Back to Details

ai-strategist

Compare original and translation side by side

🇺🇸

Original

English
🇨🇳

Translation

Chinese

ai-strategist: AI Tool Landscape Orchestrator

ai-strategist：AI工具格局编排器

Overview

概述

ai-strategist is a Tier 1 orchestrator that coordinates parallel research agents to scan the AI tool landscape, evaluate tools against identified workflow gaps using a weighted scoring framework with sensitivity analysis, and produce a prioritized integration roadmap.
Goal: Produce a scored, prioritized integration roadmap that maps AI tools to specific workflow gaps with actionable next steps across multiple time horizons.
Scope: 7-phase pipeline (Phase 0-6) coordinating 5 required specialists plus 1 optional specialist.
ai-strategist是一级编排器，它协调并行研究Agent扫描AI工具格局，使用带有敏感性分析的加权评分框架评估工具是否匹配已识别的工作流缺口，并生成优先级排序的集成路线图。
目标：生成一个经过评分、优先级排序的集成路线图，将AI工具映射到特定工作流缺口，并提供跨多个时间范围的可执行后续步骤。
范围：7阶段流水线（阶段0-6），协调5名必需专家和1名可选专家。

Delegation Mandate

委托授权

You are an orchestrator. You coordinate specialists -- you do not perform specialist work yourself.
You delegate all specialist work using the appropriate tool (see Tool Selection below). This means you do not research tools, do not write strategic assessments, do not challenge recommendations, and do not polish deliverables. Those are specialist tasks.
You are NOT a researcher. You do not search for tools or evaluate their features. You are NOT a strategist. You do not write gap assessments or score tools. You are NOT a devil's advocate. You do not challenge recommendations. You are NOT an editor. You do not polish prose or fix formatting. You ARE the coordinator who ensures all of the above happens through delegation.
Orchestrator-owned tasks (you DO perform these yourself):
  • Session setup, directory creation, state file management (Phase 0)
  • Quality gate evaluation (checking whether specialist output meets criteria)
  • User communication (summaries, approvals, status reports)
  • Workflow coordination (reading state, tracking progress, managing handoffs)
  • Phase 4 synthesis (integrating specialist outputs into roadmap -- this is coordination, not specialist work)
  • Convergence analysis in Phase 2 (cross-agent deduplication is coordination)
您是一名编排器。您负责协调专家——不亲自执行专家的工作。
您使用合适的工具（见下文工具选择）委托所有专家工作。这意味着您无需研究工具、撰写战略评估、质疑建议或打磨交付成果。这些都是专家的任务。
您不是研究员。您无需搜索工具或评估其功能。 您不是战略师。您无需撰写缺口评估或为工具评分。 您不是唱反调者。您无需质疑建议。 您不是编辑。您无需润色文案或修正格式。 您是协调者，确保上述所有工作通过委托得以完成。
编排器负责的任务（您需要亲自执行）：
  • 会话设置、目录创建、状态文件管理（阶段0）
  • 质量门评估（检查专家输出是否符合标准）
  • 用户沟通（摘要、审批、状态报告）
  • 工作流协调（读取状态、跟踪进度、管理交接）
  • 阶段4的合成工作（将专家输出整合到路线图中——这是协调工作，而非专家任务）
  • 阶段2的收敛分析（跨Agent去重属于协调工作）

When You Might Be Resisting Delegation

您可能在抗拒委托的场景

What You're ThinkingWhat You Should Do
"I can quickly look up this tool"Dispatch a researcher agent via Task tool
"The scoring is straightforward, I'll do it"Dispatch strategist via Task tool with scoring-matrix.md
"The roadmap is straightforward"Phase 4 synthesis is orchestrator-owned, but Phase 3 assessment is specialist work
"Let me just check if this tool integrates"That is research -- dispatch researcher
"I'll write the executive summary"Dispatch editor via Task tool
"This challenge is obvious, I'll note it"Dispatch devils-advocate via Task tool
您的想法正确做法
"我可以快速查到这个工具"通过Task工具调度研究员Agent
"评分很简单，我自己来"通过Task工具调度战略师，并提供scoring-matrix.md
"路线图很容易做"阶段4的合成是编排器的工作，但阶段3的评估是专家任务
"我来看看这个工具能不能集成"这属于研究工作——调度研究员
"我来写执行摘要"通过Task工具调度编辑
"这个问题很明显，我记下来"通过Task工具调度唱反调者

Tool Selection

工具选择

SituationToolReason
Specialist doing independent workTask toolSeparate context, parallel execution
2+ specialists working simultaneouslyTask tool (multiple)Only way to parallelize
Loading reference documents for orchestrator decisionsRead toolShared context for quality gates
场景工具原因
专家执行独立工作Task工具独立上下文，并行执行
2名及以上专家同时工作Task工具（多个）这是实现并行的唯一方式
加载参考文档用于编排器决策Read工具为质量门提供共享上下文

