metergraph 0.4.0 → 0.6.0
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 +83 -16
- package/dist/batch-first.d.ts +112 -0
- package/dist/batch-first.js +213 -0
- package/dist/capture.js +29 -27
- package/dist/config.js +9 -4
- package/dist/context.js +18 -10
- package/dist/failure-log.js +5 -1
- package/dist/index.d.ts +18 -5
- package/dist/index.js +86 -38
- package/dist/index.mjs +43 -0
- package/dist/package.json +1 -0
- package/dist/provider-batch.d.ts +160 -0
- package/dist/provider-batch.js +247 -0
- package/dist/repo-config.d.ts +1 -1
- package/dist/repo-config.js +43 -36
- package/dist/session.js +16 -12
- package/dist/template.js +6 -2
- package/dist/track.js +6 -3
- package/dist/transport.js +9 -5
- package/dist/vercel-ai.js +9 -6
- package/dist/version.d.ts +1 -1
- package/dist/version.js +4 -1
- package/dist/wrap.js +24 -18
- package/package.json +17 -8
package/README.md
CHANGED
|
@@ -5,14 +5,18 @@ provider-independent Vercel AI SDK language models on Node 18+, including AWS
|
|
|
5
5
|
Lambda. Worker-style `waitUntil` hooks are present but are not part of the v1
|
|
6
6
|
public-package qualification contract.
|
|
7
7
|
|
|
8
|
-
|
|
9
|
-
|
|
10
|
-
|
|
8
|
+
Initialize Metergraph once, then wrap each provider client. `init()` reads the
|
|
9
|
+
token and other omitted options from the environment.
|
|
10
|
+
|
|
11
|
+
Initialization is process-wide. The first `init()` configuration remains
|
|
12
|
+
active; later explicit calls are ignored and produce one generic warning
|
|
13
|
+
without option names, token values, or other secrets.
|
|
11
14
|
|
|
12
15
|
```ts
|
|
13
|
-
import { wrap, route, trace, modelFor, recordOutcome, setSession } from "metergraph";
|
|
16
|
+
import { init, wrap, route, trace, modelFor, recordOutcome, setSession } from "metergraph";
|
|
14
17
|
import OpenAI from "openai";
|
|
15
18
|
|
|
19
|
+
init({ repository: "owner/repository" });
|
|
16
20
|
// new Anthropic() and new GoogleGenAI({}) wrap the same way.
|
|
17
21
|
const client = wrap(new OpenAI());
|
|
18
22
|
setSession("ticket-123");
|
|
@@ -47,7 +51,7 @@ import * as mg from "metergraph";
|
|
|
47
51
|
|
|
48
52
|
const model = wrapLanguageModel({
|
|
49
53
|
model: openai("gpt-5.6-luna"),
|
|
50
|
-
middleware: mg.vercelAISDKMiddleware(),
|
|
54
|
+
middleware: mg.vercelAISDKMiddleware({ repository: "owner/repository" }),
|
|
51
55
|
});
|
|
52
56
|
|
|
53
57
|
const result = await mg.trace("support-answer", () =>
|
|
@@ -65,6 +69,19 @@ for server-side pricing. Provider options, abort signals, and transport headers
|
|
|
65
69
|
are never serialized. Calls outside `trace()` remain valid one-span traces;
|
|
66
70
|
wrap a multi-call workflow in `trace()` to group its spans.
|
|
67
71
|
|
|
72
|
+
`vercelAISDKMiddleware()` initializes Metergraph itself, so the standard
|
|
73
|
+
integration needs only the middleware call. Pass any `init()` options there,
|
|
74
|
+
as shown above, or omit them when configuration comes from environment
|
|
75
|
+
variables. Applications that initialize Metergraph centrally may continue to
|
|
76
|
+
call `init()` first and then create the middleware without initialization
|
|
77
|
+
options.
|
|
78
|
+
|
|
79
|
+
Applications with several providers should use the [examples
|
|
80
|
+
chooser](../examples/README.md): apply MeterGraph once through the AI SDK
|
|
81
|
+
provider registry, or wrap the controlled exit of a custom factory the
|
|
82
|
+
application already has. The separate examples mark original application code
|
|
83
|
+
and every MeterGraph addition.
|
|
84
|
+
|
|
68
85
|
| Vercel AI SDK | Metergraph middleware | Node.js |
|
|
69
86
|
|---|---|---|
|
|
70
87
|
| 5 | `vercelAISDKMiddleware({ aiSdkVersion: 5 })` | 18+ |
|
|
@@ -86,12 +103,22 @@ required and throws if missing so a caller error surfaces immediately
|
|
|
86
103
|
instead of an `undefined` model reaching a provider call.
|
|
87
104
|
|
|
88
105
|
Set `METERGRAPH_APP_TOKEN`; `METERGRAPH_INGEST_URL` is only needed to override
|
|
89
|
-
the hosted HTTPS endpoint.
|
|
90
|
-
|
|
91
|
-
|
|
92
|
-
|
|
93
|
-
|
|
94
|
-
|
|
106
|
+
the hosted HTTPS endpoint. Repository identity enables repository-level
|
|
107
|
+
attribution and [MeterGraph Bot](https://github.com/apps/metergraph). Choose any
|
|
108
|
+
one of these options; each is sufficient on its own:
|
|
109
|
+
|
|
110
|
+
- Pass `{ repository: "owner/repository" }` to `mg.init()`.
|
|
111
|
+
- Set `METERGRAPH_REPOSITORY=owner/repository` in the environment.
