| name | risk-pipeline-risk-pipeline-dispatch |
| description | Dispatch risk vector -> reviewer roster + impl/reviewer model + tdd_required. Triggers: s>=- security-reviewer, d>=- data-reviewer, r>=+ reversibility-reviewer, b>=+ full code-quality, u>=+ adds novelty-reviewer + research. Crit axes hard-block. Model routes impl by tier (haiku-4.5/sonnet-4.6/opus-4.7) with crit + u-crit escalation; reviewer one tier down with security/post-task same-tier overrides. TDD when d|s>=- or b|u>=+. |
| disable-model-invocation | true |
risk-pipeline-dispatch
概覽 (Overview)
風險向量轉派發規格之技藝。呼叫方：risk-classify（Phase 06），在 risk-budget（Phase 07）產出 verdict + task_risk + pipeline_tier 之後調用。輸入風險向量、裁決、管道層、trivial_bypass 旗、配置，輸出 pipeline_spec 對象：審者陣列、impl/reviewer 模型、tdd_required 旗、硬阻增量（human_checkpoint/evidence/artifacts/research_task）。配對表照 spec §Pipeline dispatch line 368–384 字字照辦；模型路由照 line 386–406；TDD 規照 line 408–417。
輸入 (Inputs)
呼叫方必供：
task_risk: {b, d, s, r, u, scalar, crit_axes} — 五軸整數（0–3）、總標量、crit 軸名（全名如 security 非 s）verdict: ok | split_required | refactor_first_required | escalate — 自 risk-budgetpipeline_tier: smoke | light | dim_matched | architectural — 自 classifytrivial_bypass: bool — 前濳八檢結果config.model_routing — 讀 .claude/rules/risk.md frontmatter；缺鍵回退 spec 默認lci_diff — 供 TDD skip 之 behavior-change 檢（無則保守視為有變動）
配對表 (Dim-matched dispatch table)
Spec §Pipeline dispatch line 368–384 十一行照錄：
| Trigger | Reviewer fires | Blocking |
|---|
s >= - | security-reviewer | yes |
s = ! | + human checkpoint | hard-block |
d >= - | data-reviewer | yes |
d = ! | + migration dry-run evidence | hard-block |
r >= + | reversibility-reviewer | yes |
r = ! | + rollback runbook + canary plan | hard-block |
b >= + | code-quality-reviewer (full) | yes |
u >= + | novelty-reviewer + research task first | yes |
Any axis >= - AND non-trivial | qa-reviewer | yes |
All axes . | no reviewers | smoke |
| Trivial bypass | no reviewers | smoke |
觸發累加：一軸滿多條即多審派發（如 s! 兼觸 s>=- 與 s=!，派 security-reviewer + human_checkpoint）。reviewers 陣列去重。
硬阻 (Hard-block augmentations)
crit 軸觸發 hard-block，pipeline_spec 增欄：
| Trigger | Augmentation |
|---|
s = ! | human_checkpoint: required |
d = ! | evidence_required: [migration_dry_run] |
r = ! | artifacts_required: [rollback_runbook, canary_plan] |
多硬阻共現則並拼（如 s!d! 則 human_checkpoint + evidence_required）。verdict: escalate 無 crit 軸時亦強置 human_checkpoint: required（findability 違反升人工檢）。
並行與重試 (Parallel execution + retry)
- 全員審者並行發火（
reviewer_parallelism: all_parallel），無序依。
- 重試：每失敗審者最多
2 次重試，同輸入重發。retry_policy: {max_retries: 2, scope: failing_only}——僅失敗者重跑，成者不動。
- 重試後仍失則硬阻管道完成，待呼叫方（loop 協調器）決斷（修任務/分拆/升級）。
- 重試機制本身非本技藝實作——此文件僅述策略。實作見 Phase 10/11 loop 協調器。
