@cap-js/agents 0.9.2 → 0.9.4
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/README.md +2 -0
- package/cds-plugin.js +2 -2
- package/lib/agents/middleware/content-filter.js +3 -2
- package/lib/agents/middleware/hitl-decision-note-injector.js +18 -0
- package/lib/agents/middleware/hitl.js +2 -2
- package/lib/agents/middleware/index.js +11 -0
- package/lib/agents/middleware/quota-enforcer.js +9 -0
- package/lib/agents/middleware/remote-mcp.js +62 -0
- package/lib/agents/middleware/tool-wrap.js +52 -0
- package/lib/compile.js +6 -2
- package/lib/eval/Judge.js +239 -0
- package/lib/eval/eval-describe.js +69 -0
- package/lib/eval/eval-run.js +147 -0
- package/lib/eval/index.js +6 -0
- package/lib/eval/metrics.js +53 -0
- package/lib/eval/span-collector.js +50 -0
- package/lib/index.js +2 -1
- package/lib/models/aicore.js +23 -6
- package/lib/models/anthropic.js +27 -12
- package/lib/preview/chat.html +373 -66
- package/lib/protocol/agent-card.js +6 -2
- package/lib/sidecar.js +1 -1
- package/lib/telemetry/chat-tracing.js +39 -8
- package/lib/telemetry/mlflow/credentials.js +79 -0
- package/lib/telemetry/mlflow/evaluation.js +38 -0
- package/lib/telemetry/mlflow/exporter/DatabricksExporter.js +75 -0
- package/lib/telemetry/mlflow/exporter/MlflowExporter.js +115 -0
- package/lib/telemetry/mlflow/exporter/index.js +23 -0
- package/lib/telemetry/mlflow/index.js +16 -0
- package/lib/telemetry/mlflow/prompts.js +123 -0
- package/lib/telemetry/mlflow/tracing.js +264 -0
- package/lib/telemetry/tool-tracing.js +27 -54
- package/lib/telemetry/tracing.js +3 -1
- package/lib/utils/markdown.js +1 -11
- package/lib/utils/message-handling.js +14 -0
- package/lib/utils/resilience.js +133 -0
- package/lib/utils/utils.js +17 -0
- package/package.json +26 -14
- package/srv/handlers/chat.js +278 -0
- package/srv/handlers/graph-executor/hitl.js +263 -0
- package/srv/handlers/graph-executor.js +100 -286
- package/srv/handlers/index.js +5 -1
- package/srv/handlers/mcp-tools.js +7 -45
- package/srv/handlers/sub-agent-tools.js +30 -12
- package/srv/handlers/tools.js +39 -5
- package/index.js +0 -0
- package/lib/agents/middleware/hitl-edit-note-injector.js +0 -18
- package/lib/telemetry/mlflow.js +0 -290
- /package/{index.cds → srv/entities.cds} +0 -0
|
@@ -2,10 +2,11 @@ import cds from "@sap/cds"
|
|
|
2
2
|
import { tool } from "@langchain/core/tools"
|
|
3
3
|
import { z } from "zod"
|
|
4
4
|
import { LangGraphExecutor } from "../langgraph-executor-srv.js"
|
|
5
|
-
import { toolName } from "../../lib/utils/utils.js"
|
|
5
|
+
import { short, toolName } from "../../lib/utils/utils.js"
|
|
6
6
|
|
|
7
|
-
const LOG = cds.log("agents:sub-agents")
|
|
7
|
+
const LOG = cds.log("agents:sub-agents|agents|sub-agents")
|
|
8
8
|
|
|
9
|
+
import { inspect } from "util"
|
|
9
10
|
/**
|
|
10
11
|
* Extract text and file parts from an A2A response (task or message).
|
|
11
12
|
*/
|
|
@@ -15,7 +16,7 @@ function extractResult(result) {
|
|
|
15
16
|
const text = []
|
|
16
17
|
const files = []
|
|
17
18
|
|
|
18
|
-
const processParts = (parts
|
|
19
|
+
const processParts = (parts) => {
|
|
19
20
|
for (const part of parts) {
|
|
20
21
|
if (part.kind === "text") text.push(part.text)
|
|
21
22
|
else if (part.kind === "file") files.push(part)
|
|
@@ -23,8 +24,11 @@ function extractResult(result) {
|
|
|
23
24
|
}
|
|
24
25
|
|
|
25
26
|
if (result.kind === "task") {
|
|
26
|
-
|
|
27
|
-
|
|
27
|
+
if (result.artifacts)
|
|
28
|
+
for (let artifact of result.artifacts) {
|
|
29
|
+
if (!artifact.artifactId.startsWith("thinking")) processParts(artifact.parts)
|
|
30
|
+
}
|
|
31
|
+
else if (result.status?.message) processParts(result.status.message.parts)
|
|
28
32
|
if (text.length === 0 && files.length === 0) {
|
|
29
33
|
return { text: `Task ${result.id}: ${result.status?.state || "unknown"}`, files: [] }
|
|
30
34
|
}
|
|
@@ -71,21 +75,29 @@ function formatToolResult({ text, files }) {
|
|
|
71
75
|
* Wrap an A2A client as a LangChain tool the agent can call.