|
|
112
|
+
- Store the identity with the source code in `.metergraph/config.json`:
|
|
113
|
+
|
|
114
|
+
```json
|
|
115
|
+
{"repository":"owner/repository"}
|
|
116
|
+
```
|
|
117
|
+
|
|
118
|
+
Resolution order is the explicit option, environment variable, then
|
|
119
|
+
configuration file. The SDK treats the file as read-only. Without repository
|
|
120
|
+
identity, it warns once and continues capture without repository attribution.
|
|
121
|
+
Metergraph captures the scrubbed provider request and a
|
|
95
122
|
normalized response envelope, including assistant content and tool calls, by
|
|
96
123
|
default. Provider credentials and transport headers are removed. Request and
|
|
97
124
|
response are each limited to 100 KiB of UTF-8 with an explicit truncation
|
|
@@ -125,6 +152,43 @@ inference while iterating a wrapped `client.messages.batches.results()` result.
|
|
|
125
152
|
Consume results inside `route()` so the asynchronous batch retains its product
|
|
126
153
|
route. Job-management polls are deliberately not counted as model calls.
|
|
127
154
|
|
|
155
|
+
## Batch-first execution (opt-in)
|
|
156
|
+
|
|
157
|
+
`batchFirst()` is a separate, explicitly opt-in code path from `wrap()`/capture: submit one request through a provider's Batch API, wait up to a caller-chosen deadline, and fall back to exactly one direct call if the batch hasn't finished in time.
|
|
158
|
+
|
|
159
|
+
```ts
|
|
160
|
+
import { batchFirst } from "metergraph";
|
|
161
|
+
import OpenAI from "openai";
|
|
162
|
+
|
|
163
|
+
const client = new OpenAI(); // unwrapped — batchFirst() drives it directly, not through wrap()
|
|
164
|
+
|
|
165
|
+
const outcome = await batchFirst(
|
|
166
|
+
client,
|
|
167
|
+
"openai",
|
|
168
|
+
{ model: "gpt-5-mini", input: "Summarize this document." },
|
|
169
|
+
{
|
|
170
|
+
deadlineMs: 60_000,
|
|
171
|
+
acceptDuplicateProviderExecution: true, // required: a missed deadline can execute the request twice
|
|
172
|
+
onLateBatchSettled: (info) => {
|
|
173
|
+
// Fires later, asynchronously, only if a losing batch eventually
|
|
174
|
+
// settles. Never blocks batchFirst()'s own return, and never exposes
|
|
175
|
+
// the late result's content — only whether it contained a
|
|
176
|
+
// tool-call plan.
|
|
177
|
+
},
|
|
178
|
+
},
|
|
179
|
+
);
|
|
180
|
+
|
|
181
|
+
outcome.source; // "batch" | "direct"
|
|
182
|
+
outcome.result; // the provider response
|
|
183
|
+
outcome.metadata.batch_outcome; // "completed" | "failed" | "expired" | "pending_at_deadline"
|
|
184
|
+
```
|
|
185
|
+
|
|
186
|
+
`provider` is `"openai" | "anthropic" | "google"`, matching `wrap()`'s own explicit-provider option — never inferred from the client instance. A request with `stream: true` is rejected before any provider call. A request carrying `tools` is rejected unless `allowDuplicateToolCallPlans: true` is also set, acknowledging that the batch result and the direct fallback are independent provider executions that may each choose a different tool-call plan. `acceptDuplicateProviderExecution` must be exactly `true` — there is no default and no environment-variable override, and a missed deadline can execute (and bill) the same request twice. Neither the batch nor the direct path executes a tool call automatically; the caller receives the tool-call plan in `outcome.result` and is responsible for executing it, exactly as with a normal (non-batch-first) provider response.
|
|
187
|
+
|
|
188
|
+
`batchFirst()` is not integrated with `wrap()`'s capture/telemetry pipeline; its result and metadata are returned directly to the caller, never enqueued for delivery. The background poll that watches a losing batch for late telemetry uses an ordinary `setTimeout`, not an unref'd one — after a deadline-triggered direct fallback it keeps the Node process alive until the batch reaches a terminal state, which can take up to 24 hours; this is currently unmitigated, so a short-lived script or CLI process should not assume it exits promptly after a fallback.
|
|
189
|
+
|
|
190
|
+
OpenAI, Anthropic, and Google Gemini all have adapters (`createOpenAIBatchAdapter`, `createAnthropicBatchAdapter`, `createGoogleBatchAdapter`), built against each provider's documented Batch API shape and covered by fake-client contract tests, but **none has been exercised against a live provider Batch API from this SDK** — treat this as a beta-quality, code-reviewed-but-not-live-verified surface.
|
|
191
|
+
|
|
128
192
|
## Set up with an AI coding agent
|
|
129
193
|
|
|
130
194
|
Paste this into Claude Code, Codex, Cursor, or any coding agent inside the
|
|
@@ -136,11 +200,14 @@ Instrument this codebase's LLM API costs with the `metergraph` npm package
|
|
|
136
200
|
then wrap every new OpenAI(), new Anthropic(), and new GoogleGenAI()
|
|
137
201
|
construction in place, e.g. const client = mg.wrap(new OpenAI()) after
|
|
138
202
|
import * as mg from "metergraph". wrap() returns the same client and
|
|
139
|
-
initializes itself from the environment
|
|
140
|
-
(
|
|
141
|
-
|
|
142
|
-
|
|
143
|
-
|
|
203
|
+
initializes itself from the environment. Before wrapping, call
|
|
204
|
+
mg.init({ repository: "owner/repository" }) using the actual GitHub owner and
|
|
205
|
+
repository name. METERGRAPH_APP_TOKEN is required; the SDK warns and disables
|
|
206
|
+
capture when it is missing. METERGRAPH_INGEST_URL is only for self-hosted
|
|
207
|
+
servers. Document variable names with placeholders in .env.example, and put
|
|
208
|
+
real values only in deployment configuration. Metergraph captures scrubbed
|
|
209
|
+
provider requests and normalized responses by default for the hosted dashboard;
|
|
210
|
+
use METERGRAPH_CAPTURE_TEXT=0 or
|
|
144
211
|
captureText: false around sensitive operations. Provider credentials and
|
|
145
212
|
transport headers must never be captured. Capture is fail-open, so do not
|
|
146
213
|
change call sites, arguments, or error handling; async and streaming work
|
|
@@ -0,0 +1,112 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* Explicit, opt-in batch-first execution: submit one request through a
|
|
3
|
+
* provider's Batch API, wait a caller-selected deadline, and fall back to
|
|
4
|
+
* a single direct call if the batch hasn't finished in time. Never enabled
|
|
5
|
+
* by wrap()/capture defaults or by an environment variable — a caller
|
|
6
|
+
* reaches this only by importing and calling batchFirst() directly, and
|
|
7
|
+
* only after explicitly acknowledging the duplicate-execution semantic
|
|
8
|
+
* below.
