mesedi 0.7.1 → 0.7.2
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 +30 -20
- package/dist/cohere_integration.d.ts +1 -1
- package/dist/cohere_integration.js +1 -1
- package/dist/errors.d.ts +4 -4
- package/dist/errors.js +7 -7
- package/dist/gemini_integration.d.ts +1 -1
- package/dist/gemini_integration.js +1 -1
- package/dist/integrations/_integration_helpers.d.ts +4 -4
- package/dist/integrations/_integration_helpers.js +3 -3
- package/dist/integrations/langchain.d.ts +2 -2
- package/dist/integrations/langchain.js +2 -2
- package/dist/integrations/langgraph.d.ts +2 -2
- package/dist/integrations/langgraph.js +2 -2
- package/dist/integrations/mastra.d.ts +4 -4
- package/dist/integrations/mastra.js +3 -3
- package/dist/integrations/vectara.d.ts +2 -2
- package/dist/integrations/vectara.js +2 -2
- package/dist/observe.d.ts +2 -2
- package/dist/observe.js +1 -1
- package/dist/ollama_integration.d.ts +4 -4
- package/dist/ollama_integration.js +6 -6
- package/dist/openai_integration.d.ts +2 -2
- package/dist/openai_integration.js +2 -2
- package/dist/tool.js +1 -1
- package/dist/truncate.d.ts +3 -3
- package/dist/truncate.js +3 -3
- package/dist/vertex_gemini_integration.d.ts +2 -2
- package/dist/vertex_gemini_integration.js +2 -2
- package/dist/wrap.d.ts +1 -1
- package/package.json +1 -1
package/README.md
CHANGED
|
@@ -111,7 +111,7 @@ Descriptions are truncated at 2000 characters with an inline
|
|
|
111
111
|
|
|
112
112
|
## Optional: hard-halt with local budgets
|
|
113
113
|
|
|
114
|
-
Cap a single execution across four axes
|
|
114
|
+
Cap a single execution across four axes: input tokens, output tokens,
|
|
115
115
|
wall-clock seconds, and step count. Pass any subset; unset fields impose
|
|
116
116
|
no limit on that axis. When any budget is exceeded, the SDK throws
|
|
117
117
|
`MesediHalt` at the next safe boundary (between LLM calls, tool calls,
|
|
@@ -137,8 +137,7 @@ export const myAgent = wrap(
|
|
|
137
137
|
When a budget is supplied, the SDK also opens an SSE subscription to
|
|
138
138
|
`GET /executions/{id}/halt-stream`. Operators can halt a running
|
|
139
139
|
execution from the dashboard. If the SSE connection fails (backend
|
|
140
|
-
unreachable, 4xx/5xx, network partition), the reader logs and returns
|
|
141
|
-
— the wrapped agent keeps running with local budgets still enforced
|
|
140
|
+
unreachable, 4xx/5xx, network partition), the reader logs and returns. The wrapped agent keeps running with local budgets still enforced
|
|
142
141
|
client-side. Mesedi never decides to halt on its own; operator intent
|
|
143
142
|
or your own budget rules are the only triggers. `MesediHalt` carries an
|
|
144
143
|
internal Symbol marker so `wrap()` detects it even if user code
|
|
@@ -158,42 +157,53 @@ opt-in.
|
|
|
158
157
|
|
|
159
158
|
```typescript
|
|
160
159
|
import { configure, wrap } from "mesedi";
|
|
161
|
-
import {
|
|
160
|
+
import { instrumentLangGraph } from "mesedi/integrations/langgraph";
|
|
162
161
|
|
|
163
162
|
configure({ apiKey: process.env.MESEDI_API_KEY! });
|
|
164
163
|
|
|
164
|
+
let graph = buildMyGraph().compile();
|
|
165
|
+
graph = instrumentLangGraph(graph);
|
|
166
|
+
|
|
165
167
|
export const runMyGraph = wrap(async (question: string) => {
|
|
166
|
-
const
|
|
167
|
-
|
|
168
|
-
const result = await graph.invoke({ input: question });
|
|
169
|
-
return result.output;
|
|
168
|
+
const result = await graph.invoke({ question });
|
|
169
|
+
return result.answer;
|
|
170
170
|
});
|
|
171
171
|
```
|
|
172
172
|
|
|
173
|
-
`
|
|
174
|
-
|
|
175
|
-
|
|
176
|
-
|
|
173
|
+
`instrumentLangGraph` returns the same graph with its invoke methods
|
|
174
|
+
patched, so re-assigning in place is the intended usage. It emits
|
|
175
|
+
`llm_call` and `tool_call` events plus a `checkpoint` at every node
|
|
176
|
+
entry, labeled with the node name, so the dashboard timeline shows the
|
|
177
|
+
graph's flow alongside per-step detail.
|
|
177
178
|
|
|
178
179
|
### OpenAI Agents SDK
|
|
179
180
|
|
|
180
181
|
```typescript
|
|
181
182
|
import { configure, wrap } from "mesedi";
|
|
182
|
-
import {
|
|
183
|
+
import { MesediRunHooks } from "mesedi/integrations/openai_agents";
|
|
184
|
+
import { Agent, Runner } from "@openai/agents";
|
|
183
185
|
|
|
184
186
|
configure({ apiKey: process.env.MESEDI_API_KEY! });
|
|
185
187
|
|
|
186
188
|
export const runMyAgent = wrap(async (question: string) => {
|
|
187
|
-
const
|
|
188
|
-
|
|
189
|
-
|
|
189
|
+
const result = await Runner.run(
|
|
190
|
+
triageAgent,
|
|
191
|
+
question,
|
|
192
|
+
{ hooks: new MesediRunHooks() },
|
|
193
|
+
);
|
|
194
|
+
return result.finalOutput;
|
|
190
195
|
});
|
|
191
196
|
```
|
|
192
197
|
|
|
193
|
-
`
|
|
194
|
-
|
|
195
|
-
|
|
196
|
-
|
|
198
|
+
`MesediRunHooks` implements the SDK's `RunHooks` interface. It emits a
|
|
199
|
+
`checkpoint` on every agent start and end, an `agent_handoff` on every
|
|
200
|
+
transfer between agents, and a `tool_call` per tool invocation.
