Remove dead telemetry stubs

This commit is contained in:
2026-04-04 01:12:54 +08:00
parent 497f81f4f9
commit a95f0a540a
7 changed files with 137 additions and 2461 deletions

View File

@@ -1,806 +0,0 @@
import type { HrTime } from '@opentelemetry/api'
import { type ExportResult, ExportResultCode } from '@opentelemetry/core'
import type {
LogRecordExporter,
ReadableLogRecord,
} from '@opentelemetry/sdk-logs'
import axios from 'axios'
import { randomUUID } from 'crypto'
import { appendFile, mkdir, readdir, unlink, writeFile } from 'fs/promises'
import * as path from 'path'
import type { CoreUserData } from 'src/utils/user.js'
import {
getIsNonInteractiveSession,
getSessionId,
} from '../../bootstrap/state.js'
import { ClaudeCodeInternalEvent } from '../../types/generated/events_mono/claude_code/v1/claude_code_internal_event.js'
import { GrowthbookExperimentEvent } from '../../types/generated/events_mono/growthbook/v1/growthbook_experiment_event.js'
import {
getClaudeAIOAuthTokens,
hasProfileScope,
isClaudeAISubscriber,
} from '../../utils/auth.js'
import { checkHasTrustDialogAccepted } from '../../utils/config.js'
import { logForDebugging } from '../../utils/debug.js'
import { getClaudeConfigHomeDir } from '../../utils/envUtils.js'
import { errorMessage, isFsInaccessible, toError } from '../../utils/errors.js'
import { getAuthHeaders } from '../../utils/http.js'
import { readJSONLFile } from '../../utils/json.js'
import { logError } from '../../utils/log.js'
import { sleep } from '../../utils/sleep.js'
import { jsonStringify } from '../../utils/slowOperations.js'
import { getClaudeCodeUserAgent } from '../../utils/userAgent.js'
import { isOAuthTokenExpired } from '../oauth/client.js'
import { stripProtoFields } from './index.js'
import { type EventMetadata, to1PEventFormat } from './metadata.js'
// Unique ID for this process run - used to isolate failed event files between runs
const BATCH_UUID = randomUUID()
// File prefix for failed event storage
const FILE_PREFIX = '1p_failed_events.'
// Storage directory for failed events - evaluated at runtime to respect CLAUDE_CONFIG_DIR in tests
function getStorageDir(): string {
return path.join(getClaudeConfigHomeDir(), 'telemetry')
}
// API envelope - event_data is the JSON output from proto toJSON()
type FirstPartyEventLoggingEvent = {
event_type: 'ClaudeCodeInternalEvent' | 'GrowthbookExperimentEvent'
event_data: unknown
}
type FirstPartyEventLoggingPayload = {
events: FirstPartyEventLoggingEvent[]
}
/**
* Exporter for 1st-party event logging to /api/event_logging/batch.
*
* Export cycles are controlled by OpenTelemetry's BatchLogRecordProcessor, which
* triggers export() when either:
* - Time interval elapses (default: 5 seconds via scheduledDelayMillis)
* - Batch size is reached (default: 200 events via maxExportBatchSize)
*
* This exporter adds resilience on top:
* - Append-only log for failed events (concurrency-safe)
* - Quadratic backoff retry for failed events, dropped after maxAttempts
* - Immediate retry of queued events when any export succeeds (endpoint is healthy)
* - Chunking large event sets into smaller batches
* - Auth fallback: retries without auth on 401 errors
*/
export class FirstPartyEventLoggingExporter implements LogRecordExporter {
private readonly endpoint: string
private readonly timeout: number
private readonly maxBatchSize: number
private readonly skipAuth: boolean
private readonly batchDelayMs: number
private readonly baseBackoffDelayMs: number
private readonly maxBackoffDelayMs: number
private readonly maxAttempts: number
private readonly isKilled: () => boolean
private pendingExports: Promise<void>[] = []
private isShutdown = false
private readonly schedule: (
fn: () => Promise<void>,
delayMs: number,
) => () => void
private cancelBackoff: (() => void) | null = null
private attempts = 0
private isRetrying = false
private lastExportErrorContext: string | undefined
constructor(
options: {
timeout?: number
maxBatchSize?: number
skipAuth?: boolean
batchDelayMs?: number
baseBackoffDelayMs?: number
maxBackoffDelayMs?: number
maxAttempts?: number
path?: string
baseUrl?: string
// Injected killswitch probe. Checked per-POST so that disabling the
// firstParty sink also stops backoff retries (not just new emits).
// Passed in rather than imported to avoid a cycle with firstPartyEventLogger.ts.
isKilled?: () => boolean
schedule?: (fn: () => Promise<void>, delayMs: number) => () => void
} = {},
) {
// Default: prod, except when ANTHROPIC_BASE_URL is explicitly staging.
// Overridable via tengu_1p_event_batch_config.baseUrl.
const baseUrl =
options.baseUrl ||
(process.env.ANTHROPIC_BASE_URL === 'https://api-staging.anthropic.com'
? 'https://api-staging.anthropic.com'
: 'https://api.anthropic.com')
this.endpoint = `${baseUrl}${options.path || '/api/event_logging/batch'}`
this.timeout = options.timeout || 10000
this.maxBatchSize = options.maxBatchSize || 200
this.skipAuth = options.skipAuth ?? false
this.batchDelayMs = options.batchDelayMs || 100
this.baseBackoffDelayMs = options.baseBackoffDelayMs || 500
this.maxBackoffDelayMs = options.maxBackoffDelayMs || 30000
this.maxAttempts = options.maxAttempts ?? 8
this.isKilled = options.isKilled ?? (() => false)
this.schedule =
options.schedule ??
