Thoroughness: very-thorough — "comprehensive analysis" + "all options" triggers maximum depth
查询：“为我提供2025年所有AI编码助手的全面分析”
详尽程度：极深度 — “全面分析” + “所有方案”触发最高深度

Wave 1: Initial Investigation

第1轮：初始调查

Decomposition (6 orthogonal facets):
  • Facet 1: Market landscape - what tools exist (names only, no features/pricing)
  • Facet 2: Feature comparison - autocomplete, chat, agents, IDE support (no pricing)
  • Facet 3: Pricing and licensing - costs, tiers, enterprise deals (no features)
  • Facet 4: Enterprise/security - compliance, SOC2, on-prem (no general features)
  • Facet 5: Developer sentiment - reviews, community feedback (no official docs)
  • Facet 6: Recent news - announcements, launches, acquisitions (no evergreen content)
Agents launched with explicit boundaries (parallel batch of 4, then 2):
  1. "AI coding assistant market landscape 2025. YOUR SCOPE: List all tools (Copilot, Cursor, Claude Code, Codeium, etc). DO NOT RESEARCH: features, pricing, reviews."
  2. "AI coding assistant features 2025. YOUR SCOPE: autocomplete, chat, agentic capabilities, IDE support. DO NOT RESEARCH: pricing, enterprise security, user reviews."
  3. "AI coding assistant pricing 2025. YOUR SCOPE: subscription costs, usage-based models, free tiers. DO NOT RESEARCH: features, security compliance."
  4. "Enterprise AI coding assistant compliance 2025. YOUR SCOPE: SOC2, HIPAA, on-premise, data residency. DO NOT RESEARCH: general features, consumer pricing."
  5. "AI coding assistant developer sentiment 2025. YOUR SCOPE: Reddit, HN, Twitter discussions, community feedback. DO NOT RESEARCH: official documentation, pricing pages."
  6. "AI coding assistant news 2025. YOUR SCOPE: recent announcements, launches, acquisitions since Jan 2025. DO NOT RESEARCH: established features, pricing."
Gap Evaluation (Wave 1):
  • Critical gaps: None (all facets had substantial findings)
  • Significant gaps:
    • Conflict: Sources disagree on which tool has best agentic capabilities
    • Partial answer: Enterprise pricing not fully detailed for all options
    • New discovery: Several sources mention "AI code review" as emerging category
  • Minor gaps: Specific latency benchmarks, rare IDE integrations
Wave Decision: CONTINUE — 3 significant gaps remain, Wave 1 was productive (18 findings), research still yielding new information
分解结果（6个独立方面）：
  • 方面1：市场格局 - 现有工具（仅名称，不含功能/定价）
  • 方面2：功能对比 - 自动补全、聊天、Agent、IDE支持（不含定价）
  • 方面3：定价与许可 - 成本、套餐、企业协议（不含功能）
  • 方面4：企业/安全性 - 合规性、SOC2、本地部署（不含通用功能）
  • 方面5：开发者反馈 - 评价、社区反馈（不含官方文档）
  • 方面6：最新动态 - 公告、发布、收购（不含常规内容）
启动带明确边界的Agent（先并行启动4个，再启动2个）：
  1. “2025年AI编码助手市场格局。你的任务范围：列出所有工具（Copilot、Cursor、Claude Code、Codeium等）。禁止研究：功能、定价、评价。”
  2. “2025年AI编码助手功能。你的任务范围：自动补全、聊天、Agent能力、IDE支持。禁止研究：定价、企业安全、用户评价。”
  3. “2025年AI编码助手定价。你的任务范围：订阅成本、基于使用量的模式、免费套餐。禁止研究：功能、安全合规。”
  4. “2025年企业级AI编码助手合规性。你的任务范围：SOC2、HIPAA、本地部署、数据驻留。禁止研究：通用功能、个人版定价。”
  5. “2025年AI编码助手开发者反馈。你的任务范围：Reddit、HN、Twitter讨论、社区反馈。禁止研究：官方文档、定价页面。”
  6. “2025年AI编码助手最新动态。你的任务范围：2025年1月以来的公告、发布、收购。禁止研究：已确立的功能、定价。”
第1轮缺口评估：
  • 关键缺口：无（所有方面均有实质性结果）
  • 重要缺口：
    • 冲突：来源对哪款工具的Agent能力最佳存在分歧
    • 部分答案不完整：部分方案的企业定价未详细说明
    • 新发现：多个来源提到“AI代码审查”为新兴类别
  • 次要缺口：特定延迟基准测试、罕见IDE集成
轮次决策：继续 — 仍存在3个重要缺口，第1轮有产出（18个结果），研究仍在产生新信息

Wave 2: Gap-Filling

第2轮：填补缺口

Focus: Resolve agentic capabilities conflict, deepen enterprise pricing, explore AI code review
Agents launched (3 focused):
  1. "Compare agentic capabilities: Cursor Composer vs Claude Code vs GitHub Copilot Workspace 2025"
  2. "Enterprise AI coding assistant pricing 2025: Copilot Business, Cursor Teams, volume discounts"
  3. "AI code review tools 2025: CodeRabbit, Sourcery, Codacy AI. Emerging category analysis."
Gap Evaluation (Wave 2):
  • Critical gaps: None
  • Significant gaps: None remaining (conflict resolved, enterprise pricing clarified)
  • Minor gaps: Some niche tools not fully covered
Wave Decision: SATISFICED — No significant gaps remaining, 2 waves complete
研究重点：解决Agent能力的冲突、深化企业定价分析、探索AI代码审查类别
启动的Agent（3个聚焦式）：
  1. “对比Agent能力：Cursor Composer vs Claude Code vs GitHub Copilot Workspace 2025”
  2. “2025年企业级AI编码助手定价：Copilot Business、Cursor Teams、批量折扣”
  3. “2025年AI代码审查工具：CodeRabbit、Sourcery、Codacy AI。新兴类别分析。”
第2轮缺口评估：
  • 关键缺口：无
  • 重要缺口：无剩余（冲突已解决，企业定价已明确）
  • 次要缺口：部分小众工具未完全覆盖
轮次决策：满足终止条件 — 无重要缺口剩余，完成2轮研究

Output Summary

输出摘要

Thoroughness: very-thorough | Waves: 2 | Researchers: 9 | Sources: 34 Satisficing: All significant gaps addressed — comprehensive coverage achieved
详尽程度：极深度 | 轮次数量：2 | 研究Agent总数：9 | 来源数量：34 终止原因：所有重要缺口已填补 — 实现全面覆盖