|
|
201
|
+
|
|
202
|
+
It does **not** emit `llm_call` events, because the Agents SDK dispatches
|
|
203
|
+
model calls through its own runner and exposes no per-call hook. The
|
|
204
|
+
detectors that read `llm_call` (drift, identical-call and similar-call
|
|
205
|
+
loops, cost velocity) therefore do not fire on an OpenAI-Agents-only
|
|
206
|
+
deployment.
|
|
197
207
|
|
|
198
208
|
### Vercel AI SDK
|
|
199
209
|
|
package/dist/errors.d.ts
CHANGED
|
@@ -5,7 +5,7 @@
|
|
|
5
5
|
* openai_integration, cohere_integration, etc.) classifies its
|
|
6
6
|
* native exception types into ONE of the values in ErrorClass. The
|
|
7
7
|
* backend reads `error_class` from llm_call payloads and uses the
|
|
8
|
-
* canonical vocabulary to cluster cross-provider signals
|
|
8
|
+
* canonical vocabulary to cluster cross-provider signals: e.g.
|
|
9
9
|
* when both Anthropic and OpenAI are rate-limiting at the same
|
|
10
10
|
* time, the `provider_incident` detector sees them under the same
|
|
11
11
|
* `rate_limited` bucket regardless of which exception class each
|
|
@@ -61,7 +61,7 @@ export declare function classifyAnthropicException(err: unknown): ErrorClassValu
|
|
|
61
61
|
* Two-step classification:
|
|
62
62
|
* 1. Look up class name in OPENAI_EXCEPTION_MAP.
|
|
63
63
|
* 2. If result is RATE_LIMITED, probe the body for
|
|
64
|
-
* `error.code === "insufficient_quota"
|
|
64
|
+
* `error.code === "insufficient_quota"`: OpenAI overloads
|
|
65
65
|
* RateLimitError to cover both true rate limit and billing-
|
|
66
66
|
* cap exhaustion. The QUOTA_EXHAUSTED bucket needs different
|
|
67
67
|
* remediation (raise quota vs. back off), so the detector
|
|
@@ -81,7 +81,7 @@ export declare function classifyCohereException(err: unknown): ErrorClassValue;
|
|
|
81
81
|
* Two-step classification:
|
|
82
82
|
* 1. Look up class name in GEMINI_EXCEPTION_MAP.
|
|
83
83
|
* 2. If result is RATE_LIMITED (ResourceExhausted), probe the
|
|
84
|
-
* exception message for substring "quota"
|
|
84
|
+
* exception message for substring "quota": Google overloads
|
|
85
85
|
* ResourceExhausted to cover both true rate-limit and
|
|
86
86
|
* quota-exhaust.
|
|
87
87
|
*
|
|
@@ -119,7 +119,7 @@ export declare function classifyByProvider(provider: string, err: unknown): Erro
|
|
|
119
119
|
/**
|
|
120
120
|
* Variant of `classifyByProvider` that dispatches on an exception
|
|
121
121
|
* class NAME (string) rather than an Error object. Mastra's
|
|
122
|
-
* SpanErrorInfo carries `.name` as a bare string
|
|
122
|
+
* SpanErrorInfo carries `.name` as a bare string: the underlying
|
|
123
123
|
* Error object is not preserved through the exporter chain. This
|
|
124
124
|
* shim wraps the name in a synthetic Error so the same class-name
|
|
125
125
|
* lookup path in the per-provider classifiers works unchanged.
|
package/dist/errors.js
CHANGED
|
@@ -5,7 +5,7 @@
|
|
|
5
5
|
* openai_integration, cohere_integration, etc.) classifies its
|
|
6
6
|
* native exception types into ONE of the values in ErrorClass. The
|
|
7
7
|
* backend reads `error_class` from llm_call payloads and uses the
|
|
8
|
-
* canonical vocabulary to cluster cross-provider signals
|
|
8
|
+
* canonical vocabulary to cluster cross-provider signals: e.g.
|
|
9
9
|
* when both Anthropic and OpenAI are rate-limiting at the same
|
|
10
10
|
* time, the `provider_incident` detector sees them under the same
|
|
11
11
|
* `rate_limited` bucket regardless of which exception class each
|
|
@@ -42,7 +42,7 @@ export const PROVIDER_SIDE_ERROR_CLASSES = PROVIDER_SIDE_ERROR_CLASS_VALUES;
|
|
|
42
42
|
/**
|
|
43
43
|
* Mapping from Anthropic exception class names (string, not the
|
|
44
44
|
* class object) to canonical ErrorClass values. Keyed by name to
|
|
45
|
-
* avoid taking a hard runtime dependency on @anthropic-ai/sdk
|
|
45
|
+
* avoid taking a hard runtime dependency on @anthropic-ai/sdk:
|
|
46
46
|
* the mesedi SDK works even if the anthropic package is not
|
|
47
47
|
* installed, and the mapping stays correct because exception class
|
|
48
48
|
* names are stable across anthropic SDK versions.
