harnessed 1.0.3 → 2.0.0

package/workflows/judgments/fallback.yaml

@@ -0,0 +1,35 @@
+# workflows/judgments/fallback.yaml
+# Fallback 三条铁律 — 机器化 ~/.claude/CLAUDE.md 「Fallback 三条铁律」 verbatim
+# Sister: 不同于 triggers (fires_when/skips_when 模式), 本 file 使用 'rules' 顶级 key
+# Phase v2.0-2.3 Wave 0 T2.3.W0.2 ship per R20.16 + D-16
+
+schema_version: harnessed.judgment.v1
+
+rules:
+  uncertain-skip-transparently:
+    description: |
+      铁律 1 — 拿不准 → 跳过 + 透明声明。
+      Workflow runtime 在 gate eval throw GateEvalError 或 evidence 不足时,
+      emit ⚠️ message 给用户, 不静默跳过。
+    fallback_action: skip_with_transparency
+    message_template: "⚠️ 跳过 {gate_name}, 因为 {reason}。如认为需要请明示。"
+
+  user-explicit-override:
+    description: |
+      铁律 2 — 用户明示 → 覆盖判据。
+      Workflow runtime 词法匹配 user input, 命中以下任一 signal 强制 invoke,
+      即使 fires_when==false / skips_when==true。
+    override_signal:
+      - "先 brainstorm"
+      - "跑 office-hours"
+      - "讨论一下"
+      - "office-hours"
+      - "brainstorm"
+      - "深度调研"
+
+  chain-isolation:
+    description: |
+      铁律 3 — 链式互不前置, 跳过战略层 ≠ 必须跳过 phase 层;
+      每层独立判断 (防"上层没跑下层不敢跑"的死板)。
+      Strategic / phase / subtask 三层 gate eval 互相独立, 无父子继承。
+    chain_isolation: true

package/workflows/judgments/parallelism-gate.yaml

@@ -0,0 +1,47 @@
+# workflows/judgments/parallelism-gate.yaml
+# Parallelism 判据 — 机器化 ~/.claude/CLAUDE.md 「子任务并行执行机制 (subagent vs Agent Teams 路由)」节
+# Sister: ~/.claude/rules/agent-teams.md (5 升级触发完整生命周期)
+# Phase v2.0-2.3 Wave 0 T2.3.W0.2 ship per D-11 (parallelism-gate) + R20.10 (ralph-loop 正交 wrapper)
+
+schema_version: harnessed.judgment.v1
+
+triggers:
+  subagent-default:
+    description: |
+      Default — subagent fan-out per CLAUDE.md 子任务并行机制
+      (Task / Agent 工具 spawn, focused 任务 research / verify / review,
+      多任务并行 fan-out, context 隔离, token 敏感)。
+    fires_when: "subtask.parallel_count <= 3 and subtask.communication_needed == false"
+    invokes:
+      - capability: superpowers-subagent-driven-development
+
+  main-session-fallback:
+    description: |
+      Downgrade — main session direct execution 主 session 直接做
+      (小任务 < 20 行 / 单命令查询, 不开 subagent 节省 token)。
+    fires_when: "subtask.lines < 20 or subtask.type == 'single_command_query'"
+
+  agent-teams-upgrade:
+    description: |
+      Upgrade — Claude Code Agent Teams 升级 per ~/.claude/rules/agent-teams.md
+      5 任一触发即可 (teammate SendMessage 互通 / subagent context overflow /
+      共享 task list 自协调 / 对立假设辩论 / 全栈三路协同)。
+      开 team 前 token 成本估算 team_cost < 2 × subagent_cost 才划算。
+    fires_when: "teammate_send_message_needed == true or subagent_context_overflow == true or shared_task_list == true or opposing_hypothesis_debate == true or fullstack_three_way == true"
+    invokes:
+      - capability: agent-teams-create
+    requires:
+      capabilities:
+        - agent-teams-create
+
+  ralph-loop-wrapper:
+    description: |
+      Orthogonal — ralph-loop 正交 wrapper, 套在 subagent / team / main session
+      外层保 completion-promise verbatim "COMPLETE"。任何执行单元均可套用,
+      并非互斥触发器 (per R20.10 + D-10)。
+    wraps:
+      - subagent-default
+      - agent-teams-upgrade
+      - main-session-fallback
+    invokes:
+      - capability: ralph-loop