|
|
9
|
+
*
|
|
10
|
+
* On a missed deadline, the request may execute twice against the
|
|
11
|
+
* provider (once via batch, once via direct) — an accepted, deliberate
|
|
12
|
+
* semantic, never silently avoided. What this module guarantees instead:
|
|
13
|
+
* exactly one direct fallback is ever issued, and a batch result that
|
|
14
|
+
* arrives after the fallback already won is never returned, never
|
|
15
|
+
* executed, and never mutates an already-resolved result.
|
|
16
|
+
*/
|
|
17
|
+
import { type BatchFirstRequest, type ProviderBatchAdapter } from "./provider-batch.js";
|
|
18
|
+
export type { BatchFirstRequest } from "./provider-batch.js";
|
|
19
|
+
/** An injectable time source. Real timers by default; tests can supply a
|
|
20
|
+
* fake to drive the deadline deterministically, with no real waiting. */
|
|
21
|
+
export interface BatchFirstClock {
|
|
22
|
+
setTimeout(handler: () => void, ms: number): unknown;
|
|
23
|
+
clearTimeout(handle: unknown): void;
|
|
24
|
+
}
|
|
25
|
+
/** Reported asynchronously, after batchFirst() has already returned via the
|
|
26
|
+
* direct fallback, if the losing batch eventually reaches a terminal
|
|
27
|
+
* state. Its actual result content is never included here and never
|
|
28
|
+
* returned from batchFirst() — only whether it happened to contain a
|
|
29
|
+
* tool-call plan, which may differ from the one the direct fallback
|
|
30
|
+
* produced (see allowDuplicateToolCallPlans on BatchFirstPolicy). */
|
|
31
|
+
export interface LateBatchInfo {
|
|
32
|
+
outcome: "completed" | "failed" | "expired";
|
|
33
|
+
containedToolCallPlan: boolean;
|
|
34
|
+
}
|
|
35
|
+
export interface BatchFirstPolicy {
|
|
36
|
+
/** How long to wait for the batch before issuing the one direct
|
|
37
|
+
* fallback. Required — there is no default deadline. */
|
|
38
|
+
deadlineMs: number;
|
|
39
|
+
/**
|
|
40
|
+
* Required, explicit acknowledgement that a missed deadline can result
|
|
41
|
+
* in the same request executing twice against the provider. Must be
|
|
42
|
+
* exactly `true`; never read from an environment variable, and never
|
|
43
|
+
* defaulted.
|
|
44
|
+
*/
|
|
45
|
+
acceptDuplicateProviderExecution: true;
|
|
46
|
+
/**
|
|
47
|
+
* Required, explicit acknowledgement for a request that includes
|
|
48
|
+
* `tools`: the batch result and the direct fallback are two independent
|
|
49
|
+
* provider executions of the same prompt, and may each choose a
|
|
50
|
+
* DIFFERENT tool-call plan (different tool, different arguments). A
|
|
51
|
+
* caller whose tools have side effects must not assume the two plans
|
|
52
|
+
* agree. Without this flag, a request carrying `tools` is rejected
|
|
53
|
+
* before any provider call is made.
|
|
54
|
+
*/
|
|
55
|
+
allowDuplicateToolCallPlans?: boolean;
|
|
56
|
+
/** Fired once, later, only when the losing batch eventually settles.