|
|
72
76
|
*/
|
|
73
77
|
function createA2ATool(client, agentCard) {
|
|
78
|
+
const subagent = agentCard.name
|
|
74
79
|
return tool(
|
|
75
80
|
async ({ message }) => {
|
|
76
81
|
try {
|
|
82
|
+
const messageId = cds.utils.uuid()
|
|
83
|
+
LOG.info(`Sending message to ${subagent}`, { messageId }, "\n\n" + message + "\n")
|
|
77
84
|
const result = await client.sendMessage({
|
|
78
85
|
message: {
|
|
79
86
|
kind: "message",
|
|
80
87
|
role: "user",
|
|
81
|
-
messageId
|
|
88
|
+
messageId,
|
|
82
89
|
parts: [{ kind: "text", text: message }],
|
|
83
90
|
},
|
|
84
91
|
})
|
|
85
|
-
|
|
92
|
+
if (LOG._debug)
|
|
93
|
+
LOG.trace(`Raw results from ${subagent}`, inspect(result, { depth: null, colors: true }))
|
|
94
|
+
let response = formatToolResult(extractResult(result))
|
|
95
|
+
if (response)
|
|
96
|
+
LOG.info(`Got response from ${subagent}`, { messageId }, "\n\n" + response + "\n")
|
|
97
|
+
return response
|
|
86
98
|
} catch (err) {
|
|
87
|
-
LOG.warn("Sub
|
|
88
|
-
return `Error communicating with ${
|
|
99
|
+
LOG.warn("Sub agent tool error", { subagent, error: err.message })
|
|
100
|
+
return `Error communicating with ${subagent}: ${err.message}`
|
|
89
101
|
}
|
|
90
102
|
},
|
|
91
103
|
{
|
|
@@ -106,7 +118,7 @@ export async function buildSubAgentToolLocally(serviceName) {
|
|
|
106
118
|
|
|
107
119
|
const { generateAgentCard } = await import("../../lib/protocol/agent-card.js")
|
|
108
120
|
const agentCard = generateAgentCard(srv)
|
|
109
|
-
LOG.info(`
|
|
121
|
+
LOG.info(`Connecting to sub agent ${serviceName}`, "(local)")
|
|
110
122
|
|
|
111
123
|
const { RequestContext, DefaultExecutionEventBus } = await import("@a2a-js/sdk/server")
|
|
112
124
|
|
|
@@ -164,6 +176,7 @@ export async function buildSubAgentToolLocally(serviceName) {
|
|
|
164
176
|
taskId,
|
|
165
177
|
contextId,
|
|
166
178
|
)
|
|
179
|
+
const truncated = message?.length > 80 ? message.slice(0, 80) + "..." : message
|
|
167
180
|
|
|
168
181
|
try {
|
|
169
182
|
// Run the sub-agent detached from the calling agent, in its own root
|
|
@@ -172,6 +185,12 @@ export async function buildSubAgentToolLocally(serviceName) {
|
|
|
172
185
|
await new Promise((resolve, reject) => {
|
|
173
186
|
cds
|
|
174
187
|
.spawn({ user: cds.context?.user, tenant: cds.context?.tenant }, async () => {
|
|
188
|
+
LOG.info("request", {
|
|
189
|
+
conversation: short(contextId),
|
|
190
|
+
service: srv.name,
|
|
191
|
+
text: truncated,
|
|
192
|
+
})
|
|
193
|
+
|
|
175
194
|
await executor.execute(requestContext, eventBus)
|
|
176
195
|
await done
|
|
177
196
|
})
|
|
@@ -273,7 +292,7 @@ export async function buildSubAgentToolFromConnection(serviceName) {
|
|
|
273
292
|
|
|
274
293
|
const path = typeof credentials === "object" ? credentials?.path : null
|
|
275
294
|
const base = agentBaseUrl.replace(/\/$/, "") + (path ? `/${path.replace(/^\//, "")}` : "")
|
|
276
|
-
LOG.info(`Connecting to sub
|
|
295
|
+
LOG.info(`Connecting to sub agent ${serviceName}`, { at: base })
|
|
277
296
|
|
|
278
297
|
// revisit: a2a agents may be tenant specific, card per tenant?