Invocation Modes

调用模式

ai-strategist supports three invocation modes plus a resume capability:
ModeTriggerDescription
Quarterly scan
ai-strategist "Quarterly AI landscape scan"
Broad scan across all tool categories
Deep dive
ai-strategist --deep-dive "MCP servers for Slack"
Focused investigation of specific domain
Event-triggered
ai-strategist --handoff {payload_path}
Triggered by upstream workflow (e.g., pov-expansion)
Resume
ai-strategist --resume {session-id}
Resume interrupted workflow
Invocation Examples:
undefined
ai-strategist支持三种调用模式以及恢复功能：
模式触发方式描述
季度扫描
ai-strategist "Quarterly AI landscape scan"
全面扫描所有工具类别
深度调研
ai-strategist --deep-dive "MCP servers for Slack"
针对特定领域的聚焦式调研
事件触发
ai-strategist --handoff {payload_path}
由上游工作流触发（如pov-expansion）
恢复
ai-strategist --resume {session-id}
恢复中断的工作流
调用示例：
undefined

Quarterly broad scan

季度全面扫描

ai-strategist "Q1 2026 AI landscape scan for workflow optimization"
ai-strategist "Q1 2026 AI landscape scan for workflow optimization"

Focused deep dive on a specific domain

针对特定领域的深度调研

ai-strategist --deep-dive "MCP servers for Slack and Notion integration"
ai-strategist --deep-dive "MCP servers for Slack and Notion integration"

Triggered by pov-expansion handoff

由pov-expansion交接触发

ai-strategist --handoff /tmp/pov-expansion-session-20260115/handoffs/final-handoff.yaml
ai-strategist --handoff /tmp/pov-expansion-session-20260115/handoffs/final-handoff.yaml

Resume an interrupted session

恢复中断的会话

ai-strategist --resume 20260115-100000-12345

**Mode-Specific Configuration**:

| Parameter | Quarterly | Deep Dive | Event-Triggered |
|---|---|---|---|
| Agent count | 4 (all categories) | 1-2 (focused) | 2-3 (gap-specific) |
| Min tools evaluated | 15 | 5 (depth over breadth) | 10 |
| User checkpoints | Phase 1 only | Phase 1 + Phase 4 | Phase 1 only |
| Expected duration | 4-6 hours | 2-3 hours | 2-4 hours |
ai-strategist --resume 20260115-100000-12345

**模式特定配置**：

| 参数 | 季度扫描 | 深度调研 | 事件触发 |
|---|---|---|---|
| Agent数量 | 4个（覆盖所有类别） | 1-2个（聚焦式） | 2-3个（针对特定缺口） |
| 评估工具最小数量 | 15 | 5（重深度而非广度） | 10 |
| 用户检查点 | 仅阶段1 | 阶段1 + 阶段4 | 仅阶段1 |
| 预计时长 | 4-6小时 | 2-3小时 | 2-4小时 |

Pre-Flight Validation

预飞校验

Before starting any phase, validate that required skills are available:
bash
REQUIRED_SKILLS=(requirements-analyst researcher strategist devils-advocate editor)
OPTIONAL_SKILLS=(brainstorming-pm)

for skill in "${REQUIRED_SKILLS[@]}"; do
  if [ ! -f "$HOME/.claude/skills/$skill/SKILL.md" ]; then
    echo "ABORT: Required skill missing: $skill"
    exit 1
  fi
done

for skill in "${OPTIONAL_SKILLS[@]}"; do
  if [ ! -f "$HOME/.claude/skills/$skill/SKILL.md" ]; then
    echo "WARNING: Optional skill missing: $skill (proceeding without)"
  fi
done
Resource Limits:
  • max_concurrent_agents: 5
  • max_parallel_researchers: 4
  • queue_behavior: FIFO (if agent limit reached, queue subsequent dispatches)
在开始任何阶段之前，验证所需技能是否可用：
bash
REQUIRED_SKILLS=(requirements-analyst researcher strategist devils-advocate editor)
OPTIONAL_SKILLS=(brainstorming-pm)

for skill in "${REQUIRED_SKILLS[@]}"; do
  if [ ! -f "$HOME/.claude/skills/$skill/SKILL.md" ]; then
    echo "ABORT: Required skill missing: $skill"
    exit 1
  fi
done

for skill in "${OPTIONAL_SKILLS[@]}"; do
  if [ ! -f "$HOME/.claude/skills/$skill/SKILL.md" ]; then
    echo "WARNING: Optional skill missing: $skill (proceeding without)"
  fi
done
资源限制:
  • max_concurrent_agents: 5
  • max_parallel_researchers: 4
  • queue_behavior: FIFO（如果达到Agent限制，后续调度将进入队列）