|
|
57
|
+
* Never blocks batchFirst()'s own return. */
|
|
58
|
+
onLateBatchSettled?: (info: LateBatchInfo) => void;
|
|
59
|
+
}
|
|
60
|
+
/** Non-public knobs for driving runBatchFirst()'s poll interval and time
|
|
61
|
+
* source from tests — deliberately kept off BatchFirstPolicy so a real
|
|
62
|
+
* caller of batchFirst() has no way to set them. */
|
|
63
|
+
interface RunBatchFirstTestControls {
|
|
64
|
+
/** How often to re-check batch status while waiting. Default 2000ms. */
|
|
65
|
+
pollIntervalMs?: number;
|
|
66
|
+
/** Injectable time source — see BatchFirstClock. */
|
|
67
|
+
clock?: BatchFirstClock;
|
|
68
|
+
}
|
|
69
|
+
export type BatchFirstSource = "batch" | "direct";
|
|
70
|
+
export interface BatchFirstMetadata {
|
|
71
|
+
execution_mode: "batch_first";
|
|
72
|
+
deadline_ms: number;
|
|
73
|
+
/** Wall-clock time from submission to the canonical result settling. */
|
|
74
|
+
batch_wait_ms: number;
|
|
75
|
+
/** The batch's own status as of when batchFirst() returned — not its
|
|
76
|
+
* eventual status if that differs (see onLateBatchSettled). */
|
|
77
|
+
batch_outcome: "completed" | "failed" | "expired" | "pending_at_deadline";
|
|
78
|
+
canonical_result: BatchFirstSource;
|
|
79
|
+
duplicate_provider_execution: boolean;
|
|
80
|
+
/** Always false when canonical_result is "batch" (no lateness is
|
|
81
|
+
* possible — the batch IS the canonical result). When
|
|
82
|
+
* canonical_result is "direct", this reflects what was known AT THE
|
|
83
|
+
* MOMENT batchFirst() returned, which is always false: a completion
|
|
84
|
+
* confirmed later only reaches the caller through
|
|
85
|
+
* onLateBatchSettled, never by mutating this object. */
|
|
86
|
+
late_batch_completed: boolean;
|
|
87
|
+
late_batch_contained_tool_call_plan: boolean;
|
|
88
|
+
}
|
|
89
|
+
export interface BatchFirstResult<TResult = unknown> {
|
|
90
|
+
source: BatchFirstSource;
|
|
91
|
+
result: TResult;
|
|
92
|
+
metadata: BatchFirstMetadata;
|
|
93
|
+
}
|
|
94
|
+
/** Thrown before any provider call when a request/policy combination is
|
|
95
|
+
* not eligible for batch-first execution — streaming, tools without
|
|
96
|
+
* acknowledgement, a missing/false acceptDuplicateProviderExecution, or
|
|
97
|
+
* an adapter-specific ineligibility. */
|
|
98
|
+
export declare class BatchFirstIneligibleError extends Error {
|
|
99
|
+
}
|
|
100
|
+
/**
|
|
101
|
+
* The adapter-injected core state machine — exported so fake-adapter,
|
|
102
|
+
* fake-clock tests can drive it directly, and used internally by the
|
|
103
|
+
* public, provider-explicit batchFirst() below.
|
|
104
|
+
*/
|
|
105
|
+
export declare function runBatchFirst<TResult>(adapter: ProviderBatchAdapter<TResult>, request: BatchFirstRequest, policy: BatchFirstPolicy, testControls?: RunBatchFirstTestControls): Promise<BatchFirstResult<TResult>>;
|
|
106
|
+
export type BatchFirstProvider = "openai" | "anthropic" | "google";
|
|
107
|
+
/**
|
|
108
|
+
* Explicit, provider-specific batch-first execution. Never inferred from
|
|
109
|
+
* the client instance — the caller states the provider, matching wrap()'s
|
|
110
|
+
* own explicit-provider option.
|
|
111
|
+
*/
|
|
112
|
+
export declare function batchFirst<TResult = unknown>(client: unknown, provider: BatchFirstProvider, request: BatchFirstRequest, policy: BatchFirstPolicy): Promise<BatchFirstResult<TResult>>;
|
|
@@ -0,0 +1,213 @@
|
|
|
1
|
+
"use strict";
|
|
2
|
+
Object.defineProperty(exports, "__esModule", { value: true });
|
|
3
|
+
exports.BatchFirstIneligibleError = void 0;
|
|
4
|
+
exports.runBatchFirst = runBatchFirst;
|
|
5
|
+
exports.batchFirst = batchFirst;
|
|
6
|
+
/**
|
|
7
|
+
* Explicit, opt-in batch-first execution: submit one request through a
|
|
8
|
+
* provider's Batch API, wait a caller-selected deadline, and fall back to
|
|
9
|
+
* a single direct call if the batch hasn't finished in time. Never enabled
|
|
10
|
+
* by wrap()/capture defaults or by an environment variable — a caller
|
|
11
|
+
* reaches this only by importing and calling batchFirst() directly, and
|
|
12
|
+
* only after explicitly acknowledging the duplicate-execution semantic
|
|
13
|
+
* below.
|
|
14
|
+
*
|
|
15
|
+
* On a missed deadline, the request may execute twice against the
|
|
16
|
+
* provider (once via batch, once via direct) — an accepted, deliberate
|
|
17
|
+
* semantic, never silently avoided. What this module guarantees instead:
|
|
18
|
+
* exactly one direct fallback is ever issued, and a batch result that
|
|
19
|
+
* arrives after the fallback already won is never returned, never
|
|
20
|
+
* executed, and never mutates an already-resolved result.
|
|
21
|
+
*/
|
|
22
|
+
const provider_batch_js_1 = require("./provider-batch.js");
|
|
23
|
+
function realClock() {
|
|
24
|
+
return {
|
|
25
|
+
setTimeout: (handler, ms) => setTimeout(handler, ms),
|
|
26
|
+
clearTimeout: (handle) => clearTimeout(handle),
|
|
27
|
+
};
|
|
28
|
+
}
|
|
29
|
+
/** Thrown before any provider call when a request/policy combination is
|
|
30
|
+
* not eligible for batch-first execution — streaming, tools without
|
|
31
|
+
* acknowledgement, a missing/false acceptDuplicateProviderExecution, or
|
|
32
|
+
* an adapter-specific ineligibility. */
|
|
33
|
+
class BatchFirstIneligibleError extends Error {
|
|
34
|
+
}
|
|
35
|
+
exports.BatchFirstIneligibleError = BatchFirstIneligibleError;
|
|
36
|
+
function hasTools(request) {
|
|
37
|
+
return Array.isArray(request.tools) && request.tools.length > 0;
|
|
38
|
+
}
|
|
39
|
+
function validate(request, policy) {
|
|
40
|
+
if (policy.acceptDuplicateProviderExecution !== true) {
|
|
41
|
+
throw new BatchFirstIneligibleError("batchFirst() requires policy.acceptDuplicateProviderExecution: true — "
|
|
42
|
+
+ "a missed deadline can execute the request twice against the provider");
|
|
43
|
+
}
|
|
44
|
+
if (!Number.isFinite(policy.deadlineMs) || policy.deadlineMs <= 0) {
|
|
45
|
+
throw new BatchFirstIneligibleError("batchFirst() requires a positive policy.deadlineMs");
|
|
46
|
+
}
|
|
47
|
+
if (request.stream === true) {
|
|
48
|
+
throw new BatchFirstIneligibleError("batchFirst() does not support streaming requests — streaming is direct-only");
|
|
49
|
+
}
|
|
50
|
+
if (hasTools(request) && policy.allowDuplicateToolCallPlans !== true) {
|
|
51
|
+
throw new BatchFirstIneligibleError("batchFirst() requires policy.allowDuplicateToolCallPlans: true for requests with "
|
|
52
|
+
+ "tools — the batch result and the direct fallback are independent provider "
|
|
53
|
+
+ "executions and may each choose a different tool call plan");
|
|
54
|
+
}
|
|
55
|
+
}
|
|
56
|
+
const DEFAULT_POLL_INTERVAL_MS = 2_000;
|
|
57
|
+
/**
|
|
58
|
+
* The adapter-injected core state machine — exported so fake-adapter,
|
|
59
|
+
* fake-clock tests can drive it directly, and used internally by the
|
|
60
|
+
* public, provider-explicit batchFirst() below.