|
|
@@ -110,7 +110,7 @@ export function classifyAnthropicException(err) {
|
|
|
110
110
|
* mapping is stable regardless of which casing the runtime presents.
|
|
111
111
|
*
|
|
112
112
|
* RateLimitError is overloaded by OpenAI to cover both true rate
|
|
113
|
-
* limiting AND insufficient_quota
|
|
113
|
+
* limiting AND insufficient_quota: classifyOpenAIException probes
|
|
114
114
|
* the body to distinguish, returning QUOTA_EXHAUSTED for the latter.
|
|
115
115
|
*/
|
|
116
116
|
const OPENAI_EXCEPTION_MAP = {
|
|
@@ -137,7 +137,7 @@ const OPENAI_EXCEPTION_MAP = {
|
|
|
137
137
|
* Two-step classification:
|
|
138
138
|
* 1. Look up class name in OPENAI_EXCEPTION_MAP.
|
|
139
139
|
* 2. If result is RATE_LIMITED, probe the body for
|
|
140
|
-
* `error.code === "insufficient_quota"
|
|
140
|
+
* `error.code === "insufficient_quota"`: OpenAI overloads
|
|
141
141
|
* RateLimitError to cover both true rate limit and billing-
|
|
142
142
|
* cap exhaustion. The QUOTA_EXHAUSTED bucket needs different
|
|
143
143
|
* remediation (raise quota vs. back off), so the detector
|
|
@@ -216,7 +216,7 @@ const GEMINI_EXCEPTION_MAP = {
|
|
|
216
216
|
* Two-step classification:
|
|
217
217
|
* 1. Look up class name in GEMINI_EXCEPTION_MAP.
|
|
218
218
|
* 2. If result is RATE_LIMITED (ResourceExhausted), probe the
|
|
219
|
-
* exception message for substring "quota"
|
|
219
|
+
* exception message for substring "quota": Google overloads
|
|
220
220
|
* ResourceExhausted to cover both true rate-limit and
|
|
221
221
|
* quota-exhaust.
|
|
222
222
|
*
|
|
@@ -242,7 +242,7 @@ export function classifyGeminiException(err) {
|
|
|
242
242
|
* QUOTA_EXHAUSTED are absent from this map by design.
|
|
243
243
|
*
|
|
244
244
|
* ResponseError (Ollama-native HTTP error) carries .status_code and
|
|
245
|
-
* is bucketed by HTTP-code range inside classifyOllamaException
|
|
245
|
+
* is bucketed by HTTP-code range inside classifyOllamaException:
|
|
246
246
|
* NOT mapped here because the class name alone doesn't carry the
|
|
247
247
|
* 4xx-vs-5xx distinction the detectors require. */
|
|
248
248
|
const OLLAMA_EXCEPTION_MAP = {
|
|
@@ -335,7 +335,7 @@ export function classifyByProvider(provider, err) {
|
|
|
335
335
|
/**
|
|
336
336
|
* Variant of `classifyByProvider` that dispatches on an exception
|
|
337
337
|
* class NAME (string) rather than an Error object. Mastra's
|
|
338
|
-
* SpanErrorInfo carries `.name` as a bare string
|
|
338
|
+
* SpanErrorInfo carries `.name` as a bare string: the underlying
|
|
339
339
|
* Error object is not preserved through the exporter chain. This
|
|
340
340
|
* shim wraps the name in a synthetic Error so the same class-name
|
|
341
341
|
* lookup path in the per-provider classifiers works unchanged.
|
|
@@ -2,7 +2,7 @@
|
|
|
2
2
|
* Shared helpers for framework-adapter integration tests
|
|
3
3
|
* (mastra.integration.test.ts, langchain.integration.test.ts).
|
|
4
4
|
*
|
|
5
|
-
* Design mirrors backend/test/integration/conftest.py
|
|
5
|
+
* Design mirrors backend/test/integration/conftest.py: same wire
|
|
6
6
|
* protocol against the real Mesedi backend, same fresh-per-session
|
|
7
7
|
* project mint, same `awaitFailureGroup` polling contract.
|
|
8
8
|
*
|
|
@@ -17,12 +17,12 @@
|
|
|
17
17
|
* if (!INTEGRATION_ENABLED) return test.skip(...)
|
|
18
18
|
*
|
|
19
19
|
* Set `RUN_INTEGRATION_TESTS=1` to opt in. Default `npm test` skips
|
|
20
|
-
* everything in this file
|
|
20
|
+
* everything in this file: unit tests stay fast.
|
|
21
21
|
*/
|
|
22
22
|
export declare const INTEGRATION_ENABLED: boolean;
|
|
23
23
|
/**
|
|
24
24
|
* Secondary gate for tests that make real LLM calls (Anthropic).
|
|
25
|
-
* These add real cost
|
|
25
|
+
* These add real cost: ~$0.01-$0.20 per test depending on payload
|
|
26
26
|
* size. Set `ANTHROPIC_API_KEY` in env to unlock. Absent → skip with
|
|
27
27
|
* a clear reason.
|
|
28
28
|
*/
|
|
@@ -58,7 +58,7 @@ export interface AwaitFailureGroupOptions {
|
|
|
58
58
|
signaturePrefix?: string;
|
|
59
59
|
/** accept any one of a list of alternative (failure_class,
|
|
60
60
|
* signaturePrefix?) pairs in addition to the primary match. Useful
|
|
61
|
-
* when a signal has multiple valid backend clustering shapes
|
|
61
|
+
* when a signal has multiple valid backend clustering shapes: e.g.
