@semiont/observability 0.5.30 → 0.5.32

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/dist/index.d.ts CHANGED
@@ -167,15 +167,8 @@ declare function recordJobOutcome(jobType: string, outcome: 'completed' | 'faile
167
167
  */
168
168
  declare function recordAppendStage(stage: 'persist' | 'materialize' | 'enrich' | 'publish', durationMs: number): void;
169
169
  /**
170
- * Record a synchronous git invocation (ARCHIVIST-STAYS-UP P7).
171
- *
172
- * These are `execFileSync`, so **the duration is event-loop blockage, not just
173
- * latency** — every concurrent `browse:*` read waits behind it. One `git add`
174
- * runs per appended event, so a detection job writing hundreds of annotations
175
- * spawns hundreds of blocking subprocesses. That is the suspected mechanism
176
- * behind "reads serializing behind the detection job's annotation writes" in
177
- * `bugs/absent-archivist-wedges-browse.md`, which recorded the symptom without
178
- * a cause. This number is what turns that from a hypothesis into a reading.
170
+ * Record a git invocation. Read the `add` count against events appended: one
171
+ * per event means deferred staging has stopped deduping.
179
172
  */
180
173
  declare function recordGitCommand(command: string, durationMs: number): void;
181
174
  /**
@@ -216,6 +209,28 @@ declare function registerJobQueueProvider(provider: () => Promise<JobQueueSnapsh
216
209
  * OOM whose cause is inferred from RSS after the fact.
217
210
  */
218
211
  declare function registerFactPumpDepthProvider(provider: () => number): void;
212
+ /**
213
+ * A staging command that could not be run. Staging the index is a CONVENIENCE
214
+ * — the event log is the system of record — so a failure here is degraded
215
+ * service, never a reason to exit. But degraded must be VISIBLE: this counter
216
+ * is what stops "the index is quietly stale" from being invisible.
217
+ */
218
+ declare function recordGitStagingFailure(reason: 'index-lock' | 'other'): void;
219
+ /** Register `semiont.process.start_time`. Called by `initObservability*`. */
220
+ declare function registerProcessLifetimeMetrics(): void;
221
+ /**
222
+ * Supply a restart count. The Archivist's supervisor is POSIX shell and cannot
223
+ * emit OTel, but it already keeps a durable event log on the state mount — so
224
+ * the supervised child reads it and reports the count on the supervisor's
225
+ * behalf.
226
+ */
227
+ declare function registerRestartCountProvider(provider: () => Promise<number> | number): void;
228
+ /**
229
+ * Record that this process is dying abnormally, and mark the active span so a
230
+ * trace shows a span that ENDED IN DEATH rather than one that simply never
231
+ * ends. Never swallows: callers re-raise, so Node's own semantics are intact.
232
+ */
233
+ declare function recordAbnormalTermination(reason: string, detail?: string): void;
219
234
  declare function registerVectorIndexSizeProvider(provider: () => Promise<number> | number): void;
220
235
  /**
221
236
  * Record an inference call. Token counts are optional — providers that
@@ -277,5 +292,5 @@ declare function recordDetectionCall(opts: {
277
292
  */
278
293
  declare function recordAnchorOutcome(label: string, method: string): void;
279
294
 
280
- export { extractTraceparent, getActiveTraceparent, getLogTraceContext, injectTraceparent, recordAnchorOutcome, recordAppendStage, recordBusEmit, recordDetectionCall, recordGatherDegrade, recordGitCommand, recordHandlerDuration, recordInferenceUsage, recordJobOutcome, recordReplySuppressed, recordResumeGap, recordSubscriberConnect, recordSubscriberDisconnect, recordUnanswerableRequest, registerCorrelationRegistryProvider, registerFactPumpDepthProvider, registerJobQueueProvider, registerVectorIndexSizeProvider, withActorSpan, withSpan, withTraceparent };
295
+ export { extractTraceparent, getActiveTraceparent, getLogTraceContext, injectTraceparent, recordAbnormalTermination, recordAnchorOutcome, recordAppendStage, recordBusEmit, recordDetectionCall, recordGatherDegrade, recordGitCommand, recordGitStagingFailure, recordHandlerDuration, recordInferenceUsage, recordJobOutcome, recordReplySuppressed, recordResumeGap, recordSubscriberConnect, recordSubscriberDisconnect, recordUnanswerableRequest, registerCorrelationRegistryProvider, registerFactPumpDepthProvider, registerJobQueueProvider, registerProcessLifetimeMetrics, registerRestartCountProvider, registerVectorIndexSizeProvider, withActorSpan, withSpan, withTraceparent };
281
296
  export type { CorrelationRegistrySnapshot, JobQueueSnapshot, TraceCarrier };
package/dist/index.js CHANGED
@@ -251,7 +251,7 @@ var _gitCommandHistogram;
251
251
  function gitCommandHistogram() {
252
252
  if (!_gitCommandHistogram) {
253
253
  _gitCommandHistogram = meter().createHistogram("semiont.git.duration", {
254
- description: "Time spent in a synchronous git subprocess. These run on the event loop, so this duration is also time no other request could be served.",
254
+ description: "Wall time of a git subprocess. Async \u2014 this is latency, not event-loop blockage. Staging is deduped, so the `add` count is far below the number of appended events.",
255
255
  unit: "ms"
256
256
  });
257
257
  }
@@ -305,6 +305,53 @@ function registerFactPumpDepthProvider(provider) {
305
305
  });
306
306
  }
307
307
  }
308
+ var _gitStagingFailureCounter;
309
+ function recordGitStagingFailure(reason) {
310
+ if (!_gitStagingFailureCounter) {
311
+ _gitStagingFailureCounter = meter().createCounter("semiont.git.staging.failures", {
312
+ description: "Staging commands abandoned after retries; the index may be stale"
313
+ });
314
+ }
315
+ _gitStagingFailureCounter.add(1, { reason });
316
+ }
317
+ var PROCESS_START_TIME_SECONDS = Math.floor(Date.now() / 1e3);
318
+ var _processStartTimeGauge;
319
+ var _restartCountGauge;
320
+ var _restartCountProvider;
321
+ var _abnormalExitCounter;
322
+ function registerProcessLifetimeMetrics() {
323
+ if (_processStartTimeGauge) return;
324
+ _processStartTimeGauge = meter().createObservableGauge("semiont.process.start_time", {
325
+ description: "Unix seconds at which this process started; a change means it restarted",
326
+ unit: "s"
327
+ });
328
+ _processStartTimeGauge.addCallback((observer) => observer.observe(PROCESS_START_TIME_SECONDS));
329
+ }
330
+ function registerRestartCountProvider(provider) {
331
+ _restartCountProvider = provider;
332
+ if (!_restartCountGauge) {
333
+ _restartCountGauge = meter().createObservableGauge("semiont.process.restarts", {
334
+ description: "Times the supervisor has restarted this service"
335
+ });
336
+ _restartCountGauge.addCallback(async (observer) => {
337
+ if (_restartCountProvider) observer.observe(await _restartCountProvider());
338
+ });
339
+ }
340
+ }
341
+ function recordAbnormalTermination(reason, detail) {
342
+ if (!_abnormalExitCounter) {
343
+ _abnormalExitCounter = meter().createCounter("semiont.process.abnormal_exit", {
344
+ description: "Process terminations that were not a clean shutdown"
345
+ });
346
+ }
347
+ _abnormalExitCounter.add(1, { reason });
348
+ const active = trace.getActiveSpan();
349
+ if (active) {
350
+ active.setStatus({ code: SpanStatusCode.ERROR, message: `${reason}: ${detail ?? ""}`.trim() });
351
+ active.setAttribute("semiont.process.abnormal_exit", reason);
352
+ active.end();
353
+ }
354
+ }
308
355
  function registerVectorIndexSizeProvider(provider) {
309
356
  _vectorIndexSizeProvider = provider;
310
357
  if (!_vectorIndexSizeGauge) {
@@ -409,6 +456,6 @@ function recordAnchorOutcome(label, method) {
409
456
  anchorOutcomeCounter().add(1, { "detection.label": label, "anchor.method": method });
410
457
  }
411
458
 
412
- export { extractTraceparent, getActiveTraceparent, getLogTraceContext, injectTraceparent, recordAnchorOutcome, recordAppendStage, recordBusEmit, recordDetectionCall, recordGatherDegrade, recordGitCommand, recordHandlerDuration, recordInferenceUsage, recordJobOutcome, recordReplySuppressed, recordResumeGap, recordSubscriberConnect, recordSubscriberDisconnect, recordUnanswerableRequest, registerCorrelationRegistryProvider, registerFactPumpDepthProvider, registerJobQueueProvider, registerVectorIndexSizeProvider, withActorSpan, withSpan, withTraceparent };
459
+ export { extractTraceparent, getActiveTraceparent, getLogTraceContext, injectTraceparent, recordAbnormalTermination, recordAnchorOutcome, recordAppendStage, recordBusEmit, recordDetectionCall, recordGatherDegrade, recordGitCommand, recordGitStagingFailure, recordHandlerDuration, recordInferenceUsage, recordJobOutcome, recordReplySuppressed, recordResumeGap, recordSubscriberConnect, recordSubscriberDisconnect, recordUnanswerableRequest, registerCorrelationRegistryProvider, registerFactPumpDepthProvider, registerJobQueueProvider, registerProcessLifetimeMetrics, registerRestartCountProvider, registerVectorIndexSizeProvider, withActorSpan, withSpan, withTraceparent };
413
460
  //# sourceMappingURL=index.js.map
414
461
  //# sourceMappingURL=index.js.map
package/dist/index.js.map CHANGED
@@ -1 +1 @@
1
- {"version":3,"sources":["../src/index.ts"],"names":[],"mappings":";;;;;AAqDA,wBAAA,CAAyB,MAAM;AAC7B,EAAA,MAAM,IAAA,GAAO,MAAM,aAAA,EAAc;AACjC,EAAA,IAAI,CAAC,MAAM,OAAO,MAAA;AAClB,EAAA,MAAM,GAAA,GAAM,KAAK,WAAA,EAAY;AAC7B,EAAA,IAAI,CAAC,kBAAA,CAAmB,GAAG,CAAA,EAAG,OAAO,MAAA;AACrC,EAAA,OAAO,GAAA,CAAI,OAAA;AACb,CAAC,CAAA;AAED,IAAM,WAAA,GAAc,SAAA;AAEpB,IAAM,MAAA,GAAS,MAAM,KAAA,CAAM,SAAA,CAAU,WAAW,CAAA;AAShD,eAAsB,QAAA,CACpB,IAAA,EACA,EAAA,EACA,OAAA,EACY;AACZ,EAAA,MAAM,IAAA,GAAO,MAAA,EAAO,CAAE,SAAA,CAAU,IAAA,EAAM;AAAA,IACpC,IAAA,EAAM,OAAA,EAAS,IAAA,IAAQ,QAAA,CAAS,QAAA;AAAA,IAChC,GAAI,SAAS,KAAA,GAAQ,EAAE,YAAY,OAAA,CAAQ,KAAA,KAAU;AAAC,GACvD,CAAA;AACD,EAAA,IAAI;AACF,IAAA,OAAO,MAAM,OAAA,CAAQ,IAAA,CAAK,KAAA,CAAM,OAAA,CAAQ,OAAA,CAAQ,MAAA,EAAO,EAAG,IAAI,CAAA,EAAG,MAAM,EAAA,CAAG,IAAI,CAAC,CAAA;AAAA,EACjF,SAAS,GAAA,EAAK;AACZ,IAAA,IAAA,CAAK,gBAAgB,GAAY,CAAA;AACjC,IAAA,IAAA,CAAK,SAAA,CAAU;AAAA,MACb,MAAM,cAAA,CAAe,KAAA;AAAA,MACrB,SAAS,GAAA,YAAe,KAAA,GAAQ,GAAA,CAAI,OAAA,GAAU,OAAO,GAAG;AAAA,KACzD,CAAA;AACD,IAAA,MAAM,GAAA;AAAA,EACR,CAAA,SAAE;AACA,IAAA,IAAA,CAAK,GAAA,EAAI;AAAA,EACX;AACF;AAIA,IAAM,WAAA,GAAc,QAAA;AAmBb,SAAS,oBAAA,GAAiD;AAC/D,EAAA,MAAM,UAAkC,EAAC;AACzC,EAAA,WAAA,CAAY,MAAA,CAAO,OAAA,CAAQ,MAAA,EAAO,EAAG,OAAO,CAAA;AAC5C,EAAA,MAAM,WAAA,GAAc,QAAQ,aAAa,CAAA;AACzC,EAAA,IAAI,CAAC,aAAa,OAAO,MAAA;AACzB,EAAA,OAAO,OAAA,CAAQ,YAAY,CAAA,GACvB,EAAE,WAAA,EAAa,UAAA,EAAY,OAAA,CAAQ,YAAY,CAAA,EAAE,GACjD,EAAE,WAAA,EAAY;AACpB;AAOO,SAAS,kBAAqD,OAAA,EAAe;AAClF,EAAA,MAAM,UAAU,oBAAA,EAAqB;AACrC,EAAA,IAAI,OAAA,EAAS;AACX,IAAC,OAAA,CAAoC,WAAW,CAAA,GAAI,OAAA;AAAA,EACtD;AACA,EAAA,OAAO,OAAA;AACT;AAMO,SAAS,mBACd,OAAA,EAC0B;AAC1B,EAAA,MAAM,OAAA,GAAW,QAAoC,WAAW,CAAA;AAGhE,EAAA,IAAI,YAAY,MAAA,EAAW;AACzB,IAAA,OAAQ,QAAoC,WAAW,CAAA;AAAA,EACzD;AACA,EAAA,IAAI,CAAC,OAAA,IAAW,OAAO,OAAA,CAAQ,WAAA,KAAgB,UAAU,OAAO,MAAA;AAChE,EAAA,OAAO,OAAA;AACT;AAOO,SAAS,eAAA,CACd,SACA,EAAA,EACG;AACH,EAAA,IAAI,CAAC,OAAA,EAAS,OAAO,EAAA,EAAG;AACxB,EAAA,MAAM,UAAA,GAAqC,EAAE,WAAA,EAAa,OAAA,CAAQ,WAAA,EAAY;AAC9E,EAAA,IAAI,OAAA,CAAQ,UAAA,EAAY,UAAA,CAAW,YAAY,IAAI,OAAA,CAAQ,UAAA;AAC3D,EAAA,MAAM,MAAM,WAAA,CAAY,OAAA,CAAQ,OAAA,CAAQ,MAAA,IAAU,UAAU,CAAA;AAC5D,EAAA,OAAO,OAAA,CAAQ,IAAA,CAAK,GAAA,EAAK,EAAE,CAAA;AAC7B;AAgBA,eAAsB,aAAA,CACpB,KAAA,EACA,OAAA,EACA,EAAA,EACA,UAAA,EACY;AACZ,EAAA,MAAM,KAAA,GAAQ,YAAY,GAAA,EAAI;AAC9B,EAAA,IAAI;AACF,IAAA,OAAO,MAAM,QAAA,CAAS,CAAA,MAAA,EAAS,KAAK,CAAA,CAAA,EAAI,OAAO,IAAI,EAAA,EAAI;AAAA,MACrD,MAAM,QAAA,CAAS,QAAA;AAAA,MACf,KAAA,EAAO;AAAA,QACL,KAAA;AAAA,QACA,aAAA,EAAe,OAAA;AAAA,QACf,GAAI,cAAc;AAAC;AACrB,KACD,CAAA;AAAA,EACH,CAAA,SAAE;AACA,IAAA,qBAAA,CAAsB,KAAA,EAAO,OAAA,EAAS,WAAA,CAAY,GAAA,KAAQ,KAAK,CAAA;AAAA,EACjE;AACF;AAaO,SAAS,kBAAA,GAAwE;AACtF,EAAA,MAAM,IAAA,GAAO,MAAM,aAAA,EAAc;AACjC,EAAA,IAAI,CAAC,MAAM,OAAO,MAAA;AAClB,EAAA,MAAM,GAAA,GAAM,KAAK,WAAA,EAAY;AAC7B,EAAA,IAAI,CAAC,kBAAA,CAAmB,GAAG,CAAA,EAAG,OAAO,MAAA;AACrC,EAAA,OAAO,EAAE,QAAA,EAAU,GAAA,CAAI,OAAA,EAAS,OAAA,EAAS,IAAI,MAAA,EAAO;AACtD;AAIA,IAAM,UAAA,GAAa,SAAA;AAEnB,IAAM,KAAA,GAAQ,MAAM,OAAA,CAAQ,QAAA,CAAS,UAAU,CAAA;AAE/C,IAAI,eAAA;AACJ,IAAI,uBAAA;AACJ,IAAI,iBAAA;AACJ,IAAI,oBAAA;AACJ,IAAI,yBAAA;AACJ,IAAI,4BAAA;AACJ,IAAI,yBAAA;AACJ,IAAI,kBAAA;AACJ,IAAI,qBAAA;AACJ,IAAI,qBAAA;AACJ,IAAI,sBAAA;AACJ,IAAI,uBAAA;AACJ,IAAI,2BAAA;AACJ,IAAI,eAAA;AACJ,IAAI,cAAA;AACJ,IAAI,iBAAA;AACJ,IAAI,qBAAA;AACJ,IAAI,mBAAA;AACJ,IAAI,sBAAA;AACJ,IAAI,wBAAA;AAWJ,SAAS,cAAA,GAA0B;AACjC,EAAA,IAAI,CAAC,eAAA,EAAiB;AACpB,IAAA,eAAA,GAAkB,KAAA,EAAM,CAAE,aAAA,CAAc,kBAAA,EAAoB;AAAA,MAC1D,WAAA,EAAa;AAAA,KACd,CAAA;AAAA,EACH;AACA,EAAA,OAAO,eAAA;AACT;AAEA,SAAS,wBAAA,GAAsC;AAC7C,EAAA,IAAI,CAAC,yBAAA,EAA2B;AAC9B,IAAA,yBAAA,GAA4B,KAAA,EAAM,CAAE,eAAA,CAAgB,0BAAA,EAA4B;AAAA,MAC9E,WAAA,EAAa,mCAAA;AAAA,MACb,IAAA,EAAM;AAAA,KACP,CAAA;AAAA,EACH;AACA,EAAA,OAAO,yBAAA;AACT;AAEA,SAAS,iBAAA,GAA6B;AACpC,EAAA,IAAI,CAAC,kBAAA,EAAoB;AACvB,IAAA,kBAAA,GAAqB,KAAA,EAAM,CAAE,aAAA,CAAc,qBAAA,EAAuB;AAAA,MAChE,WAAA,EAAa;AAAA,KACd,CAAA;AAAA,EACH;AACA,EAAA,OAAO,kBAAA;AACT;AAEA,SAAS,oBAAA,GAAkC;AACzC,EAAA,IAAI,CAAC,qBAAA,EAAuB;AAC1B,IAAA,qBAAA,GAAwB,KAAA,EAAM,CAAE,eAAA,CAAgB,sBAAA,EAAwB;AAAA,MACtE,WAAA,EAAa,6BAAA;AAAA,MACb,IAAA,EAAM;AAAA,KACP,CAAA;AAAA,EACH;AACA,EAAA,OAAO,qBAAA;AACT;AAEA,SAAS,qBAAA,GAAiC;AACxC,EAAA,IAAI,CAAC,sBAAA,EAAwB;AAC3B,IAAA,sBAAA,GAAyB,KAAA,EAAM,CAAE,aAAA,CAAc,yBAAA,EAA2B;AAAA,MACxE,WAAA,EAAa;AAAA,KACd,CAAA;AAAA,EACH;AACA,EAAA,OAAO,sBAAA;AACT;AAEA,SAAS,sBAAA,GAAkC;AACzC,EAAA,IAAI,CAAC,uBAAA,EAAyB;AAC5B,IAAA,uBAAA,GAA0B,KAAA,EAAM,CAAE,aAAA,CAAc,0BAAA,EAA4B;AAAA,MAC1E,WAAA,EAAa;AAAA,KACd,CAAA;AAAA,EACH;AACA,EAAA,OAAO,uBAAA;AACT;AAEA,SAAS,0BAAA,GAAwC;AAC/C,EAAA,IAAI,CAAC,2BAAA,EAA6B;AAChC,IAAA,2BAAA,GAA8B,KAAA,EAAM,CAAE,eAAA,CAAgB,4BAAA,EAA8B;AAAA,MAClF,WAAA,EAAa,yDAAA;AAAA,MACb,IAAA,EAAM;AAAA,KACP,CAAA;AAAA,EACH;AACA,EAAA,OAAO,2BAAA;AACT;AAEA,SAAS,qBAAA,GAAuC;AAC9C,EAAA,IAAI,CAAC,eAAA,EAAiB;AACpB,IAAA,eAAA,GAAkB,KAAA,EAAM,CAAE,mBAAA,CAAoB,yBAAA,EAA2B;AAAA,MACvE,WAAA,EAAa;AAAA,KACd,CAAA;AAAA,EACH;AACA,EAAA,OAAO,eAAA;AACT;AAEA,SAAS,sBAAA,GAAkC;AACzC,EAAA,IAAI,CAAC,uBAAA,EAAyB;AAC5B,IAAA,uBAAA,GAA0B,KAAA,EAAM,CAAE,aAAA,CAAc,8BAAA,EAAgC;AAAA,MAC9E,WAAA,EAAa;AAAA,KACd,CAAA;AAAA,EACH;AACA,EAAA,OAAO,uBAAA;AACT;AAWO,SAAS,sBAAsB,OAAA,EAAuB;AAC3D,EAAA,sBAAA,GAAyB,GAAA,CAAI,CAAA,EAAG,EAAE,aAAA,EAAe,SAAS,CAAA;AAC5D;AAEA,SAAS,gBAAA,GAA4B;AACnC,EAAA,IAAI,CAAC,iBAAA,EAAmB;AACtB,IAAA,iBAAA,GAAoB,KAAA,EAAM,CAAE,aAAA,CAAc,wBAAA,EAA0B;AAAA,MAClE,WAAA,EAAa;AAAA,KACd,CAAA;AAAA,EACH;AACA,EAAA,OAAO,iBAAA;AACT;AAQO,SAAS,gBAAgB,MAAA,EAAsB;AACpD,EAAA,gBAAA,GAAmB,GAAA,CAAI,CAAA,EAAG,EAAE,uBAAA,EAAyB,QAAQ,CAAA;AAC/D;AAEA,SAAS,mBAAA,GAA+B;AACtC,EAAA,IAAI,CAAC,oBAAA,EAAsB;AACzB,IAAA,oBAAA,GAAuB,KAAA,EAAM,CAAE,aAAA,CAAc,0BAAA,EAA4B;AAAA,MACvE,WAAA,EAAa;AAAA,KACd,CAAA;AAAA,EACH;AACA,EAAA,OAAO,oBAAA;AACT;AAQO,SAAS,0BAA0B,OAAA,EAAuB;AAC/D,EAAA,mBAAA,GAAsB,GAAA,CAAI,CAAA,EAAG,EAAE,aAAA,EAAe,SAAS,CAAA;AACzD;AAiBO,SAAS,oCACd,QAAA,EACM;AACN,EAAA,4BAAA,GAA+B,QAAA;AAC/B,EAAA,IAAI,CAAC,yBAAA,EAA2B;AAC9B,IAAA,yBAAA,GAA4B,KAAA,EAAM,CAAE,qBAAA,CAAsB,8BAAA,EAAgC;AAAA,MACxF,WAAA,EAAa;AAAA,KACd,CAAA;AACD,IAAA,yBAAA,CAA0B,WAAA,CAAY,CAAC,QAAA,KAAa;AAClD,MAAA,IAAI,CAAC,4BAAA,EAA8B;AACnC,MAAA,MAAM,OAAO,4BAAA,EAA6B;AAC1C,MAAA,QAAA,CAAS,QAAQ,IAAA,CAAK,MAAA,EAAQ,EAAE,kBAAA,EAAoB,UAAU,CAAA;AAC9D,MAAA,QAAA,CAAS,QAAQ,IAAA,CAAK,eAAA,EAAiB,EAAE,kBAAA,EAAoB,oBAAoB,CAAA;AAAA,IACnF,CAAC,CAAA;AAAA,EACH;AACF;AAGO,SAAS,aAAA,CAAc,SAAiB,KAAA,EAAsB;AACnE,EAAA,cAAA,EAAe,CAAE,IAAI,CAAA,EAAG;AAAA,IACtB,aAAA,EAAe,OAAA;AAAA,IACf,GAAI,KAAA,GAAQ,EAAE,WAAA,EAAa,KAAA,KAAU;AAAC,GACvC,CAAA;AACH;AAGO,SAAS,qBAAA,CAAsB,KAAA,EAAe,OAAA,EAAiB,UAAA,EAA0B;AAC9F,EAAA,wBAAA,EAAyB,CAAE,OAAO,UAAA,EAAY;AAAA,IAC5C,KAAA;AAAA,IACA,aAAA,EAAe;AAAA,GAChB,CAAA;AACH;AAGO,SAAS,gBAAA,CAAiB,OAAA,EAAiB,OAAA,EAAiC,UAAA,EAA0B;AAC3G,EAAA,iBAAA,EAAkB,CAAE,IAAI,CAAA,EAAG,EAAE,YAAY,OAAA,EAAS,aAAA,EAAe,SAAS,CAAA;AAC1E,EAAA,oBAAA,EAAqB,CAAE,OAAO,UAAA,EAAY,EAAE,YAAY,OAAA,EAAS,aAAA,EAAe,SAAS,CAAA;AAC3F;AAEA,IAAI,qBAAA;AACJ,SAAS,oBAAA,GAAkC;AACzC,EAAA,IAAI,CAAC,qBAAA,EAAuB;AAC1B,IAAA,qBAAA,GAAwB,KAAA,EAAM,CAAE,eAAA,CAAgB,gCAAA,EAAkC;AAAA,MAChF,WAAA,EAAa,2LAAA;AAAA,MACb,IAAA,EAAM;AAAA,KACP,CAAA;AAAA,EACH;AACA,EAAA,OAAO,qBAAA;AACT;AAYO,SAAS,iBAAA,CACd,OACA,UAAA,EACM;AACN,EAAA,oBAAA,GAAuB,MAAA,CAAO,UAAA,EAAY,EAAE,cAAA,EAAgB,OAAO,CAAA;AACrE;AAEA,IAAI,oBAAA;AACJ,SAAS,mBAAA,GAAiC;AACxC,EAAA,IAAI,CAAC,oBAAA,EAAsB;AACzB,IAAA,oBAAA,GAAuB,KAAA,EAAM,CAAE,eAAA,CAAgB,sBAAA,EAAwB;AAAA,MACrE,WAAA,EAAa,0IAAA;AAAA,MACb,IAAA,EAAM;AAAA,KACP,CAAA;AAAA,EACH;AACA,EAAA,OAAO,oBAAA;AACT;AAaO,SAAS,gBAAA,CAAiB,SAAiB,UAAA,EAA0B;AAC1E,EAAA,mBAAA,GAAsB,MAAA,CAAO,UAAA,EAAY,EAAE,aAAA,EAAe,SAAS,CAAA;AACrE;AAEA,SAAS,oBAAA,GAAgC;AACvC,EAAA,IAAI,CAAC,qBAAA,EAAuB;AAC1B,IAAA,qBAAA,GAAwB,KAAA,EAAM,CAAE,aAAA,CAAc,yBAAA,EAA2B;AAAA,MACvE,WAAA,EAAa;AAAA,KACd,CAAA;AAAA,EACH;AACA,EAAA,OAAO,qBAAA;AACT;AAUO,SAAS,oBAAoB,UAAA,EAAuC;AACzE,EAAA,oBAAA,EAAqB,CAAE,GAAA,CAAI,CAAA,EAAG,EAAE,YAAY,CAAA;AAC9C;AAGO,SAAS,uBAAA,GAAgC;AAC9C,EAAA,qBAAA,EAAsB,CAAE,IAAI,CAAC,CAAA;AAC/B;AAGO,SAAS,0BAAA,GAAmC;AACjD,EAAA,qBAAA,EAAsB,CAAE,IAAI,EAAE,CAAA;AAChC;AASO,SAAS,yBACd,QAAA,EACM;AACN,EAAA,iBAAA,GAAoB,QAAA;AACpB,EAAA,IAAI,CAAC,cAAA,EAAgB;AACnB,IAAA,cAAA,GAAiB,KAAA,EAAM,CAAE,qBAAA,CAAsB,wBAAA,EAA0B;AAAA,MACvE,WAAA,EAAa;AAAA,KACd,CAAA;AACD,IAAA,cAAA,CAAe,WAAA,CAAY,OAAO,QAAA,KAAa;AAC7C,MAAA,IAAI,CAAC,iBAAA,EAAmB;AACxB,MAAA,MAAM,IAAA,GAAO,MAAM,iBAAA,EAAkB;AACrC,MAAA,QAAA,CAAS,QAAQ,IAAA,CAAK,OAAA,EAAS,EAAE,YAAA,EAAc,WAAW,CAAA;AAC1D,MAAA,QAAA,CAAS,QAAQ,IAAA,CAAK,OAAA,EAAS,EAAE,YAAA,EAAc,WAAW,CAAA;AAC1D,MAAA,QAAA,CAAS,QAAQ,IAAA,CAAK,QAAA,EAAU,EAAE,YAAA,EAAc,YAAY,CAAA;AAC5D,MAAA,QAAA,CAAS,QAAQ,IAAA,CAAK,MAAA,EAAQ,EAAE,YAAA,EAAc,UAAU,CAAA;AACxD,MAAA,QAAA,CAAS,QAAQ,IAAA,CAAK,SAAA,EAAW,EAAE,YAAA,EAAc,aAAa,CAAA;AAAA,IAChE,CAAC,CAAA;AAAA,EACH;AACF;AAkBO,SAAS,8BAA8B,QAAA,EAA8B;AAC1E,EAAA,sBAAA,GAAyB,QAAA;AACzB,EAAA,IAAI,CAAC,mBAAA,EAAqB;AACxB,IAAA,mBAAA,GAAsB,KAAA,EAAM,CAAE,qBAAA,CAAsB,mCAAA,EAAqC;AAAA,MACvF,WAAA,EAAa;AAAA,KACd,CAAA;AACD,IAAA,mBAAA,CAAoB,WAAA,CAAY,CAAC,QAAA,KAAa;AAC5C,MAAA,IAAI,sBAAA,EAAwB,QAAA,CAAS,OAAA,CAAQ,sBAAA,EAAwB,CAAA;AAAA,IACvE,CAAC,CAAA;AAAA,EACH;AACF;AAEO,SAAS,gCACd,QAAA,EACM;AACN,EAAA,wBAAA,GAA2B,QAAA;AAC3B,EAAA,IAAI,CAAC,qBAAA,EAAuB;AAC1B,IAAA,qBAAA,GAAwB,KAAA,EAAM,CAAE,qBAAA,CAAsB,2BAAA,EAA6B;AAAA,MACjF,WAAA,EAAa;AAAA,KACd,CAAA;AACD,IAAA,qBAAA,CAAsB,WAAA,CAAY,OAAO,QAAA,KAAa;AACpD,MAAA,IAAI,wBAAA,EAA0B;AAC5B,QAAA,MAAM,KAAA,GAAQ,MAAM,wBAAA,EAAyB;AAC7C,QAAA,QAAA,CAAS,QAAQ,KAAK,CAAA;AAAA,MACxB;AAAA,IACF,CAAC,CAAA;AAAA,EACH;AACF;AAOO,SAAS,qBAAqB,IAAA,EAO5B;AACP,EAAA,MAAM,SAAA,GAAY;AAAA,IAChB,sBAAsB,IAAA,CAAK,QAAA;AAAA,IAC3B,mBAAmB,IAAA,CAAK,KAAA;AAAA,IACxB,qBAAqB,IAAA,CAAK;AAAA,GAC5B;AACA,EAAA,qBAAA,EAAsB,CAAE,GAAA,CAAI,CAAA,EAAG,SAAS,CAAA;AACxC,EAAA,0BAAA,EAA2B,CAAE,MAAA,CAAO,IAAA,CAAK,UAAA,EAAY,SAAS,CAAA;AAC9D,EAAA,IAAI,IAAA,CAAK,WAAA,IAAe,IAAA,IAAQ,IAAA,CAAK,cAAc,CAAA,EAAG;AACpD,IAAA,sBAAA,EAAuB,CAAE,GAAA,CAAI,IAAA,CAAK,WAAA,EAAa;AAAA,MAC7C,sBAAsB,IAAA,CAAK,QAAA;AAAA,MAC3B,mBAAmB,IAAA,CAAK,KAAA;AAAA,MACxB,qBAAA,EAAuB;AAAA,KACxB,CAAA;AAAA,EACH;AACA,EAAA,IAAI,IAAA,CAAK,YAAA,IAAgB,IAAA,IAAQ,IAAA,CAAK,eAAe,CAAA,EAAG;AACtD,IAAA,sBAAA,EAAuB,CAAE,GAAA,CAAI,IAAA,CAAK,YAAA,EAAc;AAAA,MAC9C,sBAAsB,IAAA,CAAK,QAAA;AAAA,MAC3B,mBAAmB,IAAA,CAAK,KAAA;AAAA,MACxB,qBAAA,EAAuB;AAAA,KACxB,CAAA;AAAA,EACH;AACF;AAEA,IAAI,qBAAA;AACJ,IAAI,2BAAA;AACJ,IAAI,wBAAA;AACJ,IAAI,yBAAA;AAEJ,SAAS,oBAAA,GAAgC;AACvC,EAAA,IAAI,CAAC,qBAAA,EAAuB;AAC1B,IAAA,qBAAA,GAAwB,KAAA,EAAM,CAAE,aAAA,CAAc,yBAAA,EAA2B;AAAA,MACvE,WAAA,EAAa;AAAA,KACd,CAAA;AAAA,EACH;AACA,EAAA,OAAO,qBAAA;AACT;AAEA,SAAS,0BAAA,GAAwC;AAC/C,EAAA,IAAI,CAAC,2BAAA,EAA6B;AAChC,IAAA,2BAAA,GAA8B,KAAA,EAAM,CAAE,eAAA,CAAgB,iCAAA,EAAmC;AAAA,MACvF,WAAA,EAAa,qGAAA;AAAA,MACb,IAAA,EAAM;AAAA,KACP,CAAA;AAAA,EACH;AACA,EAAA,OAAO,2BAAA;AACT;AAEA,SAAS,uBAAA,GAAqC;AAC5C,EAAA,IAAI,CAAC,wBAAA,EAA0B;AAC7B,IAAA,wBAAA,GAA2B,KAAA,EAAM,CAAE,eAAA,CAAgB,8BAAA,EAAgC;AAAA,MACjF,WAAA,EAAa;AAAA,KACd,CAAA;AAAA,EACH;AACA,EAAA,OAAO,wBAAA;AACT;AAEA,SAAS,wBAAA,GAAsC;AAC7C,EAAA,IAAI,CAAC,yBAAA,EAA2B;AAC9B,IAAA,yBAAA,GAA4B,KAAA,EAAM,CAAE,eAAA,CAAgB,+BAAA,EAAiC;AAAA,MACnF,WAAA,EAAa;AAAA,KACd,CAAA;AAAA,EACH;AACA,EAAA,OAAO,yBAAA;AACT;AAoBO,SAAS,oBAAoB,IAAA,EAU3B;AACP,EAAA,MAAM,KAAA,GAAQ;AAAA,IACZ,mBAAmB,IAAA,CAAK,KAAA;AAAA,IACxB,qBAAqB,IAAA,CAAK,OAAA;AAAA,IAC1B,mBAAmB,IAAA,CAAK,KAAA;AAAA,IACxB,oBAAoB,IAAA,CAAK;AAAA,GAC3B;AACA,EAAA,oBAAA,EAAqB,CAAE,GAAA,CAAI,CAAA,EAAG,KAAK,CAAA;AACnC,EAAA,0BAAA,EAA2B,CAAE,MAAA,CAAO,IAAA,CAAK,UAAA,EAAY,KAAK,CAAA;AAC1D,EAAA,uBAAA,EAAwB,CAAE,MAAA,CAAO,IAAA,CAAK,KAAA,EAAO,KAAK,CAAA;AAClD,EAAA,IAAI,IAAA,CAAK,gBAAgB,MAAA,EAAW;AAClC,IAAA,wBAAA,EAAyB,CAAE,OAAO,IAAA,CAAK,WAAA,EAAa,EAAE,GAAG,KAAA,EAAO,qBAAA,EAAuB,OAAA,EAAS,CAAA;AAAA,EAClG;AACA,EAAA,IAAI,IAAA,CAAK,iBAAiB,MAAA,EAAW;AACnC,IAAA,wBAAA,EAAyB,CAAE,OAAO,IAAA,CAAK,YAAA,EAAc,EAAE,GAAG,KAAA,EAAO,qBAAA,EAAuB,QAAA,EAAU,CAAA;AAAA,EACpG;AACF;AAEA,IAAI,qBAAA;AACJ,SAAS,oBAAA,GAAgC;AACvC,EAAA,IAAI,CAAC,qBAAA,EAAuB;AAC1B,IAAA,qBAAA,GAAwB,KAAA,EAAM,CAAE,aAAA,CAAc,2BAAA,EAA6B;AAAA,MACzE,WAAA,EAAa;AAAA,KACd,CAAA;AAAA,EACH;AACA,EAAA,OAAO,qBAAA;AACT;AAkBO,SAAS,mBAAA,CAAoB,OAAe,MAAA,EAAsB;AACvE,EAAA,oBAAA,EAAqB,CAAE,IAAI,CAAA,EAAG,EAAE,mBAAmB,KAAA,EAAO,eAAA,EAAiB,QAAQ,CAAA;AACrF","file":"index.js","sourcesContent":["/**\n * @semiont/observability — public API.\n *\n * Universal surface (works in Node + browser). For SDK *initialization*,\n * import from `@semiont/observability/node` or `/web` at the process entry\n * point. Everything else uses this module.\n *\n * Tier 2 of `.plans/OBSERVABILITY.md`. The public surface:\n *\n * - `withSpan(name, fn, options?)` — wrap an async block in a span;\n * `options` carries `kind` and `attrs`.\n * - `withActorSpan(actor, channel, fn, extraAttrs?)` — consumer-span\n * wrapper for bus-event handlers, with handler-duration recording.\n * - `injectTraceparent(payload)` / `extractTraceparent(payload)` — W3C\n * trace-context propagation across the SSE channel (the bus payload\n * gets a `_trace?: { traceparent }` sibling to `correlationId`).\n * - `withTraceparent(carrier, fn)` — run `fn` with the incoming\n * traceparent as the parent context.\n * - `getActiveTraceparent()` — read the active span's traceparent for\n * manual propagation (e.g. attaching to a fetch header or SSE field).\n * - `getLogTraceContext()` — active `trace_id` / `span_id` for log-line\n * correlation.\n * - Metric recorders (`recordBusEmit`, `recordHandlerDuration`,\n * `recordJobOutcome`, `recordSubscriberConnect` / `Disconnect`,\n * `recordInferenceUsage`) and gauge providers\n * (`registerJobQueueProvider`, `registerVectorIndexSizeProvider`).\n *\n * No-op when no exporter is configured: `@opentelemetry/api`'s default\n * tracer is a no-op, so `withSpan` is essentially free until\n * `initObservability*()` runs.\n */\n\nimport {\n context,\n isSpanContextValid,\n metrics,\n propagation,\n SpanKind,\n SpanStatusCode,\n trace,\n type Attributes,\n type Counter,\n type Histogram,\n type ObservableGauge,\n type Span,\n type UpDownCounter,\n} from '@opentelemetry/api';\nimport { setBusLogTraceIdProvider } from '@semiont/core';\n\n// Wire `busLog`'s trace-id provider once at module load. When an OTel\n// SDK is initialized (and a span is active when `busLog` fires), the\n// emitted line gets a `trace=<8hex>` suffix that correlates the\n// grep-timeline with the trace UI. No-op when no SDK is active.\nsetBusLogTraceIdProvider(() => {\n const span = trace.getActiveSpan();\n if (!span) return undefined;\n const ctx = span.spanContext();\n if (!isSpanContextValid(ctx)) return undefined;\n return ctx.traceId;\n});\n\nconst TRACER_NAME = 'semiont';\n\nconst tracer = () => trace.getTracer(TRACER_NAME);\n\n// ── withSpan ───────────────────────────────────────────────────────────\n\n/**\n * Wrap an async block in a span. The span is started before `fn` runs and\n * ended after it resolves or rejects; exceptions are recorded and the span\n * status is set to ERROR. `kind` defaults to INTERNAL.\n */\nexport async function withSpan<T>(\n name: string,\n fn: (span: Span) => Promise<T> | T,\n options?: { kind?: SpanKind; attrs?: Attributes },\n): Promise<T> {\n const span = tracer().startSpan(name, {\n kind: options?.kind ?? SpanKind.INTERNAL,\n ...(options?.attrs ? { attributes: options.attrs } : {}),\n });\n try {\n return await context.with(trace.setSpan(context.active(), span), () => fn(span));\n } catch (err) {\n span.recordException(err as Error);\n span.setStatus({\n code: SpanStatusCode.ERROR,\n message: err instanceof Error ? err.message : String(err),\n });\n throw err;\n } finally {\n span.end();\n }\n}\n\n// ── Traceparent on bus payloads ────────────────────────────────────────\n\nconst TRACE_FIELD = '_trace';\n\n/**\n * Sibling of `correlationId` on bus payloads. Lives on the SSE event body\n * because SSE has no header trailer; the SDK strips it before delivering\n * the payload to subscribers. Additive — payloads without `_trace` parse\n * unchanged.\n */\nexport interface TraceCarrier {\n /** W3C `traceparent` header value (`00-<traceId>-<spanId>-<flags>`). */\n traceparent: string;\n /** W3C `tracestate` header value (vendor-specific extensions). */\n tracestate?: string;\n}\n\n/**\n * Read the active span's W3C traceparent (and tracestate). Returns\n * `undefined` if no span is active.\n */\nexport function getActiveTraceparent(): TraceCarrier | undefined {\n const carrier: Record<string, string> = {};\n propagation.inject(context.active(), carrier);\n const traceparent = carrier['traceparent'];\n if (!traceparent) return undefined;\n return carrier['tracestate']\n ? { traceparent, tracestate: carrier['tracestate'] }\n : { traceparent };\n}\n\n/**\n * Attach the active span's trace-context to a payload object as\n * `_trace`. No-op when no span is active. Returns the same object\n * reference for chaining.\n */\nexport function injectTraceparent<T extends Record<string, unknown>>(payload: T): T {\n const carrier = getActiveTraceparent();\n if (carrier) {\n (payload as Record<string, unknown>)[TRACE_FIELD] = carrier;\n }\n return payload;\n}\n\n/**\n * Strip and return the `_trace` field from a payload. Mutates `payload`.\n * The field is internal plumbing and should not be visible to subscribers.\n */\nexport function extractTraceparent<T extends Record<string, unknown>>(\n payload: T,\n): TraceCarrier | undefined {\n const carrier = (payload as Record<string, unknown>)[TRACE_FIELD] as\n | TraceCarrier\n | undefined;\n if (carrier !== undefined) {\n delete (payload as Record<string, unknown>)[TRACE_FIELD];\n }\n if (!carrier || typeof carrier.traceparent !== 'string') return undefined;\n return carrier;\n}\n\n/**\n * Run `fn` with the given W3C traceparent set as the parent context.