helps

package
v8.0.2 Latest Latest
Warning

This package is not in the latest version of its module.

Go to latest
Published: Sep 27, 2026 License: MIT Imports: 76 Imported by: 0

Documentation

Index

Constants

View Source
const (
	ClaudeCodeSessionHeader = "X-Claude-Code-Session-Id"
	ClaudeCodeAgentHeader   = "X-Claude-Code-Agent-Id"
	ClaudeCodeMainAgentID   = "main"
)
View Source
const (
	// ConnectFlagData marks an uncompressed Connect-proto data frame.
	ConnectFlagData byte = 0x00
	// ConnectFlagCompressed marks a gzipped Connect-proto frame.
	ConnectFlagCompressed byte = 0x01
	// ConnectFlagEndStream marks the terminal Connect-proto trailer frame.
	ConnectFlagEndStream byte = 0x02

	// DevinDefaultBaseURL is the default upstream Codeium/Devin endpoint.
	DevinDefaultBaseURL = "https://server.codeium.com"
	// DevinChatPath is the Connect-RPC endpoint for chat completions.
	DevinChatPath = "/exa.api_server_pb.ApiServerService/GetChatMessage"

	// DevinDefaultClientName is the client identifier declared in metadata.
	DevinDefaultClientName = "chisel"
	// DevinDefaultClientVersion is the client version declared in metadata.
	DevinDefaultClientVersion = "3000.10.21"
	// DevinFingerprintHexLen is the mandatory 732-hex-character length for metadata #31.
	DevinFingerprintHexLen = 732

	// DevinDefaultMaxTokens is the fallback max completion tokens.
	DevinDefaultMaxTokens = 128000
)
View Source
const (
	// CodexEmptyIncompleteStreamMessage is the error message emitted when upstream ChatGPT Codex
	// terminates with response.incomplete having 0 output tokens and 0 content output.
	CodexEmptyIncompleteStreamMessage = "stream error: upstream terminated with incomplete empty response (0 tokens)"
)
View Source
const DefaultTransportCacheCapacity = 64

DefaultTransportCacheCapacity bounds how many transports a TransportCache keeps alive at once. Every cached transport owns an independent connection pool, so an unbounded cache would let idle sockets and the goroutines managing them grow without limit whenever keys churn, for example when a credential's proxy is rotated through the management API or when an SDK embedder supplies a freshly built base transport per request.

Variables

View Source
var (
	ClaudeDeviceProfileBeforeCandidateStore func(ClaudeDeviceProfile)
)

Functions

func APIKeyFromContext

func APIKeyFromContext(ctx context.Context) string

func APIKeyModelIsCompat

func APIKeyModelIsCompat(req cliproxyexecutor.Request) bool

APIKeyModelIsCompat reports whether the selected API-key model enables compatibility handling for Claude thinking blocks.

func AllocateClaudeMCPToolAlias

func AllocateClaudeMCPToolAlias(secret, original string, reserved map[string]bool) (string, bool)

AllocateClaudeMCPToolAlias picks an alias that is not already reserved. Attempts are capped at the wordlist size so names that sanitize to the same suffix cannot spin forever. ok is false only when every one-word tool ID for this semantic is already reserved.

func AppendAPIResponseChunk

func AppendAPIResponseChunk(ctx context.Context, cfg *config.Config, chunk []byte)

AppendAPIResponseChunk appends an upstream response chunk to Gin context for request logging.

func AppendAPIWebsocketResponse

func AppendAPIWebsocketResponse(ctx context.Context, cfg *config.Config, payload []byte)

AppendAPIWebsocketResponse stores an upstream websocket response frame in Gin context.

func AppendCodexAPIWebsocketResponse

func AppendCodexAPIWebsocketResponse(ctx context.Context, cfg *config.Config, payload []byte)

AppendCodexAPIWebsocketResponse stores a codex upstream websocket response frame and merges any quota event headers carried by the frame into the request log.

func ApplyClaudeCredentialMetadata

func ApplyClaudeCredentialMetadata(payload []byte, auth *cliproxyauth.Auth, sessionID string) ([]byte, string, error)

ApplyClaudeCredentialMetadata rewrites the identity exception shared by native and cloaked OAuth requests.

func ApplyClaudeDefaultDeviceProfileHeaders

func ApplyClaudeDefaultDeviceProfileHeaders(r *http.Request, cfg *config.Config)

func ApplyClaudeDeviceProfileHeaders

func ApplyClaudeDeviceProfileHeaders(r *http.Request, profile ClaudeDeviceProfile)

func ApplyClaudeLegacyDeviceHeaders

func ApplyClaudeLegacyDeviceHeaders(r *http.Request, ginHeaders http.Header, cfg *config.Config, confirmedClaudeCode bool)

func ApplyPayloadConfigWithRequest

func ApplyPayloadConfigWithRequest(cfg *config.Config, model, protocol, fromProtocol, root string, payload, original []byte, requestedModel string, requestPath string, headers http.Header) []byte

ApplyPayloadConfigWithRequest applies payload config using source protocol and request header gates.

func ApplyPayloadConfigWithRequestTracked

func ApplyPayloadConfigWithRequestTracked(cfg *config.Config, model, protocol, fromProtocol, root string, payload, original []byte, requestedModel string, requestPath string, headers http.Header, trackedPath string) ([]byte, bool)

ApplyPayloadConfigWithRequestTracked applies payload config and reports whether an applied rule targeted trackedPath or one of its descendants.

func ApplyPayloadConfigWithRoot

func ApplyPayloadConfigWithRoot(cfg *config.Config, model, protocol, root string, payload, original []byte, requestedModel string, requestPath string) []byte

ApplyPayloadConfigWithRoot behaves like applyPayloadConfig but treats all parameter paths as relative to the provided root path and restricts matches to the given protocol when supplied. Defaults are checked against the original payload when provided. requestedModel carries the client-visible model name before alias resolution so payload rules can target aliases precisely. requestPath is the inbound HTTP request path (when available) used for endpoint-scoped gates.

func ApplyPayloadConfigWithTrackedPaths

func ApplyPayloadConfigWithTrackedPaths(cfg *config.Config, model, protocol, fromProtocol, root string, payload, original []byte, requestedModel string, requestPath string, headers http.Header, trackedPaths ...string) ([]byte, map[string]bool)

ApplyPayloadConfigWithRequestTracked applies payload config and reports whether an applied rule targeted trackedPath or one of its descendants. ApplyPayloadConfigWithTrackedPaths applies payload config and reports which tracked paths (or their descendants) were targeted by an applied rule.

func ApplyRequestThinking

func ApplyRequestThinking(body []byte, req cliproxyexecutor.Request, opts cliproxyexecutor.Options, fromFormat, toFormat, provider string, normalizedUpdatesChanged ...bool) ([]byte, error)

ApplyRequestThinking preserves the registry lookup path unless the auth manager bound authoritative model capabilities to this execution attempt.

func ApplyThinkingWithSourcePayload

func ApplyThinkingWithSourcePayload(body, currentSourcePayload, originalSourcePayload []byte, model, fromFormat, toFormat, providerKey string) ([]byte, error)

ApplyThinkingWithSourcePayload preserves summary visibility from the original client payload while applying thinking configuration to its translated target payload. currentSourcePayload is the payload that was translated, while originalSourcePayload retains intent removed by an earlier interceptor.

func AugmentClaudeBuiltinToolRegistry

func AugmentClaudeBuiltinToolRegistry(body []byte, registry map[string]bool) map[string]bool

func BeginClaudeContinuity

func BeginClaudeContinuity(credentialIdentity, sessionID string, isNewPromptTurn bool, explicitPromptID string) (key string, sequence uint64, previousMessageID, previousRequestID, promptID string)

BeginClaudeContinuity starts one request generation for a stable credential identity and Claude conversation. It tracks the previous upstream message ID, previous upstream request ID (cc_prev_req), and active prompt ID (cc_prompt_id). If explicitPromptID is provided and valid, it is adopted. Otherwise if isNewPromptTurn is true (or if no prompt ID exists), a fresh UUIDv4 is generated.

func BeginClaudeDiagnostics

func BeginClaudeDiagnostics(credentialIdentity, sessionID string) (key string, sequence uint64, previousMessageID string)

BeginClaudeDiagnostics starts one request generation for a stable credential identity and Claude conversation. It returns the last successfully completed upstream message ID, if any. Only a SHA-256 digest of the credential identity and session is retained as the cache key, so access-token rotation does not interrupt continuity.

func BuildAntigravityCompactionResponse

func BuildAntigravityCompactionResponse(modelName, capsule string, inputTokens, outputTokens, totalTokens int) []byte

BuildAntigravityCompactionResponse creates JSON response for non-stream response.compaction.

func BuildAntigravityCompactionStreamChunks

func BuildAntigravityCompactionStreamChunks(modelName, capsule string, inputTokens, outputTokens, totalTokens int) [][]byte

BuildAntigravityCompactionStreamChunks creates SSE frames for compaction stream response.

func BuildDevinClientMetadataBytes

func BuildDevinClientMetadataBytes(sessionToken, deviceSeed, osName string) []byte

BuildDevinClientMetadataBytes constructs the serialized bytes for Field 1 (ClientMetadata).

func BuildDevinGetChatMessageRequest

func BuildDevinGetChatMessageRequest(
	sessionToken string,
	deviceSeed string,
	chatModelUID string,
	systemPrompt string,
	prompts []DevinPrompt,
	tools []DevinTool,
	temperature *float64,
	maxTokens int,
	sessionID string,
	cascadeID string,
	matcher *SensitiveWordMatcher,
) []byte

BuildDevinGetChatMessageRequest encodes an entire GetChatMessageRequest protobuf payload.