|
|
62
62
|
* near-duplicate LLM calls fire as EITHER `loops/similar_call_*` OR
|
|
63
63
|
* `drift/lexical_drift_*` depending on detector sensitivity. Both
|
|
64
64
|
* are correct product signals; the test succeeds if ANY of the
|
|
@@ -2,7 +2,7 @@
|
|
|
2
2
|
* Shared helpers for framework-adapter integration tests
|
|
3
3
|
* (mastra.integration.test.ts, langchain.integration.test.ts).
|
|
4
4
|
*
|
|
5
|
-
* Design mirrors backend/test/integration/conftest.py
|
|
5
|
+
* Design mirrors backend/test/integration/conftest.py: same wire
|
|
6
6
|
* protocol against the real Mesedi backend, same fresh-per-session
|
|
7
7
|
* project mint, same `awaitFailureGroup` polling contract.
|
|
8
8
|
*
|
|
@@ -18,12 +18,12 @@
|
|
|
18
18
|
* if (!INTEGRATION_ENABLED) return test.skip(...)
|
|
19
19
|
*
|
|
20
20
|
* Set `RUN_INTEGRATION_TESTS=1` to opt in. Default `npm test` skips
|
|
21
|
-
* everything in this file
|
|
21
|
+
* everything in this file: unit tests stay fast.
|
|
22
22
|
*/
|
|
23
23
|
export const INTEGRATION_ENABLED = process.env["RUN_INTEGRATION_TESTS"] === "1";
|
|
24
24
|
/**
|
|
25
25
|
* Secondary gate for tests that make real LLM calls (Anthropic).
|
|
26
|
-
* These add real cost
|
|
26
|
+
* These add real cost: ~$0.01-$0.20 per test depending on payload
|
|
27
27
|
* size. Set `ANTHROPIC_API_KEY` in env to unlock. Absent → skip with
|
|
28
28
|
* a clear reason.
|
|
29
29
|
*/
|
|
@@ -29,7 +29,7 @@
|
|
|
29
29
|
*
|
|
30
30
|
* `mesedi.wrap()` manages the execution boundary (started / completed /
|
|
31
31
|
* crash signature). The callback handler emits the intra-execution
|
|
32
|
-
* events
|
|
32
|
+
* events, `llm_call` and `tool_call`, that Mesedi's detectors
|
|
33
33
|
* consume. Splitting responsibility this way lets the customer adopt
|
|
34
34
|
* Mesedi without reshaping their LangChain code: `wrap` goes around
|
|
35
35
|
* the entry point, the handler gets attached at the existing
|
|
@@ -68,7 +68,7 @@ import type { BaseMessage } from "@langchain/core/messages";
|
|
|
68
68
|
* format from `emit_llm_call` and the Anthropic patch) and one
|
|
69
69
|
* `tool_call` event per tool invocation (matching the wire format
|
|
70
70
|
* from `@mesedi.tool`). Both event types feed the standard Mesedi
|
|
71
|
-
* detector chain
|
|
71
|
+
* detector chain: drift, identical / similar-call loops,
|
|
72
72
|
* tool-failures, cost-velocity, prompt-injection.
|
|
73
73
|
*
|
|
74
74
|
* Outside a `wrap()` execution context, all emissions silently no-op
|
|
@@ -29,7 +29,7 @@
|
|
|
29
29
|
*
|
|
30
30
|
* `mesedi.wrap()` manages the execution boundary (started / completed /
|
|
31
31
|
* crash signature). The callback handler emits the intra-execution
|
|
32
|
-
* events
|
|
32
|
+
* events, `llm_call` and `tool_call`, that Mesedi's detectors
|
|
33
33
|
* consume. Splitting responsibility this way lets the customer adopt
|
|
34
34
|
* Mesedi without reshaping their LangChain code: `wrap` goes around
|
|
35
35
|
* the entry point, the handler gets attached at the existing
|
|
@@ -94,7 +94,7 @@ const MAX_EXC_MSG = 500;
|
|
|
94
94
|
* format from `emit_llm_call` and the Anthropic patch) and one
|
|
95
95
|
* `tool_call` event per tool invocation (matching the wire format
|
|
96
96
|
* from `@mesedi.tool`). Both event types feed the standard Mesedi
|
|
97
|
-
* detector chain
|
|
97
|
+
* detector chain: drift, identical / similar-call loops,
|
|
98
98
|
* tool-failures, cost-velocity, prompt-injection.
|
|
99
99
|
*
|
|
100
100
|
* Outside a `wrap()` execution context, all emissions silently no-op
|
|
@@ -39,9 +39,9 @@
|
|
|
39
39
|
* Out of scope for v1:
|
|
40
40
|
*
|
|
41
41
|
* - Streaming hooks (`graph.streamEvents`). Covered in a follow-up.
|
|
42
|
-
* - LangGraph's persistence (checkpointer)
|
|
42
|
+
* - LangGraph's persistence (checkpointer): Mesedi emits
|
|
43
43
|
* telemetry parallel to it without touching the stored state.
|
|
44
|
-
* - LangGraph `interrupt()` auto-bridging
|
|
44
|
+
* - LangGraph `interrupt()` auto-bridging: use the existing HITL
|
|
45
45
|
* helpers (`pauseForHuman`, `requestHumanIntervention`) for now.
|
|
46
46
|
*/
|
|
47
47
|
/**
|
|
@@ -39,9 +39,9 @@
|
|
|
39
39
|
* Out of scope for v1:
|
|
40
40
|
*
|
|
41
41
|
* - Streaming hooks (`graph.streamEvents`). Covered in a follow-up.
