provider

type APIStyle string

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

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

type AttemptErrorContract struct {
	Stage        types.ProviderErrorStage
	Class        types.ProviderErrorClass
	ReplaySafety types.ProviderReplaySafety
}

type AttemptLimitError struct {
	MaxAttempts int
	Attempts    int
	Cause       error
}

type BedrockConfig struct {
	// Region is the AWS region (e.g. "us-east-1"). Defaults to AWS_REGION / AWS_DEFAULT_REGION env.
	Region string

	// Model is the Bedrock model ID (e.g. "anthropic.claude-3-5-sonnet-20241022-v2:0").
	// Supports foundation models, cross-region inference profiles (e.g. "us.anthropic.claude-..."),
	// and ARN-format inference profiles.
	Model string

	// MaxTokens overrides the default max output tokens.
	MaxTokens int

	// Timeout in seconds for requests.
	Timeout int

	// StaticCredentials, if set, skips the default credential chain.
	// Populated from AWS_ACCESS_KEY_ID + AWS_SECRET_ACCESS_KEY + AWS_SESSION_TOKEN env vars.
	AccessKeyID     string
	SecretAccessKey string
	SessionToken    string

	// BearerToken is set from AWS_BEARER_TOKEN_BEDROCK for API-key-style auth.
	BearerToken string

	// BaseURL overrides the Bedrock runtime endpoint (ANTHROPIC_BEDROCK_BASE_URL).
	// Must be empty, start with "https://", or start with "http://localhost" (for testing).
	BaseURL string
}

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

type CacheRoutingMode string

type CacheRoutingPreference string

type CapabilityProvider interface {
	Capabilities() ProviderCapabilities
}

type CapabilitySupport uint8

type CloseProvider interface {
	Close() error
}

type CompatibleModelRequest struct {
	Provider string
	APIStyle APIStyle
	BaseURL  string
	APIKey   string
}

type CompatibleProviderDefinition struct {
	Name        string
	DisplayName string
	Popularity  int
	BaseURLs    map[APIStyle]string
	UserDefined bool
}

type Config struct {
	ProviderName string
	APIStyle     APIStyle

	// APIFormat is an explicit OpenAI-family wire override. Empty selects the
	// native provider's cataloged format; compatible providers may negotiate a
	// catalog-inferred format. Supported values are responses and chat-completions.
	APIFormat string
	// ResponsesSemantics declares the upstream wire contract independently of
	// the transport URL. This is required when the official OpenAI API is
	// reached through a content-blind benchmark proxy: changing only URL/auth
	// must not silently downgrade the request to generic-compatible semantics.
	ResponsesSemantics ResponsesSemantics
	// ResponsesWebSocket is an explicit endpoint capability assertion. Unknown
	// and Unsupported keep the documented HTTP/SSE transport. Supported enables
	// the production WebSocket transport only for the public Responses profile;
	// it is never inferred from a hostname or model name.
	ResponsesWebSocket        CapabilitySupport
	APIKey                    string
	AuthToken                 string
	BaseURL                   string
	Model                     string
	MaxTokens                 int
	Headers                   map[string]string
	Timeout                   int // seconds
	DisableStrictTools        bool
	DisablePromptCacheOptions bool
	CacheRoutingPreference    CacheRoutingPreference
}

type ConnectionDetail struct {
	Provider        string
	DisplayName     string
	State           ConnectionState
	Kind            ConnectionKind
	Source          string
	DetailKey       i18n.Key
	DetailArgs      []any
	CanSelectModels bool
	CanConnect      bool
	SetupHintKey    i18n.Key
	SetupHintArgs   []any
}

type ConnectionKind string

type ConnectionState string

type CredentialEntry struct {
	// Provider is the canonical provider name (e.g. "anthropic", "openai").
	Provider string `json:"provider"`

	// AuthMethod describes how the credential was obtained.
	// Values: "api_key", "oauth"
	AuthMethod string `json:"auth_method"`