func BuildDevinUpstreamLogBody

func BuildDevinUpstreamLogBody(
	interactionsPayload []byte,
	isInteractionsSource bool,
	chatModelUID string,
	systemPrompt string,
	prompts []DevinPrompt,
	tools []DevinTool,
	temp *float64,
	maxTokens int,
	sessionID string,
	cascadeID string,
) []byte

BuildDevinUpstreamLogBody formats the intermediate interactions and the decoded Devin request into a clear, aligned log body without synthetic wrapping.

func BuildDevinUpstreamResponseLogBody

func BuildDevinUpstreamResponseLogBody(respLog *DevinUpstreamResponseLog, interactionsJSON []byte) []byte

BuildDevinUpstreamResponseLogBody formats the decoded Devin response and the intermediate interactions into a clear, aligned log body.

func BuildOpenAIUsageJSON

func BuildOpenAIUsageJSON(count int64) []byte

BuildOpenAIUsageJSON returns a minimal usage structure understood by downstream translators.

func CachedSessionID

func CachedSessionID(apiKey string) string

CachedSessionID returns a stable session UUID per apiKey, refreshing the TTL on each access.

func CachedSessionIDRequired

func CachedSessionIDRequired(ctx context.Context, apiKey string) (string, error)

CachedSessionIDRequired returns a stable session UUID per apiKey for request-time paths.

func CachedUserID

func CachedUserID(apiKey string) string

func CachedUserIDRequired

func CachedUserIDRequired(ctx context.Context, apiKey string) (string, error)

CachedUserIDRequired returns a stable fake user ID per apiKey for request-time paths.

func ClaudeAgentSessionUUID

func ClaudeAgentSessionUUID(headers http.Header, originalPayload, translatedPayload []byte, metadataSets ...map[string]any) string

ClaudeAgentSessionUUID maps the downstream agent conversation to one stable UUID, preserving native Claude Code session signals.

func ClaudeAgentSessionUUIDForRequest

func ClaudeAgentSessionUUIDForRequest(headers http.Header, originalPayload, translatedPayload []byte, confirmedClaudeCode bool, metadataSets ...map[string]any) string

ClaudeAgentSessionUUIDForRequest preserves Claude-specific session signals only for a confirmed native caller. Other callers use protocol session fields, execution metadata, or the stable derived conversation root.

func ClaudeCLIAuthIdentitySeed

func ClaudeCLIAuthIdentitySeed(auth *cliproxyauth.Auth) string

ClaudeCLIAuthIdentitySeed returns a stable credential identity that does not rotate with delegated-provider access tokens.

func ClaudeCodeExecutionScope

func ClaudeCodeExecutionScope(ctx context.Context, payload []byte, headers http.Header) (string, bool)

ClaudeCodeExecutionScope returns the stable root-session and agent identity used by Codex execution state.

func ClaudeCredentialAccountUUID

func ClaudeCredentialAccountUUID(auth *cliproxyauth.Auth) string

ClaudeCredentialAccountUUID returns the selected upstream credential's account UUID.

func ClaudeDeterministicPromptID

func ClaudeDeterministicPromptID(seed string) string

ClaudeDeterministicPromptID generates a deterministic RFC 4122 UUIDv4 from a seed string.

func ClaudeDeviceProfileStabilizationEnabled

func ClaudeDeviceProfileStabilizationEnabled(cfg *config.Config) bool

func ClaudeExecutionMetadataFromContext

func ClaudeExecutionMetadataFromContext(ctx context.Context) bool

ClaudeExecutionMetadataFromContext retrieves whether execution metadata was present from ctx.

func ClaudeHeadersIndicateUnifiedRateLimitRejection

func ClaudeHeadersIndicateUnifiedRateLimitRejection(headers http.Header) bool

ClaudeHeadersIndicateUnifiedRateLimitRejection reports whether response headers explicitly declare an Anthropic shared 5h or 7d rate-limit rejection. An overage-only or Fable-only rejection remains model-scoped when shared subscription windows are not rejected (status "allowed", "allowed_warning", or unexhausted when omitted).

func ClaudeMCPAliasWordCount

func ClaudeMCPAliasWordCount() int

ClaudeMCPAliasWordCount is the BIP-39 English dictionary size used for the virtual server pair and the one-word tool ID.

func ClaudeMCPToolAlias

func ClaudeMCPToolAlias(secret, original string, attempt uint32) string

ClaudeMCPToolAlias derives a Claude Code-style MCP tool name. Aliases from one caller share a virtual server component. The tool component combines a stable keyed ID with a truncated semantic suffix so the model can distinguish tools by name while the request-local symbol table restores the exact original. A higher attempt linearly probes the next word when a collision must be avoided. Server and tool IDs use BIP-39 English words so weak models are less likely to drift high-entropy Base32 fragments.

func ClaudePayloadHas1hTTL

func ClaudePayloadHas1hTTL(payload []byte) bool

ClaudePayloadHas1hTTL reports whether the request payload contains any cache_control block with ttl set to "1h".

func ClaudeRequestHasExecutionMetadata

func ClaudeRequestHasExecutionMetadata(metadataSets ...map[string]any) bool

ClaudeRequestHasExecutionMetadata reports whether the request explicitly carried an internal execution session metadata key.

func ClaudeSessionIDFromContext

func ClaudeSessionIDFromContext(ctx context.Context) string

ClaudeSessionIDFromContext retrieves the Claude session ID from ctx, if present.

func ClaudeSubagentRequests1h

func ClaudeSubagentRequests1h(headers http.Header, body []byte) bool

ClaudeSubagentRequests1h reports whether a subagent request explicitly requests 1h cache TTL either via a cache_control block with ttl="1h" in the payload or via extended-cache-ttl-2025-04-11 in incoming Anthropic-Beta headers.

func CodexPromptCacheKey

func CodexPromptCacheKey(modelName string, userScope string) string

CodexPromptCacheKey builds the Home KV key for a model/user prompt cache.

func CommitClaudeContinuity

func CommitClaudeContinuity(key string, sequence uint64, messageID, requestID string, promptIDs ...string)

CommitClaudeContinuity advances continuity only after a response completes. It commits the upstream message ID (msg_01...), request ID (req_01...), and prompt ID (cc_prompt_id). A non-empty messageID is required so incomplete/truncated streams do not advance sequence.

func CommitClaudeDiagnostics

func CommitClaudeDiagnostics(key string, sequence uint64, messageID string)

CommitClaudeDiagnostics advances continuity only after a response completes. A response from an older concurrently-started request cannot overwrite a newer committed generation, including after TTL expiry or capacity eviction.

func CountClaudeInputTokens

func CountClaudeInputTokens(payload []byte) (int64, error)

CountClaudeInputTokens estimates tokens for a Claude request with the O200kBase tokenizer.

func CountOpenAIChatTokens

func CountOpenAIChatTokens(enc tokenizer.Codec, payload []byte) (int64, error)

CountOpenAIChatTokens approximates prompt tokens for OpenAI chat completions payloads.

func CreditsUsed

func CreditsUsed(ctx context.Context) bool

CreditsUsed returns true if the request used AI credits.

func DefaultClaudeVersion

func DefaultClaudeVersion(cfg *config.Config) string

DefaultClaudeVersion returns the version string (e.g. "2.1.280") from the current baseline device profile. It extracts the version from the User-Agent.

func DeleteJSONField

func DeleteJSONField(body []byte, key string) []byte

DeleteJSONField removes a top-level or nested JSON field from a payload.

func DerivedAntigravitySessionID

func DerivedAntigravitySessionID(metadataSets ...map[string]any) string

DerivedAntigravitySessionID maps a derived session identity to Antigravity's negative decimal format.

func DerivedSessionID

func DerivedSessionID(metadataSets ...map[string]any) string

DerivedSessionID returns the first context-derived session identity in metadata order.

func DerivedSessionUUID

func DerivedSessionUUID(provider string, metadataSets ...map[string]any) string

DerivedSessionUUID maps a derived session identity to a provider-scoped stable UUID.

func EmitWebSocketResponseEvent

func EmitWebSocketResponseEvent(ctx context.Context, opts cliproxyexecutor.Options, auth *cliproxyauth.Auth, provider, model string, payload []byte)

EmitWebSocketResponseEvent delivers an upstream WebSocket response frame to any configured observer.

func EnsureClaudeCLIFingerprintIdentity

func EnsureClaudeCLIFingerprintIdentity(auth *cliproxyauth.Auth, seed string, synthesizeMissing bool) error

EnsureClaudeCLIFingerprintIdentity prepares auth.Metadata so the shared ApplyClaudeCredentialMetadata path can run.