|
|
42
|
-
* - LangGraph's persistence (checkpointer)
|
|
42
|
+
* - LangGraph's persistence (checkpointer): Mesedi emits
|
|
43
43
|
* telemetry parallel to it without touching the stored state.
|
|
44
|
-
* - LangGraph `interrupt()` auto-bridging
|
|
44
|
+
* - LangGraph `interrupt()` auto-bridging: use the existing HITL
|
|
45
45
|
* helpers (`pauseForHuman`, `requestHumanIntervention`) for now.
|
|
46
46
|
*/
|
|
47
47
|
import { getClient } from "../client.js";
|
|
@@ -16,7 +16,7 @@
|
|
|
16
16
|
* configure({ apiKey: process.env.MESEDI_API_KEY });
|
|
17
17
|
*
|
|
18
18
|
* // Then wire the exporter into Mastra. Mastra 1.x requires the
|
|
19
|
-
* // `configs` shape below
|
|
19
|
+
* // `configs` shape below: the `exporters: {...}` root shape from
|
|
20
20
|
* // pre-1.x is silently ignored and no spans reach the exporter.
|
|
21
21
|
* const mastra = new Mastra({
|
|
22
22
|
* observability: new Observability({
|
|
@@ -38,7 +38,7 @@
|
|
|
38
38
|
* observability integration is to implement an Observability exporter
|
|
39
39
|
* that consumes those spans. Unlike the other Mesedi integrations
|
|
40
40
|
* (LangGraph, OpenAI Agents, Vercel AI SDK), the customer does NOT
|
|
41
|
-
* need to wrap their entry function with `mesedi.wrap()
|
|
41
|
+
* need to wrap their entry function with `mesedi.wrap()`: Mastra's
|
|
42
42
|
* AGENT_RUN / WORKFLOW_RUN root span is the execution boundary. The
|
|
43
43
|
* exporter opens a Mesedi execution on root-span-start and closes it
|
|
44
44
|
* on root-span-end. All descendant spans on the same trace flow into
|
|
@@ -54,7 +54,7 @@
|
|
|
54
54
|
* - Other span types (RAG_*, MEMORY_OPERATION, WORKSPACE_ACTION,
|
|
55
55
|
* SCORER_*, MAPPING, PROCESSOR_RUN, MODEL_STEP / MODEL_INFERENCE /
|
|
56
56
|
* MODEL_CHUNK) are silently ignored in v1. MODEL_GENERATION is
|
|
57
|
-
* the aggregating span for a full model call
|
|
57
|
+
* the aggregating span for a full model call: the finer-grained
|
|
58
58
|
* step / inference / chunk spans are noise for Mesedi's detectors
|
|
59
59
|
* and are already rolled up by Mastra.
|
|
60
60
|
*
|
|
@@ -66,7 +66,7 @@ import { BaseExporter } from "@mastra/observability";
|
|
|
66
66
|
import type { TracingEvent } from "@mastra/core/observability";
|
|
67
67
|
/**
|
|
68
68
|
* Configuration accepted by `new MesediExporter(...)`. Kept tiny on
|
|
69
|
-
* purpose
|
|
69
|
+
* purpose: the SDK's `configure()` already owns apiKey / baseUrl /
|
|
70
70
|
* shipper knobs globally; passing them per-exporter would produce two
|
|
71
71
|
* shippers if the customer ever instantiates two Mastra apps in the
|
|
72
72
|
* same process.
|
|
@@ -16,7 +16,7 @@
|
|
|
16
16
|
* configure({ apiKey: process.env.MESEDI_API_KEY });
|
|
17
17
|
*
|
|
18
18
|
* // Then wire the exporter into Mastra. Mastra 1.x requires the
|
|
19
|
-
* // `configs` shape below
|
|
19
|
+
* // `configs` shape below: the `exporters: {...}` root shape from
|
|
20
20
|
* // pre-1.x is silently ignored and no spans reach the exporter.
|
|
21
21
|
* const mastra = new Mastra({
|
|
22
22
|
* observability: new Observability({
|
|
@@ -38,7 +38,7 @@
|
|
|
38
38
|
* observability integration is to implement an Observability exporter
|
|
39
39
|
* that consumes those spans. Unlike the other Mesedi integrations
|
|
40
40
|
* (LangGraph, OpenAI Agents, Vercel AI SDK), the customer does NOT
|
|
41
|
-
* need to wrap their entry function with `mesedi.wrap()
|
|
41
|
+
* need to wrap their entry function with `mesedi.wrap()`: Mastra's
|
|
42
42
|
* AGENT_RUN / WORKFLOW_RUN root span is the execution boundary. The
|
|
43
43
|
* exporter opens a Mesedi execution on root-span-start and closes it
|
|
44
44
|
* on root-span-end. All descendant spans on the same trace flow into
|
|
@@ -54,7 +54,7 @@
|
|
|
54
54
|
* - Other span types (RAG_*, MEMORY_OPERATION, WORKSPACE_ACTION,
|
|
55
55
|
* SCORER_*, MAPPING, PROCESSOR_RUN, MODEL_STEP / MODEL_INFERENCE /
|
|
56
56
|
* MODEL_CHUNK) are silently ignored in v1. MODEL_GENERATION is
|
|
57
|
-
* the aggregating span for a full model call
|
|
57
|
+
* the aggregating span for a full model call: the finer-grained
|
|
58
58
|
* step / inference / chunk spans are noise for Mesedi's detectors
|
|
59
59
|
* and are already rolled up by Mastra.
