feat(copilot): server-side mothership tool/vfs/file metrics + $env tagging

- copilot.tool.duration/calls, vfs.materialize, file.read metrics on Sim's
  MeterProvider (shared buckets, bounded labels, name-capped) at the
  tool/VFS/file-read boundaries
- metrics-v1 TS mirror + sync-metrics-contract script + GENERATORS entry;
  regen trace-attributes/events mirrors (adds gen_ai.system)
- deployment.environment from APPCONFIG_ENVIRONMENT (production -> prod) so a
  single Grafana $env spans Sim + Go without a new infra env var

Author: Sg312

apps/sim/instrumentation-node.ts

@@ -83,6 +83,17 @@ function normalizeOtlpMetricsUrl(url: string): string {
   }
 }
 
+// deployment.environment in the GO value space (dev | staging | prod) without
+// any new infra env var. Every deployed Sim tier already gets
+// APPCONFIG_ENVIRONMENT = the infra env name (dev | staging | production), so we
+// reuse it and map production -> prod to match Go's `OTEL_DEPLOYMENT_ENVIRONMENT`
+// (and thus a single Grafana $env filter spans Sim + Go). Returns undefined when
+// unset (local dev) so the OTEL_/NODE_ENV fallbacks still apply.
+function deploymentEnvFromAppConfig(v: string | undefined): string | undefined {
+  if (!v) return undefined
+  return v === 'production' ? 'prod' : v
+}
+
 // Sampling ratio from env (mirrors Go's `samplerFromEnv`); fallback
 // is 100% everywhere. Retention caps cost, not sampling.
 function resolveSamplingRatio(_isLocalEndpoint: boolean): number {
@@ -270,11 +281,15 @@ async function initializeOpenTelemetry() {
       resourceFromAttributes({
         [ATTR_SERVICE_NAME]: telemetryConfig.serviceName,
         [ATTR_SERVICE_VERSION]: telemetryConfig.serviceVersion,
-        // OTEL_ → DEPLOYMENT_ENVIRONMENT → NODE_ENV; matches Go's
-        // `resourceEnvFromEnv()` so both halves tag the same value.
+        // OTEL_ → DEPLOYMENT_ENVIRONMENT → APPCONFIG_ENVIRONMENT (mapped to the
+        // Go value space) → NODE_ENV. Matches Go's `resourceEnvFromEnv()` so a
+        // single $env spans Sim + Go. APPCONFIG_ENVIRONMENT (already set on every
+        // deployed tier) is the fix that stops deployed Sim tagging everything
+        // "production" via the NODE_ENV fallback — no new infra env var needed.
         [ATTR_DEPLOYMENT_ENVIRONMENT]:
           process.env.OTEL_DEPLOYMENT_ENVIRONMENT ||
           process.env.DEPLOYMENT_ENVIRONMENT ||
+          deploymentEnvFromAppConfig(process.env.APPCONFIG_ENVIRONMENT) ||
           env.NODE_ENV ||
           'development',
         'service.namespace': 'mothership',