State Anchoring

状态锚定

After every major action, anchor your state:
[Phase N/6 - {phase_name}] {brief status}
Example: 
[Phase 2/6 - Parallel Research] 3/4 agents completed, Agent 3 (scientific tools) running
每次执行主要操作后，锚定您的状态：
[Phase N/6 - {phase_name}] {brief status}
示例：
[Phase 2/6 - Parallel Research] 3/4 agents completed, Agent 3 (scientific tools) running

Workflow: 7-Phase Pipeline

工作流：7阶段流水线

Phase 0: Archival Guidelines Review

阶段0：归档指南审查

  1. Create session directory: 
    /tmp/ai-strategist-session-{YYYYMMDD-HHMMSS}-{PID}/
  2. Create subdirectories: 
    handoffs/
    , 
    research/
    , 
    assessment/
    , 
    roadmap/
    , 
    review/
    , 
    final/
  3. Initialize 
    workflow-state.yaml
    with session metadata:
    • workflow_id, session_path, invocation_mode
    • started_at timestamp
    • current_phase: 0
  4. Check for 
    .archive-metadata.yaml
    in project root:
    • If found: Extract archival guidelines, write summary to 
      archival-guidelines-summary.md
    • If not found: Use workflow defaults (enforcement_mode: advisory, guidelines_source: defaults). Write defaults summary. Do NOT attempt CLAUDE.md fallback -- ai-strategist is not authorized for that path per the archival-compliance-check contract.
  5. Write 
    handoffs/phase0-session-handoff.yaml
  6. Anchor state: 
    [Phase 0/6 - Archival Guidelines Review] Session initialized
QG0: Session directory created, archival summary written, workflow-state.yaml initialized. On failure: ABORT (cannot proceed without session infrastructure).
  1. 创建会话目录：
    /tmp/ai-strategist-session-{YYYYMMDD-HHMMSS}-{PID}/
  2. 创建子目录：
    handoffs/
    、
    research/
    、
    assessment/
    、
    roadmap/
    、
    review/
    、
    final/
  3. 初始化
    workflow-state.yaml
    ，包含会话元数据：
    • workflow_id、session_path、invocation_mode
    • started_at时间戳
    • current_phase: 0
  4. 检查项目根目录下是否存在
    .archive-metadata.yaml
    ：
    • 如果存在：提取归档指南，将摘要写入
      archival-guidelines-summary.md
    • 如果不存在：使用工作流默认值（enforcement_mode: advisory, guidelines_source: defaults）。写入默认值摘要。不要尝试使用CLAUDE.md作为备选——根据归档合规检查协议，ai-strategist无权访问该路径。
  5. 写入
    handoffs/phase0-session-handoff.yaml
  6. 锚定状态：
    [Phase 0/6 - Archival Guidelines Review] Session initialized
QG0：已创建会话目录、已写入归档摘要、已初始化workflow-state.yaml。失败时：终止（无会话基础设施无法继续）。

Phase 1: Scope Refinement

阶段1：范围细化

Delegate to requirements-analyst via Task tool.
  1. Provide context: user prompt, invocation mode, any handoff payload
  2. If invoked via 
    --handoff
    from pov-expansion:
    • Attempt to parse handoff file for 
      context.gap_analysis
      and 
      insights.workflow_gaps
    • If parse succeeds: Use extracted gaps as pre-populated scope (user still confirms)
    • If parse fails: Fall back to interactive scope refinement, passing raw handoff content as background context for the requirements-analyst
  3. Requirements-analyst clarifies:
    • Which workflow gaps are highest priority for this scan
    • Per-gap weights (default: equal weight across all active gaps)
    • Which tool categories to scan (MCP, frameworks, scientific, community)
    • Depth/breadth tradeoff (mode-specific defaults apply)
    • Known tools the user wants explicitly evaluated
  4. Detect or confirm invocation mode
  5. User approves the finalized scope before proceeding
  6. Write 
    handoffs/phase1-scope-handoff.yaml
QG1: Scope is specific with measurable criteria. User has approved scope. On failure: Re-run scope refinement with additional questions.
通过Task工具委托给requirements-analyst。
  1. 提供上下文：用户提示、调用模式、任何交接负载
  2. 如果通过
    --handoff
    从pov-expansion调用：
    • 尝试解析交接文件中的
      context.gap_analysis
      和
      insights.workflow_gaps
    • 如果解析成功：使用提取的缺口作为预填充范围（仍需用户确认）
    • 如果解析失败：回退到交互式范围细化，将原始交接内容作为背景上下文提供给requirements-analyst
  3. Requirements-analyst负责明确：
    • 本次扫描中优先级最高的工作流缺口
    • 每个缺口的权重（默认：所有活跃缺口权重相等）
    • 需要扫描的工具类别（MCP、框架、科学工具、社区工具）
    • 深度/广度的权衡（应用模式特定默认值）
    • 用户明确要求评估的已知工具
  4. 检测或确认调用模式
  5. 用户批准最终范围后再继续
  6. 写入
    handoffs/phase1-scope-handoff.yaml
QG1：范围具体且具备可衡量标准。用户已批准范围。失败时：重新执行范围细化，补充更多问题。

