Querystrategist

QueryStrategist（文献检索策略师）

QueryStrategist（文献检索策略师）是一款面向科研人员的交互式文献检索 Skill。你只需提供研究方向，它会运用 LLM 语义理解自动识别研究对象、技术方法、任务指标和范围边界，生成适用于 Web of Science、Scopus、IEEE Xplore、Google Scholar、CNKI 和万方的可复制高级检索式。经授权后，还可通过 OpenAlex 收集候选文献，并使用 Crossref 核验 DOI。最终交付范围卡、六库检索式、候选文献清单和使用说明，适用于综述、论文、学位论文、开题报告和基金申请。

从一句研究方向，到六库高级检索式、候选文献收割与核验，再到可离线浏览的检索策略工作台。

安装提示：请安装或提交完整的 QueryStrategist 目录。本项目依赖 11 个子模块、运行脚本和交付模板；只有根 SKILL.md 无法运行完整流程。

快速开始

完成安装后，发送：

开始文献检索

已经有明确研究方向时，发送：

开始文献检索，我的研究方向是：水产养殖鱼类光谱成像品质鉴定与规格分级

只需要六库检索式、不需要 API 收割时，发送：

只启动 Search A，为以下方向生成六库检索式：水产养殖鱼类光谱成像品质鉴定与规格分级

选择使用方式

使用前准备

完整流程只要求提供一句话研究方向。写作类型、目标语言、目标期刊、期刊层级、时间范围、中文文献和行业报告均使用安全默认值，不再逐项询问。用户在入口中明确给出的期刊、年份、写作类型、平台或排除要求会覆盖默认值。

如有明确边界，也可以一并提供必须包含或不纳入的内容；未提供时，系统会先生成宽范围卡供一次性确认。交付物默认保存到项目目录，无需提前选择路径。

一次完整使用会经历什么

1. 自动建立默认配置

系统根据研究方向自动建档，默认采用通用检索策略、生成六库检索式，并提供近 10 年、近 5 年和近 2 年的数据库筛选预设；Search A 检索式本身不写死年份。G0 仅执行内部完整性校验，不要求用户确认。

2. 界定研究范围

系统从研究方向生成研究对象、必需技术、支持方法、任务指标和中英文同义词，并整理为范围卡。自动推断的排除词不进入 NOT；只有用户明确提出且适合作为强排除的内容才可进入。用户只需在 G1 选择直接生成完整策略包或调整范围。

3. 生成并交付检索策略

生成 Web of Science、Scopus、IEEE Xplore、Google Scholar、CNKI 和万方检索式，并自动执行 Query QA，检查括号、引号、平台字段、查询长度、IEEE clause、排除词风险和综述限定。经用户明确授权后，可继续通过 OpenAlex 的 2-3 个梯度查询收集候选文献，合并去重后再使用 Crossref 按 DOI 核验。

流程保留两个人工业务决策门：G1 确认范围，G2 确认交付。Search B 联网前另行请求一次隐私授权；这项授权不是业务决策门，拒绝后仍会交付六库 Search A 和完整工作台。

如何使用 A0、A1 和 B

推荐按照 A0 → A1 → B 的顺序测试。不要一开始只使用 B。若 IEEE 返回零结果或结果很少，应先测试 A0；若 A0 仍然过少，应扩大对象中心词、减少固定短语或暂时移除排除项。

最终会收到什么

普通阅读只需打开 index.html。HTML 可离线使用；其他 Markdown、CSV 和子页面作为导出或审计备份保留。聊天默认只展示摘要，用户明确要求审计模式时才完整展开全部检索式与候选条目。

使用示例

研究方向：水产养殖鱼类光谱成像品质鉴定与规格分级

系统会协助明确：

• 对象：养殖鱼类、fish、farmed fish、aquaculture fish。
• 技术：高光谱成像、多光谱成像、光谱成像。
• 任务：新鲜度、脂肪、水分、蛋白、纹理和规格分级。
• 排除：虾蟹贝类、水质监测、病原检测和传统人工感官评价。