package/workflows/judgments/phase-gate.yaml

@@ -0,0 +1,17 @@
+# workflows/judgments/phase-gate.yaml
+# Phase 层 (Phase tier) 判据 — 机器化 ~/.claude/CLAUDE.md 「澄清/审查触发判据」节
+# Sister: GSD /gsd-discuss-phase orchestration gate (also /gsd-new-project, /gsd-next)
+# Phase v2.0-2.3 Wave 0 T2.3.W0.2 ship per D-04 + D-16 multi-file 分类
+
+schema_version: harnessed.judgment.v1
+
+triggers:
+  gsd-discuss-phase:
+    description: |
+      GSD /gsd-discuss-phase — 新 phase 且 ≥2 个 open implementation decisions,
+      或跨 phase 数据流 / API contract / 依赖不清, 或 phase scope > 1 天 / > 5 文件,
+      或存在灰色地带 (多种合理实现都讲得通) 时激活。
+    fires_when: "phase.open_decisions >= 2 or phase.has_cross_phase_data_flow == true or phase.scope_days > 1"
+    skips_when: "phase.single_task == true or phase.scope_days < 1"
+    invokes:
+      - capability: gsd-discuss-phase

package/workflows/judgments/strategic-gate.yaml

@@ -0,0 +1,25 @@
+# workflows/judgments/strategic-gate.yaml
+# 战略层 (Strategy tier) 判据 — 机器化 ~/.claude/CLAUDE.md 「澄清/审查触发判据」节
+# Sister: gstack /office-hours + /plan-ceo-review governance gate
+# Phase v2.0-2.3 Wave 0 T2.3.W0.2 ship per D-04 + D-16 multi-file 分类
+
+schema_version: harnessed.judgment.v1
+
+triggers:
+  office-hours:
+    description: |
+      gstack /office-hours — 新项目 / 新 milestone / 新产品方向时激活,
+      用户原话出现"启动 / 想做个 X / 新功能"触发商业 scope 澄清。
+    fires_when: "phase.type in ['new_project', 'new_milestone', 'new_feature']"
+    skips_when: "phase.type in ['bug_fix', 'tech_debt', 'continuing_phase'] or phase.scope_locked_in_history == true"
+    invokes:
+      - capability: gstack-office-hours
+
+  plan-ceo-review:
+    description: |
+      gstack /plan-ceo-review — 新项目启动 / 主要发布前强制治理关卡,
+      CEO 视角商业 positioning 与 scope 评审。
+    fires_when: "phase.type == 'new_project' or phase.is_major_release == true"
+    skips_when: "phase.scope_locked_in_history == true or phase.type in ['bug_fix', 'tech_debt']"
+    invokes:
+      - capability: gstack-plan-ceo-review

package/workflows/judgments/subtask-gate.yaml

@@ -0,0 +1,17 @@
+# workflows/judgments/subtask-gate.yaml
+# 子任务层 (Subtask tier) 判据 — 机器化 ~/.claude/CLAUDE.md 「澄清/审查触发判据」节
+# Sister: superpowers:brainstorming skill (与 superpowers TDD red-green-refactor 并行评估)
+# Phase v2.0-2.3 Wave 0 T2.3.W0.2 ship per D-04 + D-16 multi-file 分类
+
+schema_version: harnessed.judgment.v1
+
+triggers:
+  brainstorming:
+    description: |
+      superpowers:brainstorming — 子任务有 ≥2 个明显不同实现方案,
+      或涉及核心算法 / 数据结构 / API contract 设计,
+      或无现成 pattern 可抄, 或错误成本高 (数据迁移 / 并发 / 性能关键路径) 时激活。
+    fires_when: "subtask.approaches >= 2 or subtask.core_algorithm == true or subtask.has_api_contract == true or subtask.error_cost == 'high'"
+    skips_when: "subtask.type in ['crud', 'standard_lib_call'] or subtask.lines < 20"
+    invokes:
+      - capability: superpowers-brainstorming