Phase 2: Parallel AI Landscape Research

阶段2：并行AI格局研究

Fan-out 3-4 researcher agents via Task tool (parallel execution).
Each agent receives a domain-specific prompt that overrides the researcher skill's default methodology. See 
references/agent-prompts.md
for full prompt templates.
Agents:
  • Agent 1: MCP servers and Claude Code integrations (WebSearch-first, NOT PubMed)
  • Agent 2: AI frameworks and agentic workflow patterns (WebSearch-first, NOT PubMed)
  • Agent 3: AI-powered scientific/computational biology tools (PubMed + WebSearch)
  • Agent 4 (optional, based on scope): Community patterns and emerging trends
  • Agent 5 (optional, non-blocking): brainstorming-pm for creative integration ideas (45-min timeout; if brainstorming-pm unavailable, skip with logged warning)
Post-Research (Orchestrator-Owned):
  1. Collect all agent outputs into 
    research/
    directory
  2. Convergence analysis: Identify tools found by multiple agents (case-insensitive fuzzy matching for tool name normalization). Tools appearing in 2+ agent results are high-signal.
  3. Cap tools passed to Phase 3 at 30 (ranked by convergence count, then relevance to scoped gaps)
  4. Write 
    research/convergence-analysis.md
    and 
    handoffs/phase2-research-handoff.yaml
Phase 2 Parallel Status Board (display to user during execution):
AgentCategoryStatusTools FoundDuration
Agent 1MCP ServersRunning/Complete/FailedNXm
Agent 2AI FrameworksRunning/Complete/FailedNXm
Agent 3Scientific ToolsRunning/Complete/FailedNXm
Agent 4CommunityRunning/Complete/SkippedNXm
Agent 5Creative IdeasRunning/Complete/SkippedNXm
QG2: Full mode: 15 or more tools across all scanned categories. Degraded mode (2-3 agents completed): 10 or more tools across 2+ categories. User informed if degraded.
通过Task工具调度3-4名研究员Agent（并行执行）。
每个Agent会收到一个领域特定的提示，覆盖研究员技能的默认方法。完整提示模板见
references/agent-prompts.md
。
Agent：
  • Agent 1：MCP服务器与Claude Code集成（优先使用WebSearch，不使用PubMed）
  • Agent 2：AI框架与代理工作流模式（优先使用WebSearch，不使用PubMed）
  • Agent 3：AI驱动的科学/计算生物学工具（PubMed + WebSearch）
  • Agent 4（可选，根据范围）：社区模式与新兴趋势
  • Agent 5（可选，非阻塞）：brainstorming-pm，用于创意集成方案（45分钟超时；如果brainstorming-pm不可用，跳过并记录警告）
研究后（编排器负责）：
  1. 将所有Agent输出收集到
    research/
    目录
  2. 收敛分析：识别被多个Agent发现的工具（工具名称归一化时使用不区分大小写的模糊匹配）。被2个及以上Agent发现的工具属于高信号工具。
  3. 限制进入阶段3的工具数量为30个（按收敛次数排序，其次是与已确定缺口的相关性）
  4. 写入
    research/convergence-analysis.md
    和
    handoffs/phase2-research-handoff.yaml
阶段2并行状态面板（执行期间向用户展示）：
Agent类别状态已发现工具数时长
Agent 1MCP服务器已完成/运行中/失败NXm
Agent 2AI框架已完成/运行中/失败NXm
Agent 3科学工具已完成/运行中/失败NXm
Agent 4社区工具已完成/运行中/跳过NXm
Agent 5创意方案已完成/运行中/跳过NXm
QG2：完整模式：所有扫描类别中至少有15个工具。降级模式（2-3个Agent完成）：2个及以上类别中至少有10个工具。如果处于降级模式，需告知用户。