最终生成六个平台的 A0、A1 和 B 检索式，各平台的粘贴位置与筛选建议，经 DOI 核验的候选文献清单，以及可离线浏览、筛选和复制的 HTML 工作台。

常见问题

可以只生成检索式、不联网吗？

可以。使用“只启动 Search A”指令即可。生成六库检索式不需要访问 OpenAlex 或 Crossref。

检索结果为零或很少怎么办？

先测试 A0；扩大对象中心词，减少过长的固定短语，暂时移除任务词或排除项。IEEE 等工程数据库尤其应避免一开始使用过窄的标题检索。

检索结果太多怎么办？

从 A0 切换到 A1，再使用 B；增加具体任务词、合理排除词、年份筛选或平台支持的标题与邻近字段。

为什么六个平台的检索式不完全相同？

各数据库的字段代码、布尔规则、邻近算符、通配符和长度限制不同。QueryStrategist 按平台规则分别生成，不采用一条检索式机械套用所有平台。

候选文献可以直接写进论文吗？

不可以。候选清单经过 DOI 核验，但仍需用户检查题名、作者、年份、研究内容和全文，再决定是否纳入论文。

会自动下载论文全文吗？

不会。本 Skill 只生成检索策略并收集、核验候选元数据，不代替数据库订阅，也不自动下载 PDF。

需要 OpenAlex 或 Crossref API Key 吗？

不需要。只有执行 Search B 时才会请求一次联网授权；标准流程不提交邮箱或其他个人信息。

中文文件出现乱码时怎么办？

优先打开交付目录中的 index.html。HTML 是默认阅读入口；Markdown 和 CSV 已使用适合 Windows 的 UTF-8 编码。

中途退出后可以继续吗？

可以。项目状态按独立目录保存；重新启动后选择已有项目继续，避免与其他检索项目混用。

已确认的范围还能调整吗？

可以。说明需要重做的步骤或直接提出修改项，系统会返回相应确认阶段重新生成后续结果。

联网、隐私与使用边界

• 只有获得明确授权后，Search B 才访问 api.openalex.org 和 api.crossref.org。
• 标准流程不下载论文全文，不提交邮箱或其他个人信息。
• API 获取的文献仅作为候选，最终纳入决定由用户人工完成。
• 六库实际命中量以各数据库网站为准；数据库订阅和机构访问权限不由本 Skill 提供。
• 本 Skill 负责检索策略，不替代系统综述的人工筛选、质量评价和全文证据核验。

完整包检查

下载后应至少看到：根 SKILL.md、VERSION、11 个子模块目录、query_generator.py、harvest.py、HTML 生成脚本和检索策略包模板。若只有一个 SKILL.md，说明下载的是入口文件而不是完整包。

当前版本

• v1.6.5（2026-09-09）

以下内容供运行本 Skill 的 Agent 使用。普通用户无需阅读。

本文件是 QueryStrategist 的主 Skill（唯一入口），承担编排器职责：驱动 Step 0–2 状态机（Setup Wizard → Scope Definer → Search Strategist V1）。G0 为内部自动校验，人工业务决策门仅保留 G1 与 G2，最终交付检索策略包（范围卡 + 6 库检索式 + 文献候选清单 + 使用说明）。

本包内还包含 11 个子模块目录（setup_wizard/、scope_definer/、search_strategist_v1/、query_crafter/、6 个平台检索器、literature_harvester/）。每个子模块的指令文件统一为 SKILL.sub.md，由主 Skill 按“子模块执行机制”读取并执行。

子模块执行机制（MANDATORY — 禁止幽灵动作）

本主 Skill 是唯一入口，11 个子模块（<module>/SKILL.sub.md）不是独立 Skill，不会被单独注册。因此任何"调用子能力"的步骤都是执行指令，必须真实执行，执行通道按优先级：

1. (a) 平台 Skill 工具调用 — 若运行环境中存在已注册的同名独立 Skill（如用户本地安装了子 Skill），可先尝试 Skill 工具调用。
2. (b) 读取执行（单包默认通道） — 读取对应子模块的 SKILL.sub.md，按其指令真实执行：运行其 scripts/ 脚本、遵循其门控、产出其交付物。读取并执行才算执行；仅写一句"加载 X 子技能"就结束回合 = 幽灵动作，严格禁止。
3. (c) Agent 委派 — 若环境提供子代理/队友机制，可将子模块委派给 agent 并等待其返回。

自检：任何回合结束前，对自己描述的每个"已完成/进行中"动作，确认有匹配的真实执行（工具调用 / agent 返回 / 已按子模块指令运行脚本产出文件）。若写了"加载 X"却三种通道都没有执行，立即补齐后再回复。