模型路由 — impl (Model routing: implementer)
決策表，由上而下求交：
| Condition | Model |
|---|
trivial_bypass == true | haiku-4.5 |
| scalar 0–4 | haiku-4.5 |
| scalar 5–9 | sonnet-4.6 |
| scalar 10–14 | sonnet-4.6 |
scalar >= 15 | opus-4.7 |
any axis = ! | opus-4.7 (override) |
u = ! | opus-4.7 (research phase) |
優先序明文：crit 軸覆寫 > scalar 層。即任軸 = ! 則 opus-4.7，無論 scalar 多低（除 trivial_bypass 不可能與 crit 共存——前濳檢查 8 安全路徑 + 無 schema 變動皆排除 crit 可能）。配置缺鍵以 spec 默認補。
模型路由 — reviewer (Model routing: reviewer)
| Case | Model |
|---|
| 默認 | one tier down from impl（地板 haiku-4.5） |
security-reviewer | same tier as impl（特例不降） |
post-task-reviewer | same tier as impl（特例不降） |
tier 階梯：haiku-4.5 → sonnet-4.6 → opus-4.7。「一級下」映射：
opus-4.7 → sonnet-4.6sonnet-4.6 → haiku-4.5haiku-4.5 → haiku-4.5（已地板）
其他 reviewer（qa-reviewer、data-reviewer、reversibility-reviewer、novelty-reviewer、code-quality-reviewer）皆走默認規則。
TDD 必規 (TDD required)
必備於任一成：
d >= -（測 migration/transform）s >= -（測 auth 路徑）b >= +（回歸面）u >= +（探索前 pin 行為）
僅可跳於全備成：
- 上四條件皆不中（
d == . 且 s == . 且 b < + 且 u < +）
trivial_bypass == true 或 純樣式/文案/配置改動lci_diff 示無行為變（無新符號、無函數簽名變、無控制流變；lci_diff 不可得時保守視為有變）
默認：required。跳 TDD 為顯性 opt-in，非 opt-out。三條任一不成即 tdd_required: true。
輸出 (pipeline_spec schema)
pipeline_spec:
impl_model: haiku-4.5 | sonnet-4.6 | opus-4.7
required_reviewers:
- {name: security-reviewer, model: sonnet-4.6, blocking: true}
- {name: code-quality-reviewer, model: haiku-4.5, blocking: true}
- {name: qa-reviewer, model: haiku-4.5, blocking: true}
reviewer_parallelism: all_parallel
retry_policy: {max_retries: 2, scope: failing_only}
tdd_required: true | false
human_checkpoint: required | false
evidence_required: [migration_dry_run]
artifacts_required: [rollback_runbook, canary_plan]
research_task_required:
欄義：impl_model 必出三檔之一。required_reviewers 元素含 name/model/blocking；reviewer 模型由上節規則獨立計算。reviewer_parallelism 恆 all_parallel。retry_policy 恆 {2, failing_only}。human_checkpoint 省時等同 false。evidence_required artifacts_required 僅 crit 時出。research_task_required u>=+ 時 true（Phase 11 workflow 創任務）。
範例 (Examples)
重用 Phase 07 例三則。
例 1 — b.d.s.r.u. + smoke（trivial bypass）：
task_risk: {b:0, d:0, s:0, r:0, u:0, scalar:0, crit_axes:[]}
pipeline_tier: smoke
trivial_bypass: true
pipeline_spec:
impl_model: haiku-4.5 # trivial_bypass → haiku
required_reviewers: [] # "All axes ." 行 → no reviewers
reviewer_parallelism: all_parallel
retry_policy: {max_retries: 2, scope: failing_only}
tdd_required: false # 四條件皆 false + trivial + 假設 lci_diff 無行為變
human_checkpoint: false
research_task_required: false
例 2 — b+d.s.r.u. scalar 6（blast 高）：
task_risk: {b:2 (effective 3 after amp), d:0, s:0, r:0, u:0, scalar:6, crit_axes:[]}
pipeline_tier: light
trivial_bypass: false