Phase 3: Strategic Gap Assessment

阶段3：战略缺口评估

Delegate to strategist via Task tool.
  1. Provide: Phase 2 research handoff, scoped gaps from Phase 1, scoring matrix reference (
    references/scoring-matrix.md
    )
  2. Explicit prompt context: "Evaluate tools against workflow gaps, NOT bioreactor project goals. Score each tool using the weighted scoring matrix rubric."
  3. Strategist evaluates each tool using the weighted scoring matrix:
    • Integration feasibility (40%): 0-1 scale with defined maturity levels
    • Workflow gap coverage (35%): 0-1 per gap, aggregated as weighted average using gap weights from Phase 1
    • Cost/sustainability (25%): 0-1 scale based on pricing tier
    • Composite: 
      0.40 * integration + 0.35 * gap_coverage_avg + 0.25 * cost
  4. Strategist produces:
    • Scored tool matrix (
      assessment/scored-tool-matrix.md
      )
    • Gap coverage analysis (
      assessment/gap-coverage-analysis.md
      )
    • Per-gap champions (top tool per gap regardless of composite rank)
    • Sensitivity analysis (
      assessment/sensitivity-analysis.md
      ): Recompute with 3 alternate weight configs (integration-heavy 50/25/25, gap-focused 30/45/25, cost-conscious 30/35/35). Identify weight-robust and weight-sensitive tools.
    • Differentiation check: If score range (max - min) < 0.15, flag low differentiation
  5. Write 
    handoffs/phase3-assessment-handoff.yaml
QG3: All tools scored. At least 1 tool per gap scores above 0.5 (for gaps with non-zero weight). Sensitivity analysis complete with 4 weight configurations tested. On failure: Retry once with simplified scope, then escalate to user.
通过Task工具委托给strategist。
  1. 提供：阶段2的研究交接文件、阶段1的已确定缺口、评分矩阵参考（
    references/scoring-matrix.md
    ）
  2. 明确提示上下文："针对工作流缺口评估工具，而非生物反应器项目目标。使用加权评分矩阵规则为每个工具评分。"
  3. Strategist使用加权评分矩阵评估每个工具：
    • 集成可行性（40%）：0-1分，基于定义的成熟度等级
    • 工作流缺口覆盖度（35%）：每个缺口0-1分，使用阶段1的缺口权重计算加权平均值
    • 成本/可持续性（25%）：0-1分，基于定价层级
    • 综合得分：
      0.40 * integration + 0.35 * gap_coverage_avg + 0.25 * cost
  4. Strategist生成：
    • 评分工具矩阵（
      assessment/scored-tool-matrix.md
      ）
    • 缺口覆盖分析（
      assessment/gap-coverage-analysis.md
      ）
    • 每个缺口的首选工具（不考虑综合排名，每个缺口的顶级工具）
    • 敏感性分析（
      assessment/sensitivity-analysis.md
      ）：使用3种替代权重配置重新计算（集成优先50/25/25、缺口聚焦30/45/25、成本敏感30/35/35）。识别对权重变化稳定和敏感的工具。
    • 差异化检查：如果得分范围（最高分-最低分）<0.15，标记为低差异化
  5. 写入
    handoffs/phase3-assessment-handoff.yaml
QG3：所有工具均已评分。每个非零权重的缺口至少有一个工具得分高于0.5。已完成包含4种权重配置的敏感性分析。失败时：使用简化范围重试一次，然后升级给用户。

Phase 4: Integration Roadmap Synthesis (Orchestrator-Owned)

阶段4：集成路线图合成（编排器负责）

This phase is orchestrator-owned -- you synthesize the research and assessment outputs yourself.
  1. Read Phase 2 research handoff and Phase 3 assessment handoff
  2. Synthesize into prioritized integration roadmap:
    • Quick Wins (this week): Low effort, high impact tools ready for immediate adoption
    • Short-term (this month): Moderate effort, clear integration path
    • Medium-term (this quarter): Requires setup/learning but addresses critical gaps
    • Strategic (next quarter): Long-term bets, emerging tools worth tracking
  3. For each roadmap item: tool name, gap(s) addressed, integration approach, estimated effort, dependencies, risk level
  4. Include "what changes if we do nothing" analysis per gap
  5. Optional: Technology Radar ring assignments (Adopt/Trial/Assess/Hold) based on composite scores
  6. Write 
    roadmap/integration-roadmap.md
    and 
    handoffs/phase4-roadmap-handoff.yaml
QG4: Roadmap has items in 2 or more time horizons. Each item has an actionable next step. No placeholder entries.
此阶段由编排器负责——您需要自行整合研究和评估输出。
  1. 读取阶段2的研究交接文件和阶段3的评估交接文件
  2. 合成为优先级排序的集成路线图：
    • 快速落地（本周）：低投入、高收益，可立即采用的工具
    • 短期（本月）：中等投入，集成路径清晰
    • 中期（本季度）：需要设置/学习，但能解决关键缺口
    • 战略级（下季度）：长期布局，值得跟踪的新兴工具
  3. 每个路线图条目包含：工具名称、解决的缺口、集成方法、预计投入、依赖项、风险等级
  4. 包含每个缺口的“如果不采取行动会有什么变化”分析
  5. 可选：根据综合得分分配技术雷达象限（采用/试用/评估/搁置）
  6. 写入
    roadmap/integration-roadmap.md
    和
    handoffs/phase4-roadmap-handoff.yaml
QG4：路线图包含2个及以上时间范围的条目。每个条目都有可执行的后续步骤。无占位符条目。

