@cap-js/agents 0.0.0 → 0.9.1

package/lib/telemetry/chat-tracing.js

@@ -0,0 +1,342 @@
+import cds from "@sap/cds"
+import * as metrics from "./metrics.js"
+import { mlflowAttrs, setSpanAttrs } from "./mlflow.js"
+import { audit } from "../utils/utils.js"
+
+const LOG = cds.log("agents")
+
+const PATCHED = Symbol.for("@cap-js/agents:patched")
+const SPAN_KIND_CLIENT = 3
+
+// ─── Public API ──────────────────────────────────────────────────────────────
+
+export async function patchChatModel() {
+  try {
+    const mod = await import("@langchain/core/language_models/chat_models")
+    if (mod.BaseChatModel?.prototype) _patchChatModelProto(mod.BaseChatModel.prototype)
+
+    // Also patch CJS prototype (different from ESM in dual-package modules)
+    try {
+      const { createRequire } = await import("node:module")
+      const req = createRequire(import.meta.url)
+      const cjs = req("@langchain/core/language_models/chat_models")
+      if (cjs.BaseChatModel?.prototype) _patchChatModelProto(cjs.BaseChatModel.prototype)
+    } catch {
+      /* CJS not available */
+    }
+  } catch (err) {
+    LOG.error("Failed to patch BaseChatModel for tracing", { error: err.message })
+  }
+}
+
+// ─── Prototype patch ─────────────────────────────────────────────────────────
+
+export function _patchChatModelProto(proto) {
+  if (proto[PATCHED]) return
+  const original = proto.invoke
+  if (typeof original !== "function") {
+    LOG.warn("BaseChatModel.invoke not found — tracing patch skipped")
+    return
+  }
+
+  proto.invoke = async function (input, opts) {
+    // Skip instances explicitly opted out (e.g. content-filter probe model)
+    const INSTRUMENTED = Symbol.for("@cap-js/agents:instrumented")
+    if (this[INSTRUMENTED]) return original.call(this, input, opts)
+
+    const tracer = metrics.getTracer()
+    if (!tracer) return original.call(this, input, opts)
+
+    const model = this.options?.model || this.model || this.constructor.name
+    const provider = _detectProvider(this)
+    const node = opts?.runName || "agent"
+    const messages = Array.isArray(input) ? input : undefined
+    const cacheControl = _detectCacheControl(messages)
+    const streaming = this.streaming || this.orchestrationConfig?.streaming || false
+    const mAttrs = _metricAttrs(model, node)
+
+    const invoke = async (span) => {
+      if (span) {
+        setLLMSpanStartAttrs(span, {
+          model,
+          provider,
+          params: this.options?.params,
+          streaming,
+          cacheControl,
+          node,
+          messages,
+        })
+      }
+      const t0 = Date.now()
+      try {
+        const result = await original.call(this, input, opts)
+        const duration = Date.now() - t0
+        /** @type {import('@langchain/core/messages').UsageMetadata} */
+        const usage = result.usage_metadata
+        const finishReason =
+          result.response_metadata?.finish_reason ||
+          result.response_metadata?.stop_reason ||
+          (result.tool_calls?.length > 0 ? "tool_use" : "stop")
+
+        _handleSuccess(span, {
+          model,
+          provider,
+          mAttrs,
+          opts,
+          params: this.options?.params,
+          node,
+          tokenUsage: convertUsageData(usage),
+          outputContent: result.content ? extractText(result.content) : undefined,
+          response: {
+            finishReason,
+            model: result.response_metadata?.model_name || result.response_metadata?.model || model,
+            id: result.response_metadata?.id || result.id,
+            toolCalls: result.tool_calls?.map((tc) => ({ name: tc.name, args: tc.args })),
+          },
+          duration,
+        })
+        return result
+      } catch (err) {
+        _handleError(span, { err, mAttrs, model, node, messages })
+        throw err
+      } finally {
+        if (span) span.end()
+      }
+    }
+
+    return tracer.startActiveSpan(`chat ${model}`, { kind: SPAN_KIND_CLIENT }, (span) =>
+      invoke(span),
+    )
+  }
+  proto[PATCHED] = true
+}
+
+// ─── Internal helpers ────────────────────────────────────────────────────────
+
+/** Detect LLM provider from model instance properties. */
