llm

type AnthropicProvider struct {
	// contains filtered or unexported fields
}

type ChatGPTConfig struct {
	// Name shows up in logs and Provider.Name(). Defaults to "chatgpt".
	Name string
	// BaseURL e.g. "https://chatgpt.com/backend-api/codex". Path "/responses"
	// is appended automatically.
	BaseURL string
	// ClientID is the OAuth client id used to refresh tokens on 401.
	ClientID string
	// TokenEndpoint is the OAuth token URL for refresh.
	TokenEndpoint string
	// AccountIDHeader is the HTTP header that carries the per-account id
	// (e.g. "chatgpt-account-id"). Empty disables injection.
	AccountIDHeader string
	// HTTPClient is overridable for tests. Defaults to a long-timeout client.
	HTTPClient *http.Client
	// ExtraHeaders are merged into every request.
	ExtraHeaders map[string]string
}

type ChatGPTProvider struct {
	// contains filtered or unexported fields
}

type ClaudeConfig struct {
	// Name shows up in logs and Provider.Name(). Defaults to "claude".
	Name string
	// BaseURL e.g. "https://api.anthropic.com/v1". The "/messages" path is
	// appended automatically.
	BaseURL string
	// EnvKey is the API-key env var (e.g. ANTHROPIC_API_KEY).
	EnvKey string
	// HTTPClient is overridable for tests. Defaults to a long-timeout client.
	HTTPClient *http.Client
	// ExtraHeaders are merged into every request.
	ExtraHeaders map[string]string
}

type ClaudeProvider struct {
	// contains filtered or unexported fields
}

type Event struct {
	Type       EventType
	Delta      string         // for EventTextDelta
	ToolUse    *types.ToolUse // for EventToolUse
	StopReason string         // for EventDone
	Err        error          // for EventError
	Usage      *Usage         // optional, on EventDone
}

type EventType string

type GeminiConfig struct {
	BaseURL    string
	APIKey     string
	HTTPClient *http.Client
}

type GeminiProvider struct {
	// contains filtered or unexported fields
}

type Model struct {
	ID            string
	Name          string
	ContextLength int
	Pricing       string
}

type OpenAICompatConfig struct {
	// Name shows up in logs and Provider.Name(). Defaults to "openai-compat".
	Name string
	// BaseURL e.g. "https://openrouter.ai/api/v1". The "/chat/completions"
	// suffix is appended automatically.
	BaseURL string
	// EnvKey is the environment variable holding the bearer token. Optional —
	// when blank (local servers like Ollama), no Authorization header is sent.
	EnvKey string
	// HTTPClient is overridable for tests. Defaults to a long-timeout client.
	HTTPClient *http.Client
	// ExtraHeaders are merged into every request. Useful for OpenRouter's
	// HTTP-Referer + X-Title attribution headers.
	ExtraHeaders map[string]string
	// StallTimeout overrides the per-chunk SSE inactivity timer. 0 falls back
	// to the package default (streamStallTimeout). Negative disables the
	// watchdog (test-only).
	StallTimeout time.Duration
	// ChatTemplateKwargs flows into the MLX/vllm `chat_template_kwargs` body
	// field. Lets local servers flip Qwen3 template switches that change tool
	// emission shape. Nil = omit. Per-provider (TOML `chat_template_kwargs`).
	ChatTemplateKwargs map[string]any
	// ReportsCost opts the provider into a request flag asking for actual spend
	// in the usage block. Off by default; strict endpoints would reject the
	// extra field, so only set it for services that return real per-call cost.
	ReportsCost bool
}

type OpenAICompatProvider struct {
	// contains filtered or unexported fields
}

type Provider interface {
	Name() string
	Stream(ctx context.Context, req Request) (<-chan Event, error)
}

type Request struct {
	Model       string
	System      string
	Messages    []types.Message
	Tools       []ToolSpec
	MaxTokens   int
	Temperature float64
	// TopP pins nucleus sampling. 0 = use provider default (omit on wire).
	// Set by the active profile (tiny pins 0.8 for 4-bit MoE).
	TopP float64
	// Stop is the optional list of stop sequences. textmode wraps tool calls
	// in `<bash>{…}</bash>`-style envelopes; passing the closing tag halts
	// decode immediately after one tool call, saving 50–300 tokens per turn.
	Stop   []string
	Stream bool
	// Thinking selects the extended-reasoning budget for providers that
	// support it. Off means omit the field entirely.
	Thinking Thinking
	// ChatTemplateKwargs overrides/merges with the provider's static
	// chat_template_kwargs for this request. Used to flip Qwen3's
	// enable_thinking switch per-turn from the effort level. Nil = no override.
	ChatTemplateKwargs map[string]any
}

type TextModeOptions struct {
	// ExtraHint is appended verbatim after the auto-generated tool block.
	// Useful for caveman-style brevity nudges.
	ExtraHint string
}

type TextModeProvider struct {
	// contains filtered or unexported fields
}

type Thinking string

type ToolSpec struct {
	Name          string         `json:"name"`
	Description   string         `json:"description"`
	PromptSnippet string         `json:"-"`
	Schema        map[string]any `json:"schema"`
}

type Usage struct {
	InputTokens  int
	OutputTokens int
	// CachedTokens is the cached prompt-token count, when the provider reports
	// it. Zero otherwise — most providers omit it.
	CachedTokens int
	// CostUSD is the provider-reported actual spend for this turn, when the
	// service returns it (some routed aggregators do). Zero when unreported;
	// callers then fall back to a static price-table estimate.
	CostUSD float64
}