Neptune810/dsh-model-router

English | 中文

A DeepSeek Harness plugin that keeps a pool of models and routes every step to one of them with its reasoning effort set. Thinking stays on — DeepSeek rejects a thinking-enabled request whose history contains a tool call made with thinking off.

Routing table

Step class	Decided by	Reasoning effort	Internal point
trivial	a clearly cheap intent (translate, rename, reformat) in a short step	low	~50
standard	a plain short request with no engineering cue	low	~50
engineering	engineering cues, code/diff/XML structure, or an agent tool loop	high	~75
hard	a dense engineering brief, or failures earned inside the task	high (max when allowMax)	~75 / 100

high is the default ceiling. V4.1-Flash drives effort from an internal 1–100 scalar and exposes three preset names; high corresponds to roughly 75, which is where the published effort curve is still steep. Pushing past it costs about 1.6–1.8x the output tokens for a marginal gain.

Authorization scopes, the model pool and task presets (v0.5)

control decides what the router owns, and a manual pick only counts in the scope that owns it: an effort change stands down full/effort but not model, and a model switch stands down full/model but not effort. While the router is down it stays out until the next command or the composer control's resume action.

Scope	Owns
full — 模型+思考 / Model+Effort	model + reasoning effort
effort — 思考 / Effort	one model you pick in the panel: the router only sets that model's thinking level
model — 模型 / Model	model only; the session keeps its thinking level

In effort scope the panel first asks you to pick exactly one model (POST /model-router/effort-model, reported by /state as effortModel / effectiveEffortModel). Until you do, the router stays out entirely — it changes neither the model nor the effort. Once one is bound, a manual model switch does not stand it down; changing the thinking level by hand does.

The client half registers its control into the shipped conversation.input.model seat at priority -1, taking over the composer model cell; it carries the official model + effort sections plus the scope switch, the resume action, the task-type picker, the model pool and the task presets, and talks to the host over same-origin /model-router/* routes.

Model pool. Only pooled models are ever selected.

pool:
  - { id: deepseek-official/deepseek-flash, cost: 1, tier: cheap }
  - { id: deepseek-official/deepseek-v4-pro, cost: 8, tier: strong, maxPerTask: 3 }
  - { id: deepseek-official/deepseek-v4-flash-vision-exp, cost: 1, tags: [vision] }
  - { id: vendor-x/writer-pro, cost: 20, weights: { 小说续写: 95 } }

tier: strong is what hard steps and evidence escalation prefer; cost feeds scoring.costPenalty; maxPerTask caps how many steps of one task may spend that entry; a vision tag is required for image steps once any pooled model carries one.

Task presets. A user-defined task type with rules and per-model weights:

presets:
  小说续写:
    match: ["续写", "小说", "文风", "/chapter\\s+\\d+/i"]
    weights: { "vendor-x/writer-pro": 95, "deepseek-official/deepseek-v4-pro": 60 }
  代码重构:
    match: ["重构", "refactor", "架构"]
    weights: { "deepseek-official/deepseek-v4-pro": 90 }

A type pinned in the composer always wins; with nothing pinned the keyword rules run (classifier: rules). Third-party models keep their own effort vocabulary: the router asks the provider what it supports, picks the nearest level, and sends no effort field when the model advertises none.

Opaque by default (the theme menu colour is translucent, which let the conversation show through):

The router control is the composer model cell (v0.11.0)

The router's control registers into the shipped conversation.input.model seat at priority -1, below the shipped ModelSelect (priority 0, from @deepseek-ai/dsh-client-ui-model-selection). The slots engine renders the lowest priority, so the router's merged control takes over the composer model cell and the composer shows a single control instead of a second chip beside the official picker.