apps/sim/lib/copilot/generated/metrics-v1.ts

@@ -0,0 +1,58 @@
+// AUTO-GENERATED FILE. DO NOT EDIT.
+//
+// Source: copilot/copilot/contracts/metrics-v1.schema.json
+// Regenerate with: bun run metrics-contract:generate
+//
+// Canonical mothership OTel metric names. Call sites should reference
+// `Metric.<Identifier>` (e.g. `Metric.CopilotToolDuration`) rather than raw
+// string literals, so the Go-side contract is the single source of truth and
+// typos become compile errors.
+//
+// NAMES ONLY. Label keys and histogram bucket boundaries are NOT in this
+// contract — Go owns the label-cardinality allowlist and the shared bucket
+// constant, and the Sim emitter MUST mirror those by hand so the Go∪Sim metric
+// union is queryable as one series set.
+
+export const Metric = {
+  CopilotCacheAttempted: 'copilot.cache.attempted',
+  CopilotCacheHit: 'copilot.cache.hit',
+  CopilotCacheWrite: 'copilot.cache.write',
+  CopilotFileReadDuration: 'copilot.file.read.duration',
+  CopilotFileReadSize: 'copilot.file.read.size',
+  CopilotMessagesSerializeDuration: 'copilot.messages.serialize.duration',
+  CopilotRequestCount: 'copilot.request.count',
+  CopilotRequestDuration: 'copilot.request.duration',
+  CopilotToolCalls: 'copilot.tool.calls',
+  CopilotToolDuration: 'copilot.tool.duration',
+  CopilotVfsMaterializeDuration: 'copilot.vfs.materialize.duration',
+  GenAiClientCacheTokenUsage: 'gen_ai.client.cache.token.usage',
+  GenAiClientTokenUsage: 'gen_ai.client.token.usage',
+  LlmClientErrors: 'llm.client.errors',
+  LlmClientOutputCutoff: 'llm.client.output_cutoff',
+  LlmClientStreamDuration: 'llm.client.stream.duration',
+  LlmClientTimeToFirstToken: 'llm.client.time_to_first_token',
+} as const
+
+export type MetricKey = keyof typeof Metric
+export type MetricValue = (typeof Metric)[MetricKey]
+
+/** Readonly sorted list of every canonical mothership metric name. */
+export const MetricValues: readonly MetricValue[] = [
+  'copilot.cache.attempted',
+  'copilot.cache.hit',
+  'copilot.cache.write',
+  'copilot.file.read.duration',
+  'copilot.file.read.size',
+  'copilot.messages.serialize.duration',
+  'copilot.request.count',
+  'copilot.request.duration',
+  'copilot.tool.calls',
+  'copilot.tool.duration',
+  'copilot.vfs.materialize.duration',
+  'gen_ai.client.cache.token.usage',
+  'gen_ai.client.token.usage',
+  'llm.client.errors',
+  'llm.client.output_cutoff',
+  'llm.client.stream.duration',
+  'llm.client.time_to_first_token',
+] as const