((fn, ms) => {
const t = setTimeout(fn, ms)
return () => clearTimeout(t)
})
// Retry any failed events from previous runs of this session (in background)
void this.retryPreviousBatches()
}
// Expose for testing
async getQueuedEventCount(): Promise<number> {
return (await this.loadEventsFromCurrentBatch()).length
}
// --- Storage helpers ---
private getCurrentBatchFilePath(): string {
return path.join(
getStorageDir(),
`${FILE_PREFIX}${getSessionId()}.${BATCH_UUID}.json`,
)
}
private async loadEventsFromFile(
filePath: string,
): Promise<FirstPartyEventLoggingEvent[]> {
try {
return await readJSONLFile<FirstPartyEventLoggingEvent>(filePath)
} catch {
return []
}
}
private async loadEventsFromCurrentBatch(): Promise<
FirstPartyEventLoggingEvent[]
> {
return this.loadEventsFromFile(this.getCurrentBatchFilePath())
}
private async saveEventsToFile(
filePath: string,
events: FirstPartyEventLoggingEvent[],
): Promise<void> {
try {
if (events.length === 0) {
try {
await unlink(filePath)
} catch {
// File doesn't exist, nothing to delete
}
} else {
// Ensure storage directory exists
await mkdir(getStorageDir(), { recursive: true })
// Write as JSON lines (one event per line)
const content = events.map(e => jsonStringify(e)).join('\n') + '\n'
await writeFile(filePath, content, 'utf8')
}
} catch (error) {
logError(error)
}
}
private async appendEventsToFile(
filePath: string,
events: FirstPartyEventLoggingEvent[],
): Promise<void> {
if (events.length === 0) return
try {
// Ensure storage directory exists
await mkdir(getStorageDir(), { recursive: true })
// Append as JSON lines (one event per line) - atomic on most filesystems
const content = events.map(e => jsonStringify(e)).join('\n') + '\n'
await appendFile(filePath, content, 'utf8')
} catch (error) {
logError(error)
}
}
private async deleteFile(filePath: string): Promise<void> {
try {
await unlink(filePath)
} catch {
// File doesn't exist or can't be deleted, ignore
}
}
// --- Previous batch retry (startup) ---
private async retryPreviousBatches(): Promise<void> {
try {
const prefix = `${FILE_PREFIX}${getSessionId()}.`
let files: string[]
try {
files = (await readdir(getStorageDir()))
.filter((f: string) => f.startsWith(prefix) && f.endsWith('.json'))
.filter((f: string) => !f.includes(BATCH_UUID)) // Exclude current batch
} catch (e) {
if (isFsInaccessible(e)) return
throw e
}
for (const file of files) {
const filePath = path.join(getStorageDir(), file)
void this.retryFileInBackground(filePath)
}
} catch (error) {
logError(error)
}
}
private async retryFileInBackground(filePath: string): Promise<void> {
if (this.attempts >= this.maxAttempts) {
await this.deleteFile(filePath)
return
}
const events = await this.loadEventsFromFile(filePath)
if (events.length === 0) {
await this.deleteFile(filePath)
return
}
if (process.env.USER_TYPE === 'ant') {
logForDebugging(
`1P event logging: retrying ${events.length} events from previous batch`,
)
}
const failedEvents = await this.sendEventsInBatches(events)
if (failedEvents.length === 0) {
await this.deleteFile(filePath)
if (process.env.USER_TYPE === 'ant') {
logForDebugging('1P event logging: previous batch retry succeeded')
}
} else {
// Save only the failed events back (not all original events)
await this.saveEventsToFile(filePath, failedEvents)
if (process.env.USER_TYPE === 'ant') {
logForDebugging(
`1P event logging: previous batch retry failed, ${failedEvents.length} events remain`,
)
}
}
}
async export(
logs: ReadableLogRecord[],
resultCallback: (result: ExportResult) => void,
): Promise<void> {
if (this.isShutdown) {
if (process.env.USER_TYPE === 'ant') {
logForDebugging(
'1P event logging export failed: Exporter has been shutdown',
)
}
resultCallback({
code: ExportResultCode.FAILED,
error: new Error('Exporter has been shutdown'),
})
return
}
const exportPromise = this.doExport(logs, resultCallback)
this.pendingExports.push(exportPromise)
// Clean up completed exports
void exportPromise.finally(() => {
const index = this.pendingExports.indexOf(exportPromise)
if (index > -1) {
void this.pendingExports.splice(index, 1)
}
})
}
private async doExport(
logs: ReadableLogRecord[],
resultCallback: (result: ExportResult) => void,
): Promise<void> {
try {
// Filter for event logs only (by scope name)
const eventLogs = logs.filter(
log =>
log.instrumentationScope?.name === 'com.anthropic.claude_code.events',
)
if (eventLogs.length === 0) {
resultCallback({ code: ExportResultCode.SUCCESS })
return
}
// Transform new logs (failed events are retried independently via backoff)
const events = this.transformLogsToEvents(eventLogs).events
if (events.length === 0) {
resultCallback({ code: ExportResultCode.SUCCESS })
return
}
if (this.attempts >= this.maxAttempts) {
resultCallback({
code: ExportResultCode.FAILED,
error: new Error(
`Dropped ${events.length} events: max attempts (${this.maxAttempts}) reached`,
),
})
return
}
// Send events
const failedEvents = await this.sendEventsInBatches(events)
this.attempts++
if (failedEvents.length > 0) {
await this.queueFailedEvents(failedEvents)
this.scheduleBackoffRetry()
const context = this.lastExportErrorContext
? ` (${this.lastExportErrorContext})`
: ''
resultCallback({
code: ExportResultCode.FAILED,
error: new Error(
`Failed to export ${failedEvents.length} events${context}`,
),
})
return
}
// Success - reset backoff and immediately retry any queued events
this.resetBackoff()
if ((await this.getQueuedEventCount()) > 0 && !this.isRetrying) {
void this.retryFailedEvents()
}
resultCallback({ code: ExportResultCode.SUCCESS })
} catch (error) {
if (process.env.USER_TYPE === 'ant') {
logForDebugging(
`1P event logging export failed: ${errorMessage(error)}`,
)
}
logError(error)
resultCallback({
code: ExportResultCode.FAILED,
error: toError(error),
})
}
}
private async sendEventsInBatches(
events: FirstPartyEventLoggingEvent[],
): Promise<FirstPartyEventLoggingEvent[]> {
// Chunk events into batches
const batches: FirstPartyEventLoggingEvent[][] = []
for (let i = 0; i < events.length; i += this.maxBatchSize) {
batches.push(events.slice(i, i + this.maxBatchSize))
}
if (process.env.USER_TYPE === 'ant') {
logForDebugging(
`1P event logging: exporting ${events.length} events in ${batches.length} batch(es)`,
)
}
// Send each batch with delay between them. On first failure, assume the
// endpoint is down and short-circuit: queue the failed batch plus all
// remaining unsent batches without POSTing them. The backoff retry will
// probe again with a single batch next tick.
const failedBatchEvents: FirstPartyEventLoggingEvent[] = []
let lastErrorContext: string | undefined
for (let i = 0; i < batches.length; i++) {
const batch = batches[i]!
try {
await this.sendBatchWithRetry({ events: batch })
} catch (error) {
lastErrorContext = getAxiosErrorContext(error)
for (let j = i; j < batches.length; j++) {
failedBatchEvents.push(...batches[j]!)