When synthesizeMissing is false (real OAuth), this is a no-op: missing account or device data must surface as credential errors. When synthesizeMissing is true (fingerprint-profile=claude-code-cli on API keys), missing account_uuid / device pool are filled with stable values derived from seed. Callers that hold a shared Auth must use PrepareClaudeCLIFingerprintAuth instead.

func EnsureClaudeCredentialDevicePoolRequired

func EnsureClaudeCredentialDevicePoolRequired(ctx context.Context, auth *cliproxyauth.Auth) ([]string, error)

EnsureClaudeCredentialDevicePoolRequired initializes a credential pool locally, or coordinates it through Home KV when the selected auth is a remote dispatch clone.

func EnsureGeminiBoundaryUserContent

func EnsureGeminiBoundaryUserContent(payload []byte, path string) []byte

EnsureGeminiBoundaryUserContent ensures that the contents array at the given path both starts and ends with a user turn when sending to Gemini/Antigravity upstreams.

func EnsureGeminiLeadingUserContent

func EnsureGeminiLeadingUserContent(payload []byte, path string) []byte

EnsureGeminiLeadingUserContent ensures that the contents array at the given path starts with a user turn when sending to Gemini/Antigravity upstreams.

func EnsureGeminiTrailingUserContent

func EnsureGeminiTrailingUserContent(payload []byte, path string) []byte

EnsureGeminiTrailingUserContent ensures that the contents array at the given path ends with a user turn when sending to Gemini/Antigravity upstreams. If the final turn contains a functionResponse, it is preserved without appending an empty user turn because upstream expects the model to generate the response.

func EnsureResponsesUsageDetails

func EnsureResponsesUsageDetails(payload []byte) []byte

EnsureResponsesUsageDetails ensures that Responses usage objects contain output_tokens_details (defaulting reasoning_tokens to 0) and input_tokens_details (defaulting cached_tokens to 0). It supports plain JSON payloads, single-line SSE data: lines, and multi-line SSE frames (e.g. event: ...\ndata: ...).

func EnsureSessionContext

func EnsureSessionContext(ctx context.Context, opts cliproxyexecutor.Options, payload []byte) context.Context

EnsureSessionContext ensures that ctx carries the internal session identity for $CPA-SESSION-ID expansion in custom headers.

func ExecutorTypeName

func ExecutorTypeName(executor any) string

func ExpandAntigravityCompactionCapsules

func ExpandAntigravityCompactionCapsules(payload []byte) ([]byte, error)

ExpandAntigravityCompactionCapsules expands compaction items in input into developer context.

func ExtractAntigravitySummaryText

func ExtractAntigravitySummaryText(respPayload []byte) (string, error)

ExtractAntigravitySummaryText extracts summary text from Antigravity or Responses response.

func ExtractClaudeBillingTags

func ExtractClaudeBillingTags(body []byte) (prevReq, promptID string)

ExtractClaudeBillingTags extracts existing cc_prev_req and cc_prompt_id values from a billing header in system text, if present.

func ExtractClaudeCodeAgentID

func ExtractClaudeCodeAgentID(ctx context.Context, headers http.Header) string

ExtractClaudeCodeAgentID resolves the Claude Code agent ID and uses a stable sentinel for the root agent.

func ExtractClaudeCodeSessionID

func ExtractClaudeCodeSessionID(ctx context.Context, payload []byte, headers http.Header) string

ExtractClaudeCodeSessionID resolves a Claude Code session ID, preferring X-Claude-Code-Session-Id over payload metadata.

func ExtractStreamJSONPayload

func ExtractStreamJSONPayload(line []byte) []byte

ExtractStreamJSONPayload extracts SSE data/json payload from a raw line.

func FilterSSEUsageMetadata

func FilterSSEUsageMetadata(payload []byte) []byte

FilterSSEUsageMetadata removes usageMetadata from SSE events that are not terminal (finishReason != "stop"). Stop chunks are left untouched. This function is shared between aistudio and antigravity executors.

func GenerateDevinDeviceFingerprint

func GenerateDevinDeviceFingerprint(seed string) string

GenerateDevinDeviceFingerprint generates a 732-character hex device fingerprint. When seed is empty, it generates a cryptographically random 732-character hex string per request (matching native devin-cli). When seed is provided, it derives a deterministic 732-character hex fingerprint.

func GenerateDevinSentryTrace

func GenerateDevinSentryTrace() string

GenerateDevinSentryTrace generates a Sentry distributed tracing header in the format: "<32-hex-trace-id>-<16-hex-span-id>-1"

func GenerateFakeUserID

func GenerateFakeUserID() string

func GenerateFakeUserIDWithSessionID

func GenerateFakeUserIDWithSessionID(sessionID string) string

func HasCodexMultiAgentV2NamespaceConflict

func HasCodexMultiAgentV2NamespaceConflict(payload []byte) bool

HasCodexMultiAgentV2NamespaceConflict reports whether the request defines the reserved optimized namespace, which must remain untouched.

func HasDevinEffortSuffix

func HasDevinEffortSuffix(model string) bool

HasDevinEffortSuffix reports whether the model name already ends with a known Devin effort suffix.

func HasMeaningfulCodexOutputDelta

func HasMeaningfulCodexOutputDelta(eventData []byte) bool

HasMeaningfulCodexOutputDelta reports whether an event carries non-empty generated content (such as non-empty text, reasoning, or function call arguments delta).

func HasResponsesCompactionItem

func HasResponsesCompactionItem(payload []byte) bool

HasResponsesCompactionItem checks whether input contains a compaction item.

func HasResponsesCompactionTrigger

func HasResponsesCompactionTrigger(payload []byte) bool

HasResponsesCompactionTrigger checks whether input contains a compaction_trigger item.

func HeaderValueCaseInsensitive

func HeaderValueCaseInsensitive(headers http.Header, name string) string

HeaderValueCaseInsensitive returns the first non-empty header value matching name case-insensitively.

func HeaderValuesCaseInsensitive

func HeaderValuesCaseInsensitive(headers http.Header, name string) []string

HeaderValuesCaseInsensitive returns all non-empty header values matching name case-insensitively.

func IncomingHeadersFromContext

func IncomingHeadersFromContext(ctx context.Context) http.Header

IncomingHeadersFromContext retrieves incoming HTTP headers from ctx, if present.

func InjectClaudeBillingTags

func InjectClaudeBillingTags(body []byte, prevReq, promptID string) []byte

InjectClaudeBillingTags appends cc_prev_req and cc_prompt_id to the billing header in body if present.

func IsAnthropicUpstreamURL

func IsAnthropicUpstreamURL(u *url.URL) bool

IsAnthropicUpstreamURL reports whether a resolved request targets Anthropic's first-party API origin. Claude-specific body, header, HTTP, and TLS behavior must all use this gate so they cannot drift onto custom ports or userinfo URLs.

func IsChatTokenEvent

func IsChatTokenEvent(payload []byte) bool

IsChatTokenEvent reports whether the given OpenAI-compatible Chat Completions SSE or JSON chunk carries substantive output token content (such as text delta, reasoning content, or tool call arguments). It filters out container metadata, role announcements, and empty delta frames.

func IsClaudeMCPToolName

func IsClaudeMCPToolName(name string) bool

IsClaudeMCPToolName reports whether name follows Claude Code's MCP tool convention and contains only characters accepted by Anthropic tool names.

func IsClaudeNewPromptTurn

func IsClaudeNewPromptTurn(body []byte) bool

IsClaudeNewPromptTurn inspects the request messages to determine whether this request starts a new user prompt turn (requiring a new cc_prompt_id) versus a tool continuation step (which reuses the current cc_prompt_id).

func IsClaudeProbeOrHelperRequest

func IsClaudeProbeOrHelperRequest(body []byte) bool

IsClaudeProbeOrHelperRequest reports whether the request is a minimal probe/preflight (max_tokens: 1) or an automated title generation helper, which in native Claude Code do not emit cc_prompt_id or cc_prev_req.

func IsClaudeServerToolType

func IsClaudeServerToolType(toolType string) bool

IsClaudeServerToolType reports whether a typed declaration is a recognized Anthropic-operated tool. Client-defined type:"custom" declarations are not server tools and must remain eligible for MCP aliasing.

func IsClaudeSubagentRequest

func IsClaudeSubagentRequest(headers http.Header, body []byte) bool

IsClaudeSubagentRequest reports whether the incoming request originates from or represents a Claude Code subagent.

func IsClaudeTokenEvent

func IsClaudeTokenEvent(payload []byte) bool

IsClaudeTokenEvent reports whether an Anthropic Messages SSE chunk carries substantive output token content (such as text delta, thinking delta, or tool input delta). It filters out message_start, ping, content_block_stop, and empty containers.

func IsCodexModelSubstituted

func IsCodexModelSubstituted(requested, served string) bool

IsCodexModelSubstituted reports whether the upstream served a model other than the requested one; a dated alias pins a snapshot of the same model and is accepted.

func IsCodexTerminalEmptyIncomplete

func IsCodexTerminalEmptyIncomplete(eventData []byte, outputItemsCount int, sawOutputDelta bool) bool

IsCodexTerminalEmptyIncomplete reports whether a response.incomplete event represents an upstream silent abort where explicitly zero output tokens were generated and zero output content (no output items and no non-empty output deltas) was produced.

func IsGeminiTokenEvent

func IsGeminiTokenEvent(payload []byte) bool

IsGeminiTokenEvent reports whether a Google Gemini / Antigravity streaming chunk carries substantive output token content (such as text parts, thought traces, or function calls). It filters out metadata-only chunks (e.g. standalone usageMetadata) and empty parts.

func IsModelSubstituted

func IsModelSubstituted(requested, served string) bool

IsModelSubstituted reports whether the upstream served a model other than the requested one for any provider; dated aliases and snapshot pins are accepted.

func IsNativeCodexRequest

func IsNativeCodexRequest(body []byte, opts cliproxyexecutor.Options) bool

IsNativeCodexRequest checks the client dialect for use inside Codex executors.

func IsResponsesTokenEvent

func IsResponsesTokenEvent(payload []byte) bool

IsResponsesTokenEvent reports whether the given OpenAI/xAI Responses API WebSocket or SSE event payload carries an actual output token, reasoning trace, tool call argument, or multimodal delta, according to the official Responses API streaming / websocket specification plus Codex-compatible private events.

func IsValidClaudePromptID

func IsValidClaudePromptID(id string) bool

IsValidClaudePromptID verifies whether an explicit prompt ID adheres to strict RFC 4122 UUIDv4 semantics.

func IsValidUserID

func IsValidUserID(userID string) bool

func IterateStreamLines

func IterateStreamLines(payload []byte, fn func(line []byte))