+function _detectProvider(instance) {
+  if (instance.orchestrationConfig) return "sap-ai-core"
+  return "langchain"
+}
+
+function _metricAttrs(model, node) {
+  return {
+    "sap.tenantId": cds.context?.tenant || "anonymous",
+    "agent.service": cds.context?.["agent.service"],
+    model,
+    node,
+  }
+}
+
+/** Detect cache_control presence in messages (set by injectCacheControl in aicore). */
+function _detectCacheControl(messages) {
+  if (!Array.isArray(messages)) return false
+  return messages.some((m) => {
+    if (!Array.isArray(m.content)) return false
+    return m.content.some((b) => b.cache_control)
+  })
+}
+
+/** Handle successful LLM call: metrics, span end attrs, truncation warning, audit. */
+function _handleSuccess(
+  span,
+  { model, provider, mAttrs, opts, params, node, tokenUsage, outputContent, response, duration },
+) {
+  metrics.llmInvocations.add(1, { ...mAttrs, outcome: "success" })
+  if (tokenUsage?.input_tokens) metrics.llmInputTokens.add(tokenUsage.input_tokens, mAttrs)
+  if (tokenUsage?.output_tokens) metrics.llmOutputTokens.add(tokenUsage.output_tokens, mAttrs)
+
+  if (span) {
+    setLLMSpanEndAttrs(span, { model, provider, tokenUsage, outputContent, response })
+    span.setStatus({ code: 1 })
+  }
+
+  if (response?.finishReason === "length" || response?.finishReason === "max_tokens") {
+    LOG.warn("LLM response truncated: output_tokens reached max_tokens limit", {
+      model,
+      node,
+      max_tokens: params?.max_tokens,
+      output_tokens: tokenUsage?.output_tokens,
+    })
+  }
+
+  const taskId = opts?.configurable?._taskId || cds.context?.["agent.task.id"]
+  if (taskId) {
+    audit("AgentDecision", {
+      data: {
+        taskId,
+        contextId:
+          opts?.configurable?.thread_id?.split(":")[1] || cds.context?.["agent.context.id"],
+        service: opts?.configurable?._service || cds.context?.["agent.service"],
+        model,
+        iteration: opts?.configurable?._iteration ?? cds.context?.["agent.iteration"],
+        toolCalls: response?.toolCalls,
+        tokenUsage,
+        duration,
+      },
+    })
+  }
+}
+
+/** Handle LLM error: content-filter warnings, metrics, span error attrs. */
+function _handleError(span, { err, mAttrs, model, node, messages }) {
+  const status = err.rootCause?.status
+  const data = err.rootCause?.response?.data
+  const headers = err.rootCause?.response?.headers
+  const isFilterModule = /Filtering Module/i.test(data?.error?.location || "")
+  const isExternalFailure = headers?.["ai-external-failure"] === "true"
+
+  if (isFilterModule && status === 503 && isExternalFailure) {
+    LOG.warn(
+      "Content filter service rejected the request (likely payload too large for prompt_shield). " +
+        "Disable content filtering via the buildContentFilter event handler — see README → Content Filter → Limitations.",
+      { model, node, status, location: data?.error?.location, messageCount: messages?.length },
+    )
+  } else if (isFilterModule && status === 400) {
+    LOG.warn("Content filter blocked the request", {
+      model,
+      node,
+      status,
+      reason: data?.error?.message,
+    })
+  }
+
+  metrics.llmInvocations.add(1, { ...mAttrs, outcome: "error" })
+  if (span) {
+    span.setAttribute("error.type", err.constructor?.name || "Error")
+    span.setStatus({ code: 2, message: err.message })
+    span.recordException(err)
+  }
+}
+
+// ─── Shared LLM span helpers ─────────────────────────────────────────────────
+
+/** Extract plain text from LangChain message content (string or content-block array). */
+export function extractText(content) {
+  if (typeof content === "string") return content
+  if (Array.isArray(content)) {
+    return content
+      .filter((b) => b.type === "text")
+      .map((b) => b.text || "")
+      .join("")
+  }
+  return String(content ?? "")
+}
+
+/** Map LangChain message role to OpenAI role string. */
+export function toRole(msg) {
+  const t = msg._getType?.()
+  if (t === "human") return "user"
+  if (t === "ai") return "assistant"
+  return t || "user"
+}
+
+/**
+ * Set common LLM span start attributes.