package/workflows/judgments/tdd-gate.yaml

@@ -0,0 +1,17 @@
+# workflows/judgments/tdd-gate.yaml
+# TDD 判据 — 机器化 ~/.claude/CLAUDE.md 「Execute 阶段 TDD 强烈建议开启」节
+# Sister: superpowers:test-driven-development skill (alias mattpocock /tdd per D-13)
+# Phase v2.0-2.3 Wave 0 T2.3.W0.2 ship per D-13 (6 fires + 3 skips)
+
+schema_version: harnessed.judgment.v1
+
+triggers:
+  tdd-strongly-suggested:
+    description: |
+      superpowers:test-driven-development (red-green-refactor) — 子任务为
+      核心业务逻辑 / 算法 / 数据处理, 或回归 bug 风险高, 或需要高可靠性时
+      强烈建议开启 TDD。Alias /tdd (mattpocock-skills) 可替代。
+    fires_when: "subtask.is_core_business_logic == true or subtask.is_algorithm == true or subtask.is_data_processing == true or subtask.regression_risk == 'high' or subtask.reliability_required == true"
+    skips_when: "subtask.type in ['crud', 'ui_polish', 'docs_only']"
+    invokes:
+      - capability: tdd

package/workflows/plan-feature/SKILL.md

@@ -1,8 +1,11 @@
 ---
 name: plan-feature
 description: |
-  plan-feature workflow — 5-phase three-layer-stack composition (gstack governance →
-  superpowers brainstorm → GSD discuss → GSD plan → planning-with-files persist).
+  plan-feature workflow v2 — 5-phase 三层栈 composition (gstack governance gate →
+  superpowers brainstorm → GSD /gsd-discuss-phase → GSD /gsd-plan-phase →
+  planning-with-files Claude Code plugin slash cmd `/plan` 持久化 task_plan.md +
+  progress.md + findings.md). schema_version: harnessed.workflow.v2 with
+  capability template interpolation + 4-level gate refs + conditional `on` clauses.
   Triggered by harnessed CLI `harnessed plan-feature --task <text>` or slash command
   `/plan-feature` after `harnessed setup`.
 trigger_phrases:
@@ -12,26 +15,61 @@ trigger_phrases:
   - "跑 plan-feature"
 ---
 
-# plan-feature workflow
+# plan-feature workflow (v2)
 
 ## Overview
 
 5-phase chain mapping the user's CLAUDE.md Discuss + Plan discipline onto the harnessed
-runtime (ADR 0011 — SDK + ralph-loop integration):
+runtime, upgraded to `harnessed.workflow.v2` schema (Phase v2.0-2.4 W1 T2.4.W1.3 —
+D-15 + Q-AUDIT-5a Option A: planning-with-files = Claude Code plugin slash cmd
+`/plan` 真接, **NOT** npm SDK call, **NOT** fs.writeFile self-impl).
 
-| phase | id | upstream | model | rationale |
-| ----- | -- | -------- | ----- | --------- |
-| 1 | `01-gstack-decision` | `gstack` `/office-hours` governance gate | opus | 多角色决策关卡 (CEO/EM/Designer/Paranoid/QA/CSO) |
-| 2 | `02-brainstorm` | `superpowers brainstorming` + ui-ux-pro-max | sonnet | 任务复杂度澄清 + UI 设计方案 |
-| 3 | `03-gsd-discuss` | `gsd` `/gsd-discuss-phase` | sonnet | 灰色地带澄清 4-question batch |
-| 4 | `04-gsd-plan` | `gsd` `/gsd-plan-phase` | sonnet | Wave A research + Wave B planner + Wave C checker |
-| 5 | `05-persist` | `planning-with-files` | haiku | task_plan.md + progress.md + findings.md 持久化 |
+| phase | id | upstream | model | capability / invokes | gate / on |
+| ----- | -- | -------- | ----- | -------------------- | --------- |
+| 1 | `01-gstack-decision` | gstack | opus | `{{ capabilities.gstack-office-hours.cmd }}` / `{{ gstack_prefix }}office-hours` | `gate: judgments.strategic-gate.office-hours.fires` |
+| 2 | `02-brainstorm` | superpowers | sonnet | `{{ capabilities.superpowers-brainstorming.cmd }}` | `on: judgments.subtask-gate.brainstorming.fires → invoke` |
+| 3 | `03-gsd-discuss` | gsd | sonnet | `{{ capabilities.gsd-discuss-phase.cmd }}` | `on: judgments.phase-gate.gsd-discuss-phase.fires → invoke` |
+| 4 | `04-gsd-plan` | gsd | sonnet | `invokes: gsd-plan-phase` (literal — v2.x patch followup capabilities entry) | — |
+| 5 | `05-persist` | planning-with-files | haiku | `{{ capabilities.planning-with-files.cmd }}` / `invokes: /plan` | `on: phase.scope_days > 1 or phase.is_critical_module → invoke` |
 
 Per-phase config loads from `workflows/plan-feature/workflow.yaml`; engine.runRouting
 spawns each phase as a sub-agent via `@anthropic-ai/claude-agent-sdk` 0.3.142+.
 