}
if (process.env.USER_TYPE === 'ant') {
const skipped = batches.length - 1 - i
logForDebugging(
`1P event logging: batch ${i + 1}/${batches.length} failed (${lastErrorContext}); short-circuiting ${skipped} remaining batch(es)`,
)
}
break
}
if (i < batches.length - 1 && this.batchDelayMs > 0) {
await sleep(this.batchDelayMs)
}
}
if (failedBatchEvents.length > 0 && lastErrorContext) {
this.lastExportErrorContext = lastErrorContext
}
return failedBatchEvents
}
private async queueFailedEvents(
events: FirstPartyEventLoggingEvent[],
): Promise<void> {
const filePath = this.getCurrentBatchFilePath()
// Append-only: just add new events to file (atomic on most filesystems)
await this.appendEventsToFile(filePath, events)
const context = this.lastExportErrorContext
? ` (${this.lastExportErrorContext})`
: ''
const message = `1P event logging: ${events.length} events failed to export${context}`
logError(new Error(message))
}
private scheduleBackoffRetry(): void {
// Don't schedule if already retrying or shutdown
if (this.cancelBackoff || this.isRetrying || this.isShutdown) {
return
}
// Quadratic backoff (matching Statsig SDK): base * attempts²
const delay = Math.min(
this.baseBackoffDelayMs * this.attempts * this.attempts,
this.maxBackoffDelayMs,
)
if (process.env.USER_TYPE === 'ant') {
logForDebugging(
`1P event logging: scheduling backoff retry in ${delay}ms (attempt ${this.attempts})`,
)
}
this.cancelBackoff = this.schedule(async () => {
this.cancelBackoff = null
await this.retryFailedEvents()
}, delay)
}
private async retryFailedEvents(): Promise<void> {
const filePath = this.getCurrentBatchFilePath()
// Keep retrying while there are events and endpoint is healthy
while (!this.isShutdown) {
const events = await this.loadEventsFromFile(filePath)
if (events.length === 0) break
if (this.attempts >= this.maxAttempts) {
if (process.env.USER_TYPE === 'ant') {
logForDebugging(
`1P event logging: max attempts (${this.maxAttempts}) reached, dropping ${events.length} events`,
)
}
await this.deleteFile(filePath)
this.resetBackoff()
return
}
this.isRetrying = true
// Clear file before retry (we have events in memory now)
await this.deleteFile(filePath)
if (process.env.USER_TYPE === 'ant') {
logForDebugging(
`1P event logging: retrying ${events.length} failed events (attempt ${this.attempts + 1})`,
)
}
const failedEvents = await this.sendEventsInBatches(events)
this.attempts++
this.isRetrying = false
if (failedEvents.length > 0) {
// Write failures back to disk
await this.saveEventsToFile(filePath, failedEvents)
this.scheduleBackoffRetry()
return // Failed - wait for backoff
}
// Success - reset backoff and continue loop to drain any newly queued events
this.resetBackoff()
if (process.env.USER_TYPE === 'ant') {
logForDebugging('1P event logging: backoff retry succeeded')
}
}
}
private resetBackoff(): void {
this.attempts = 0
if (this.cancelBackoff) {
this.cancelBackoff()
this.cancelBackoff = null
}
}
private async sendBatchWithRetry(
payload: FirstPartyEventLoggingPayload,
): Promise<void> {
if (this.isKilled()) {
// Throw so the caller short-circuits remaining batches and queues
// everything to disk. Zero network traffic while killed; the backoff
// timer keeps ticking and will resume POSTs as soon as the GrowthBook
// cache picks up the cleared flag.
throw new Error('firstParty sink killswitch active')
}
const baseHeaders: Record<string, string> = {
'Content-Type': 'application/json',
'User-Agent': getClaudeCodeUserAgent(),
'x-service-name': 'claude-code',
}
// Skip auth if trust hasn't been established yet
// This prevents executing apiKeyHelper commands before the trust dialog
// Non-interactive sessions implicitly have workspace trust
const hasTrust =
checkHasTrustDialogAccepted() || getIsNonInteractiveSession()
if (process.env.USER_TYPE === 'ant' && !hasTrust) {
logForDebugging('1P event logging: Trust not accepted')
}
// Skip auth when the OAuth token is expired or lacks user:profile
// scope (service key sessions). Falls through to unauthenticated send.