\n * Any spans started inside `fn` will be children of the incoming trace.\n * No-op if `carrier` is undefined.\n */\nexport function withTraceparent<T>(\n carrier: TraceCarrier | undefined,\n fn: () => T,\n): T {\n if (!carrier) return fn();\n const carrierObj: Record<string, string> = { traceparent: carrier.traceparent };\n if (carrier.tracestate) carrierObj['tracestate'] = carrier.tracestate;\n const ctx = propagation.extract(context.active(), carrierObj);\n return context.with(ctx, fn);\n}\n\n// ── Actor handler convenience ──────────────────────────────────────────\n\n/**\n * Wrap a bus-event handler in an `actor.<name>:<channel>` consumer span.\n * Used at every `eventBus.get(channel).subscribe(handler)` site inside\n * an actor (Stower, Gatherer, Matcher, Browser, Smelter), to attribute\n * each in-process subscriber's work to a span without scattering manual\n * `withSpan` calls across handler bodies.\n *\n * The span's parent is the active context at the time the handler\n * fires — which is the `bus.dispatch:<channel>` span on the gateway\n * (Subject.next runs synchronously inside the dispatch span), or the\n * `bus.emit:<channel>` span when an actor emits to itself.\n */\nexport async function withActorSpan<T>(\n actor: string,\n channel: string,\n fn: (span: Span) => Promise<T> | T,\n extraAttrs?: Attributes,\n): Promise<T> {\n const start = performance.now();\n try {\n return await withSpan(`actor.${actor}:${channel}`, fn, {\n kind: SpanKind.CONSUMER,\n attrs: {\n actor,\n 'bus.channel': channel,\n ...(extraAttrs ?? {}),\n },\n });\n } finally {\n recordHandlerDuration(actor, channel, performance.now() - start);\n }\n}\n\n// ── Log correlation ────────────────────────────────────────────────────\n\n/**\n * Read the active span's `trace_id` / `span_id` for log-line correlation.\n * Tier 3 of `.plans/OBSERVABILITY.md`. Each structured log line gets\n * tagged with these so a log query in CloudWatch / Loki / Datadog can\n * jump to the trace in Tempo / Jaeger / X-Ray.\n *\n * Returns `undefined` if no span is active, or if the active span's\n * context is invalid (uninitialized SDK, no-op tracer).\n */\nexport function getLogTraceContext(): { trace_id: string; span_id: string } | undefined {\n const span = trace.getActiveSpan();\n if (!span) return undefined;\n const ctx = span.spanContext();\n if (!isSpanContextValid(ctx)) return undefined;\n return { trace_id: ctx.traceId, span_id: ctx.spanId };\n}\n\n// ── Metrics — Tier 3 ───────────────────────────────────────────────────\n\nconst METER_NAME = 'semiont';\n\nconst meter = () => metrics.getMeter(METER_NAME);\n\nlet _busEmitCounter: Counter | undefined;\nlet _replySuppressedCounter: Counter | undefined;\nlet _resumeGapCounter: Counter | undefined;\nlet _unanswerableCounter: Counter | undefined;\nlet _correlationRegistryGauge: ObservableGauge | undefined;\nlet _correlationRegistryProvider: (() => CorrelationRegistrySnapshot) | undefined;\nlet _handlerDurationHistogram: Histogram | undefined;\nlet _jobOutcomeCounter: Counter | undefined;\nlet _jobDurationHistogram: Histogram | undefined;\nlet _gatherDegradeCounter: Counter | undefined;\nlet _inferenceCallsCounter: Counter | undefined;\nlet _inferenceTokensCounter: Counter | undefined;\nlet _inferenceDurationHistogram: Histogram | undefined;\nlet _sseSubscribers: UpDownCounter | undefined;\nlet _jobQueueGauge: ObservableGauge | undefined;\nlet _jobQueueProvider: (() => Promise<JobQueueSnapshot> | JobQueueSnapshot) | undefined;\nlet _vectorIndexSizeGauge: ObservableGauge | undefined;\nlet _factPumpDepthGauge: ObservableGauge | undefined;\nlet _factPumpDepthProvider: (() => number) | undefined;\nlet _vectorIndexSizeProvider: (() => Promise<number> | number) | undefined;\n\n/** Snapshot of job-queue contents by status. Match `JobQueue.getStats()`. */\nexport interface JobQueueSnapshot {\n pending: number;\n running: number;\n complete: number;\n failed: number;\n cancelled: number;\n}\n\nfunction busEmitCounter(): Counter {\n if (!_busEmitCounter) {\n _busEmitCounter = meter().createCounter('semiont.bus.emit', {\n description: 'Bus emits by channel and scope',\n });\n }\n return _busEmitCounter;\n}\n\nfunction handlerDurationHistogram(): Histogram {\n if (!_handlerDurationHistogram) {\n _handlerDurationHistogram = meter().createHistogram('semiont.handler.duration', {\n description: 'In-process actor handler duration',\n unit: 'ms',\n });\n }\n return _handlerDurationHistogram;\n}\n\nfunction jobOutcomeCounter(): Counter {\n if (!_jobOutcomeCounter) {\n _jobOutcomeCounter = meter().createCounter('semiont.job.outcome', {\n description: 'Worker job completions by type and outcome',\n });\n }\n return _jobOutcomeCounter;\n}\n\nfunction jobDurationHistogram(): Histogram {\n if (!_jobDurationHistogram) {\n _jobDurationHistogram = meter().createHistogram('semiont.job.duration', {\n description: 'Worker job duration by type',\n unit: 'ms',\n });\n }\n return _jobDurationHistogram;\n}\n\nfunction inferenceCallsCounter(): Counter {\n if (!_inferenceCallsCounter) {\n _inferenceCallsCounter = meter().createCounter('semiont.inference.calls', {\n description: 'Inference API calls by provider, model, and outcome',\n });\n }\n return _inferenceCallsCounter;\n}\n\nfunction inferenceTokensCounter(): Counter {\n if (!_inferenceTokensCounter) {\n _inferenceTokensCounter = meter().createCounter('semiont.inference.tokens', {\n description: 'Inference token usage by provider, model, and direction',\n });\n }\n return _inferenceTokensCounter;\n}\n\nfunction inferenceDurationHistogram(): Histogram {\n if (!_inferenceDurationHistogram) {\n _inferenceDurationHistogram = meter().createHistogram('semiont.inference.duration', {\n description: 'Inference call duration by provider, model, and outcome',\n unit: 'ms',\n });\n }\n return _inferenceDurationHistogram;\n}\n\nfunction sseSubscribersCounter(): UpDownCounter {\n if (!_sseSubscribers) {\n _sseSubscribers = meter().createUpDownCounter('semiont.sse.subscribers', {\n description: 'Active SSE subscribers',\n });\n }\n return _sseSubscribers;\n}\n\nfunction replySuppressedCounter(): Counter {\n if (!_replySuppressedCounter) {\n _replySuppressedCounter = meter().createCounter('semiont.bus.reply.suppressed', {\n description: 'Correlated replies withheld from a non-owning subscriber',\n });\n }\n return _replySuppressedCounter;\n}\n\n/**\n * A correlated reply was withheld from a subscriber that does not own its\n * correlationId (CORRELATED-REPLY-ROUTING P5).\n *\n * Counts ONLY that case. A frame with no correlationId is a shape violation\n * (warned, not counted), and a cid nobody claimed is the structural in-process\n * case that fires constantly — counting either would drown the signal this\n * metric exists to show: the fan-out amplification the delivery filter removes.\n */\nexport function recordReplySuppressed(channel: string): void {\n replySuppressedCounter().add(1, { 'bus.channel': channel });\n}\n\nfunction resumeGapCounter(): Counter {\n if (!_resumeGapCounter) {\n _resumeGapCounter = meter().createCounter('semiont.bus.resume_gap', {\n description: 'SSE resumes that degraded to a gap because replay was unavailable',\n });\n }\n return _resumeGapCounter;\n}\n\n/**\n * An SSE resume could not be served and the client was told to fall back to\n * cache. This degradation is CORRECT by design and therefore silent — which is\n * exactly why it needs a number. A rising rate means clients are losing\n * history, and nothing else in the stack says so.\n */\nexport function recordResumeGap(reason: string): void {\n resumeGapCounter().add(1, { 'bus.resume_gap.reason': reason });\n}\n\nfunction unanswerableCounter(): Counter {\n if (!_unanswerableCounter) {\n _unanswerableCounter = meter().createCounter('semiont.bus.unanswerable', {\n description: 'Request emits that reached zero subscribers and were failed at the gateway',\n });\n }\n return _unanswerableCounter;\n}\n\n/**\n * A request-shaped emit reached no subscriber, so the gateway synthesized its\n * mapped failure (ARCHIVIST-STAYS-UP P3). By channel, this is the absence rate\n * of the service that answers it — the difference between \"it went down once\"\n * and \"it is flapping.\"\n */\nexport function recordUnanswerableRequest(channel: string): void {\n unanswerableCounter().add(1, { 'bus.channel': channel });\n}\n\n/** Claims held and reply payloads retained by a gateway's correlation registry. */\nexport interface CorrelationRegistrySnapshot {\n claims: number;\n retainedReplies: number;\n}\n\n/**\n * Register a callback returning the gateway's correlation-registry occupancy.\n *\n * COUNTS, not bytes: retention is count-budgeted today (byte-budgeting is a\n * known limit in CORRELATED-REPLY-ROUTING), so `retainedReplies` is a proxy for\n * heap, not a measure of it. It is still the closest observable to the question\n * two OOM investigations keep asking — a browse result can be 1-2 MB, and up to\n * REPLY_RETENTION_MAX of them are held at once.\n */\nexport function registerCorrelationRegistryProvider(\n provider: () => CorrelationRegistrySnapshot,\n): void {\n _correlationRegistryProvider = provider;\n if (!_correlationRegistryGauge) {\n _correlationRegistryGauge = meter().createObservableGauge('semiont.bus.correlation.size', {\n description: 'Correlation registry occupancy: live claims and retained reply payloads',\n });\n _correlationRegistryGauge.addCallback((observer) => {\n if (!_correlationRegistryProvider) return;\n const snap = _correlationRegistryProvider();\n observer.observe(snap.claims, { 'correlation.kind': 'claims' });\n observer.observe(snap.retainedReplies, { 'correlation.kind': 'retained_replies' });\n });\n }\n}\n\n/** Increment the bus-emit counter. Called at every transport `emit` site. */\nexport function recordBusEmit(channel: string, scope?: string): void {\n busEmitCounter().add(1, {\n 'bus.channel': channel,\n ...(scope ? { 'bus.scope': scope } : {}),\n });\n}\n\n/** Record an in-process actor handler's duration. */\nexport function recordHandlerDuration(actor: string, channel: string, durationMs: number): void {\n handlerDurationHistogram().record(durationMs, {\n actor,\n 'bus.channel': channel,\n });\n}\n\n/** Record a worker job's outcome and duration. */\nexport function recordJobOutcome(jobType: string, outcome: 'completed' | 'failed', durationMs: number): void {\n jobOutcomeCounter().add(1, { 'job.type': jobType, 'job.outcome': outcome });\n jobDurationHistogram().record(durationMs, { 'job.type': jobType, 'job.outcome': outcome });\n}\n\nlet _appendStageHistogram: Histogram | undefined;\nfunction appendStageHistogram(): Histogram {\n if (!_appendStageHistogram) {\n _appendStageHistogram = meter().createHistogram('semiont.record.append.duration', {\n description: 'Time spent in one stage of appending an event to the record, labeled by stage: persist (JSONL write + git), materialize (view rebuild), enrich, publish. The Archivist\\'s core write path.',\n unit: 'ms',\n });\n }\n return _appendStageHistogram;\n}\n\n/**\n * Record one stage of `EventStore.appendEvent` (ARCHIVIST-STAYS-UP P7).\n *\n * The append path is the one operation only the Archivist can perform, and it\n * was entirely dark: reads had `recordHandlerDuration` and the bus had its own\n * counters, while writes had nothing. Stage-labeled because the useful\n * question is never \"was the append slow\" but WHICH PART — and `materialize`\n * in particular does work proportional to a resource's annotation count, so it\n * degrades with history rather than with load.\n */\nexport function recordAppendStage(\n stage: 'persist' | 'materialize' | 'enrich' | 'publish',\n durationMs: number,\n): void {\n appendStageHistogram().record(durationMs, { 'record.stage': stage });\n}\n\nlet _gitCommandHistogram: Histogram | undefined;\nfunction gitCommandHistogram(): Histogram {\n if (!_gitCommandHistogram) {\n _gitCommandHistogram = meter().createHistogram('semiont.git.duration', {\n description: 'Time spent in a synchronous git subprocess. These run on the event loop, so this duration is also time no other request could be served.',\n unit: 'ms',\n });\n }\n return _gitCommandHistogram;\n}\n\n/**\n * Record a synchronous git invocation (ARCHIVIST-STAYS-UP P7).\n *\n * These are `execFileSync`, so **the duration is event-loop blockage, not just\n * latency** — every concurrent `browse:*` read waits behind it. One `git add`\n * runs per appended event, so a detection job writing hundreds of annotations\n * spawns hundreds of blocking subprocesses. That is the suspected mechanism\n * behind \"reads serializing behind the detection job's annotation writes\" in\n * `bugs/absent-archivist-wedges-browse.md`, which recorded the symptom without\n * a cause. This number is what turns that from a hypothesis into a reading.\n */\nexport function recordGitCommand(command: string, durationMs: number): void {\n gitCommandHistogram().record(durationMs, { 'git.command': command });\n}\n\nfunction gatherDegradeCounter(): Counter {\n if (!_gatherDegradeCounter) {\n _gatherDegradeCounter = meter().createCounter('semiont.gather.degraded', {\n description: 'Gathers that degraded because an eventually-consistent projection did not catch up within its read barrier (vectors: absent semanticContext; graph: projection-lag failure). Labeled by projection.',\n });\n }\n return _gatherDegradeCounter;\n}\n\n/**\n * Record a gather degraded by a projection read barrier: `'vectors'` — the\n * Smelter settle barrier timed out (semanticContext shipped absent);\n * `'graph'` — the Weaver applied barrier + poll floor exhausted (projection\n * lag surfaced as a distinct failure). Fleet-alertable counterpart of the\n * `[gather DEGRADED]` L4 breadcrumbs — a rising rate on either label means\n * that pipeline is not keeping up.\n */\nexport function recordGatherDegrade(projection: 'graph' | 'vectors'): void {\n gatherDegradeCounter().add(1, { projection });\n}\n\n/** Increment the SSE subscriber gauge — call on `/bus/subscribe` open. */\nexport function recordSubscriberConnect(): void {\n sseSubscribersCounter().add(1);\n}\n\n/** Decrement on disconnect. Pair with `recordSubscriberConnect`. */\nexport function recordSubscriberDisconnect(): void {\n sseSubscribersCounter().add(-1);\n}\n\n/**\n * Register a callback that returns the current job-queue snapshot.\n * Polled at the SDK's metric-collection interval. The single gauge\n * emits one observation per status (`pending`, `running`, …) tagged\n * with the `job.status` attribute. Idempotent — last registered\n * provider wins.\n */\nexport function registerJobQueueProvider(\n provider: () => Promise<JobQueueSnapshot> | JobQueueSnapshot,\n): void {\n _jobQueueProvider = provider;\n if (!_jobQueueGauge) {\n _jobQueueGauge = meter().createObservableGauge('semiont.job.queue.size', {\n description: 'Job queue size by status',\n });\n _jobQueueGauge.addCallback(async (observer) => {\n if (!_jobQueueProvider) return;\n const snap = await _jobQueueProvider();\n observer.observe(snap.pending, { 'job.status': 'pending' });\n observer.observe(snap.running, { 'job.status': 'running' });\n observer.observe(snap.complete, { 'job.status': 'complete' });\n observer.observe(snap.failed, { 'job.status': 'failed' });\n observer.observe(snap.cancelled, { 'job.status': 'cancelled' });\n });\n }\n}\n\n/**\n * Register a callback that returns the current vector-index size\n * (point count). Async to allow remote queries (Qdrant). Polled at\n * the metric-collection interval.\n */\n/**\n * Register the Archivist's fact-pump backlog — facts appended to the record\n * but not yet republished onto the bus.\n *\n * At rest this is zero. A value that climbs and does not come back means the\n * pump is outrunning its transport, which is the leading hypothesis for the\n * load-correlated heap growth in `bugs/absent-archivist-wedges-browse.md`\n * (ARCHIVIST-STAYS-UP P5). The backlog is deliberately unbounded today, so\n * this number is the only thing standing between \"the pump is behind\" and an\n * OOM whose cause is inferred from RSS after the fact.\n */\nexport function registerFactPumpDepthProvider(provider: () => number): void {\n _factPumpDepthProvider = provider;\n if (!_factPumpDepthGauge) {\n _factPumpDepthGauge = meter().createObservableGauge('semiont.archivist.fact_pump.depth', {\n description: 'Facts appended to the record but not yet published to the bus. Zero at rest; a rising floor means the pump is behind its transport.',\n });\n _factPumpDepthGauge.addCallback((observer) => {\n if (_factPumpDepthProvider) observer.observe(_factPumpDepthProvider());\n });\n }\n}\n\nexport function registerVectorIndexSizeProvider(\n provider: () => Promise<number> | number,\n): void {\n _vectorIndexSizeProvider = provider;\n if (!_vectorIndexSizeGauge) {\n _vectorIndexSizeGauge = meter().createObservableGauge('semiont.vector.index.size', {\n description: 'Vector store point count',\n });\n _vectorIndexSizeGauge.addCallback(async (observer) => {\n if (_vectorIndexSizeProvider) {\n const value = await _vectorIndexSizeProvider();\n observer.observe(value);\n }\n });\n }\n}\n\n/**\n * Record an inference call. Token counts are optional — providers that\n * don't expose them (or fail before generating) record only call count\n * and duration.\n */\nexport function recordInferenceUsage(opts: {\n provider: string;\n model: string;\n durationMs: number;\n outcome: 'success' | 'error';\n inputTokens?: number;\n outputTokens?: number;\n}): void {\n const baseAttrs = {\n 'inference.provider': opts.provider,\n 'inference.model': opts.model,\n 'inference.outcome': opts.outcome,\n };\n inferenceCallsCounter().add(1, baseAttrs);\n inferenceDurationHistogram().record(opts.durationMs, baseAttrs);\n if (opts.inputTokens != null && opts.inputTokens > 0) {\n inferenceTokensCounter().add(opts.inputTokens, {\n 'inference.provider': opts.provider,\n 'inference.model': opts.model,\n 'inference.direction': 'input',\n });\n }\n if (opts.outputTokens != null && opts.outputTokens > 0) {\n inferenceTokensCounter().add(opts.outputTokens, {\n 'inference.provider': opts.provider,\n 'inference.model': opts.model,\n 'inference.direction': 'output',\n });\n }\n}\n\nlet _detectionCallCounter: Counter | undefined;\nlet _detectionDurationHistogram: Histogram | undefined;\nlet _detectionItemsHistogram: Histogram | undefined;\nlet _detectionTokensHistogram: Histogram | undefined;\n\nfunction detectionCallCounter(): Counter {\n if (!_detectionCallCounter) {\n _detectionCallCounter = meter().createCounter('semiont.detection.calls', {\n description: 'Detection model calls, labeled by motivation, outcome, subdivision depth and whether this was the floor re-roll.',\n });\n }\n return _detectionCallCounter;\n}\n\nfunction detectionDurationHistogram(): Histogram {\n if (!_detectionDurationHistogram) {\n _detectionDurationHistogram = meter().createHistogram('semiont.detection.call.duration', {\n description: 'Wall time of one detection model call, including the attempts that failed and were retried smaller.',\n unit: 'ms',\n });\n }\n return _detectionDurationHistogram;\n}\n\nfunction detectionItemsHistogram(): Histogram {\n if (!_detectionItemsHistogram) {\n _detectionItemsHistogram = meter().createHistogram('semiont.detection.call.items', {\n description: 'Annotations returned by one detection call. Against the input size on the same record, this is yield.',\n });\n }\n return _detectionItemsHistogram;\n}\n\nfunction detectionTokensHistogram(): Histogram {\n if (!_detectionTokensHistogram) {\n _detectionTokensHistogram = meter().createHistogram('semiont.detection.call.tokens', {\n description: \"Provider-reported tokens for one detection call, by direction. Deliberately separate from semiont.inference.tokens: that series is the authoritative total but carries no subdivision depth, and 'what does a depth-2 call cost' is the question every sizing decision asks.\",\n });\n }\n return _detectionTokensHistogram;\n}\n\n/**\n * Record one detection model call (DETECTION-QUALITY-THROUGHPUT P1).\n *\n * The adapters already record provider/model/duration/tokens for every\n * inference call. What they cannot know is the detection shape around it:\n * which motivation asked, how big the piece was, how many annotations came\n * back, how deep subdivision had descended, and whether this was the floor\n * re-roll. Those are the facts that distinguish a healthy call from a\n * expensive descent, and without them a slow detection run is one\n * undifferentiated number.\n *\n * FAILED attempts are recorded too, and that is the point: the calls paid for\n * and thrown away during a descent are exactly the cost later phases exist to\n * avoid, so a record only of successes would hide the thing being optimized.\n *\n * Tokens are the PROVIDER's counts, passed through — never estimated. Absent\n * means the provider did not report them.\n */\nexport function recordDetectionCall(opts: {\n label: string;\n pieceChars: number;\n durationMs: number;\n items: number;\n depth: number;\n reroll: boolean;\n outcome: 'success' | 'truncated' | 'timeout' | 'collapsed' | 'error';\n inputTokens?: number;\n outputTokens?: number;\n}): void {\n const attrs = {\n 'detection.label': opts.label,\n 'detection.outcome': opts.outcome,\n 'detection.depth': opts.depth,\n 'detection.reroll': opts.reroll,\n };\n detectionCallCounter().add(1, attrs);\n detectionDurationHistogram().record(opts.durationMs, attrs);\n detectionItemsHistogram().record(opts.items, attrs);\n if (opts.inputTokens !== undefined) {\n detectionTokensHistogram().record(opts.inputTokens, { ...attrs, 'detection.direction': 'input' });\n }\n if (opts.outputTokens !== undefined) {\n detectionTokensHistogram().record(opts.outputTokens, { ...attrs, 'detection.direction': 'output' });\n }\n}\n\nlet _anchorOutcomeCounter: Counter | undefined;\nfunction anchorOutcomeCounter(): Counter {\n if (!_anchorOutcomeCounter) {\n _anchorOutcomeCounter = meter().createCounter('semiont.detection.anchors', {\n description: 'Every annotation anchoring, labeled by the method that resolved it. EVERY outcome is counted, not just the risky ones, because a bare count of degraded anchors has no denominator — the rate is the precision signal.',\n });\n }\n return _anchorOutcomeCounter;\n}\n\n/**\n * Record how one annotation got anchored (DETECTION-QUALITY-THROUGHPUT P5).\n *\n * The selector-vs-source check is already a WRITE-TIME INVARIANT — both\n * `buildTextAnnotation` and `buildPdfAnnotation` throw on a selector that does\n * not match its source — so mechanical correctness is guaranteed rather than\n * sampled, and auditing it would measure a constant.\n *\n * What is genuinely uncertain is which anchoring METHOD got there. An `exact`\n * the model quoted verbatim and that appears once is certain; one resolved by\n * `first-of-many` (several occurrences, no usable context) or `fuzzy-match`\n * picked a plausible occurrence and may have picked wrong. Those were visible\n * only as log warnings — countable by a human reading worker output, which is\n * how 47 of them went unreviewed. As a rate they are the precision number that\n * sits beside the yield numbers.\n */\nexport function recordAnchorOutcome(label: string, method: string): void {\n anchorOutcomeCounter().add(1, { 'detection.label': label, 'anchor.method': method });\n}\n\n// ── Re-exports from @opentelemetry/api ─────────────────────────────────\n\nexport { SpanKind, SpanStatusCode, type Attributes, type Span } from '@opentelemetry/api';\n"]}
1