	// APIKey is the provider API key, or an optional exchanged API key for OAuth entries.
	APIKey string `json:"api_key,omitempty"`

	// AccessToken is the OAuth access token (for "oauth" auth method).
	AccessToken string `json:"access_token,omitempty"`

	// RefreshToken is the OAuth refresh token (for "oauth" auth method).
	RefreshToken string `json:"refresh_token,omitempty"`

	// ExpiresAt is the expiry time for OAuth tokens.
	ExpiresAt time.Time `json:"expires_at,omitempty"`

	// AccountID is the ChatGPT workspace/account identifier used by the Codex backend.
	AccountID string `json:"account_id,omitempty"`

	// ChatGPTPlanType is informational metadata parsed from the OpenAI ID token.
	ChatGPTPlanType string `json:"chatgpt_plan_type,omitempty"`

	// ChatGPTAccountIsFedRAMP routes workspace traffic through the FedRAMP edge.
	ChatGPTAccountIsFedRAMP bool `json:"chatgpt_account_is_fedramp,omitempty"`

	// BaseURL is an optional custom API base URL.
	BaseURL string `json:"base_url,omitempty"`

	// APIStyle selects the compatibility protocol used for model discovery and
	// inference. Empty entries predate compatible aggregate providers.
	APIStyle APIStyle `json:"api_style,omitempty"`
	// APIFormat is an explicit OpenAI-family wire override. It is deliberately
	// separate from Models[].APIFormat, which is discovered catalog metadata.
	APIFormat string `json:"api_format,omitempty"`
	// DisableStrictTools is an explicit compatibility escape hatch for an
	// OpenAI Responses endpoint that is reached through a gateway which rejects
	// strict function schemas. The endpoint identity remains OpenAI; only schema
	// strictness is relaxed.
	DisableStrictTools bool `json:"disable_strict_tools,omitempty"`
	// DisablePromptCacheOptions preserves prompt_cache_key routing while
	// omitting public-API cache option and breakpoint extensions rejected by
	// some content-blind gateways.
	DisablePromptCacheOptions bool `json:"disable_prompt_cache_options,omitempty"`

	// DisplayName and UserDefined persist providers created through the generic
	// gateway flow without requiring a second configuration file.
	DisplayName string `json:"display_name,omitempty"`
	UserDefined bool   `json:"user_defined,omitempty"`

	// Models caches the most recently discovered catalog so a configured
	// provider remains selectable before its next online refresh.
	Models []ModelInfo `json:"models,omitempty"`

	// LastUsed records when this credential was last used.
	LastUsed time.Time `json:"last_used"`
}

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

type DebugCallKind string

type DebugEvent struct {
	ID       uint64         `json:"id"`
	Phase    DebugPhase     `json:"phase"`
	Kind     DebugCallKind  `json:"kind"`
	Provider string         `json:"provider"`
	Model    string         `json:"model"`
	Metadata map[string]any `json:"metadata,omitempty"`
	Request  *DebugRequest  `json:"request,omitempty"`
	Response *DebugResponse `json:"response,omitempty"`
}

type DebugObserver func(DebugEvent)

type DebugPhase string

type DebugRequest struct {
	Stream                   bool                         `json:"stream"`
	Model                    string                       `json:"model"`
	MaxTokens                int                          `json:"max_tokens"`
	MaxOutputTokensOverride  int                          `json:"max_output_tokens_override,omitempty"`
	EffectiveMaxOutputTokens int                          `json:"effective_max_output_tokens,omitempty"`
	CatalogMaxOutputTokens   int                          `json:"catalog_max_output_tokens,omitempty"`
	System                   string                       `json:"system,omitempty"`
	SystemBlocks             []prompt.SystemPromptBlock   `json:"system_blocks,omitempty"`
	Messages                 []types.Message              `json:"messages"`
	Tools                    []types.ToolDefinition       `json:"tools,omitempty"`
	ExtraToolSchemas         []types.ServerToolDefinition `json:"extra_tool_schemas,omitempty"`
	ToolChoice               *ToolChoice                  `json:"tool_choice,omitempty"`
	Thinking                 *ThinkingConfig              `json:"thinking,omitempty"`
	TaskBudget               *TaskBudget                  `json:"task_budget,omitempty"`
	Conversation             string                       `json:"conversation,omitempty"`
	PreviousResponseID       string                       `json:"previous_response_id,omitempty"`
	Truncation               string                       `json:"truncation,omitempty"`
	PromptCacheKey           string                       `json:"prompt_cache_key,omitempty"`
	ReasoningEffort          string                       `json:"reasoning_effort,omitempty"`
	ServiceTier              ServiceTier                  `json:"service_tier,omitempty"`
	UsePromptCache           bool                         `json:"use_prompt_cache,omitempty"`
}