apps/sim/lib/copilot/generated/trace-attributes-v1.ts

@@ -54,10 +54,8 @@ export const TraceAttr = {
   AuthProvider: 'auth.provider',
   AuthValidateStatusCode: 'auth.validate.status_code',
   AwsRegion: 'aws.region',
-  BedrockErrorCode: 'bedrock.error_code',
-  BedrockModelId: 'bedrock.model_id',
-  BedrockRequestBodyBytesRetry: 'bedrock.request.body_bytes_retry',
   BillingAttempts: 'billing.attempts',
+  BillingByok: 'billing.byok',
   BillingChangeType: 'billing.change_type',
   BillingCostInputUsd: 'billing.cost.input_usd',
   BillingCostOutputUsd: 'billing.cost.output_usd',
@@ -159,6 +157,14 @@ export const TraceAttr = {
   ContextReduced: 'context.reduced',
   ContextSummarizeInputChars: 'context.summarize.input_chars',
   ContextSummarizeOutputChars: 'context.summarize.output_chars',
+  ContextTransformCaller: 'context.transform.caller',
+  ContextTransformCharsIn: 'context.transform.chars_in',
+  ContextTransformCharsOut: 'context.transform.chars_out',
+  ContextTransformDropCount: 'context.transform.drop_count',
+  ContextTransformDrops: 'context.transform.drops',
+  ContextTransformMessagesIn: 'context.transform.messages_in',
+  ContextTransformMessagesOut: 'context.transform.messages_out',
+  ContextTransformStage: 'context.transform.stage',
   CopilotAbortControllerFired: 'copilot.abort.controller_fired',
   CopilotAbortGoMarkerOk: 'copilot.abort.go_marker_ok',
   CopilotAbortLocalAborted: 'copilot.abort.local_aborted',
@@ -276,6 +282,7 @@ export const TraceAttr = {
   CopilotVfsOutcome: 'copilot.vfs.outcome',
   CopilotVfsOutputBytes: 'copilot.vfs.output.bytes',
   CopilotVfsOutputMediaType: 'copilot.vfs.output.media_type',
+  CopilotVfsPhase: 'copilot.vfs.phase',
   CopilotVfsReadImageResized: 'copilot.vfs.read.image.resized',
   CopilotVfsReadOutcome: 'copilot.vfs.read.outcome',
   CopilotVfsReadOutputBytes: 'copilot.vfs.read.output.bytes',
@@ -389,6 +396,7 @@ export const TraceAttr = {
   GenAiRequestToolUseBlocks: 'gen_ai.request.tool_use_blocks',
   GenAiRequestToolsCount: 'gen_ai.request.tools.count',
   GenAiRequestUserMessages: 'gen_ai.request.user_messages',
+  GenAiResponseFinishReasons: 'gen_ai.response.finish_reasons',
   GenAiResponseModel: 'gen_ai.response.model',
   GenAiStreamPhaseTextBytes: 'gen_ai.stream.phase.text.bytes',
   GenAiStreamPhaseTextChunks: 'gen_ai.stream.phase.text.chunks',
@@ -434,7 +442,9 @@ export const TraceAttr = {
   InvitationRole: 'invitation.role',
   KnowledgeBaseId: 'knowledge_base.id',
   KnowledgeBaseName: 'knowledge_base.name',
+  LlmBackend: 'llm.backend',
   LlmErrorStage: 'llm.error_stage',
+  LlmProtocol: 'llm.protocol',
   LlmRequestBodyBytes: 'llm.request.body_bytes',
   LlmStreamBytes: 'llm.stream.bytes',
   LlmStreamChunks: 'llm.stream.chunks',
@@ -460,6 +470,10 @@ export const TraceAttr = {
   MemoryPath: 'memory.path',
   MemoryRowCount: 'memory.row_count',
   MessageId: 'message.id',
+  MessagesDeserializeMs: 'messages.deserialize_ms',
+  MessagesSerializeOp: 'messages.serialize.op',
+  MessagesSerializeSite: 'messages.serialize.site',
+  MessagesSerializeMs: 'messages.serialize_ms',
   MessagingDestinationName: 'messaging.destination.name',
   MessagingSystem: 'messaging.system',
   ModelDurationMs: 'model.duration_ms',
@@ -495,14 +509,15 @@ export const TraceAttr = {
   ResumeResultsFailureCount: 'resume.results.failure_count',