+ */
+export function setLLMSpanStartAttrs(
+  span,
+  { model, provider, params, streaming, cacheControl, node, messages },
+) {
+  span.setAttribute("gen_ai.operation.name", "chat")
+  span.setAttribute("gen_ai.provider.name", provider)
+  span.setAttribute("gen_ai.request.model", model)
+  span.setAttribute("mlflow.message.format", "langchain-js")
+  if (params?.temperature != null)
+    span.setAttribute("gen_ai.request.temperature", params.temperature)
+  if (params?.max_tokens != null) span.setAttribute("gen_ai.request.max_tokens", params.max_tokens)
+  if (cds.context?.["agent.context.id"])
+    span.setAttribute("gen_ai.conversation.id", cds.context["agent.context.id"])
+  if (streaming) span.setAttribute("gen_ai.request.stream", true)
+  if (node) span.setAttribute("agent.llm.node", node)
+  if (cacheControl) span.setAttribute("gen_ai.request.cache_control", true)
+
+  // MLflow inputs: chat messages for Chat tab
+  if ((cds.env.agents?.mlflow || LOG._debug) && messages) {
+    const mlMessages = messages.map((m) => ({ role: toRole(m), content: extractText(m.content) }))
+    setSpanAttrs(span, mlflowAttrs("LLM", { model, provider, inputs: { messages: mlMessages } }))
+    if (LOG._debug) {
+      span.setAttribute("gen_ai.input.messages", JSON.stringify(messages.map((m) => m.content)))
+    }
+  }
+}
+
+/**
+ * Set gen_ai.response.*, gen_ai.usage.*, MLflow outputs on span end.
+ */
+export function setLLMSpanEndAttrs(span, { model, provider, tokenUsage, outputContent, response }) {
+  if (LOG._debug && outputContent) {
+    span.setAttribute("gen_ai.output.messages", outputContent)
+  }
+
+  // gen_ai.response.* attributes
+  if (response) {
+    if (response.finishReason)
+      span.setAttribute("gen_ai.response.finish_reasons", [response.finishReason])
+    if (response.model) span.setAttribute("gen_ai.response.model", response.model)
+    if (response.id) span.setAttribute("gen_ai.response.id", response.id)
+    if (response.finishReason === "length" || response.finishReason === "max_tokens") {
+      span.setAttribute("gen_ai.response.truncated", true)
+    }
+    if (response.toolCalls?.length > 0) {
+      span.setAttribute("gen_ai.response.tool_calls", JSON.stringify(response.toolCalls))
+    }
+  }
+
+  if (tokenUsage) {
+    setTokenUsage(span, tokenUsage)
+    if (cds.env.agents?.mlflow) {
+      const mlOpts = { model, provider, tokenUsage: { ...tokenUsage, reasoning_tokens: undefined } }
+      if (outputContent) {
+        mlOpts.outputs = {
+          choices: [{ message: { role: "assistant", content: outputContent } }],
+        }
+      }
+      setSpanAttrs(span, mlflowAttrs("LLM", mlOpts))
+      return
+    }
+  }
+
+  // MLflow outputs without token usage
+  if (cds.env.agents?.mlflow && outputContent) {
+    setSpanAttrs(
+      span,
+      mlflowAttrs("LLM", {
+        model,
+        provider,
+        outputs: { choices: [{ message: { role: "assistant", content: outputContent } }] },
+      }),
+    )
+  }
+}
+
+export function setTokenUsage(span, tokenUsage) {
+  span.setAttribute("gen_ai.usage.input_tokens", tokenUsage.input_tokens)
+  span.setAttribute("gen_ai.usage.output_tokens", tokenUsage.output_tokens)
+  span.setAttribute("gen_ai.usage.total_tokens", tokenUsage.total_tokens)
+
+  if (tokenUsage.cache_read_input_tokens != null)
+    span.setAttribute("gen_ai.usage.cache_read.input_tokens", tokenUsage.cache_read_input_tokens)
+  if (tokenUsage.cache_creation_input_tokens != null)
+    span.setAttribute(
+      "gen_ai.usage.cache_creation.input_tokens",
+      tokenUsage.cache_creation_input_tokens,
+    )
+  if (tokenUsage.reasoning_tokens != null)
+    span.setAttribute("gen_ai.usage.reasoning.output_tokens", tokenUsage.reasoning_tokens)
+}
+
+/**
+ * Convert LangChain UsageMetadata to flat token usage format.