觸發分析：
- b >= +（b=3）→ code-quality-reviewer (full)
- 任軸 >= - AND 非 trivial → qa-reviewer
- s/d/r/u 皆 < 觸發閾 → 其他 reviewer 不派
pipeline_spec:
impl_model: sonnet-4.6 # scalar 5-9
required_reviewers:
- {name: code-quality-reviewer, model: haiku-4.5, blocking: true} # 一級下
- {name: qa-reviewer, model: haiku-4.5, blocking: true}
reviewer_parallelism: all_parallel
retry_policy: {max_retries: 2, scope: failing_only}
tdd_required: true # b >= + 觸
human_checkpoint: false
research_task_required: false
例 3 — b+d.s!r-u. scalar 19 + crit security + escalate：
task_risk: {b:2, d:0, s:3, r:1, u:0, scalar:19, crit_axes:[security]}
verdict: escalate
pipeline_tier: architectural
trivial_bypass: false
觸發分析：
- s >= -（s=3）→ security-reviewer
- s = ! → human_checkpoint
- r >= +（r=1? no, r=1=-，不觸 r>=+ 閾 2）→ 無 reversibility-reviewer
※ 重審：r=1 即 `-` 級，`r >= +` 需 r>=2。此例 r 不觸。
- b >= +（b=2）→ code-quality-reviewer
- 任軸 >= - → qa-reviewer
- u < + → 無 novelty；無 research_task
pipeline_spec:
impl_model: opus-4.7 # crit 軸覆寫
required_reviewers:
- {name: security-reviewer, model: opus-4.7, blocking: true} # same-tier
- {name: code-quality-reviewer, model: sonnet-4.6, blocking: true} # 一級下
- {name: qa-reviewer, model: sonnet-4.6, blocking: true}
reviewer_parallelism: all_parallel
retry_policy: {max_retries: 2, scope: failing_only}
tdd_required: true # s >= -, b >= + 皆觸
human_checkpoint: required # s=! crit + escalate 雙觸
research_task_required: false
notes: "crit_security hard-block"
錯誤與回退 (Error + fallback)
- 配置
model_routing 缺：用 spec §Model routing 默認（上節表格所載）；notes: "model_routing_fallback"。
- 審者名未注於代理註冊表（Phase 14 未落地前可能）：降級為
code-quality-reviewer，記 warning notes: "reviewer <name> unregistered, downgraded"。呼叫方不阻。
verdict == escalate 無 crit 軸（findability 違反所致）：仍強置 human_checkpoint: required；reviewer 集全派架構層羅盤（security+data+reversibility+novelty+qa+code-quality）。pipeline_tier: smoke 與任一觸發條件衝突（前濳應排除，但程式漏洞可能致）：記 integrity violation 警告，守保守——reviewers 發火，忽略 smoke 層 bypass。notes: "integrity_violation_smoke_vs_trigger"。lci_diff 不可得於 TDD skip 判斷：保守視為有行為變，tdd_required: true。
依賴 (Dependencies)
- 消費：
risk-budget（Phase 07，已建，task spRTJjH4LPHS）之 task_risk + verdict。
- 被調：
risk-classify（Phase 06，已建，task bjKHROwLCMaq）——Phase 06 當前使用 dispatch fallback stub，本技藝落地後 classify 可移除 fallback。
- 審者名引：spec §Reviewer roster line 357–367 七員（
code-quality-reviewer、qa-reviewer、post-task-reviewer 三舊；security-reviewer、data-reviewer、reversibility-reviewer、novelty-reviewer 四新）。四新者代理體 Phase 14（emlONCR1TRH4）創建；本技藝僅引名，不實作。
- 配置：
.claude/rules/risk.md frontmatter 之 model_routing 塊（Phase 09，已建）。
- loop 協調：重試機制、research 任務創建、hook 觸發皆 Phase 10（dartai）/ Phase 11（workflow）/ Phase 04（PostToolUse hook）之責，本技藝不實作。