触发方式与入口路由（MANDATORY）

主 Skill 必须先识别以下五种入口，再执行对应流程。不得把直接模式误路由到完整状态机。

1. 完整流水线：开始文献检索 / Start QueryStrategist。若未提供研究方向，只询问这一项；取得方向后执行 Step 0–2。Step 0 自动建立默认配置并完成 G0 内部校验，不询问七项配置。
2. 带方向完整流水线：开始文献检索，我的研究方向是：[topic]。把原文记录为 research_direction，标记 research_direction_source=user_provided，直接执行零配置建档与范围生成。用户明确给出的写作类型、期刊、年份、平台或边界优先于系统默认值；不得重复询问已经提供的信息。
3. 仅 Search A（六库）：只启动 Search A，为……生成六库检索式。直接读取并执行 query_crafter/SKILL.sub.md 的 search_a_all 模式；仅补问生成检索式所必需的范围信息，不执行 Setup Wizard、G0–G2 或 Search B，不访问 OpenAlex/Crossref。
4. 单平台检索式：为……生成 IEEE Xplore 检索式（平台可替换）。直接读取并执行对应平台构建器的 single_platform 模式；仅输出该平台结果，不执行完整流水线或联网收割。
5. 调整已有检索式：帮我调宽/调窄以下检索式：……。读取 query_crafter/SKILL.sub.md 的 adjust_existing 模式；识别目标平台，无法可靠识别时先询问；输出诊断、修改对照、预计影响和 Query QA，不访问 OpenAlex/Crossref。

直接模式不得声称已完成 Step 0–2 或交付完整策略包。用户随后明确要求升级为完整检索时，再从 Step 0 开始。

演示模式提示：为录屏或评审演示，建议使用「开始文献检索，我的研究方向是：……」。主题明确时系统自动完成 Step 0 与 G0，只在 G1 展示范围卡，可稳定复现短流程。

包内子模块一览

流水线终点：检索策略包

G2 确认后，自动产出四份相互衔接的文件（模板见 search_strategist_v1/assets/search_strategy_pack_template.md）：

1. scope_card.md/.html — 范围界定卡（三级关键词 + 排除词分级 + 写作类型 + 策略权重）；
2. query_pack.md/.html — 多平台检索式合集（6 库，每库 A0 召回基线 + A1 主题式 + B 精准式；附 Query QA 状态；检索式代码块原样保留）；
3. candidate_list.csv/.md/.html — 文献候选清单（去重元数据 + OA 状态 + DOI 链接，标注"候选清单、非最终语料"）；
4. usage_guide.md/.html — 使用说明（检索式填入位置 + 命中量级预估 + 调宽/调窄方法 + 按写作类型建议）。

Markdown 和 CSV 统一写为 UTF-8 BOM；HTML 为默认阅读入口、可离线打开。HTML 必须由 _shared_tools/scripts/render_deliverables.py 从同名 Markdown 生成，禁止维护第二套内容。 默认交付目录为 projects/<active_project_id>/deliverables/；用户已明确提供自定义路径时直接采用。最终交付目录必须生成 index.html，该文件是唯一默认阅读入口；各页面在 JavaScript 不可用时仍须完整展示原始内容。

所有字段继承 Step 0–2 上游选择（【继承自 …】 标注），禁止凭空生成。

核心红线

• 收割 ≠ 语料：API 收割的元数据仅作候选清单，绝不自动进入下游当作全文语料，需用户自行下载验证。
• 人机闸门：G0 为内部自动校验；G1 与 G2 是人工确认点，范围与最终交付确认始终由人类决定。
• 语言锁：回复语言严格跟随用户当前消息语言；检索式按各数据库适用的原生语言与官方语法生成，文献元数据保持来源原文。

Response Language Rule (Mandatory — Hard Language Lock)

Language Lock: The reply language is strictly bound to the language the user is writing in their current message. No other criterion applies. This is a hard constraint that overrides any default model tendency or system hint. ALL natural languages are supported — not only Chinese and English.