|
|
279
298
|
const initialHeaders = await resolveHeaders()
|
|
@@ -287,7 +306,6 @@ export async function buildSubAgentToolFromConnection(serviceName) {
|
|
|
287
306
|
)
|
|
288
307
|
}
|
|
289
308
|
const agentCard = await cardRes.json()
|
|
290
|
-
LOG.info(`Connected to sub-agent "${agentCard.name}" (${serviceName})`)
|
|
291
309
|
|
|
292
310
|
const { ClientFactory, ClientFactoryOptions, JsonRpcTransportFactory, RestTransportFactory } =
|
|
293
311
|
await import("@a2a-js/sdk/client")
|
package/srv/handlers/tools.js
CHANGED
|
@@ -11,7 +11,7 @@ import {
|
|
|
11
11
|
executeCallActionTool,
|
|
12
12
|
executePerActionTool,
|
|
13
13
|
} from "@cap-js/mcp/lib/tools.js"
|
|
14
|
-
import { getFilteredEntities, getFilteredActions } from "../../lib/utils/utils.js"
|
|
14
|
+
import { getFilteredEntities, getFilteredActions, getAgentLogger } from "../../lib/utils/utils.js"
|
|
15
15
|
import { isTextMime } from "../../lib/agents/markdown/backends/mime-utils.js"
|
|
16
16
|
import { checkAuthorization } from "@cap-js/mcp/lib/auth.js"
|
|
17
17
|
|
|
@@ -31,6 +31,7 @@ function cachedAuth(srv) {
|
|
|
31
31
|
|
|
32
32
|
class GenericReadTool extends DynamicStructuredTool {
|
|
33
33
|
constructor(srv, entities) {
|
|
34
|
+
const log = getAgentLogger(srv)
|
|
34
35
|
const def = createGenericReadToolDefinition(Object.keys(entities), srv.name, "")
|
|
35
36
|
super({
|
|
36
37
|
name: def.name,
|
|
@@ -38,7 +39,7 @@ class GenericReadTool extends DynamicStructuredTool {
|
|
|
38
39
|
schema: def.inputSchema,
|
|
39
40
|
responseFormat: "content_and_artifact",
|
|
40
41
|
func: async (args) => {
|
|
41
|
-
return unwrap(await executeGenericReadTool(srv, entities, args, { log
|
|
42
|
+
return unwrap(await executeGenericReadTool(srv, entities, args, { log }))
|
|
42
43
|
},
|
|
43
44
|
})
|
|
44
45
|
this.srv = srv
|
|
@@ -67,6 +68,7 @@ class GenericReadTool extends DynamicStructuredTool {
|
|
|
67
68
|
|
|
68
69
|
class DescribeTool extends DynamicStructuredTool {
|
|
69
70
|
constructor(srv, entities, actions) {
|
|
71
|
+
const log = getAgentLogger(srv)
|
|
70
72
|
const def = createDescribeToolDefinition(
|
|
71
73
|
Object.keys(entities),
|
|
72
74
|
Object.keys(actions),
|
|
@@ -79,7 +81,7 @@ class DescribeTool extends DynamicStructuredTool {
|
|
|
79
81
|
schema: def.inputSchema,
|
|
80
82
|
responseFormat: "content_and_artifact",
|
|
81
83
|
func: async (args) => {
|
|
82
|
-
return unwrap(await executeDescribe(srv, entities, actions, args, { log
|
|
84
|
+
return unwrap(await executeDescribe(srv, entities, actions, args, { log }))
|
|
83
85
|
},
|
|
84
86
|
})
|
|
85
87
|
this.srv = srv
|
|
@@ -119,6 +121,7 @@ class DescribeTool extends DynamicStructuredTool {
|
|
|
119
121
|
|
|
120
122
|
class PerActionTool extends DynamicStructuredTool {
|
|
121
123
|
constructor(srv, actionName, action) {
|
|
124
|
+
const log = getAgentLogger(srv)
|
|
122
125
|
const def = createPerActionToolDefinition(actionName, action, srv.name, srv.model, "")
|
|
123
126
|
super({
|
|
124
127
|
name: def.name,
|
|
@@ -126,7 +129,7 @@ class PerActionTool extends DynamicStructuredTool {
|
|
|
126
129
|
schema: def.inputSchema,
|
|
127
130
|
responseFormat: "content_and_artifact",
|
|
128
131
|
func: async (args) => {
|
|
129
|
-
return unwrap(await executePerActionTool(srv, actionName, action, args, { log
|
|
132
|
+
