返回目录
部署运维 插件

dsh-autotier

PerryLink/dsh-autotier

Automatic strong/cheap model-tier routing for DeepSeek Harness: intent-gated tier landing on the agent/request waterfall, plan-mode handoff (strong plans, cheap implements), deterministic high-risk guards on tools/pre-execute, failure escalation with TTL fallback, a /tier command, and tier status tools

Stars
1
Forks
0
Issues
0
更新
今天

PROJECT TOPICS

项目标签

INSTALL REFERENCE

安装参考

未验证
dsh plugin --profile web add github:PerryLink/dsh-autotier

该命令指向仓库当前默认分支;尚无绑定当前 commit 的完整验证结果。

PROJECT README

README

dsh-autotier

Gitee OpenSSF Scorecard License DSH plugin dsh-doctor DSH Market Node CI Version dshfind

English | 简体中文 | Español | Português | हिन्दी

Automatic model-tier routing for DeepSeek Harness: one user instruction enters, one tier decision comes out — no manual model switching.

Complex intent (architecture, planning, debugging, multi-step engineering) is planned on the strong tier and then implemented on the cheap tier. Simple intent (questions, retrieval, batch chores, daily work) is designed and implemented on the cheap tier directly. While the cheap tier executes, high-risk tool calls are denied by a deterministic guard, and repeated failures escalate to the strong tier with a TTL fallback.

Compatibility

Harness Status
@deepseek-ai/dsh 0.1.2-rc.1 no longer supported; that line predates the SettingsForms contract this plugin now targets
@deepseek-ai/dsh 0.1.5-rc.2 no longer supported; the settings.register / settings/updated contract it exposes was removed upstream
@deepseek-ai/dsh 0.1.6-alpha.2 no longer supported; same removal, first line to ship SettingsForms
@deepseek-ai/dsh 0.1.7-alpha.1 required; verified against the matching checkout (typecheck) and the published packages (typecheck:ci, 214 tests)
@deepseek-ai/cordis ^4.0.3, @deepseek-ai/cosmokit ^1.8.4, @deepseek-ai/schemastery ^3.18.3 peer baseline

Peer ranges name all published lines explicitly (>=0.1.2-rc.1 <0.2.0 || >=0.1.5-alpha.1 <0.2.0 || >=0.1.6-0 <0.2.0 || >=0.1.7-0 <0.2.0), because a semver range whose only prerelease comparator sits on an earlier version tuple does not admit a later alpha. They are refreshed per published wave. The declared range is deliberately wider than the verified one: it documents what the manifest accepts, not what has been tested.

This release is a breaking adaptation. The 0.1.6-generation harness removed the settings provider seam this plugin was built on: the @deepseek-ai/dsh-settings-file package is gone, ctx.settings is now SettingsForms (a schema→form projector, with no register), and settings/updated no longer exists. There is no shared surface that lets one plugin observe another plugin's settings document, so no version of this plugin can support both contracts at once. Configuration now flows through the host's volatile-config mechanism, and the per-session routing mode — the setting that actually changes behaviour at runtime — is unchanged.

The plugin is host-plane only. It needs no agent preset of its own: the host row applies to every session. A one-line prompt section in your preset is optional and only makes the router's decisions visible to the model (see Install & uninstall).

What you get

  • Intent gate — every turn is classified from deterministic signals (message text, tool names, image presence, conversation length). The zero-token rule layer decides when it is confident; only a low-confidence turn calls the cheap judge model, and never on a cooldown.
  • Tier landing on the official seam — the decision is applied on the agent/request waterfall by returning a replacement provider/model/effort triple. Sampling scalars the session already chose (temperature, maxTokens, stop) are preserved.
  • Plan-mode handoff — a complex instruction enters plan mode on the strong tier; leaving plan mode drops back to the cheap tier for implementation.
  • High-risk guard — while the cheap tier executes, destructive commands (rm -rf, sudo, mkfs, git push --force, credential-file writes, …) are denied with a corrective message telling the model to escalate instead.
  • Failure escalation — repeated failures (optionally same-signature) raise the tier for a TTL; a model/route failure walks the configured fallback chain.
  • Manual escape hatches/tier auto|strong|cheap|off and the tier_status / tier_route tools. Setting routingMode: delegated (or /tier off) stops routing for a session that must keep its own model.
  • ctx.autotier service — a small read surface (status) plus the autotier/route veto waterfall and autotier/tier-changed event, so other plugins can observe or override a decision.