That one control keeps both halves. The official half is 本会话模型 / Session model and 推理等级 / Reasoning effort, driven by the real ModelDirectory through props.directory.store subscribe, props.load() and props.select({ provider, model, reasoningEffort }): provider/model groups, the current row's check, the pending spinner, the provider-default row, catalog loading / error / retry, and the official "model · effort" trigger label. Below it sit the router's own groups: the 接管 status, 控制范围 (模型+思考 / 思考 / 模型) and 任务类型, then the model list itself — the pool is folded into the model rows: every session-model row carries a 模型池 / Model pool chip on its right, tapping it pools or unpools that model, a pooled row expands its 定位 / 价格 / 视觉 controls underneath, and pool entries whose provider group is not in the directory gather under 其他模型 / Other models with the manual "type a model id" row. Then 任务预设 and 更多. Every mutation goes to the host over same-origin /model-router/* routes, and the per-call route row in the conversation is unchanged.

The old conversation.input.right chip (model-router:composer-control, order 20) is still registered, but it self-retracts — it disposes itself as soon as the seat registration lands — so a normal host renders an empty .right column and only ever sees the merged cell. A host that rejects or renames the seat keeps the chip as its fallback.

0.10.0 tried this first and was rolled back in 0.10.1: the seat rendered blank and a restart failed to boot. The root cause: cordis resolves nested services such as remote.session by walking fibers upward from the calling fiber, so the client plugin itself must declare sessions, remote and remote.session before a method may touch ModelDirectory.directoryFor(). The plugin-level inject is now ["slots", "uiConversation", "sessions", "remote", "remote.session"].

Limits. This is a registration into the shipped seat, not a patch of the official component: the official popup's internals (portal menu, search box, keyboard navigation, module CSS) are not replicated, only its selection semantics. A DSH that stops declaring conversation.input.model makes the seat registration fail silently and the small fallback chip takes over. Plugin source changes still need a full dsh web/desktop restart.

The panel matches the host (v0.6.4)

The composer control is styled with the harness's own tokens — the same 34px rows, hover wash, glyph checks, segmented controls and popover elevation the built-in menus use. Open docs/ui-preview.html in a browser (or see docs/preview-dark.png / docs/preview-light.png) to look at it without installing anything.

Everything is a click (v0.6.2)

The composer control is the whole configuration surface — no JSON or YAML:

Section	What you do
mode	click 模型+思考 / 思考 / 模型; in 思考, pick the one model the router may tune; a manual pick adds a resume button
task type	pick auto (keyword rules) or one of your presets
model pool	tick models from the live catalog, then set tier / cost / vision with dropdowns
task presets	+ 新建预设 → click a keyword pack (写作 / 代码 / 翻译 / 分析) or add a word → choose each pooled model's weight
more	image steps, task detection (rules / semantic), context pressure

Everything persists in <profile>/.model-router/state.json through the same-origin /model-router/* routes, so the profile row config stays optional.

Every call shows its route (v0.8.0)

Every tool call in the conversation now carries one compact line with the model and thinking level that actually ran that segment:

deepseek-official/deepseek-v4-pro · high

The host folds the session's own events into the modelRouterRoute session projection (stateVersion 1): request/header gives the provider, model and effort in force, step/start gives the turn and step, and tool/call records that pair under the call id. The effort is null when the adapter owns the default, and those rows show the localized 默认 / default label. The projection keeps the last 200 calls per session and drops the oldest.

The client half registers a model-router-route conversation Definition on every tool/call and renders the row through the conversation.chat.node slot, anchored right after the call (+0.1). It reads the projection defensively: a call id the host has not folded yet renders nothing rather than borrowing another call's route. The line is a separate compact row next to the tool call — see Limitations for why it is not a badge inside the tool row.

Images, session signals, subagents and the classifier (v0.6.0)

Images. imagePolicy: vision sends a task that carries an image to a vision model: the pool entry tagged vision, or visionModel when set explicitly. The image is task-scoped — the next command without one goes back to the normal tier.