+ {"version":3,"sources":["../src/index.ts"],"names":[],"mappings":";;;;;AAqDA,wBAAA,CAAyB,MAAM;AAC7B,EAAA,MAAM,IAAA,GAAO,MAAM,aAAA,EAAc;AACjC,EAAA,IAAI,CAAC,MAAM,OAAO,MAAA;AAClB,EAAA,MAAM,GAAA,GAAM,KAAK,WAAA,EAAY;AAC7B,EAAA,IAAI,CAAC,kBAAA,CAAmB,GAAG,CAAA,EAAG,OAAO,MAAA;AACrC,EAAA,OAAO,GAAA,CAAI,OAAA;AACb,CAAC,CAAA;AAED,IAAM,WAAA,GAAc,SAAA;AAEpB,IAAM,MAAA,GAAS,MAAM,KAAA,CAAM,SAAA,CAAU,WAAW,CAAA;AAShD,eAAsB,QAAA,CACpB,IAAA,EACA,EAAA,EACA,OAAA,EACY;AACZ,EAAA,MAAM,IAAA,GAAO,MAAA,EAAO,CAAE,SAAA,CAAU,IAAA,EAAM;AAAA,IACpC,IAAA,EAAM,OAAA,EAAS,IAAA,IAAQ,QAAA,CAAS,QAAA;AAAA,IAChC,GAAI,SAAS,KAAA,GAAQ,EAAE,YAAY,OAAA,CAAQ,KAAA,KAAU;AAAC,GACvD,CAAA;AACD,EAAA,IAAI;AACF,IAAA,OAAO,MAAM,OAAA,CAAQ,IAAA,CAAK,KAAA,CAAM,OAAA,CAAQ,OAAA,CAAQ,MAAA,EAAO,EAAG,IAAI,CAAA,EAAG,MAAM,EAAA,CAAG,IAAI,CAAC,CAAA;AAAA,EACjF,SAAS,GAAA,EAAK;AACZ,IAAA,IAAA,CAAK,gBAAgB,GAAY,CAAA;AACjC,IAAA,IAAA,CAAK,SAAA,CAAU;AAAA,MACb,MAAM,cAAA,CAAe,KAAA;AAAA,MACrB,SAAS,GAAA,YAAe,KAAA,GAAQ,GAAA,CAAI,OAAA,GAAU,OAAO,GAAG;AAAA,KACzD,CAAA;AACD,IAAA,MAAM,GAAA;AAAA,EACR,CAAA,SAAE;AACA,IAAA,IAAA,CAAK,GAAA,EAAI;AAAA,EACX;AACF;AAIA,IAAM,WAAA,GAAc,QAAA;AAmBb,SAAS,oBAAA,GAAiD;AAC/D,EAAA,MAAM,UAAkC,EAAC;AACzC,EAAA,WAAA,CAAY,MAAA,CAAO,OAAA,CAAQ,MAAA,EAAO,EAAG,OAAO,CAAA;AAC5C,EAAA,MAAM,WAAA,GAAc,QAAQ,aAAa,CAAA;AACzC,EAAA,IAAI,CAAC,aAAa,OAAO,MAAA;AACzB,EAAA,OAAO,OAAA,CAAQ,YAAY,CAAA,GACvB,EAAE,WAAA,EAAa,UAAA,EAAY,OAAA,CAAQ,YAAY,CAAA,EAAE,GACjD,EAAE,WAAA,EAAY;AACpB;AAOO,SAAS,kBAAqD,OAAA,EAAe;AAClF,EAAA,MAAM,UAAU,oBAAA,EAAqB;AACrC,EAAA,IAAI,OAAA,EAAS;AACX,IAAC,OAAA,CAAoC,WAAW,CAAA,GAAI,OAAA;AAAA,EACtD;AACA,EAAA,OAAO,OAAA;AACT;AAMO,SAAS,mBACd,OAAA,EAC0B;AAC1B,EAAA,MAAM,OAAA,GAAW,QAAoC,WAAW,CAAA;AAGhE,EAAA,IAAI,YAAY,MAAA,EAAW;AACzB,IAAA,OAAQ,QAAoC,WAAW,CAAA;AAAA,EACzD;AACA,EAAA,IAAI,CAAC,OAAA,IAAW,OAAO,OAAA,CAAQ,WAAA,KAAgB,UAAU,OAAO,MAAA;AAChE,EAAA,OAAO,OAAA;AACT;AAOO,SAAS,eAAA,CACd,SACA,EAAA,EACG;AACH,EAAA,IAAI,CAAC,OAAA,EAAS,OAAO,EAAA,EAAG;AACxB,EAAA,MAAM,UAAA,GAAqC,EAAE,WAAA,EAAa,OAAA,CAAQ,WAAA,EAAY;AAC9E,EAAA,IAAI,OAAA,CAAQ,UAAA,EAAY,UAAA,CAAW,YAAY,IAAI,OAAA,CAAQ,UAAA;AAC3D,EAAA,MAAM,MAAM,WAAA,CAAY,OAAA,CAAQ,OAAA,CAAQ,MAAA,IAAU,UAAU,CAAA;AAC5D,EAAA,OAAO,OAAA,CAAQ,IAAA,CAAK,GAAA,EAAK,EAAE,CAAA;AAC7B;AAgBA,eAAsB,aAAA,CACpB,KAAA,EACA,OAAA,EACA,EAAA,EACA,UAAA,EACY;AACZ,EAAA,MAAM,KAAA,GAAQ,YAAY,GAAA,EAAI;AAC9B,EAAA,IAAI;AACF,IAAA,OAAO,MAAM,QAAA,CAAS,CAAA,MAAA,EAAS,KAAK,CAAA,CAAA,EAAI,OAAO,IAAI,EAAA,EAAI;AAAA,MACrD,MAAM,QAAA,CAAS,QAAA;AAAA,MACf,KAAA,EAAO;AAAA,QACL,KAAA;AAAA,QACA,aAAA,EAAe,OAAA;AAAA,QACf,GAAI,cAAc;AAAC;AACrB,KACD,CAAA;AAAA,EACH,CAAA,SAAE;AACA,IAAA,qBAAA,CAAsB,KAAA,EAAO,OAAA,EAAS,WAAA,CAAY,GAAA,KAAQ,KAAK,CAAA;AAAA,EACjE;AACF;AAaO,SAAS,kBAAA,GAAwE;AACtF,EAAA,MAAM,IAAA,GAAO,MAAM,aAAA,EAAc;AACjC,EAAA,IAAI,CAAC,MAAM,OAAO,MAAA;AAClB,EAAA,MAAM,GAAA,GAAM,KAAK,WAAA,EAAY;AAC7B,EAAA,IAAI,CAAC,kBAAA,CAAmB,GAAG,CAAA,EAAG,OAAO,MAAA;AACrC,EAAA,OAAO,EAAE,QAAA,EAAU,GAAA,CAAI,OAAA,EAAS,OAAA,EAAS,IAAI,MAAA,EAAO;AACtD;AAIA,IAAM,UAAA,GAAa,SAAA;AAEnB,IAAM,KAAA,GAAQ,MAAM,OAAA,CAAQ,QAAA,CAAS,UAAU,CAAA;AAE/C,IAAI,eAAA;AACJ,IAAI,uBAAA;AACJ,IAAI,iBAAA;AACJ,IAAI,oBAAA;AACJ,IAAI,yBAAA;AACJ,IAAI,4BAAA;AACJ,IAAI,yBAAA;AACJ,IAAI,kBAAA;AACJ,IAAI,qBAAA;AACJ,IAAI,qBAAA;AACJ,IAAI,sBAAA;AACJ,IAAI,uBAAA;AACJ,IAAI,2BAAA;AACJ,IAAI,eAAA;AACJ,IAAI,cAAA;AACJ,IAAI,iBAAA;AACJ,IAAI,qBAAA;AACJ,IAAI,mBAAA;AACJ,IAAI,sBAAA;AACJ,IAAI,wBAAA;AAWJ,SAAS,cAAA,GAA0B;AACjC,EAAA,IAAI,CAAC,eAAA,EAAiB;AACpB,IAAA,eAAA,GAAkB,KAAA,EAAM,CAAE,aAAA,CAAc,kBAAA,EAAoB;AAAA,MAC1D,WAAA,EAAa;AAAA,KACd,CAAA;AAAA,EACH;AACA,EAAA,OAAO,eAAA;AACT;AAEA,SAAS,wBAAA,GAAsC;AAC7C,EAAA,IAAI,CAAC,yBAAA,EAA2B;AAC9B,IAAA,yBAAA,GAA4B,KAAA,EAAM,CAAE,eAAA,CAAgB,0BAAA,EAA4B;AAAA,MAC9E,WAAA,EAAa,mCAAA;AAAA,MACb,IAAA,EAAM;AAAA,KACP,CAAA;AAAA,EACH;AACA,EAAA,OAAO,yBAAA;AACT;AAEA,SAAS,iBAAA,GAA6B;AACpC,EAAA,IAAI,CAAC,kBAAA,EAAoB;AACvB,IAAA,kBAAA,GAAqB,KAAA,EAAM,CAAE,aAAA,CAAc,qBAAA,EAAuB;AAAA,MAChE,WAAA,EAAa;AAAA,KACd,CAAA;AAAA,EACH;AACA,EAAA,OAAO,kBAAA;AACT;AAEA,SAAS,oBAAA,GAAkC;AACzC,EAAA,IAAI,CAAC,qBAAA,EAAuB;AAC1B,IAAA,qBAAA,GAAwB,KAAA,EAAM,CAAE,eAAA,CAAgB,sBAAA,EAAwB;AAAA,MACtE,WAAA,EAAa,6BAAA;AAAA,MACb,IAAA,EAAM;AAAA,KACP,CAAA;AAAA,EACH;AACA,EAAA,OAAO,qBAAA;AACT;AAEA,SAAS,qBAAA,GAAiC;AACxC,EAAA,IAAI,CAAC,sBAAA,EAAwB;AAC3B,IAAA,sBAAA,GAAyB,KAAA,EAAM,CAAE,aAAA,CAAc,yBAAA,EAA2B;AAAA,MACxE,WAAA,EAAa;AAAA,KACd,CAAA;AAAA,EACH;AACA,EAAA,OAAO,sBAAA;AACT;AAEA,SAAS,sBAAA,GAAkC;AACzC,EAAA,IAAI,CAAC,uBAAA,EAAyB;AAC5B,IAAA,uBAAA,GAA0B,KAAA,EAAM,CAAE,aAAA,CAAc,0BAAA,EAA4B;AAAA,MAC1E,WAAA,EAAa;AAAA,KACd,CAAA;AAAA,EACH;AACA,EAAA,OAAO,uBAAA;AACT;AAEA,SAAS,0BAAA,GAAwC;AAC/C,EAAA,IAAI,CAAC,2BAAA,EAA6B;AAChC,IAAA,2BAAA,GAA8B,KAAA,EAAM,CAAE,eAAA,CAAgB,4BAAA,EAA8B;AAAA,MAClF,WAAA,EAAa,yDAAA;AAAA,MACb,IAAA,EAAM;AAAA,KACP,CAAA;AAAA,EACH;AACA,EAAA,OAAO,2BAAA;AACT;AAEA,SAAS,qBAAA,GAAuC;AAC9C,EAAA,IAAI,CAAC,eAAA,EAAiB;AACpB,IAAA,eAAA,GAAkB,KAAA,EAAM,CAAE,mBAAA,CAAoB,yBAAA,EAA2B;AAAA,MACvE,WAAA,EAAa;AAAA,KACd,CAAA;AAAA,EACH;AACA,EAAA,OAAO,eAAA;AACT;AAEA,SAAS,sBAAA,GAAkC;AACzC,EAAA,IAAI,CAAC,uBAAA,EAAyB;AAC5B,IAAA,uBAAA,GAA0B,KAAA,EAAM,CAAE,aAAA,CAAc,8BAAA,EAAgC;AAAA,MAC9E,WAAA,EAAa;AAAA,KACd,CAAA;AAAA,EACH;AACA,EAAA,OAAO,uBAAA;AACT;AAWO,SAAS,sBAAsB,OAAA,EAAuB;AAC3D,EAAA,sBAAA,GAAyB,GAAA,CAAI,CAAA,EAAG,EAAE,aAAA,EAAe,SAAS,CAAA;AAC5D;AAEA,SAAS,gBAAA,GAA4B;AACnC,EAAA,IAAI,CAAC,iBAAA,EAAmB;AACtB,IAAA,iBAAA,GAAoB,KAAA,EAAM,CAAE,aAAA,CAAc,wBAAA,EAA0B;AAAA,MAClE,WAAA,EAAa;AAAA,KACd,CAAA;AAAA,EACH;AACA,EAAA,OAAO,iBAAA;AACT;AAQO,SAAS,gBAAgB,MAAA,EAAsB;AACpD,EAAA,gBAAA,GAAmB,GAAA,CAAI,CAAA,EAAG,EAAE,uBAAA,EAAyB,QAAQ,CAAA;AAC/D;AAEA,SAAS,mBAAA,GAA+B;AACtC,EAAA,IAAI,CAAC,oBAAA,EAAsB;AACzB,IAAA,oBAAA,GAAuB,KAAA,EAAM,CAAE,aAAA,CAAc,0BAAA,EAA4B;AAAA,MACvE,WAAA,EAAa;AAAA,KACd,CAAA;AAAA,EACH;AACA,EAAA,OAAO,oBAAA;AACT;AAQO,SAAS,0BAA0B,OAAA,EAAuB;AAC/D,EAAA,mBAAA,GAAsB,GAAA,CAAI,CAAA,EAAG,EAAE,aAAA,EAAe,SAAS,CAAA;AACzD;AAiBO,SAAS,oCACd,QAAA,EACM;AACN,EAAA,4BAAA,GAA+B,QAAA;AAC/B,EAAA,IAAI,CAAC,yBAAA,EAA2B;AAC9B,IAAA,yBAAA,GAA4B,KAAA,EAAM,CAAE,qBAAA,CAAsB,8BAAA,EAAgC;AAAA,MACxF,WAAA,EAAa;AAAA,KACd,CAAA;AACD,IAAA,yBAAA,CAA0B,WAAA,CAAY,CAAC,QAAA,KAAa;AAClD,MAAA,IAAI,CAAC,4BAAA,EAA8B;AACnC,MAAA,MAAM,OAAO,4BAAA,EAA6B;AAC1C,MAAA,QAAA,CAAS,QAAQ,IAAA,CAAK,MAAA,EAAQ,EAAE,kBAAA,EAAoB,UAAU,CAAA;AAC9D,MAAA,QAAA,CAAS,QAAQ,IAAA,CAAK,eAAA,EAAiB,EAAE,kBAAA,EAAoB,oBAAoB,CAAA;AAAA,IACnF,CAAC,CAAA;AAAA,EACH;AACF;AAGO,SAAS,aAAA,CAAc,SAAiB,KAAA,EAAsB;AACnE,EAAA,cAAA,EAAe,CAAE,IAAI,CAAA,EAAG;AAAA,IACtB,aAAA,EAAe,OAAA;AAAA,IACf,GAAI,KAAA,GAAQ,EAAE,WAAA,EAAa,KAAA,KAAU;AAAC,GACvC,CAAA;AACH;AAGO,SAAS,qBAAA,CAAsB,KAAA,EAAe,OAAA,EAAiB,UAAA,EAA0B;AAC9F,EAAA,wBAAA,EAAyB,CAAE,OAAO,UAAA,EAAY;AAAA,IAC5C,KAAA;AAAA,IACA,aAAA,EAAe;AAAA,GAChB,CAAA;AACH;AAGO,SAAS,gBAAA,CAAiB,OAAA,EAAiB,OAAA,EAAiC,UAAA,EAA0B;AAC3G,EAAA,iBAAA,EAAkB,CAAE,IAAI,CAAA,EAAG,EAAE,YAAY,OAAA,EAAS,aAAA,EAAe,SAAS,CAAA;AAC1E,EAAA,oBAAA,EAAqB,CAAE,OAAO,UAAA,EAAY,EAAE,YAAY,OAAA,EAAS,aAAA,EAAe,SAAS,CAAA;AAC3F;AAEA,IAAI,qBAAA;AACJ,SAAS,oBAAA,GAAkC;AACzC,EAAA,IAAI,CAAC,qBAAA,EAAuB;AAC1B,IAAA,qBAAA,GAAwB,KAAA,EAAM,CAAE,eAAA,CAAgB,gCAAA,EAAkC;AAAA,MAChF,WAAA,EAAa,2LAAA;AAAA,MACb,IAAA,EAAM;AAAA,KACP,CAAA;AAAA,EACH;AACA,EAAA,OAAO,qBAAA;AACT;AAYO,SAAS,iBAAA,CACd,OACA,UAAA,EACM;AACN,EAAA,oBAAA,GAAuB,MAAA,CAAO,UAAA,EAAY,EAAE,cAAA,EAAgB,OAAO,CAAA;AACrE;AAEA,IAAI,oBAAA;AACJ,SAAS,mBAAA,GAAiC;AACxC,EAAA,IAAI,CAAC,oBAAA,EAAsB;AACzB,IAAA,oBAAA,GAAuB,KAAA,EAAM,CAAE,eAAA,CAAgB,sBAAA,EAAwB;AAAA,MACrE,WAAA,EAAa,0KAAA;AAAA,MACb,IAAA,EAAM;AAAA,KACP,CAAA;AAAA,EACH;AACA,EAAA,OAAO,oBAAA;AACT;AAMO,SAAS,gBAAA,CAAiB,SAAiB,UAAA,EAA0B;AAC1E,EAAA,mBAAA,GAAsB,MAAA,CAAO,UAAA,EAAY,EAAE,aAAA,EAAe,SAAS,CAAA;AACrE;AAEA,SAAS,oBAAA,GAAgC;AACvC,EAAA,IAAI,CAAC,qBAAA,EAAuB;AAC1B,IAAA,qBAAA,GAAwB,KAAA,EAAM,CAAE,aAAA,CAAc,yBAAA,EAA2B;AAAA,MACvE,WAAA,EAAa;AAAA,KACd,CAAA;AAAA,EACH;AACA,EAAA,OAAO,qBAAA;AACT;AAUO,SAAS,oBAAoB,UAAA,EAAuC;AACzE,EAAA,oBAAA,EAAqB,CAAE,GAAA,CAAI,CAAA,EAAG,EAAE,YAAY,CAAA;AAC9C;AAGO,SAAS,uBAAA,GAAgC;AAC9C,EAAA,qBAAA,EAAsB,CAAE,IAAI,CAAC,CAAA;AAC/B;AAGO,SAAS,0BAAA,GAAmC;AACjD,EAAA,qBAAA,EAAsB,CAAE,IAAI,EAAE,CAAA;AAChC;AASO,SAAS,yBACd,QAAA,EACM;AACN,EAAA,iBAAA,GAAoB,QAAA;AACpB,EAAA,IAAI,CAAC,cAAA,EAAgB;AACnB,IAAA,cAAA,GAAiB,KAAA,EAAM,CAAE,qBAAA,CAAsB,wBAAA,EAA0B;AAAA,MACvE,WAAA,EAAa;AAAA,KACd,CAAA;AACD,IAAA,cAAA,CAAe,WAAA,CAAY,OAAO,QAAA,KAAa;AAC7C,MAAA,IAAI,CAAC,iBAAA,EAAmB;AACxB,MAAA,MAAM,IAAA,GAAO,MAAM,iBAAA,EAAkB;AACrC,MAAA,QAAA,CAAS,QAAQ,IAAA,CAAK,OAAA,EAAS,EAAE,YAAA,EAAc,WAAW,CAAA;AAC1D,MAAA,QAAA,CAAS,QAAQ,IAAA,CAAK,OAAA,EAAS,EAAE,YAAA,EAAc,WAAW,CAAA;AAC1D,MAAA,QAAA,CAAS,QAAQ,IAAA,CAAK,QAAA,EAAU,EAAE,YAAA,EAAc,YAAY,CAAA;AAC5D,MAAA,QAAA,CAAS,QAAQ,IAAA,CAAK,MAAA,EAAQ,EAAE,YAAA,EAAc,UAAU,CAAA;AACxD,MAAA,QAAA,CAAS,QAAQ,IAAA,CAAK,SAAA,EAAW,EAAE,YAAA,EAAc,aAAa,CAAA;AAAA,IAChE,CAAC,CAAA;AAAA,EACH;AACF;AAkBO,SAAS,8BAA8B,QAAA,EAA8B;AAC1E,EAAA,sBAAA,GAAyB,QAAA;AACzB,EAAA,IAAI,CAAC,mBAAA,EAAqB;AACxB,IAAA,mBAAA,GAAsB,KAAA,EAAM,CAAE,qBAAA,CAAsB,mCAAA,EAAqC;AAAA,MACvF,WAAA,EAAa;AAAA,KACd,CAAA;AACD,IAAA,mBAAA,CAAoB,WAAA,CAAY,CAAC,QAAA,KAAa;AAC5C,MAAA,IAAI,sBAAA,EAAwB,QAAA,CAAS,OAAA,CAAQ,sBAAA,EAAwB,CAAA;AAAA,IACvE,CAAC,CAAA;AAAA,EACH;AACF;AAEA,IAAI,yBAAA;AAQG,SAAS,wBAAwB,MAAA,EAAsC;AAC5E,EAAA,IAAI,CAAC,yBAAA,EAA2B;AAC9B,IAAA,yBAAA,GAA4B,KAAA,EAAM,CAAE,aAAA,CAAc,8BAAA,EAAgC;AAAA,MAChF,WAAA,EAAa;AAAA,KACd,CAAA;AAAA,EACH;AACA,EAAA,yBAAA,CAA0B,GAAA,CAAI,CAAA,EAAG,EAAE,MAAA,EAAQ,CAAA;AAC7C;AAcA,IAAM,6BAA6B,IAAA,CAAK,KAAA,CAAM,IAAA,CAAK,GAAA,KAAQ,GAAI,CAAA;AAC/D,IAAI,sBAAA;AACJ,IAAI,kBAAA;AACJ,IAAI,qBAAA;AACJ,IAAI,oBAAA;AAGG,SAAS,8BAAA,GAAuC;AACrD,EAAA,IAAI,sBAAA,EAAwB;AAC5B,EAAA,sBAAA,GAAyB,KAAA,EAAM,CAAE,qBAAA,CAAsB,4BAAA,EAA8B;AAAA,IACnF,WAAA,EAAa,yEAAA;AAAA,IACb,IAAA,EAAM;AAAA,GACP,CAAA;AACD,EAAA,sBAAA,CAAuB,YAAY,CAAC,QAAA,KAAa,QAAA,CAAS,OAAA,CAAQ,0BAA0B,CAAC,CAAA;AAC/F;AAQO,SAAS,6BACd,QAAA,EACM;AACN,EAAA,qBAAA,GAAwB,QAAA;AACxB,EAAA,IAAI,CAAC,kBAAA,EAAoB;AACvB,IAAA,kBAAA,GAAqB,KAAA,EAAM,CAAE,qBAAA,CAAsB,0BAAA,EAA4B;AAAA,MAC7E,WAAA,EAAa;AAAA,KACd,CAAA;AACD,IAAA,kBAAA,CAAmB,WAAA,CAAY,OAAO,QAAA,KAAa;AACjD,MAAA,IAAI,qBAAA,EAAuB,QAAA,CAAS,OAAA,CAAQ,MAAM,uBAAuB,CAAA;AAAA,IAC3E,CAAC,CAAA;AAAA,EACH;AACF;AAOO,SAAS,yBAAA,CAA0B,QAAgB,MAAA,EAAuB;AAC/E,EAAA,IAAI,CAAC,oBAAA,EAAsB;AACzB,IAAA,oBAAA,GAAuB,KAAA,EAAM,CAAE,aAAA,CAAc,+BAAA,EAAiC;AAAA,MAC5E,WAAA,EAAa;AAAA,KACd,CAAA;AAAA,EACH;AACA,EAAA,oBAAA,CAAqB,GAAA,CAAI,CAAA,EAAG,EAAE,MAAA,EAAQ,CAAA;AACtC,EAAA,MAAM,MAAA,GAAS,MAAM,aAAA,EAAc;AACnC,EAAA,IAAI,MAAA,EAAQ;AACV,IAAA,MAAA,CAAO,SAAA,CAAU,EAAE,IAAA,EAAM,cAAA,CAAe,OAAO,OAAA,EAAS,CAAA,EAAG,MAAM,CAAA,EAAA,EAAK,MAAA,IAAU,EAAE,CAAA,CAAA,CAAG,IAAA,IAAQ,CAAA;AAC7F,IAAA,MAAA,CAAO,YAAA,CAAa,iCAAiC,MAAM,CAAA;AAC3D,IAAA,MAAA,CAAO,GAAA,EAAI;AAAA,EACb;AACF;AAEO,SAAS,gCACd,QAAA,EACM;AACN,EAAA,wBAAA,GAA2B,QAAA;AAC3B,EAAA,IAAI,CAAC,qBAAA,EAAuB;AAC1B,IAAA,qBAAA,GAAwB,KAAA,EAAM,CAAE,qBAAA,CAAsB,2BAAA,EAA6B;AAAA,MACjF,WAAA,EAAa;AAAA,KACd,CAAA;AACD,IAAA,qBAAA,CAAsB,WAAA,CAAY,OAAO,QAAA,KAAa;AACpD,MAAA,IAAI,wBAAA,EAA0B;AAC5B,QAAA,MAAM,KAAA,GAAQ,MAAM,wBAAA,EAAyB;AAC7C,QAAA,QAAA,CAAS,QAAQ,KAAK,CAAA;AAAA,MACxB;AAAA,IACF,CAAC,CAAA;AAAA,EACH;AACF;AAOO,SAAS,qBAAqB,IAAA,EAO5B;AACP,EAAA,MAAM,SAAA,GAAY;AAAA,IAChB,sBAAsB,IAAA,CAAK,QAAA;AAAA,IAC3B,mBAAmB,IAAA,CAAK,KAAA;AAAA,IACxB,qBAAqB,IAAA,CAAK;AAAA,GAC5B;AACA,EAAA,qBAAA,EAAsB,CAAE,GAAA,CAAI,CAAA,EAAG,SAAS,CAAA;AACxC,EAAA,0BAAA,EAA2B,CAAE,MAAA,CAAO,IAAA,CAAK,UAAA,EAAY,SAAS,CAAA;AAC9D,EAAA,IAAI,IAAA,CAAK,WAAA,IAAe,IAAA,IAAQ,IAAA,CAAK,cAAc,CAAA,EAAG;AACpD,IAAA,sBAAA,EAAuB,CAAE,GAAA,CAAI,IAAA,CAAK,WAAA,EAAa;AAAA,MAC7C,sBAAsB,IAAA,CAAK,QAAA;AAAA,MAC3B,mBAAmB,IAAA,CAAK,KAAA;AAAA,MACxB,qBAAA,EAAuB;AAAA,KACxB,CAAA;AAAA,EACH;AACA,EAAA,IAAI,IAAA,CAAK,YAAA,IAAgB,IAAA,IAAQ,IAAA,CAAK,eAAe,CAAA,EAAG;AACtD,IAAA,sBAAA,EAAuB,CAAE,GAAA,CAAI,IAAA,CAAK,YAAA,EAAc;AAAA,MAC9C,sBAAsB,IAAA,CAAK,QAAA;AAAA,MAC3B,mBAAmB,IAAA,CAAK,KAAA;AAAA,MACxB,qBAAA,EAAuB;AAAA,KACxB,CAAA;AAAA,EACH;AACF;AAEA,IAAI,qBAAA;AACJ,IAAI,2BAAA;AACJ,IAAI,wBAAA;AACJ,IAAI,yBAAA;AAEJ,SAAS,oBAAA,GAAgC;AACvC,EAAA,IAAI,CAAC,qBAAA,EAAuB;AAC1B,IAAA,qBAAA,GAAwB,KAAA,EAAM,CAAE,aAAA,CAAc,yBAAA,EAA2B;AAAA,MACvE,WAAA,EAAa;AAAA,KACd,CAAA;AAAA,EACH;AACA,EAAA,OAAO,qBAAA;AACT;AAEA,SAAS,0BAAA,GAAwC;AAC/C,EAAA,IAAI,CAAC,2BAAA,EAA6B;AAChC,IAAA,2BAAA,GAA8B,KAAA,EAAM,CAAE,eAAA,CAAgB,iCAAA,EAAmC;AAAA,MACvF,WAAA,EAAa,qGAAA;AAAA,MACb,IAAA,EAAM;AAAA,KACP,CAAA;AAAA,EACH;AACA,EAAA,OAAO,2BAAA;AACT;AAEA,SAAS,uBAAA,GAAqC;AAC5C,EAAA,IAAI,CAAC,wBAAA,EAA0B;AAC7B,IAAA,wBAAA,GAA2B,KAAA,EAAM,CAAE,eAAA,CAAgB,8BAAA,EAAgC;AAAA,MACjF,WAAA,EAAa;AAAA,KACd,CAAA;AAAA,EACH;AACA,EAAA,OAAO,wBAAA;AACT;AAEA,SAAS,wBAAA,GAAsC;AAC7C,EAAA,IAAI,CAAC,yBAAA,EAA2B;AAC9B,IAAA,yBAAA,GAA4B,KAAA,EAAM,CAAE,eAAA,CAAgB,+BAAA,EAAiC;AAAA,MACnF,WAAA,EAAa;AAAA,KACd,CAAA;AAAA,EACH;AACA,EAAA,OAAO,yBAAA;AACT;AAoBO,SAAS,oBAAoB,IAAA,EAU3B;AACP,EAAA,MAAM,KAAA,GAAQ;AAAA,IACZ,mBAAmB,IAAA,CAAK,KAAA;AAAA,IACxB,qBAAqB,IAAA,CAAK,OAAA;AAAA,IAC1B,mBAAmB,IAAA,CAAK,KAAA;AAAA,IACxB,oBAAoB,IAAA,CAAK;AAAA,GAC3B;AACA,EAAA,oBAAA,EAAqB,CAAE,GAAA,CAAI,CAAA,EAAG,KAAK,CAAA;AACnC,EAAA,0BAAA,EAA2B,CAAE,MAAA,CAAO,IAAA,CAAK,UAAA,EAAY,KAAK,CAAA;AAC1D,EAAA,uBAAA,EAAwB,CAAE,MAAA,CAAO,IAAA,CAAK,KAAA,EAAO,KAAK,CAAA;AAClD,EAAA,IAAI,IAAA,CAAK,gBAAgB,MAAA,EAAW;AAClC,IAAA,wBAAA,EAAyB,CAAE,OAAO,IAAA,CAAK,WAAA,EAAa,EAAE,GAAG,KAAA,EAAO,qBAAA,EAAuB,OAAA,EAAS,CAAA;AAAA,EAClG;AACA,EAAA,IAAI,IAAA,CAAK,iBAAiB,MAAA,EAAW;AACnC,IAAA,wBAAA,EAAyB,CAAE,OAAO,IAAA,CAAK,YAAA,EAAc,EAAE,GAAG,KAAA,EAAO,qBAAA,EAAuB,QAAA,EAAU,CAAA;AAAA,EACpG;AACF;AAEA,IAAI,qBAAA;AACJ,SAAS,oBAAA,GAAgC;AACvC,EAAA,IAAI,CAAC,qBAAA,EAAuB;AAC1B,IAAA,qBAAA,GAAwB,KAAA,EAAM,CAAE,aAAA,CAAc,2BAAA,EAA6B;AAAA,MACzE,WAAA,EAAa;AAAA,KACd,CAAA;AAAA,EACH;AACA,EAAA,OAAO,qBAAA;AACT;AAkBO,SAAS,mBAAA,CAAoB,OAAe,MAAA,EAAsB;AACvE,EAAA,oBAAA,EAAqB,CAAE,IAAI,CAAA,EAAG,EAAE,mBAAmB,KAAA,EAAO,eAAA,EAAiB,QAAQ,CAAA;AACrF","file":"index.js","sourcesContent":["/**\n * @semiont/observability — public API.\n *\n * Universal surface (works in Node + browser). For SDK *initialization*,\n * import from `@semiont/observability/node` or `/web` at the process entry\n * point. Everything else uses this module.\n *\n * Tier 2 of `.plans/OBSERVABILITY.md`. The public surface:\n *\n * - `withSpan(name, fn, options?)` — wrap an async block in a span;\n * `options` carries `kind` and `attrs`.\n * - `withActorSpan(actor, channel, fn, extraAttrs?)` — consumer-span\n * wrapper for bus-event handlers, with handler-duration recording.\n * - `injectTraceparent(payload)` / `extractTraceparent(payload)` — W3C\n * trace-context propagation across the SSE channel (the bus payload\n * gets a `_trace?: { traceparent }` sibling to `correlationId`).\n * - `withTraceparent(carrier, fn)` — run `fn` with the incoming\n * traceparent as the parent context.\n * - `getActiveTraceparent()` — read the active span's traceparent for\n * manual propagation (e.g. attaching to a fetch header or SSE field).\n * - `getLogTraceContext()` — active `trace_id` / `span_id` for log-line\n * correlation.\n * - Metric recorders (`recordBusEmit`, `recordHandlerDuration`,\n * `recordJobOutcome`, `recordSubscriberConnect` / `Disconnect`,\n * `recordInferenceUsage`) and gauge providers\n * (`registerJobQueueProvider`, `registerVectorIndexSizeProvider`).\n *\n * No-op when no exporter is configured: `@opentelemetry/api`'s default\n * tracer is a no-op, so `withSpan` is essentially free until\n * `initObservability*()` runs.\n */\n\nimport {\n context,\n isSpanContextValid,\n metrics,\n propagation,\n SpanKind,\n SpanStatusCode,\n trace,\n type Attributes,\n type Counter,\n type Histogram,\n type ObservableGauge,\n type Span,\n type UpDownCounter,\n} from '@opentelemetry/api';\nimport { setBusLogTraceIdProvider } from '@semiont/core';\n\n// Wire `busLog`'s trace-id provider once at module load. When an OTel\n// SDK is initialized (and a span is active when `busLog` fires), the\n// emitted line gets a `trace=<8hex>` suffix that correlates the\n// grep-timeline with the trace UI. No-op when no SDK is active.\nsetBusLogTraceIdProvider(() => {\n const span = trace.getActiveSpan();\n if (!span) return undefined;\n const ctx = span.spanContext();\n if (!isSpanContextValid(ctx)) return undefined;\n return ctx.traceId;\n});\n\nconst TRACER_NAME = 'semiont';\n\nconst tracer = () => trace.getTracer(TRACER_NAME);\n\n// ── withSpan ───────────────────────────────────────────────────────────\n\n/**\n * Wrap an async block in a span. The span is started before `fn` runs and\n * ended after it resolves or rejects; exceptions are recorded and the span\n * status is set to ERROR. `kind` defaults to INTERNAL.\n */\nexport async function withSpan<T>(\n name: string,\n fn: (span: Span) => Promise<T> | T,\n options?: { kind?: SpanKind; attrs?: Attributes },\n): Promise<T> {\n const span = tracer().startSpan(name, {\n kind: options?.kind ?? SpanKind.INTERNAL,\n ...(options?.attrs ? { attributes: options.attrs } : {}),\n });\n try {\n return await context.with(trace.setSpan(context.active(), span), () => fn(span));\n } catch (err) {\n span.recordException(err as Error);\n span.setStatus({\n code: SpanStatusCode.ERROR,\n message: err instanceof Error ? err.message : String(err),\n });\n throw err;\n } finally {\n span.end();\n }\n}\n\n// ── Traceparent on bus payloads ────────────────────────────────────────\n\nconst TRACE_FIELD = '_trace';\n\n/**\n * Sibling of `correlationId` on bus payloads. Lives on the SSE event body\n * because SSE has no header trailer; the SDK strips it before delivering\n * the payload to subscribers. Additive — payloads without `_trace` parse\n * unchanged.\n */\nexport interface TraceCarrier {\n /** W3C `traceparent` header value (`00-<traceId>-<spanId>-<flags>`). */\n traceparent: string;\n /** W3C `tracestate` header value (vendor-specific extensions). */\n tracestate?: string;\n}\n\n/**\n * Read the active span's W3C traceparent (and tracestate). Returns\n * `undefined` if no span is active.\n */\nexport function getActiveTraceparent(): TraceCarrier | undefined {\n const carrier: Record<string, string> = {};\n propagation.inject(context.active(), carrier);\n const traceparent = carrier['traceparent'];\n if (!traceparent) return undefined;\n return carrier['tracestate']\n ? { traceparent, tracestate: carrier['tracestate'] }\n : { traceparent };\n}\n\n/**\n * Attach the active span's trace-context to a payload object as\n * `_trace`. No-op when no span is active. Returns the same object\n * reference for chaining.\n */\nexport function injectTraceparent<T extends Record<string, unknown>>(payload: T): T {\n const carrier = getActiveTraceparent();\n if (carrier) {\n (payload as Record<string, unknown>)[TRACE_FIELD] = carrier;\n }\n return payload;\n}\n\n/**\n * Strip and return the `_trace` field from a payload. Mutates `payload`.\n * The field is internal plumbing and should not be visible to subscribers.\n */\nexport function extractTraceparent<T extends Record<string, unknown>>(\n payload: T,\n): TraceCarrier | undefined {\n const carrier = (payload as Record<string, unknown>)[TRACE_FIELD] as\n | TraceCarrier\n | undefined;\n if (carrier !== undefined) {\n delete (payload as Record<string, unknown>)[TRACE_FIELD];\n }\n if (!carrier || typeof carrier.traceparent !== 'string') return undefined;\n return carrier;\n}\n\n/**\n * Run `fn` with the given W3C traceparent set as the parent context.\n * Any spans started inside `fn` will be children of the incoming trace.\n * No-op if `carrier` is undefined.\n */\nexport function withTraceparent<T>(\n carrier: TraceCarrier | undefined,\n fn: () => T,\n): T {\n if (!carrier) return fn();\n const carrierObj: Record<string, string> = { traceparent: carrier.traceparent };\n if (carrier.tracestate) carrierObj['tracestate'] = carrier.tracestate;\n const ctx = propagation.extract(context.active(), carrierObj);\n return context.with(ctx, fn);\n}\n\n// ── Actor handler convenience ──────────────────────────────────────────\n\n/**\n * Wrap a bus-event handler in an `actor.<name>:<channel>` consumer span.\n * Used at every `eventBus.get(channel).subscribe(handler)` site inside\n * an actor (Stower, Gatherer, Matcher, Browser, Smelter), to attribute\n * each in-process subscriber's work to a span without scattering manual\n * `withSpan` calls across handler bodies.\n *\n * The span's parent is the active context at the time the handler\n * fires — which is the `bus.dispatch:<channel>` span on the gateway\n * (Subject.next runs synchronously inside the dispatch span), or the\n * `bus.emit:<channel>` span when an actor emits to itself.\n */\nexport async function withActorSpan<T>(\n actor: string,\n channel: string,\n fn: (span: Span) => Promise<T> | T,\n extraAttrs?: Attributes,\n): Promise<T> {\n const start = performance.now();\n try {\n return await withSpan(`actor.${actor}:${channel}`, fn, {\n kind: SpanKind.CONSUMER,\n attrs: {\n actor,\n 'bus.channel': channel,\n ...(extraAttrs ?? {}),\n },\n });\n } finally {\n recordHandlerDuration(actor, channel, performance.now() - start);\n }\n}\n\n// ── Log correlation ────────────────────────────────────────────────────\n\n/**\n * Read the active span's `trace_id` / `span_id` for log-line correlation.\n * Tier 3 of `.plans/OBSERVABILITY.md`. Each structured log line gets\n * tagged with these so a log query in CloudWatch / Loki / Datadog can\n * jump to the trace in Tempo / Jaeger / X-Ray.\n *\n * Returns `undefined` if no span is active, or if the active span's\n * context is invalid (uninitialized SDK, no-op tracer).\n */\nexport function getLogTraceContext(): { trace_id: string; span_id: string } | undefined {\n const span = trace.getActiveSpan();\n if (!span) return undefined;\n const ctx = span.spanContext();\n if (!isSpanContextValid(ctx)) return undefined;\n return { trace_id: ctx.traceId, span_id: ctx.spanId };\n}\n\n// ── Metrics — Tier 3 ───────────────────────────────────────────────────\n\nconst METER_NAME = 'semiont';\n\nconst meter = () => metrics.getMeter(METER_NAME);\n\nlet _busEmitCounter: Counter | undefined;\nlet _replySuppressedCounter: Counter | undefined;\nlet _resumeGapCounter: Counter | undefined;\nlet _unanswerableCounter: Counter | undefined;\nlet _correlationRegistryGauge: ObservableGauge | undefined;\nlet _correlationRegistryProvider: (() => CorrelationRegistrySnapshot) | undefined;\nlet _handlerDurationHistogram: Histogram | undefined;\nlet _jobOutcomeCounter: Counter | undefined;\nlet _jobDurationHistogram: Histogram | undefined;\nlet _gatherDegradeCounter: Counter | undefined;\nlet _inferenceCallsCounter: Counter | undefined;\nlet _inferenceTokensCounter: Counter | undefined;\nlet _inferenceDurationHistogram: Histogram | undefined;\nlet _sseSubscribers: UpDownCounter | undefined;\nlet _jobQueueGauge: ObservableGauge | undefined;\nlet _jobQueueProvider: (() => Promise<JobQueueSnapshot> | JobQueueSnapshot) | undefined;\nlet _vectorIndexSizeGauge: ObservableGauge | undefined;\nlet _factPumpDepthGauge: ObservableGauge | undefined;\nlet _factPumpDepthProvider: (() => number) | undefined;\nlet _vectorIndexSizeProvider: (() => Promise<number> | number) | undefined;\n\n/** Snapshot of job-queue contents by status. Match `JobQueue.getStats()`. */\nexport interface JobQueueSnapshot {\n pending: number;\n running: number;\n complete: number;\n failed: number;\n cancelled: number;\n}\n\nfunction busEmitCounter(): Counter {\n if (!_busEmitCounter) {\n _busEmitCounter = meter().createCounter('semiont.bus.emit', {\n description: 'Bus emits by channel and scope',\n });\n }\n return _busEmitCounter;\n}\n\nfunction handlerDurationHistogram(): Histogram {\n if (!_handlerDurationHistogram) {\n _handlerDurationHistogram = meter().createHistogram('semiont.handler.duration', {\n description: 'In-process actor handler duration',\n unit: 'ms',\n });\n }\n return _handlerDurationHistogram;\n}\n\nfunction jobOutcomeCounter(): Counter {\n if (!_jobOutcomeCounter) {\n _jobOutcomeCounter = meter().createCounter('semiont.job.outcome', {\n description: 'Worker job completions by type and outcome',\n });\n }\n return _jobOutcomeCounter;\n}\n\nfunction jobDurationHistogram(): Histogram {\n if (!_jobDurationHistogram) {\n _jobDurationHistogram = meter().createHistogram('semiont.job.duration', {\n description: 'Worker job duration by type',\n unit: 'ms',\n });\n }\n return _jobDurationHistogram;\n}\n\nfunction inferenceCallsCounter(): Counter {\n if (!_inferenceCallsCounter) {\n _inferenceCallsCounter = meter().createCounter('semiont.inference.calls', {\n description: 'Inference API calls by provider, model, and outcome',\n });\n }\n return _inferenceCallsCounter;\n}\n\nfunction inferenceTokensCounter(): Counter {\n if (!_inferenceTokensCounter) {\n _inferenceTokensCounter = meter().createCounter('semiont.inference.tokens', {\n description: 'Inference token usage by provider, model, and direction',\n });\n }\n return _inferenceTokensCounter;\n}\n\nfunction inferenceDurationHistogram(): Histogram {\n if (!_inferenceDurationHistogram) {\n _inferenceDurationHistogram = meter().createHistogram('semiont.inference.duration', {\n description: 'Inference call duration by provider, model, and outcome',\n unit: 'ms',\n });\n }\n return _inferenceDurationHistogram;\n}\n\nfunction sseSubscribersCounter(): UpDownCounter {\n if (!_sseSubscribers) {\n _sseSubscribers = meter().createUpDownCounter('semiont.sse.subscribers', {\n description: 'Active SSE subscribers',\n });\n }\n return _sseSubscribers;\n}\n\nfunction replySuppressedCounter(): Counter {\n if (!_replySuppressedCounter) {\n _replySuppressedCounter = meter().createCounter('semiont.bus.reply.suppressed', {\n description: 'Correlated replies withheld from a non-owning subscriber',\n });\n }\n return _replySuppressedCounter;\n}\n\n/**\n * A correlated reply was withheld from a subscriber that does not own its\n * correlationId (CORRELATED-REPLY-ROUTING P5).\n *\n * Counts ONLY that case. A frame with no correlationId is a shape violation\n * (warned, not counted), and a cid nobody claimed is the structural in-process\n * case that fires constantly — counting either would drown the signal this\n * metric exists to show: the fan-out amplification the delivery filter removes.\n */\nexport function recordReplySuppressed(channel: string): void {\n replySuppressedCounter().add(1, { 'bus.channel': channel });\n}\n\nfunction resumeGapCounter(): Counter {\n if (!_resumeGapCounter) {\n _resumeGapCounter = meter().createCounter('semiont.bus.resume_gap', {\n description: 'SSE resumes that degraded to a gap because replay was unavailable',\n });\n }\n return _resumeGapCounter;\n}\n\n/**\n * An SSE resume could not be served and the client was told to fall back to\n * cache. This degradation is CORRECT by design and therefore silent — which is\n * exactly why it needs a number. A rising rate means clients are losing\n * history, and nothing else in the stack says so.\n */\nexport function recordResumeGap(reason: string): void {\n resumeGapCounter().add(1, { 'bus.resume_gap.reason': reason });\n}\n\nfunction unanswerableCounter(): Counter {\n if (!_unanswerableCounter) {\n _unanswerableCounter = meter().createCounter('semiont.bus.unanswerable', {\n description: 'Request emits that reached zero subscribers and were failed at the gateway',\n });\n }\n return _unanswerableCounter;\n}\n\n/**\n * A request-shaped emit reached no subscriber, so the gateway synthesized its\n * mapped failure (ARCHIVIST-STAYS-UP P3). By channel, this is the absence rate\n * of the service that answers it — the difference between \"it went down once\"\n * and \"it is flapping.\"\n */\nexport function recordUnanswerableRequest(channel: string): void {\n unanswerableCounter().add(1, { 'bus.channel': channel });\n}\n\n/** Claims held and reply payloads retained by a gateway's correlation registry. */\nexport interface CorrelationRegistrySnapshot {\n claims: number;\n retainedReplies: number;\n}\n\n/**\n * Register a callback returning the gateway's correlation-registry occupancy.\n *\n * COUNTS, not bytes: retention is count-budgeted today (byte-budgeting is a\n * known limit in CORRELATED-REPLY-ROUTING), so `retainedReplies` is a proxy for\n * heap, not a measure of it. It is still the closest observable to the question\n * two OOM investigations keep asking — a browse result can be 1-2 MB, and up to\n * REPLY_RETENTION_MAX of them are held at once.\n */\nexport function registerCorrelationRegistryProvider(\n provider: () => CorrelationRegistrySnapshot,\n): void {\n _correlationRegistryProvider = provider;\n if (!_correlationRegistryGauge) {\n _correlationRegistryGauge = meter().createObservableGauge('semiont.bus.correlation.size', {\n description: 'Correlation registry occupancy: live claims and retained reply payloads',\n });\n _correlationRegistryGauge.addCallback((observer) => {\n if (!_correlationRegistryProvider) return;\n const snap = _correlationRegistryProvider();\n observer.observe(snap.claims, { 'correlation.kind': 'claims' });\n observer.observe(snap.retainedReplies, { 'correlation.kind': 'retained_replies' });\n });\n }\n}\n\n/** Increment the bus-emit counter. Called at every transport `emit` site. */\nexport function recordBusEmit(channel: string, scope?: string): void {\n busEmitCounter().add(1, {\n 'bus.channel': channel,\n ...(scope ? { 'bus.scope': scope } : {}),\n });\n}\n\n/** Record an in-process actor handler's duration. */\nexport function recordHandlerDuration(actor: string, channel: string, durationMs: number): void {\n handlerDurationHistogram().record(durationMs, {\n actor,\n 'bus.channel': channel,\n });\n}\n\n/** Record a worker job's outcome and duration. */\nexport function recordJobOutcome(jobType: string, outcome: 'completed' | 'failed', durationMs: number): void {\n jobOutcomeCounter().add(1, { 'job.type': jobType, 'job.outcome': outcome });\n jobDurationHistogram().record(durationMs, { 'job.type': jobType, 'job.outcome': outcome });\n}\n\nlet _appendStageHistogram: Histogram | undefined;\nfunction appendStageHistogram(): Histogram {\n if (!