let shouldSkipAuth = this.skipAuth || !hasTrust
if (!shouldSkipAuth && isClaudeAISubscriber()) {
const tokens = getClaudeAIOAuthTokens()
if (!hasProfileScope()) {
shouldSkipAuth = true
} else if (tokens && isOAuthTokenExpired(tokens.expiresAt)) {
shouldSkipAuth = true
if (process.env.USER_TYPE === 'ant') {
logForDebugging(
'1P event logging: OAuth token expired, skipping auth to avoid 401',
)
}
}
}
// Try with auth headers first (unless trust not established or token is known to be expired)
const authResult = shouldSkipAuth
? { headers: {}, error: 'trust not established or Oauth token expired' }
: getAuthHeaders()
const useAuth = !authResult.error
if (!useAuth && process.env.USER_TYPE === 'ant') {
logForDebugging(
`1P event logging: auth not available, sending without auth`,
)
}
const headers = useAuth
? { ...baseHeaders, ...authResult.headers }
: baseHeaders
try {
const response = await axios.post(this.endpoint, payload, {
timeout: this.timeout,
headers,
})
this.logSuccess(payload.events.length, useAuth, response.data)
return
} catch (error) {
// Handle 401 by retrying without auth
if (
useAuth &&
axios.isAxiosError(error) &&
error.response?.status === 401
) {
if (process.env.USER_TYPE === 'ant') {
logForDebugging(
'1P event logging: 401 auth error, retrying without auth',
)
}
const response = await axios.post(this.endpoint, payload, {
timeout: this.timeout,
headers: baseHeaders,
})
this.logSuccess(payload.events.length, false, response.data)
return
}
throw error
}
}
private logSuccess(
eventCount: number,
withAuth: boolean,
responseData: unknown,
): void {
if (process.env.USER_TYPE === 'ant') {
logForDebugging(
`1P event logging: ${eventCount} events exported successfully${withAuth ? ' (with auth)' : ' (without auth)'}`,
)
logForDebugging(`API Response: ${jsonStringify(responseData, null, 2)}`)
}
}
private hrTimeToDate(hrTime: HrTime): Date {
const [seconds, nanoseconds] = hrTime
return new Date(seconds * 1000 + nanoseconds / 1000000)
}
private transformLogsToEvents(
logs: ReadableLogRecord[],
): FirstPartyEventLoggingPayload {
const events: FirstPartyEventLoggingEvent[] = []
for (const log of logs) {
const attributes = log.attributes || {}
// Check if this is a GrowthBook experiment event
if (attributes.event_type === 'GrowthbookExperimentEvent') {
const timestamp = this.hrTimeToDate(log.hrTime)
const account_uuid = attributes.account_uuid as string | undefined
const organization_uuid = attributes.organization_uuid as
| string
| undefined
events.push({
event_type: 'GrowthbookExperimentEvent',
event_data: GrowthbookExperimentEvent.toJSON({
event_id: attributes.event_id as string,
timestamp,
experiment_id: attributes.experiment_id as string,
variation_id: attributes.variation_id as number,
environment: attributes.environment as string,
user_attributes: attributes.user_attributes as string,
experiment_metadata: attributes.experiment_metadata as string,
device_id: attributes.device_id as string,
session_id: attributes.session_id as string,
auth:
account_uuid || organization_uuid
? { account_uuid, organization_uuid }
: undefined,
}),
})
continue
}
// Extract event name
const eventName =
(attributes.event_name as string) || (log.body as string) || 'unknown'
// Extract metadata objects directly (no JSON parsing needed)
const coreMetadata = attributes.core_metadata as EventMetadata | undefined
const userMetadata = attributes.user_metadata as CoreUserData
const eventMetadata = (attributes.event_metadata || {}) as Record<
string,
unknown
>
if (!coreMetadata) {
// Emit partial event if core metadata is missing
if (process.env.USER_TYPE === 'ant') {
logForDebugging(
`1P event logging: core_metadata missing for event ${eventName}`,
)
}
events.push({
event_type: 'ClaudeCodeInternalEvent',
event_data: ClaudeCodeInternalEvent.toJSON({
event_id: attributes.event_id as string | undefined,
event_name: eventName,
client_timestamp: this.hrTimeToDate(log.hrTime),
session_id: getSessionId(),
additional_metadata: Buffer.from(
jsonStringify({
transform_error: 'core_metadata attribute is missing',
}),
).toString('base64'),
}),
})
continue
}
// Transform to 1P format
const formatted = to1PEventFormat(
coreMetadata,
userMetadata,
eventMetadata,
)
// _PROTO_* keys are PII-tagged values meant only for privileged BQ
// columns. Hoist known keys to proto fields, then defensively strip any
// remaining _PROTO_* so an unrecognized future key can't silently land
// in the general-access additional_metadata blob. sink.ts applies the
// same strip before Datadog; this closes the 1P side.
const {
_PROTO_skill_name,
_PROTO_plugin_name,
_PROTO_marketplace_name,
...rest
} = formatted.additional
const additionalMetadata = stripProtoFields(rest)
events.push({
event_type: 'ClaudeCodeInternalEvent',
event_data: ClaudeCodeInternalEvent.toJSON({
event_id: attributes.event_id as string | undefined,
event_name: eventName,
client_timestamp: this.hrTimeToDate(log.hrTime),
device_id: attributes.user_id as string | undefined,
email: userMetadata?.email,
auth: formatted.auth,
...formatted.core,
env: formatted.env,
process: formatted.process,
skill_name:
typeof _PROTO_skill_name === 'string'
? _PROTO_skill_name
: undefined,
plugin_name:
typeof _PROTO_plugin_name === 'string'
? _PROTO_plugin_name
: undefined,
marketplace_name:
typeof _PROTO_marketplace_name === 'string'
? _PROTO_marketplace_name
: undefined,
additional_metadata:
Object.keys(additionalMetadata).length > 0
? Buffer.from(jsonStringify(additionalMetadata)).toString(
'base64',
)
: undefined,
}),
})
}
return { events }
}
async shutdown(): Promise<void> {
this.isShutdown = true
this.resetBackoff()
await this.forceFlush()
if (process.env.USER_TYPE === 'ant') {
logForDebugging('1P event logging exporter shutdown complete')
}
}
async forceFlush(): Promise<void> {
await Promise.all(this.pendingExports)
if (process.env.USER_TYPE === 'ant') {
logForDebugging('1P event logging exporter flush complete')
}
}
}
function getAxiosErrorContext(error: unknown): string {
if (!axios.isAxiosError(error)) {
return errorMessage(error)
}
const parts: string[] = []
const requestId = error.response?.headers?.['request-id']
if (requestId) {
parts.push(`request-id=${requestId}`)
}
if (error.response?.status) {
parts.push(`status=${error.response.status}`)
}
if (error.code) {
parts.push(`code=${error.code}`)
}
if (error.message) {
parts.push(error.message)
}
return parts.join(', ')
}

View File

@@ -1,105 +1,33 @@
/**
* Beta Session Tracing for Claude Code
* Detailed beta tracing egress is disabled in this build.
*
* This module contains beta tracing features enabled when
* ENABLE_BETA_TRACING_DETAILED=1 and BETA_TRACING_ENDPOINT are set.
*
* For external users, tracing is enabled in SDK/headless mode, or in
* interactive mode when the org is allowlisted via the
* tengu_trace_lantern GrowthBook gate.
* For ant users, tracing is enabled in all modes.
*
* Visibility Rules:
* | Content | External | Ant |
* |------------------|----------|------|
* | System prompts | ✅ | ✅ |
* | Model output | ✅ | ✅ |
* | Thinking output | ❌ | ✅ |
* | Tools | ✅ | ✅ |
* | new_context | ✅ | ✅ |
*
* Features:
* - Per-agent message tracking with hash-based deduplication
* - System prompt logging (once per unique hash)
* - Hook execution spans
* - Detailed new_context attributes for LLM requests
* The exported helpers remain for compile-time compatibility, but do not
* retain tracing state or emit tracing attributes.