-Phase 1 (gstack governance) uses `on_veto: halt_workflow` — any CEO/EM veto halts the
-entire workflow before proceeding to brainstorm or planning.
+### v1 → v2 字段 delta
+
+- ADD `schema_version: harnessed.workflow.v2` (T2.4.W0.1 16th surface — first .v2)
+- ADD `description` root field (workflow purpose summary)
+- ADD `phase.capability: '{{ capabilities.<name>.cmd }}'` template interpolation (D-10)
+- ADD `phase.gate: judgments.<file>.<trigger>.fires` 4-level ref (D-04, pre-resolved by `judgmentResolver`)
+- ADD `phase.on: [{if, invoke|action}]` conditional clause (D-09)
+- ADD `phase.artifacts_expected: [...]` 3-file persistence list (D-15)
+- CHANGE `max_iterations` 改用 `{{ defaults.ralph_max_iterations.plan-feature.<phase> }}` jinja ref (T2.3.W1.2 ship)
+- KEEP `invokes: '/plan'` literal at 05-persist (Q-AUDIT-5a anti-pattern guard)
+
+### Phase 1 governance gate
+
+Phase 1 (gstack governance) uses workflow-level `on_veto: halt_workflow` — any
+CEO/EM/Designer/Paranoid/QA/CSO veto halts the entire workflow before proceeding
+to brainstorm or planning (D-04 PUSH 任 1 phase 转换前 read = vetoed → 全 halt).
+
+### Phase 5 planning-with-files plugin 真接 (Q-AUDIT-5a LOCKED Option A)
+
+05-persist invokes the **Claude Code plugin** slash command `/plan` via plugin
+cache at `~/.claude/plugins/cache/planning-with-files/planning-with-files/2.34.0/`
+(verified 2026-05-20). The plugin generates 3 markdown files in
+`.planning/<phase-id>/`:
+
+- `task_plan.md` — 主计划 (task 列表 + 文件路径 + 依赖顺序 + 验收标准)
+- `progress.md` — 跨 session 进度跟踪
+- `findings.md` — 调研发现 / 知识沉淀
+
+Q-AUDIT-5a rejected alternatives (anti-pattern guard verified by
+`tests/workflow/plan-feature-v2.test.ts`):
+
+- (b) plugin script direct spawn — Windows PowerShell 跨平台不友好
+- (c) `fs.writeFile` self-implementation — duplicates plugin functionality, breaks
+  multi-IDE compatibility (sister .continue / .factory / .gemini / .codebuddy /
+  .cursor structure 已就绪)
 
 ## CLI invocation
 
@@ -52,5 +90,9 @@ command `/plan-feature` automatically (Gap B fix — v1.0.2).
 ## References
 
 - ADR 0011 — SDK + ralph-loop integration
-- `workflows/plan-feature/workflow.yaml` — 5-phase config
+- `workflows/plan-feature/workflow.yaml` — 5-phase config (v2 schema)
+- `workflows/capabilities.yaml` — capability manifest (35 entry baseline)
+- `workflows/judgments/{strategic-gate,phase-gate,subtask-gate}.yaml` — gate refs
+- `workflows/defaults.yaml` — ralph_max_iterations.plan-feature.* values
 - `docs/WORKFLOW.md` — 4-stage workflow mermaid + gap analysis
+- `.planning/phase-v2.0-2.2/RESEARCH.md` § 5.3 — D-15 实装路径 reframe