Quick start

dsh plugin --profile web add dsh-autotier
npm i -g dsh1024
dsh1024 plugin --profile web add dsh-autotier

Then start (or restart) the harness. The row is appended to your profile's cordis.patch.yml; routing starts on the next turn with no further setup.

Install & uninstall

npm channel

dsh plugin --profile web add dsh-autotier
npm i -g dsh1024
dsh1024 plugin --profile web add dsh-autotier

git channel

dsh plugin --profile web add "github:PerryLink/dsh-autotier#main"
git clone https://github.com/PerryLink/dsh-autotier.git
cd dsh-autotier && pnpm install && pnpm run build
dsh plugin --profile web add .

Optional preset prompt section. The router works without it. To let the model know which tier it is running on, add one row to your agent preset (docs/preset-row.md has the exact block):

- insert:
    - id: autotier-prompt
      name: '@deepseek-ai/dsh-system-prompt'
      # sections: [...]  — see docs/preset-row.md

Uninstall

dsh plugin --profile web remove dsh-autotier

The row, its command, its tools and its listeners are all removed with the plugin. Configuration a user saved through the Plugins page lives in the active profile patch and belongs to that profile, not to this plugin; removing the row leaves it there untouched.

Configuration

Every key is validated at load time; an invalid value fails loudly instead of silently disabling routing. cordis.patch.yml in this repository documents the same keys inline.

Key Default Meaning
tiers.strong.provider deepseek-official Provider for the planning/review tier.
tiers.strong.model deepseek-v4-pro Catalog id of the strong model.
tiers.strong.effort high Adapter vocabulary off | low | high | max.
tiers.strong.followSession false false = this tier's effort overrides the session's.
tiers.strong.fallback [] Ordered provider/model landings when the tier is unavailable.
tiers.cheap.provider deepseek-official Provider for the implementation tier.
tiers.cheap.model deepseek-flash Catalog id of the cheap model.
tiers.cheap.effort low Adapter vocabulary off | low | high | max.
tiers.cheap.followSession true true = inherit the session's effort so an explicit choice wins.
tiers.cheap.fallback [] Ordered provider/model landings when the tier is unavailable.
tiers.vision.provider deepseek-official Provider for image-carrying turns.
tiers.vision.model deepseek-flash The catalog's image-capable model.
intent.ruleThreshold 0.7 Confidence at or above which the rule layer decides alone.
intent.attemptBand.enabled false Start the middle band on the cheap tier and escalate on a signal.
intent.attemptBand.tauLow 0.45 Lower bound of the attempt-first band.
intent.hysteresis.toStrong 0.8 Score that switches a cheap turn to strong.
intent.hysteresis.toCheap 0.6 Score below which a strong turn returns to cheap.
intent.rules [] Declarative rule table (when.patterns / when.tools / when.cwd, tier, priority).
intent.judge.enabled true Allow the low-confidence judge.
intent.judge.model '' Judge model id; empty = first catalog model containing flash.
intent.judge.temperature 0 Judge sampling temperature.
intent.judge.maxTokens 16 Judge output cap (it answers with one word).
intent.judge.cooldownMs 30000 Minimum gap between two judge calls.
intent.judge.timeoutMs 2000 Judge call timeout.
intent.judge.unavailableSkip 2 Consecutive judge failures after which the turn skips it.
intent.scenarios all true Per-scenario switches: coding, review, planning, retrieval, batch, daily, longText, multimodal.
intent.costMode balanced Ambiguity arbitration: cost-first | quality-first | balanced.
guard.enabled true Enable the deterministic high-risk guard.
guard.tiers [cheap] Tiers the guard protects.
guard.whitelist [] Commands, tools or path prefixes that never trip the guard.
guard.protectedPaths ['.dsh','AGENTS.md','package.json','.github/workflows'] Self-modification surfaces that force strong-tier review.
guard.interopDefend auto Relationship with dsh-defend: auto audits coexistence, none stays silent.
escalation.threshold 2 Failures within the window that raise the tier.
escalation.windowMs 60000 Failure-counting window.
escalation.ttlMs 180000 How long an escalation stays in effect.
escalation.fallbackTtlMs 300000 TTL used after a fallback landing was taken.
escalation.signature true Count same-signature recurrences instead of every failure.
routingMode auto auto | strong | cheap | delegated | off.

