dsh-tokenslash

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)
pinnedTools	array of strings	[]	Explicit tool names to preserve regardless of intent or extreme pruning
failOpen	boolean	true	Preserve all tool groups if triage endpoint fails or returns invalid response
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.