IterateStreamLines splits payload by newline and invokes fn for non-empty lines.

func JSONPayload

func JSONPayload(line []byte) []byte

func JoinRawJSONArray

func JoinRawJSONArray(items [][]byte) []byte

JoinRawJSONArray joins validated raw JSON array items without re-encoding them.

func JoinRawJSONStrings

func JoinRawJSONStrings(items []string) []byte

JoinRawJSONStrings joins raw JSON array items held as strings.

func LogWithRequestID

func LogWithRequestID(ctx context.Context) *log.Entry

logWithRequestID returns a logrus Entry with request_id field populated from context. If no request ID is found in context, it returns the standard logger.

func MapStainlessArch

func MapStainlessArch() string

func MapStainlessOS

func MapStainlessOS() string

func MarkCreditsUsed

func MarkCreditsUsed(ctx context.Context)

MarkCreditsUsed flags the request as having used AI credits for billing.

func MergeStreamUsageDetail

func MergeStreamUsageDetail(existing, update usage.Detail) usage.Detail

MergeStreamUsageDetail merges existing stream usage with a newer update.

func NewDevinHTTPClient

func NewDevinHTTPClient(ctx context.Context, cfg *config.Config, auth *cliproxyauth.Auth, timeout time.Duration) *http.Client

NewDevinHTTPClient creates an HTTP client customized for Devin Connect-RPC upstream. Suppresses automatic Accept-Encoding: gzip while preserving connection reuse across requests.

func NewProxyAwareHTTPClient

func NewProxyAwareHTTPClient(ctx context.Context, cfg *config.Config, auth *cliproxyauth.Auth, timeout time.Duration) *http.Client

NewProxyAwareHTTPClient creates an HTTP client with proper proxy configuration priority: 1. Use the execution-scoped request proxy if configured (highest priority) 2. Use auth.ProxyURL if configured 3. Use cfg.ProxyURL if auth proxy is not configured 4. Use RoundTripper from context if none are configured

Parameters:

  • ctx: The context containing optional RoundTripper
  • cfg: The application configuration
  • auth: The authentication information
  • timeout: The client timeout (0 means no timeout)

Returns:

  • *http.Client: An HTTP client with configured proxy or transport

func NewUtlsHTTPClient

func NewUtlsHTTPClient(ctx context.Context, cfg *config.Config, auth *cliproxyauth.Auth, timeout time.Duration) *http.Client

NewUtlsHTTPClient creates an HTTP client using provider-specific TLS fingerprints for protected hosts. It uses Claude Code's Node/OpenSSL profile for Anthropic and a Chrome profile for ChatGPT, with a standard-transport fallback for other hosts.

func NextDevinSessionTurnIndex

func NextDevinSessionTurnIndex(sessionID string) int

NextDevinSessionTurnIndex returns the next 0-based request ordinal for a session (Field 15.2). In native devin-cli, the counter is process-scoped per session: First request in a session returns 0 (which is omitted on the wire). Subsequent requests return 1, 2, 3... monotonically.

func NormalizeCodexToolSchemas

func NormalizeCodexToolSchemas(body []byte) []byte

NormalizeCodexToolSchemas inspects function tools in a Codex request payload and simplifies pure constant union combinations (e.g. large oneOf branch sets representing enums with descriptions, as emitted by MCP servers) into semantically equivalent enum lists. Only unions mathematically proven to be semantically equivalent to enum definitions are modified; all other structures, property names, types, and constraints remain untouched.

func NormalizeOpenAIMaxTokens

func NormalizeOpenAIMaxTokens(payload []byte, useMaxCompletionTokens bool) []byte

NormalizeOpenAIMaxTokens normalizes max_tokens and max_completion_tokens according to the model's preference. When useMaxCompletionTokens is true, it ensures max_completion_tokens is set and max_tokens is removed. When useMaxCompletionTokens is false, it ensures max_tokens is used and max_completion_tokens is removed.

func NormalizeOpenAIToolResultsTextOnly

func NormalizeOpenAIToolResultsTextOnly(payload []byte) []byte

NormalizeOpenAIToolResultsTextOnly converts tool message content to strings and strips relayed tool result images for text-only compatibility models. Text parts are preserved and image parts are replaced with a short marker.

func NormalizeThinkingLevel

func NormalizeThinkingLevel(level string, budgetTokens int) string

NormalizeThinkingLevel converts numeric budgets or loose effort strings to canonical Devin efforts.

func ObfuscateSensitiveWords

func ObfuscateSensitiveWords(payload []byte, matcher *SensitiveWordMatcher) []byte

ObfuscateSensitiveWords processes the payload and obfuscates sensitive words in system blocks and message content.

func ObfuscateSensitiveWordsInSystemInstruction

func ObfuscateSensitiveWordsInSystemInstruction(payload []byte, matcher *SensitiveWordMatcher) []byte

ObfuscateSensitiveWordsInSystemInstruction obfuscates sensitive words in an Antigravity system instruction.

func ObserveChatTokenEvent

func ObserveChatTokenEvent(reporter *UsageReporter, payload []byte)

ObserveChatTokenEvent inspects an OpenAI Chat Completions chunk and records TTFT if the frame represents the first meaningful token event. It records first-packet arrival time as fallback and returns immediately with zero allocations once effective token TTFT is set.

func ObserveClaudeTokenEvent

func ObserveClaudeTokenEvent(reporter *UsageReporter, payload []byte)

ObserveClaudeTokenEvent inspects an Anthropic Messages SSE chunk and records TTFT if the frame represents the first meaningful token event. It records first-packet arrival time as fallback and returns immediately with zero allocations once effective token TTFT is set.

func ObserveGeminiTokenEvent

func ObserveGeminiTokenEvent(reporter *UsageReporter, payload []byte)

ObserveGeminiTokenEvent inspects a Google Gemini / Antigravity streaming chunk and records TTFT if the frame represents the first meaningful token event. It records first-packet arrival time as fallback and returns immediately with zero allocations once effective token TTFT is set.

func ObserveMergedStreamUsage

func ObserveMergedStreamUsage(buffer *StreamUsageBuffer, update usage.Detail)

ObserveMergedStreamUsage updates buffer with merged usage details.

func ObservePluginExecutorStreamTTFT

func ObservePluginExecutorStreamTTFT(protocol string, reporter *UsageReporter, payload []byte)

ObservePluginExecutorStreamTTFT inspects a streaming payload and records TTFT on the reporter.

func ObservePluginExecutorStreamUsage

func ObservePluginExecutorStreamUsage(protocol string, payload []byte, buffer *StreamUsageBuffer)

ObservePluginExecutorStreamUsage parses streaming chunks and updates a stream usage buffer.

func ObserveResponsesTokenEvent

func ObserveResponsesTokenEvent(reporter *UsageReporter, payload []byte)

ObserveResponsesTokenEvent inspects a Responses API frame payload and records TTFT if the frame represents the first meaningful token event. It records the first packet arrival time as a fallback and exits immediately with zero allocations once effective token TTFT is set.

func OptimizeCodexMultiAgentV2Request

func OptimizeCodexMultiAgentV2Request(ctx context.Context, headers http.Header, payload []byte, cfg *config.Config) ([]byte, bool)

OptimizeCodexMultiAgentV2Request rewrites an eligible spawn_agent request and reports whether the collaboration namespace was renamed for upstream use.

func OptimizeCodexMultiAgentV2RequestForAuth

func OptimizeCodexMultiAgentV2RequestForAuth(ctx context.Context, headers http.Header, payload []byte, cfg *config.Config, auth *cliproxyauth.Auth, model string) ([]byte, bool)

OptimizeCodexMultiAgentV2RequestForAuth applies the standard Codex MultiAgentV2 request optimization and, when the selected codex-api-key model has is-compat enabled, also converts agent_message items into portable message/user input, proactively stripping author, recipient, and internal passthrough metadata.

func ParseAntigravityStreamUsage

func ParseAntigravityStreamUsage(line []byte) (usage.Detail, bool)

func ParseAntigravityUsage

func ParseAntigravityUsage(data []byte) usage.Detail

func ParseClaudeRateLimitReset

func ParseClaudeRateLimitReset(headers http.Header, now time.Time) *time.Duration

ParseClaudeRateLimitReset inspects Anthropic response headers for shared and Fable-specific unified rate-limit and standard Retry-After reset information, returning the conservative cooldown duration including a bounded non-negative random grace period. If no valid future reset information is present, it returns nil.

func ParseClaudeStreamUsage

func ParseClaudeStreamUsage(line []byte) (usage.Detail, bool)

func ParseClaudeUsage

func ParseClaudeUsage(data []byte) usage.Detail

func ParseCodexImageToolUsage

func ParseCodexImageToolUsage(data []byte) (usage.Detail, bool)

func ParseCodexQuotaEventHeaders

func ParseCodexQuotaEventHeaders(payload []byte) http.Header

ParseCodexQuotaEventHeaders converts one Codex websocket quota event into the same bounded header representation used by HTTP usage observations. Codex sends additional_rate_limits as an object on the websocket path, while the /wham/usage probe exposes the equivalent data as an array; both forms are normalized into namespaced X-Codex headers here.