Session signals. The router reads core session projections and lets them sharpen the choice:

Signal	Effect
active todo (todos)	feeds task-type matching — a todo list states the work better than the prompt
context pressure	above signals.contextPressure (0–1) the cheap tier is preferred
token total	above signals.sessionTokens (0 = off) the cheap tier is preferred
delegationDepth	subagents prefer the cheap tier (subagent.preferCheap)

Every decision records the signals it saw, so /router and the composer control can explain it.

/router prints the scope, engagement, pool, task type, the recent decisions and the task's strong-tier spend.

Classifier. classifier: llm makes one small call per turn to pick a preset (classifierModel, effort off, classifierTimeoutMs), caches it for the turn and falls back to the keyword rules on any failure. The default stays rules.

The five rules

1. No ratchet. Turn depth contributes no score by default (scoring.turnPerPoint: 0) and tool calls are counted per task, so a long agent run does not drift toward the most expensive effort. A long run is not a harder task. Since 0.9.0 a tool loop is a score signal only — it no longer pins the class at engineering (scoring.toolCallClass).
2. Escalation needs evidence. Inside the current task, escalateOnErrors failing tool results, or the same tool call retried with identical arguments escalateOnRepeats times, step the class up — at most maxEscalations times. Nothing else moves it.
3. max is opt-in. Automatic routing never emits max unless allowMax is set. The ceiling is enforced at the effort level, so even a hand-written route table asking for max is clamped.
4. A manually selected max is demoted too (demoteManualMax), but only for models this plugin manages. Other models are left alone. Enabling allowMax turns both clamps off.
5. Thinking stays on. off is not an automatic route. DeepSeek's Messages API rejects a thinking-enabled request whose history contains an assistant tool call produced while thinking was disabled (The content[].thinking in the thinking mode must be passed back to the API). A cheap step can still call a tool, and the next step of the same task is an engineering tool loop, so one off step followed by a high step failed the turn. trivial now routes to low, and any configured or manually selected off is raised to low unless allowThinkingOff is set. Recovery: if a conversation already contains such a turn (from an earlier version, or from a hand-selected off), the router detects it and keeps that session at off instead of failing every step, and logs one warning. Compact the session or start a new one to get thinking back.

Task boundaries come from agent/inbox/claimed, which is what actually opens a new piece of work. When a task ends on an unresolved failure, the next one inherits a single hesitant step up — and only if it is engineering or hard work. A one-line "translate this" never inherits a crash.

Tool loops no longer pin the effort at high (v0.9.0)

Before 0.9.0, classifyStep classified a step as engineering whenever its task had already seen one tool call (cue agent tool loop), and engineering routes to high. Because the per-task tool counter only resets when a new command is claimed, every step from the first tool call to the end of the Agent run stayed at high. There was no way back down mid-loop: trivial needs a cheap cue and at most 60 tokens, and standard required zero tool calls. And with allowMax: false (the default), max is demoted to high, so hard and engineering applied the same effort — evidence escalation changed nothing for a task that was already engineering.

Measured before: 帮我把这个文件里的日志改成中文 ran standard/low at 0 tools, then engineering/high at 1, 3, 6, 12 and 25 tools for the rest of the task; 翻译一下这句话 with one tool also went to high.

scoring.toolCallClass (default standard) decides whether a tool loop can earn the engineering class on its own. With standard, tool activity still adds score (toolCallBase = 5, toolCallAt = 3, toolCallBig = 8) but the brief — strong cues or structural text — decides the class, and escalateOnErrors / escalateOnRepeats still escalate on evidence. Set it to engineering to restore the old rule; any other value falls back to standard. Measured after: a casual brief with tools stays standard/low throughout, an engineering brief stays engineering/high throughout. hard still collapses to high unless allowMax: true, and routes.<class>.effort (for example routes.engineering.effort: low) is the direct way to say how expensive a class may get.