*/
import type { Span } from '@opentelemetry/api'
import { createHash } from 'crypto'
import { getIsNonInteractiveSession } from '../../bootstrap/state.js'
import { getFeatureValue_CACHED_MAY_BE_STALE } from '../../services/analytics/growthbook.js'
import { sanitizeToolNameForAnalytics } from '../../services/analytics/metadata.js'
import type { AssistantMessage, UserMessage } from '../../types/message.js'
import { isEnvTruthy } from '../envUtils.js'
import { jsonParse, jsonStringify } from '../slowOperations.js'
import { logOTelEvent } from './events.js'
type AttributeValue = string | number | boolean
// Message type for API calls (UserMessage or AssistantMessage)
type APIMessage = UserMessage | AssistantMessage
export interface SpanAttributeWriter {
setAttribute?(_key: string, _value: AttributeValue): void
setAttributes?(_attributes: Record<string, AttributeValue>): void
}
/**
* Track hashes we've already logged this session (system prompts, tools, etc).
*
* WHY: System prompts and tool schemas are large and rarely change within a session.
* Sending full content on every request would be wasteful. Instead, we hash and
* only log the full content once per unique hash.
*/
const seenHashes = new Set<string>()
export interface LLMRequestNewContext {
systemPrompt?: string
querySource?: string
tools?: string
}
/**
* Track the last reported message hash per querySource (agent) for incremental context.
*
* WHY: When debugging traces, we want to see what NEW information was added each turn,
* not the entire conversation history (which can be huge). By tracking the last message
* we reported per agent, we can compute and send only the delta (new messages since
* the last request). This is tracked per-agent (querySource) because different agents
* (main thread, subagents, warmup requests) have independent conversation contexts.
*/
const lastReportedMessageHash = new Map<string, string>()
const MAX_CONTENT_SIZE = 60 * 1024
/**
* Clear tracking state after compaction.
* Old hashes are irrelevant once messages have been replaced.
*/
export function clearBetaTracingState(): void {
seenHashes.clear()
lastReportedMessageHash.clear()
return
}
const MAX_CONTENT_SIZE = 60 * 1024 // 60KB (Honeycomb limit is 64KB, staying safe)
/**
* Check if beta detailed tracing is enabled.
* - Requires ENABLE_BETA_TRACING_DETAILED=1 and BETA_TRACING_ENDPOINT
* - For external users, enabled in SDK/headless mode OR when org is
* allowlisted via the tengu_trace_lantern GrowthBook gate
*/
export function isBetaTracingEnabled(): boolean {
const baseEnabled =
isEnvTruthy(process.env.ENABLE_BETA_TRACING_DETAILED) &&
Boolean(process.env.BETA_TRACING_ENDPOINT)
if (!baseEnabled) {
return false
}
// For external users, enable in SDK/headless mode OR when org is allowlisted.
// Gate reads from disk cache, so first run after allowlisting returns false;
// works from second run onward (same behavior as enhanced_telemetry_beta).
if (process.env.USER_TYPE !== 'ant') {
return (
getIsNonInteractiveSession() ||
getFeatureValue_CACHED_MAY_BE_STALE('tengu_trace_lantern', false)
)
}
return true
return false
}
/**
* Truncate content to fit within Honeycomb limits.
*/
export function truncateContent(
content: string,
maxSize: number = MAX_CONTENT_SIZE,
@@ -116,376 +44,43 @@ export function truncateContent(
}
}
/**
* Generate a short hash (first 12 hex chars of SHA-256).
*/
function shortHash(content: string): string {
return createHash('sha256').update(content).digest('hex').slice(0, 12)
}
/**
* Generate a hash for a system prompt.
*/
function hashSystemPrompt(systemPrompt: string): string {
return `sp_${shortHash(systemPrompt)}`
}
/**
* Generate a hash for a message based on its content.
*/
function hashMessage(message: APIMessage): string {
const content = jsonStringify(message.message.content)
return `msg_${shortHash(content)}`
}
// Regex to detect content wrapped in <system-reminder> tags
const SYSTEM_REMINDER_REGEX =
/^<system-reminder>\n?([\s\S]*?)\n?<\/system-reminder>$/
/**
* Check if text is entirely a system reminder (wrapped in <system-reminder> tags).
* Returns the inner content if it is, null otherwise.
*/
function extractSystemReminderContent(text: string): string | null {
const match = text.trim().match(SYSTEM_REMINDER_REGEX)
return match && match[1] ? match[1].trim() : null
}
/**
* Result of formatting messages - separates regular content from system reminders.
*/
interface FormattedMessages {
contextParts: string[]
systemReminders: string[]
}
/**
* Format user messages for new_context display, separating system reminders.
* Only handles user messages (assistant messages are filtered out before this is called).
*/
function formatMessagesForContext(messages: UserMessage[]): FormattedMessages {
const contextParts: string[] = []
const systemReminders: string[] = []
for (const message of messages) {
const content = message.message.content
if (typeof content === 'string') {
const reminderContent = extractSystemReminderContent(content)
if (reminderContent) {
systemReminders.push(reminderContent)
} else {
contextParts.push(`[USER]\n${content}`)
}
} else if (Array.isArray(content)) {
for (const block of content) {
if (block.type === 'text') {
const reminderContent = extractSystemReminderContent(block.text)
if (reminderContent) {
systemReminders.push(reminderContent)
} else {
contextParts.push(`[USER]\n${block.text}`)
}
} else if (block.type === 'tool_result') {
const resultContent =
typeof block.content === 'string'
? block.content
: jsonStringify(block.content)
// Tool results can also contain system reminders (e.g., malware warning)
const reminderContent = extractSystemReminderContent(resultContent)
if (reminderContent) {
systemReminders.push(reminderContent)
} else {
contextParts.push(
`[TOOL RESULT: ${block.tool_use_id}]\n${resultContent}`,
)
}
}
}
}
}
return { contextParts, systemReminders }
}
export interface LLMRequestNewContext {
/** System prompt (typically only on first request or if changed) */
systemPrompt?: string
/** Query source identifying the agent/purpose (e.g., 'repl_main_thread', 'agent:builtin') */
querySource?: string
/** Tool schemas sent with the request */
tools?: string
}
/**
* Add beta attributes to an interaction span.
