dsh-tokenslash
TokenSlash optimization suite for DeepSeek Harness (DSH). Routes subagents to cheaper model tiers, prunes noisy tool output, compacts context, and guards goal budgets via TypeSafe Jev triage decisions.
Features
- System prompt & tool schema pruning —
system-prompt/assemble waterfall strips unused tool schemas and related verbose prompt sections based on intent, saving 3K-8K input tokens per turn.
- Subagent routing —
tokenslash_triage decisions fork work to cheap/medium/smart tiers, saves tokens per decouple.
- Tool output pruning — post-execute hook strips oversized JSON payloads over 10KB to compact summaries.
- Output compacting — prompt-assemble hook summarizes history when context exceeds ~8K tokens.
- Goal guard — warns at 80% of max goal rounds, emits notice via
ctx.app.notice/ctx.notice.
- Provider fan-out — OpenCode Zen, 9Router, OpenRouter, TypeSafe Jev, Custom endpoint with cached model discovery.
- Telemetry — tokens saved, USD estimate, per-module counts, rolling history, JSON persistence via
toJSON/fromJSON.
- Settings UI — host settings section + web client (
lib/client.js, ModelPicker/SettingsCard components).
Installation
Requires Node.js >=20, ESM ("type": "module").
npm install dsh-tokenslash
# or drop into DSH packages dir and restart host
Plugin mounts via cordis.patch.yml (id: tokenslash). Host entry lib/index.js exports apply/inject; browser entry lib/client.js re-exports web slots.
Configuration
| Key | Type | Default | Purpose |
|---|
enabled | boolean | true | Master switch |
provider | string | opencode | Active Jev provider id |
customBaseUrl | string | http://127.0.0.1:20128/v1/systemone | Base URL when provider=custom |
apiKey | string | "" | Bearer token for auth providers |
model | string | jev-1.13-free | Default triage model |
decoupleThreshold | number (0-1) | 0.85 | Confidence cutoff for subagent fork |
modelTiers.cheap | string (CSV) | gemini-2.5-flash, deepseek-chat, gpt-4o-mini | Cheap tier list |
modelTiers.medium | string (CSV) | gemini-2.5-pro, claude-3-5-haiku | Medium tier list |
modelTiers.smart | string (CSV) | deepseek-reasoner, claude-3-7-sonnet, gpt-4o | Smart tier list |
toolPruningMode | string | "normal" | Pruning mode: "normal" (intent-based, keep core tools), "extreme" (prune all tools), "off" (no tool pruning) |
modules.subagentRouting | boolean | true | Enable routing hook + auto-route handler |
modules.toolPruning | boolean | true | Enable tool output prune hook |
modules.promptPruning | boolean | true | Enable system prompt & tool schema pruning before model request |
modules.outputCompacting | boolean | true | Enable compact hook |
modules.goalGuard | boolean | true | Enable goal budget warnings |
Helpers:
parseConfig(raw) — schema-validated config with defaults.parseModelTiers(tiers) — CSV/array → { cheap, medium, smart } arrays.getConfigForProvider(provider, overrides) — resolves baseUrl/model per provider.getProviderConfig(provider, config) / listProviders() / fetchModels(provider, opts) — cached (5 min TTL) model discovery; clearModelsCache() busts cache.
API Endpoints
Base prefix: /api/dsh-tokenslash
| Method | Path | Body / Query | Response |
|---|
| GET | /config | — | Current config JSON |
| POST | /config | SaveSchema (provider, customBaseUrl, apiKey, model, decoupleThreshold, modelTiers, modules, enabled) | { ok, config } |
| POST | /triage | { questions, model? } | Jev triage JSON |
| POST | /describe | { task, context?, model? } | Triage describe result |
| POST | /test | { provider?, customBaseUrl?, apiKey?, model?, timeoutMs? } | { ok, provider, model } or error |
| GET | /models | — | { models, current, provider } |
| GET | /stats | — | Telemetry stats |
| GET | /history?limit=N | limit ≤ 500 | { history } |
Host tool: tokenslash_triage — input { questions: object, model?: string }, calls jevClient.triage.
Architecture
lib/index.js apply()/inject() — bridge routes, settings section, tool, agent hooks
lib/config.js ConfigSchema, parseConfig, parseModelTiers, getConfigForProvider
lib/providers.js getProviderConfig, listProviders, fetchModels (TTL cache)
lib/jev-client.js JevClient — buildUrl/buildHeaders/request(+retry/timeout), triage, testConnection, getModels
lib/telemetry.js TokenslashTelemetry — record/getStats/getHistory/reset/toJSON/fromJSON
lib/routes/bridge.js createBridgeRoutes — config/triage/describe/test/models/stats/history
lib/hooks/routing.js pre-execute: triage route/prune questions, veto on "no"/"false"
lib/hooks/prune.js post-execute: compact payloads >10KB via triage
lib/hooks/compact.js prompt-assemble: summarize when tokens >8K
lib/hooks/goal-guard.js goal hook: notice at ≥80% round budget
lib/client.js browser entry, re-exports client/slots apply/inject + CSS
lib/types.d.ts Config, TriageResult/TriageDecision, TelemetryStats/Entry, ProviderConfig, ModelTiers
Retry/timeout defaults: JevClient 15s timeout, 2 retries; testConnection 8s/1 retry; fetchModels 10s.
Development
node --test test/config.test.js test/providers.test.js test/jev-client.test.js test/telemetry.test.js
node --test test/hooks/
fetch mocked globally per test; no network calls. Types: import from ./lib/types.d.ts.
License
MIT
