Documentation
¶
Index ¶
- Variables
- func ClampThinkingLevel(level string, available []string) (string, bool)
- func CreateModel(prov, name, apiKey, baseURL string, providerExtra map[string]any) (agentcore.ChatModel, error)
- func IsExplicitThinkingLevel(s string) bool
- func IsSupportedType(name string) bool
- func IsValidThinkingLevel(s string) bool
- func NewModelFactory(clientOpts ...litellm.ClientOption) ...
- func ResolveThinkingLevel(model agentcore.ChatModel, level string) (string, bool)
- func StartPricingRefresh(registry *ModelRegistry, cacheDir string)
- func SupportedTypeNames() []string
- func ThinkingLevelsForModel(model agentcore.ChatModel) []string
- func WrapStreamSafe(m agentcore.ChatModel) agentcore.ChatModel
- type ModelEntry
- type ModelRegistry
- func (r *ModelRegistry) AvailableThinkingLevels(modelID string) []string
- func (r *ModelRegistry) CostRates(modelID string) (inputPer1M, outputPer1M, cacheReadPer1M, cacheWritePer1M float64)
- func (r *ModelRegistry) FindByProvider(prov string) []ModelEntry
- func (r *ModelRegistry) List(filter string) []ModelEntry
- func (r *ModelRegistry) MergeModels(fetched []ModelEntry)
- func (r *ModelRegistry) Resolve(pattern string) (*ModelEntry, agentcore.ThinkingLevel, error)
Constants ¶
This section is empty.
Variables ¶
var ThinkingLevelOrder = []string{"off", "low", "medium", "high", "xhigh", "max"}
ThinkingLevelOrder defines the ordered progression of thinking levels.
Functions ¶
func ClampThinkingLevel ¶
ClampThinkingLevel resolves a supported level against the available set. Unsupported or unavailable levels are rejected instead of downgraded.
func CreateModel ¶
func CreateModel(prov, name, apiKey, baseURL string, providerExtra map[string]any) (agentcore.ChatModel, error)
CreateModel creates a ChatModel for the given provider, model name, API key, optional base URL, and provider-level extra config.
func IsExplicitThinkingLevel ¶ added in v0.3.0
func IsSupportedType ¶ added in v0.1.0
IsSupportedType reports whether the given provider type is registered in litellm (built-in or custom). The check is delegated to agentcore/litellm so codebot does not maintain a duplicate whitelist.
func IsValidThinkingLevel ¶ added in v0.0.2
IsValidThinkingLevel reports whether s is a recognized thinking level.
func NewModelFactory ¶ added in v0.2.0
func NewModelFactory(clientOpts ...litellm.ClientOption) func(prov, name, apiKey, baseURL string, providerExtra map[string]any) (agentcore.ChatModel, error)
NewModelFactory returns a model factory that forwards the given litellm ClientOptions (e.g. litellm.WithHook for telemetry) into every model it builds. The return type structurally matches agent.ModelFactory without importing that package, avoiding an import cycle.
func ResolveThinkingLevel ¶ added in v0.3.0
func StartPricingRefresh ¶
func StartPricingRefresh(registry *ModelRegistry, cacheDir string)
StartPricingRefresh launches a background goroutine that refreshes model data from the OpenRouter API. Disk cache (cacheDir/models-cache.json) is checked first with a 24h TTL. Fresh data is merged into the registry, overlaying the compiled-in baseline from models_generated.go.
func SupportedTypeNames ¶ added in v0.1.0
func SupportedTypeNames() []string
SupportedTypeNames returns all provider names known to litellm, sorted.
func ThinkingLevelsForModel ¶ added in v0.3.0
Types ¶
type ModelEntry ¶
type ModelEntry struct {
Provider string `json:"provider"`
ID string `json:"id"`
Name string `json:"name"`
ContextWindow int `json:"context_window"`
MaxTokens int `json:"max_tokens"`
Reasoning bool `json:"reasoning"`
InputCostPer1M float64 `json:"input_cost_per_1m"`
OutputCostPer1M float64 `json:"output_cost_per_1m"`
CacheReadCostPer1M float64 `json:"cache_read_cost_per_1m"`
CacheWriteCostPer1M float64 `json:"cache_write_cost_per_1m"`
}
ModelEntry describes a known LLM model.
type ModelRegistry ¶
type ModelRegistry struct {
// contains filtered or unexported fields
}
ModelRegistry holds known models and provides resolution/cycling.
func NewModelRegistry ¶
func NewModelRegistry() *ModelRegistry
NewModelRegistry creates a registry loaded from generated model data.
func (*ModelRegistry) AvailableThinkingLevels ¶
func (r *ModelRegistry) AvailableThinkingLevels(modelID string) []string
AvailableThinkingLevels returns valid thinking levels for a model. Non-reasoning models only support "off"; reasoning models support all levels.
func (*ModelRegistry) CostRates ¶
func (r *ModelRegistry) CostRates(modelID string) (inputPer1M, outputPer1M, cacheReadPer1M, cacheWritePer1M float64)
CostRates returns the per-1M-token cost rates for a model. Uses Resolve for matching, so partial names like "sonnet" work.
func (*ModelRegistry) FindByProvider ¶
func (r *ModelRegistry) FindByProvider(prov string) []ModelEntry
FindByProvider returns all models for a given provider.
func (*ModelRegistry) List ¶
func (r *ModelRegistry) List(filter string) []ModelEntry
List returns models matching the filter (empty filter = all).
func (*ModelRegistry) MergeModels ¶
func (r *ModelRegistry) MergeModels(fetched []ModelEntry)
MergeModels updates existing models and adds new ones from fetched data. Matching by Provider+ID (case-insensitive). Updates cost/context/name fields.
func (*ModelRegistry) Resolve ¶
func (r *ModelRegistry) Resolve(pattern string) (*ModelEntry, agentcore.ThinkingLevel, error)
Resolve parses a model pattern and returns the matched entry and optional thinking level. Pattern formats:
- "claude-sonnet-4-5" → exact/partial match
- "anthropic/claude-sonnet-4-5" → provider/model
- "claude-sonnet-4-5:high" → model + thinking level
- "sonnet" → partial match
Source Files