|
|
61
|
+
*/
|
|
62
|
+
async function runBatchFirst(adapter, request, policy, testControls) {
|
|
63
|
+
validate(request, policy);
|
|
64
|
+
const eligibility = adapter.eligibility(request);
|
|
65
|
+
if (!eligibility.eligible) {
|
|
66
|
+
throw new BatchFirstIneligibleError(`request is not eligible for batch-first execution: ${eligibility.reason ?? "unsupported by this adapter"}`);
|
|
67
|
+
}
|
|
68
|
+
const clock = testControls?.clock ?? realClock();
|
|
69
|
+
const pollIntervalMs = testControls?.pollIntervalMs ?? DEFAULT_POLL_INTERVAL_MS;
|
|
70
|
+
const startedAt = Date.now();
|
|
71
|
+
const handle = await adapter.submitOne(request);
|
|
72
|
+
let settleDeadline;
|
|
73
|
+
const deadlineSettled = new Promise((resolve) => {
|
|
74
|
+
settleDeadline = resolve;
|
|
75
|
+
});
|
|
76
|
+
const deadlineTimer = clock.setTimeout(() => settleDeadline?.(), policy.deadlineMs);
|
|
77
|
+
// Resolves at most once, and only from inside the poll loop below, the
|
|
78
|
+
// moment a real terminal status is observed — never resolves at all if
|
|
79
|
+
// the deadline wins first. Racing this directly (rather than racing the
|
|
80
|
+
// loop's own completion promise) means there is no "loop stopped without
|
|
81
|
+
// a terminal status" case to fake a result for: that case simply never
|
|
82
|
+
// resolves this promise, which is exactly the semantics Promise.race
|
|
83
|
+
// needs.
|
|
84
|
+
let resolveBatchTerminal;
|
|
85
|
+
const batchTerminalSettled = new Promise((resolve) => {
|
|
86
|
+
resolveBatchTerminal = resolve;
|
|
87
|
+
});
|
|
88
|
+
let stopPolling = false;
|
|
89
|
+
const pollUntilTerminal = (async () => {
|
|
90
|
+
while (!stopPolling) {
|
|
91
|
+
const outcome = await adapter.poll(handle);
|
|
92
|
+
if (outcome.status !== "pending") {
|
|
93
|
+
resolveBatchTerminal?.(outcome.status);
|
|
94
|
+
return outcome.status;
|
|
95
|
+
}
|
|
96
|
+
if (stopPolling)
|
|
97
|
+
return undefined;
|
|
98
|
+
await new Promise((resolve) => clock.setTimeout(resolve, pollIntervalMs));
|
|
99
|
+
}
|
|
100
|
+
return undefined;
|
|
101
|
+
})();
|
|
102
|
+
const winner = await Promise.race([
|
|
103
|
+
batchTerminalSettled.then((status) => ({ kind: "batch", status })),
|
|
104
|
+
deadlineSettled.then(() => ({ kind: "deadline" })),
|
|
105
|
+
]);
|
|
106
|
+
clock.clearTimeout(deadlineTimer);
|
|
107
|
+
// A batch reported "completed" but whose result cannot be read (a
|
|
108
|
+
// missing output file, an item-level provider error, a malformed or
|
|
109
|
+
// missing matching line, a transient read failure) is neither a valid
|
|
110
|
+
// canonical batch result nor grounds to throw out of batchFirst()
|
|
111
|
+
// instead of the promised fallback — it is treated exactly like a batch
|
|
112
|
+
// that reported "failed": exactly one direct fallback, never a second
|
|
113
|
+
// read attempt, never surfaced as a rejection.