* Adds new_context with the user prompt.
*/
export function addBetaInteractionAttributes(
span: Span,
userPrompt: string,
_span: SpanAttributeWriter,
_userPrompt: string,
): void {
if (!isBetaTracingEnabled()) {
return
}
const { content: truncatedPrompt, truncated } = truncateContent(
`[USER PROMPT]\n${userPrompt}`,
)
span.setAttributes({
new_context: truncatedPrompt,
...(truncated && {
new_context_truncated: true,
new_context_original_length: userPrompt.length,
}),
})
return
}
/**
* Add beta attributes to an LLM request span.
* Handles system prompt logging and new_context computation.
*/
export function addBetaLLMRequestAttributes(
span: Span,
newContext?: LLMRequestNewContext,
messagesForAPI?: APIMessage[],
_span: SpanAttributeWriter,
_newContext?: LLMRequestNewContext,
_messagesForAPI?: unknown[],
): void {
if (!isBetaTracingEnabled()) {
return
}
// Add system prompt info to the span
if (newContext?.systemPrompt) {
const promptHash = hashSystemPrompt(newContext.systemPrompt)
const preview = newContext.systemPrompt.slice(0, 500)
// Always add hash, preview, and length to the span
span.setAttribute('system_prompt_hash', promptHash)
span.setAttribute('system_prompt_preview', preview)
span.setAttribute('system_prompt_length', newContext.systemPrompt.length)
// Log the full system prompt only once per unique hash this session
if (!seenHashes.has(promptHash)) {
seenHashes.add(promptHash)
// Truncate for the log if needed
const { content: truncatedPrompt, truncated } = truncateContent(
newContext.systemPrompt,
)
void logOTelEvent('system_prompt', {
system_prompt_hash: promptHash,
system_prompt: truncatedPrompt,
system_prompt_length: String(newContext.systemPrompt.length),
...(truncated && { system_prompt_truncated: 'true' }),
})
}
}
// Add tools info to the span
if (newContext?.tools) {
try {
const toolsArray = jsonParse(newContext.tools) as Record<
string,
unknown
>[]
// Build array of {name, hash} for each tool
const toolsWithHashes = toolsArray.map(tool => {
const toolJson = jsonStringify(tool)
const toolHash = shortHash(toolJson)
return {
name: typeof tool.name === 'string' ? tool.name : 'unknown',
hash: toolHash,
json: toolJson,
}
})
// Set span attribute with array of name/hash pairs
span.setAttribute(
'tools',
jsonStringify(
toolsWithHashes.map(({ name, hash }) => ({ name, hash })),
),
)
span.setAttribute('tools_count', toolsWithHashes.length)
// Log each tool's full description once per unique hash
for (const { name, hash, json } of toolsWithHashes) {
if (!seenHashes.has(`tool_${hash}`)) {
seenHashes.add(`tool_${hash}`)
const { content: truncatedTool, truncated } = truncateContent(json)
void logOTelEvent('tool', {
tool_name: sanitizeToolNameForAnalytics(name),
tool_hash: hash,
tool: truncatedTool,
...(truncated && { tool_truncated: 'true' }),
})
}
}
} catch {
// If parsing fails, log the raw tools string
span.setAttribute('tools_parse_error', true)
}
}
// Add new_context using hash-based tracking (visible to all users)
if (messagesForAPI && messagesForAPI.length > 0 && newContext?.querySource) {
const querySource = newContext.querySource
const lastHash = lastReportedMessageHash.get(querySource)
// Find where the last reported message is in the array
let startIndex = 0
if (lastHash) {
for (let i = 0; i < messagesForAPI.length; i++) {
const msg = messagesForAPI[i]
if (msg && hashMessage(msg) === lastHash) {
startIndex = i + 1 // Start after the last reported message
break
}
}
// If lastHash not found, startIndex stays 0 (send everything)
}
// Get new messages (filter out assistant messages - we only want user input/tool results)
const newMessages = messagesForAPI
.slice(startIndex)
.filter((m): m is UserMessage => m.type === 'user')
if (newMessages.length > 0) {
// Format new messages, separating system reminders from regular content
const { contextParts, systemReminders } =
formatMessagesForContext(newMessages)
// Set new_context (regular user content and tool results)
if (contextParts.length > 0) {
const fullContext = contextParts.join('\n\n---\n\n')
const { content: truncatedContext, truncated } =
truncateContent(fullContext)
span.setAttributes({
new_context: truncatedContext,
new_context_message_count: newMessages.length,
...(truncated && {
new_context_truncated: true,
new_context_original_length: fullContext.length,
}),
})
}
// Set system_reminders as a separate attribute
if (systemReminders.length > 0) {
const fullReminders = systemReminders.join('\n\n---\n\n')
const { content: truncatedReminders, truncated: remindersTruncated } =
truncateContent(fullReminders)
span.setAttributes({
system_reminders: truncatedReminders,
system_reminders_count: systemReminders.length,
...(remindersTruncated && {
system_reminders_truncated: true,
system_reminders_original_length: fullReminders.length,
}),
})
}
// Update last reported hash to the last message in the array
const lastMessage = messagesForAPI[messagesForAPI.length - 1]
if (lastMessage) {
lastReportedMessageHash.set(querySource, hashMessage(lastMessage))
}
}
}
return
}
/**
* Add beta attributes to endLLMRequestSpan.
* Handles model_output and thinking_output truncation.
*/
export function addBetaLLMResponseAttributes(
endAttributes: Record<string, string | number | boolean>,
metadata?: {
_attributes: Record<string, AttributeValue>,
_metadata?: {
modelOutput?: string
thinkingOutput?: string
},
): void {
if (!isBetaTracingEnabled() || !metadata) {
return
}
// Add model_output (text content) - visible to all users
if (metadata.modelOutput !== undefined) {
const { content: modelOutput, truncated: outputTruncated } =
truncateContent(metadata.modelOutput)
endAttributes['response.model_output'] = modelOutput
if (outputTruncated) {
endAttributes['response.model_output_truncated'] = true
endAttributes['response.model_output_original_length'] =
metadata.modelOutput.length
}
}
// Add thinking_output - ant-only
if (
process.env.USER_TYPE === 'ant' &&
metadata.thinkingOutput !== undefined
) {
const { content: thinkingOutput, truncated: thinkingTruncated } =
truncateContent(metadata.thinkingOutput)
endAttributes['response.thinking_output'] = thinkingOutput
if (thinkingTruncated) {
endAttributes['response.thinking_output_truncated'] = true
endAttributes['response.thinking_output_original_length'] =
metadata.thinkingOutput.length
}
}
return
}
/**
* Add beta attributes to startToolSpan.