• Default from Step 0: Setup Wizard (Step 0) detects the user's first-message language and records it as interaction_language (ISO 639-1 code) in the Project Configuration Profile. This value is the default interaction language passed to every sub-module, ensuring sub-modules that do not directly see the user's latest message still reply in the correct language.
• Detect per message, not just at entry: For EVERY user message, detect its language and reply in exactly that language. Do NOT base the reply language on the entry command alone — if the user switches languages mid-conversation, the reply language switches with them. The Step 0 interaction_language is the default; the per-message detection is the active enforcer.
• Any language → same-language reply: When the user writes in any language (Chinese, English, Japanese, French, German, Korean, Spanish, Russian, Arabic, Portuguese, etc.), ALL output — greetings, configuration prompts, scope questions, status displays, decision-gate confirmations, completion messages — MUST be in that same language. Brand names / technical terms may stay in their original form; the explanatory prose MUST match the user's language.
• Mixed input: If a single message mixes languages, follow the dominant (majority) language of that message.
• NEVER switch unprompted: Under no circumstances reply in a language the user did not use in their latest message (e.g., never answer a Chinese message in Japanese or English, and never answer an English message in Chinese). If unsure of the language, default to matching the most recent user message, or fall back to the Step 0 interaction_language.
• Sub-module binding: This lock binds every delegated sub-module. When the main Skill invokes a sub-module, it MUST explicitly instruct that sub-module to reply in the locked language (passing the interaction_language value from the config profile). Sub-modules must not introduce a different language.
• Low-resource languages: For the ~30–40 major world languages, language following is highly reliable. For very low-resource languages, if the model occasionally slips, default to the Step 0 interaction_language and inform the user they may switch to a major language if they prefer.

This rule takes priority over any conflicting language preference in downstream modules or system defaults.

Hard Rule — No Phantom Actions (禁止幽灵动作)

This is a MANDATORY, non-negotiable rule, on par with the Language Lock above.

The problem it prevents: A workflow step says "invoke / call / load / execute / delegate to the X sub-module", but the assistant only writes that sentence (often ending with a colon, e.g. "先加载 X 子模块以获取…") and then ends the turn — without ever actually executing the sub-module's logic (no tool call, no script run, no inline execution). The instruction exists, the action does not. This is a "phantom action" and is a critical failure.

The binding rule:

1. Any instruction in this SKILL that reads "invoke / call / load / execute / delegate to a sub-module" is an EXECUTION DIRECTIVE, not narration. When you reach such a step, you MUST actually run the sub-module's logic through whichever channel your environment provides, in this order of preference:
   • (a) Skill tool call — if the platform exposes a Skill tool and the sub-module is installed as an independent Skill (e.g. Skill: "literature_harvester"), emit the real call.
   • (b) Read-and-execute — READ the sub-module's SKILL.sub.md (and any scripts/ helpers it references) and EXECUTE its instructions inline: run its scripts, follow its gates, produce its artifacts. Importing the module's content as context and acting on it counts as execution; merely narrating that you will do it does not.
   • (c) Agent delegation — if a sub-agent / teammate mechanism is available, delegate the sub-module to an agent and wait for its result. Do not describe it; do it.
2. You are strictly forbidden from writing a transitional sentence about loading/calling a sub-module and then ending the turn without the actual tool call. Examples of forbidden behavior:
   • ❌ "先加载 Literature Harvester 子模块以获取收割脚本与参数规范：" → [turn ends, no call]
   • ❌ "Now invoking Query Crafter…" → [turn ends, no call]
3. Self-check before ending ANY turn: For every action you described as done or in-progress, verify a matching execution actually happened (a Skill call, an agent delegation that returned, or an inline run of the sub-module's instructions/scripts). If you wrote "loading X" but executed none of the three channels above, you have a phantom action — fix it NOW by executing it before replying.
4. If you genuinely must wait for user input (e.g., a decision gate, an AskUserQuestion), do NOT narrate a sub-module load you have not made. Instead state what the NEXT step will be after the user responds (e.g., "下一步将读取 Literature Harvester 的 SKILL.sub.md 执行 Search B"). The narration must reflect reality.
5. This rule binds every delegated sub-module call across Steps 0–2, including search_strategist_v1 → (Query Crafter, Literature Harvester) and query_crafter → (platform Query Crafters) delegations.

Hard Rule — Project Isolation (项目隔离)

This is a MANDATORY, non-negotiable rule, on par with the Language Lock and No Phantom Actions above.

The problem it prevents: A pipeline run begins by selecting/creating a project in Step 0 (Setup Wizard's Project Workspace Gate). Without an isolation rule, the agent may accidentally read, modify, or borrow files from a different project's folder (projects/<other_id>/), corrupting one project with another's data, or leaking a past project into a "fresh start".