   ResumeResultsSuccessCount: 'resume.results.success_count',
   RouterBackendName: 'router.backend_name',
-  RouterBedrockEnabled: 'router.bedrock_enabled',
-  RouterBedrockSupportedModel: 'router.bedrock_supported_model',
+  RouterConfigVersion: 'router.config_version',
   RouterId: 'router.id',
   RouterName: 'router.name',
+  RouterRouteReason: 'router.route_reason',
   RouterSelectedBackend: 'router.selected_backend',
   RouterSelectedPath: 'router.selected_path',
   RunId: 'run.id',
   SearchResultsCount: 'search.results_count',
+  ServerAddress: 'server.address',
   ServiceInstanceId: 'service.instance.id',
   ServiceName: 'service.name',
   ServiceNamespace: 'service.namespace',
@@ -663,10 +678,8 @@ export const TraceAttrValues: readonly TraceAttrValue[] = [
   'auth.provider',
   'auth.validate.status_code',
   'aws.region',
-  'bedrock.error_code',
-  'bedrock.model_id',
-  'bedrock.request.body_bytes_retry',
   'billing.attempts',
+  'billing.byok',
   'billing.change_type',
   'billing.cost.input_usd',
   'billing.cost.output_usd',
@@ -768,6 +781,14 @@ export const TraceAttrValues: readonly TraceAttrValue[] = [
   'context.reduced',
   'context.summarize.input_chars',
   'context.summarize.output_chars',
+  'context.transform.caller',
+  'context.transform.chars_in',
+  'context.transform.chars_out',
+  'context.transform.drop_count',
+  'context.transform.drops',
+  'context.transform.messages_in',
+  'context.transform.messages_out',
+  'context.transform.stage',
   'copilot.abort.controller_fired',
   'copilot.abort.go_marker_ok',
   'copilot.abort.local_aborted',
@@ -885,6 +906,7 @@ export const TraceAttrValues: readonly TraceAttrValue[] = [
   'copilot.vfs.outcome',
   'copilot.vfs.output.bytes',
   'copilot.vfs.output.media_type',
+  'copilot.vfs.phase',
   'copilot.vfs.read.image.resized',
   'copilot.vfs.read.outcome',
   'copilot.vfs.read.output.bytes',
@@ -987,6 +1009,7 @@ export const TraceAttrValues: readonly TraceAttrValue[] = [
   'gen_ai.request.tool_use_blocks',
   'gen_ai.request.tools.count',
   'gen_ai.request.user_messages',
+  'gen_ai.response.finish_reasons',
   'gen_ai.response.model',
   'gen_ai.stream.phase.text.bytes',
   'gen_ai.stream.phase.text.chunks',
@@ -1032,7 +1055,9 @@ export const TraceAttrValues: readonly TraceAttrValue[] = [
   'invitation.role',
   'knowledge_base.id',
   'knowledge_base.name',
+  'llm.backend',
   'llm.error_stage',
+  'llm.protocol',
   'llm.request.body_bytes',
   'llm.stream.bytes',
   'llm.stream.chunks',
@@ -1058,6 +1083,10 @@ export const TraceAttrValues: readonly TraceAttrValue[] = [
   'memory.path',
   'memory.row_count',
   'message.id',
+  'messages.deserialize_ms',
+  'messages.serialize.op',
+  'messages.serialize.site',
+  'messages.serialize_ms',
   'messaging.destination.name',
   'messaging.system',
   'model.duration_ms',
@@ -1093,14 +1122,15 @@ export const TraceAttrValues: readonly TraceAttrValue[] = [
   'resume.results.failure_count',
   'resume.results.success_count',
   'router.backend_name',
-  'router.bedrock_enabled',
-  'router.bedrock_supported_model',
+  'router.config_version',
   'router.id',
   'router.name',
+  'router.route_reason',
   'router.selected_backend',
   'router.selected_path',
   'run.id',
   'search.results_count',
+  'server.address',
   'service.instance.id',
   'service.name',
   'service.namespace',