|
|
60
60
|
*
|
|
@@ -9,13 +9,13 @@
|
|
|
9
9
|
* use this helper to report scores into Mesedi without hand-writing
|
|
10
10
|
* the boilerplate around {@link emitEvalScore}.
|
|
11
11
|
*
|
|
12
|
-
* IMPORTANT
|
|
12
|
+
* IMPORTANT: direction note. HHEM scores have INVERTED semantics
|
|
13
13
|
* compared to Ragas / Promptfoo metrics: HIGHER means MORE faithful
|
|
14
14
|
* to context (less hallucination). The helper passes
|
|
15
15
|
* `higherIsBetter: true` to the backend and the parameter is named
|
|
16
16
|
* `faithfulnessScore` to mirror HHEM's documented output. Some HHEM
|
|
17
17
|
* consumers transform the output into a "hallucination_score" where
|
|
18
|
-
* LOWER is better
|
|
18
|
+
* LOWER is better: if your pipeline does that transformation,
|
|
19
19
|
* invert your value before calling this helper (i.e. pass
|
|
20
20
|
* `1 - hallucinationScore`).
|
|
21
21
|
*
|
|
@@ -9,13 +9,13 @@
|
|
|
9
9
|
* use this helper to report scores into Mesedi without hand-writing
|
|
10
10
|
* the boilerplate around {@link emitEvalScore}.
|
|
11
11
|
*
|
|
12
|
-
* IMPORTANT
|
|
12
|
+
* IMPORTANT: direction note. HHEM scores have INVERTED semantics
|
|
13
13
|
* compared to Ragas / Promptfoo metrics: HIGHER means MORE faithful
|
|
14
14
|
* to context (less hallucination). The helper passes
|
|
15
15
|
* `higherIsBetter: true` to the backend and the parameter is named
|
|
16
16
|
* `faithfulnessScore` to mirror HHEM's documented output. Some HHEM
|
|
17
17
|
* consumers transform the output into a "hallucination_score" where
|
|
18
|
-
* LOWER is better
|
|
18
|
+
* LOWER is better: if your pipeline does that transformation,
|
|
19
19
|
* invert your value before calling this helper (i.e. pass
|
|
20
20
|
* `1 - hallucinationScore`).
|
|
21
21
|
*
|
package/dist/observe.d.ts
CHANGED
|
@@ -262,7 +262,7 @@ export interface AgentHandoffOptions {
|
|
|
262
262
|
/**
|
|
263
263
|
* Object-form options for the new {@link emitAgentHandoff} overload
|
|
264
264
|
* (). Use this when calling inside a `wrap()` whose
|
|
265
|
-
* `agentName` option supplies the source agent
|
|
265
|
+
* `agentName` option supplies the source agent: the `fromAgent`
|
|
266
266
|
* field becomes optional and falls back to the wrapped execution's
|
|
267
267
|
* agent name.
|
|
268
268
|
*/
|
|
@@ -293,7 +293,7 @@ export interface EmitAgentHandoffArgs extends AgentHandoffOptions {
|
|
|
293
293
|
* `wrap({ agentName: ... }, fn)` is used.
|
|
294
294
|
* 3. If neither is supplied while inside `wrap()`, this function
|
|
295
295
|
* throws Error rather than silently emitting an `unknown` source
|
|
296
|
-
* agent
|
|
296
|
+
* agent: polluting the topology graph would degrade
|
|
297
297
|
* cascading_failure / coordination_deadlock clustering at scale.
|
|
298
298
|
*
|
|
299
299
|
* Outside `wrap()` / `runInExecutionContext`: no-op (matches the
|
package/dist/observe.js
CHANGED
|
@@ -106,7 +106,7 @@ export function validatorResult(name, passed, opts = {}) {
|
|
|
106
106
|
client.submitEvent(event);
|
|
107
107
|
}
|
|
108
108
|
/**
|
|
109
|
-
* Canonical error_class values that signal per-tenant throttling
|
|
109
|
+
* Canonical error_class values that signal per-tenant throttling: used
|
|
110
110
|
* by the instrument_* modules to decide whether to auto-emit an
|
|
111
111
|
* infrastructure_event alongside their failed llm_call event so the
|
|
112
112
|
* infrastructure_throttled detector picks up the pattern. Deliberately
|
|
@@ -1,16 +1,16 @@
|
|
|
1
1
|
/**
|
|
2
|
-
* Ollama SDK monkey-patch
|
|
2
|
+
* Ollama SDK monkey-patch: auto-emit llm_call events for every
|
|
3
3
|
* `Ollama.chat()` call inside a wrap()'d execution.
|
|
4
4
|
*
|
|
5
5
|
* Mirrors instrumentOpenAI / instrumentAnthropic shape (same
|
|
6
6
|
* fail-open semantics, same payload contract, same idempotency
|
|
7
7
|
* guarantee), with Ollama-shape adjustments:
|
|
8
8
|
*
|
|
9
|
-
* - Token field names: prompt_eval_count / eval_count
|
|
9
|
+
* - Token field names: prompt_eval_count / eval_count: translated
|
|
10
10
|
* to canonical input_tokens / output_tokens on payload.
|
|
11
11
|
* - Response text is `response.message.content` (not `choices[0]...`).
|
|
12
12
|
* - No API-key auth; customers point at localhost or a remote
|
|
13
|
-
* Ollama host. No throttling auto-emit in this sub-wave
|
|
13
|
+
* Ollama host. No throttling auto-emit in this sub-wave: for a
|
|
14
14
|
* local runtime there is no upstream rate-limit signal worth
|
|
15
15
|
* surfacing.