package/workflows/plan-feature/workflow.yaml

@@ -1,40 +1,81 @@
-# workflows/plan-feature/workflow.yaml — Phase 3.2 W2 T2.4 (D-03 WIRED LOCKED)
-# 5-phase reference implementation; Phase 3.3+ dogfood 时换真 gsd-discuss/plan/execute spawn.
-# R7.1 acceptance: 30 plan-feature 场景跑通 + CEO veto halt_workflow
-# R7.4 acceptance: 用户三种前缀场景任一都跑通 (JINJA `{{ gstack_prefix }}` 插值)
+# workflows/plan-feature/workflow.yaml — Phase v2.0-2.4 W1 T2.4.W1.3 (D-15 + Q-AUDIT-5a LOCKED)
+# 5-phase 三层栈 composition: gstack governance → superpowers brainstorm → GSD discuss/plan → planning-with-files plugin /plan persist
+# v1 → v2 delta (sister L1-40):
+#   ADD `schema_version: harnessed.workflow.v2`              (T2.4.W0.1 16th surface)
+#   ADD `description` root field                              (workflow purpose)
+#   ADD `capability: '{{ capabilities.<name>.cmd }}'` 各 phase (D-10 template interpolation)
+#   ADD `gate: judgments.<file>.<trigger>.fires` 01-gstack-decision (D-04 gate reference)
+#   ADD `on: [{if, invoke}]` 02-brainstorm + 03-gsd-discuss   (D-09 conditional clause)
+#   ADD `artifacts_expected: [...]` 05-persist                (D-15 planning-with-files 3 file 持久化)
+#   ADD `max_iterations: '{{ defaults.ralph_max_iterations.plan-feature.<phase> }}'` (T2.3.W1.2 ship)
+#   KEEP `invokes: '/plan'` literal at 05-persist             (Q-AUDIT-5a Claude Code plugin slash cmd, NOT npm SDK NOT fs.writeFile)
+# Q-AUDIT-5a anti-pattern guard (verified by tests/workflow/plan-feature-v2.test.ts):
+#   - NO npm SDK call (no `import` / `require` planning-with-files — yaml syntax natively absent)
+#   - NO fs.writeFile self-impl (no `fs.writeFile` / `writeFile` reference — anti-pattern reviewer guard)
+#   - YES `invokes: '/plan'` literal at 05-persist            (plugin slash cmd 真接 path a)
+#   - YES `capability: '{{ capabilities.planning-with-files.cmd }}'` template ref (resolve to /plan via capabilities.yaml L317)
+# 04-gsd-plan phase: `gsd-plan-phase` 在 capabilities.yaml 中暂未单独 entry — 沿袭 v1 L32 literal pattern (D-08 v2.x patch followup
+# 添加 gsd-plan-phase capability entry; v2.0 baseline 不阻塞 ship per T2.4.W1.3 Implementation Tasks #4-#5 scope)
+
+schema_version: harnessed.workflow.v2
 workflow: plan-feature
-on_veto: halt_workflow  # workflow-level (NOT per-phase per RESEARCH § 5.2 DRY) — D-04 PUSH 任 1 phase 转换前 read = vetoed → 全 halt
+description: |
+  plan-feature 5-phase 三层栈 workflow — gstack governance gate (CEO/EM/Designer/Paranoid/QA/CSO)
+  → superpowers brainstorm 子任务级澄清 → GSD /gsd-discuss-phase 灰色澄清 → GSD /gsd-plan-phase
+  Wave A/B/C planning → planning-with-files Claude Code plugin slash cmd `/plan` 持久化 3 file
+  (task_plan.md + progress.md + findings.md) 在 .planning/<phase-id>/.
+  Phase 1 gstack governance veto halts the entire workflow (D-04 PUSH) — engine-level
+  on_veto wiring lives in runtime, NOT yaml v2 root schema (WorkflowSchemaV2 strict
+  additionalProperties:false per T2.4.W0.1 + STRIDE T-2.2-02).
+
 phases:
   - id: 01-gstack-decision
     name: gstack-decision (governance gate — CEO/EM/Designer/Paranoid/QA/CSO)
     upstream: gstack
-    model: opus              # 高层决策 = opus
-    invokes: '{{ gstack_prefix }}office-hours'  # D-02 JINJA 插值 (gstack-office-hours OR office-hours)
-    skills: ['plan-feature-decision']
-    max_iterations: 1
+    capability: '{{ capabilities.gstack-office-hours.cmd }}'
+    invokes: '{{ gstack_prefix }}office-hours'
+    model: opus
+    gate: judgments.strategic-gate.office-hours.fires
+    max_iterations: '{{ defaults.ralph_max_iterations.plan-feature.01-gstack-decision }}'
+
   - id: 02-brainstorm
-    name: brainstorm (execution + ui-ux-pro-max UI 任务)
+    name: brainstorm (子任务级澄清 multi-approach)
     upstream: superpowers
-    model: sonnet            # 设计澄清 = sonnet
-    skills: ['plan-feature-brainstorm']
-    max_iterations: 5
+    capability: '{{ capabilities.superpowers-brainstorming.cmd }}'
+    model: sonnet
+    on:
+      - if: judgments.subtask-gate.brainstorming.fires
+        invoke: '{{ capabilities.superpowers-brainstorming.cmd }}'
+    max_iterations: '{{ defaults.ralph_max_iterations.plan-feature.02-brainstorm }}'
+
   - id: 03-gsd-discuss
-    name: gsd-discuss (orchestration — phase discussion)
+    name: gsd-discuss (orchestration — phase 灰色澄清 4-question batch)
     upstream: gsd
+    capability: '{{ capabilities.gsd-discuss-phase.cmd }}'
     model: sonnet
-    invokes: 'gsd-discuss-phase'  # GSD 命令固定无 prefix
-    skills: ['plan-feature-discuss']
-    max_iterations: 3
+    on:
+      - if: judgments.phase-gate.gsd-discuss-phase.fires
+        invoke: '{{ capabilities.gsd-discuss-phase.cmd }}'
+    max_iterations: '{{ defaults.ralph_max_iterations.plan-feature.03-gsd-discuss }}'
+
   - id: 04-gsd-plan
-    name: gsd-plan (orchestration — phase planning)
+    name: gsd-plan (orchestration — Wave A research + Wave B planner + Wave C checker)
     upstream: gsd
-    model: sonnet
     invokes: 'gsd-plan-phase'
-    skills: ['plan-feature-plan']
-    max_iterations: 3
+    model: sonnet
+    max_iterations: '{{ defaults.ralph_max_iterations.plan-feature.04-gsd-plan }}'
+
   - id: 05-persist
-    name: persist (planning-with-files MD persistence)
+    name: persist (planning-with-files plugin /plan — Manus-style markdown persistence)
     upstream: planning-with-files
-    model: haiku             # 文档写入 = haiku 省 token (sister Phase 2.2 CD-2 04-deliver=haiku)
-    skills: ['plan-feature-persist']
-    max_iterations: 5
+    capability: '{{ capabilities.planning-with-files.cmd }}'
+    invokes: '/plan'
+    model: haiku
+    on:
+      - if: phase.scope_days > 1 or phase.is_critical_module == true
+        action: invoke
+    artifacts_expected:
+      - task_plan.md
+      - progress.md
+      - findings.md
+    max_iterations: '{{ defaults.ralph_max_iterations.plan-feature.05-persist }}'