+ */
+export function convertUsageData(usage) {
+  if (!usage) return undefined
+  return {
+    input_tokens: usage.input_tokens,
+    output_tokens: usage.output_tokens,
+    total_tokens: usage.total_tokens,
+    cache_creation_input_tokens: usage.input_token_details?.cache_creation,
+    cache_read_input_tokens: usage.input_token_details?.cache_read,
+    reasoning_tokens: usage.output_token_details?.reasoning,
+  }
+}

package/lib/telemetry/metrics.js

@@ -0,0 +1,85 @@
+import cds from "@sap/cds"
+import { createRequire } from "node:module"
+
+const NOOP_METER = (() => {
+  const noop = () => ({ add() {}, record() {} })
+  return { createCounter: noop, createHistogram: noop, createUpDownCounter: noop }
+})()
+
+let _otel
+try {
+  // Synchronous require avoids top-level await which blocks module graph
+  // resolution and causes issues with CDS plugin loading order.
+  const require = createRequire(import.meta.url)
+  _otel = require("@opentelemetry/api")
+} catch {
+  _otel = null
+}
+
+function getMeter() {
+  return _otel ? _otel.metrics.getMeter("@cap-js/agents") : NOOP_METER
+}
+
+const meter = getMeter()
+
+export const requestDuration = meter.createHistogram("agent.request.duration", {
+  description: "End-to-end agent request duration",
+  unit: "ms",
+})
+export const requestsTotal = meter.createCounter("agent.requests.total", {
+  description: "Total inbound agent requests",
+})
+export const errorsTotal = meter.createCounter("agent.errors.total", {
+  description: "Agent requests resulting in error",
+})
+export const concurrentExecutions = meter.createUpDownCounter("agent.executions.concurrent", {
+  description: "Currently active workflow executions",
+})
+export const workflowsCompleted = meter.createCounter("agent.workflows.completed", {
+  description: "Number of completed agent workflows",
+})
+export const agentActions = meter.createCounter("agent_actions", {
+  description: "LLM invocations (agent node calls) per tenant",
+})
+export const llmInputTokens = meter.createCounter("agent.llm.input_tokens", {
+  description: "LLM input tokens consumed",
+})
+export const llmOutputTokens = meter.createCounter("agent.llm.output_tokens", {
+  description: "LLM output tokens generated",
+})
+export const llmInvocations = meter.createCounter("agent.llm.invocations", {
+  description: "LLM invocation count",
+})
+export const toolInvocations = meter.createCounter("agent.tool.invocations", {
+  description: "Tool invocation count",
+})
+
+/** Common attributes for all agent metrics */
+export function attrs(srv) {
+  return {
+    "sap.tenantId": cds.context?.tenant || "anonymous",
+    "agent.service": typeof srv === "string" ? srv : srv.name,
+  }
+}
+
+/**
+ * Get active OTel span. Returns null if OTel is unavailable or no span is active.
+ */
+export function getActiveSpan() {
+  return _otel?.trace.getActiveSpan() || null
+}
+
+/** Get OTel tracer for @cap-js/agents (or null if unavailable) */
+export function getTracer() {
+  return _otel?.trace.getTracer("@cap-js/agents") || null
+}
+
+/** Create the active_users ObservableGauge with given observation callback */
+export function createActiveUsersGauge(callback) {
+  if (!_otel) return null
+  const gauge = meter.createObservableGauge("active_users", {
+    description: "Active users per tenant and agent service (24h rolling window)",
+  })
+  gauge.addCallback(callback)
+  return gauge
+}