return unwrap(await executePerActionTool(srv, actionName, action, args, { log }))
|
|
130
133
|
},
|
|
131
134
|
})
|
|
132
135
|
this.srv = srv
|
|
@@ -142,6 +145,7 @@ class PerActionTool extends DynamicStructuredTool {
|
|
|
142
145
|
|
|
143
146
|
class CallActionTool extends DynamicStructuredTool {
|
|
144
147
|
constructor(srv, actions) {
|
|
148
|
+
const log = getAgentLogger(srv)
|
|
145
149
|
const def = createCallActionToolDefinition(Object.keys(actions), srv.name, "")
|
|
146
150
|
super({
|
|
147
151
|
name: def.name,
|
|
@@ -149,7 +153,7 @@ class CallActionTool extends DynamicStructuredTool {
|
|
|
149
153
|
schema: def.inputSchema,
|
|
150
154
|
responseFormat: "content_and_artifact",
|
|
151
155
|
func: async (args) => {
|
|
152
|
-
return unwrap(await executeCallActionTool(srv, actions, args, { log
|
|
156
|
+
return unwrap(await executeCallActionTool(srv, actions, args, { log }))
|
|
153
157
|
},
|
|
154
158
|
})
|
|
155
159
|
this.srv = srv
|
|
@@ -218,6 +222,10 @@ export function generateTools(srv) {
|
|
|
218
222
|
tools.push(createEmitFilePartTool())
|
|
219
223
|
}
|
|
220
224
|
|
|
225
|
+
if (cds.env.agents?.emitDataParts) {
|
|
226
|
+
tools.push(createEmitDataPartTool())
|
|
227
|
+
}
|
|
228
|
+
|
|
221
229
|
return tools
|
|
222
230
|
}
|
|
223
231
|
|
|
@@ -364,3 +372,29 @@ export function createReadFileTool(fileStore, contextId, userId) {
|
|
|
364
372
|
},
|
|
365
373
|
)
|
|
366
374
|
}
|
|
375
|
+
|
|
376
|
+
/**
|
|
377
|
+
* Create a tool that emits a DataPart in the A2A response.
|
|
378
|
+
* The executor's toolResults collection detects `kind: "data"`
|
|
379
|
+
* and emits the tool result as a data part.
|
|
380
|
+
*/
|
|
381
|
+
export function createEmitDataPartTool() {
|
|
382
|
+
return tool(
|
|
383
|
+
async ({ data, mediaType }) => {
|
|
384
|
+
return {
|
|
385
|
+
kind: "data",
|
|
386
|
+
data,
|
|
387
|
+
mediaType: mediaType ?? "application/json",
|
|
388
|
+
}
|
|
389
|
+
},
|
|
390
|
+
{
|
|
391
|
+
name: "emit_data_part",
|
|
392
|
+
description: "Emit a structured A2A DataPart. Only use when instructed.",
|
|
393
|
+
schema: z.object({
|
|
394
|
+
// A2A DataPart is specified to be an object in A2A 0.3
|
|
395
|
+
// https://a2a-protocol.org/v0.3.0/specification/#653-datapart-object
|
|
396
|
+
data: z.looseObject().describe("Structured object"),
|
|
397
|
+
}),
|
|
398
|
+
},
|
|
399
|
+
)
|
|
400
|
+
}
|
package/index.js
DELETED
|
File without changes
|
|
@@ -1,18 +0,0 @@
|
|
|
1
|
-
import { createMiddleware } from "langchain"
|
|
2
|
-
import { HumanMessage } from "@langchain/core/messages"
|
|
3
|
-
import { z } from "zod"
|
|
4
|
-
|
|
5
|
-
// Injects a HumanMessage for a pending _hitlEditNote before the next model turn.
|
|
6
|
-
export function hitlEditNoteInjectorMiddleware() {
|
|
7
|
-
return createMiddleware({
|
|
8
|
-
name: "hitlEditNoteInjectorMiddleware",
|
|
9
|
-
stateSchema: z.object({ _hitlEditNote: z.string().optional() }),
|
|
10
|
-
beforeModel: async (state) => {
|
|
11
|
-
if (!state._hitlEditNote) return
|
|
12
|
-
return {
|
|
13
|
-
messages: [new HumanMessage(state._hitlEditNote)],
|
|
14
|
-
_hitlEditNote: undefined,
|
|
15
|
-
}
|
|
16
|
-
},
|
|
17
|
-
})
|
|
18
|
-
}
|
package/lib/telemetry/mlflow.js
DELETED
|
@@ -1,290 +0,0 @@
|
|
|
1
|
-
/**
|
|
2
|
-
* MLflow span attribute helpers for @cap-js/agents.