The websocket payload identifies each additional limit only by limit name ("GPT-5.3-Codex-Spark"), while HTTP responses namespace the same limit by a short name (x-codex-bengalfox-*). The two forms therefore cannot produce identical header names; the X-Codex-Additional- prefix marks the websocket origin, and the snapshot-replacement semantics in QuotaState keep the two spellings from accumulating side by side.

func ParseCodexUsage

func ParseCodexUsage(data []byte) (usage.Detail, bool)

func ParseDevinResponseDimensionGroups

func ParseDevinResponseDimensionGroups(groups ...[]byte) (promptTokens, completionTokens, cachedTokens int64, found bool)

ParseDevinResponseDimensionGroups parses Field 28 (ResponseDimensionGroups) entries to extract Token Usage metrics: input_tokens, output_tokens, cached_input_tokens. Accepts one or more group payloads (each corresponding to a Field 28 value), or an outer envelope containing Tag 28.

func ParseDevinTrailerError

func ParseDevinTrailerError(payload []byte) (statusCode int, err error)

ParseDevinTrailerError inspects Connect-RPC EOS trailer frames and maps error status codes.

func ParseGeminiStreamUsage

func ParseGeminiStreamUsage(line []byte) (usage.Detail, bool)

func ParseGeminiUsage

func ParseGeminiUsage(data []byte) usage.Detail

func ParseInteractionsStreamUsage

func ParseInteractionsStreamUsage(line []byte) (usage.Detail, bool)

func ParseInteractionsUsage

func ParseInteractionsUsage(data []byte) usage.Detail

func ParseOpenAIStreamUsage

func ParseOpenAIStreamUsage(line []byte) (usage.Detail, bool)

func ParseOpenAIUsage

func ParseOpenAIUsage(data []byte) usage.Detail

func ParsePluginExecutorResponseUsage

func ParsePluginExecutorResponseUsage(protocol string, payload []byte) usage.Detail

ParsePluginExecutorResponseUsage extracts token usage from a non-streaming plugin executor response.

func ParseRetryDelay

func ParseRetryDelay(errorBody []byte) (*time.Duration, error)

ParseRetryDelay extracts the retry delay from a Google API 429 error response.

func PayloadRequestPath

func PayloadRequestPath(opts cliproxyexecutor.Options) string

func PayloadRequestedModel

func PayloadRequestedModel(opts cliproxyexecutor.Options, fallback string) string

func PrepareAntigravityCompactionSummaryPayload

func PrepareAntigravityCompactionSummaryPayload(payload []byte, modelName string) []byte

PrepareAntigravityCompactionSummaryPayload prepares a payload for non-stream Antigravity summary generation.

func PrepareClaudeCLIFingerprintAuth

func PrepareClaudeCLIFingerprintAuth(auth *cliproxyauth.Auth, seed string, synthesizeMissing bool) (*cliproxyauth.Auth, error)

PrepareClaudeCLIFingerprintAuth returns the auth object that should receive ApplyClaudeCredentialMetadata. Synthesized API-key / delegated-provider identity is written to a clone so the shared credential metadata map is not mutated on the request path.

func ProviderSessionUUID

func ProviderSessionUUID(provider string, metadataSets ...map[string]any) string

ProviderSessionUUID prefers a long-lived execution session and falls back to the derived identity.

func ReadConnectFrame

func ReadConnectFrame(r io.Reader) (flag byte, payload []byte, err error)

ReadConnectFrame reads a single framed message from a Connect-proto stream.

func RecordAPIRequest

func RecordAPIRequest(ctx context.Context, cfg *config.Config, info UpstreamRequestLog)

RecordAPIRequest stores the upstream request metadata in Gin context for request logging.

func RecordAPIResponseError

func RecordAPIResponseError(ctx context.Context, cfg *config.Config, err error)

RecordAPIResponseError adds an error entry for the latest attempt when no HTTP response is available.

func RecordAPIResponseMetadata

func RecordAPIResponseMetadata(ctx context.Context, cfg *config.Config, status int, headers http.Header)

RecordAPIResponseMetadata captures upstream response status/header information for the latest attempt.

func RecordAPIWebsocketError

func RecordAPIWebsocketError(ctx context.Context, cfg *config.Config, stage string, err error)

RecordAPIWebsocketError stores an upstream websocket error event in Gin context.

func RecordAPIWebsocketHandshake

func RecordAPIWebsocketHandshake(ctx context.Context, cfg *config.Config, status int, headers http.Header)

RecordAPIWebsocketHandshake stores the upstream websocket handshake response metadata.

func RecordAPIWebsocketRequest

func RecordAPIWebsocketRequest(ctx context.Context, cfg *config.Config, info UpstreamRequestLog)

RecordAPIWebsocketRequest stores an upstream websocket request event in Gin context.

func RecordAPIWebsocketUpgradeRejection

func RecordAPIWebsocketUpgradeRejection(ctx context.Context, cfg *config.Config, info UpstreamRequestLog, status int, headers http.Header, body []byte)

RecordAPIWebsocketUpgradeRejection stores a rejected websocket upgrade as an HTTP attempt.

func RefreshAuthViaHome

func RefreshAuthViaHome(ctx context.Context, cfg *config.Config, auth *cliproxyauth.Auth) (*cliproxyauth.Auth, bool, error)

RefreshAuthViaHome replaces local refresh logic when home control plane integration is enabled. It returns (updatedAuth, true, nil) when home refresh succeeds; (nil, true, err) when home is enabled but refresh fails; and (nil, false, nil) when home is disabled.

func ResetClaudeDeviceProfileCache

func ResetClaudeDeviceProfileCache()

func ResetClaudeDiagnosticsForTest

func ResetClaudeDiagnosticsForTest()

ResetClaudeDiagnosticsForTest resets in-memory continuity state for test isolation.

func ResetDevinSessionTurnIndex

func ResetDevinSessionTurnIndex(sessionID string)

ResetDevinSessionTurnIndex clears the session counter (used for testing or explicit session reset).

func ResolveAntigravityGroundingURLs

func ResolveAntigravityGroundingURLs(ctx context.Context, cfg *config.Config, auth *cliproxyauth.Auth, payload []byte) []byte

ResolveAntigravityGroundingURLs replaces Vertex Search redirect URLs in grounding chunks with their target URLs.

func ResolveDevinChatModelUID

func ResolveDevinChatModelUID(rawModel string, thinkingLevel string, budgetTokens int) string

ResolveDevinChatModelUID resolves a model identifier into a valid upstream Devin chat_model_uid. It prioritizes dynamic catalog metadata from devin_models.json, automatically clamps requested efforts to supported levels, applies sensible default efforts for thinking models, and ensures bare non-thinking models remain bare.

func ResolveKimiBaseURL

func ResolveKimiBaseURL(auth *cliproxyauth.Auth) string

ResolveKimiBaseURL resolves the upstream API base URL for Kimi API requests based on auth.

func ResolveKimiChatURL

func ResolveKimiChatURL(auth *cliproxyauth.Auth) string

ResolveKimiChatURL resolves the upstream URL for Kimi Chat Completions requests.

func ResolveKimiClaudeBaseURL

func ResolveKimiClaudeBaseURL(auth *cliproxyauth.Auth) string

ResolveKimiClaudeBaseURL resolves the base URL for Kimi Claude Messages delegation.

func ResolveKimiResponsesURL

func ResolveKimiResponsesURL(auth *cliproxyauth.Auth) string

ResolveKimiResponsesURL resolves the upstream URL for Kimi Responses API requests.

func RestoreCodexMultiAgentV2Response

func RestoreCodexMultiAgentV2Response(payload []byte, optimized bool) []byte

RestoreCodexMultiAgentV2Response restores optimized collaboration namespace values before an upstream response is translated and returned to the client.

func RewriteCodexMultiAgentV2Input

func RewriteCodexMultiAgentV2Input(ctx context.Context, headers http.Header, payload []byte, cfg *config.Config) []byte

RewriteCodexMultiAgentV2Input converts official Codex multi-agent input into standard Responses API messages when multi-agent v2 optimization is enabled.

func RewriteCodexOrphanDelegationInput

func RewriteCodexOrphanDelegationInput(ctx context.Context, headers http.Header, payload []byte, cfg *config.Config) []byte

RewriteCodexOrphanDelegationInput converts orphan Codex delegation outputs into standard user messages when orphan delegation compatibility is enabled and the request carries the X-Openai-Subagent: collab_spawn header.

func RewriteCodexSpawnAgentDescription

func RewriteCodexSpawnAgentDescription(ctx context.Context, headers http.Header, payload []byte, cfg *config.Config) []byte

RewriteCodexSpawnAgentDescription optimizes spawn_agent definitions for official Codex clients when multi-agent v2 optimization is enabled.

func SanitizeCodexInputItemIDs

func SanitizeCodexInputItemIDs(body []byte) []byte

SanitizeCodexInputItemIDs normalizes supported input item IDs for Codex, removes encrypted reasoning items whose IDs exceed the Codex limit, and deterministically shortens other overlong input item IDs.

func SanitizeDevinSystemPrompt

func SanitizeDevinSystemPrompt(prompt string, matcher *SensitiveWordMatcher) string

SanitizeDevinSystemPrompt strips Claude Code attribution and CLI identity headers (referencing Antigravity conventions), and applies zero-width obfuscation for configured sensitive words.

func SealAntigravityCompaction

func SealAntigravityCompaction(summary, modelName string) (string, error)

SealAntigravityCompaction encrypts summary data into an opaque capsule using AES-GCM.

func SetBoolIfDifferent

func SetBoolIfDifferent(payload []byte, path string, value bool) []byte

SetBoolIfDifferent updates path only when its value is not already the canonical JSON boolean. Values with another JSON type are still normalized.

func SetCodexCache

func SetCodexCache(key string, cache CodexCache)

SetCodexCache stores a cache entry.

func SetCodexCacheBestEffort

func SetCodexCacheBestEffort(ctx context.Context, key string, cache CodexCache) bool

SetCodexCacheBestEffort stores a cache entry without failing completed responses.

func SetCodexCacheRequired

func SetCodexCacheRequired(ctx context.Context, key string, cache CodexCache) error