|
|
114
|
+
let unreadableCompletedBatch = false;
|
|
115
|
+
if (winner.kind === "batch" && winner.status === "completed") {
|
|
116
|
+
stopPolling = true;
|
|
117
|
+
try {
|
|
118
|
+
const { result } = await adapter.readResult(handle);
|
|
119
|
+
return {
|
|
120
|
+
source: "batch",
|
|
121
|
+
result,
|
|
122
|
+
metadata: {
|
|
123
|
+
execution_mode: "batch_first",
|
|
124
|
+
deadline_ms: policy.deadlineMs,
|
|
125
|
+
batch_wait_ms: Date.now() - startedAt,
|
|
126
|
+
batch_outcome: "completed",
|
|
127
|
+
canonical_result: "batch",
|
|
128
|
+
duplicate_provider_execution: false,
|
|
129
|
+
late_batch_completed: false,
|
|
130
|
+
late_batch_contained_tool_call_plan: false,
|
|
131
|
+
},
|
|
132
|
+
};
|
|
133
|
+
}
|
|
134
|
+
catch {
|
|
135
|
+
unreadableCompletedBatch = true;
|
|
136
|
+
}
|
|
137
|
+
}
|
|
138
|
+
// Either the deadline fired first, the batch reached a non-completed
|
|
139
|
+
// terminal status before the deadline, or the batch completed but its
|
|
140
|
+
// result could not be read — either way, issue exactly one direct
|
|
141
|
+
// fallback now, and never wait further (or retry a read) on the batch
|
|
142
|
+
// for the canonical result. Only stop the poll loop when the batch
|
|
143
|
+
// itself already produced a terminal status (it has nothing left to do,
|
|
144
|
+
// so this is a no-op) — when the DEADLINE won, deliberately leave
|
|
145
|
+
// polling running in the background: "keep polling only to write
|
|
146
|
+
// terminal telemetry" requires the loop to keep going, not stop here.
|
|
147
|
+
const batchOutcomeAtFallback = unreadableCompletedBatch
|
|
148
|
+
? "failed"
|
|
149
|
+
: winner.kind === "batch"
|
|
150
|
+
? winner.status
|
|
151
|
+
: "pending_at_deadline";
|
|
152
|
+
if (winner.kind === "batch")
|
|
153
|
+
stopPolling = true;
|
|
154
|
+
const directResultPromise = adapter.direct(request);
|
|
155
|
+
if (winner.kind === "deadline") {
|
|
156
|
+
// Observe the batch purely for telemetry — never read for its content
|
|
157
|
+
// to be returned, executed, or used to mutate the already-in-flight
|
|
158
|
+
// direct result.
|
|
159
|
+
pollUntilTerminal
|
|
160
|
+
.then(async (status) => {
|
|
161
|
+
if (status !== "completed" && status !== "failed" && status !== "expired")
|
|
162
|
+
return;
|
|
163
|
+
let containedToolCallPlan = false;
|
|
164
|
+
if (status === "completed") {
|
|
165
|
+
try {
|
|
166
|
+
const late = await adapter.readResult(handle);
|
|
167
|
+
containedToolCallPlan = late.containedToolCallPlan;
|
|
168
|
+
}
|
|
169
|
+
catch {
|
|
170
|
+
/* telemetry only — never thrown, never surfaced */
|
|
171
|
+
}
|
|
172
|
+
}
|
|
173
|
+
policy.onLateBatchSettled?.({ outcome: status, containedToolCallPlan });
|
|
174
|
+
})
|
|
175
|
+
.catch(() => { });
|
|
176
|
+
}
|
|
177
|
+
const { result } = await directResultPromise;
|
|
178
|
+
return {
|
|
179
|
+
source: "direct",
|
|
180
|
+
result,
|
|
181
|
+
metadata: {
|
|
182
|
+
execution_mode: "batch_first",
|
|
183
|
+
deadline_ms: policy.deadlineMs,
|
|
184
|
+
batch_wait_ms: Date.now() - startedAt,
|
|
185
|
+
batch_outcome: batchOutcomeAtFallback,
|
|
186
|
+
canonical_result: "direct",
|
|
187
|
+
duplicate_provider_execution: true,
|
|
188
|
+
late_batch_completed: false,
|
|
189
|
+
late_batch_contained_tool_call_plan: false,
|
|
190
|
+
},
|
|
191
|
+
};
|
|
192
|
+
}
|
|
193
|
+
function resolveAdapter(client, provider) {
|
|
194
|
+
if (provider === "openai") {
|
|
195
|
+
return (0, provider_batch_js_1.createOpenAIBatchAdapter)(client);
|
|
196
|
+
}
|
|
197
|
+
if (provider === "anthropic") {
|
|
198
|
+
return (0, provider_batch_js_1.createAnthropicBatchAdapter)(client);
|
|
199
|
+
}
|
|
200
|
+
if (provider === "google") {
|
|
201
|
+
return (0, provider_batch_js_1.createGoogleBatchAdapter)(client);
|
|
202
|
+
}
|
|
203
|
+
throw new BatchFirstIneligibleError(`batchFirst() has no adapter for provider "${provider}" yet`);
|
|
204
|
+
}
|
|
205
|
+
/**
|
|
206
|
+
* Explicit, provider-specific batch-first execution. Never inferred from
|
|
207
|
+
* the client instance — the caller states the provider, matching wrap()'s
|
|
208
|
+
* own explicit-provider option.