package/workflows/research/SKILL.md

@@ -0,0 +1,35 @@
+---
+name: research
+description: 多源调研 workflow — Tavily/Exa/ctx7 多源 fan-out + GSD discuss synth aggregate; harnessed v2.0 NEW per R20.7 (Stage ① Discuss 独立 call). Triggered by harnessed CLI `harnessed research --topic <text>` or slash command `/research` after `harnessed setup`.
+preamble-tier: 2
+---
+
+# research workflow
+
+Multi-source research workflow shipped with harnessed v2.0 (Stage ① Discuss).
+
+## Phases (2)
+
+| # | Phase | Upstream | Capability | Model | Description |
+|---|-------|----------|-----------|-------|-------------|
+| 01 | `01-fan-out` | web-search | (route-by-subtask) | sonnet | 3 source fan-out (Tavily MCP / Exa MCP / ctx7 CLI per ~/.claude/rules/web-search.md + context7.md routing) |
+| 02 | `02-synth` | gsd | `gsd-discuss-phase` | opus | GSD discuss-phase aggregate + dedup + reconcile |
+
+## Capability refs
+
+Sister `workflows/capabilities.yaml` entries:
+- `tavily-mcp` (impl: mcp-server, default 关键词)
+- `exa-mcp` (impl: mcp-server, 描述式 / 学术)
+- `ctx7` (impl: cli-tool, 库 API 文档)
+- `gsd-discuss-phase` (synth aggregate)
+
+## Invocation
+- CLI: `harnessed research --topic "<topic>"`
+- Slash command: `/research <topic>` (after `harnessed setup`)
+
+## Routing rules (sister ~/.claude/rules/web-search.md)
+- 描述式查询 ("找一篇对比 X 和 Y 的博客") → Exa MCP
+- 学术 / 论文 → Exa MCP
+- 库 / API 文档 → ctx7 CLI (per ~/.claude/rules/context7.md)
+- 关键词 / 时效内容 → Tavily MCP (默认)
+- 抓整站 / 站点结构 → Tavily crawl/map