_appendStageHistogram) {\n _appendStageHistogram = meter().createHistogram('semiont.record.append.duration', {\n description: 'Time spent in one stage of appending an event to the record, labeled by stage: persist (JSONL write + git), materialize (view rebuild), enrich, publish. The Archivist\\'s core write path.',\n unit: 'ms',\n });\n }\n return _appendStageHistogram;\n}\n\n/**\n * Record one stage of `EventStore.appendEvent` (ARCHIVIST-STAYS-UP P7).\n *\n * The append path is the one operation only the Archivist can perform, and it\n * was entirely dark: reads had `recordHandlerDuration` and the bus had its own\n * counters, while writes had nothing. Stage-labeled because the useful\n * question is never \"was the append slow\" but WHICH PART — and `materialize`\n * in particular does work proportional to a resource's annotation count, so it\n * degrades with history rather than with load.\n */\nexport function recordAppendStage(\n stage: 'persist' | 'materialize' | 'enrich' | 'publish',\n durationMs: number,\n): void {\n appendStageHistogram().record(durationMs, { 'record.stage': stage });\n}\n\nlet _gitCommandHistogram: Histogram | undefined;\nfunction gitCommandHistogram(): Histogram {\n if (!_gitCommandHistogram) {\n _gitCommandHistogram = meter().createHistogram('semiont.git.duration', {\n description: 'Wall time of a git subprocess. Async — this is latency, not event-loop blockage. Staging is deduped, so the `add` count is far below the number of appended events.',\n unit: 'ms',\n });\n }\n return _gitCommandHistogram;\n}\n\n/**\n * Record a git invocation. Read the `add` count against events appended: one\n * per event means deferred staging has stopped deduping.\n */\nexport function recordGitCommand(command: string, durationMs: number): void {\n gitCommandHistogram().record(durationMs, { 'git.command': command });\n}\n\nfunction gatherDegradeCounter(): Counter {\n if (!_gatherDegradeCounter) {\n _gatherDegradeCounter = meter().createCounter('semiont.gather.degraded', {\n description: 'Gathers that degraded because an eventually-consistent projection did not catch up within its read barrier (vectors: absent semanticContext; graph: projection-lag failure). Labeled by projection.',\n });\n }\n return _gatherDegradeCounter;\n}\n\n/**\n * Record a gather degraded by a projection read barrier: `'vectors'` — the\n * Smelter settle barrier timed out (semanticContext shipped absent);\n * `'graph'` — the Weaver applied barrier + poll floor exhausted (projection\n * lag surfaced as a distinct failure). Fleet-alertable counterpart of the\n * `[gather DEGRADED]` L4 breadcrumbs — a rising rate on either label means\n * that pipeline is not keeping up.\n */\nexport function recordGatherDegrade(projection: 'graph' | 'vectors'): void {\n gatherDegradeCounter().add(1, { projection });\n}\n\n/** Increment the SSE subscriber gauge — call on `/bus/subscribe` open. */\nexport function recordSubscriberConnect(): void {\n sseSubscribersCounter().add(1);\n}\n\n/** Decrement on disconnect. Pair with `recordSubscriberConnect`. */\nexport function recordSubscriberDisconnect(): void {\n sseSubscribersCounter().add(-1);\n}\n\n/**\n * Register a callback that returns the current job-queue snapshot.\n * Polled at the SDK's metric-collection interval. The single gauge\n * emits one observation per status (`pending`, `running`, …) tagged\n * with the `job.status` attribute. Idempotent — last registered\n * provider wins.\n */\nexport function registerJobQueueProvider(\n provider: () => Promise<JobQueueSnapshot> | JobQueueSnapshot,\n): void {\n _jobQueueProvider = provider;\n if (!_jobQueueGauge) {\n _jobQueueGauge = meter().createObservableGauge('semiont.job.queue.size', {\n description: 'Job queue size by status',\n });\n _jobQueueGauge.addCallback(async (observer) => {\n if (!_jobQueueProvider) return;\n const snap = await _jobQueueProvider();\n observer.observe(snap.pending, { 'job.status': 'pending' });\n observer.observe(snap.running, { 'job.status': 'running' });\n observer.observe(snap.complete, { 'job.status': 'complete' });\n observer.observe(snap.failed, { 'job.status': 'failed' });\n observer.observe(snap.cancelled, { 'job.status': 'cancelled' });\n });\n }\n}\n\n/**\n * Register a callback that returns the current vector-index size\n * (point count). Async to allow remote queries (Qdrant). Polled at\n * the metric-collection interval.\n */\n/**\n * Register the Archivist's fact-pump backlog — facts appended to the record\n * but not yet republished onto the bus.\n *\n * At rest this is zero. A value that climbs and does not come back means the\n * pump is outrunning its transport, which is the leading hypothesis for the\n * load-correlated heap growth in `bugs/absent-archivist-wedges-browse.md`\n * (ARCHIVIST-STAYS-UP P5). The backlog is deliberately unbounded today, so\n * this number is the only thing standing between \"the pump is behind\" and an\n * OOM whose cause is inferred from RSS after the fact.\n */\nexport function registerFactPumpDepthProvider(provider: () => number): void {\n _factPumpDepthProvider = provider;\n if (!_factPumpDepthGauge) {\n _factPumpDepthGauge = meter().createObservableGauge('semiont.archivist.fact_pump.depth', {\n description: 'Facts appended to the record but not yet published to the bus. Zero at rest; a rising floor means the pump is behind its transport.',\n });\n _factPumpDepthGauge.addCallback((observer) => {\n if (_factPumpDepthProvider) observer.observe(_factPumpDepthProvider());\n });\n }\n}\n\nlet _gitStagingFailureCounter: Counter | undefined;\n\n/**\n * A staging command that could not be run. Staging the index is a CONVENIENCE\n * — the event log is the system of record — so a failure here is degraded\n * service, never a reason to exit. But degraded must be VISIBLE: this counter\n * is what stops \"the index is quietly stale\" from being invisible.\n */\nexport function recordGitStagingFailure(reason: 'index-lock' | 'other'): void {\n if (!_gitStagingFailureCounter) {\n _gitStagingFailureCounter = meter().createCounter('semiont.git.staging.failures', {\n description: 'Staging commands abandoned after retries; the index may be stale',\n });\n }\n _gitStagingFailureCounter.add(1, { reason });\n}\n\n/**\n * Process lifetime telemetry (ARCHIVIST-GIT-STAGER-CRASH).\n *\n * A supervised process that dies and comes back is INVISIBLE in logs unless\n * someone greps for boot lines, and every request in flight when it died looks\n * to its caller like a hang. On 2026-09-08 that cost a multi-hour hunt through\n * search, qdrant, neo4j and the SSE transport for a crash loop that one metric\n * would have named immediately.\n *\n * `start_time` is the diagnostic, not uptime: a CHANGE in it is unambiguous\n * proof of a restart, and uptime is derivable from it.\n */\nconst PROCESS_START_TIME_SECONDS = Math.floor(Date.now() / 1000);\nlet _processStartTimeGauge: ObservableGauge | undefined;\nlet _restartCountGauge: ObservableGauge | undefined;\nlet _restartCountProvider: (() => Promise<number> | number) | undefined;\nlet _abnormalExitCounter: Counter | undefined;\n\n/** Register `semiont.process.start_time`. Called by `initObservability*`. */\nexport function registerProcessLifetimeMetrics(): void {\n if (_processStartTimeGauge) return;\n _processStartTimeGauge = meter().createObservableGauge('semiont.process.start_time', {\n description: 'Unix seconds at which this process started; a change means it restarted',\n unit: 's',\n });\n _processStartTimeGauge.addCallback((observer) => observer.observe(PROCESS_START_TIME_SECONDS));\n}\n\n/**\n * Supply a restart count. The Archivist's supervisor is POSIX shell and cannot\n * emit OTel, but it already keeps a durable event log on the state mount — so\n * the supervised child reads it and reports the count on the supervisor's\n * behalf.\n */\nexport function registerRestartCountProvider(\n provider: () => Promise<number> | number,\n): void {\n _restartCountProvider = provider;\n if (!_restartCountGauge) {\n _restartCountGauge = meter().createObservableGauge('semiont.process.restarts', {\n description: 'Times the supervisor has restarted this service',\n });\n _restartCountGauge.addCallback(async (observer) => {\n if (_restartCountProvider) observer.observe(await _restartCountProvider());\n });\n }\n}\n\n/**\n * Record that this process is dying abnormally, and mark the active span so a\n * trace shows a span that ENDED IN DEATH rather than one that simply never\n * ends. Never swallows: callers re-raise, so Node's own semantics are intact.\n */\nexport function recordAbnormalTermination(reason: string, detail?: string): void {\n if (!_abnormalExitCounter) {\n _abnormalExitCounter = meter().createCounter('semiont.process.abnormal_exit', {\n description: 'Process terminations that were not a clean shutdown',\n });\n }\n _abnormalExitCounter.add(1, { reason });\n const active = trace.getActiveSpan();\n if (active) {\n active.setStatus({ code: SpanStatusCode.ERROR, message: `${reason}: ${detail ?? ''}`.trim() });\n active.setAttribute('semiont.process.abnormal_exit', reason);\n active.end();\n }\n}\n\nexport function registerVectorIndexSizeProvider(\n provider: () => Promise<number> | number,\n): void {\n _vectorIndexSizeProvider = provider;\n if (!_vectorIndexSizeGauge) {\n _vectorIndexSizeGauge = meter().createObservableGauge('semiont.vector.index.size', {\n description: 'Vector store point count',\n });\n _vectorIndexSizeGauge.addCallback(async (observer) => {\n if (_vectorIndexSizeProvider) {\n const value = await _vectorIndexSizeProvider();\n observer.observe(value);\n }\n });\n }\n}\n\n/**\n * Record an inference call. Token counts are optional — providers that\n * don't expose them (or fail before generating) record only call count\n * and duration.\n */\nexport function recordInferenceUsage(opts: {\n provider: string;\n model: string;\n durationMs: number;\n outcome: 'success' | 'error';\n inputTokens?: number;\n outputTokens?: number;\n}): void {\n const baseAttrs = {\n 'inference.provider': opts.provider,\n 'inference.model': opts.model,\n 'inference.outcome': opts.outcome,\n };\n inferenceCallsCounter().add(1, baseAttrs);\n inferenceDurationHistogram().record(opts.durationMs, baseAttrs);\n if (opts.inputTokens != null && opts.inputTokens > 0) {\n inferenceTokensCounter().add(opts.inputTokens, {\n 'inference.provider': opts.provider,\n 'inference.model': opts.model,\n 'inference.direction': 'input',\n });\n }\n if (opts.outputTokens != null && opts.outputTokens > 0) {\n inferenceTokensCounter().add(opts.outputTokens, {\n 'inference.provider': opts.provider,\n 'inference.model': opts.model,\n 'inference.direction': 'output',\n });\n }\n}\n\nlet _detectionCallCounter: Counter | undefined;\nlet _detectionDurationHistogram: Histogram | undefined;\nlet _detectionItemsHistogram: Histogram | undefined;\nlet _detectionTokensHistogram: Histogram | undefined;\n\nfunction detectionCallCounter(): Counter {\n if (!_detectionCallCounter) {\n _detectionCallCounter = meter().createCounter('semiont.detection.calls', {\n description: 'Detection model calls, labeled by motivation, outcome, subdivision depth and whether this was the floor re-roll.',\n });\n }\n return _detectionCallCounter;\n}\n\nfunction detectionDurationHistogram(): Histogram {\n if (!_detectionDurationHistogram) {\n _detectionDurationHistogram = meter().createHistogram('semiont.detection.call.duration', {\n description: 'Wall time of one detection model call, including the attempts that failed and were retried smaller.',\n unit: 'ms',\n });\n }\n return _detectionDurationHistogram;\n}\n\nfunction detectionItemsHistogram(): Histogram {\n if (!_detectionItemsHistogram) {\n _detectionItemsHistogram = meter().createHistogram('semiont.detection.call.items', {\n description: 'Annotations returned by one detection call. Against the input size on the same record, this is yield.',\n });\n }\n return _detectionItemsHistogram;\n}\n\nfunction detectionTokensHistogram(): Histogram {\n if (!_detectionTokensHistogram) {\n _detectionTokensHistogram = meter().createHistogram('semiont.detection.call.tokens', {\n description: \"Provider-reported tokens for one detection call, by direction. Deliberately separate from semiont.inference.tokens: that series is the authoritative total but carries no subdivision depth, and 'what does a depth-2 call cost' is the question every sizing decision asks.\",\n });\n }\n return _detectionTokensHistogram;\n}\n\n/**\n * Record one detection model call (DETECTION-QUALITY-THROUGHPUT P1).\n *\n * The adapters already record provider/model/duration/tokens for every\n * inference call. What they cannot know is the detection shape around it:\n * which motivation asked, how big the piece was, how many annotations came\n * back, how deep subdivision had descended, and whether this was the floor\n * re-roll. Those are the facts that distinguish a healthy call from a\n * expensive descent, and without them a slow detection run is one\n * undifferentiated number.\n *\n * FAILED attempts are recorded too, and that is the point: the calls paid for\n * and thrown away during a descent are exactly the cost later phases exist to\n * avoid, so a record only of successes would hide the thing being optimized.\n *\n * Tokens are the PROVIDER's counts, passed through — never estimated. Absent\n * means the provider did not report them.\n */\nexport function recordDetectionCall(opts: {\n label: string;\n pieceChars: number;\n durationMs: number;\n items: number;\n depth: number;\n reroll: boolean;\n outcome: 'success' | 'truncated' | 'timeout' | 'collapsed' | 'error';\n inputTokens?: number;\n outputTokens?: number;\n}): void {\n const attrs = {\n 'detection.label': opts.label,\n 'detection.outcome': opts.outcome,\n 'detection.depth': opts.depth,\n 'detection.reroll': opts.reroll,\n };\n detectionCallCounter().add(1, attrs);\n detectionDurationHistogram().record(opts.durationMs, attrs);\n detectionItemsHistogram().record(opts.items, attrs);\n if (opts.inputTokens !== undefined) {\n detectionTokensHistogram().record(opts.inputTokens, { ...attrs, 'detection.direction': 'input' });\n }\n if (opts.outputTokens !== undefined) {\n detectionTokensHistogram().record(opts.outputTokens, { ...attrs, 'detection.direction': 'output' });\n }\n}\n\nlet _anchorOutcomeCounter: Counter | undefined;\nfunction anchorOutcomeCounter(): Counter {\n if (!_anchorOutcomeCounter) {\n _anchorOutcomeCounter = meter().createCounter('semiont.detection.anchors', {\n description: 'Every annotation anchoring, labeled by the method that resolved it. EVERY outcome is counted, not just the risky ones, because a bare count of degraded anchors has no denominator — the rate is the precision signal.',\n });\n }\n return _anchorOutcomeCounter;\n}\n\n/**\n * Record how one annotation got anchored (DETECTION-QUALITY-THROUGHPUT P5).\n *\n * The selector-vs-source check is already a WRITE-TIME INVARIANT — both\n * `buildTextAnnotation` and `buildPdfAnnotation` throw on a selector that does\n * not match its source — so mechanical correctness is guaranteed rather than\n * sampled, and auditing it would measure a constant.\n *\n * What is genuinely uncertain is which anchoring METHOD got there. An `exact`\n * the model quoted verbatim and that appears once is certain; one resolved by\n * `first-of-many` (several occurrences, no usable context) or `fuzzy-match`\n * picked a plausible occurrence and may have picked wrong. Those were visible\n * only as log warnings — countable by a human reading worker output, which is\n * how 47 of them went unreviewed. As a rate they are the precision number that\n * sits beside the yield numbers.\n */\nexport function recordAnchorOutcome(label: string, method: string): void {\n anchorOutcomeCounter().add(1, { 'detection.label': label, 'anchor.method': method });\n}\n\n// ── Re-exports from @opentelemetry/api ─────────────────────────────────\n\nexport { SpanKind, SpanStatusCode, type Attributes, type Span } from '@opentelemetry/api';\n"]}
package/dist/node.js CHANGED
@@ -1,6 +1,7 @@
1
+ import { trace, isSpanContextValid, context, propagation, metrics, SpanStatusCode } from '@opentelemetry/api';
2
+ import { setBusLogTraceIdProvider } from '@semiont/core';
1
3
  import { monitorEventLoopDelay } from 'perf_hooks';
2
4
  import { getHeapStatistics } from 'v8';
3
- import { context, trace, propagation, metrics } from '@opentelemetry/api';
4
5
  import { AsyncLocalStorageContextManager } from '@opentelemetry/context-async-hooks';
5
6
  import { W3CTraceContextPropagator } from '@opentelemetry/core';
6
7
  import { OTLPMetricExporter } from '@opentelemetry/exporter-metrics-otlp-http';
@@ -10,7 +11,41 @@ import { ConsoleMetricExporter, MeterProvider, PeriodicExportingMetricReader } f
10
11
  import { ConsoleSpanExporter, BasicTracerProvider, BatchSpanProcessor } from '@opentelemetry/sdk-trace-base';
11
12
  import { ATTR_SERVICE_VERSION, ATTR_SERVICE_NAME } from '@opentelemetry/semantic-conventions';
12
13
 
13
- // src/node.ts
14
+ // src/index.ts
15
+ setBusLogTraceIdProvider(() => {
16
+ const span = trace.getActiveSpan();
17
+ if (!span) return void 0;
18
+ const ctx = span.spanContext();
19
+ if (!isSpanContextValid(ctx)) return void 0;
20
+ return ctx.traceId;
21
+ });
22
+ var METER_NAME = "semiont";
23
+ var meter = () => metrics.getMeter(METER_NAME);
24
+ var PROCESS_START_TIME_SECONDS = Math.floor(Date.now() / 1e3);
25
+ var _processStartTimeGauge;
26
+ var _abnormalExitCounter;
27
+ function registerProcessLifetimeMetrics() {
28
+ if (_processStartTimeGauge) return;
29
+ _processStartTimeGauge = meter().createObservableGauge("semiont.process.start_time", {
30
+ description: "Unix seconds at which this process started; a change means it restarted",
31
+ unit: "s"
32
+ });
33
+ _processStartTimeGauge.addCallback((observer) => observer.observe(PROCESS_START_TIME_SECONDS));
34
+ }
35
+ function recordAbnormalTermination(reason, detail) {
36
+ if (!_abnormalExitCounter) {
37
+ _abnormalExitCounter = meter().createCounter("semiont.process.abnormal_exit", {
38
+ description: "Process terminations that were not a clean shutdown"
39
+ });
40
+ }
41
+ _abnormalExitCounter.add(1, { reason });
42
+ const active = trace.getActiveSpan();
43
+ if (active) {
44
+ active.setStatus({ code: SpanStatusCode.ERROR, message: `${reason}: ${detail ?? ""}`.trim() });
45
+ active.setAttribute("semiont.process.abnormal_exit", reason);
46
+ active.end();
47
+ }
48
+ }
14
49
  function heapStats() {
15
50
  const mem = process.memoryUsage();
16
51
  return {
@@ -93,6 +128,27 @@ function initObservabilityNode(config) {
93
128
  };
94
129
  process.once("SIGTERM", shutdown);
95
130
  process.once("SIGINT", shutdown);
131
+ registerProcessLifetimeMetrics();
132
+ const fatal = (reason) => (err) => {
133
+ const detail = err instanceof Error ? err.message : String(err);
134
+ try {
135
+ recordAbnormalTermination(reason, detail);
136
+ } catch {
137
+ }
138
+ const flushed = Promise.all([
139
+ tracerProviderInstance?.forceFlush().catch(() => {
140
+ }),
141
+ meterProviderInstance?.forceFlush().catch(() => {
142
+ })
143
+ ]);
144
+ const bounded = new Promise((resolve) => setTimeout(resolve, 2e3).unref?.());
145
+ void Promise.race([flushed, bounded]).finally(() => {
146
+ console.error(`[fatal] ${reason}: ${detail}`);
147
+ process.exit(1);
148
+ });
149
+ };
150
+ process.on("unhandledRejection", fatal("unhandledRejection"));
151
+ process.on("uncaughtException", fatal("uncaughtException"));
96
152
  return true;
97
153
  }
98
154
  async function shutdownObservabilityNode() {
package/dist/node.js.map CHANGED
@@ -1 +1 @@
1
- {"version":3,"sources":["../src/runtime-stats.ts","../src/node.ts"],"names":[],"mappings":";;;;;;;;;;;;;AA8BO,SAAS,SAAA,GAAuB;AACrC,EAAA,MAAM,GAAA,GAAM,QAAQ,WAAA,EAAY;AAChC,EAAA,OAAO;AAAA,IACL,UAAU,GAAA,CAAI,QAAA;AAAA,IACd,WAAW,GAAA,CAAI,SAAA;AAAA,IACf,SAAA,EAAW,mBAAkB,CAAE,eAAA;AAAA,IAC/B,KAAK,GAAA,CAAI;AAAA,GACX;AACF;ACmBA,IAAI,sBAAA;AACJ,IAAI,qBAAA;AAMJ,IAAM,iCAAA,GAAoC,GAAA;AASnC,SAAS,mBAAA,CACd,UACA,SAAA,EACoB;AACpB,EAAA,IAAI,SAAA,KAAc,WAAW,OAAO,SAAA;AACpC,EAAA,OAAO,WAAW,MAAA,GAAS,SAAA;AAC7B;AAWO,SAAS,sBAAsB,MAAA,EAA0C;AAC9E,EAAA,IAAI,wBAAwB,OAAO,KAAA;AACnC,EAAA,IAAI,OAAA,CAAQ,GAAA,CAAI,mBAAmB,CAAA,KAAM,QAAQ,OAAO,KAAA;AAExD,EAAA,MAAM,QAAA,GAAW,OAAA,CAAQ,GAAA,CAAI,6BAA6B,CAAA;AAC1D,EAAA,MAAM,UAAA,GAAa,OAAA,CAAQ,GAAA,CAAI,uBAAuB,CAAA,KAAM,MAAA;AAM5D,EAAA,IAAI,CAAC,QAAA,IAAY,CAAC,UAAA,EAAY,OAAO,KAAA;AAErC,EAAA,MAAM,WAAW,sBAAA,CAAuB;AAAA,IACtC,CAAC,iBAAiB,GAAG,QAAQ,GAAA,CAAI,mBAAmB,KAAK,MAAA,CAAO,WAAA;AAAA,IAChE,CAAC,oBAAoB,GAAG,MAAA,CAAO,cAAA,IAAkB;AAAA,GAClD,CAAA;AAGD,EAAA,MAAM,gBAAgB,QAAA,GAAW,IAAI,iBAAA,EAAkB,GAAI,IAAI,mBAAA,EAAoB;AACnF,EAAA,sBAAA,GAAyB,IAAI,mBAAA,CAAoB;AAAA,IAC/C,QAAA;AAAA,IACA,cAAA,EAAgB,CAAC,IAAI,kBAAA,CAAmB,aAAa,CAAC;AAAA,GACvD,CAAA;AAKD,EAAA,OAAA,CAAQ,uBAAA,CAAwB,IAAI,+BAAA,EAAgC,CAAE,QAAQ,CAAA;AAC9E,EAAA,KAAA,CAAM,wBAAwB,sBAAsB,CAAA;AAQpD,EAAA,WAAA,CAAY,mBAAA,CAAoB,IAAI,yBAAA,EAA2B,CAAA;AAG/D,EAAA,MAAM,cAAA,GACJ,mBAAA,CAAoB,QAAA,EAAU,OAAA,CAAQ,GAAA,CAAI,uBAAuB,CAAC,CAAA,KAAM,SAAA,GACpE,IAAI,qBAAA,EAAsB,GAC1B,IAAI,kBAAA,EAAmB;AAC7B,EAAA,MAAM,WAAA,GAAc,OAAA,CAAQ,GAAA,CAAI,6BAA6B,CAAA;AAC7D,EAAA,MAAM,uBAAuB,WAAA,GACzB,MAAA,CAAO,QAAA,CAAS,WAAA,EAAa,EAAE,CAAA,GAC/B,iCAAA;AAEJ,EAAA,qBAAA,GAAwB,IAAI,aAAA,CAAc;AAAA,IACxC,QAAA;AAAA,IACA,OAAA,EAAS;AAAA,MACP,IAAI,6BAAA,CAA8B;AAAA,QAChC,QAAA,EAAU,cAAA;AAAA,QACV,sBAAsB,MAAA,CAAO,QAAA,CAAS,oBAAoB,CAAA,IAAK,oBAAA,GAAuB,IAClF,oBAAA,GACA;AAAA,OACL;AAAA;AACH,GACD,CAAA;AACD,EAAA,OAAA,CAAQ,uBAAuB,qBAAqB,CAAA;AAUpD,EAAA,MAAM,SAAA,GAAY,qBAAA,CAAsB,EAAE,UAAA,EAAY,IAAI,CAAA;AAC1D,EAAA,SAAA,CAAU,MAAA,EAAO;AACjB,EAAA,MAAM,YAAA,GAAe,qBAAA,CAAsB,QAAA,CAAS,iBAAiB,CAAA;AACrE,EAAA,YAAA,CACG,sBAAsB,gCAAA,EAAkC;AAAA,IACvD,WAAA,EAAa,wGAAA;AAAA,IACb,IAAA,EAAM;AAAA,GACP,CAAA,CACA,WAAA,CAAY,CAAC,QAAA,KAAa;AACzB,IAAA,QAAA,CAAS,QAAQ,SAAA,CAAU,IAAA,GAAO,KAAK,EAAE,UAAA,EAAY,QAAQ,CAAA;AAC7D,IAAA,QAAA,CAAS,OAAA,CAAQ,UAAU,UAAA,CAAW,EAAE,IAAI,GAAA,EAAK,EAAE,UAAA,EAAY,KAAA,EAAO,CAAA;AACtE,IAAA,QAAA,CAAS,QAAQ,SAAA,CAAU,GAAA,GAAM,KAAK,EAAE,UAAA,EAAY,OAAO,CAAA;AAC3D,IAAA,SAAA,CAAU,KAAA,EAAM;AAAA,EAClB,CAAC,CAAA;AAWH,EAAA,YAAA,CACG,sBAAsB,sBAAA,EAAwB;AAAA,IAC7C,WAAA,EAAa,+HAAA;AAAA,IACb,IAAA,EAAM;AAAA,GACP,CAAA,CACA,WAAA,CAAY,CAAC,QAAA,KAAa;AACzB,IAAA,MAAM,IAAI,SAAA,EAAU;AACpB,IAAA,QAAA,CAAS,QAAQ,CAAA,CAAE,QAAA,EAAU,EAAE,WAAA,EAAa,QAAQ,CAAA;AACpD,IAAA,QAAA,CAAS,QAAQ,CAAA,CAAE,SAAA,EAAW,EAAE,WAAA,EAAa,SAAS,CAAA;AACtD,IAAA,QAAA,CAAS,QAAQ,CAAA,CAAE,SAAA,EAAW,EAAE,WAAA,EAAa,SAAS,CAAA;AACtD,IAAA,QAAA,CAAS,QAAQ,CAAA,CAAE,GAAA,EAAK,EAAE,WAAA,EAAa,OAAO,CAAA;AAAA,EAChD,CAAC,CAAA;AAGH,EAAA,MAAM,WAAW,MAAM;AACrB,IAAA,OAAA,CAAQ,GAAA,CAAI;AAAA,MACV,sBAAA,EAAwB,QAAA,EAAS,CAAE,KAAA,CAAM,MAAM;AAAA,MAAC,CAAC,CAAA;AAAA,MACjD,qBAAA,EAAuB,QAAA,EAAS,CAAE,KAAA,CAAM,MAAM;AAAA,MAAC,CAAC;AAAA,KACjD,CAAA,CAAE,OAAA,CAAQ,MAAM;AACf,MAAA,sBAAA,GAAyB,MAAA;AACzB,MAAA,qBAAA,GAAwB,MAAA;AAAA,IAC1B,CAAC,CAAA;AAAA,EACH,CAAA;AACA,EAAA,OAAA,CAAQ,IAAA,CAAK,WAAW,QAAQ,CAAA;AAChC,EAAA,OAAA,CAAQ,IAAA,CAAK,UAAU,QAAQ,CAAA;AAE/B,EAAA,OAAO,IAAA;AACT;AAGA,eAAsB,yBAAA,GAA2C;AAC/D,EAAA,MAAM,QAAQ,GAAA,CAAI;AAAA,IAChB,wBAAwB,QAAA,EAAS;AAAA,IACjC,uBAAuB,QAAA;AAAS,GACjC,CAAA;AACD,EAAA,sBAAA,GAAyB,MAAA;AACzB,EAAA,qBAAA,GAAwB,MAAA;AAC1B","file":"node.js","sourcesContent":["/**\n * Process-runtime readings (ARCHIVIST-STAYS-UP P4).\n *\n * Node-only, and deliberately a plain function rather than a gauge callback:\n * the numbers are the thing worth testing, and a callback registered inside\n * an SDK is awkward to assert against.\n *\n * **Why `heapLimit` is the field that matters.** The Archivist died at\n * ~1016 MB inside a 2048 MB container (`bugs/absent-archivist-wedges-browse.md`)\n * — not because it exhausted the container, but because it hit V8's OWN\n * default old-space ceiling, which is derived from visible memory and lands\n * well under it. `heapUsed` alone cannot express \"how close to death is\n * this\"; only the pair can. It is also how a configured\n * `--max-old-space-size` is verified to have taken effect, rather than\n * assumed from the fact that someone set an env var.\n */\n\nimport { getHeapStatistics } from 'node:v8';\n\nexport interface HeapStats {\n /** Live heap in use. */\n heapUsed: number;\n /** Heap V8 has currently reserved. */\n heapTotal: number;\n /** The ceiling V8 will die at — NOT the container's limit. */\n heapLimit: number;\n /** Resident set: everything, including buffers outside the JS heap. */\n rss: number;\n}\n\nexport function heapStats(): HeapStats {\n const mem = process.memoryUsage();\n return {\n heapUsed: mem.heapUsed,\n heapTotal: mem.heapTotal,\n heapLimit: getHeapStatistics().heap_size_limit,\n rss: mem.rss,\n };\n}\n","/**\n * Node SDK initialization. Call once at the process entry point\n * (gateway `index.ts`, `worker-main.ts`, `smelter-main.ts`).\n *\n * Configuration is via standard `OTEL_*` env vars:\n * - `OTEL_SERVICE_NAME` — service identity (e.g. `semiont-gateway`)\n * - `OTEL_EXPORTER_OTLP_ENDPOINT` — collector endpoint (HTTP)\n * - `OTEL_TRACES_SAMPLER` — sampler (default: `parentbased_always_on`)\n * - `OTEL_TRACES_SAMPLER_ARG` — sampler ratio (default: `1.0`)\n * - `OTEL_CONSOLE_EXPORTER=true` — dev-only: emit spans + metrics to stderr\n * - `OTEL_SDK_DISABLED=true` — skip initialization entirely\n *\n * **Off-by-default invariant**: with neither `OTEL_EXPORTER_OTLP_ENDPOINT`\n * nor `OTEL_CONSOLE_EXPORTER=true` set, this function is a no-op and the\n * `@opentelemetry/api` no-op tracer takes over. This avoids accidentally\n * flooding production stderr (and CloudWatch) when an operator deploys\n * without configuring an exporter.\n *\n * Implementation note: this module wires `BasicTracerProvider` and\n * `MeterProvider` (both from the stable `@opentelemetry/sdk-trace-base`\n * 2.x line) directly, plus `AsyncLocalStorageContextManager` for Node\n * async-context propagation. We deliberately avoid `@opentelemetry/sdk-node`\n * because its `0.x` experimental versions cross-depend on older 2.0.x\n * SDKs, forcing npm to nest duplicate copies of the stable packages and\n * blowing up bundles for every consumer.