|
|
209
|
+
*/
|
|
210
|
+
async function batchFirst(client, provider, request, policy) {
|
|
211
|
+
const adapter = resolveAdapter(client, provider);
|
|
212
|
+
return runBatchFirst(adapter, request, policy);
|
|
213
|
+
}
|
package/dist/capture.js
CHANGED
|
@@ -1,7 +1,11 @@
|
|
|
1
|
-
|
|
2
|
-
|
|
3
|
-
|
|
4
|
-
|
|
1
|
+
"use strict";
|
|
2
|
+
Object.defineProperty(exports, "__esModule", { value: true });
|
|
3
|
+
exports.CaptureRuntime = void 0;
|
|
4
|
+
exports.chunkText = chunkText;
|
|
5
|
+
const node_crypto_1 = require("node:crypto");
|
|
6
|
+
const context_js_1 = require("./context.js");
|
|
7
|
+
const template_js_1 = require("./template.js");
|
|
8
|
+
const version_js_1 = require("./version.js");
|
|
5
9
|
function get(value, key) {
|
|
6
10
|
return value && typeof value === "object" ? value[key] : undefined;
|
|
7
11
|
}
|
|
@@ -94,7 +98,7 @@ function responseText(response) {
|
|
|
94
98
|
}
|
|
95
99
|
return undefined;
|
|
96
100
|
}
|
|
97
|
-
|
|
101
|
+
function chunkText(chunk) {
|
|
98
102
|
const delta = get(first(get(chunk, "choices")), "delta") ?? get(chunk, "delta");
|
|
99
103
|
if (typeof delta === "string" && get(chunk, "type") === "text-delta")
|
|
100
104
|
return delta;
|
|
@@ -126,7 +130,7 @@ function toolArgument(value) {
|
|
|
126
130
|
return value;
|
|
127
131
|
}
|
|
128
132
|
}
|
|
129
|
-
return scrub(value);
|
|
133
|
+
return (0, template_js_1.scrub)(value);
|
|
130
134
|
}
|
|
131
135
|
function toolEvents(request, response, streamChunks = []) {
|
|
132
136
|
const calls = new Map();
|
|
@@ -297,7 +301,7 @@ function toolEvents(request, response, streamChunks = []) {
|
|
|
297
301
|
arguments: input && typeof input === "object"
|
|
298
302
|
&& Object.keys(input).length === 0
|
|
299
303
|
? ""
|
|
300
|
-
: JSON.stringify(scrub(input ?? {})),
|
|
304
|
+
: JSON.stringify((0, template_js_1.scrub)(input ?? {})),
|
|
301
305
|
});
|
|
302
306
|
}
|
|
303
307
|
}
|
|
@@ -327,25 +331,25 @@ function responseContent(response, aggregate) {
|
|
|
327
331
|
return aggregate;
|
|
328
332
|
const direct = get(response, "output_text") ?? get(response, "text");
|
|
329
333
|
if (direct !== undefined)
|
|
330
|
-
return scrub(direct);
|
|
334
|
+
return (0, template_js_1.scrub)(direct);
|
|
331
335
|
const message = get(first(get(response, "choices")), "message");
|
|
332
336
|
const content = get(message, "content");
|
|
333
337
|
if (content !== undefined)
|
|
334
|
-
return scrub(content);
|
|
338
|
+
return (0, template_js_1.scrub)(content);
|
|
335
339
|
const parsed = get(message, "parsed");
|
|
336
340
|
if (parsed !== undefined)
|
|
337
|
-
return scrub(parsed);
|
|
341
|
+
return (0, template_js_1.scrub)(parsed);
|
|
338
342
|
const normalizedText = responseText(response);
|
|
339
343
|
if (normalizedText !== undefined)
|
|
340
344
|
return normalizedText;
|
|
341
345
|
const blocks = get(response, "content");
|
|
342
346
|
if (blocks !== undefined)
|
|
343
|
-
return scrub(blocks);
|
|
347
|
+
return (0, template_js_1.scrub)(blocks);
|
|
344
348
|
const output = get(response, "output");
|
|
345
349
|
if (output !== undefined)
|
|
346
|
-
return scrub(output);
|
|
350
|
+
return (0, template_js_1.scrub)(output);
|
|
347
351
|
const candidates = get(response, "candidates");
|
|
348
|
-
return candidates === undefined ? undefined : scrub(candidates);
|
|
352
|
+
return candidates === undefined ? undefined : (0, template_js_1.scrub)(candidates);
|
|
349
353
|
}
|
|
350
354
|
function responseEnvelope(response, aggregate, tools, error, status) {
|
|
351
355
|
const responseMetadata = get(response, "response");
|
|
@@ -369,12 +373,9 @@ function responseEnvelope(response, aggregate, tools, error, status) {
|
|
|
369
373
|
return Object.fromEntries(Object.entries(envelope).filter(([, value]) => value !== undefined));
|
|
370
374
|
}
|
|
371
375
|
const SDK_DIR = (() => {
|
|
372
|
-
|
|
373
|
-
|
|
374
|
-
|
|
375
|
-
catch {
|
|
376
|
-
return "metergraph-sdk-dir-unavailable";
|
|
377
|
-
}
|
|
376
|
+
const frame = new Error().stack?.split("\n")[1] ?? "";
|
|
377
|
+
const match = frame.match(/^\s*at\s+(?:.*?\s+\()?(.+?)[/\\][^/\\:]+:\d+:\d+\)?$/);
|
|
378
|
+
return match?.[1]?.replace(/^file:\/\//, "") ?? "metergraph-sdk-dir-unavailable";
|
|
378
379
|
})();
|
|
379
380
|
function frames(stack, appRoot, skip, repoRoot) {
|
|
380
381
|
if (!stack)
|
|
@@ -405,14 +406,14 @@ function frames(stack, appRoot, skip, repoRoot) {
|
|
|
405
406
|
}
|
|
406
407
|
return result;
|
|
407
408
|
}
|
|
408
|
-
|
|
409
|
+
class CaptureRuntime {
|
|
409
410
|
transport;
|
|
410
411
|
options;
|
|
411
412
|
constructor(transport, options) {
|
|
412
413
|
this.transport = transport;
|
|
413
414
|
this.options = options;