The binding rule:

1. Step 0 sets the active project. After the user chooses "继续 <项目>" or "新建项目", the active project is fixed for the rest of the session and recorded in active_project_dir (Pipeline Context) + workspace-root active_project.json.
2. All file reads/writes are confined to active_project_dir. Every downstream module (Steps 1–2) MUST resolve paths relative to projects/<active_id>/ — its pipeline_state/, Step2*/, config.json, and .workbuddy/memory/. The agent MUST NOT read, write, move, copy, or otherwise reference any projects/<other_id>/.
3. "新建项目" severs access to prior projects. Once the user picks a fresh project, the previous project's workspace is effectively disconnected for the remainder of the session. The agent MUST treat projects/<other_id>/ as non-existent — it must not peek at old configs, PDFs, reports, or memory to "help" the new project.
4. Memory is isolated too. The workspace-root .workbuddy/memory/ is the live mirror of the ACTIVE project only (restored at Step 0). The agent MUST NOT read another project's .workbuddy/memory/; it must not write cross-project notes. Switching projects at Step 0 archives/restores memory per the Setup Wizard's "记忆隔离" procedure.
5. Self-check before any file operation: Before reading or writing any path, verify it is under projects/<active_id>/. If a tool call targets projects/<other_id>/ (or the workspace root outside active_project_dir), STOP — it is a violation of this rule.
6. This rule binds every sub-module call across Steps 0–2.

Hard Rule — No Long-Term Memory Borrowing (禁止借用长期记忆/跨项目背景)

This is a MANDATORY, non-negotiable rule, on par with the Language Lock, No Phantom Actions, and Project Isolation above.

The problem it prevents: At pipeline activation, the assistant uses the user's long-term / cross-project memory (e.g., a past research direction, target journal, or review topic from previous projects) to pre-fill configuration options or scope questions — even though the current workspace may have no project and the user expects a fresh start. This contaminates a new project with stale context and violates the "fresh start" expectation.

The binding rule:

1. At activation, consult ONLY the workspace. When the pipeline starts (Step 0 / Step 0.5), the ONLY sources of context are: (a) the current workspace's projects/ directory (existing QueryStrategist projects) and (b) the user's explicit input in the current session (entry command, answers to configuration questions, direction sentence).
2. NEVER pre-fill from long-term memory. The assistant MUST NOT use the user's long-term research background, past project topics, target journals, or any cross-project memory to pre-fill configuration options, to hint at scope directions, or to "help" the user describe their research direction. No "基于你的长期研究背景，我已预填…" type of behavior.
3. Fresh workspace = fresh start. If projects/ is empty (or contains no real project), treat the session as a brand-new project. Ask only for a missing research direction, then create the documented system defaults; do not borrow values from older projects or long-term memory.
4. Scope provenance must be explicit. The direction sentence comes from the current session. Scope fields may be safely inferred from that direction and must be marked system_inferred_safe; explicit user input is marked user_explicit and always takes precedence.
5. This rule binds every sub-module call across Steps 0–2.

Hard Rule — Bash Sandbox Consistency & Safe File Persistence

This is a MANDATORY operational guard, on par with the Language Lock, No Phantom Actions, and Project Isolation above. It is the global fix for the "run hangs / deliverable file ends up empty" failure mode observed in real runs.

The two failure modes it prevents:

1. Sandbox / real-FS split-brain. Bash runs sandboxed by default. A file written by a normal (sandboxed) Bash call (cat > file, python -c "open('file','w')") lands in a throwaway sandbox FS layer that is invisible to: (a) a later dangerouslyDisableSandbox: true Bash call, and (b) the Read / Write / Edit tools (which always use the real FS). The result is a "successful" write no other command or tool can see → silently produces a missing or 0-byte file. The sibling file that was written via a non-sandboxed call survives, while the sandboxed one vanishes — a classic split-brain.
2. Bash command-length cap (~1.5–2 KB). Embedding multi-KB content (a long markdown report, large JSON, or a big inline Python block) inside a Bash heredoc (cat <<'EOF' / python <<'PYEOF') causes the command to be truncated mid-stream — the closing delimiter is never sent → shell error unexpected EOF while looking for matching ''. The agent tends to retry the same failing pattern, which looks like "running a long time without completing" but is actually doing no useful work.