\n */\n\nimport { monitorEventLoopDelay } from 'node:perf_hooks';\nimport { heapStats } from './runtime-stats';\nimport { context, metrics, propagation, trace } from '@opentelemetry/api';\nimport { AsyncLocalStorageContextManager } from '@opentelemetry/context-async-hooks';\nimport { W3CTraceContextPropagator } from '@opentelemetry/core';\nimport { OTLPMetricExporter } from '@opentelemetry/exporter-metrics-otlp-http';\nimport { OTLPTraceExporter } from '@opentelemetry/exporter-trace-otlp-http';\nimport { resourceFromAttributes } from '@opentelemetry/resources';\nimport {\n ConsoleMetricExporter,\n MeterProvider,\n PeriodicExportingMetricReader,\n} from '@opentelemetry/sdk-metrics';\nimport {\n BasicTracerProvider,\n BatchSpanProcessor,\n ConsoleSpanExporter,\n} from '@opentelemetry/sdk-trace-base';\nimport {\n ATTR_SERVICE_NAME,\n ATTR_SERVICE_VERSION,\n} from '@opentelemetry/semantic-conventions';\n\nexport interface NodeObservabilityConfig {\n /** Service identity (e.g. `semiont-gateway`). Overridden by `OTEL_SERVICE_NAME`. */\n serviceName: string;\n /** Service version. Defaults to `0.0.0` if omitted. */\n serviceVersion?: string;\n}\n\nlet tracerProviderInstance: BasicTracerProvider | undefined;\nlet meterProviderInstance: MeterProvider | undefined;\n\n/**\n * Default metric export interval. 30s mirrors the SDK default and gives\n * operators enough granularity without flooding the collector.\n */\nconst DEFAULT_METRIC_EXPORT_INTERVAL_MS = 30_000;\n\n/**\n * Which exporter metrics use. `console` wins even with an OTLP endpoint set —\n * the readout for a bare process or CI with no collector; traces unaffected.\n * Any other value (incl. unrecognised) leaves the endpoint to decide, so an\n * unknown value cannot silently disable metrics. Exported and pure so tests\n * assert the choice without mocking the exporter modules.\n */\nexport function metricsExporterKind(\n endpoint: string | undefined,\n requested: string | undefined,\n): 'otlp' | 'console' {\n if (requested === 'console') return 'console';\n return endpoint ? 'otlp' : 'console';\n}\n\n/**\n * Initialize OTel for the current process. Wires up both tracing and\n * metrics. Idempotent — calling twice is a no-op. Returns `true` if the\n * SDK started, `false` if disabled, no exporter is configured, or\n * already initialized.\n *\n * Metrics export at `OTEL_METRIC_EXPORT_INTERVAL` ms (default 30s) to\n * the same `OTEL_EXPORTER_OTLP_ENDPOINT` as traces.\n */\nexport function initObservabilityNode(config: NodeObservabilityConfig): boolean {\n if (tracerProviderInstance) return false;\n if (process.env['OTEL_SDK_DISABLED'] === 'true') return false;\n\n const endpoint = process.env['OTEL_EXPORTER_OTLP_ENDPOINT'];\n const useConsole = process.env['OTEL_CONSOLE_EXPORTER'] === 'true';\n\n // No exporter configured = no SDK init. The `@opentelemetry/api`\n // no-op tracer takes over; `withSpan` still runs `fn` but emits\n // nothing, `getActiveSpan()` returns a sentinel. Avoids flooding\n // production stderr when no collector endpoint is set.\n if (!endpoint && !useConsole) return false;\n\n const resource = resourceFromAttributes({\n [ATTR_SERVICE_NAME]: process.env['OTEL_SERVICE_NAME'] ?? config.serviceName,\n [ATTR_SERVICE_VERSION]: config.serviceVersion ?? '0.0.0',\n });\n\n // Trace SDK\n const traceExporter = endpoint ? new OTLPTraceExporter() : new ConsoleSpanExporter();\n tracerProviderInstance = new BasicTracerProvider({\n resource,\n spanProcessors: [new BatchSpanProcessor(traceExporter)],\n });\n\n // Async-context propagation across `await` boundaries — equivalent\n // to what NodeSDK installs internally, but pinned to the stable 2.x\n // cohort.\n context.setGlobalContextManager(new AsyncLocalStorageContextManager().enable());\n trace.setGlobalTracerProvider(tracerProviderInstance);\n\n // W3C trace-context propagator. Without this, `propagation.inject`\n // and `propagation.extract` walk an empty propagator chain and\n // silently do nothing — which means traceparent never makes it onto\n // outgoing HTTP headers or SSE `_trace` payloads, and cross-service\n // traces stay disconnected. NodeSDK registers this for you; bare\n // BasicTracerProvider does not.\n propagation.setGlobalPropagator(new W3CTraceContextPropagator());\n\n // Traces follow the endpoint; metrics additionally honour OTEL_METRICS_EXPORTER.\n const metricExporter =\n metricsExporterKind(endpoint, process.env['OTEL_METRICS_EXPORTER']) === 'console'\n ? new ConsoleMetricExporter()\n : new OTLPMetricExporter();\n const intervalRaw = process.env['OTEL_METRIC_EXPORT_INTERVAL'];\n const exportIntervalMillis = intervalRaw\n ? Number.parseInt(intervalRaw, 10)\n : DEFAULT_METRIC_EXPORT_INTERVAL_MS;\n\n meterProviderInstance = new MeterProvider({\n resource,\n readers: [\n new PeriodicExportingMetricReader({\n exporter: metricExporter,\n exportIntervalMillis: Number.isFinite(exportIntervalMillis) && exportIntervalMillis > 0\n ? exportIntervalMillis\n : DEFAULT_METRIC_EXPORT_INTERVAL_MS,\n }),\n ],\n });\n metrics.setGlobalMeterProvider(meterProviderInstance);\n\n // Event-loop lag (ARCHIVIST-STAYS-UP P7). The cause-AGNOSTIC detector for a\n // blocked process: it rises whether the cause is a synchronous git\n // subprocess, a large JSON parse, or GC. Cheap — libuv keeps the histogram;\n // we read percentiles and reset once per export interval.\n //\n // It earns its place because the Archivist runs `execFileSync('git', …)` on\n // this loop once per appended event: while that blocks, every concurrent\n // `browse:*` read waits, and nothing else in the stack says so.\n const loopDelay = monitorEventLoopDelay({ resolution: 10 });\n loopDelay.enable();\n const runtimeMeter = meterProviderInstance.getMeter('semiont-runtime');\n runtimeMeter\n .createObservableGauge('semiont.runtime.event_loop.lag', {\n description: 'Event-loop delay percentiles over the last export interval. Time the process could not serve anything.',\n unit: 'ms',\n })\n .addCallback((observer) => {\n observer.observe(loopDelay.mean / 1e6, { 'lag.stat': 'mean' });\n observer.observe(loopDelay.percentile(99) / 1e6, { 'lag.stat': 'p99' });\n observer.observe(loopDelay.max / 1e6, { 'lag.stat': 'max' });\n loopDelay.reset();\n });\n\n // Heap (ARCHIVIST-STAYS-UP P4), on the SAME registration as lag rather than\n // a second mechanism — they are read together when diagnosing a process\n // that stopped answering, and splitting them would mean two things to wire.\n //\n // `limit` is the field that repays the effort: the Archivist died at\n // ~1016 MB inside a 2048 MB container because V8's own default ceiling sits\n // well under the container's. `used` alone cannot say how close to death a\n // process is, and `limit` is what a configured --max-old-space-size changes\n // — so this is also how that setting is VERIFIED rather than assumed.\n runtimeMeter\n .createObservableGauge('semiont.runtime.heap', {\n description: \"Process memory by kind. `limit` is V8's own ceiling, which is what the process dies at — not the container's allocation.\",\n unit: 'By',\n })\n .addCallback((observer) => {\n const s = heapStats();\n observer.observe(s.heapUsed, { 'heap.stat': 'used' });\n observer.observe(s.heapTotal, { 'heap.stat': 'total' });\n observer.observe(s.heapLimit, { 'heap.stat': 'limit' });\n observer.observe(s.rss, { 'heap.stat': 'rss' });\n });\n\n // Flush traces + metrics on shutdown so nothing is lost on SIGTERM/SIGINT.\n const shutdown = () => {\n Promise.all([\n tracerProviderInstance?.shutdown().catch(() => {}),\n meterProviderInstance?.shutdown().catch(() => {}),\n ]).finally(() => {\n tracerProviderInstance = undefined;\n meterProviderInstance = undefined;\n });\n };\n process.once('SIGTERM', shutdown);\n process.once('SIGINT', shutdown);\n\n return true;\n}\n\n/** Force-flush + shutdown both SDKs. Test cleanup, not production. */\nexport async function shutdownObservabilityNode(): Promise<void> {\n await Promise.all([\n tracerProviderInstance?.shutdown(),\n meterProviderInstance?.shutdown(),\n ]);\n tracerProviderInstance = undefined;\n meterProviderInstance = undefined;\n}\n"]}
1
+ {"version":3,"sources":["../src/index.ts","../src/runtime-stats.ts","../src/node.ts"],"names":["context","trace","propagation","metrics"],"mappings":";;;;;;;;;;;;;;AAqDA,wBAAA,CAAyB,MAAM;AAC7B,EAAA,MAAM,IAAA,GAAO,MAAM,aAAA,EAAc;AACjC,EAAA,IAAI,CAAC,MAAM,OAAO,MAAA;AAClB,EAAA,MAAM,GAAA,GAAM,KAAK,WAAA,EAAY;AAC7B,EAAA,IAAI,CAAC,kBAAA,CAAmB,GAAG,CAAA,EAAG,OAAO,MAAA;AACrC,EAAA,OAAO,GAAA,CAAI,OAAA;AACb,CAAC,CAAA;AAyKD,IAAM,UAAA,GAAa,SAAA;AAEnB,IAAM,KAAA,GAAQ,MAAM,OAAA,CAAQ,QAAA,CAAS,UAAU,CAAA;AA+X/C,IAAM,6BAA6B,IAAA,CAAK,KAAA,CAAM,IAAA,CAAK,GAAA,KAAQ,GAAI,CAAA;AAC/D,IAAI,sBAAA;AAGJ,IAAI,oBAAA;AAGG,SAAS,8BAAA,GAAuC;AACrD,EAAA,IAAI,sBAAA,EAAwB;AAC5B,EAAA,sBAAA,GAAyB,KAAA,EAAM,CAAE,qBAAA,CAAsB,4BAAA,EAA8B;AAAA,IACnF,WAAA,EAAa,yEAAA;AAAA,IACb,IAAA,EAAM;AAAA,GACP,CAAA;AACD,EAAA,sBAAA,CAAuB,YAAY,CAAC,QAAA,KAAa,QAAA,CAAS,OAAA,CAAQ,0BAA0B,CAAC,CAAA;AAC/F;AA2BO,SAAS,yBAAA,CAA0B,QAAgB,MAAA,EAAuB;AAC/E,EAAA,IAAI,CAAC,oBAAA,EAAsB;AACzB,IAAA,oBAAA,GAAuB,KAAA,EAAM,CAAE,aAAA,CAAc,+BAAA,EAAiC;AAAA,MAC5E,WAAA,EAAa;AAAA,KACd,CAAA;AAAA,EACH;AACA,EAAA,oBAAA,CAAqB,GAAA,CAAI,CAAA,EAAG,EAAE,MAAA,EAAQ,CAAA;AACtC,EAAA,MAAM,MAAA,GAAS,MAAM,aAAA,EAAc;AACnC,EAAA,IAAI,MAAA,EAAQ;AACV,IAAA,MAAA,CAAO,SAAA,CAAU,EAAE,IAAA,EAAM,cAAA,CAAe,OAAO,OAAA,EAAS,CAAA,EAAG,MAAM,CAAA,EAAA,EAAK,MAAA,IAAU,EAAE,CAAA,CAAA,CAAG,IAAA,IAAQ,CAAA;AAC7F,IAAA,MAAA,CAAO,YAAA,CAAa,iCAAiC,MAAM,CAAA;AAC3D,IAAA,MAAA,CAAO,GAAA,EAAI;AAAA,EACb;AACF;AC7nBO,SAAS,SAAA,GAAuB;AACrC,EAAA,MAAM,GAAA,GAAM,QAAQ,WAAA,EAAY;AAChC,EAAA,OAAO;AAAA,IACL,UAAU,GAAA,CAAI,QAAA;AAAA,IACd,WAAW,GAAA,CAAI,SAAA;AAAA,IACf,SAAA,EAAW,mBAAkB,CAAE,eAAA;AAAA,IAC/B,KAAK,GAAA,CAAI;AAAA,GACX;AACF;ACoBA,IAAI,sBAAA;AACJ,IAAI,qBAAA;AAMJ,IAAM,iCAAA,GAAoC,GAAA;AASnC,SAAS,mBAAA,CACd,UACA,SAAA,EACoB;AACpB,EAAA,IAAI,SAAA,KAAc,WAAW,OAAO,SAAA;AACpC,EAAA,OAAO,WAAW,MAAA,GAAS,SAAA;AAC7B;AAWO,SAAS,sBAAsB,MAAA,EAA0C;AAC9E,EAAA,IAAI,wBAAwB,OAAO,KAAA;AACnC,EAAA,IAAI,OAAA,CAAQ,GAAA,CAAI,mBAAmB,CAAA,KAAM,QAAQ,OAAO,KAAA;AAExD,EAAA,MAAM,QAAA,GAAW,OAAA,CAAQ,GAAA,CAAI,6BAA6B,CAAA;AAC1D,EAAA,MAAM,UAAA,GAAa,OAAA,CAAQ,GAAA,CAAI,uBAAuB,CAAA,KAAM,MAAA;AAM5D,EAAA,IAAI,CAAC,QAAA,IAAY,CAAC,UAAA,EAAY,OAAO,KAAA;AAErC,EAAA,MAAM,WAAW,sBAAA,CAAuB;AAAA,IACtC,CAAC,iBAAiB,GAAG,QAAQ,GAAA,CAAI,mBAAmB,KAAK,MAAA,CAAO,WAAA;AAAA,IAChE,CAAC,oBAAoB,GAAG,MAAA,CAAO,cAAA,IAAkB;AAAA,GAClD,CAAA;AAGD,EAAA,MAAM,gBAAgB,QAAA,GAAW,IAAI,iBAAA,EAAkB,GAAI,IAAI,mBAAA,EAAoB;AACnF,EAAA,sBAAA,GAAyB,IAAI,mBAAA,CAAoB;AAAA,IAC/C,QAAA;AAAA,IACA,cAAA,EAAgB,CAAC,IAAI,kBAAA,CAAmB,aAAa,CAAC;AAAA,GACvD,CAAA;AAKD,EAAAA,QAAQ,uBAAA,CAAwB,IAAI,+BAAA,EAAgC,CAAE,QAAQ,CAAA;AAC9E,EAAAC,KAAAA,CAAM,wBAAwB,sBAAsB,CAAA;AAQpD,EAAAC,WAAAA,CAAY,mBAAA,CAAoB,IAAI,yBAAA,EAA2B,CAAA;AAG/D,EAAA,MAAM,cAAA,GACJ,mBAAA,CAAoB,QAAA,EAAU,OAAA,CAAQ,GAAA,CAAI,uBAAuB,CAAC,CAAA,KAAM,SAAA,GACpE,IAAI,qBAAA,EAAsB,GAC1B,IAAI,kBAAA,EAAmB;AAC7B,EAAA,MAAM,WAAA,GAAc,OAAA,CAAQ,GAAA,CAAI,6BAA6B,CAAA;AAC7D,EAAA,MAAM,uBAAuB,WAAA,GACzB,MAAA,CAAO,QAAA,CAAS,WAAA,EAAa,EAAE,CAAA,GAC/B,iCAAA;AAEJ,EAAA,qBAAA,GAAwB,IAAI,aAAA,CAAc;AAAA,IACxC,QAAA;AAAA,IACA,OAAA,EAAS;AAAA,MACP,IAAI,6BAAA,CAA8B;AAAA,QAChC,QAAA,EAAU,cAAA;AAAA,QACV,sBAAsB,MAAA,CAAO,QAAA,CAAS,oBAAoB,CAAA,IAAK,oBAAA,GAAuB,IAClF,oBAAA,GACA;AAAA,OACL;AAAA;AACH,GACD,CAAA;AACD,EAAAC,OAAAA,CAAQ,uBAAuB,qBAAqB,CAAA;AAKpD,EAAA,MAAM,SAAA,GAAY,qBAAA,CAAsB,EAAE,UAAA,EAAY,IAAI,CAAA;AAC1D,EAAA,SAAA,CAAU,MAAA,EAAO;AACjB,EAAA,MAAM,YAAA,GAAe,qBAAA,CAAsB,QAAA,CAAS,iBAAiB,CAAA;AACrE,EAAA,YAAA,CACG,sBAAsB,gCAAA,EAAkC;AAAA,IACvD,WAAA,EAAa,wGAAA;AAAA,IACb,IAAA,EAAM;AAAA,GACP,CAAA,CACA,WAAA,CAAY,CAAC,QAAA,KAAa;AACzB,IAAA,QAAA,CAAS,QAAQ,SAAA,CAAU,IAAA,GAAO,KAAK,EAAE,UAAA,EAAY,QAAQ,CAAA;AAC7D,IAAA,QAAA,CAAS,OAAA,CAAQ,UAAU,UAAA,CAAW,EAAE,IAAI,GAAA,EAAK,EAAE,UAAA,EAAY,KAAA,EAAO,CAAA;AACtE,IAAA,QAAA,CAAS,QAAQ,SAAA,CAAU,GAAA,GAAM,KAAK,EAAE,UAAA,EAAY,OAAO,CAAA;AAC3D,IAAA,SAAA,CAAU,KAAA,EAAM;AAAA,EAClB,CAAC,CAAA;AAWH,EAAA,YAAA,CACG,sBAAsB,sBAAA,EAAwB;AAAA,IAC7C,WAAA,EAAa,+HAAA;AAAA,IACb,IAAA,EAAM;AAAA,GACP,CAAA,CACA,WAAA,CAAY,CAAC,QAAA,KAAa;AACzB,IAAA,MAAM,IAAI,SAAA,EAAU;AACpB,IAAA,QAAA,CAAS,QAAQ,CAAA,CAAE,QAAA,EAAU,EAAE,WAAA,EAAa,QAAQ,CAAA;AACpD,IAAA,QAAA,CAAS,QAAQ,CAAA,CAAE,SAAA,EAAW,EAAE,WAAA,EAAa,SAAS,CAAA;AACtD,IAAA,QAAA,CAAS,QAAQ,CAAA,CAAE,SAAA,EAAW,EAAE,WAAA,EAAa,SAAS,CAAA;AACtD,IAAA,QAAA,CAAS,QAAQ,CAAA,CAAE,GAAA,EAAK,EAAE,WAAA,EAAa,OAAO,CAAA;AAAA,EAChD,CAAC,CAAA;AAGH,EAAA,MAAM,WAAW,MAAM;AACrB,IAAA,OAAA,CAAQ,GAAA,CAAI;AAAA,MACV,sBAAA,EAAwB,QAAA,EAAS,CAAE,KAAA,CAAM,MAAM;AAAA,MAAC,CAAC,CAAA;AAAA,MACjD,qBAAA,EAAuB,QAAA,EAAS,CAAE,KAAA,CAAM,MAAM;AAAA,MAAC,CAAC;AAAA,KACjD,CAAA,CAAE,OAAA,CAAQ,MAAM;AACf,MAAA,sBAAA,GAAyB,MAAA;AACzB,MAAA,qBAAA,GAAwB,MAAA;AAAA,IAC1B,CAAC,CAAA;AAAA,EACH,CAAA;AACA,EAAA,OAAA,CAAQ,IAAA,CAAK,WAAW,QAAQ,CAAA;AAChC,EAAA,OAAA,CAAQ,IAAA,CAAK,UAAU,QAAQ,CAAA;AAE/B,EAAA,8BAAA,EAA+B;AAS/B,EAAA,MAAM,KAAA,GAAQ,CAAC,MAAA,KAAmB,CAAC,GAAA,KAAiB;AAClD,IAAA,MAAM,SAAS,GAAA,YAAe,KAAA,GAAQ,GAAA,CAAI,OAAA,GAAU,OAAO,GAAG,CAAA;AAC9D,IAAA,IAAI;AACF,MAAA,yBAAA,CAA0B,QAAQ,MAAM,CAAA;AAAA,IAC1C,CAAA,CAAA,MAAQ;AAAA,IAER;AAGA,IAAA,MAAM,OAAA,GAAU,QAAQ,GAAA,CAAI;AAAA,MAC1B,sBAAA,EAAwB,UAAA,EAAW,CAAE,KAAA,CAAM,MAAM;AAAA,MAAC,CAAC,CAAA;AAAA,MACnD,qBAAA,EAAuB,UAAA,EAAW,CAAE,KAAA,CAAM,MAAM;AAAA,MAAC,CAAC;AAAA,KACnD,CAAA;AACD,IAAA,MAAM,OAAA,GAAU,IAAI,OAAA,CAAQ,CAAC,OAAA,KAAY,WAAW,OAAA,EAAS,GAAK,CAAA,CAAE,KAAA,IAAS,CAAA;AAC7E,IAAA,KAAK,OAAA,CAAQ,KAAK,CAAC,OAAA,EAAS,OAAO,CAAC,CAAA,CAAE,QAAQ,MAAM;AAClD,MAAA,OAAA,CAAQ,KAAA,CAAM,CAAA,QAAA,EAAW,MAAM,CAAA,EAAA,EAAK,MAAM,CAAA,CAAE,CAAA;AAC5C,MAAA,OAAA,CAAQ,KAAK,CAAC,CAAA;AAAA,IAChB,CAAC,CAAA;AAAA,EACH,CAAA;AACA,EAAA,OAAA,CAAQ,EAAA,CAAG,oBAAA,EAAsB,KAAA,CAAM,oBAAoB,CAAC,CAAA;AAC5D,EAAA,OAAA,CAAQ,EAAA,CAAG,mBAAA,EAAqB,KAAA,CAAM,mBAAmB,CAAC,CAAA;AAE1D,EAAA,OAAO,IAAA;AACT;AAGA,eAAsB,yBAAA,GAA2C;AAC/D,EAAA,MAAM,QAAQ,GAAA,CAAI;AAAA,IAChB,wBAAwB,QAAA,EAAS;AAAA,IACjC,uBAAuB,QAAA;AAAS,GACjC,CAAA;AACD,EAAA,sBAAA,GAAyB,MAAA;AACzB,EAAA,qBAAA,GAAwB,MAAA;AAC1B","file":"node.js","sourcesContent":["/**\n * @semiont/observability — public API.\n *\n * Universal surface (works in Node + browser). For SDK *initialization*,\n * import from `@semiont/observability/node` or `/web` at the process entry\n * point. Everything else uses this module.\n *\n * Tier 2 of `.plans/OBSERVABILITY.md`. The public surface:\n *\n * - `withSpan(name, fn, options?)` — wrap an async block in a span;\n * `options` carries `kind` and `attrs`.\n * - `withActorSpan(actor, channel, fn, extraAttrs?)` — consumer-span\n * wrapper for bus-event handlers, with handler-duration recording.\n * - `injectTraceparent(payload)` / `extractTraceparent(payload)` — W3C\n * trace-context propagation across the SSE channel (the bus payload\n * gets a `_trace?: { traceparent }` sibling to `correlationId`).\n * - `withTraceparent(carrier, fn)` — run `fn` with the incoming\n * traceparent as the parent context.\n * - `getActiveTraceparent()` — read the active span's traceparent for\n * manual propagation (e.g. attaching to a fetch header or SSE field).\n * - `getLogTraceContext()` — active `trace_id` / `span_id` for log-line\n * correlation.\n * - Metric recorders (`recordBusEmit`, `recordHandlerDuration`,\n * `recordJobOutcome`, `recordSubscriberConnect` / `Disconnect`,\n * `recordInferenceUsage`) and gauge providers\n * (`registerJobQueueProvider`, `registerVectorIndexSizeProvider`).\n *\n * No-op when no exporter is configured: `@opentelemetry/api`'s default\n * tracer is a no-op, so `withSpan` is essentially free until\n * `initObservability*()` runs.\n */\n\nimport {\n context,\n isSpanContextValid,\n metrics,\n propagation,\n SpanKind,\n SpanStatusCode,\n trace,\n type Attributes,\n type Counter,\n type Histogram,\n type ObservableGauge,\n type Span,\n type UpDownCounter,\n} from '@opentelemetry/api';\nimport { setBusLogTraceIdProvider } from '@semiont/core';\n\n// Wire `busLog`'s trace-id provider once at module load. When an OTel\n// SDK is initialized (and a span is active when `busLog` fires), the\n// emitted line gets a `trace=<8hex>` suffix that correlates the\n// grep-timeline with the trace UI. No-op when no SDK is active.\nsetBusLogTraceIdProvider(() => {\n const span = trace.getActiveSpan();\n if (!span) return undefined;\n const ctx = span.spanContext();\n if (!isSpanContextValid(ctx)) return undefined;\n return ctx.traceId;\n});\n\nconst TRACER_NAME = 'semiont';\n\nconst tracer = () => trace.getTracer(TRACER_NAME);\n\n// ── withSpan ───────────────────────────────────────────────────────────\n\n/**\n * Wrap an async block in a span. The span is started before `fn` runs and\n * ended after it resolves or rejects; exceptions are recorded and the span\n * status is set to ERROR. `kind` defaults to INTERNAL.\n */\nexport async function withSpan<T>(\n name: string,\n fn: (span: Span) => Promise<T> | T,\n options?: { kind?: SpanKind; attrs?: Attributes },\n): Promise<T> {\n const span = tracer().startSpan(name, {\n kind: options?.kind ?? SpanKind.INTERNAL,\n ...(options?.attrs ? { attributes: options.attrs } : {}),\n });\n try {\n return await context.with(trace.setSpan(context.active(), span), () => fn(span));\n } catch (err) {\n span.recordException(err as Error);\n span.setStatus({\n code: SpanStatusCode.ERROR,\n message: err instanceof Error ? err.message : String(err),\n });\n throw err;\n } finally {\n span.end();\n }\n}\n\n// ── Traceparent on bus payloads ────────────────────────────────────────\n\nconst TRACE_FIELD = '_trace';\n\n/**\n * Sibling of `correlationId` on bus payloads. Lives on the SSE event body\n * because SSE has no header trailer; the SDK strips it before delivering\n * the payload to subscribers. Additive — payloads without `_trace` parse\n * unchanged.\n */\nexport interface TraceCarrier {\n /** W3C `traceparent` header value (`00-<traceId>-<spanId>-<flags>`). */\n traceparent: string;\n /** W3C `tracestate` header value (vendor-specific extensions). */\n tracestate?: string;\n}\n\n/**\n * Read the active span's W3C traceparent (and tracestate). Returns\n * `undefined` if no span is active.\n */\nexport function getActiveTraceparent(): TraceCarrier | undefined {\n const carrier: Record<string, string> = {};\n propagation.inject(context.active(), carrier);\n const traceparent = carrier['traceparent'];\n if (!traceparent) return undefined;\n return carrier['tracestate']\n ? { traceparent, tracestate: carrier['tracestate'] }\n : { traceparent };\n}\n\n/**\n * Attach the active span's trace-context to a payload object as\n * `_trace`. No-op when no span is active. Returns the same object\n * reference for chaining.\n */\nexport function injectTraceparent<T extends Record<string, unknown>>(payload: T): T {\n const carrier = getActiveTraceparent();\n if (carrier) {\n (payload as Record<string, unknown>)[TRACE_FIELD] = carrier;\n }\n return payload;\n}\n\n/**\n * Strip and return the `_trace` field from a payload. Mutates `payload`.\n * The field is internal plumbing and should not be visible to subscribers.\n */\nexport function extractTraceparent<T extends Record<string, unknown>>(\n payload: T,\n): TraceCarrier | undefined {\n const carrier = (payload as Record<string, unknown>)[TRACE_FIELD] as\n | TraceCarrier\n | undefined;\n if (carrier !== undefined) {\n delete (payload as Record<string, unknown>)[TRACE_FIELD];\n }\n if (!carrier || typeof carrier.traceparent !== 'string') return undefined;\n return carrier;\n}\n\n/**\n * Run `fn` with the given W3C traceparent set as the parent context.\n * Any spans started inside `fn` will be children of the incoming trace.\n * No-op if `carrier` is undefined.\n */\nexport function withTraceparent<T>(\n carrier: TraceCarrier | undefined,\n fn: () => T,\n): T {\n if (!carrier) return fn();\n const carrierObj: Record<string, string> = { traceparent: carrier.traceparent };\n if (carrier.tracestate) carrierObj['tracestate'] = carrier.tracestate;\n const ctx = propagation.extract(context.active(), carrierObj);\n return context.with(ctx, fn);\n}\n\n// ── Actor handler convenience ──────────────────────────────────────────\n\n/**\n * Wrap a bus-event handler in an `actor.<name>:<channel>` consumer span.\n * Used at every `eventBus.get(channel).subscribe(handler)` site inside\n * an actor (Stower, Gatherer, Matcher, Browser, Smelter), to attribute\n * each in-process subscriber's work to a span without scattering manual\n * `withSpan` calls across handler bodies.\n *\n * The span's parent is the active context at the time the handler\n * fires — which is the `bus.dispatch:<channel>` span on the gateway\n * (Subject.next runs synchronously inside the dispatch span), or the\n * `bus.emit:<channel>` span when an actor emits to itself.\n */\nexport async function withActorSpan<T>(\n actor: string,\n channel: string,\n fn: (span: Span) => Promise<T> | T,\n extraAttrs?: Attributes,\n): Promise<T> {\n const start = performance.now();\n try {\n return await withSpan(`actor.${actor}:${channel}`, fn, {\n kind: SpanKind.CONSUMER,\n attrs: {\n actor,\n 'bus.channel': channel,\n ...(extraAttrs ?? {}),\n },\n });\n } finally {\n recordHandlerDuration(actor, channel, performance.now() - start);\n }\n}\n\n// ── Log correlation ────────────────────────────────────────────────────\n\n/**\n * Read the active span's `trace_id` / `span_id` for log-line correlation.\n * Tier 3 of `.plans/OBSERVABILITY.md`. Each structured log line gets\n * tagged with these so a log query in CloudWatch / Loki / Datadog can\n * jump to the trace in Tempo / Jaeger / X-Ray.\n *\n * Returns `undefined` if no span is active, or if the active span's\n * context is invalid (uninitialized SDK, no-op tracer).\n */\nexport function getLogTraceContext(): { trace_id: string; span_id: string } | undefined {\n const span = trace.getActiveSpan();\n if (!span) return undefined;\n const ctx = span.spanContext();\n if (!isSpanContextValid(ctx)) return undefined;\n return { trace_id: ctx.traceId, span_id: ctx.spanId };\n}\n\n// ── Metrics — Tier 3 ───────────────────────────────────────────────────\n\nconst METER_NAME = 'semiont';\n\nconst meter = () => metrics.getMeter(METER_NAME);\n\nlet _busEmitCounter: Counter | undefined;\nlet _replySuppressedCounter: Counter | undefined;\nlet _resumeGapCounter: Counter | undefined;\nlet _unanswerableCounter: Counter | undefined;\nlet _correlationRegistryGauge: ObservableGauge | undefined;\nlet _correlationRegistryProvider: (() => CorrelationRegistrySnapshot) | undefined;\nlet _handlerDurationHistogram: Histogram | undefined;\nlet _jobOutcomeCounter: Counter | undefined;\nlet _jobDurationHistogram: Histogram | undefined;\nlet _gatherDegradeCounter: Counter | undefined;\nlet _inferenceCallsCounter: Counter | undefined;\nlet _inferenceTokensCounter: Counter | undefined;\nlet _inferenceDurationHistogram: Histogram | undefined;\nlet _sseSubscribers: UpDownCounter | undefined;\nlet _jobQueueGauge: ObservableGauge | undefined;\nlet _jobQueueProvider: (() => Promise<JobQueueSnapshot> | JobQueueSnapshot) | undefined;\nlet _vectorIndexSizeGauge: ObservableGauge | undefined;\nlet _factPumpDepthGauge: ObservableGauge | undefined;\nlet _factPumpDepthProvider: (() => number) | undefined;\nlet _vectorIndexSizeProvider: (() => Promise<number> | number) | undefined;\n\n/** Snapshot of job-queue contents by status. Match `JobQueue.getStats()`. */\nexport interface JobQueueSnapshot {\n pending: number;\n running: number;\n complete: number;\n failed: number;\n cancelled: number;\n}\n\nfunction busEmitCounter(): Counter {\n if (!_busEmitCounter) {\n _busEmitCounter = meter().createCounter('semiont.bus.emit', {\n description: 'Bus emits by channel and scope',\n });\n }\n return _busEmitCounter;\n}\n\nfunction handlerDurationHistogram(): Histogram {\n if (!_handlerDurationHistogram) {\n _handlerDurationHistogram = meter().createHistogram('semiont.handler.duration', {\n description: 'In-process actor handler duration',\n unit: 'ms',\n });\n }\n return _handlerDurationHistogram;\n}\n\nfunction jobOutcomeCounter(): Counter {\n if (!_jobOutcomeCounter) {\n _jobOutcomeCounter = meter().createCounter('semiont.job.outcome', {\n description: 'Worker job completions by type and outcome',\n });\n }\n return _jobOutcomeCounter;\n}\n\nfunction jobDurationHistogram(): Histogram {\n if (!_jobDurationHistogram) {\n _jobDurationHistogram = meter().createHistogram('semiont.job.duration', {\n description: 'Worker job duration by type',\n unit: 'ms',\n });\n }\n return _jobDurationHistogram;\n}\n\nfunction inferenceCallsCounter(): Counter {\n if (!_inferenceCallsCounter) {\n _inferenceCallsCounter = meter().createCounter('semiont.inference.calls', {\n description: 'Inference API calls by provider, model, and outcome',\n });\n }\n return _inferenceCallsCounter;\n}\n\nfunction inferenceTokensCounter(): Counter {\n if (!_inferenceTokensCounter) {\n _inferenceTokensCounter = meter().createCounter('semiont.inference.tokens', {\n description: 'Inference token usage by provider, model, and direction',\n });\n }\n return _inferenceTokensCounter;\n}\n\nfunction inferenceDurationHistogram(): Histogram {\n if (!_inferenceDurationHistogram) {\n _inferenceDurationHistogram = meter().createHistogram('semiont.inference.duration', {\n description: 'Inference call duration by provider, model, and outcome',\n unit: 'ms',\n });\n }\n return _inferenceDurationHistogram;\n}\n\nfunction sseSubscribersCounter(): UpDownCounter {\n if (!_sseSubscribers) {\n _sseSubscribers = meter().createUpDownCounter('semiont.sse.subscribers', {\n description: 'Active SSE subscribers',\n });\n }\n return _sseSubscribers;\n}\n\nfunction replySuppressedCounter(): Counter {\n if (!_replySuppressedCounter) {\n _replySuppressedCounter = meter().createCounter('semiont.bus.reply.suppressed', {\n description: 'Correlated replies withheld from a non-owning subscriber',\n });\n }\n return _replySuppressedCounter;\n}\n\n/**\n * A correlated reply was withheld from a subscriber that does not own its\n * correlationId (CORRELATED-REPLY-ROUTING P5).\n *\n * Counts ONLY that case. A frame with no correlationId is a shape violation\n * (warned, not counted), and a cid nobody claimed is the structural in-process\n * case that fires constantly — counting either would drown the signal this\n * metric exists to show: the fan-out amplification the delivery filter removes.\n */\nexport function recordReplySuppressed(channel: string): void {\n replySuppressedCounter().add(1, { 'bus.channel': channel });\n}\n\nfunction resumeGapCounter(): Counter {\n if (!_resumeGapCounter) {\n _resumeGapCounter = meter().createCounter('semiont.bus.resume_gap', {\n description: 'SSE resumes that degraded to a gap because replay was unavailable',\n });\n }\n return _resumeGapCounter;\n}\n\n/**\n * An SSE resume could not be served and the client was told to fall back to\n * cache. This degradation is CORRECT by design and therefore silent — which is\n * exactly why it needs a number. A rising rate means clients are losing\n * history, and nothing else in the stack says so.