|
|
414
415
|
}
|
|
415
|
-
start(provider, endpoint, request, stack, context = contextSnapshot()) {
|
|
416
|
+
start(provider, endpoint, request, stack, context = (0, context_js_1.contextSnapshot)()) {
|
|
416
417
|
return {
|
|
417
418
|
provider,
|
|
418
419
|
endpoint,
|
|
@@ -421,8 +422,8 @@ export class CaptureRuntime {
|
|
|
421
422
|
started: performance.now(),
|
|
422
423
|
ts: new Date().toISOString(),
|
|
423
424
|
frames: frames(stack, this.options.appRoot, this.options.skipFrames, this.options.repoRoot),
|
|
424
|
-
traceId: context.traceId ?? randomBytes(16).toString("hex"),
|
|
425
|
-
spanId: randomBytes(8).toString("hex"),
|
|
425
|
+
traceId: context.traceId ?? (0, node_crypto_1.randomBytes)(16).toString("hex"),
|
|
426
|
+
spanId: (0, node_crypto_1.randomBytes)(8).toString("hex"),
|
|
426
427
|
parentSpanId: context.parentSpanId,
|
|
427
428
|
traceName: context.traceName ?? context.route ?? context.funcName
|
|
428
429
|
?? endpoint,
|
|
@@ -456,8 +457,8 @@ export class CaptureRuntime {
|
|
|
456
457
|
}
|
|
457
458
|
return { value: `${clipped}${marker}`, truncated: true };
|
|
458
459
|
};
|
|
459
|
-
const request = text(JSON.stringify(scrub(state.request)), "request");
|
|
460
|
-
const fullTools = toolEvents(scrub(state.request), response, extra.responseChunks);
|
|
460
|
+
const request = text(JSON.stringify((0, template_js_1.scrub)(state.request)), "request");
|
|
461
|
+
const fullTools = toolEvents((0, template_js_1.scrub)(state.request), response, extra.responseChunks);
|
|
461
462
|
const status = extra.status ?? (extra.error ? "error" : stopReason(response) ?? "success");
|
|
462
463
|
const output = text(JSON.stringify(responseEnvelope(response, extra.responseText, fullTools, extra.error, status)), "response");
|
|
463
464
|
let toolTruncated = false;
|
|
@@ -503,7 +504,7 @@ export class CaptureRuntime {
|
|
|
503
504
|
span_id: state.spanId,
|
|
504
505
|
parent_span_id: state.parentSpanId,
|
|
505
506
|
trace_name: state.traceName,
|
|
506
|
-
template_hash: templateHash(state.request),
|
|
507
|
+
template_hash: (0, template_js_1.templateHash)(state.request),
|
|
507
508
|
unit_name: state.context.unitName,
|
|
508
509
|
unit_count: state.context.unitCount,
|
|
509
510
|
tool_calls: fullTools ? persistedTools : toolNames(state.request),
|
|
@@ -531,8 +532,9 @@ export class CaptureRuntime {
|
|
|
531
532
|
error: Boolean(extra.error),
|
|
532
533
|
error_type: extra.error instanceof Error ? extra.error.name : undefined,
|
|
533
534
|
sdk: "js",
|
|
534
|
-
sdk_version: SDK_VERSION,
|
|
535
|
+
sdk_version: version_js_1.SDK_VERSION,
|
|
535
536
|
runtime: typeof process === "undefined" ? "edge" : `node-${process.versions.node}`,
|
|
536
537
|
});
|
|
537
538
|
}
|
|
538
539
|
}
|
|
540
|
+
exports.CaptureRuntime = CaptureRuntime;
|
package/dist/config.js
CHANGED
|
@@ -1,4 +1,8 @@
|
|
|
1
|
-
|
|
1
|
+
"use strict";
|
|
2
|
+
Object.defineProperty(exports, "__esModule", { value: true });
|
|
3
|
+
exports.ConfigPoller = void 0;
|
|
4
|
+
exports.chooseModel = chooseModel;
|
|
5
|
+
const failure_log_js_1 = require("./failure-log.js");
|
|
2
6
|
function bucket(value) {
|
|
3
7
|
let hash = 0xcbf29ce484222325n;
|
|
4
8
|
for (const byte of new TextEncoder().encode(value)) {
|
|
@@ -7,7 +11,7 @@ function bucket(value) {
|
|
|
7
11
|
}
|
|
8
12
|
return Number(hash) / 2 ** 64 * 100;
|
|
9
13
|
}
|
|
10
|
-
|
|
14
|
+
function chooseModel(route, fallback, sessionKey, config) {
|
|
11
15
|
if (!config || config.enabled === false)
|
|
12
16
|
return fallback;
|
|
13
17
|
const incumbent = config.incumbent_model ?? fallback;
|
|
@@ -18,7 +22,7 @@ export function chooseModel(route, fallback, sessionKey, config) {
|
|
|
18
22
|
const seed = `${route}:${config.version ?? ""}:${config.salt ?? ""}:${sessionKey}`;
|
|
19
23
|
return bucket(seed) < percentage ? challenger : incumbent;
|
|
20
24
|
}
|
|
21
|
-
|
|
25
|
+
class ConfigPoller {
|
|
22
26
|
token;
|
|
23
27
|
baseUrl;
|
|
24
28
|
pollMs;
|
|
@@ -28,7 +32,7 @@ export class ConfigPoller {
|
|
|
28
32
|
lastSuccess = 0;
|
|
29
33
|
timer;
|
|
30
34
|
stopped = false;
|
|
31
|
-
failureLog = new FailureLogger();
|
|
35
|
+
failureLog = new failure_log_js_1.FailureLogger();
|
|
32
36
|
/** Resolves once the constructor's initial poll attempt has completed (success or failure). Exposed primarily so tests can synchronize deterministically instead of guessing a delay; not required for normal use. */
|
|
33
37
|
ready;
|
|
34
38
|
constructor(token, baseUrl, pollMs = 30_000, hardTtlMs = 120_000) {
|
|
@@ -86,3 +90,4 @@ export class ConfigPoller {
|
|
|
86
90
|
clearInterval(this.timer);
|
|
87
91
|
}
|
|
88
92
|
}
|
|
93
|
+
exports.ConfigPoller = ConfigPoller;
|