The binding rule:

1. Be consistent about the sandbox. For ANY Bash command that reads/writes files under the workspace (projects/<id>/, or the workspace root), use dangerouslyDisableSandbox: true every time. Never mix sandboxed and non-sandboxed Bash in the same task. Preferred alternative: use the Read / Write / Edit tools for file I/O (they always hit the real FS and have no length cap), and reserve Bash for pure compute / external commands.
2. Never inline multi-KB content in Bash. If you need to generate a large file (report, JSON, code), write a generator script via the Write tool (no length cap), then run it with a short Bash command (python gen.py). Keep inline Bash commands < 1.5 KB; if unavoidable, split with cat >> chunks or a temp file.
3. Write-then-verify. After any step that writes a deliverable, immediately verify the file exists with byte size > 0 (e.g. wc -c file). A 0-byte file is the signature of the two bugs above — never assume a write succeeded; never proceed silently past an empty output.
4. This rule binds every sub-module call across Steps 0–2.

Entry Command

The user initiates the pipeline with:

""Start QueryStrategist."" ""开始文献检索"" ""启动检索策略"" ""建检索式""

Step 0 启动行为（重要）：入口指令触发后，主 Skill 第一步读取并执行 setup_wizard/SKILL.sub.md。若研究方向缺失，只询问研究方向；否则自动建立项目与默认配置并执行 G0 内部校验。不得逐项询问写作类型、目标语言、目标期刊、期刊层级、时间范围、中文文献或行业报告。

Alternatively, the user can specify a research direction directly:

""Start QueryStrategist. My research direction is: [topic]"" ""开始文献检索，我的研究方向是：[topic]""

Workflow

State Machine Definition

The main Skill maintains an internal state that tracks:

• Current Step: Which module is active (0–2).
• Step Status: pending, in_progress, awaiting_confirmation, completed, skipped, failed.
• User Context: All confirmed outputs from previous steps (configuration profile, scope document, folder paths, confirmed topic).

Sequential Execution Logic

For each step, the main Skill:

1. Check preconditions: Ensure all required inputs from previous steps are available.
2. Invoke the sub-module: Read the corresponding SKILL.sub.md, pass the relevant context, and execute its defined workflow (or call the registered independent Skill if installed — see 子模块执行机制).
3. Present output: Display the sub-module's output to the user in a clear, summarized format.
4. Apply the gate policy: G0 is an automatic validation and does not pause. Pause only at G1 and G2; request separate network consent immediately before Search B.
5. Record state: Mark the step as completed and log the outputs for downstream use.

Decision Gates and Network Consent

G0 is an internal validation and MUST NOT interrupt the user. The main Skill pauses at G1 and G2 only. Immediately before Search B, it requests a separate one-time network authorization. All prompts use the user's interaction language:

Before Search B, use the consent wording defined in search_strategist_v1/SKILL.sub.md: explain that only OpenAlex and Crossref HTTPS APIs are accessed, no full text is downloaded, and no personal information is submitted. A refusal sets Search B to skipped_by_user and does not block Search A or delivery.

Pipeline Step Map (0–2)

终点（Step 2 完成后）：Search Strategist V1 交付检索策略包作为最终交付物：范围卡、6 库检索式合集、全量文献候选清单和使用说明。Markdown/CSV 为 UTF-8 BOM，同时生成对应离线 HTML；用户在 G2 确认后流水线结束。

Error Handling

If a sub-module encounters an error or cannot complete:

1. Log the error: Record which step failed and any available error details.
2. Notify the user: Present the error clearly and offer three options:
   • Retry: Re-run the current step.
   • Skip: Skip this step and continue (only if the step is optional or can be completed later).
   • Abort: Terminate the pipeline.
3. Respect user decision: Do not proceed until the user chooses an option.

Pipeline Context Schema (Central Definition)

Purpose: All fields shared between steps are defined here. Each downstream module's Input Requirements section MUST reference this schema by field name.