\n */\nexport function recordResumeGap(reason: string): void {\n resumeGapCounter().add(1, { 'bus.resume_gap.reason': reason });\n}\n\nfunction unanswerableCounter(): Counter {\n if (!_unanswerableCounter) {\n _unanswerableCounter = meter().createCounter('semiont.bus.unanswerable', {\n description: 'Request emits that reached zero subscribers and were failed at the gateway',\n });\n }\n return _unanswerableCounter;\n}\n\n/**\n * A request-shaped emit reached no subscriber, so the gateway synthesized its\n * mapped failure (ARCHIVIST-STAYS-UP P3). By channel, this is the absence rate\n * of the service that answers it — the difference between \"it went down once\"\n * and \"it is flapping.\"\n */\nexport function recordUnanswerableRequest(channel: string): void {\n unanswerableCounter().add(1, { 'bus.channel': channel });\n}\n\n/** Claims held and reply payloads retained by a gateway's correlation registry. */\nexport interface CorrelationRegistrySnapshot {\n claims: number;\n retainedReplies: number;\n}\n\n/**\n * Register a callback returning the gateway's correlation-registry occupancy.\n *\n * COUNTS, not bytes: retention is count-budgeted today (byte-budgeting is a\n * known limit in CORRELATED-REPLY-ROUTING), so `retainedReplies` is a proxy for\n * heap, not a measure of it. It is still the closest observable to the question\n * two OOM investigations keep asking — a browse result can be 1-2 MB, and up to\n * REPLY_RETENTION_MAX of them are held at once.\n */\nexport function registerCorrelationRegistryProvider(\n provider: () => CorrelationRegistrySnapshot,\n): void {\n _correlationRegistryProvider = provider;\n if (!_correlationRegistryGauge) {\n _correlationRegistryGauge = meter().createObservableGauge('semiont.bus.correlation.size', {\n description: 'Correlation registry occupancy: live claims and retained reply payloads',\n });\n _correlationRegistryGauge.addCallback((observer) => {\n if (!_correlationRegistryProvider) return;\n const snap = _correlationRegistryProvider();\n observer.observe(snap.claims, { 'correlation.kind': 'claims' });\n observer.observe(snap.retainedReplies, { 'correlation.kind': 'retained_replies' });\n });\n }\n}\n\n/** Increment the bus-emit counter. Called at every transport `emit` site. */\nexport function recordBusEmit(channel: string, scope?: string): void {\n busEmitCounter().add(1, {\n 'bus.channel': channel,\n ...(scope ? { 'bus.scope': scope } : {}),\n });\n}\n\n/** Record an in-process actor handler's duration. */\nexport function recordHandlerDuration(actor: string, channel: string, durationMs: number): void {\n handlerDurationHistogram().record(durationMs, {\n actor,\n 'bus.channel': channel,\n });\n}\n\n/** Record a worker job's outcome and duration. */\nexport function recordJobOutcome(jobType: string, outcome: 'completed' | 'failed', durationMs: number): void {\n jobOutcomeCounter().add(1, { 'job.type': jobType, 'job.outcome': outcome });\n jobDurationHistogram().record(durationMs, { 'job.type': jobType, 'job.outcome': outcome });\n}\n\nlet _appendStageHistogram: Histogram | undefined;\nfunction appendStageHistogram(): Histogram {\n if (!_appendStageHistogram) {\n _appendStageHistogram = meter().createHistogram('semiont.record.append.duration', {\n description: 'Time spent in one stage of appending an event to the record, labeled by stage: persist (JSONL write + git), materialize (view rebuild), enrich, publish. The Archivist\\'s core write path.',\n unit: 'ms',\n });\n }\n return _appendStageHistogram;\n}\n\n/**\n * Record one stage of `EventStore.appendEvent` (ARCHIVIST-STAYS-UP P7).\n *\n * The append path is the one operation only the Archivist can perform, and it\n * was entirely dark: reads had `recordHandlerDuration` and the bus had its own\n * counters, while writes had nothing. Stage-labeled because the useful\n * question is never \"was the append slow\" but WHICH PART — and `materialize`\n * in particular does work proportional to a resource's annotation count, so it\n * degrades with history rather than with load.\n */\nexport function recordAppendStage(\n stage: 'persist' | 'materialize' | 'enrich' | 'publish',\n durationMs: number,\n): void {\n appendStageHistogram().record(durationMs, { 'record.stage': stage });\n}\n\nlet _gitCommandHistogram: Histogram | undefined;\nfunction gitCommandHistogram(): Histogram {\n if (!_gitCommandHistogram) {\n _gitCommandHistogram = meter().createHistogram('semiont.git.duration', {\n description: 'Wall time of a git subprocess. Async — this is latency, not event-loop blockage. Staging is deduped, so the `add` count is far below the number of appended events.',\n unit: 'ms',\n });\n }\n return _gitCommandHistogram;\n}\n\n/**\n * Record a git invocation. Read the `add` count against events appended: one\n * per event means deferred staging has stopped deduping.\n */\nexport function recordGitCommand(command: string, durationMs: number): void {\n gitCommandHistogram().record(durationMs, { 'git.command': command });\n}\n\nfunction gatherDegradeCounter(): Counter {\n if (!_gatherDegradeCounter) {\n _gatherDegradeCounter = meter().createCounter('semiont.gather.degraded', {\n description: 'Gathers that degraded because an eventually-consistent projection did not catch up within its read barrier (vectors: absent semanticContext; graph: projection-lag failure). Labeled by projection.',\n });\n }\n return _gatherDegradeCounter;\n}\n\n/**\n * Record a gather degraded by a projection read barrier: `'vectors'` — the\n * Smelter settle barrier timed out (semanticContext shipped absent);\n * `'graph'` — the Weaver applied barrier + poll floor exhausted (projection\n * lag surfaced as a distinct failure). Fleet-alertable counterpart of the\n * `[gather DEGRADED]` L4 breadcrumbs — a rising rate on either label means\n * that pipeline is not keeping up.\n */\nexport function recordGatherDegrade(projection: 'graph' | 'vectors'): void {\n gatherDegradeCounter().add(1, { projection });\n}\n\n/** Increment the SSE subscriber gauge — call on `/bus/subscribe` open. */\nexport function recordSubscriberConnect(): void {\n sseSubscribersCounter().add(1);\n}\n\n/** Decrement on disconnect. Pair with `recordSubscriberConnect`. */\nexport function recordSubscriberDisconnect(): void {\n sseSubscribersCounter().add(-1);\n}\n\n/**\n * Register a callback that returns the current job-queue snapshot.\n * Polled at the SDK's metric-collection interval. The single gauge\n * emits one observation per status (`pending`, `running`, …) tagged\n * with the `job.status` attribute. Idempotent — last registered\n * provider wins.\n */\nexport function registerJobQueueProvider(\n provider: () => Promise<JobQueueSnapshot> | JobQueueSnapshot,\n): void {\n _jobQueueProvider = provider;\n if (!_jobQueueGauge) {\n _jobQueueGauge = meter().createObservableGauge('semiont.job.queue.size', {\n description: 'Job queue size by status',\n });\n _jobQueueGauge.addCallback(async (observer) => {\n if (!_jobQueueProvider) return;\n const snap = await _jobQueueProvider();\n observer.observe(snap.pending, { 'job.status': 'pending' });\n observer.observe(snap.running, { 'job.status': 'running' });\n observer.observe(snap.complete, { 'job.status': 'complete' });\n observer.observe(snap.failed, { 'job.status': 'failed' });\n observer.observe(snap.cancelled, { 'job.status': 'cancelled' });\n });\n }\n}\n\n/**\n * Register a callback that returns the current vector-index size\n * (point count). Async to allow remote queries (Qdrant). Polled at\n * the metric-collection interval.\n */\n/**\n * Register the Archivist's fact-pump backlog — facts appended to the record\n * but not yet republished onto the bus.\n *\n * At rest this is zero. A value that climbs and does not come back means the\n * pump is outrunning its transport, which is the leading hypothesis for the\n * load-correlated heap growth in `bugs/absent-archivist-wedges-browse.md`\n * (ARCHIVIST-STAYS-UP P5). The backlog is deliberately unbounded today, so\n * this number is the only thing standing between \"the pump is behind\" and an\n * OOM whose cause is inferred from RSS after the fact.\n */\nexport function registerFactPumpDepthProvider(provider: () => number): void {\n _factPumpDepthProvider = provider;\n if (!_factPumpDepthGauge) {\n _factPumpDepthGauge = meter().createObservableGauge('semiont.archivist.fact_pump.depth', {\n description: 'Facts appended to the record but not yet published to the bus. Zero at rest; a rising floor means the pump is behind its transport.',\n });\n _factPumpDepthGauge.addCallback((observer) => {\n if (_factPumpDepthProvider) observer.observe(_factPumpDepthProvider());\n });\n }\n}\n\nlet _gitStagingFailureCounter: Counter | undefined;\n\n/**\n * A staging command that could not be run. Staging the index is a CONVENIENCE\n * — the event log is the system of record — so a failure here is degraded\n * service, never a reason to exit. But degraded must be VISIBLE: this counter\n * is what stops \"the index is quietly stale\" from being invisible.\n */\nexport function recordGitStagingFailure(reason: 'index-lock' | 'other'): void {\n if (!_gitStagingFailureCounter) {\n _gitStagingFailureCounter = meter().createCounter('semiont.git.staging.failures', {\n description: 'Staging commands abandoned after retries; the index may be stale',\n });\n }\n _gitStagingFailureCounter.add(1, { reason });\n}\n\n/**\n * Process lifetime telemetry (ARCHIVIST-GIT-STAGER-CRASH).\n *\n * A supervised process that dies and comes back is INVISIBLE in logs unless\n * someone greps for boot lines, and every request in flight when it died looks\n * to its caller like a hang. On 2026-09-08 that cost a multi-hour hunt through\n * search, qdrant, neo4j and the SSE transport for a crash loop that one metric\n * would have named immediately.\n *\n * `start_time` is the diagnostic, not uptime: a CHANGE in it is unambiguous\n * proof of a restart, and uptime is derivable from it.\n */\nconst PROCESS_START_TIME_SECONDS = Math.floor(Date.now() / 1000);\nlet _processStartTimeGauge: ObservableGauge | undefined;\nlet _restartCountGauge: ObservableGauge | undefined;\nlet _restartCountProvider: (() => Promise<number> | number) | undefined;\nlet _abnormalExitCounter: Counter | undefined;\n\n/** Register `semiont.process.start_time`. Called by `initObservability*`. */\nexport function registerProcessLifetimeMetrics(): void {\n if (_processStartTimeGauge) return;\n _processStartTimeGauge = meter().createObservableGauge('semiont.process.start_time', {\n description: 'Unix seconds at which this process started; a change means it restarted',\n unit: 's',\n });\n _processStartTimeGauge.addCallback((observer) => observer.observe(PROCESS_START_TIME_SECONDS));\n}\n\n/**\n * Supply a restart count. The Archivist's supervisor is POSIX shell and cannot\n * emit OTel, but it already keeps a durable event log on the state mount — so\n * the supervised child reads it and reports the count on the supervisor's\n * behalf.\n */\nexport function registerRestartCountProvider(\n provider: () => Promise<number> | number,\n): void {\n _restartCountProvider = provider;\n if (!_restartCountGauge) {\n _restartCountGauge = meter().createObservableGauge('semiont.process.restarts', {\n description: 'Times the supervisor has restarted this service',\n });\n _restartCountGauge.addCallback(async (observer) => {\n if (_restartCountProvider) observer.observe(await _restartCountProvider());\n });\n }\n}\n\n/**\n * Record that this process is dying abnormally, and mark the active span so a\n * trace shows a span that ENDED IN DEATH rather than one that simply never\n * ends. Never swallows: callers re-raise, so Node's own semantics are intact.\n */\nexport function recordAbnormalTermination(reason: string, detail?: string): void {\n if (!_abnormalExitCounter) {\n _abnormalExitCounter = meter().createCounter('semiont.process.abnormal_exit', {\n description: 'Process terminations that were not a clean shutdown',\n });\n }\n _abnormalExitCounter.add(1, { reason });\n const active = trace.getActiveSpan();\n if (active) {\n active.setStatus({ code: SpanStatusCode.ERROR, message: `${reason}: ${detail ?? ''}`.trim() });\n active.setAttribute('semiont.process.abnormal_exit', reason);\n active.end();\n }\n}\n\nexport function registerVectorIndexSizeProvider(\n provider: () => Promise<number> | number,\n): void {\n _vectorIndexSizeProvider = provider;\n if (!_vectorIndexSizeGauge) {\n _vectorIndexSizeGauge = meter().createObservableGauge('semiont.vector.index.size', {\n description: 'Vector store point count',\n });\n _vectorIndexSizeGauge.addCallback(async (observer) => {\n if (_vectorIndexSizeProvider) {\n const value = await _vectorIndexSizeProvider();\n observer.observe(value);\n }\n });\n }\n}\n\n/**\n * Record an inference call. Token counts are optional — providers that\n * don't expose them (or fail before generating) record only call count\n * and duration.\n */\nexport function recordInferenceUsage(opts: {\n provider: string;\n model: string;\n durationMs: number;\n outcome: 'success' | 'error';\n inputTokens?: number;\n outputTokens?: number;\n}): void {\n const baseAttrs = {\n 'inference.provider': opts.provider,\n 'inference.model': opts.model,\n 'inference.outcome': opts.outcome,\n };\n inferenceCallsCounter().add(1, baseAttrs);\n inferenceDurationHistogram().record(opts.durationMs, baseAttrs);\n if (opts.inputTokens != null && opts.inputTokens > 0) {\n inferenceTokensCounter().add(opts.inputTokens, {\n 'inference.provider': opts.provider,\n 'inference.model': opts.model,\n 'inference.direction': 'input',\n });\n }\n if (opts.outputTokens != null && opts.outputTokens > 0) {\n inferenceTokensCounter().add(opts.outputTokens, {\n 'inference.provider': opts.provider,\n 'inference.model': opts.model,\n 'inference.direction': 'output',\n });\n }\n}\n\nlet _detectionCallCounter: Counter | undefined;\nlet _detectionDurationHistogram: Histogram | undefined;\nlet _detectionItemsHistogram: Histogram | undefined;\nlet _detectionTokensHistogram: Histogram | undefined;\n\nfunction detectionCallCounter(): Counter {\n if (!_detectionCallCounter) {\n _detectionCallCounter = meter().createCounter('semiont.detection.calls', {\n description: 'Detection model calls, labeled by motivation, outcome, subdivision depth and whether this was the floor re-roll.',\n });\n }\n return _detectionCallCounter;\n}\n\nfunction detectionDurationHistogram(): Histogram {\n if (!_detectionDurationHistogram) {\n _detectionDurationHistogram = meter().createHistogram('semiont.detection.call.duration', {\n description: 'Wall time of one detection model call, including the attempts that failed and were retried smaller.',\n unit: 'ms',\n });\n }\n return _detectionDurationHistogram;\n}\n\nfunction detectionItemsHistogram(): Histogram {\n if (!_detectionItemsHistogram) {\n _detectionItemsHistogram = meter().createHistogram('semiont.detection.call.items', {\n description: 'Annotations returned by one detection call. Against the input size on the same record, this is yield.',\n });\n }\n return _detectionItemsHistogram;\n}\n\nfunction detectionTokensHistogram(): Histogram {\n if (!_detectionTokensHistogram) {\n _detectionTokensHistogram = meter().createHistogram('semiont.detection.call.tokens', {\n description: \"Provider-reported tokens for one detection call, by direction. Deliberately separate from semiont.inference.tokens: that series is the authoritative total but carries no subdivision depth, and 'what does a depth-2 call cost' is the question every sizing decision asks.\",\n });\n }\n return _detectionTokensHistogram;\n}\n\n/**\n * Record one detection model call (DETECTION-QUALITY-THROUGHPUT P1).\n *\n * The adapters already record provider/model/duration/tokens for every\n * inference call. What they cannot know is the detection shape around it:\n * which motivation asked, how big the piece was, how many annotations came\n * back, how deep subdivision had descended, and whether this was the floor\n * re-roll. Those are the facts that distinguish a healthy call from a\n * expensive descent, and without them a slow detection run is one\n * undifferentiated number.\n *\n * FAILED attempts are recorded too, and that is the point: the calls paid for\n * and thrown away during a descent are exactly the cost later phases exist to\n * avoid, so a record only of successes would hide the thing being optimized.\n *\n * Tokens are the PROVIDER's counts, passed through — never estimated. Absent\n * means the provider did not report them.\n */\nexport function recordDetectionCall(opts: {\n label: string;\n pieceChars: number;\n durationMs: number;\n items: number;\n depth: number;\n reroll: boolean;\n outcome: 'success' | 'truncated' | 'timeout' | 'collapsed' | 'error';\n inputTokens?: number;\n outputTokens?: number;\n}): void {\n const attrs = {\n 'detection.label': opts.label,\n 'detection.outcome': opts.outcome,\n 'detection.depth': opts.depth,\n 'detection.reroll': opts.reroll,\n };\n detectionCallCounter().add(1, attrs);\n detectionDurationHistogram().record(opts.durationMs, attrs);\n detectionItemsHistogram().record(opts.items, attrs);\n if (opts.inputTokens !== undefined) {\n detectionTokensHistogram().record(opts.inputTokens, { ...attrs, 'detection.direction': 'input' });\n }\n if (opts.outputTokens !== undefined) {\n detectionTokensHistogram().record(opts.outputTokens, { ...attrs, 'detection.direction': 'output' });\n }\n}\n\nlet _anchorOutcomeCounter: Counter | undefined;\nfunction anchorOutcomeCounter(): Counter {\n if (!_anchorOutcomeCounter) {\n _anchorOutcomeCounter = meter().createCounter('semiont.detection.anchors', {\n description: 'Every annotation anchoring, labeled by the method that resolved it. EVERY outcome is counted, not just the risky ones, because a bare count of degraded anchors has no denominator — the rate is the precision signal.',\n });\n }\n return _anchorOutcomeCounter;\n}\n\n/**\n * Record how one annotation got anchored (DETECTION-QUALITY-THROUGHPUT P5).\n *\n * The selector-vs-source check is already a WRITE-TIME INVARIANT — both\n * `buildTextAnnotation` and `buildPdfAnnotation` throw on a selector that does\n * not match its source — so mechanical correctness is guaranteed rather than\n * sampled, and auditing it would measure a constant.\n *\n * What is genuinely uncertain is which anchoring METHOD got there. An `exact`\n * the model quoted verbatim and that appears once is certain; one resolved by\n * `first-of-many` (several occurrences, no usable context) or `fuzzy-match`\n * picked a plausible occurrence and may have picked wrong. Those were visible\n * only as log warnings — countable by a human reading worker output, which is\n * how 47 of them went unreviewed. As a rate they are the precision number that\n * sits beside the yield numbers.\n */\nexport function recordAnchorOutcome(label: string, method: string): void {\n anchorOutcomeCounter().add(1, { 'detection.label': label, 'anchor.method': method });\n}\n\n// ── Re-exports from @opentelemetry/api ─────────────────────────────────\n\nexport { SpanKind, SpanStatusCode, type Attributes, type Span } from '@opentelemetry/api';\n","/**\n * Process-runtime readings (ARCHIVIST-STAYS-UP P4).\n *\n * Node-only, and deliberately a plain function rather than a gauge callback:\n * the numbers are the thing worth testing, and a callback registered inside\n * an SDK is awkward to assert against.\n *\n * **Why `heapLimit` is the field that matters.** The Archivist died at\n * ~1016 MB inside a 2048 MB container (`bugs/absent-archivist-wedges-browse.md`)\n * — not because it exhausted the container, but because it hit V8's OWN\n * default old-space ceiling, which is derived from visible memory and lands\n * well under it. `heapUsed` alone cannot express \"how close to death is\n * this\"; only the pair can. It is also how a configured\n * `--max-old-space-size` is verified to have taken effect, rather than\n * assumed from the fact that someone set an env var.\n */\n\nimport { getHeapStatistics } from 'node:v8';\n\nexport interface HeapStats {\n /** Live heap in use. */\n heapUsed: number;\n /** Heap V8 has currently reserved. */\n heapTotal: number;\n /** The ceiling V8 will die at — NOT the container's limit. */\n heapLimit: number;\n /** Resident set: everything, including buffers outside the JS heap. */\n rss: number;\n}\n\nexport function heapStats(): HeapStats {\n const mem = process.memoryUsage();\n return {\n heapUsed: mem.heapUsed,\n heapTotal: mem.heapTotal,\n heapLimit: getHeapStatistics().heap_size_limit,\n rss: mem.rss,\n };\n}\n","/**\n * Node SDK initialization. Call once at the process entry point\n * (gateway `index.ts`, `worker-main.ts`, `smelter-main.ts`).\n *\n * Configuration is via standard `OTEL_*` env vars:\n * - `OTEL_SERVICE_NAME` — service identity (e.g. `semiont-gateway`)\n * - `OTEL_EXPORTER_OTLP_ENDPOINT` — collector endpoint (HTTP)\n * - `OTEL_TRACES_SAMPLER` — sampler (default: `parentbased_always_on`)\n * - `OTEL_TRACES_SAMPLER_ARG` — sampler ratio (default: `1.0`)\n * - `OTEL_CONSOLE_EXPORTER=true` — dev-only: emit spans + metrics to stderr\n * - `OTEL_SDK_DISABLED=true` — skip initialization entirely\n *\n * **Off-by-default invariant**: with neither `OTEL_EXPORTER_OTLP_ENDPOINT`\n * nor `OTEL_CONSOLE_EXPORTER=true` set, this function is a no-op and the\n * `@opentelemetry/api` no-op tracer takes over. This avoids accidentally\n * flooding production stderr (and CloudWatch) when an operator deploys\n * without configuring an exporter.\n *\n * Implementation note: this module wires `BasicTracerProvider` and\n * `MeterProvider` (both from the stable `@opentelemetry/sdk-trace-base`\n * 2.x line) directly, plus `AsyncLocalStorageContextManager` for Node\n * async-context propagation. We deliberately avoid `@opentelemetry/sdk-node`\n * because its `0.x` experimental versions cross-depend on older 2.0.x\n * SDKs, forcing npm to nest duplicate copies of the stable packages and\n * blowing up bundles for every consumer.\n */\n\nimport { recordAbnormalTermination, registerProcessLifetimeMetrics } from './index.js';\nimport { monitorEventLoopDelay } from 'node:perf_hooks';\nimport { heapStats } from './runtime-stats';\nimport { context, metrics, propagation, trace } from '@opentelemetry/api';\nimport { AsyncLocalStorageContextManager } from '@opentelemetry/context-async-hooks';\nimport { W3CTraceContextPropagator } from '@opentelemetry/core';\nimport { OTLPMetricExporter } from '@opentelemetry/exporter-metrics-otlp-http';\nimport { OTLPTraceExporter } from '@opentelemetry/exporter-trace-otlp-http';\nimport { resourceFromAttributes } from '@opentelemetry/resources';\nimport {\n ConsoleMetricExporter,\n MeterProvider,\n PeriodicExportingMetricReader,\n} from '@opentelemetry/sdk-metrics';\nimport {\n BasicTracerProvider,\n BatchSpanProcessor,\n ConsoleSpanExporter,\n} from '@opentelemetry/sdk-trace-base';\nimport {\n ATTR_SERVICE_NAME,\n ATTR_SERVICE_VERSION,\n} from '@opentelemetry/semantic-conventions';\n\nexport interface NodeObservabilityConfig {\n /** Service identity (e.g. `semiont-gateway`). Overridden by `OTEL_SERVICE_NAME`. */\n serviceName: string;\n /** Service version. Defaults to `0.0.0` if omitted. */\n serviceVersion?: string;\n}\n\nlet tracerProviderInstance: BasicTracerProvider | undefined;\nlet meterProviderInstance: MeterProvider | undefined;\n\n/**\n * Default metric export interval. 30s mirrors the SDK default and gives\n * operators enough granularity without flooding the collector.\n */\nconst DEFAULT_METRIC_EXPORT_INTERVAL_MS = 30_000;\n\n/**\n * Which exporter metrics use. `console` wins even with an OTLP endpoint set —\n * the readout for a bare process or CI with no collector; traces unaffected.\n * Any other value (incl. unrecognised) leaves the endpoint to decide, so an\n * unknown value cannot silently disable metrics. Exported and pure so tests\n * assert the choice without mocking the exporter modules.\n */\nexport function metricsExporterKind(\n endpoint: string | undefined,\n requested: string | undefined,\n): 'otlp' | 'console' {\n if (requested === 'console') return 'console';\n return endpoint ? 'otlp' : 'console';\n}\n\n/**\n * Initialize OTel for the current process. Wires up both tracing and\n * metrics. Idempotent — calling twice is a no-op. Returns `true` if the\n * SDK started, `false` if disabled, no exporter is configured, or\n * already initialized.\n *\n * Metrics export at `OTEL_METRIC_EXPORT_INTERVAL` ms (default 30s) to\n * the same `OTEL_EXPORTER_OTLP_ENDPOINT` as traces.\n */\nexport function initObservabilityNode(config: NodeObservabilityConfig): boolean {\n if (tracerProviderInstance) return false;\n if (process.env['OTEL_SDK_DISABLED'] === 'true') return false;\n\n const endpoint = process.env['OTEL_EXPORTER_OTLP_ENDPOINT'];\n const useConsole = process.env['OTEL_CONSOLE_EXPORTER'] === 'true';\n\n // No exporter configured = no SDK init. The `@opentelemetry/api`\n // no-op tracer takes over; `withSpan` still runs `fn` but emits\n // nothing, `getActiveSpan()` returns a sentinel. Avoids flooding\n // production stderr when no collector endpoint is set.\n if (!endpoint && !useConsole) return false;\n\n const resource = resourceFromAttributes({\n [ATTR_SERVICE_NAME]: process.env['OTEL_SERVICE_NAME'] ?? config.serviceName,\n [ATTR_SERVICE_VERSION]: config.serviceVersion ?? '0.0.0',\n });\n\n // Trace SDK\n const traceExporter = endpoint ? new OTLPTraceExporter() : new ConsoleSpanExporter();\n tracerProviderInstance = new BasicTracerProvider({\n resource,\n spanProcessors: [new BatchSpanProcessor(traceExporter)],\n });\n\n // Async-context propagation across `await` boundaries — equivalent\n // to what NodeSDK installs internally, but pinned to the stable 2.x\n // cohort.\n context.setGlobalContextManager(new AsyncLocalStorageContextManager().enable());\n trace.setGlobalTracerProvider(tracerProviderInstance);\n\n // W3C trace-context propagator. Without this, `propagation.inject`\n // and `propagation.extract` walk an empty propagator chain and\n // silently do nothing — which means traceparent never makes it onto\n // outgoing HTTP headers or SSE `_trace` payloads, and cross-service\n // traces stay disconnected. NodeSDK registers this for you; bare\n // BasicTracerProvider does not.\n propagation.setGlobalPropagator(new W3CTraceContextPropagator());\n\n // Traces follow the endpoint; metrics additionally honour OTEL_METRICS_EXPORTER.