|
|
3
|
-
*
|
|
4
|
-
* Adds mlflow.* attributes to existing OTel spans so the MLflow OTLP
|
|
5
|
-
* ingestion endpoint can assemble them into proper MLflow traces.
|
|
6
|
-
*
|
|
7
|
-
* All functions return {} when cds.env.agents.mlflow is falsy (zero overhead).
|
|
8
|
-
*
|
|
9
|
-
* The exporter itself is configured via setupMlflowExporter() as a
|
|
10
|
-
* RoutingSpanProcessor that lazily creates per-experiment-ID exporters —
|
|
11
|
-
* existing exporters (Dynatrace, Cloud Logging, Grafana, etc.) continue
|
|
12
|
-
* to work unchanged.
|
|
13
|
-
*/
|
|
14
|
-
import cds from "@sap/cds"
|
|
15
|
-
|
|
16
|
-
/**
|
|
17
|
-
* Resolve the MLflow experiment ID for the current request context.
|
|
18
|
-
*
|
|
19
|
-
* Resolution order:
|
|
20
|
-
* 1. @Core.SchemaVersion annotation on the service (feature-toggleable via cds.context.model)
|
|
21
|
-
* 2. cds.env.requires["mlflow"].credentials.MLFLOW_EXPERIMENT_ID
|
|
22
|
-
* 3. process.env.MLFLOW_EXPERIMENT_ID
|
|
23
|
-
*
|
|
24
|
-
* MLflow requires experiment IDs to be numeric (int64). Throws if annotation
|
|
25
|
-
* value is not a valid integer string.
|
|
26
|
-
*/
|
|
27
|
-
function resolveExperimentId() {
|
|
28
|
-
// Read from @Core.SchemaVersion on the active service (feature-toggle-aware)
|
|
29
|
-
const serviceName = cds.context?.["agent.service"]
|
|
30
|
-
if (serviceName) {
|
|
31
|
-
const definition =
|
|
32
|
-
cds.context?.model?.definitions?.[serviceName] || cds.services?.[serviceName]?.definition
|
|
33
|
-
const annotated = definition?.["@Core.SchemaVersion"]
|
|
34
|
-
if (annotated) {
|
|
35
|
-
const id = String(annotated)
|
|
36
|
-
if (!/^\d+$/.test(id)) {
|
|
37
|
-
throw new Error(
|
|
38
|
-
`@Core.SchemaVersion on "${serviceName}" must be a numeric string (MLflow experiment ID requires int64). Got: "${id}"`,
|
|
39
|
-
)
|
|
40
|
-
}
|
|
41
|
-
return id
|
|
42
|
-
}
|
|
43
|
-
}
|
|
44
|
-
|
|
45
|
-
// Fallback: credentials / env
|
|
46
|
-
const creds = cds.env.requires?.mlflow?.credentials || cds.env.requires?.mlflow?.credentials
|
|
47
|
-
return creds?.MLFLOW_EXPERIMENT_ID || process.env.MLFLOW_EXPERIMENT_ID
|
|
48
|
-
}
|
|
49
|
-
|
|
50
|
-
/**
|
|
51
|
-
* Build mlflow.* span attributes for a given span type.
|
|
52
|
-
* @param {"LLM"|"AGENT"|"TOOL"|"CHAIN"|"RETRIEVER"} spanType
|
|
53
|
-
* @param {{ inputs?: any, outputs?: any, model?: string, provider?: string, functionName?: string, tokenUsage?: object }} [opts]
|
|
54
|
-
* @returns {Record<string, string>} Attributes to set on span, or {} when disabled
|
|
55
|
-
*/
|
|
56
|
-
export function mlflowAttrs(spanType, opts = {}) {
|
|
57
|
-
if (!cds.env.agents?.mlflow) return {}
|
|
58
|
-
const experimentId = resolveExperimentId()
|
|
59
|
-
const attrs = {}
|
|
60
|
-
if (experimentId) attrs["mlflow.experimentId"] = String(experimentId)
|
|
61
|
-
attrs["mlflow.spanType"] = spanType
|
|
62
|
-
if (opts.model) attrs["mlflow.llm.model"] = String(opts.model)
|
|
63
|
-
if (opts.provider) attrs["mlflow.llm.provider"] = String(opts.provider)
|
|
64
|
-
if (opts.functionName) attrs["mlflow.spanFunctionName"] = String(opts.functionName)
|
|
65
|
-
if (opts.inputs !== undefined) attrs["mlflow.spanInputs"] = JSON.stringify(opts.inputs)
|
|
66
|
-
if (opts.outputs !== undefined) attrs["mlflow.spanOutputs"] = JSON.stringify(opts.outputs)
|
|
67
|
-
if (opts.tokenUsage) {
|
|
68
|
-
attrs["mlflow.chat.tokenUsage"] = JSON.stringify({
|
|
69
|
-
input_tokens: opts.tokenUsage.input_tokens,
|
|
70
|
-
output_tokens: opts.tokenUsage.output_tokens,
|
|
71
|
-
total_tokens: opts.tokenUsage.total_tokens,
|
|
72
|
-
cache_creation_input_tokens: opts.tokenUsage.cache_creation_input_tokens,
|
|
73
|
-
cache_read_input_tokens: opts.tokenUsage.cache_read_input_tokens,
|
|
74
|
-
})
|
|
75
|
-
}
|
|
76
|
-
return attrs
|
|
77
|
-
}
|
|
78
|
-
|
|
79
|
-
/**
|
|
80
|
-
* Build mlflow trace-level tag attributes (set on root/workflow span).