SetCodexCacheRequired stores a cache entry for request-time paths.

func SetRawIfDifferent

func SetRawIfDifferent(payload []byte, path string, value []byte) []byte

SetRawIfDifferent updates path only when the existing raw JSON is identical.

func SetStringIfDifferent

func SetStringIfDifferent(payload []byte, path, value string) []byte

SetStringIfDifferent updates path only when its value is not already the canonical JSON string. Values with another JSON type are still normalized.

func ShouldNormalizeOpenAIToolResultsForModel

func ShouldNormalizeOpenAIToolResultsForModel(compat *config.OpenAICompatibility, upstreamModel, requestedModel string) bool

ShouldNormalizeOpenAIToolResultsForModel reports whether the selected model explicitly excludes image input through its input-modalities configuration.

func ShouldUseMaxCompletionTokensForModel

func ShouldUseMaxCompletionTokensForModel(compat *config.OpenAICompatibility, upstreamModel, requestedModel string) bool

ShouldUseMaxCompletionTokensForModel reports whether the resolved model under an OpenAI compatibility configuration has use-max-completion-tokens set to true.

func StableClaudeCLIAccountUUID

func StableClaudeCLIAccountUUID(seed string) string

StableClaudeCLIAccountUUID returns a deterministic UUIDv5 account ID derived from a seed.

func StableClaudeCLIDeviceID

func StableClaudeCLIDeviceID(seed string) string

StableClaudeCLIDeviceID returns a deterministic device ID derived from a seed.

func StripClaudeBillingTags

func StripClaudeBillingTags(body []byte) []byte

StripClaudeBillingTags removes cc_prev_req and cc_prompt_id from the billing header in body.

func StripUsageMetadataFromJSON

func StripUsageMetadataFromJSON(rawJSON []byte) ([]byte, bool)

StripUsageMetadataFromJSON drops usageMetadata unless finishReason is present (terminal). It handles both formats: - Aistudio: candidates.0.finishReason - Antigravity: response.candidates.0.finishReason

func StripVertexOpenAIResponsesToolCallIDs

func StripVertexOpenAIResponsesToolCallIDs(payload []byte, sourceFormat string) []byte

StripVertexOpenAIResponsesToolCallIDs removes OpenAI Responses call IDs that Vertex rejects in Gemini functionCall/functionResponse payloads.

func SummarizeErrorBody

func SummarizeErrorBody(contentType string, body []byte) string

func TokenizerForModel

func TokenizerForModel(model string) (tokenizer.Codec, error)

TokenizerForModel returns a tokenizer codec suitable for an OpenAI-style model id.

func TranslateRequestEnvelopePairWithCodexMultiAgentV2

func TranslateRequestEnvelopePairWithCodexMultiAgentV2(ctx context.Context, headers http.Header, cfg *config.Config, from, to sdktranslator.Format, req sdktranslator.RequestEnvelope, originalPayload, requestPayload []byte) (original, working []byte)

TranslateRequestEnvelopePairWithCodexMultiAgentV2 translates the baseline and working payload while preserving request-scoped metadata in req.

func TranslateRequestEnvelopeWithCodexMultiAgentV2

func TranslateRequestEnvelopeWithCodexMultiAgentV2(ctx context.Context, headers http.Header, cfg *config.Config, from, to sdktranslator.Format, req sdktranslator.RequestEnvelope) sdktranslator.RequestEnvelope

TranslateRequestEnvelopeWithCodexMultiAgentV2 normalizes official Codex multi-agent input while preserving the complete request envelope.

func TranslateRequestPairWithAPIKeyModelCompatibility

func TranslateRequestPairWithAPIKeyModelCompatibility(ctx context.Context, headers http.Header, cfg *config.Config, from, to sdktranslator.Format, model string, originalPayload, requestPayload []byte, stream, isCompat bool) (original, working []byte)

TranslateRequestPairWithAPIKeyModelCompatibility avoids translating identical inputs twice while retaining separate buffers and stateful plugin invocations.

func TranslateRequestPairWithAPIKeyModelCompatibilityAndUpdateIntent

func TranslateRequestPairWithAPIKeyModelCompatibilityAndUpdateIntent(ctx context.Context, headers http.Header, cfg *config.Config, from, to sdktranslator.Format, model string, originalPayload, requestPayload []byte, stream, isCompat bool) (original, working []byte, updatesChanged bool)

TranslateRequestPairWithAPIKeyModelCompatibilityAndUpdateIntent returns the normalizer decision for the working payload, not for the baseline payload.

func TranslateRequestPairWithCodexMultiAgentV2

func TranslateRequestPairWithCodexMultiAgentV2(ctx context.Context, headers http.Header, cfg *config.Config, from, to sdktranslator.Format, model string, originalPayload, requestPayload []byte, stream bool) (original, working []byte)

TranslateRequestPairWithCodexMultiAgentV2 translates the untouched baseline payload and the working payload that later stages mutate in place. Executors normally assign the original payload to the request before translating, so both translations would rescan the same bytes and produce the same result. Built-in request translation is deterministic, so that case is translated once and duplicated when no plugin hooks are installed. Hooks retain two invocations because they may have request-scoped output or side effects. This removes a full extra pass over payloads that can reach tens of megabytes.

func TranslateRequestWithAPIKeyModelCompatibility

func TranslateRequestWithAPIKeyModelCompatibility(ctx context.Context, headers http.Header, cfg *config.Config, from, to sdktranslator.Format, model string, payload []byte, stream, isCompat bool) []byte

TranslateRequestWithAPIKeyModelCompatibility applies compatibility-aware request translators when a configured API-key model enables compatibility mode.

func TranslateRequestWithAPIKeyModelCompatibilityAndUpdateIntent

func TranslateRequestWithAPIKeyModelCompatibilityAndUpdateIntent(ctx context.Context, headers http.Header, cfg *config.Config, from, to sdktranslator.Format, model string, payload []byte, stream, isCompat bool) ([]byte, bool)

TranslateRequestWithAPIKeyModelCompatibilityAndUpdateIntent returns the plugin normalizer's request-scoped update decision for Responses targets.

func TranslateRequestWithCodexMultiAgentV2

func TranslateRequestWithCodexMultiAgentV2(ctx context.Context, headers http.Header, cfg *config.Config, from, to sdktranslator.Format, model string, payload []byte, stream bool) []byte

TranslateRequestWithCodexMultiAgentV2 normalizes official Codex multi-agent input before translating it to a non-Codex target protocol.

func TranslateStreamWithClaudeInputTokens

func TranslateStreamWithClaudeInputTokens(
	ctx context.Context,
	upstreamFormat, responseFormat sdktranslator.Format,
	model string,
	originalRequestRawJSON, requestRawJSON, rawJSON []byte,
	param *any,
	state *ClaudeInputTokenState,
) [][]byte

TranslateStreamWithClaudeInputTokens translates a stream chunk and estimates Claude message_start input usage once.

func UnsealAntigravityCompaction

func UnsealAntigravityCompaction(encryptedContent string) (string, error)

UnsealAntigravityCompaction decrypts and validates an opaque capsule.

func WebsocketUpgradeRequestURL

func WebsocketUpgradeRequestURL(rawURL string) string

WebsocketUpgradeRequestURL converts a websocket URL back to its HTTP handshake URL for logging.

func WithClaudeContinuityContext

func WithClaudeContinuityContext(ctx context.Context, cc *ClaudeContinuityContext) context.Context

WithClaudeContinuityContext attaches a mutable ClaudeContinuityContext to ctx.

func WithClaudeExecutionMetadata

func WithClaudeExecutionMetadata(ctx context.Context, present bool) context.Context

WithClaudeExecutionMetadata attaches whether the request had an explicit execution session metadata to ctx.

func WithClaudeSessionID

func WithClaudeSessionID(ctx context.Context, sessionID string) context.Context

WithClaudeSessionID attaches a known Claude session ID to ctx.

func WithIncomingHeaders

func WithIncomingHeaders(ctx context.Context, headers http.Header) context.Context

WithIncomingHeaders attaches incoming HTTP headers to ctx.

func WrapConnectEnvelope

func WrapConnectEnvelope(protoBytes []byte) []byte

WrapConnectEnvelope wraps raw payload bytes into a standard 5-byte Connect envelope: [1 byte flag: 0x00] + [4 byte big-endian length] + [payload].

func WrapConnectEnvelopeWithFlag

func WrapConnectEnvelopeWithFlag(flag byte, protoBytes []byte) []byte

WrapConnectEnvelopeWithFlag wraps payload bytes with an explicit Connect flag.