type DebugResponse struct {
	Message           *types.Message                   `json:"message,omitempty"`
	Usage             *types.Usage                     `json:"usage,omitempty"`
	StopReason        *types.StopReason                `json:"stop_reason,omitempty"`
	ResponseID        string                           `json:"response_id,omitempty"`
	SystemFingerprint string                           `json:"system_fingerprint,omitempty"`
	Error             string                           `json:"error,omitempty"`
	FailureDiagnostic *types.ProviderFailureDiagnostic `json:"failure_diagnostic,omitempty"`
}

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

type FallbackTriggeredError struct {
	OriginalModel string
	FallbackModel string
	Cause         error
}

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

type ModelInfo struct {
	// ID is the canonical model identifier (e.g. "claude-sonnet-4-20250514").
	ID string

	// Aliases lists current provider identifiers that resolve to ID without
	// being shown as duplicate entries in model pickers.
	Aliases []string

	// Name is the human-readable display name (e.g. "Claude Sonnet 4").
	Name string

	// Provider is the provider that owns this model (e.g. "anthropic").
	Provider string

	// ContextWindow is the maximum input context window in tokens.
	ContextWindow int

	// MaxOutput is the maximum output tokens the model can generate.
	// 0 means unknown / provider-default.
	MaxOutput int

	// CostPer1MIn is the cost per 1M input tokens in the model's native billing currency.
	CostPer1MIn float64

	// CostPer1MOut is the cost per 1M output tokens in the model's native billing currency.
	CostPer1MOut float64

	// CacheReadPer1M is the cost per 1M cache-read tokens in the model's native billing currency. 0 = no caching.
	CacheReadPer1M float64

	// CacheCreatePer1M is the cost per 1M cache-creation tokens in the model's native billing currency. 0 = no caching.
	CacheCreatePer1M float64

	// CostCurrency is the ISO 4217 billing currency for all Cost* fields.
	CostCurrency string

	// CanReason indicates the model supports extended thinking / reasoning.
	CanReason bool

	// ReasoningEfforts lists selectable reasoning effort tiers for providers
	// whose API supports a discrete effort parameter. Empty means no secondary
	// effort picker should be shown for this model.
	ReasoningEfforts []string

	// DefaultReasoningEffort is the provider-documented default selectable
	// effort. Empty retains the shared catalog fallback.
	DefaultReasoningEffort string

	// CanUseTools indicates the model supports function/tool calling.
	CanUseTools bool

	// CanSeeImages indicates the model supports vision / image inputs.
	CanSeeImages bool

	// CacheControl indicates the model supports prompt caching.
	CacheControl bool

	// APIFormat is the API protocol: "messages", "chat-completions", "responses".
	APIFormat string

	// IsDefault is true if this is the default model for its provider.
	IsDefault bool

	// ContextOverridden is true when ContextWindow was changed by user settings.
	ContextOverridden bool
}

type ModelOverride struct {
	ContextWindow *int `json:"context_window,omitempty"`
	MaxOutput     *int `json:"max_output,omitempty"`
}

type ModelOverrides map[string]ModelOverride

type OAuthHook interface {
	// LoadAccessToken returns a valid access token, refreshing if needed.
	// Returns ("", nil) if no credentials are stored.
	LoadAccessToken(ctx context.Context) (string, error)