|
|
81
|
-
* Uses mlflow.traceTag.* prefix so MLflow server extracts them as trace tags.
|
|
82
|
-
* Also sets OTel semconv user.id and session.id which MLflow reads natively.
|
|
83
|
-
* @returns {Record<string, string>} Attributes to set on root span, or {} when disabled
|
|
84
|
-
*/
|
|
85
|
-
export function mlflowTraceAttrs() {
|
|
86
|
-
if (!cds.env.agents?.mlflow) return {}
|
|
87
|
-
const session = String(cds.context?.["agent.context.id"] || "")
|
|
88
|
-
const user = String(cds.context?.user?.id || "")
|
|
89
|
-
const tenant = String(cds.context?.tenant || "")
|
|
90
|
-
return {
|
|
91
|
-
// OTel semconv keys MLflow reads natively for user/session display
|
|
92
|
-
"session.id": session,
|
|
93
|
-
"user.id": user,
|
|
94
|
-
// mlflow.traceTag.* prefix for custom trace tags
|
|
95
|
-
"mlflow.traceTag.tenant": tenant,
|
|
96
|
-
}
|
|
97
|
-
}
|
|
98
|
-
|
|
99
|
-
/**
|
|
100
|
-
* Spread attributes object onto an OTel span (null-safe).
|
|
101
|
-
* Coerces all values to strings to prevent [object Object] in downstream systems.
|
|
102
|
-
* @param {import("@opentelemetry/api").Span | null} span
|
|
103
|
-
* @param {Record<string, string>} attrs
|
|
104
|
-
*/
|
|
105
|
-
export function setSpanAttrs(span, attrs) {
|
|
106
|
-
if (!cds.env.agents?.mlflow) return
|
|
107
|
-
if (!span) return
|
|
108
|
-
for (const [k, v] of Object.entries(attrs)) {
|
|
109
|
-
if (v !== undefined && v !== null) span.setAttribute(k, v)
|
|
110
|
-
}
|
|
111
|
-
}
|
|
112
|
-
|
|
113
|
-
/**
|
|
114
|
-
* SpanProcessor that routes spans to per-experiment-ID BatchSpanProcessors.
|
|
115
|
-
* Lazily creates an OTLPTraceExporter for each unique mlflow.experimentId
|
|
116
|
-
* seen on spans, so different services (or feature-toggled @Core.SchemaVersion
|
|
117
|
-
* values) export to the correct MLflow experiment.
|
|
118
|
-
*
|
|
119
|
-
* Supports both static headers (PAT token) and async header factories (OAuth).