apps/sim/lib/copilot/generated/trace-events-v1.ts

@@ -10,7 +10,7 @@
 // become compile errors.
 
 export const TraceEvent = {
-  BedrockInvokeRetryWithoutImages: 'bedrock.invoke.retry_without_images',
+  ContextTransform: 'context.transform',
   CopilotOutputFileError: 'copilot.output_file.error',
   CopilotSseFirstEvent: 'copilot.sse.first_event',
   CopilotSseIdleGapExceeded: 'copilot.sse.idle_gap_exceeded',
@@ -33,7 +33,7 @@ export type TraceEventValue = (typeof TraceEvent)[TraceEventKey]
 
 /** Readonly sorted list of every canonical event name. */
 export const TraceEventValues: readonly TraceEventValue[] = [
-  'bedrock.invoke.retry_without_images',
+  'context.transform',
   'copilot.output_file.error',
   'copilot.sse.first_event',
   'copilot.sse.idle_gap_exceeded',

apps/sim/lib/copilot/request/metrics.ts

@@ -0,0 +1,101 @@
+// Sim server-side copilot metrics (U17). Sim's MeterProvider is wired in
+// instrumentation-node.ts (OTLP → Mimir, 60s) but had no copilot instruments;
+// this module is its first consumer. We emit the SAME metric names + label keys
+// + histogram bucket boundaries as the Go side (copilot internal/telemetry +
+// contracts/metrics_v1.go) so the Go∪Sim union is queryable as one series set
+// — e.g. `copilot.tool.duration` split by `tool.executor` (go|client|sim).
+//
+// Bounded cardinality only: tool.name is capped to the shared tool catalog
+// (else "other"); vfs phase / file-read outcome are bounded sets. NEVER a
+// user/chat/request id (those explode Prometheus series).
+import { type Counter, type Histogram, metrics } from '@opentelemetry/api'
+import { Metric } from '@/lib/copilot/generated/metrics-v1'
+import { TOOL_CATALOG } from '@/lib/copilot/generated/tool-catalog-v1'
+import { TraceAttr } from '@/lib/copilot/generated/trace-attributes-v1'
+
+// MUST match Go's copilot/internal/telemetry/metrics.go LatencyBucketsMs
+// exactly — a histogram_quantile(sum by (le) …) over the Go∪Sim union is only
+// valid with identical boundaries. If you change one side, change the other.
+const LATENCY_BUCKETS_MS = [
+  50, 100, 250, 500, 1000, 2000, 5000, 10000, 20000, 30000, 60000, 120000, 300000,
+]
+
+// File sizes span KB→tens of MB; a bytes-appropriate bucket set (not latency).
+const BYTE_BUCKETS = [1024, 8192, 65536, 262144, 1048576, 4194304, 16777216, 67108864, 268435456]
+
+interface CopilotMeterInstruments {
+  toolDuration: Histogram
+  toolCalls: Counter
+  vfsMaterializeDuration: Histogram
+  fileReadDuration: Histogram
+  fileReadBytes: Histogram
+}
+
+let cached: CopilotMeterInstruments | undefined
+
+// Lazy init: Turbopack/Next can evaluate this module before the NodeSDK
+// installs the real MeterProvider, so resolve instruments on first use (a
+// no-op meter before then simply drops records — same pattern as getCopilotTracer).
+function instruments(): CopilotMeterInstruments {
+  if (cached) return cached
+  const meter = metrics.getMeter('sim-copilot')
+  cached = {
+    toolDuration: meter.createHistogram(Metric.CopilotToolDuration, {
+      unit: 'ms',
+      advice: { explicitBucketBoundaries: LATENCY_BUCKETS_MS },
+    }),
+    toolCalls: meter.createCounter(Metric.CopilotToolCalls),
+    vfsMaterializeDuration: meter.createHistogram(Metric.CopilotVfsMaterializeDuration, {
+      unit: 'ms',
+      advice: { explicitBucketBoundaries: LATENCY_BUCKETS_MS },
+    }),
+    fileReadDuration: meter.createHistogram(Metric.CopilotFileReadDuration, {
+      unit: 'ms',
+      advice: { explicitBucketBoundaries: LATENCY_BUCKETS_MS },
+    }),
+    fileReadBytes: meter.createHistogram(Metric.CopilotFileReadSize, {
+      unit: 'By',
+      advice: { explicitBucketBoundaries: BYTE_BUCKETS },
+    }),
+  }
+  return cached
+}
+
+// Caps tool.name to the shared catalog (matches Go's cappedToolName): a
+// catalog tool keeps its name, everything else (user MCP/custom/unknown)
+// collapses to "other" so series count stays finite.
+function cappedToolName(name: string): string {
+  return TOOL_CATALOG[name] ? name : 'other'
+}
+
+// recordSimToolMetric emits copilot.tool.calls (+1) and copilot.tool.duration
+// for one server-side Sim tool dispatch (executor=sim). outcome is the bounded
+// tool outcome (success/error/…). Pure telemetry.
+export function recordSimToolMetric(name: string, outcome: string, durationMs: number): void {
+  const { toolDuration, toolCalls } = instruments()
+  const attrs = {
+    [TraceAttr.ToolName]: cappedToolName(name),
+    [TraceAttr.ToolExecutor]: 'sim',
+    [TraceAttr.ToolOutcome]: outcome,
+  }
+  toolCalls.add(1, attrs)
+  if (durationMs >= 0) toolDuration.record(durationMs, attrs)
+}
+
+// recordVfsMaterialize records VFS materialization time. Call once per phase
+// with that phase's duration and once with phase="total" for the whole op, so
+// the dashboard can show total + per-phase. phase must be a bounded value.
+export function recordVfsMaterialize(phase: string, durationMs: number): void {
+  if (durationMs < 0) return
+  instruments().vfsMaterializeDuration.record(durationMs, {
+    [TraceAttr.CopilotVfsPhase]: phase,
+  })
+}
+
+// recordFileRead records server-side file-read duration + size by outcome.
+export function recordFileRead(outcome: string, durationMs: number, bytes: number): void {
+  const { fileReadDuration, fileReadBytes } = instruments()
+  const attrs = { [TraceAttr.CopilotVfsReadOutcome]: outcome }
+  if (durationMs >= 0) fileReadDuration.record(durationMs, attrs)
+  if (bytes >= 0) fileReadBytes.record(bytes, attrs)
+}