\n const metricExporter =\n metricsExporterKind(endpoint, process.env['OTEL_METRICS_EXPORTER']) === 'console'\n ? new ConsoleMetricExporter()\n : new OTLPMetricExporter();\n const intervalRaw = process.env['OTEL_METRIC_EXPORT_INTERVAL'];\n const exportIntervalMillis = intervalRaw\n ? Number.parseInt(intervalRaw, 10)\n : DEFAULT_METRIC_EXPORT_INTERVAL_MS;\n\n meterProviderInstance = new MeterProvider({\n resource,\n readers: [\n new PeriodicExportingMetricReader({\n exporter: metricExporter,\n exportIntervalMillis: Number.isFinite(exportIntervalMillis) && exportIntervalMillis > 0\n ? exportIntervalMillis\n : DEFAULT_METRIC_EXPORT_INTERVAL_MS,\n }),\n ],\n });\n metrics.setGlobalMeterProvider(meterProviderInstance);\n\n // The cause-AGNOSTIC detector for a blocked process: it rises whether the\n // cause is a large JSON parse, GC, or a sync subprocess. Cheap — libuv keeps\n // the histogram; we read percentiles and reset once per export interval.\n const loopDelay = monitorEventLoopDelay({ resolution: 10 });\n loopDelay.enable();\n const runtimeMeter = meterProviderInstance.getMeter('semiont-runtime');\n runtimeMeter\n .createObservableGauge('semiont.runtime.event_loop.lag', {\n description: 'Event-loop delay percentiles over the last export interval. Time the process could not serve anything.',\n unit: 'ms',\n })\n .addCallback((observer) => {\n observer.observe(loopDelay.mean / 1e6, { 'lag.stat': 'mean' });\n observer.observe(loopDelay.percentile(99) / 1e6, { 'lag.stat': 'p99' });\n observer.observe(loopDelay.max / 1e6, { 'lag.stat': 'max' });\n loopDelay.reset();\n });\n\n // Heap (ARCHIVIST-STAYS-UP P4), on the SAME registration as lag rather than\n // a second mechanism — they are read together when diagnosing a process\n // that stopped answering, and splitting them would mean two things to wire.\n //\n // `limit` is the field that repays the effort: the Archivist died at\n // ~1016 MB inside a 2048 MB container because V8's own default ceiling sits\n // well under the container's. `used` alone cannot say how close to death a\n // process is, and `limit` is what a configured --max-old-space-size changes\n // — so this is also how that setting is VERIFIED rather than assumed.\n runtimeMeter\n .createObservableGauge('semiont.runtime.heap', {\n description: \"Process memory by kind. `limit` is V8's own ceiling, which is what the process dies at — not the container's allocation.\",\n unit: 'By',\n })\n .addCallback((observer) => {\n const s = heapStats();\n observer.observe(s.heapUsed, { 'heap.stat': 'used' });\n observer.observe(s.heapTotal, { 'heap.stat': 'total' });\n observer.observe(s.heapLimit, { 'heap.stat': 'limit' });\n observer.observe(s.rss, { 'heap.stat': 'rss' });\n });\n\n // Flush traces + metrics on shutdown so nothing is lost on SIGTERM/SIGINT.\n const shutdown = () => {\n Promise.all([\n tracerProviderInstance?.shutdown().catch(() => {}),\n meterProviderInstance?.shutdown().catch(() => {}),\n ]).finally(() => {\n tracerProviderInstance = undefined;\n meterProviderInstance = undefined;\n });\n };\n process.once('SIGTERM', shutdown);\n process.once('SIGINT', shutdown);\n\n registerProcessLifetimeMetrics();\n\n // A fatal path must leave a RECORD, not just a stack trace on stdout.\n //\n // Semantics are deliberately unchanged: Node treats an unhandled rejection\n // and an uncaught exception as fatal, and so do we. Registering a listener\n // would normally SUPPRESS that, which is why each handler re-raises after\n // recording — swallowing here would convert a loud crash into a silent\n // wedged process, which is strictly worse than the bug that motivated this.\n const fatal = (reason: string) => (err: unknown) => {\n const detail = err instanceof Error ? err.message : String(err);\n try {\n recordAbnormalTermination(reason, detail);\n } catch {\n // Telemetry must never be the reason a crash report is lost.\n }\n // Best-effort export before the process goes; bounded so a dead collector\n // cannot hold a dying process open.\n const flushed = Promise.all([\n tracerProviderInstance?.forceFlush().catch(() => {}),\n meterProviderInstance?.forceFlush().catch(() => {}),\n ]);\n const bounded = new Promise((resolve) => setTimeout(resolve, 2_000).unref?.());\n void Promise.race([flushed, bounded]).finally(() => {\n console.error(`[fatal] ${reason}: ${detail}`);\n process.exit(1);\n });\n };\n process.on('unhandledRejection', fatal('unhandledRejection'));\n process.on('uncaughtException', fatal('uncaughtException'));\n\n return true;\n}\n\n/** Force-flush + shutdown both SDKs. Test cleanup, not production. */\nexport async function shutdownObservabilityNode(): Promise<void> {\n await Promise.all([\n tracerProviderInstance?.shutdown(),\n meterProviderInstance?.shutdown(),\n ]);\n tracerProviderInstance = undefined;\n meterProviderInstance = undefined;\n}\n"]}
@@ -1 +1 @@
1
- {"version":3,"sources":["../src/index.ts","../src/process-logger.ts"],"names":["trace"],"mappings":";;;;;AAqDA,wBAAA,CAAyB,MAAM;AAC7B,EAAA,MAAM,IAAA,GAAO,MAAM,aAAA,EAAc;AACjC,EAAA,IAAI,CAAC,MAAM,OAAO,MAAA;AAClB,EAAA,MAAM,GAAA,GAAM,KAAK,WAAA,EAAY;AAC7B,EAAA,IAAI,CAAC,kBAAA,CAAmB,GAAG,CAAA,EAAG,OAAO,MAAA;AACrC,EAAA,OAAO,GAAA,CAAI,OAAA;AACb,CAAC,CAAA;AA+JM,SAAS,kBAAA,GAAwE;AACtF,EAAA,MAAM,IAAA,GAAO,MAAM,aAAA,EAAc;AACjC,EAAA,IAAI,CAAC,MAAM,OAAO,MAAA;AAClB,EAAA,MAAM,GAAA,GAAM,KAAK,WAAA,EAAY;AAC7B,EAAA,IAAI,CAAC,kBAAA,CAAmB,GAAG,CAAA,EAAG,OAAO,MAAA;AACrC,EAAA,OAAO,EAAE,QAAA,EAAU,GAAA,CAAI,OAAA,EAAS,OAAA,EAAS,IAAI,MAAA,EAAO;AACtD;;;AC1MA,IAAM,kBAAA,GAAqB,OAAA,CAAQ,MAAA,CAAO,CAAC,IAAA,KAAS;AAClD,EAAA,MAAMA,SAAQ,kBAAA,EAAmB;AACjC,EAAA,IAAIA,MAAAA,EAAO;AACT,IAAA,IAAA,CAAK,WAAWA,MAAAA,CAAM,QAAA;AACtB,IAAA,IAAA,CAAK,UAAUA,MAAAA,CAAM,OAAA;AAAA,EACvB;AACA,EAAA,OAAO,IAAA;AACT,CAAC,CAAA,EAAE;AAEI,SAAS,oBAAoB,SAAA,EAA2B;AAC7D,EAAA,MAAM,KAAA,GAAQ,OAAA,CAAQ,GAAA,CAAI,SAAA,IAAa,MAAA;AACvC,EAAA,MAAM,SAAS,OAAA,CAAQ,GAAA,CAAI,UAAA,KAAe,QAAA,GACtC,QAAQ,MAAA,CAAO,OAAA;AAAA,IACb,QAAQ,MAAA,CAAO,SAAA,CAAU,EAAE,MAAA,EAAQ,uBAAuB,CAAA;AAAA,IAC1D,QAAQ,MAAA,CAAO,MAAA,CAAO,EAAE,KAAA,EAAO,MAAM,CAAA;AAAA,IACrC,kBAAA;AAAA,IACA,OAAA,CAAQ,MAAA,CAAO,MAAA,CAAO,CAAC,EAAE,KAAA,EAAO,GAAA,EAAK,OAAA,EAAS,SAAA,EAAW,GAAG,IAAA,EAAK,KAAM;AACrE,MAAA,MAAM,OAAA,GAAU,MAAA,CAAO,IAAA,CAAK,IAAI,CAAA,CAAE,MAAA,GAAS,CAAA,GAAI,CAAA,CAAA,EAAI,IAAA,CAAK,SAAA,CAAU,IAAI,CAAC,CAAA,CAAA,GAAK,EAAA;AAC5E,MAAA,OAAO,CAAA,EAAG,SAAS,CAAA,EAAA,EAAK,GAAA,CAAI,WAAA,EAAa,CAAA,GAAA,EAAM,SAAS,CAAA,EAAA,EAAK,OAAO,CAAA,EAAG,OAAO,CAAA,CAAA;AAAA,IAChF,CAAC;AAAA,GACH,GACA,QAAQ,MAAA,CAAO,OAAA;AAAA,IACb,OAAA,CAAQ,OAAO,SAAA,EAAU;AAAA,IACzB,QAAQ,MAAA,CAAO,MAAA,CAAO,EAAE,KAAA,EAAO,MAAM,CAAA;AAAA,IACrC,kBAAA;AAAA,IACA,OAAA,CAAQ,OAAO,IAAA;AAAK,GACtB;AAEJ,EAAA,MAAM,MAAA,GAAS,QAAQ,YAAA,CAAa;AAAA,IAClC,KAAA;AAAA,IACA,WAAA,EAAa,EAAE,SAAA,EAAU;AAAA,IACzB,MAAA;AAAA,IACA,YAAY,CAAC,IAAI,OAAA,CAAQ,UAAA,CAAW,SAAS;AAAA,GAC9C,CAAA;AAED,EAAA,OAAO,MAAA;AACT","file":"process-logger.js","sourcesContent":["/**\n * @semiont/observability — public API.\n *\n * Universal surface (works in Node + browser). For SDK *initialization*,\n * import from `@semiont/observability/node` or `/web` at the process entry\n * point. Everything else uses this module.\n *\n * Tier 2 of `.plans/OBSERVABILITY.md`. The public surface:\n *\n * - `withSpan(name, fn, options?)` — wrap an async block in a span;\n * `options` carries `kind` and `attrs`.\n * - `withActorSpan(actor, channel, fn, extraAttrs?)` — consumer-span\n * wrapper for bus-event handlers, with handler-duration recording.\n * - `injectTraceparent(payload)` / `extractTraceparent(payload)` — W3C\n * trace-context propagation across the SSE channel (the bus payload\n * gets a `_trace?: { traceparent }` sibling to `correlationId`).\n * - `withTraceparent(carrier, fn)` — run `fn` with the incoming\n * traceparent as the parent context.\n * - `getActiveTraceparent()` — read the active span's traceparent for\n * manual propagation (e.g. attaching to a fetch header or SSE field).\n * - `getLogTraceContext()` — active `trace_id` / `span_id` for log-line\n * correlation.\n * - Metric recorders (`recordBusEmit`, `recordHandlerDuration`,\n * `recordJobOutcome`, `recordSubscriberConnect` / `Disconnect`,\n * `recordInferenceUsage`) and gauge providers\n * (`registerJobQueueProvider`, `registerVectorIndexSizeProvider`).\n *\n * No-op when no exporter is configured: `@opentelemetry/api`'s default\n * tracer is a no-op, so `withSpan` is essentially free until\n * `initObservability*()` runs.\n */\n\nimport {\n context,\n isSpanContextValid,\n metrics,\n propagation,\n SpanKind,\n SpanStatusCode,\n trace,\n type Attributes,\n type Counter,\n type Histogram,\n type ObservableGauge,\n type Span,\n type UpDownCounter,\n} from '@opentelemetry/api';\nimport { setBusLogTraceIdProvider } from '@semiont/core';\n\n// Wire `busLog`'s trace-id provider once at module load. When an OTel\n// SDK is initialized (and a span is active when `busLog` fires), the\n// emitted line gets a `trace=<8hex>` suffix that correlates the\n// grep-timeline with the trace UI. No-op when no SDK is active.\nsetBusLogTraceIdProvider(() => {\n const span = trace.getActiveSpan();\n if (!span) return undefined;\n const ctx = span.spanContext();\n if (!isSpanContextValid(ctx)) return undefined;\n return ctx.traceId;\n});\n\nconst TRACER_NAME = 'semiont';\n\nconst tracer = () => trace.getTracer(TRACER_NAME);\n\n// ── withSpan ───────────────────────────────────────────────────────────\n\n/**\n * Wrap an async block in a span. The span is started before `fn` runs and\n * ended after it resolves or rejects; exceptions are recorded and the span\n * status is set to ERROR. `kind` defaults to INTERNAL.\n */\nexport async function withSpan<T>(\n name: string,\n fn: (span: Span) => Promise<T> | T,\n options?: { kind?: SpanKind; attrs?: Attributes },\n): Promise<T> {\n const span = tracer().startSpan(name, {\n kind: options?.kind ?? SpanKind.INTERNAL,\n ...(options?.attrs ? { attributes: options.attrs } : {}),\n });\n try {\n return await context.with(trace.setSpan(context.active(), span), () => fn(span));\n } catch (err) {\n span.recordException(err as Error);\n span.setStatus({\n code: SpanStatusCode.ERROR,\n message: err instanceof Error ? err.message : String(err),\n });\n throw err;\n } finally {\n span.end();\n }\n}\n\n// ── Traceparent on bus payloads ────────────────────────────────────────\n\nconst TRACE_FIELD = '_trace';\n\n/**\n * Sibling of `correlationId` on bus payloads. Lives on the SSE event body\n * because SSE has no header trailer; the SDK strips it before delivering\n * the payload to subscribers. Additive — payloads without `_trace` parse\n * unchanged.\n */\nexport interface TraceCarrier {\n /** W3C `traceparent` header value (`00-<traceId>-<spanId>-<flags>`). */\n traceparent: string;\n /** W3C `tracestate` header value (vendor-specific extensions). */\n tracestate?: string;\n}\n\n/**\n * Read the active span's W3C traceparent (and tracestate). Returns\n * `undefined` if no span is active.\n */\nexport function getActiveTraceparent(): TraceCarrier | undefined {\n const carrier: Record<string, string> = {};\n propagation.inject(context.active(), carrier);\n const traceparent = carrier['traceparent'];\n if (!traceparent) return undefined;\n return carrier['tracestate']\n ? { traceparent, tracestate: carrier['tracestate'] }\n : { traceparent };\n}\n\n/**\n * Attach the active span's trace-context to a payload object as\n * `_trace`. No-op when no span is active. Returns the same object\n * reference for chaining.\n */\nexport function injectTraceparent<T extends Record<string, unknown>>(payload: T): T {\n const carrier = getActiveTraceparent();\n if (carrier) {\n (payload as Record<string, unknown>)[TRACE_FIELD] = carrier;\n }\n return payload;\n}\n\n/**\n * Strip and return the `_trace` field from a payload. Mutates `payload`.\n * The field is internal plumbing and should not be visible to subscribers.\n */\nexport function extractTraceparent<T extends Record<string, unknown>>(\n payload: T,\n): TraceCarrier | undefined {\n const carrier = (payload as Record<string, unknown>)[TRACE_FIELD] as\n | TraceCarrier\n | undefined;\n if (carrier !== undefined) {\n delete (payload as Record<string, unknown>)[TRACE_FIELD];\n }\n if (!carrier || typeof carrier.traceparent !== 'string') return undefined;\n return carrier;\n}\n\n/**\n * Run `fn` with the given W3C traceparent set as the parent context.\n * Any spans started inside `fn` will be children of the incoming trace.\n * No-op if `carrier` is undefined.\n */\nexport function withTraceparent<T>(\n carrier: TraceCarrier | undefined,\n fn: () => T,\n): T {\n if (!carrier) return fn();\n const carrierObj: Record<string, string> = { traceparent: carrier.traceparent };\n if (carrier.tracestate) carrierObj['tracestate'] = carrier.tracestate;\n const ctx = propagation.extract(context.active(), carrierObj);\n return context.with(ctx, fn);\n}\n\n// ── Actor handler convenience ──────────────────────────────────────────\n\n/**\n * Wrap a bus-event handler in an `actor.<name>:<channel>` consumer span.\n * Used at every `eventBus.get(channel).subscribe(handler)` site inside\n * an actor (Stower, Gatherer, Matcher, Browser, Smelter), to attribute\n * each in-process subscriber's work to a span without scattering manual\n * `withSpan` calls across handler bodies.\n *\n * The span's parent is the active context at the time the handler\n * fires — which is the `bus.dispatch:<channel>` span on the gateway\n * (Subject.next runs synchronously inside the dispatch span), or the\n * `bus.emit:<channel>` span when an actor emits to itself.\n */\nexport async function withActorSpan<T>(\n actor: string,\n channel: string,\n fn: (span: Span) => Promise<T> | T,\n extraAttrs?: Attributes,\n): Promise<T> {\n const start = performance.now();\n try {\n return await withSpan(`actor.${actor}:${channel}`, fn, {\n kind: SpanKind.CONSUMER,\n attrs: {\n actor,\n 'bus.channel': channel,\n ...(extraAttrs ?? {}),\n },\n });\n } finally {\n recordHandlerDuration(actor, channel, performance.now() - start);\n }\n}\n\n// ── Log correlation ────────────────────────────────────────────────────\n\n/**\n * Read the active span's `trace_id` / `span_id` for log-line correlation.\n * Tier 3 of `.plans/OBSERVABILITY.md`. Each structured log line gets\n * tagged with these so a log query in CloudWatch / Loki / Datadog can\n * jump to the trace in Tempo / Jaeger / X-Ray.\n *\n * Returns `undefined` if no span is active, or if the active span's\n * context is invalid (uninitialized SDK, no-op tracer).\n */\nexport function getLogTraceContext(): { trace_id: string; span_id: string } | undefined {\n const span = trace.getActiveSpan();\n if (!span) return undefined;\n const ctx = span.spanContext();\n if (!isSpanContextValid(ctx)) return undefined;\n return { trace_id: ctx.traceId, span_id: ctx.spanId };\n}\n\n// ── Metrics — Tier 3 ───────────────────────────────────────────────────\n\nconst METER_NAME = 'semiont';\n\nconst meter = () => metrics.getMeter(METER_NAME);\n\nlet _busEmitCounter: Counter | undefined;\nlet _replySuppressedCounter: Counter | undefined;\nlet _resumeGapCounter: Counter | undefined;\nlet _unanswerableCounter: Counter | undefined;\nlet _correlationRegistryGauge: ObservableGauge | undefined;\nlet _correlationRegistryProvider: (() => CorrelationRegistrySnapshot) | undefined;\nlet _handlerDurationHistogram: Histogram | undefined;\nlet _jobOutcomeCounter: Counter | undefined;\nlet _jobDurationHistogram: Histogram | undefined;\nlet _gatherDegradeCounter: Counter | undefined;\nlet _inferenceCallsCounter: Counter | undefined;\nlet _inferenceTokensCounter: Counter | undefined;\nlet _inferenceDurationHistogram: Histogram | undefined;\nlet _sseSubscribers: UpDownCounter | undefined;\nlet _jobQueueGauge: ObservableGauge | undefined;\nlet _jobQueueProvider: (() => Promise<JobQueueSnapshot> | JobQueueSnapshot) | undefined;\nlet _vectorIndexSizeGauge: ObservableGauge | undefined;\nlet _factPumpDepthGauge: ObservableGauge | undefined;\nlet _factPumpDepthProvider: (() => number) | undefined;\nlet _vectorIndexSizeProvider: (() => Promise<number> | number) | undefined;\n\n/** Snapshot of job-queue contents by status. Match `JobQueue.getStats()`. */\nexport interface JobQueueSnapshot {\n pending: number;\n running: number;\n complete: number;\n failed: number;\n cancelled: number;\n}\n\nfunction busEmitCounter(): Counter {\n if (!_busEmitCounter) {\n _busEmitCounter = meter().createCounter('semiont.bus.emit', {\n description: 'Bus emits by channel and scope',\n });\n }\n return _busEmitCounter;\n}\n\nfunction handlerDurationHistogram(): Histogram {\n if (!_handlerDurationHistogram) {\n _handlerDurationHistogram = meter().createHistogram('semiont.handler.duration', {\n description: 'In-process actor handler duration',\n unit: 'ms',\n });\n }\n return _handlerDurationHistogram;\n}\n\nfunction jobOutcomeCounter(): Counter {\n if (!_jobOutcomeCounter) {\n _jobOutcomeCounter = meter().createCounter('semiont.job.outcome', {\n description: 'Worker job completions by type and outcome',\n });\n }\n return _jobOutcomeCounter;\n}\n\nfunction jobDurationHistogram(): Histogram {\n if (!_jobDurationHistogram) {\n _jobDurationHistogram = meter().createHistogram('semiont.job.duration', {\n description: 'Worker job duration by type',\n unit: 'ms',\n });\n }\n return _jobDurationHistogram;\n}\n\nfunction inferenceCallsCounter(): Counter {\n if (!_inferenceCallsCounter) {\n _inferenceCallsCounter = meter().createCounter('semiont.inference.calls', {\n description: 'Inference API calls by provider, model, and outcome',\n });\n }\n return _inferenceCallsCounter;\n}\n\nfunction inferenceTokensCounter(): Counter {\n if (!_inferenceTokensCounter) {\n _inferenceTokensCounter = meter().createCounter('semiont.inference.tokens', {\n description: 'Inference token usage by provider, model, and direction',\n });\n }\n return _inferenceTokensCounter;\n}\n\nfunction inferenceDurationHistogram(): Histogram {\n if (!_inferenceDurationHistogram) {\n _inferenceDurationHistogram = meter().createHistogram('semiont.inference.duration', {\n description: 'Inference call duration by provider, model, and outcome',\n unit: 'ms',\n });\n }\n return _inferenceDurationHistogram;\n}\n\nfunction sseSubscribersCounter(): UpDownCounter {\n if (!_sseSubscribers) {\n _sseSubscribers = meter().createUpDownCounter('semiont.sse.subscribers', {\n description: 'Active SSE subscribers',\n });\n }\n return _sseSubscribers;\n}\n\nfunction replySuppressedCounter(): Counter {\n if (!_replySuppressedCounter) {\n _replySuppressedCounter = meter().createCounter('semiont.bus.reply.suppressed', {\n description: 'Correlated replies withheld from a non-owning subscriber',\n });\n }\n return _replySuppressedCounter;\n}\n\n/**\n * A correlated reply was withheld from a subscriber that does not own its\n * correlationId (CORRELATED-REPLY-ROUTING P5).\n *\n * Counts ONLY that case. A frame with no correlationId is a shape violation\n * (warned, not counted), and a cid nobody claimed is the structural in-process\n * case that fires constantly — counting either would drown the signal this\n * metric exists to show: the fan-out amplification the delivery filter removes.\n */\nexport function recordReplySuppressed(channel: string): void {\n replySuppressedCounter().add(1, { 'bus.channel': channel });\n}\n\nfunction resumeGapCounter(): Counter {\n if (!_resumeGapCounter) {\n _resumeGapCounter = meter().createCounter('semiont.bus.resume_gap', {\n description: 'SSE resumes that degraded to a gap because replay was unavailable',\n });\n }\n return _resumeGapCounter;\n}\n\n/**\n * An SSE resume could not be served and the client was told to fall back to\n * cache. This degradation is CORRECT by design and therefore silent — which is\n * exactly why it needs a number. A rising rate means clients are losing\n * history, and nothing else in the stack says so.\n */\nexport function recordResumeGap(reason: string): void {\n resumeGapCounter().add(1, { 'bus.resume_gap.reason': reason });\n}\n\nfunction unanswerableCounter(): Counter {\n if (!_unanswerableCounter) {\n _unanswerableCounter = meter().createCounter('semiont.bus.unanswerable', {\n description: 'Request emits that reached zero subscribers and were failed at the gateway',\n });\n }\n return _unanswerableCounter;\n}\n\n/**\n * A request-shaped emit reached no subscriber, so the gateway synthesized its\n * mapped failure (ARCHIVIST-STAYS-UP P3). By channel, this is the absence rate\n * of the service that answers it — the difference between \"it went down once\"\n * and \"it is flapping.\"\n */\nexport function recordUnanswerableRequest(channel: string): void {\n unanswerableCounter().add(1, { 'bus.channel': channel });\n}\n\n/** Claims held and reply payloads retained by a gateway's correlation registry. */\nexport interface CorrelationRegistrySnapshot {\n claims: number;\n retainedReplies: number;\n}\n\n/**\n * Register a callback returning the gateway's correlation-registry occupancy.\n *\n * COUNTS, not bytes: retention is count-budgeted today (byte-budgeting is a\n * known limit in CORRELATED-REPLY-ROUTING), so `retainedReplies` is a proxy for\n * heap, not a measure of it. It is still the closest observable to the question\n * two OOM investigations keep asking — a browse result can be 1-2 MB, and up to\n * REPLY_RETENTION_MAX of them are held at once.\n */\nexport function registerCorrelationRegistryProvider(\n provider: () => CorrelationRegistrySnapshot,\n): void {\n _correlationRegistryProvider = provider;\n if (!_correlationRegistryGauge) {\n _correlationRegistryGauge = meter().createObservableGauge('semiont.bus.correlation.size', {\n description: 'Correlation registry occupancy: live claims and retained reply payloads',\n });\n _correlationRegistryGauge.addCallback((observer) => {\n if (!_correlationRegistryProvider) return;\n const snap = _correlationRegistryProvider();\n observer.observe(snap.claims, { 'correlation.kind': 'claims' });\n observer.observe(snap.retainedReplies, { 'correlation.kind': 'retained_replies' });\n });\n }\n}\n\n/** Increment the bus-emit counter. Called at every transport `emit` site. */\nexport function recordBusEmit(channel: string, scope?: string): void {\n busEmitCounter().add(1, {\n 'bus.channel': channel,\n ...(scope ? { 'bus.scope': scope } : {}),\n });\n}\n\n/** Record an in-process actor handler's duration. */\nexport function recordHandlerDuration(actor: string, channel: string, durationMs: number): void {\n handlerDurationHistogram().record(durationMs, {\n actor,\n 'bus.channel': channel,\n });\n}\n\n/** Record a worker job's outcome and duration. */\nexport function recordJobOutcome(jobType: string, outcome: 'completed' | 'failed', durationMs: number): void {\n jobOutcomeCounter().add(1, { 'job.type': jobType, 'job.outcome': outcome });\n jobDurationHistogram().record(durationMs, { 'job.type': jobType, 'job.outcome': outcome });\n}\n\nlet _appendStageHistogram: Histogram | undefined;\nfunction appendStageHistogram(): Histogram {\n if (!_appendStageHistogram) {\n _appendStageHistogram = meter().createHistogram('semiont.record.append.duration', {\n description: 'Time spent in one stage of appending an event to the record, labeled by stage: persist (JSONL write + git), materialize (view rebuild), enrich, publish. The Archivist\\'s core write path.',\n unit: 'ms',\n });\n }\n return _appendStageHistogram;\n}\n\n/**\n * Record one stage of `EventStore.appendEvent` (ARCHIVIST-STAYS-UP P7).\n *\n * The append path is the one operation only the Archivist can perform, and it\n * was entirely dark: reads had `recordHandlerDuration` and the bus had its own\n * counters, while writes had nothing. Stage-labeled because the useful\n * question is never \"was the append slow\" but WHICH PART — and `materialize`\n * in particular does work proportional to a resource's annotation count, so it\n * degrades with history rather than with load.\n */\nexport function recordAppendStage(\n stage: 'persist' | 'materialize' | 'enrich' | 'publish',\n durationMs: number,\n): void {\n appendStageHistogram().record(durationMs, { 'record.stage': stage });\n}\n\nlet _gitCommandHistogram: Histogram | undefined;\nfunction gitCommandHistogram(): Histogram {\n if (!_gitCommandHistogram) {\n _gitCommandHistogram = meter().createHistogram('semiont.git.duration', {\n description: 'Time spent in a synchronous git subprocess. These run on the event loop, so this duration is also time no other request could be served.',\n unit: 'ms',\n });\n }\n return _gitCommandHistogram;\n}\n\n/**\n * Record a synchronous git invocation (ARCHIVIST-STAYS-UP P7).\n *\n * These are `execFileSync`, so **the duration is event-loop blockage, not just\n * latency** — every concurrent `browse:*` read waits behind it. One `git add`\n * runs per appended event, so a detection job writing hundreds of annotations\n * spawns hundreds of blocking subprocesses. That is the suspected mechanism\n * behind \"reads serializing behind the detection job's annotation writes\" in\n * `bugs/absent-archivist-wedges-browse.md`, which recorded the symptom without\n * a cause. This number is what turns that from a hypothesis into a reading.\n */\nexport function recordGitCommand(command: string, durationMs: number): void {\n gitCommandHistogram().record(durationMs, { 'git.command': command });\n}\n\nfunction gatherDegradeCounter(): Counter {\n if (!_gatherDegradeCounter) {\n _gatherDegradeCounter = meter().createCounter('semiont.gather.degraded', {\n description: 'Gathers that degraded because an eventually-consistent projection did not catch up within its read barrier (vectors: absent semanticContext; graph: projection-lag failure). Labeled by projection.',\n });\n }\n return _gatherDegradeCounter;\n}\n\n/**\n * Record a gather degraded by a projection read barrier: `'vectors'` — the\n * Smelter settle barrier timed out (semanticContext shipped absent);\n * `'graph'` — the Weaver applied barrier + poll floor exhausted (projection\n * lag surfaced as a distinct failure). Fleet-alertable counterpart of the\n * `[gather DEGRADED]` L4 breadcrumbs — a rising rate on either label means\n * that pipeline is not keeping up.\n */\nexport function recordGatherDegrade(projection: 'graph' | 'vectors'): void {\n gatherDegradeCounter().add(1, { projection });\n}\n\n/** Increment the SSE subscriber gauge — call on `/bus/subscribe` open. */\nexport function recordSubscriberConnect(): void {\n sseSubscribersCounter().add(1);\n}\n\n/** Decrement on disconnect. Pair with `recordSubscriberConnect`. */\nexport function recordSubscriberDisconnect(): void {\n sseSubscribersCounter().add(-1);\n}\n\n/**\n * Register a callback that returns the current job-queue snapshot.