|
|
16
16
|
*
|
|
@@ -118,7 +118,7 @@ export declare const _testing: {
|
|
|
118
118
|
* full wrap() + chat() integration scaffold. Not part of the
|
|
119
119
|
* public API. */
|
|
120
120
|
maybeWarnStreamingUnsupported: typeof maybeWarnStreamingUnsupported;
|
|
121
|
-
|
|
121
|
+
/**: exposed for unit tests of the streaming chunk-
|
|
122
122
|
* aggregation logic. Not part of the public API. */
|
|
123
123
|
accumulateChunk: typeof accumulateChunk;
|
|
124
124
|
};
|
|
@@ -1,16 +1,16 @@
|
|
|
1
1
|
/**
|
|
2
|
-
* Ollama SDK monkey-patch
|
|
2
|
+
* Ollama SDK monkey-patch: auto-emit llm_call events for every
|
|
3
3
|
* `Ollama.chat()` call inside a wrap()'d execution.
|
|
4
4
|
*
|
|
5
5
|
* Mirrors instrumentOpenAI / instrumentAnthropic shape (same
|
|
6
6
|
* fail-open semantics, same payload contract, same idempotency
|
|
7
7
|
* guarantee), with Ollama-shape adjustments:
|
|
8
8
|
*
|
|
9
|
-
* - Token field names: prompt_eval_count / eval_count
|
|
9
|
+
* - Token field names: prompt_eval_count / eval_count: translated
|
|
10
10
|
* to canonical input_tokens / output_tokens on payload.
|
|
11
11
|
* - Response text is `response.message.content` (not `choices[0]...`).
|
|
12
12
|
* - No API-key auth; customers point at localhost or a remote
|
|
13
|
-
* Ollama host. No throttling auto-emit in this sub-wave
|
|
13
|
+
* Ollama host. No throttling auto-emit in this sub-wave: for a
|
|
14
14
|
* local runtime there is no upstream rate-limit signal worth
|
|
15
15
|
* surfacing.
|
|
16
16
|
*
|
|
@@ -47,11 +47,11 @@ const MAX_RESPONSE = 1000;
|
|
|
47
47
|
const MAX_EXC_MSG = 500;
|
|
48
48
|
/** Stable lowercase provider identifier shipped on every llm_call
|
|
49
49
|
* event. Backend's provider_incident detector clusters cross-tenant
|
|
50
|
-
* signals on (provider, error_class)
|
|
50
|
+
* signals on (provider, error_class), though for local runtimes the
|
|
51
51
|
* cross-tenant signal is necessarily quiet. */
|
|
52
52
|
const PROVIDER = "ollama";
|
|
53
53
|
const _patched = new WeakSet();
|
|
54
|
-
|
|
54
|
+
/**: one-time guard for streaming calls. The
|
|
55
55
|
* chunk-aggregating wrapper for streaming responses ships in a
|
|
56
56
|
* follow-up sub-wave; until then, calls with stream: true emit NO
|
|
57
57
|
* llm_call event AND log a single console.info so customers
|
|
@@ -397,7 +397,7 @@ export const _testing = {
|
|
|
397
397
|
* full wrap() + chat() integration scaffold. Not part of the
|
|
398
398
|
* public API. */
|
|
399
399
|
maybeWarnStreamingUnsupported,
|
|
400
|
-
|
|
400
|
+
/**: exposed for unit tests of the streaming chunk-
|
|
401
401
|
* aggregation logic. Not part of the public API. */
|
|
402
402
|
accumulateChunk,
|
|
403
403
|
};
|
|
@@ -1,5 +1,5 @@
|
|
|
1
1
|
/**
|
|
2
|
-
* OpenAI SDK monkey-patch
|
|
2
|
+
* OpenAI SDK monkey-patch: auto-emit llm_call events for every
|
|
3
3
|
* chat.completions.create() and responses.create() call inside a
|
|
4
4
|
* wrap()'d execution.
|
|
5
5
|
*
|
|
@@ -12,7 +12,7 @@
|
|
|
12
12
|
* - Response text is `choices[0].message.content` for chat
|
|
13
13
|
* completions, OR `output_text` for the Responses API.
|
|
14
14
|
* - Token names: `prompt_tokens` / `completion_tokens` for chat,
|
|
15
|
-
* `input_tokens` / `output_tokens` for Responses
|
|
15
|
+
* `input_tokens` / `output_tokens` for Responses: both
|
|
16
16
|
* normalized to the canonical input_tokens/output_tokens on
|
|
17
17
|
* payload.
|
|
18
18
|
*
|
|
@@ -1,5 +1,5 @@
|
|
|
1
1
|
/**
|
|
2
|
-
* OpenAI SDK monkey-patch
|
|
2
|
+
* OpenAI SDK monkey-patch: auto-emit llm_call events for every
|
|
3
3
|
* chat.completions.create() and responses.create() call inside a
|
|
4
4
|
* wrap()'d execution.
|
|
5
5
|
*
|
|
@@ -12,7 +12,7 @@
|
|
|
12
12
|
* - Response text is `choices[0].message.content` for chat
|
|
13
13
|
* completions, OR `output_text` for the Responses API.
|
|
14
14
|
* - Token names: `prompt_tokens` / `completion_tokens` for chat,
|
|
15
|
-
* `input_tokens` / `output_tokens` for Responses
|
|
15
|
+
* `input_tokens` / `output_tokens` for Responses: both
|
|
16
16
|
* normalized to the canonical input_tokens/output_tokens on
|
|
17
17
|
* payload.