	// OnAuthError is called when a 401 is received. Returns true if
	// credentials were successfully refreshed and a retry is warranted.
	OnAuthError() bool
}

type OpenAIDialect string

type OpenAIOAuthResult struct {
	APIKey              string
	APIKeyExchangeError string
	AccessToken         string
	RefreshToken        string
	ExpiresAt           time.Time
	AccountID           string
	PlanType            string
	AccountIsFedRAMP    bool
}

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

type Params struct {
	Model     string
	MaxTokens int
	// MaxOutputTokensOverride temporarily overrides MaxTokens for output-limit
	// recovery paths. Zero preserves the provider's normal MaxTokens behavior.
	MaxOutputTokensOverride int
	System                  string                     // single system prompt block
	SystemBlocks            []prompt.SystemPromptBlock // ordered system prompt blocks with optional metadata
	Messages                []types.Message
	Tools                   []types.ToolDefinition
	ExtraToolSchemas        []types.ServerToolDefinition
	ToolChoice              *ToolChoice     // nil = auto (default)
	Thinking                *ThinkingConfig // nil = thinking disabled
	TaskBudget              *TaskBudget     // nil = API-side task budget disabled

	// Responses API specific fields
	Conversation       string // Session ID for conversation tracking
	PreviousResponseID string // Link to previous response for context continuity
	Truncation         string // "auto", "disabled", "required_or_error"

	// Prompt cache routing fields shared by Responses and compatible Chat APIs.
	PromptCacheKey string // Stable session lineage lowered to a user-level routing shard by capable providers
	UsePromptCache bool   // Mirror Codex CLI: keep prompt_cache_key enabled even after previous_response_id fallback
	// PromptCacheTTL is provider-selected. Empty preserves the provider's
	// default; "1h" requests the longest documented Anthropic-compatible TTL.
	PromptCacheTTL  string
	ReasoningEffort string // "low", "medium", "high" for reasoning models
	ServiceTier     ServiceTier

	// ContinuationLineage/Epoch are content-free stateless replay evidence.
	// A higher epoch with ContinuationReset proves an intentional context
	// replacement (for example compaction) without exposing message content.
	ContinuationLineage string
	ContinuationEpoch   uint64
	ContinuationReset   bool
	// contains filtered or unexported fields
}

type Provider interface {
	// Name returns the provider identifier
	Name() string

	// CreateStream sends a streaming message request and returns events via channel
	CreateStream(ctx context.Context, params Params) (<-chan types.StreamEvent, error)

	// ModelID returns the default model for this provider
	ModelID() string
}

type ProviderCapabilities struct {
	Thinking           bool
	ToolUse            bool
	CustomTools        CapabilitySupport
	ResponsesWebSocket CapabilitySupport
	ServiceTier        CapabilitySupport
	CacheControl       bool
	CacheRouting       CacheRoutingMode
	SystemParts        bool
	Vision             bool
	MaxContext         int // token limit; 0 = unknown
}

type ProviderFactory func(cfg Config, modelOverride string) (Provider, error)

type ProviderInfo struct {
	// Name is the registered provider identifier (e.g. "anthropic", "openai").
	Name string

	// DisplayName is the human-readable name (e.g. "Anthropic", "OpenAI").
	DisplayName string

	// EnvKey is the primary environment variable for authentication (e.g. "ANTHROPIC_API_KEY").
	// Empty for providers that don't use an API key (e.g. "ollama").
	EnvKey string

	// Models lists the canonical model IDs available for this provider.
	Models []string

	// DefaultModel is the model used when none is specified.
	DefaultModel string

	// AuthMethods describes supported authentication methods.
	// Values: "api_key", "oauth_pkce", "device_code", "aws_credentials", "gcp_adc"
	AuthMethods []string

	// Hidden marks internal providers omitted from user-facing pickers.
	Hidden bool

	// Popularity is a sorting weight (higher = more popular, shown first).
	Popularity int

	// RequiresContext indicates the factory needs a context.Context (e.g. Bedrock, Vertex).
	RequiresContext bool