* Adds tool_input with the serialized tool input.
*/
export function addBetaToolInputAttributes(
span: Span,
toolName: string,
toolInput: string,
_span: SpanAttributeWriter,
_toolName: string,
_toolInput: string,
): void {
if (!isBetaTracingEnabled()) {
return
}
const { content: truncatedInput, truncated } = truncateContent(
`[TOOL INPUT: ${toolName}]\n${toolInput}`,
)
span.setAttributes({
tool_input: truncatedInput,
...(truncated && {
tool_input_truncated: true,
tool_input_original_length: toolInput.length,
}),
})
return
}
/**
* Add beta attributes to endToolSpan.
* Adds new_context with the tool result.
*/
export function addBetaToolResultAttributes(
endAttributes: Record<string, string | number | boolean>,
toolName: string | number | boolean,
toolResult: string,
_attributes: Record<string, AttributeValue>,
_toolName: string | number | boolean,
_toolResult: string,
): void {
if (!isBetaTracingEnabled()) {
return
}
const { content: truncatedResult, truncated } = truncateContent(
`[TOOL RESULT: ${toolName}]\n${toolResult}`,
)
endAttributes['new_context'] = truncatedResult
if (truncated) {
endAttributes['new_context_truncated'] = true
endAttributes['new_context_original_length'] = toolResult.length
}
return
}

View File

@@ -1,252 +0,0 @@
import type { Attributes, HrTime } from '@opentelemetry/api'
import { type ExportResult, ExportResultCode } from '@opentelemetry/core'
import {
AggregationTemporality,
type MetricData,
type DataPoint as OTelDataPoint,
type PushMetricExporter,
type ResourceMetrics,
} from '@opentelemetry/sdk-metrics'
import axios from 'axios'
import { checkMetricsEnabled } from 'src/services/api/metricsOptOut.js'
import { getIsNonInteractiveSession } from '../../bootstrap/state.js'
import { getSubscriptionType, isClaudeAISubscriber } from '../auth.js'
import { checkHasTrustDialogAccepted } from '../config.js'
import { logForDebugging } from '../debug.js'
import { errorMessage, toError } from '../errors.js'
import { getAuthHeaders } from '../http.js'
import { logError } from '../log.js'
import { jsonStringify } from '../slowOperations.js'
import { getClaudeCodeUserAgent } from '../userAgent.js'
type DataPoint = {
attributes: Record<string, string>
value: number
timestamp: string
}
type Metric = {
name: string
description?: string
unit?: string
data_points: DataPoint[]
}
type InternalMetricsPayload = {
resource_attributes: Record<string, string>
metrics: Metric[]
}
export class BigQueryMetricsExporter implements PushMetricExporter {
private readonly endpoint: string
private readonly timeout: number
private pendingExports: Promise<void>[] = []
private isShutdown = false
constructor(options: { timeout?: number } = {}) {
const defaultEndpoint = 'https://api.anthropic.com/api/claude_code/metrics'
if (
process.env.USER_TYPE === 'ant' &&
process.env.ANT_CLAUDE_CODE_METRICS_ENDPOINT
) {
this.endpoint =
process.env.ANT_CLAUDE_CODE_METRICS_ENDPOINT +
'/api/claude_code/metrics'
} else {
this.endpoint = defaultEndpoint
}
this.timeout = options.timeout || 5000
}
async export(
metrics: ResourceMetrics,
resultCallback: (result: ExportResult) => void,
): Promise<void> {
if (this.isShutdown) {
resultCallback({
code: ExportResultCode.FAILED,
error: new Error('Exporter has been shutdown'),
})
return
}
const exportPromise = this.doExport(metrics, resultCallback)
this.pendingExports.push(exportPromise)
// Clean up completed exports
void exportPromise.finally(() => {
const index = this.pendingExports.indexOf(exportPromise)
if (index > -1) {
void this.pendingExports.splice(index, 1)
}
})
}
private async doExport(
metrics: ResourceMetrics,
resultCallback: (result: ExportResult) => void,
): Promise<void> {
try {
// Skip if trust not established in interactive mode
// This prevents triggering apiKeyHelper before trust dialog
const hasTrust =
checkHasTrustDialogAccepted() || getIsNonInteractiveSession()
if (!hasTrust) {
logForDebugging(
'BigQuery metrics export: trust not established, skipping',
)
resultCallback({ code: ExportResultCode.SUCCESS })
return
}
// Check organization-level metrics opt-out
const metricsStatus = await checkMetricsEnabled()
if (!metricsStatus.enabled) {
logForDebugging('Metrics export disabled by organization setting')
resultCallback({ code: ExportResultCode.SUCCESS })
return
}
const payload = this.transformMetricsForInternal(metrics)
const authResult = getAuthHeaders()
if (authResult.error) {
logForDebugging(`Metrics export failed: ${authResult.error}`)
resultCallback({
code: ExportResultCode.FAILED,
error: new Error(authResult.error),
})
return
}
const headers: Record<string, string> = {
'Content-Type': 'application/json',
'User-Agent': getClaudeCodeUserAgent(),
...authResult.headers,
}
const response = await axios.post(this.endpoint, payload, {
timeout: this.timeout,
headers,
})
logForDebugging('BigQuery metrics exported successfully')
logForDebugging(
`BigQuery API Response: ${jsonStringify(response.data, null, 2)}`,
)
resultCallback({ code: ExportResultCode.SUCCESS })
} catch (error) {
logForDebugging(`BigQuery metrics export failed: ${errorMessage(error)}`)
logError(error)
resultCallback({
code: ExportResultCode.FAILED,
error: toError(error),
})
}
}
private transformMetricsForInternal(
metrics: ResourceMetrics,
): InternalMetricsPayload {
const attrs = metrics.resource.attributes
const resourceAttributes: Record<string, string> = {
'service.name': (attrs['service.name'] as string) || 'claude-code',
'service.version': (attrs['service.version'] as string) || 'unknown',
'os.type': (attrs['os.type'] as string) || 'unknown',