Configuration

The plugin reads its config from its row in the profile's cordis.patch.yml:

- id: model-router
  config:
    mode: auto            # auto | off
    model: deepseek-flash
    allowMax: false       # true enables max for auto routing and manual selection alike
    maxFallback: high     # where max collapses when allowMax is false
    allowThinkingOff: false # true re-enables effort "off" (see rule 5: it breaks tool loops)
    control: full         # full | effort | model — what the router may choose
    manualOverride:
      yieldOnManual: true        # a manual pick stands; routing stands down
      resumeOnNextCommand: true  # …and re-engages on your next command
    pool: []              # whitelist; empty keeps the single `model` above
    presets: {}           # task type -> { match: [...], weights: { modelId: 0-100 } }
    scoring:
      costPenalty: 0.4    # how strongly relative cost subtracts from a pool score
      turnPerPoint: 0     # raise this to let long sessions weigh more (not recommended)
      toolCallClass: standard  # class a tool loop earns: standard (0.9.0) | engineering (pre-0.9.0)
    hysteresis:
      downAfter: 2        # quiet steps required before the effort steps down
    imagePolicy: keep     # keep | vision — route image tasks to a vision model
    visionModel: null     # explicit vision model; the pool's `vision` tag is used otherwise
    signals:
      todos: true         # let the active todo feed task-type matching
      contextPressure: 0.75  # above this occupancy, prefer the cheap tier (0 disables)
      sessionTokens: 0    # above this token total, prefer the cheap tier (0 disables)
    subagent:
      preferCheap: true   # delegated work stays cheap unless a preset says otherwise
    classifier: rules     # rules | llm — llm makes one small call per turn
    classifierModel: null # model used by the llm classifier (defaults to the first pool entry)
    escalateOnErrors: 2   # failed tool results needed to step up
    escalateOnRepeats: 3  # identical retries needed to step up
    routes:
      trivial:     { effort: low }   # "off" here is raised to low unless allowThinkingOff
      standard:    { effort: low }
      engineering: { effort: high }
      hard:        { effort: max }

DSH_MODEL_ROUTER=off|auto overrides mode at boot.

Key	Default	Meaning
mode	auto	off passes every request through, except the manual-max clamp
provider	deepseek-official	only this provider is ever touched
model	deepseek-flash	the one model this plugin drives
familyPattern	^deepseek-(flash|v4)	conversation models it may take over; a pro session is pulled back to flash
allowMax	false	whether max is reachable at all
maxFallback	high	what max collapses to
demoteManualMax	true	also demote a manually selected max
allowThinkingOff	false	allow effort off again; only for sessions that stay non-thinking
control	full	what the router owns: full, effort or model
manualOverride.yieldOnManual	true	a manual pick in the scope that owns it makes the router stand down
manualOverride.resumeOnNextCommand	true	the next command re-engages it
pool	[]	provider/model whitelist with cost/tier/tags/weights/maxPerTask
presets	{}	task types with match rules and per-model weights
scoring.costPenalty	0.4	relative-cost weight in the pool score
scoring.toolCallClass	standard	class a tool loop earns on its own: standard (0.9.0) or engineering (pre-0.9.0 rule); any other value falls back to standard
hysteresis.downAfter	2	quiet steps before a downgrade applies
classifier	rules	rules or llm (one small call per turn, rules as fallback)
classifierModel	first pool entry	model the LLM classifier uses
imagePolicy	keep	vision routes a task carrying an image to a vision model
visionModel	null	explicit vision model id
signals.todos	true	let the active todo feed task-type matching
signals.contextPressure	0.75	occupancy above which the cheap tier is preferred
signals.sessionTokens	0	token total above which the cheap tier is preferred
subagent.preferCheap	true	delegated work prefers the cheap tier
leaveImageSteps	true	steps carrying images keep the caller's model
imagePolicy	keep	set to flash to route image steps too (the flash model has native vision)
escalateOnErrors	2	failure-evidence threshold
escalateOnRepeats	3	repeated-call threshold
maxEscalations	2	most classes a single task may climb
carryUnresolved	true	carry one step up from an unresolved failure

Install

dsh plugin is a pnpm forwarder, so any pnpm specifier works. From the npm registry:

dsh plugin --profile web add @neptune810/dsh-model-router

or straight from the repository:

dsh plugin --profile web add github:Neptune810/dsh-model-router

Restart dsh web afterwards. Because the router registers listeners at boot, a reload is not enough.

