BruceLanLan/dsh-tier-router1

dsh-tier-router

Tiered model routing for DeepSeek Harness: a strong tier for planning/architecture/review, a cheap tier for implementation, with plan-mode-aware auto routing, advisor/review consultations, a high-impact escalation guard, failure auto-escalation, durable tier configuration, and subagent tiering.

AI 分析

核心用途是实现多模型分层自动路由。适合需要平衡大模型规划能力与执行成本的用户,通过将规划与执行任务分流至不同档次模型来优化开销。需在 agent preset 中进行组合配置。

パッケージ
dsh-tier-router
バージョン
0.4.7
ライセンス
MIT
最終更新
2026/08/16

インストール

$npx -p @deepseek-ai/dsh dsh plugin --profile web add github:BruceLanLan/dsh-tier-router

ドキュメント

README 全文を読む ↗

Usage

Slash commands (typed in the composer; they affect only the current session)

/advisor                           # one strong-tier consultation
/tier status                                # routing state, escalation state, diagnostics
/tier strong | cheap                        # force one tier for this session (optionally persist as session default)
/tier auto                                  # restore auto for this session (plan -> strong, execution -> cheap)
/tier off                                   # disable routing for this session, restore its default model (other sessions unaffected)
/tier plan                                  # auto + enter plan mode, apply the strong header immediately
/tier models                                # list registered providers and their models
/tier set  
  [effort]
/tier subagent        # policy used to classify subagent execution (guard)
/tier review                         # strong-tier review

Sample output (/tier status):

Tiered model routing
  mode: global=auto, this session=auto (per-session via /tier strong|cheap|auto|off; escalate: 2 errors / 60s window -> 180s strong)
  strong: deepseek-official/deepseek-v4-pro (max)
  cheap:  deepseek-official/deepseek-v4-flash (high)
  subagents: inherit
  diag: requestSteps=42 guardChecks=31 guardDenies=3 headerWrites=4 errorsSeen=0 escalations=0
  session default: deepseek-official/deepseek-v4-pro (max)
  providers: deepseek-official, opencode-go, minimax

Model tools (called by the model when needed)

ToolPurpose
tier_advisorStrong-tier consultation: one question + evidence -> advice / risks / acceptance criteria
tier_reviewStrong-tier review: change set + validation results -> verdict with ranked issues
tier_routeSet this session's tier (strong/cheap/auto/off, optionally persisted)
tier_configureReconfigure either tier's provider/model/effort and the subagent policy
tier_workerDispatch a bounded task packet to a subagent on a chosen tier; supports outputSchema / toolFilter / maxDepth / persona
tier_statusRead-only diagnostics: global & session tiers, escalation state, listener counters, effective tier

Configuration

Runtime configuration (no restart needed):

/tier set strong deepseek-official deepseek-v4-pro max
/tier set cheap deepseek-official deepseek-v4-flash high

/tier set and tier_configure persist the tier configuration in the tier-router settings namespace (survives restarts; pass sessionOnly: true for a transient change). tier_route strong|cheap scopes to the current session and does NOT persist by default; pass persist: true to also write the session default model (agent-default-model). Failure auto-escalation parameters (threshold / window / TTL) are currently built-in constants; configurable in a later version.