'os.version': (attrs['os.version'] as string) || 'unknown',
'host.arch': (attrs['host.arch'] as string) || 'unknown',
'aggregation.temporality':
this.selectAggregationTemporality() === AggregationTemporality.DELTA
? 'delta'
: 'cumulative',
}
// Only add wsl.version if it exists (omit instead of default)
if (attrs['wsl.version']) {
resourceAttributes['wsl.version'] = attrs['wsl.version'] as string
}
// Add customer type and subscription type
if (isClaudeAISubscriber()) {
resourceAttributes['user.customer_type'] = 'claude_ai'
const subscriptionType = getSubscriptionType()
if (subscriptionType) {
resourceAttributes['user.subscription_type'] = subscriptionType
}
} else {
resourceAttributes['user.customer_type'] = 'api'
}
const transformed = {
resource_attributes: resourceAttributes,
metrics: metrics.scopeMetrics.flatMap(scopeMetric =>
scopeMetric.metrics.map(metric => ({
name: metric.descriptor.name,
description: metric.descriptor.description,
unit: metric.descriptor.unit,
data_points: this.extractDataPoints(metric),
})),
),
}
return transformed
}
private extractDataPoints(metric: MetricData): DataPoint[] {
const dataPoints = metric.dataPoints || []
return dataPoints
.filter(
(point): point is OTelDataPoint<number> =>
typeof point.value === 'number',
)
.map(point => ({
attributes: this.convertAttributes(point.attributes),
value: point.value,
timestamp: this.hrTimeToISOString(
point.endTime || point.startTime || [Date.now() / 1000, 0],
),
}))
}
async shutdown(): Promise<void> {
this.isShutdown = true
await this.forceFlush()
logForDebugging('BigQuery metrics exporter shutdown complete')
}
async forceFlush(): Promise<void> {
await Promise.all(this.pendingExports)
logForDebugging('BigQuery metrics exporter flush complete')
}
private convertAttributes(
attributes: Attributes | undefined,
): Record<string, string> {
const result: Record<string, string> = {}
if (attributes) {
for (const [key, value] of Object.entries(attributes)) {
if (value !== undefined && value !== null) {
result[key] = String(value)
}
}
}
return result
}
private hrTimeToISOString(hrTime: HrTime): string {
const [seconds, nanoseconds] = hrTime
const date = new Date(seconds * 1000 + nanoseconds / 1000000)
return date.toISOString()
}
selectAggregationTemporality(): AggregationTemporality {
// DO NOT CHANGE THIS TO CUMULATIVE
// It would mess up the aggregation of metrics
// for CC Productivity metrics dashboard
return AggregationTemporality.DELTA
}
}

View File

@@ -1,14 +1,5 @@
export function bootstrapTelemetry(): void {}
export function parseExporterTypes(value: string | undefined): string[] {
return (value || '')
.trim()
.split(',')
.filter(Boolean)
.map(t => t.trim())
.filter(t => t !== 'none')
}
export function isTelemetryEnabled(): boolean {
return false
}

View File

@@ -1,26 +0,0 @@
import type { DiagLogger } from '@opentelemetry/api'
import { logForDebugging } from '../debug.js'
import { logError } from '../log.js'
export class ClaudeCodeDiagLogger implements DiagLogger {
error(message: string, ..._: unknown[]) {
logError(new Error(message))
logForDebugging(`[3P telemetry] OTEL diag error: ${message}`, {
level: 'error',
})
}
warn(message: string, ..._: unknown[]) {
logError(new Error(message))
logForDebugging(`[3P telemetry] OTEL diag warn: ${message}`, {
level: 'warn',
})
}
info(_message: string, ..._args: unknown[]) {
return
}
debug(_message: string, ..._args: unknown[]) {
return
}
verbose(_message: string, ..._args: unknown[]) {
return
}
}

File diff suppressed because it is too large Load Diff

View File

@@ -1,71 +0,0 @@
import type { Attributes } from '@opentelemetry/api'
import { getSessionId } from 'src/bootstrap/state.js'
import { getOauthAccountInfo } from './auth.js'
import { getOrCreateUserID } from './config.js'
import { envDynamic } from './envDynamic.js'
import { isEnvTruthy } from './envUtils.js'
import { toTaggedId } from './taggedId.js'
// Default configuration for metrics cardinality
const METRICS_CARDINALITY_DEFAULTS = {
OTEL_METRICS_INCLUDE_SESSION_ID: true,
OTEL_METRICS_INCLUDE_VERSION: false,
OTEL_METRICS_INCLUDE_ACCOUNT_UUID: true,
}
function shouldIncludeAttribute(
envVar: keyof typeof METRICS_CARDINALITY_DEFAULTS,
): boolean {
const defaultValue = METRICS_CARDINALITY_DEFAULTS[envVar]
const envValue = process.env[envVar]
if (envValue === undefined) {
return defaultValue
}
return isEnvTruthy(envValue)
}
export function getTelemetryAttributes(): Attributes {
const userId = getOrCreateUserID()
const sessionId = getSessionId()
const attributes: Attributes = {
'user.id': userId,
}
if (shouldIncludeAttribute('OTEL_METRICS_INCLUDE_SESSION_ID')) {
attributes['session.id'] = sessionId
}
if (shouldIncludeAttribute('OTEL_METRICS_INCLUDE_VERSION')) {
attributes['app.version'] = MACRO.VERSION
}
// Only include OAuth account data when actively using OAuth authentication
const oauthAccount = getOauthAccountInfo()
if (oauthAccount) {
const orgId = oauthAccount.organizationUuid
const email = oauthAccount.emailAddress
const accountUuid = oauthAccount.accountUuid
if (orgId) attributes['organization.id'] = orgId
if (email) attributes['user.email'] = email
if (
accountUuid &&
shouldIncludeAttribute('OTEL_METRICS_INCLUDE_ACCOUNT_UUID')
) {
attributes['user.account_uuid'] = accountUuid
attributes['user.account_id'] =
process.env.CLAUDE_CODE_ACCOUNT_TAGGED_ID ||
toTaggedId('user', accountUuid)
}
}
// Add terminal type if available
if (envDynamic.terminal) {
attributes['terminal.type'] = envDynamic.terminal
}
return attributes
}