All keys can also be edited live from the plugin's card on the Plugins page; the Host validates the new value against the schema and commits it with loader/volatile-update, and this plugin then re-judges the whole configuration through the same cross-field judge the mount path uses. A value that violates a cross-field requirement (two tiers landing on the same route, a hysteresis pair that cannot stop flapping, a rule with neither a pattern nor a tool) leaves the last good policy routing instead of swapping in something unroutable.

routingMode is the one key that is not live: it is the composition default, and the runtime switch is the session-scoped override written by /tier, the composer pill and the card's selector. Making it live as well would give the same behaviour two owners.

Tools & surfaces

Surface Kind Purpose
/tier command auto | strong | cheap | off | status; session-scoped override.
tier_status tool Current tier, mode, escalation TTL and guard state.
tier_route tool Route one intent string without sending a request (dry run).
ctx.autotier service status() read surface for other plugins.
autotier/route serial event Third parties may veto a proposed tier.
autotier/tier-changed emit event Observability when the effective tier changes.

Permissions & data

  • Files — the plugin reads no file and writes none. Configuration is the host's: a value saved on the Plugins page is persisted by the Host into the active profile patch, and this plugin only reads the live snapshot it is handed.
  • Network — the only outbound traffic is the judge call, which goes through the normal ctx.llm path and the configured provider.
  • Session log — the plugin appends no custom session events. The routing trail is the plugin logger plus the live autotier/tier-changed bus event; the only append it makes is the plan/mode fallback when the plan-mode service is absent. Custom event types are fail-closed on 0.1.2-alpha.1 and later, so no durable plugin-owned record is written.
  • Secrets — no credential is read, logged or stored by this plugin.

Security boundaries

  • The guard is a defence in depth, not a sandbox. It denies the patterns it knows on the cheap tier and never weakens dsh-defend, the approval service, or the sandbox policy. Keep those enabled.
  • The guard protects only the tiers listed in guard.tiers (cheap by default). A strong-tier turn is not blocked by design: the strong model is the reviewer.
  • If the guard itself throws, the call is escalated to the strong tier rather than allowed — a broken guard must not become an open door.
  • /tier off disables routing entirely; the harness then behaves exactly as it did before the plugin was installed.

Known limitations

  • The rule layer is deterministic and therefore finite: a novel phrasing of a complex request may start on the cheap tier and escalate only after a failure or a guard denial. The judge call covers the low-confidence middle.
  • Escalation is per-agent and in-memory; a harness restart starts from auto.
  • Tier switching resets the provider prompt cache for the changed request, so very chatty sessions may see a small cache-miss cost on the switch turn. The hysteresis thresholds exist to keep that rare.
  • The plugin routes conversation requests. Compaction and title generation are separate host seams; align their own model settings with the cheap tier if you want the same cost profile (docs/supporting-lanes.md).
  • followSession: true on the cheap tier means an explicit session model choice wins; in that case the cheap tier cannot force its own model.
  • The Settings card and composer tier pill shipped in 0.2.0. The card (routing mode, live tier landings, model catalog) lives in the Plugins settings section and the pill cycles the session mode from the composer.
  • A model picked in the GUI is not detected automatically. The router does not watch the agent-default-model document, so switching models there does not by itself stop routing — use routingMode: delegated or /tier off.
  • Fingerprint posteriors are in-memory. They reset on restart and re-learn from the judge's cold-start fallback; persisting them through the settings document is planned for v0.2.
  • The attempt-first middle band ships disabled. intent.attemptBand.enabled stays false until the calibration corpus and its metric gate land (v0.2).

Development

pnpm install
pnpm run typecheck      # against the local harness checkout type faces
pnpm run typecheck:ci   # against the published 0.1.7-alpha.1 faces (what CI runs)
pnpm test
pnpm run build
pnpm run verify:self-contained
pnpm run verify:artifacts
pnpm pack

pnpm run build emits lib/types (tsc declarations) and lib/index.js (tsdown bundle). Tests use the published host packages directly — real Context, real session/tools/commands services — plus one real Loader composition over a temporary cordis.yml. The settings service is the one stand-in: the host's real SettingsForms binds to configEditor, profileContext and the Loader's fiber graph, none of which this plugin consumes, while the plugin's whole contract with it is configure({ auto }) plus the Loader's loader/volatile-update event.

Topics

dsh, dsh-plugin, deepseek-harness, deepseek, cordis, router, model-tier, cost, auto.

Contributors

PerryLink. Issues and pull requests are welcome at https://github.com/PerryLink/dsh-autotier/issues.

Applicable DSH version: verified against dsh-v0.1.7-alpha.1 (the host release this build targets); requires >=0.1.7-alpha.1 <0.2.0.