|
|
18
18
|
*
|
package/dist/tool.js
CHANGED
|
@@ -26,7 +26,7 @@ const MAX_EXC_MSG = 500;
|
|
|
26
26
|
* The SDK uses this as a generous upper bound; the backend applies
|
|
27
27
|
* its OWN per-project cap (default 8 KB) at fingerprint time so
|
|
28
28
|
* customers tune storage / fingerprint policy from the dashboard
|
|
29
|
-
* without an SDK redeploy (see
|
|
29
|
+
* without an SDK redeploy (see migration 042).
|
|
30
30
|
*
|
|
31
31
|
* Raised from the v0.5.0 hardcoded 2048 to 16384 (16 KB) so the
|
|
32
32
|
* SDK no longer aggressively truncates returns that the backend
|
package/dist/truncate.d.ts
CHANGED
|
@@ -24,13 +24,13 @@
|
|
|
24
24
|
*
|
|
25
25
|
* **Markers added on truncation:**
|
|
26
26
|
*
|
|
27
|
-
* - `_truncated` (boolean)
|
|
27
|
+
* - `_truncated` (boolean): always `true` when this
|
|
28
28
|
* helper altered the payload.
|
|
29
|
-
* - `_original_payload_bytes` (number)
|
|
29
|
+
* - `_original_payload_bytes` (number): serialized byte count of
|
|
30
30
|
* the original payload, so the
|
|
31
31
|
* customer can see how much got
|
|
32
32
|
* dropped.
|
|
33
|
-
* - `_truncated_fields` (string[])
|
|
33
|
+
* - `_truncated_fields` (string[]): names of the top-level
|
|
34
34
|
* fields whose values got shortened.
|
|
35
35
|
*
|
|
36
36
|
* **Fail-open.** Any exception during truncation falls back to the
|
package/dist/truncate.js
CHANGED
|
@@ -24,13 +24,13 @@
|
|
|
24
24
|
*
|
|
25
25
|
* **Markers added on truncation:**
|
|
26
26
|
*
|
|
27
|
-
* - `_truncated` (boolean)
|
|
27
|
+
* - `_truncated` (boolean): always `true` when this
|
|
28
28
|
* helper altered the payload.
|
|
29
|
-
* - `_original_payload_bytes` (number)
|
|
29
|
+
* - `_original_payload_bytes` (number): serialized byte count of
|
|
30
30
|
* the original payload, so the
|
|
31
31
|
* customer can see how much got
|
|
32
32
|
* dropped.
|
|
33
|
-
* - `_truncated_fields` (string[])
|
|
33
|
+
* - `_truncated_fields` (string[]): names of the top-level
|
|
34
34
|
* fields whose values got shortened.
|
|
35
35
|
*
|
|
36
36
|
* **Fail-open.** Any exception during truncation falls back to the
|
|
@@ -1,12 +1,12 @@
|
|
|
1
1
|
/**
|
|
2
|
-
* Vertex AI Gemini TypeScript SDK monkey-patch
|
|
2
|
+
* Vertex AI Gemini TypeScript SDK monkey-patch: auto-emit llm_call
|
|
3
3
|
* events for every GenerativeModel.generateContent() call inside a
|
|
4
4
|
* wrap()'d execution against the Vertex AI surface
|
|
5
5
|
* (@google-cloud/vertexai package).
|
|
6
6
|
*
|
|
7
7
|
* Same provider tag ("gemini") and same canonical error_class map
|
|
8
8
|
* as instrumentGemini. Customers running both surfaces call both
|
|
9
|
-
* instrumentGemini() and instrumentVertexGemini()
|
|
9
|
+
* instrumentGemini() and instrumentVertexGemini(): each patches a
|
|
10
10
|
* different package's class. Provider_incident detection sees one
|
|
11
11
|
* signal regardless of which surface was hit.
|
|
12
12
|
*
|
|
@@ -1,12 +1,12 @@
|
|
|
1
1
|
/**
|
|
2
|
-
* Vertex AI Gemini TypeScript SDK monkey-patch
|
|
2
|
+
* Vertex AI Gemini TypeScript SDK monkey-patch: auto-emit llm_call
|
|
3
3
|
* events for every GenerativeModel.generateContent() call inside a
|
|
4
4
|
* wrap()'d execution against the Vertex AI surface
|
|
5
5
|
* (@google-cloud/vertexai package).
|
|
6
6
|
*
|
|
7
7
|
* Same provider tag ("gemini") and same canonical error_class map
|
|
8
8
|
* as instrumentGemini. Customers running both surfaces call both
|
|
9
|
-
* instrumentGemini() and instrumentVertexGemini()
|
|
9
|
+
* instrumentGemini() and instrumentVertexGemini(): each patches a
|
|
10
10
|
* different package's class. Provider_incident detection sees one
|
|
11
11
|
* signal regardless of which surface was hit.
|
|
12
12
|
*
|
package/dist/wrap.d.ts
CHANGED
|
@@ -91,7 +91,7 @@ export interface WrapOptions {
|
|
|
91
91
|
*
|
|
92
92
|
* Explicit `fromAgent` always wins. If neither is supplied while
|
|
93
93
|
* inside `wrap()`, `emitAgentHandoff` throws an Error rather than
|
|
94
|
-
* silently emitting an `unknown` source agent
|
|
94
|
+
* silently emitting an `unknown` source agent: the topology graph
|
|
95
95
|
* powers cascading_failure / coordination_deadlock clustering and
|
|
96
96
|
* must not be polluted.
|
|
97
97
|
*/
|
package/package.json
CHANGED