|
|
120
|
-
*/
|
|
121
|
-
export class RoutingSpanProcessor {
|
|
122
|
-
constructor({ url, headersConfig, ucTableName, BatchSpanProcessor, OTLPTraceExporter }) {
|
|
123
|
-
this._url = url
|
|
124
|
-
this._headersConfig = headersConfig
|
|
125
|
-
this._ucTableName = ucTableName
|
|
126
|
-
this._BatchSpanProcessor = BatchSpanProcessor
|
|
127
|
-
this._OTLPTraceExporter = OTLPTraceExporter
|
|
128
|
-
this._processors = new Map()
|
|
129
|
-
}
|
|
130
|
-
|
|
131
|
-
onStart() {}
|
|
132
|
-
|
|
133
|
-
onEnd(span) {
|
|
134
|
-
const expId = span.attributes?.["mlflow.experimentId"]
|
|
135
|
-
if (!expId) return
|
|
136
|
-
this._getOrCreate(expId).onEnd(span)
|
|
137
|
-
}
|
|
138
|
-
|
|
139
|
-
_getOrCreate(experimentId) {
|
|
140
|
-
if (this._processors.has(experimentId)) return this._processors.get(experimentId)
|
|
141
|
-
const ucHeader = this._ucTableName
|
|
142
|
-
? { "X-Databricks-UC-Table-Name": this._ucTableName }
|
|
143
|
-
: undefined
|
|
144
|
-
// Build headers — async factory or static object
|
|
145
|
-
const headers =
|
|
146
|
-
typeof this._headersConfig === "function"
|
|
147
|
-
? async () => ({
|
|
148
|
-
...(await this._headersConfig()),
|
|
149
|
-
"x-mlflow-experiment-id": experimentId,
|
|
150
|
-
...ucHeader,
|
|
151
|
-
})
|
|
152
|
-
: {
|
|
153
|
-
...this._headersConfig,
|
|
154
|
-
"x-mlflow-experiment-id": experimentId,
|
|
155
|
-
...ucHeader,
|
|
156
|
-
}
|
|
157
|
-
const exporter = new this._OTLPTraceExporter({ url: this._url, headers })
|
|
158
|
-
const proc = new this._BatchSpanProcessor(exporter)
|
|
159
|
-
this._processors.set(experimentId, proc)
|
|
160
|
-
return proc
|
|
161
|
-
}
|
|
162
|
-
|
|
163
|
-
async forceFlush() {
|
|
164
|
-
await Promise.all([...this._processors.values()].map((p) => p.forceFlush()))
|
|
165
|
-
}
|
|
166
|
-
|
|
167
|
-
async shutdown() {
|
|
168
|
-
await Promise.all([...this._processors.values()].map((p) => p.shutdown()))
|
|
169
|
-
this._processors.clear()
|
|
170
|
-
}
|
|
171
|
-
}
|
|
172
|
-
|
|
173
|
-
/**
|
|
174
|
-
* Register MLflow OTLP RoutingSpanProcessor. No-op when credentials missing.
|
|
175
|
-
* Reads from cds.env.requires["mlflow"].credentials
|
|
176
|
-
*
|
|
177
|
-
* Auth: OAuth (clientid+clientsecret+url) takes precedence over static MLFLOW_TOKEN.
|
|
178
|
-
* UC: set UC_CATALOG + UC_SCHEMA + UC_TABLE_PREFIX for Unity Catalog trace storage.
|
|
179
|
-
*/
|
|
180
|
-
export async function setupMlflowExporter() {
|
|
181
|
-
if (!cds.env.agents?.mlflow) return
|
|
182
|
-
|
|
183
|
-
const LOG = cds.log("agents")
|
|
184
|
-
try {
|
|
185
|
-
const creds = cds.env.requires?.mlflow?.credentials || {}
|
|
186
|
-
const host = creds.MLFLOW_HOST
|
|
187
|
-
const token = creds.MLFLOW_TOKEN
|
|
188
|
-
const mlflowEndpoint =
|
|
189
|
-
creds.MLFLOW_OTLP_ENDPOINT || (host && `${host.replace(/\/$/, "")}/v1/traces`)
|
|
190
|
-
|
|
191
|
-
// Resolve authentication mode
|
|
192
|
-
let headersConfig
|
|
193
|
-
if (creds.clientid && creds.clientsecret) {
|
|
194
|
-
const tokenUrl = `${creds.url ?? host.replace(/\/$/, "")}/oidc/v1/token`
|
|
195
|
-
let cached = null
|
|
196
|
-
let expiresAt = 0
|
|
197
|
-
headersConfig = async function fetchToken() {
|
|
198
|
-
// Return cached token if still valid (60s buffer before expiry)
|
|
199
|
-
if (cached && Date.now() < expiresAt) return cached
|
|
200
|
-
const res = await fetch(tokenUrl, {
|
|
201
|
-
method: "POST",
|
|
202
|
-
headers: {
|
|
203
|
-
Authorization:
|
|
204
|
-
"Basic " + Buffer.from(`${creds.clientid}:${creds.clientsecret}`).toString("base64"),
|
|
205
|
-