Types

type ClaudeCodeRequestDetection

type ClaudeCodeRequestDetection struct {
	Confirmed       bool
	StrongSignals   bool
	NativeClient    bool
	XAppCLI         bool
	UserAgent       bool
	BetasPresent    bool
	MetadataUserID  bool
	HelperProfile   bool
	Entrypoint      string
	Subclient       string
	AgentSDKVersion string
}

ClaudeCodeRequestDetection records the strong signals and first-party subclient identity used to distinguish an official Claude Code request from a client that only copied its User-Agent.

func DetectClaudeCodeRequest

func DetectClaudeCodeRequest(headers http.Header, payload []byte, countTokens bool, configs ...*config.Config) ClaudeCodeRequestDetection

DetectClaudeCodeRequest first mirrors CCH's strong-signal contract, then applies CPA's native-client policy. Standard Messages requests require all four strong signals; count_tokens omits metadata.user_id. A separate narrow profile recognizes measured native Haiku helper requests that intentionally omit claude-code-20250219. Generic sdk-ts/sdk-py Agent SDK entrypoints remain unconfirmed and receive CLI cloaking.

type ClaudeContinuityContext

type ClaudeContinuityContext struct {
	Key               string
	Sequence          uint64
	PreviousMessageID string
	PreviousRequestID string
	PromptID          string
	Initialized       bool
}

ClaudeContinuityContext holds request-scoped continuity state across cloaking and execution.

func ClaudeContinuityContextFromContext

func ClaudeContinuityContextFromContext(ctx context.Context) *ClaudeContinuityContext

ClaudeContinuityContextFromContext retrieves the ClaudeContinuityContext from ctx, if present.

type ClaudeDeviceProfile

type ClaudeDeviceProfile struct {
	UserAgent      string
	PackageVersion string
	RuntimeVersion string
	OS             string
	Arch           string
	// contains filtered or unexported fields
}

func ResolveClaudeDeviceProfile

func ResolveClaudeDeviceProfile(auth *cliproxyauth.Auth, apiKey string, headers http.Header, cfg *config.Config) ClaudeDeviceProfile

func ResolveClaudeDeviceProfileRequired

func ResolveClaudeDeviceProfileRequired(ctx context.Context, auth *cliproxyauth.Auth, apiKey string, headers http.Header, cfg *config.Config) (ClaudeDeviceProfile, error)

ResolveClaudeDeviceProfileRequired resolves a stable Claude Code device profile for request-time paths.

type ClaudeInputTokenState

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

ClaudeInputTokenState tracks the one-time input token update for a translated Claude stream.

func NewClaudeInputTokenState

func NewClaudeInputTokenState(sourceFormat, upstreamFormat, responseFormat sdktranslator.Format, originalRequest []byte) *ClaudeInputTokenState

NewClaudeInputTokenState creates request-scoped state for translated Claude input token usage.

type CodexCache

type CodexCache struct {
	ID     string
	Expire time.Time
}

func ClaudeCodePromptCache

func ClaudeCodePromptCache(ctx context.Context, modelName string, payload []byte, headers http.Header) (CodexCache, bool, error)

ClaudeCodePromptCache derives a deterministic upstream prompt_cache_key for one Claude Code agent.

func GetCodexCache

func GetCodexCache(key string) (CodexCache, bool)

GetCodexCache retrieves a cached entry, returning ok=false if not found or expired.

func GetCodexCacheRequired

func GetCodexCacheRequired(ctx context.Context, key string) (CodexCache, bool, error)

GetCodexCacheRequired retrieves a cached entry for request-time paths.

type DevinFrameResult

type DevinFrameResult struct {
	OutputID                string
	Timestamp               uint64
	ContentText             string
	DeltaTokens             uint64
	StopReason              uint64 // 2/4=stop, 10=tool_calls
	ToolCallDeltas          []DevinToolCallDelta
	ThinkingText            string
	DeltaSignature          []byte
	DeltaSignatureType      string
	Latency                 float64
	MessageID               string
	Usage                   *DevinUsage
	ResponseDimensionGroups [][]byte
	UnknownFieldNumbers     []int
}

DevinFrameResult represents decoded content from a single Connect-proto frame.

func ParseDevinFrame

func ParseDevinFrame(payload []byte) (DevinFrameResult, error)

ParseDevinFrame extracts deltas, tool calls, thinking, signatures, and usage from a single response frame.

type DevinImage

type DevinImage struct {
	Base64Data string
	MimeType   string
}

DevinImage represents an image attachment in a DevinPrompt (Prompt #10).

type DevinImageLogItem

type DevinImageLogItem struct {
	MimeType string `json:"mime_type"`
	DataLen  int    `json:"data_len"`
}

DevinImageLogItem represents an image attached to a prompt in the request log.

type DevinPrompt

type DevinPrompt struct {
	MessageID          string
	Source             int // 1=user, 2=assistant, 4=tool
	Content            string
	Images             []DevinImage
	ToolCalls          []DevinToolCall
	ToolCallID         string // For source=4 (tool result)
	OriginalToolCallID string // Retained when downgraded from source=4 to source=1
	IsOrphanedTool     bool   // Explicit flag marking downgraded tool results
	Thinking           string
	Signature          []byte
	SignatureType      string
}

DevinPrompt represents a single turn in the request history (repeated Field 3).

type DevinPromptLogItem

type DevinPromptLogItem struct {
	ID            string              `json:"id,omitempty"`
	Source        int                 `json:"source"`
	Role          string              `json:"role,omitempty"`
	Content       string              `json:"content,omitempty"`
	Thinking      string              `json:"thinking,omitempty"`
	Signature     string              `json:"signature,omitempty"`
	SignatureType string              `json:"signature_type,omitempty"`
	ToolCalls     []DevinToolCall     `json:"tool_calls,omitempty"`
	ToolCallID    string              `json:"tool_call_id,omitempty"`
	Images        []DevinImageLogItem `json:"images,omitempty"`
}

DevinPromptLogItem represents a single prompt item in the request log.

type DevinTool

type DevinTool struct {
	Name        string
	Description string
	Parameters  []byte
}

DevinTool represents a tool definition in GetChatMessageRequest.

type DevinToolCall

type DevinToolCall struct {
	ID        string
	Name      string
	Arguments string
}

DevinToolCall represents a tool call in a ChatMessagePrompt.

type DevinToolCallDelta

type DevinToolCallDelta struct {
	ID               string
	Name             string
	Arguments        string
	InvalidJSONStr   string
	InvalidJSONErr   string
	IsCustomToolCall bool
}

DevinToolCallDelta represents a streaming tool call chunk from response Field 6.

type DevinToolLogItem

type DevinToolLogItem struct {
	Name        string          `json:"name"`
	Description string          `json:"description,omitempty"`
	Parameters  json.RawMessage `json:"parameters,omitempty"`
}

DevinToolLogItem represents a tool declared in the request log.

type DevinUpstreamRequestLog

type DevinUpstreamRequestLog struct {
	Model        string               `json:"model"`
	SessionID    string               `json:"session_id,omitempty"`
	CascadeID    string               `json:"cascade_id,omitempty"`
	SystemPrompt string               `json:"system_prompt,omitempty"`
	Temperature  *float64             `json:"temperature,omitempty"`
	MaxTokens    int                  `json:"max_tokens,omitempty"`
	Prompts      []DevinPromptLogItem `json:"prompts,omitempty"`
	Tools        []DevinToolLogItem   `json:"tools,omitempty"`
}

DevinUpstreamRequestLog represents the human-readable JSON representation of GetChatMessageRequest for request logs.

type DevinUpstreamResponseLog

type DevinUpstreamResponseLog struct {
	Status        string          `json:"status,omitempty"`
	FramesCount   int             `json:"frames_count"`
	Content       string          `json:"content,omitempty"`
	Thinking      string          `json:"thinking,omitempty"`
	Signature     string          `json:"signature,omitempty"`
	SignatureType string          `json:"signature_type,omitempty"`
	ToolCalls     []DevinToolCall `json:"tool_calls,omitempty"`
	Usage         *DevinUsage     `json:"usage,omitempty"`
	UnknownFields []int           `json:"unknown_fields,omitempty"`
}

DevinUpstreamResponseLog represents the decoded response frames from Devin Connect-RPC.

type DevinUsage

type DevinUsage struct {
	PromptTokens     int64             `json:"prompt_tokens"`
	CompletionTokens int64             `json:"completion_tokens"`
	CachedTokens     int64             `json:"cached_tokens"`
	CacheWriteTokens int64             `json:"cache_write_tokens,omitempty"`
	StatusCode       uint64            `json:"status_code,omitempty"`
	RequestID        string            `json:"request_id,omitempty"`
	ModelName        string            `json:"model_name,omitempty"`
	Headers          map[string]string `json:"headers,omitempty"`
}

DevinUsage captures token accounting from response Field 7.

type SensitiveWordMatcher

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

SensitiveWordMatcher holds the compiled regex for matching sensitive words.

func BuildSensitiveWordMatcher

func BuildSensitiveWordMatcher(words []string) *SensitiveWordMatcher

BuildSensitiveWordMatcher compiles a regex from the word list. Words are sorted by length (longest first) for proper matching.

func (*SensitiveWordMatcher) Matches

func (m *SensitiveWordMatcher) Matches(text string) bool

Matches reports whether the text contains any of the configured sensitive words.

func (*SensitiveWordMatcher) ObfuscateText

func (m *SensitiveWordMatcher) ObfuscateText(text string) string

ObfuscateText replaces all sensitive words in the text.

type StreamResponseModelObserver

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

StreamResponseModelObserver accumulates streaming chunks using a bounded buffer and extracts complete SSE lines or events to observe the response model, ensuring extraction does not depend on network chunk boundaries.

func NewStreamResponseModelObserver

func NewStreamResponseModelObserver(reporter *UsageReporter) *StreamResponseModelObserver

NewStreamResponseModelObserver creates an observer for tracking response models from arbitrary streaming network chunks.

func (*StreamResponseModelObserver) Feed

func (o *StreamResponseModelObserver) Feed(chunk []byte)

Feed ingests a raw chunk of bytes from a stream, splitting into lines and events.

func (*StreamResponseModelObserver) Finish

func (o *StreamResponseModelObserver) Finish()

Finish flushes any pending buffered line or event at the end of the stream.

type StreamUsageBuffer

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

StreamUsageBuffer keeps the latest usage detail observed in a stream.

func (*StreamUsageBuffer) Detail

func (b *StreamUsageBuffer) Detail() (usage.Detail, bool)

Detail returns the latest observed usage detail.

func (*StreamUsageBuffer) Observe

func (b *StreamUsageBuffer) Observe(detail usage.Detail, ok bool)

Observe records detail when ok is true, allowing the final stream usage to win.

func (*StreamUsageBuffer) ObserveClaudeStream

func (b *StreamUsageBuffer) ObserveClaudeStream(line []byte)

ObserveClaudeStream records and merges usage from a Claude SSE line.

func (*StreamUsageBuffer) ObserveOpenAIStream

func (b *StreamUsageBuffer) ObserveOpenAIStream(line []byte)

ObserveOpenAIStream records response-tier state and the latest usage from an OpenAI-style stream while avoiding JSON parsing for irrelevant chunks.

func (*StreamUsageBuffer) Publish

func (b *StreamUsageBuffer) Publish(ctx context.Context, reporter *UsageReporter) bool

Publish emits the latest observed usage detail, if any.

func (*StreamUsageBuffer) PublishFailure

func (b *StreamUsageBuffer) PublishFailure(ctx context.Context, reporter *UsageReporter, errs ...error) bool

PublishFailure emits the latest observed usage detail together with failure details.

func (*StreamUsageBuffer) ResponseModel

func (b *StreamUsageBuffer) ResponseModel() string

ResponseModel returns the latest model observed in the stream buffer.

type TransportCache

type TransportCache[K comparable] struct {
	// contains filtered or unexported fields
}

TransportCache memoizes HTTP transports under a comparable key using a bounded LRU. Evicting an entry closes its idle connections so neither the pool nor its background goroutines outlive the cache entry.