package/workflows/research/workflow.yaml

@@ -0,0 +1,18 @@
+schema_version: harnessed.workflow.v2
+workflow: research
+description: 多源调研 workflow — Tavily MCP (默认关键词) + Exa MCP (描述式/学术) + ctx7 CLI (库 API 文档) fan-out + GSD discuss synth aggregate; sister ~/.claude/rules/web-search.md + context7.md routing 机器化 per R20.7 + D-08
+
+phases:
+  # 01-fan-out: 多源 fan-out — Tavily / Exa / ctx7 各源并行查询 (subagent fan-out per CLAUDE.md 子任务并行执行机制); generic shape — workflow engine 按 capability route 触发 3 source sequentially
+  - id: 01-fan-out
+    upstream: web-search
+    model: sonnet
+    max_iterations: '{{ defaults.ralph_max_iterations.research.01-fan-out }}'
+    parallelism: judgments.parallelism-gate.subagent-default.fires
+
+  # 02-synth: GSD discuss synth aggregate — 多源结果 dedup + reconcile + GSD discuss-phase format
+  - id: 02-synth
+    upstream: gsd
+    capability: '{{ capabilities.gsd-discuss-phase.cmd }}'
+    model: opus
+    max_iterations: '{{ defaults.ralph_max_iterations.research.02-synth }}'