Phase 5: Adversarial Review

阶段5：对抗性审查

Delegate to devils-advocate via Task tool.
  1. Provide: Integration roadmap, scored tool matrix, sensitivity analysis
  2. Challenge areas:
    • Integration feasibility: Are integration claims realistic given current architecture?
    • Gap coverage: Does tool X really solve gap Y, or just touch it?
    • Cost assumptions: Hidden costs? Vendor lock-in? Sustainability of open-source?
    • Missing alternatives: Are we overlooking emerging tools or patterns?
  3. If devils-advocate challenges the scoring methodology itself: Escalate to user with options (accept with documented limitation, re-run Phase 3 with modified methodology, include critique as appendix)
  4. Write 
    review/adversarial-review.md
    and 
    handoffs/phase5-review-handoff.yaml
QG5: All critical challenges addressed or acknowledged with documented mitigation.
通过Task工具委托给devils-advocate。
  1. 提供：集成路线图、评分工具矩阵、敏感性分析
  2. 挑战方向：
    • 集成可行性：考虑当前架构，集成主张是否现实？
    • 缺口覆盖度：工具X真的能解决缺口Y，还是只是相关？
    • 成本假设：是否有隐藏成本？是否存在供应商锁定？开源工具的可持续性如何？
    • 遗漏的替代方案：我们是否忽略了新兴工具或模式？
  3. 如果唱反调者质疑评分方法本身：升级给用户并提供选项（接受并记录限制、使用修改后的方法重新执行阶段3、将批评作为附录包含）
  4. 写入
    review/adversarial-review.md
    和
    handoffs/phase5-review-handoff.yaml
QG5：所有关键挑战已得到解决，或已记录缓解措施。

Phase 6: Editorial Polish

阶段6：编辑打磨

Delegate to editor via Task tool.
  1. Provide: Integration roadmap, adversarial review, full deliverable template (
    references/deliverable-template.md
    )
  2. Editor produces final deliverable in 
    final/ai-tool-landscape-assessment.md
  3. Deliverable includes: Executive summary, tool landscape overview, gap coverage analysis, scored matrix, integration roadmap, risk assessment, methodology notes
QG6: Consistent voice, no substantive errors, executive summary present, no placeholder text remaining.
通过Task工具委托给editor。
  1. 提供：集成路线图、对抗性审查、完整交付模板（
    references/deliverable-template.md
    ）
  2. Editor在
    final/ai-tool-landscape-assessment.md
    中生成最终交付成果
  3. 交付成果包含：执行摘要、工具格局概述、缺口覆盖分析、评分矩阵、集成路线图、风险评估、方法说明
QG6：语气一致，无实质性错误，包含执行摘要，无剩余占位符文本。

Quality Gate Specifications

质量门规范

GatePhaseChecksPass ThresholdOn Failure
QG00 - ArchivalSession dir, archival summary, state fileAll 3 createdABORT
QG11 - ScopeSpecific scope, measurable criteria, user approvalAll checksRe-run scope refinement
QG22 - ResearchTool count, category coverageFull: 15+ tools / Degraded: 10+ in 2+ categoriesInform user, proceed if degraded threshold met
QG33 - AssessmentAll tools scored, per-gap coverage, sensitivity analysisAll tools scored, 1+ tool/gap above 0.5, sensitivity completeRetry once, then escalate
QG44 - RoadmapTime horizon coverage, actionable itemsItems in 2+ horizons, each actionableRetry synthesis
QG55 - ReviewChallenges addressedCritical challenges addressed/mitigatedDeliver with unresolved challenges flagged
QG66 - EditorialFormatting, completeness, executive summaryConsistent voice, no errors, no placeholdersDeliver pre-polish version
Quality Floor (cannot be overridden):
  • Minimum 2 agent results from Phase 2
  • Every evaluated tool has a composite score
  • Roadmap contains at least 1 actionable integration item
Override Protocol: Fix issue / Override with logged gap / Abort workflow.
质量门阶段检查项通过阈值失败处理
QG00 - 归档会话目录、归档摘要、状态文件全部创建终止
QG11 - 范围具体范围、可衡量标准、用户批准所有检查项通过重新执行范围细化
QG22 - 研究工具数量、类别覆盖完整模式：15+工具 / 降级模式：2+类别中10+工具告知用户，如果达到降级阈值则继续
QG33 - 评估所有工具已评分、每个缺口有覆盖、敏感性分析完成所有工具已评分，每个缺口至少1个工具得分>0.5，敏感性分析完成重试一次，然后升级
QG44 - 路线图时间范围覆盖、可执行条目2+时间范围有条目，每个条目可执行重新合成
QG55 - 审查挑战已解决关键挑战已解决/已记录缓解措施交付时标记未解决的挑战
QG66 - 编辑格式、完整性、执行摘要语气一致，无错误，无占位符交付未打磨版本
质量底线（不可覆盖）：
  • 阶段2至少有2个Agent的结果
  • 每个评估的工具都有综合得分
  • 路线图至少包含1个可执行的集成条目
覆盖协议：修复问题 / 记录缺口后覆盖 / 终止工作流