	// DefaultBaseURL is the default API base URL for this provider.
	// Empty means the provider does not have a well-known default or uses
	// the provider-specific SDK default.
	DefaultBaseURL string

	// APIStyles lists the wire protocols offered by a compatible aggregate
	// provider. Empty means the provider uses its native, fixed protocol.
	APIStyles []APIStyle

	// DefaultBaseURLs maps each compatible protocol to its default endpoint.
	// It is populated for named aggregate providers and empty for user-defined
	// gateways, which must supply their own Base URL.
	DefaultBaseURLs map[APIStyle]string

	// DynamicModels marks providers whose model catalog is discovered from the
	// configured endpoint after authentication.
	DynamicModels bool

	// UserDefined marks providers created through the "other gateway" flow.
	UserDefined bool
}

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

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

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

type ResponsesSemantics string

type RetryConfig struct {
	// MaxAttempts is the total HTTP request-attempt budget used before a
	// response stream is established. The default is five (the initial request
	// plus four retries), matching Codex CLI's request_max_retries default.
	MaxAttempts int
	// StreamMaxAttempts is the independent response-stream budget. The default
	// is six (the initial stream plus five reconnects), matching Codex CLI's
	// stream_max_retries default. Keeping the budgets separate prevents a few
	// connection failures from consuming all stream-recovery capacity.
	StreamMaxAttempts int
	BaseDelay         time.Duration // default 200 ms
	MaxDelay          time.Duration // default 3.2 s for the bounded default budgets
	// OnRetry is called before each retry attempt. nil = no logging.
	// attempt is the one-based failed raw attempt that triggered the retry.
	OnRetry func(attempt, maxRetries int, delay time.Duration, err error)
	// OnAuthError is called when a 401 Unauthorized error is encountered.
	// If it returns true, credentials were refreshed and one retry is allowed.
	// If nil or it returns false, the 401 is surfaced immediately (fail fast).
	OnAuthError func() bool
	// contains filtered or unexported fields
}

type RetryEvent struct {
	Attempt    int
	MaxRetries int
	Delay      time.Duration
	Err        error
	Kind       string
	// DroppedField is an implementation-owned allowlisted request field removed
	// by a compatible gateway fallback. It never contains provider text.
	DroppedField string
}

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

type RuntimeOverrides struct {
	APIFormat          string
	ResponsesWebSocket CapabilitySupport
}

type ServiceTier string

type StreamIdleTimeoutError struct {
	Phase   streamWatchdogPhase
	IdleFor time.Duration
}

type StreamTransportFallback interface {
	TryFallbackTransport() (from, to string, activated bool)
}

type TaskBudget struct {
	Total     int
	Remaining *int
}

type ThinkingConfig struct {
	Enabled      bool
	BudgetTokens int // token budget for thinking; 0 means use model default (1024 minimum required by API)
}

type TokenCountingProvider interface {
	CountTokens(ctx context.Context, content string) (int, error)
}

type ToolChoice struct {
	Type string // "auto", "any", "tool"
	Name string // only used when Type == "tool"
}

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

type VertexConfig struct {
	// ProjectID is the GCP project ID (GOOGLE_CLOUD_PROJECT / ANTHROPIC_VERTEX_PROJECT_ID).
	ProjectID string

	// Region is the GCP region (e.g. "us-east5"). Defaults to GOOGLE_CLOUD_REGION /
	// CLOUD_ML_REGION / ANTHROPIC_VERTEX_REGION env vars, falls back to "us-east5".
	Region string

	// Model is the Vertex model ID (e.g. "claude-sonnet-4-20250514").
	// Vertex uses the same short model names as the direct API (no "anthropic." prefix).
	Model string

	// MaxTokens overrides the default max output tokens.
	MaxTokens int

	// Timeout in seconds for requests.
	Timeout int

	// BaseURL overrides only the Vertex transport endpoint. Authentication and
	// request semantics remain Vertex-native.
	BaseURL string
}

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