PerryLink DSH Plugin Family

This project is one of the 45 DeepSeek Harness plugins maintained by PerryLink. If this one helps you, the others likely will too:

Plugin One-liner
dsh-auto-review Second-model auto-review on the approval chain, fail-closed by default
dsh-autotier Automatic strong/cheap model-tier routing with deterministic risk guards and a /tier command
dsh-background-agents Durable background child agents with a Web UI sidebar, messaging and interrupt
dsh-budget Cost governance for DeepSeek Harness: budgets, carbon, and latency in one panel.
dsh-catalog DSH Desktop Market standard catalog source for the PerryLink family
dsh-cert-mcp Read-only MCP server exposing the certification registry: grades, snapshots and five-dimension evidence
dsh-checkpoint-rewind Claude Code /rewind-equivalent: snapshots, session forks, one-shot restore
dsh-claude-move Migrate Claude Code sessions, memory, skills and CLAUDE.md into DSH
dsh-click Cross-platform native desktop control for DeepSeek Harness — Windows first.
dsh-composer-history Terminal-style input history for the web composer: arrows, Ctrl+R search
dsh-data-quality Dataset quality checks and citation cross-checks (the optional numeric bridge consumed here)
dsh-defend Prompt-injection, jailbreak, and secret-leak defense for DeepSeek Harness.
dsh-doublecheck Engineering-discipline guard: requirements grill, test gates, adversary review
dsh-draw Unified static-image generation routing for DeepSeek Harness.
dsh-fast Read-only performance diagnostics for DeepSeek Harness.
dsh-fund-research Deterministic research reports for Chinese public mutual funds
dsh-github GitHub PR/issues integration for DSH, every write gated by approval
dsh-industry-research Industry research orchestration that seals its deliverables through this plugin's ctx.researchReport.assemble
dsh-laya Laya typed decisions (noul/choice/score) as a first-class Cordis service and model-visible tools
dsh-library Local document knowledge base for DeepSeek Harness.
dsh-local-ai Local-model (Ollama) integration for DeepSeek Harness.
dsh-lsp-actions LSP diagnostics, formatting, completion, code actions and rename over language servers
dsh-mask PII masking middleware: anonymize at the model boundary, restore at the display layer
dsh-mcp-panel Read-only MCP runtime panel: /mcp command + Settings tab with status, tools and errors
dsh-memento Approval-gated cross-session memory: ctx.memory seam + SQLite + memory tool
dsh-observe OpenTelemetry and Langfuse observability exporter for DeepSeek Harness.
dsh-output-styles Claude Code outputStyles-equivalent runtime style switching
dsh-permission-rules Claude Code-style declarative allow/deny/ask permission rules with audit
dsh-plugin-certification Community certification registry with repro-checkable grades and badges
dsh-plugin-doctor Zero-dependency static + sandbox smoke detector for DSH plugins
dsh-plugin-guide Plugin-development knowledge base as an on-demand agent skill
dsh-plugin-kit Shared zero-runtime-dependency toolkit for the PerryLink DSH plugins
dsh-plugin-upgrade One-package, one-corridor-index plugin upgrade skill: routes a repository to the matching closed corridor card
dsh-plugin-upgrade-015 Merged 0.1.3-alpha.10.1.5-rc.1 upgrade corridor card plus a zero-dependency seam scanner
dsh-reach Multi-channel approval/question bridge: WeChat/Telegram/Feishu, session console
dsh-research-report Verifiable research-report engine: content-addressed evidence ledger and sealed versions
dsh-score Multi-dimensional quality scoring for DeepSeek Harness plugins.
dsh-session-pin Pin sessions in the Web sidebar with durable ordering
dsh-session-sync Cross-device session sync for DeepSeek Harness — a dedicated git mirror of your session store.
dsh-skill-pack-security Security-audit skill pack: secret scan, dependency and supply-chain review
dsh-talk Voice-first session loop for DeepSeek Harness: talk to it, hear it answer.
dsh-team-rooms Cross-session team rooms: shared message bus, task board and timeline
dsh-test-drive Isolated install-and-smoke test drives for DeepSeek Harness plugins.
dsh-ticktick TickTick/Dida365 task bridge: session-header panel + 11 tools
dsh-translate Vendor parameter translation and deterministic JSON repair for DeepSeek Harness.

License

Apache-2.0. See LICENSE and THIRD_PARTY_NOTICES.md.

CLASSIFICATION EVIDENCE

分类依据

项目类型插件
功能分类部署运维
规则置信度

系统优先读取 GitHub Topics,再与站内分类词典和词根规则比对。当前命中: cost。