The key type is generic so callers can mix value identity (a normalized proxy URL) with pointer identity (a base transport supplied by the caller) without the cache retaining either beyond the LRU window.

func NewTransportCache

func NewTransportCache[K comparable](capacity int) *TransportCache[K]

NewTransportCache returns a cache holding at most capacity transports. A non-positive capacity falls back to DefaultTransportCacheCapacity.

func (*TransportCache[K]) CloseKey

func (c *TransportCache[K]) CloseKey(key K) bool

CloseKey removes the transport cached under key (if present) and closes its idle connections. Returns true if the entry was found and closed.

func (*TransportCache[K]) CloseMatching

func (c *TransportCache[K]) CloseMatching(predicate func(key K) bool) int

CloseMatching removes and closes idle connections for every cached transport whose key satisfies predicate. Returns the number of closed transports.

func (*TransportCache[K]) Contains

func (c *TransportCache[K]) Contains(key K) bool

Contains reports whether key exists in the cache.

func (*TransportCache[K]) Get

func (c *TransportCache[K]) Get(key K, build func() (*http.Transport, error)) (*http.Transport, error)

Get returns the transport cached under key, calling build on the first use of that key. Concurrent callers observe the same instance.

A build error is propagated without being cached, so a later call can retry and a failed lookup never occupies a cache slot. build must not call back into the same cache.

func (*TransportCache[K]) Len

func (c *TransportCache[K]) Len() int

Len reports how many transports the cache currently holds.

func (*TransportCache[K]) Purge

func (c *TransportCache[K]) Purge()

Purge drops every entry and closes the idle connections it was holding.

type UTF8SplitBuffer

type UTF8SplitBuffer struct {
	// contains filtered or unexported fields
}

UTF8SplitBuffer buffers incomplete UTF-8 byte sequences across chunk boundaries.

func (*UTF8SplitBuffer) Feed

func (b *UTF8SplitBuffer) Feed(chunk []byte) string

Feed consumes a byte chunk, prepending any pending remainder, and returns complete UTF-8 strings.

type UpstreamRequestLog

type UpstreamRequestLog struct {
	URL       string
	Method    string
	Headers   http.Header
	Body      []byte
	Provider  string
	AuthID    string
	AuthLabel string
	AuthType  string
	AuthValue string
}

UpstreamRequestLog captures the outbound upstream request details for logging.

type UsageReporter

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

func NewExecutorUsageReporter

func NewExecutorUsageReporter(ctx context.Context, executor usageExecutor, model string, auth *cliproxyauth.Auth) *UsageReporter

func NewUsageReporter

func NewUsageReporter(ctx context.Context, provider, model string, auth *cliproxyauth.Auth) *UsageReporter

func (*UsageReporter) EnsurePublished

func (r *UsageReporter) EnsurePublished(ctx context.Context)

ensurePublished guarantees that a usage record is emitted exactly once. It is safe to call multiple times; only the first call wins due to once.Do. This is used to ensure request counting even when upstream responses do not include any usage fields (tokens), especially for streaming paths.

func (*UsageReporter) IsFirstPacketSet

func (r *UsageReporter) IsFirstPacketSet() bool

func (*UsageReporter) IsResponseModelFinal

func (r *UsageReporter) IsResponseModelFinal() bool

IsResponseModelFinal reports whether the response model was already finalized by a terminal event.

func (*UsageReporter) IsTTFTSet

func (r *UsageReporter) IsTTFTSet() bool

func (*UsageReporter) MarkFirstResponseByte

func (r *UsageReporter) MarkFirstResponseByte()

func (*UsageReporter) ObserveCodexResponseModel

func (r *UsageReporter) ObserveCodexResponseModel(payload []byte)

ObserveCodexResponseModel stores the model reported by a codex upstream event and ignores payloads without one; the substitution warning is emitted at publish time.

func (*UsageReporter) ObserveResponse

func (r *UsageReporter) ObserveResponse(resp *http.Response)

func (*UsageReporter) ObserveResponseModel

func (r *UsageReporter) ObserveResponseModel(payload []byte)

ObserveResponseModel stores the model reported by an upstream response or event and ignores payloads without one; the substitution warning is emitted at publish time.

func (*UsageReporter) ObserveResponsePacketOnly

func (r *UsageReporter) ObserveResponsePacketOnly(resp *http.Response)

func (*UsageReporter) ObserveTokenEvent

func (r *UsageReporter) ObserveTokenEvent(isToken bool)

ObserveTokenEvent records the first packet fallback time on the first frame, and if isToken is true, records the effective TTFT. It uses a fast-path read check to return immediately with zero write lock contention once TTFT is set or when subsequent non-token metadata frames arrive after the first packet.

func (*UsageReporter) Publish

func (r *UsageReporter) Publish(ctx context.Context, detail usage.Detail)

func (*UsageReporter) PublishAdditionalModel

func (r *UsageReporter) PublishAdditionalModel(ctx context.Context, model string, detail usage.Detail)

func (*UsageReporter) PublishFailure

func (r *UsageReporter) PublishFailure(ctx context.Context, errs ...error)

func (*UsageReporter) PublishFailureWithDetail

func (r *UsageReporter) PublishFailureWithDetail(ctx context.Context, detail usage.Detail, errs ...error)

func (*UsageReporter) RecordFirstPacket

func (r *UsageReporter) RecordFirstPacket()

RecordFirstPacket records the arrival time of the first packet/chunk from upstream as a fallback.

func (*UsageReporter) RequestID

func (r *UsageReporter) RequestID() string

RequestID returns the execution instance request ID for this reporter.

func (*UsageReporter) ResponseModel

func (r *UsageReporter) ResponseModel() string

ResponseModel returns the latest model reported by the upstream response.

func (*UsageReporter) SetResponseModel

func (r *UsageReporter) SetResponseModel(model string)

SetResponseModel sets the reported model directly if valid and not already marked final.

func (*UsageReporter) SetSessionHierarchy

func (r *UsageReporter) SetSessionHierarchy(sessionID, parentSessionID string)

SetSessionHierarchy sets the explicit session and parent session identifiers. Callers should invoke this method before Publish or EnsurePublished on the request thread.

func (*UsageReporter) SetStream

func (r *UsageReporter) SetStream(stream bool)

SetStream records whether the request was executed in streaming mode.

func (*UsageReporter) SetTranslatedReasoningEffort

func (r *UsageReporter) SetTranslatedReasoningEffort(payload []byte, format string)

func (*UsageReporter) SetUpstreamModel

func (r *UsageReporter) SetUpstreamModel(model string)

SetUpstreamModel records the upstream model expected to be served when it differs from the requested model (e.g. due to provider-specific mapping or canonicalization). Model substitution detection compares the response against this upstream model, while usage accounting preserves the client's requested model.

func (*UsageReporter) StartResponseTTFT

func (r *UsageReporter) StartResponseTTFT()

func (*UsageReporter) TraceID

func (r *UsageReporter) TraceID() string

TraceID returns the parent inbound request ID for this reporter.

func (*UsageReporter) TrackFailure

func (r *UsageReporter) TrackFailure(ctx context.Context, errPtr *error)

func (*UsageReporter) TrackHTTPClient

func (r *UsageReporter) TrackHTTPClient(client *http.Client) *http.Client

func (*UsageReporter) TrackHTTPClientRoundTripOnly

func (r *UsageReporter) TrackHTTPClientRoundTripOnly(client *http.Client) *http.Client

TrackHTTPClientRoundTripOnly records the TTFT start time upon sending the request and captures first-packet arrival fallback on initial body reads, while keeping effective TTFT unset. This allows protocol-aware streaming executors (like Codex SSE) to mark effective TTFT explicitly upon receiving substantive token events, while preserving first-packet fallback metrics for non-2xx error bodies or keepalive streams.

func (*UsageReporter) UpdateAccessTokenFingerprint

func (r *UsageReporter) UpdateAccessTokenFingerprint(auth *cliproxyauth.Auth)

UpdateAccessTokenFingerprint records the token version actually used upstream.

func (*UsageReporter) UpstreamModel

func (r *UsageReporter) UpstreamModel() string

UpstreamModel returns the expected upstream model, or an empty string if not explicitly set.

Jump to

Keyboard shortcuts

? : This menu
/ : Search site
f or F : Jump to
y or Y : Canonical URL