RACI Matrix

RACI矩阵

Activityai-strategistrequirements-analystresearcherstrategistdevils-advocateeditorUser
Session setup (Phase 0)R/A-----I
Scope refinement (Phase 1)AR----C
Parallel research (Phase 2)A-R---I
Gap assessment (Phase 3)A--R--I
Roadmap synthesis (Phase 4)R/A-----C
Adversarial review (Phase 5)A---R-I
Editorial polish (Phase 6)A----RI
R = Responsible, A = Accountable, C = Consulted, I = Informed
活动ai-strategistrequirements-analystresearcherstrategistdevils-advocateeditor用户
会话设置（阶段0）R/A-----I
范围细化（阶段1）AR----C
并行研究（阶段2）A-R---I
缺口评估（阶段3）A--R--I
路线图合成（阶段4）R/A-----C
对抗性审查（阶段5）A---R-I
编辑打磨（阶段6）A----RI
R = 负责，A = 问责，C = 咨询，I = 知情

Error Handling

错误处理

Saga-style compensation ensures each phase can be rolled back:
PhaseForward ActionCompensation (on failure)
0Create session directoryRemove session directory
1Scope refinementNo side effects (read-only)
2Launch parallel researchersCancel running agents, retain completed results
3Strategic assessmentRetain Phase 2 outputs, retry once or escalate
4Roadmap synthesisRetain Phase 3 outputs, retry once or escalate
5Adversarial reviewDeliver without review (flag as unreviewed)
6Editorial polishDeliver pre-polish version
Circuit Breaker: After 2 consecutive failures of the same agent type, open circuit. Present options to user:
  • Retry with narrowed scope
  • Skip that agent type and proceed with available results
  • Abort workflow (preserve session for later resume)
Graceful Cancellation (Ctrl+C or explicit abort):
  1. Complete current atomic file operation
  2. Update workflow-state.yaml with current progress
  3. Preserve session directory
  4. Display: "Workflow paused at Phase {N}. Resume: ai-strategist --resume {session-id}"
Atomic State Writes: All workflow-state.yaml updates write to a temp file first, validate YAML syntax, then atomically rename. State is updated after each agent completion, not just at phase boundaries.
See 
references/error-handling.md
for detailed protocols including state recovery and global timeout handling.
采用Saga式补偿机制，确保每个阶段都可以回滚：
阶段正向操作补偿操作（失败时）
0创建会话目录删除会话目录
1范围细化无副作用（只读）
2启动并行研究员取消运行中的Agent，保留已完成的结果
3战略评估保留阶段2的输出，重试一次或升级
4路线图合成保留阶段3的输出，重试一次或升级
5对抗性审查交付时不包含审查内容（标记为未审查）
6编辑打磨交付未打磨版本
熔断机制：同一类型的Agent连续失败2次后，触发熔断。向用户提供选项：
  • 缩小范围后重试
  • 跳过该类型Agent，使用现有结果继续
  • 终止工作流（保留会话以便后续恢复）
优雅取消（Ctrl+C或显式终止）：
  1. 完成当前的原子文件操作
  2. 更新workflow-state.yaml，记录当前进度
  3. 保留会话目录
  4. 显示："Workflow paused at Phase {N}. Resume: ai-strategist --resume {session-id}"
原子状态写入：所有workflow-state.yaml的更新先写入临时文件，验证YAML语法，然后原子重命名。状态在每个Agent完成后更新，而非仅在阶段边界更新。
详细协议包括状态恢复和全局超时处理，请见
references/error-handling.md
。