"Content-Type": "application/x-www-form-urlencoded",
|
|
206
|
-
},
|
|
207
|
-
body: `grant_type=client_credentials&scope=all-apis`,
|
|
208
|
-
signal: AbortSignal.timeout(10_000),
|
|
209
|
-
})
|
|
210
|
-
if (!res.ok) {
|
|
211
|
-
const body = await res.text().catch(() => "")
|
|
212
|
-
const msg = `MLFlow oauth export failed: HTTP ${res.status} ${res.statusText}`
|
|
213
|
-
throw cds.error({ message: msg, response: body })
|
|
214
|
-
}
|
|
215
|
-
const { access_token, expires_in } = await res.json()
|
|
216
|
-
cached = { Authorization: `Bearer ${access_token}` }
|
|
217
|
-
expiresAt = Date.now() + (expires_in - 60) * 1000
|
|
218
|
-
return cached
|
|
219
|
-
}
|
|
220
|
-
LOG.debug("MLflow: using OAuth client credentials authentication")
|
|
221
|
-
} else if (token) {
|
|
222
|
-
headersConfig = { Authorization: `Bearer ${token}` }
|
|
223
|
-
} else {
|
|
224
|
-
headersConfig = {}
|
|
225
|
-
LOG.debug("MLflow: no auth credentials — assuming unauthenticated MLflow server")
|
|
226
|
-
}
|
|
227
|
-
|
|
228
|
-
if (!mlflowEndpoint) {
|
|
229
|
-
LOG.warn(
|
|
230
|
-
"MLflow: no endpoint configured (MLFLOW_HOST or MLFLOW_OTLP_ENDPOINT) — export disabled",
|
|
231
|
-
)
|
|
232
|
-
return
|
|
233
|
-
}
|
|
234
|
-
|
|
235
|
-
const ucCatalog = creds.UC_CATALOG
|
|
236
|
-
const ucSchema = creds.UC_SCHEMA
|
|
237
|
-
const ucTablePrefix = creds.UC_TABLE_PREFIX
|
|
238
|
-
const ucTableName =
|
|
239
|
-
ucCatalog && ucSchema && ucTablePrefix
|
|
240
|
-
? `${ucCatalog}.${ucSchema}.${ucTablePrefix}_otel_spans`
|
|
241
|
-
: undefined
|
|
242
|
-
|
|
243
|
-
const { trace } = await import("@opentelemetry/api")
|
|
244
|
-
const provider = trace.getTracerProvider()
|
|
245
|
-
const delegate = provider.getDelegate?.() || provider
|
|
246
|
-
|
|
247
|
-
let BatchSpanProcessor, OTLPTraceExporter
|
|
248
|
-
try {
|
|
249
|
-
const { createRequire } = await import("node:module")
|
|
250
|
-
const req = createRequire(process.cwd() + "/")
|
|
251
|
-
;({ BatchSpanProcessor } = req("@opentelemetry/sdk-trace-base"))
|
|
252
|
-
;({ OTLPTraceExporter } = req("@opentelemetry/exporter-trace-otlp-proto"))
|
|
253
|
-
} catch (err) {
|
|
254
|
-
LOG.warn(
|
|
255
|
-
"MLflow: @opentelemetry/exporter-trace-otlp-proto not resolvable from the application. " +
|
|
256
|
-
"Install a version matching your @cap-js/telemetry major " +
|
|
257
|
-
"(v1 → ^0.57, v2 → ^0.221). Export disabled.",
|
|
258
|
-
{ error: err.message },
|
|
259
|
-
)
|
|
260
|
-
return
|
|
261
|
-
}
|
|
262
|
-
|
|
263
|
-
const routing = new RoutingSpanProcessor({
|
|
264
|
-
url: mlflowEndpoint,
|
|
265
|
-
headersConfig,
|
|
266
|
-
ucTableName,
|
|
267
|
-
BatchSpanProcessor,
|
|
268
|
-
OTLPTraceExporter,
|
|
269
|
-
})
|
|
270
|
-
|
|
271
|
-
if (delegate.addSpanProcessor) {
|
|
272
|
-
// OTEL SDK v1
|
|
273
|
-
delegate.addSpanProcessor(routing)
|
|
274
|
-
} else if (delegate._activeSpanProcessor?._spanProcessors) {
|
|
275
|
-
// OTEL SDK v2: no addSpanProcessor — push into MultiSpanProcessor
|
|
276
|
-
delegate._activeSpanProcessor._spanProcessors.push(routing)
|
|
277
|
-
} else {
|
|
278
|
-
LOG.warn(
|
|
279
|
-
"MLflow: no TracerProvider with addSpanProcessor — ensure @cap-js/telemetry is loaded",
|
|
280
|
-
)
|
|
281
|
-
return
|
|
282
|
-
}
|
|
283
|
-
LOG.info("MLflow: routing span processor added (per-experiment export)", {
|
|
284
|
-
endpoint: mlflowEndpoint,
|
|
285
|
-
...(ucTableName && { ucTableName }),
|
|
286
|
-
})
|
|
287
|
-
} catch (err) {
|
|
288
|
-
LOG.error("MLflow: failed to configure OTLP export", { error: err.message })
|
|
289
|
-
}
|
|
290
|
-
}
|
|
File without changes
|