package/workflows/verify-work/SKILL.md

@@ -0,0 +1,92 @@
+---
+name: verify-work
+description: |
+  verify-work workflow v2 — Stage ④ Verify 9-phase composition per CLAUDE.md verbatim
+  (gsd-verify-work + gsd-progress 必跑串行 → code-review 多 agent 并行 fan-out → gstack /review
+  关键模块强制 → 可选 /qa /cso /design-review conditional → code-simplifier 末尾串行 →
+  关键发布/大重构 PR 升级 4-specialist Agent Team Pattern C 多维度审查 互相质询)。
+  schema_version: harnessed.workflow.v2 with capability template interpolation + parallelism
+  gate refs + conditional `on` clauses. Triggered by harnessed CLI `harnessed verify-work
+  --phase <num>` or slash command `/verify-work` after `harnessed setup`.
+trigger_phrases:
+  - "verify work"
+  - "verify-work workflow"
+  - "Stage 4 verify"
+  - "跑 verify-work"
+---
+
+# verify-work workflow (v2)
+
+## Overview
+
+9-phase chain mapping the user's CLAUDE.md Stage ④ Verify discipline onto the harnessed
+runtime, fully `harnessed.workflow.v2` schema (Phase v2.0-2.4 W2 T2.4.W2.2 — D-12
+Q-AUDIT-2 amend + R20.12 + R20.11 + R20.14).
+
+| phase | id | upstream | model | capability / invokes | gate / on / parallelism |
+| ----- | -- | -------- | ----- | -------------------- | ----------------------- |
+| 1 | `01-gsd-verify-work` | gsd | sonnet | `{{ capabilities.gsd-verify-work.cmd }}` | serial — `max_iterations: 3` |
+| 2 | `02-gsd-progress` | gsd | haiku | `{{ capabilities.gsd-progress.cmd }}` | serial — `max_iterations: 2` |
+| 3 | `03-code-review-parallel` | mattpocock-skills | sonnet | `{{ capabilities.code-review.cmd }}` | `parallelism: judgments.parallelism-gate.subagent-default.fires` |
+| 4 | `04-gstack-review-conditional` | gstack | opus | `{{ capabilities.gstack-review.cmd }}` | `on: phase.is_critical_module → invoke / else skip` |
+| 5 | `05-qa-conditional` | gstack | sonnet | `{{ capabilities.gstack-qa.cmd }}` | `on: phase.has_ui_changes → invoke / else skip` |
+| 6 | `06-cso-conditional` | gstack | opus | `{{ capabilities.gstack-cso.cmd }}` | `on: phase.has_auth_or_secrets → invoke / else skip` |
+| 7 | `07-design-review-conditional` | gstack | sonnet | `{{ capabilities.gstack-design-review.cmd }}` | `on: phase.has_design_changes → invoke / else skip` |
+| 8 | `08-code-simplifier` | mattpocock-skills | sonnet | `{{ capabilities.code-simplifier.cmd }}` | serial 末尾 — `max_iterations: 5` |
+| 9 | `09-agent-team-multispecialist` | claude-platform | opus | `{{ capabilities.agent-teams-create.cmd }}` | `parallelism: judgments.parallelism-gate.agent-teams-upgrade.fires`; `on: is_major_release OR is_large_refactor → invoke` |
+
+Per-phase config loads from `workflows/verify-work/workflow.yaml`; engine.runRouting
+spawns each phase as a sub-agent via `@anthropic-ai/claude-agent-sdk` 0.3.142+.
+
+## Phase 9 Pattern C 4-specialist Agent Team (D-11 + R20.11)
+
+The Agent Team upgrade phase (`09-agent-team-multispecialist`) follows
+`~/.claude/rules/agent-teams.md` Pattern C 多维度审查 verbatim — 4 teammate
+(code-review + gstack-review + gstack-cso + gstack-qa) 互相 SendMessage 质询
+findings 是否真问题 (NOT fire-and-forget subagent fan-out)。Token estimate
+`team_cost < 2 × subagent_cost` is a prereq per agent-teams.md L34 (engine-level
+check, NOT yaml schema scope). Cleanup mandatory: `SendMessage shutdown_request`
++ `TeamDelete` after round-trip complete (防呆清单 per agent-teams.md L46-L48).
+
+Trigger condition: `phase.is_major_release == true OR phase.is_large_refactor
+== true`. 常规 PR / 单点任务**不触发** Pattern C — code-review fan-out (phase 3)
++ gstack-review conditional (phase 4) 已够用且更省 token (per CLAUDE.md "Verify
+阶段" 末段 "关键发布 / 大重构 PR" 限定语)。
+
+## R20.16 chain_isolation 实装 (3 铁律 fallback)
+
+每个 conditional phase (4-7, 9) 都实装 `on: [{if: ..., action: invoke}, {if:
+..., action: skip}]` — skip 04-gstack-review **不**级联跳过 06-cso-conditional,
+等等。验证脚本 `node scripts/check-workflow-schema.mjs` 严格 schema 校验 +
+runtime engine 独立判 condition。
+
+## CLI invocation
+
+```bash
+# Dry-run preview — arbitrate-only, never spawns SDK.
+harnessed verify-work --phase <num> --dry-run --non-interactive
+
+# Apply path — real SDK spawn + 9-phase chain (conditional phases evaluate by phase fact context).
+harnessed verify-work --phase <num> --apply
+```
+
+## Forward-looking note
+
+The `trigger_phrases:` frontmatter is active after `harnessed setup` copies this
+SKILL.md to `~/.claude/skills/verify-work/` — Claude Code then loads the slash
+command `/verify-work` automatically (Gap B fix — v1.0.2 sister mechanism).
+
+## References
+
+- D-12 (Q-AUDIT-2 amend) — verify-work 7+ phase 完整 4-stage 重定
+- R20.12 — verify-work full scope acceptance
+- R20.11 — 4-specialist Agent Team upgrade
+- R20.14 — special-purpose /qa /cso /design-review conditional
+- R20.7 — D-08 verify-work NEW 2 之一
+- R20.16 — fallback 3 铁律 chain_isolation
+- ~/.claude/CLAUDE.md "Verify 阶段" 7-step verbatim
+- ~/.claude/rules/agent-teams.md Pattern C 多维度审查
+- `workflows/capabilities.yaml` — gsd-verify-work / gsd-progress / code-review / gstack-{review,qa,cso,design-review} / code-simplifier / agent-teams-create entries
+- `workflows/judgments/parallelism-gate.yaml` — subagent-default + agent-teams-upgrade.fires
+- `workflows/defaults.yaml` — ralph_max_iterations.verify-work.* values
+- `docs/WORKFLOW.md` — 4-stage workflow mermaid + Stage ④ Verify 章节