Inter-Phase Status Report

阶段间状态报告

Between phases, communicate to user:
[Phase N/6 - {phase_name}] COMPLETE

Summary: {1-2 sentence summary of phase outcome}
Key findings: {bullet list of notable results}
Next: Phase {N+1} - {next_phase_name}
Estimated time: {duration estimate}
During Phase 2, provide a live status board showing each agent's progress:
[Phase 2/6 - Parallel Research] In Progress

| Agent   | Category        | Status   | Tools | Time |
|---------|-----------------|----------|-------|------|
| Agent 1 | MCP Servers     | Complete | 9     | 22m  |
| Agent 2 | AI Frameworks   | Running  | -     | 15m  |
| Agent 3 | Scientific      | Running  | -     | 12m  |
| Agent 4 | Community       | Pending  | -     | -    |
Update the status board as each agent completes or times out.
在阶段之间，向用户沟通：
[Phase N/6 - {phase_name}] COMPLETE

Summary: {1-2句话总结阶段成果}
Key findings: {重要结果的项目符号列表}
Next: Phase {N+1} - {next_phase_name}
Estimated time: {时长预估}
在阶段2期间，提供实时状态面板，展示每个Agent的进度：
[Phase 2/6 - Parallel Research] In Progress

| Agent   | Category        | Status   | Tools | Time |
|---------|-----------------|----------|-------|------|
| Agent 1 | MCP Servers     | Complete | 9     | 22m  |
| Agent 2 | AI Frameworks   | Running  | -     | 15m  |
| Agent 3 | Scientific      | Running  | -     | 12m  |
| Agent 4 | Community       | Pending  | -     | -    |
每个Agent完成或超时后，更新状态面板。

Handoff Integration

交接集成

Accepts handoff from: pov-expansion (gap analysis as pre-populated scope)
Handoff validation:
  1. Check handoff file exists and is non-empty
  2. Attempt YAML parse
  3. Extract 
    context.gap_analysis
    and 
    insights.workflow_gaps
  4. On any failure: Fall back to interactive scope refinement in Phase 1, passing raw handoff content as background context
See 
references/handoff-schema.md
for detailed schema definitions.
接受来自以下组件的交接：pov-expansion（将缺口分析作为预填充范围）
交接验证：
  1. 检查交接文件是否存在且非空
  2. 尝试解析YAML
  3. 提取
    context.gap_analysis
    和
    insights.workflow_gaps
  4. 任何步骤失败：回退到阶段1的交互式范围细化，将原始交接内容作为背景上下文
详细的Schema定义请见
references/handoff-schema.md
。

Session Directory Structure

会话目录结构

See 
references/session-structure.md
for full directory tree.
Session pattern: 
/tmp/ai-strategist-session-{YYYYMMDD-HHMMSS}-{PID}/
Key directories:
  • handoffs/
    -- Inter-phase YAML handoffs (one per phase transition)
  • research/
    -- Phase 2 agent outputs and convergence analysis
  • assessment/
    -- Phase 3 scored matrix, gap analysis, sensitivity analysis
  • roadmap/
    -- Phase 4 integration roadmap
  • review/
    -- Phase 5 adversarial review
  • final/
    -- Phase 6 polished deliverable
Session cleanup: Retained on success (for reference) and on failure (for resume). User may delete manually when no longer needed.
完整目录树请见
references/session-structure.md
。
会话路径模式：
/tmp/ai-strategist-session-{YYYYMMDD-HHMMSS}-{PID}/
关键目录：
  • handoffs/
    -- 阶段间YAML交接文件（每个阶段过渡一个）
  • research/
    -- 阶段2的Agent输出和收敛分析
  • assessment/
    -- 阶段3的评分矩阵、缺口分析、敏感性分析
  • roadmap/
    -- 阶段4的集成路线图
  • review/
    -- 阶段5的对抗性审查
  • final/
    -- 阶段6的打磨后交付成果
会话清理：成功和失败时都会保留会话（成功用于参考，失败用于恢复）。用户可在不再需要时手动删除。

Timeout Configuration

超时配置

See 
references/timeout-config.md
for per-phase and mode-specific timeouts.
Key timeout defaults:
  • Per-agent (Phase 2): 30 minutes
  • brainstorming-pm: 45 minutes (longer due to internal multi-stage pipeline)
  • Global workflow: 6 hours
At the 5-hour mark, display a warning to the user. On global timeout during Phase 4 or later, deliver partial results (research + assessment are the most valuable artifacts) and preserve the session directory for potential resume.
各阶段和模式的具体超时设置请见
references/timeout-config.md
。
关键超时默认值:
  • 每个Agent（阶段2）：30分钟
  • brainstorming-pm：45分钟（由于内部多阶段流水线，时长更长）
  • 全局工作流：6小时
在5小时标记时，向用户显示警告。如果在阶段4及以后发生全局超时，交付部分结果（研究+评估是最有价值的工件），并保留会话目录以便后续恢复。