CRITICAL — Data Flow Principle (harvested ≠ corpus):

• API 收割的文献只是"候选参考清单"，不是语料。 AI 通过 API / 检索式收割到的元数据（标题、作者、DOI、期刊、年份）经常包含虚构、错位或拼错的内容，不能直接当作可信事实。
• 下载决策权完全在用户。 Search Strategist V1 只负责把候选清单（含检索式、下载链接、OA 状态）呈现给用户；由用户自行核对、筛选、决定下载哪些 PDF。AI 不替用户自动下载，也不替用户决定纳入。
• 检索策略包是终点交付物，不再串联下游选题。 用户据此自行到各平台验证检索式、下载文献。综述选题与大纲等下游写作环节由用户自行决定工具与流程。
• 违反此原则的代价： 把 AI 虚构的文献当事实写进论文 → 事实错误乃至学术不端。因此"人是最终把关者（human-in-the-loop）"是硬性红线，不可被任何"自动化便利"绕过。

Deliverable: The pipeline output is a search strategy pack — scope card + multi-platform query pack (6 databases, each with recall-optimized A and precision-optimized B) + candidate literature list (deduped metadata + OA status + DOI links) + usage guide. Every produced artifact carries source-citation anchors for full traceability.

Completion

After the final step (Search Strategist V1, Step 2) — once the user confirms the search strategy pack at G2 — the main Skill outputs a completion summary in the user's interaction language (detected in Step 0, locked per the Hard Language Lock). The reference versions below are in English and Chinese; translate into any other language the user is writing in:

English: "QueryStrategist pipeline complete (Steps 0–2). Search strategy pack delivered: scope card + 6-database queries + candidate list + usage guide. Thank you for using QueryStrategist!" Chinese: "QueryStrategist 流水线已完成（Step 0–2）。检索策略包已交付：范围卡 + 6 库检索式 + 文献候选清单 + 使用说明。感谢使用 QueryStrategist！"

Output Format

Pipeline Status Display

After each step, the main Skill displays a concise status update (use Chinese labels when entry is Chinese, English labels otherwise):

English version:

===== QueryStrategist Pipeline Status (Steps 0–2) ===== Steps: ██████░░░░ (2/3 completed) Current: Step 2 – Search Strategist V1 Status: in_progress

Completed: ✅ Step 0 – Setup Wizard ✅ Step 1 – Scope Definer ✅ Step 2 – Search Strategist V1 (awaiting G2 strategy pack confirmation) Next: deliver search strategy pack after confirmation

Chinese version:

===== QueryStrategist 流水线状态（Step 0–2）===== 进度: ██████░░░░ (2/3 已完成) 当前: 第 2 步 – 检索策略 V1 状态: 执行中

已完成: ✅ 第 0 步 – 配置 ✅ 第 1 步 – 范围界定 ✅ 第 2 步 – 检索策略 V1（等待 G2 检索策略包确认） 下一步: 确认后交付检索策略包

Important Notes

• If the user specifies a research direction in the entry command (e.g., ""Start QueryStrategist. My direction is: deep learning for medical image segmentation""), the main Skill feeds this direction directly into Step 0 Setup Wizard and Step 1 Scope Definer as pre-populated context, rather than asking the user to restate it.
• The main Skill does not execute any sub-module's internal logic itself out of context; it reads and executes the sub-module's SKILL.sub.md instructions (or delegates via Skill tool if installed) and reports the results. If a sub-module's file is missing, the main Skill informs the user and asks whether to skip or pause until it is available.
• The user can abort at any time by saying ""stop pipeline"" or ""abort"".
• The user can skip backward to redo any completed step by saying ""re-run Step X"".
• Scope: QueryStrategist delivers the Step 0–2 pipeline (Setup Wizard → Scope Definer → Search Strategist V1); the final deliverable is a search strategy pack (scope card + queries + candidate list + usage guide). The pipeline ends here — downstream review-topic/outline generation is out of scope.

正式单包结构

querystrategist/                    # 正式发布目录
├── SKILL.md                        # 主 Skill（本文件，唯一入口，承担编排器职责）
├── VERSION                         # 发布版本
├── LICENSE                         # MIT
├── setup_wizard/                   # Step 0 子模块（指令: setup_wizard/SKILL.sub.md）
├── scope_definer/                  # Step 1 子模块
├── search_strategist_v1/           # Step 2 子模块（含 assets/search_strategy_pack_template.md）
├── query_crafter/                  # 6 平台检索式总控子模块
├── wos_query_crafter/  scopus_query_crafter/  ieee_query_crafter/
├── google_scholar_query_crafter/  cnki_query_crafter/  wanfang_query_crafter/
├── literature_harvester/           # API 收割子模块（OpenAlex 收割 + Crossref 验证）
└── _shared_tools/                  # HTML 交付生成与项目状态校验脚本

每个子模块目录均含 SKILL.sub.md；仅保留实际被运行流程引用的 scripts/ 与 assets/ 内容。