Requirements

• Node 20 or newer.
• Verified against @deepseek-ai/dsh 0.1.7-rc.2. The plugin uses agent/request, agent/inbox/claimed, and session.deriveMessages(). No engines.dsh range is declared, so the plugin market keeps this entry visible rather than guessing it incompatible.

Limitations

• No config schema. Settings are read from the profile patch layer shown above and do not render as a form in the settings UI. This is deliberate: a schema would require importing @deepseek-ai/* packages, which a plugin installed beside the profile cannot resolve.
• The router only touches the deepseek-official provider and models matching familyPattern.
• The router control takes over the official model cell by registration, not by patching the component. It registers into conversation.input.model at priority -1 (below the shipped ModelSelect at 0), keeping the official model + effort sections inside its own panel. The old conversation.input.right chip (model-router:composer-control, order 20) self-retracts and is only a fallback for a host that rejects the seat. 0.10.0's first attempt shipped a blank cell and was rolled back in 0.10.1 because it had not declared sessions / remote / remote.session on the plugin root.
• The per-call route line is a separate small row anchored after the tool call, not a badge inside the harness's own tool row or inside the thinking chain: a keyed slot replaces whatever it targets, and the thinking chain exposes no child slot. A call the host has not folded yet shows nothing rather than guessing. Plugin source changes only appear after you restart dsh web.

Tests

node --test

131 tests. test/policy.test.js (34) covers classification, the absence of a ratchet, the unreachable max, the refusal of off, evidence escalation, effort clamping, the poisoned-history detector, tool-result error parsing, and the toolCallClass knob (default standard, engineering restores the pre-0.9.0 rule, any other value coerces to standard); test/routing.test.js (12) covers the pool, preset weights, vision filtering, maxPerTask, effort-vocabulary mapping and hysteresis; test/plugin.test.js (16) drives the host wiring with ctx/agent doubles, including a quiet tool loop that stays at its brief's class instead of climbing to high; test/modes.test.js (21) covers the three scopes, manual yield + resume, same-origin route guards (including a request whose Origin the Desktop proxy stripped, and one carrying Origin: null), pool/preset editing, a pinned task type, a third-party effort vocabulary, vision routing, subagent frugality, todo-driven task types, context pressure, the /router command and the LLM classifier; test/effort-model.test.js (10) covers effort scope bound to one model (unbound stands off, binding clears the pending flag, only the bound model's effort moves, the vocabulary comes from its own pool entry, a session-less binding becomes the default) and the scope-aware manual yield (a model switch in effort scope is not an override, an effort change is, model scope ignores effort, full scope reacts to both) plus the route's clearing and validation; test/route-projection.test.js (12) covers the projection contract (key/stateVersion, plain-JSON state, header folding with absent vs explicit null effort, turn/step tracking, the header in force at a call, the fallback to the call's own values, repeated call ids, the 200-entry cap in insertion order, integer-like call-id ordering, purity and a JSON round-trip) plus the host registration and a host without the service; test/client.test.js (26) loads the shipped browser bundle in a VM and asserts the registration (the route row, the composer chip as a self-retracting fallback, and the merged seat on conversation.input.model at priority -1), the closed trigger, the documented host routes, the route Definition's match/anchor and its defensive read of the folded projection, the localized default label, the seat face delegating to the session directory, the official selection semantics (model rows without reasoningEffort, effort rows with it, the provider default, the current-row no-op, a locked session, an unavailable subagent cell, a failed pick keeping its message, catalog errors offering a retry, and group order), the router half's control/pool/settings payloads and the /effort-model binding, the pool control that collapses the list and reopens it, and the declared ui-chat seat and uiConversation service.

License

MIT