\n * Polled at the SDK's metric-collection interval. The single gauge\n * emits one observation per status (`pending`, `running`, …) tagged\n * with the `job.status` attribute. Idempotent — last registered\n * provider wins.\n */\nexport function registerJobQueueProvider(\n provider: () => Promise<JobQueueSnapshot> | JobQueueSnapshot,\n): void {\n _jobQueueProvider = provider;\n if (!_jobQueueGauge) {\n _jobQueueGauge = meter().createObservableGauge('semiont.job.queue.size', {\n description: 'Job queue size by status',\n });\n _jobQueueGauge.addCallback(async (observer) => {\n if (!_jobQueueProvider) return;\n const snap = await _jobQueueProvider();\n observer.observe(snap.pending, { 'job.status': 'pending' });\n observer.observe(snap.running, { 'job.status': 'running' });\n observer.observe(snap.complete, { 'job.status': 'complete' });\n observer.observe(snap.failed, { 'job.status': 'failed' });\n observer.observe(snap.cancelled, { 'job.status': 'cancelled' });\n });\n }\n}\n\n/**\n * Register a callback that returns the current vector-index size\n * (point count). Async to allow remote queries (Qdrant). Polled at\n * the metric-collection interval.\n */\n/**\n * Register the Archivist's fact-pump backlog — facts appended to the record\n * but not yet republished onto the bus.\n *\n * At rest this is zero. A value that climbs and does not come back means the\n * pump is outrunning its transport, which is the leading hypothesis for the\n * load-correlated heap growth in `bugs/absent-archivist-wedges-browse.md`\n * (ARCHIVIST-STAYS-UP P5). The backlog is deliberately unbounded today, so\n * this number is the only thing standing between \"the pump is behind\" and an\n * OOM whose cause is inferred from RSS after the fact.\n */\nexport function registerFactPumpDepthProvider(provider: () => number): void {\n _factPumpDepthProvider = provider;\n if (!_factPumpDepthGauge) {\n _factPumpDepthGauge = meter().createObservableGauge('semiont.archivist.fact_pump.depth', {\n description: 'Facts appended to the record but not yet published to the bus. Zero at rest; a rising floor means the pump is behind its transport.',\n });\n _factPumpDepthGauge.addCallback((observer) => {\n if (_factPumpDepthProvider) observer.observe(_factPumpDepthProvider());\n });\n }\n}\n\nexport function registerVectorIndexSizeProvider(\n provider: () => Promise<number> | number,\n): void {\n _vectorIndexSizeProvider = provider;\n if (!_vectorIndexSizeGauge) {\n _vectorIndexSizeGauge = meter().createObservableGauge('semiont.vector.index.size', {\n description: 'Vector store point count',\n });\n _vectorIndexSizeGauge.addCallback(async (observer) => {\n if (_vectorIndexSizeProvider) {\n const value = await _vectorIndexSizeProvider();\n observer.observe(value);\n }\n });\n }\n}\n\n/**\n * Record an inference call. Token counts are optional — providers that\n * don't expose them (or fail before generating) record only call count\n * and duration.\n */\nexport function recordInferenceUsage(opts: {\n provider: string;\n model: string;\n durationMs: number;\n outcome: 'success' | 'error';\n inputTokens?: number;\n outputTokens?: number;\n}): void {\n const baseAttrs = {\n 'inference.provider': opts.provider,\n 'inference.model': opts.model,\n 'inference.outcome': opts.outcome,\n };\n inferenceCallsCounter().add(1, baseAttrs);\n inferenceDurationHistogram().record(opts.durationMs, baseAttrs);\n if (opts.inputTokens != null && opts.inputTokens > 0) {\n inferenceTokensCounter().add(opts.inputTokens, {\n 'inference.provider': opts.provider,\n 'inference.model': opts.model,\n 'inference.direction': 'input',\n });\n }\n if (opts.outputTokens != null && opts.outputTokens > 0) {\n inferenceTokensCounter().add(opts.outputTokens, {\n 'inference.provider': opts.provider,\n 'inference.model': opts.model,\n 'inference.direction': 'output',\n });\n }\n}\n\nlet _detectionCallCounter: Counter | undefined;\nlet _detectionDurationHistogram: Histogram | undefined;\nlet _detectionItemsHistogram: Histogram | undefined;\nlet _detectionTokensHistogram: Histogram | undefined;\n\nfunction detectionCallCounter(): Counter {\n if (!_detectionCallCounter) {\n _detectionCallCounter = meter().createCounter('semiont.detection.calls', {\n description: 'Detection model calls, labeled by motivation, outcome, subdivision depth and whether this was the floor re-roll.',\n });\n }\n return _detectionCallCounter;\n}\n\nfunction detectionDurationHistogram(): Histogram {\n if (!_detectionDurationHistogram) {\n _detectionDurationHistogram = meter().createHistogram('semiont.detection.call.duration', {\n description: 'Wall time of one detection model call, including the attempts that failed and were retried smaller.',\n unit: 'ms',\n });\n }\n return _detectionDurationHistogram;\n}\n\nfunction detectionItemsHistogram(): Histogram {\n if (!_detectionItemsHistogram) {\n _detectionItemsHistogram = meter().createHistogram('semiont.detection.call.items', {\n description: 'Annotations returned by one detection call. Against the input size on the same record, this is yield.',\n });\n }\n return _detectionItemsHistogram;\n}\n\nfunction detectionTokensHistogram(): Histogram {\n if (!_detectionTokensHistogram) {\n _detectionTokensHistogram = meter().createHistogram('semiont.detection.call.tokens', {\n description: \"Provider-reported tokens for one detection call, by direction. Deliberately separate from semiont.inference.tokens: that series is the authoritative total but carries no subdivision depth, and 'what does a depth-2 call cost' is the question every sizing decision asks.\",\n });\n }\n return _detectionTokensHistogram;\n}\n\n/**\n * Record one detection model call (DETECTION-QUALITY-THROUGHPUT P1).\n *\n * The adapters already record provider/model/duration/tokens for every\n * inference call. What they cannot know is the detection shape around it:\n * which motivation asked, how big the piece was, how many annotations came\n * back, how deep subdivision had descended, and whether this was the floor\n * re-roll. Those are the facts that distinguish a healthy call from a\n * expensive descent, and without them a slow detection run is one\n * undifferentiated number.\n *\n * FAILED attempts are recorded too, and that is the point: the calls paid for\n * and thrown away during a descent are exactly the cost later phases exist to\n * avoid, so a record only of successes would hide the thing being optimized.\n *\n * Tokens are the PROVIDER's counts, passed through — never estimated. Absent\n * means the provider did not report them.\n */\nexport function recordDetectionCall(opts: {\n label: string;\n pieceChars: number;\n durationMs: number;\n items: number;\n depth: number;\n reroll: boolean;\n outcome: 'success' | 'truncated' | 'timeout' | 'collapsed' | 'error';\n inputTokens?: number;\n outputTokens?: number;\n}): void {\n const attrs = {\n 'detection.label': opts.label,\n 'detection.outcome': opts.outcome,\n 'detection.depth': opts.depth,\n 'detection.reroll': opts.reroll,\n };\n detectionCallCounter().add(1, attrs);\n detectionDurationHistogram().record(opts.durationMs, attrs);\n detectionItemsHistogram().record(opts.items, attrs);\n if (opts.inputTokens !== undefined) {\n detectionTokensHistogram().record(opts.inputTokens, { ...attrs, 'detection.direction': 'input' });\n }\n if (opts.outputTokens !== undefined) {\n detectionTokensHistogram().record(opts.outputTokens, { ...attrs, 'detection.direction': 'output' });\n }\n}\n\nlet _anchorOutcomeCounter: Counter | undefined;\nfunction anchorOutcomeCounter(): Counter {\n if (!_anchorOutcomeCounter) {\n _anchorOutcomeCounter = meter().createCounter('semiont.detection.anchors', {\n description: 'Every annotation anchoring, labeled by the method that resolved it. EVERY outcome is counted, not just the risky ones, because a bare count of degraded anchors has no denominator — the rate is the precision signal.',\n });\n }\n return _anchorOutcomeCounter;\n}\n\n/**\n * Record how one annotation got anchored (DETECTION-QUALITY-THROUGHPUT P5).\n *\n * The selector-vs-source check is already a WRITE-TIME INVARIANT — both\n * `buildTextAnnotation` and `buildPdfAnnotation` throw on a selector that does\n * not match its source — so mechanical correctness is guaranteed rather than\n * sampled, and auditing it would measure a constant.\n *\n * What is genuinely uncertain is which anchoring METHOD got there. An `exact`\n * the model quoted verbatim and that appears once is certain; one resolved by\n * `first-of-many` (several occurrences, no usable context) or `fuzzy-match`\n * picked a plausible occurrence and may have picked wrong. Those were visible\n * only as log warnings — countable by a human reading worker output, which is\n * how 47 of them went unreviewed. As a rate they are the precision number that\n * sits beside the yield numbers.\n */\nexport function recordAnchorOutcome(label: string, method: string): void {\n anchorOutcomeCounter().add(1, { 'detection.label': label, 'anchor.method': method });\n}\n\n// ── Re-exports from @opentelemetry/api ─────────────────────────────────\n\nexport { SpanKind, SpanStatusCode, type Attributes, type Span } from '@opentelemetry/api';\n","/**\n * Process-level structured logger for Node entry points.\n *\n * Used by long-lived Node processes (gateway, workers, smelter) that\n * want JSON-structured stdout with active-span trace correlation. The\n * `trace_id` / `span_id` fields are populated from the current OTel\n * span context via `getLogTraceContext` — this is the same Tier 3\n * correlation that lets a grep through stdout line up with the trace\n * UI without manual stitching.\n *\n * Reads `LOG_LEVEL` (default `info`) and `LOG_FORMAT` (`json` default,\n * `simple` for human-friendly dev output).\n *\n * Co-located with `getLogTraceContext` deliberately: this is the\n * only reasonably-shaped consumer of that helper, and putting them in\n * the same package keeps the trace-id wiring in one place.\n */\n\nimport winston from 'winston';\nimport type { Logger } from '@semiont/core';\nimport { getLogTraceContext } from './index.js';\n\nconst traceContextFormat = winston.format((info) => {\n const trace = getLogTraceContext();\n if (trace) {\n info.trace_id = trace.trace_id;\n info.span_id = trace.span_id;\n }\n return info;\n})();\n\nexport function createProcessLogger(component: string): Logger {\n const level = process.env.LOG_LEVEL ?? 'info';\n const format = process.env.LOG_FORMAT === 'simple'\n ? winston.format.combine(\n winston.format.timestamp({ format: 'YYYY-MM-DD HH:mm:ss' }),\n winston.format.errors({ stack: true }),\n traceContextFormat,\n winston.format.printf(({ level: lvl, message, timestamp, ...meta }) => {\n const metaStr = Object.keys(meta).length > 0 ? ` ${JSON.stringify(meta)}` : '';\n return `${timestamp} [${lvl.toUpperCase()}] [${component}] ${message}${metaStr}`;\n }),\n )\n : winston.format.combine(\n winston.format.timestamp(),\n winston.format.errors({ stack: true }),\n traceContextFormat,\n winston.format.json(),\n );\n\n const logger = winston.createLogger({\n level,\n defaultMeta: { component },\n format,\n transports: [new winston.transports.Console()],\n });\n\n return logger;\n}\n"]}
1
+ {"version":3,"sources":["../src/index.ts","../src/process-logger.ts"],"names":["trace"],"mappings":";;;;;AAqDA,wBAAA,CAAyB,MAAM;AAC7B,EAAA,MAAM,IAAA,GAAO,MAAM,aAAA,EAAc;AACjC,EAAA,IAAI,CAAC,MAAM,OAAO,MAAA;AAClB,EAAA,MAAM,GAAA,GAAM,KAAK,WAAA,EAAY;AAC7B,EAAA,IAAI,CAAC,kBAAA,CAAmB,GAAG,CAAA,EAAG,OAAO,MAAA;AACrC,EAAA,OAAO,GAAA,CAAI,OAAA;AACb,CAAC,CAAA;AA+JM,SAAS,kBAAA,GAAwE;AACtF,EAAA,MAAM,IAAA,GAAO,MAAM,aAAA,EAAc;AACjC,EAAA,IAAI,CAAC,MAAM,OAAO,MAAA;AAClB,EAAA,MAAM,GAAA,GAAM,KAAK,WAAA,EAAY;AAC7B,EAAA,IAAI,CAAC,kBAAA,CAAmB,GAAG,CAAA,EAAG,OAAO,MAAA;AACrC,EAAA,OAAO,EAAE,QAAA,EAAU,GAAA,CAAI,OAAA,EAAS,OAAA,EAAS,IAAI,MAAA,EAAO;AACtD;;;AC1MA,IAAM,kBAAA,GAAqB,OAAA,CAAQ,MAAA,CAAO,CAAC,IAAA,KAAS;AAClD,EAAA,MAAMA,SAAQ,kBAAA,EAAmB;AACjC,EAAA,IAAIA,MAAAA,EAAO;AACT,IAAA,IAAA,CAAK,WAAWA,MAAAA,CAAM,QAAA;AACtB,IAAA,IAAA,CAAK,UAAUA,MAAAA,CAAM,OAAA;AAAA,EACvB;AACA,EAAA,OAAO,IAAA;AACT,CAAC,CAAA,EAAE;AAEI,SAAS,oBAAoB,SAAA,EAA2B;AAC7D,EAAA,MAAM,KAAA,GAAQ,OAAA,CAAQ,GAAA,CAAI,SAAA,IAAa,MAAA;AACvC,EAAA,MAAM,SAAS,OAAA,CAAQ,GAAA,CAAI,UAAA,KAAe,QAAA,GACtC,QAAQ,MAAA,CAAO,OAAA;AAAA,IACb,QAAQ,MAAA,CAAO,SAAA,CAAU,EAAE,MAAA,EAAQ,uBAAuB,CAAA;AAAA,IAC1D,QAAQ,MAAA,CAAO,MAAA,CAAO,EAAE,KAAA,EAAO,MAAM,CAAA;AAAA,IACrC,kBAAA;AAAA,IACA,OAAA,CAAQ,MAAA,CAAO,MAAA,CAAO,CAAC,EAAE,KAAA,EAAO,GAAA,EAAK,OAAA,EAAS,SAAA,EAAW,GAAG,IAAA,EAAK,KAAM;AACrE,MAAA,MAAM,OAAA,GAAU,MAAA,CAAO,IAAA,CAAK,IAAI,CAAA,CAAE,MAAA,GAAS,CAAA,GAAI,CAAA,CAAA,EAAI,IAAA,CAAK,SAAA,CAAU,IAAI,CAAC,CAAA,CAAA,GAAK,EAAA;AAC5E,MAAA,OAAO,CAAA,EAAG,SAAS,CAAA,EAAA,EAAK,GAAA,CAAI,WAAA,EAAa,CAAA,GAAA,EAAM,SAAS,CAAA,EAAA,EAAK,OAAO,CAAA,EAAG,OAAO,CAAA,CAAA;AAAA,IAChF,CAAC;AAAA,GACH,GACA,QAAQ,MAAA,CAAO,OAAA;AAAA,IACb,OAAA,CAAQ,OAAO,SAAA,EAAU;AAAA,IACzB,QAAQ,MAAA,CAAO,MAAA,CAAO,EAAE,KAAA,EAAO,MAAM,CAAA;AAAA,IACrC,kBAAA;AAAA,IACA,OAAA,CAAQ,OAAO,IAAA;AAAK,GACtB;AAEJ,EAAA,MAAM,MAAA,GAAS,QAAQ,YAAA,CAAa;AAAA,IAClC,KAAA;AAAA,IACA,WAAA,EAAa,EAAE,SAAA,EAAU;AAAA,IACzB,MAAA;AAAA,IACA,YAAY,CAAC,IAAI,OAAA,CAAQ,UAAA,CAAW,SAAS;AAAA,GAC9C,CAAA;AAED,EAAA,OAAO,MAAA;AACT","file":"process-logger.js","sourcesContent":["/**\n * @semiont/observability — public API.\n *\n * Universal surface (works in Node + browser). For SDK *initialization*,\n * import from `@semiont/observability/node` or `/web` at the process entry\n * point. Everything else uses this module.\n *\n * Tier 2 of `.plans/OBSERVABILITY.md`. The public surface:\n *\n * - `withSpan(name, fn, options?)` — wrap an async block in a span;\n * `options` carries `kind` and `attrs`.\n * - `withActorSpan(actor, channel, fn, extraAttrs?)` — consumer-span\n * wrapper for bus-event handlers, with handler-duration recording.\n * - `injectTraceparent(payload)` / `extractTraceparent(payload)` — W3C\n * trace-context propagation across the SSE channel (the bus payload\n * gets a `_trace?: { traceparent }` sibling to `correlationId`).\n * - `withTraceparent(carrier, fn)` — run `fn` with the incoming\n * traceparent as the parent context.\n * - `getActiveTraceparent()` — read the active span's traceparent for\n * manual propagation (e.g. attaching to a fetch header or SSE field).\n * - `getLogTraceContext()` — active `trace_id` / `span_id` for log-line\n * correlation.\n * - Metric recorders (`recordBusEmit`, `recordHandlerDuration`,\n * `recordJobOutcome`, `recordSubscriberConnect` / `Disconnect`,\n * `recordInferenceUsage`) and gauge providers\n * (`registerJobQueueProvider`, `registerVectorIndexSizeProvider`).\n *\n * No-op when no exporter is configured: `@opentelemetry/api`'s default\n * tracer is a no-op, so `withSpan` is essentially free until\n * `initObservability*()` runs.\n */\n\nimport {\n context,\n isSpanContextValid,\n metrics,\n propagation,\n SpanKind,\n SpanStatusCode,\n trace,\n type Attributes,\n type Counter,\n type Histogram,\n type ObservableGauge,\n type Span,\n type UpDownCounter,\n} from '@opentelemetry/api';\nimport { setBusLogTraceIdProvider } from '@semiont/core';\n\n// Wire `busLog`'s trace-id provider once at module load. When an OTel\n// SDK is initialized (and a span is active when `busLog` fires), the\n// emitted line gets a `trace=<8hex>` suffix that correlates the\n// grep-timeline with the trace UI. No-op when no SDK is active.\nsetBusLogTraceIdProvider(() => {\n const span = trace.getActiveSpan();\n if (!span) return undefined;\n const ctx = span.spanContext();\n if (!isSpanContextValid(ctx)) return undefined;\n return ctx.traceId;\n});\n\nconst TRACER_NAME = 'semiont';\n\nconst tracer = () => trace.getTracer(TRACER_NAME);\n\n// ── withSpan ───────────────────────────────────────────────────────────\n\n/**\n * Wrap an async block in a span. The span is started before `fn` runs and\n * ended after it resolves or rejects; exceptions are recorded and the span\n * status is set to ERROR. `kind` defaults to INTERNAL.\n */\nexport async function withSpan<T>(\n name: string,\n fn: (span: Span) => Promise<T> | T,\n options?: { kind?: SpanKind; attrs?: Attributes },\n): Promise<T> {\n const span = tracer().startSpan(name, {\n kind: options?.kind ?? SpanKind.INTERNAL,\n ...(options?.attrs ? { attributes: options.attrs } : {}),\n });\n try {\n return await context.with(trace.setSpan(context.active(), span), () => fn(span));\n } catch (err) {\n span.recordException(err as Error);\n span.setStatus({\n code: SpanStatusCode.ERROR,\n message: err instanceof Error ? err.message : String(err),\n });\n throw err;\n } finally {\n span.end();\n }\n}\n\n// ── Traceparent on bus payloads ────────────────────────────────────────\n\nconst TRACE_FIELD = '_trace';\n\n/**\n * Sibling of `correlationId` on bus payloads. Lives on the SSE event body\n * because SSE has no header trailer; the SDK strips it before delivering\n * the payload to subscribers. Additive — payloads without `_trace` parse\n * unchanged.\n */\nexport interface TraceCarrier {\n /** W3C `traceparent` header value (`00-<traceId>-<spanId>-<flags>`). */\n traceparent: string;\n /** W3C `tracestate` header value (vendor-specific extensions). */\n tracestate?: string;\n}\n\n/**\n * Read the active span's W3C traceparent (and tracestate). Returns\n * `undefined` if no span is active.\n */\nexport function getActiveTraceparent(): TraceCarrier | undefined {\n const carrier: Record<string, string> = {};\n propagation.inject(context.active(), carrier);\n const traceparent = carrier['traceparent'];\n if (!traceparent) return undefined;\n return carrier['tracestate']\n ? { traceparent, tracestate: carrier['tracestate'] }\n : { traceparent };\n}\n\n/**\n * Attach the active span's trace-context to a payload object as\n * `_trace`. No-op when no span is active. Returns the same object\n * reference for chaining.\n */\nexport function injectTraceparent<T extends Record<string, unknown>>(payload: T): T {\n const carrier = getActiveTraceparent();\n if (carrier) {\n (payload as Record<string, unknown>)[TRACE_FIELD] = carrier;\n }\n return payload;\n}\n\n/**\n * Strip and return the `_trace` field from a payload. Mutates `payload`.\n * The field is internal plumbing and should not be visible to subscribers.\n */\nexport function extractTraceparent<T extends Record<string, unknown>>(\n payload: T,\n): TraceCarrier | undefined {\n const carrier = (payload as Record<string, unknown>)[TRACE_FIELD] as\n | TraceCarrier\n | undefined;\n if (carrier !== undefined) {\n delete (payload as Record<string, unknown>)[TRACE_FIELD];\n }\n if (!carrier || typeof carrier.traceparent !== 'string') return undefined;\n return carrier;\n}\n\n/**\n * Run `fn` with the given W3C traceparent set as the parent context.\n * Any spans started inside `fn` will be children of the incoming trace.\n * No-op if `carrier` is undefined.\n */\nexport function withTraceparent<T>(\n carrier: TraceCarrier | undefined,\n fn: () => T,\n): T {\n if (!carrier) return fn();\n const carrierObj: Record<string, string> = { traceparent: carrier.traceparent };\n if (carrier.tracestate) carrierObj['tracestate'] = carrier.tracestate;\n const ctx = propagation.extract(context.active(), carrierObj);\n return context.with(ctx, fn);\n}\n\n// ── Actor handler convenience ──────────────────────────────────────────\n\n/**\n * Wrap a bus-event handler in an `actor.<name>:<channel>` consumer span.\n * Used at every `eventBus.get(channel).subscribe(handler)` site inside\n * an actor (Stower, Gatherer, Matcher, Browser, Smelter), to attribute\n * each in-process subscriber's work to a span without scattering manual\n * `withSpan` calls across handler bodies.\n *\n * The span's parent is the active context at the time the handler\n * fires — which is the `bus.dispatch:<channel>` span on the gateway\n * (Subject.next runs synchronously inside the dispatch span), or the\n * `bus.emit:<channel>` span when an actor emits to itself.\n */\nexport async function withActorSpan<T>(\n actor: string,\n channel: string,\n fn: (span: Span) => Promise<T> | T,\n extraAttrs?: Attributes,\n): Promise<T> {\n const start = performance.now();\n try {\n return await withSpan(`actor.${actor}:${channel}`, fn, {\n kind: SpanKind.CONSUMER,\n attrs: {\n actor,\n 'bus.channel': channel,\n ...(extraAttrs ?? {}),\n },\n });\n } finally {\n recordHandlerDuration(actor, channel, performance.now() - start);\n }\n}\n\n// ── Log correlation ────────────────────────────────────────────────────\n\n/**\n * Read the active span's `trace_id` / `span_id` for log-line correlation.\n * Tier 3 of `.plans/OBSERVABILITY.md`. Each structured log line gets\n * tagged with these so a log query in CloudWatch / Loki / Datadog can\n * jump to the trace in Tempo / Jaeger / X-Ray.\n *\n * Returns `undefined` if no span is active, or if the active span's\n * context is invalid (uninitialized SDK, no-op tracer).\n */\nexport function getLogTraceContext(): { trace_id: string; span_id: string } | undefined {\n const span = trace.getActiveSpan();\n if (!span) return undefined;\n const ctx = span.spanContext();\n if (!isSpanContextValid(ctx)) return undefined;\n return { trace_id: ctx.traceId, span_id: ctx.spanId };\n}\n\n// ── Metrics — Tier 3 ───────────────────────────────────────────────────\n\nconst METER_NAME = 'semiont';\n\nconst meter = () => metrics.getMeter(METER_NAME);\n\nlet _busEmitCounter: Counter | undefined;\nlet _replySuppressedCounter: Counter | undefined;\nlet _resumeGapCounter: Counter | undefined;\nlet _unanswerableCounter: Counter | undefined;\nlet _correlationRegistryGauge: ObservableGauge | undefined;\nlet _correlationRegistryProvider: (() => CorrelationRegistrySnapshot) | undefined;\nlet _handlerDurationHistogram: Histogram | undefined;\nlet _jobOutcomeCounter: Counter | undefined;\nlet _jobDurationHistogram: Histogram | undefined;\nlet _gatherDegradeCounter: Counter | undefined;\nlet _inferenceCallsCounter: Counter | undefined;\nlet _inferenceTokensCounter: Counter | undefined;\nlet _inferenceDurationHistogram: Histogram | undefined;\nlet _sseSubscribers: UpDownCounter | undefined;\nlet _jobQueueGauge: ObservableGauge | undefined;\nlet _jobQueueProvider: (() => Promise<JobQueueSnapshot> | JobQueueSnapshot) | undefined;\nlet _vectorIndexSizeGauge: ObservableGauge | undefined;\nlet _factPumpDepthGauge: ObservableGauge | undefined;\nlet _factPumpDepthProvider: (() => number) | undefined;\nlet _vectorIndexSizeProvider: (() => Promise<number> | number) | undefined;\n\n/** Snapshot of job-queue contents by status. Match `JobQueue.getStats()`. */\nexport interface JobQueueSnapshot {\n pending: number;\n running: number;\n complete: number;\n failed: number;\n cancelled: number;\n}\n\nfunction busEmitCounter(): Counter {\n if (!_busEmitCounter) {\n _busEmitCounter = meter().createCounter('semiont.bus.emit', {\n description: 'Bus emits by channel and scope',\n });\n }\n return _busEmitCounter;\n}\n\nfunction handlerDurationHistogram(): Histogram {\n if (!_handlerDurationHistogram) {\n _handlerDurationHistogram = meter().createHistogram('semiont.handler.duration', {\n description: 'In-process actor handler duration',\n unit: 'ms',\n });\n }\n return _handlerDurationHistogram;\n}\n\nfunction jobOutcomeCounter(): Counter {\n if (!_jobOutcomeCounter) {\n _jobOutcomeCounter = meter().createCounter('semiont.job.outcome', {\n description: 'Worker job completions by type and outcome',\n });\n }\n return _jobOutcomeCounter;\n}\n\nfunction jobDurationHistogram(): Histogram {\n if (!_jobDurationHistogram) {\n _jobDurationHistogram = meter().createHistogram('semiont.job.duration', {\n description: 'Worker job duration by type',\n unit: 'ms',\n });\n }\n return _jobDurationHistogram;\n}\n\nfunction inferenceCallsCounter(): Counter {\n if (!_inferenceCallsCounter) {\n _inferenceCallsCounter = meter().createCounter('semiont.inference.calls', {\n description: 'Inference API calls by provider, model, and outcome',\n });\n }\n return _inferenceCallsCounter;\n}\n\nfunction inferenceTokensCounter(): Counter {\n if (!_inferenceTokensCounter) {\n _inferenceTokensCounter = meter().createCounter('semiont.inference.tokens', {\n description: 'Inference token usage by provider, model, and direction',\n });\n }\n return _inferenceTokensCounter;\n}\n\nfunction inferenceDurationHistogram(): Histogram {\n if (!_inferenceDurationHistogram) {\n _inferenceDurationHistogram = meter().createHistogram('semiont.inference.duration', {\n description: 'Inference call duration by provider, model, and outcome',\n unit: 'ms',\n });\n }\n return _inferenceDurationHistogram;\n}\n\nfunction sseSubscribersCounter(): UpDownCounter {\n if (!_sseSubscribers) {\n _sseSubscribers = meter().createUpDownCounter('semiont.sse.subscribers', {\n description: 'Active SSE subscribers',\n });\n }\n return _sseSubscribers;\n}\n\nfunction replySuppressedCounter(): Counter {\n if (!_replySuppressedCounter) {\n _replySuppressedCounter = meter().createCounter('semiont.bus.reply.suppressed', {\n description: 'Correlated replies withheld from a non-owning subscriber',\n });\n }\n return _replySuppressedCounter;\n}\n\n/**\n * A correlated reply was withheld from a subscriber that does not own its\n * correlationId (CORRELATED-REPLY-ROUTING P5).\n *\n * Counts ONLY that case. A frame with no correlationId is a shape violation\n * (warned, not counted), and a cid nobody claimed is the structural in-process\n * case that fires constantly — counting either would drown the signal this\n * metric exists to show: the fan-out amplification the delivery filter removes.\n */\nexport function recordReplySuppressed(channel: string): void {\n replySuppressedCounter().add(1, { 'bus.channel': channel });\n}\n\nfunction resumeGapCounter(): Counter {\n if (!_resumeGapCounter) {\n _resumeGapCounter = meter().createCounter('semiont.bus.resume_gap', {\n description: 'SSE resumes that degraded to a gap because replay was unavailable',\n });\n }\n return _resumeGapCounter;\n}\n\n/**\n * An SSE resume could not be served and the client was told to fall back to\n * cache. This degradation is CORRECT by design and therefore silent — which is\n * exactly why it needs a number. A rising rate means clients are losing\n * history, and nothing else in the stack says so.\n */\nexport function recordResumeGap(reason: string): void {\n resumeGapCounter().add(1, { 'bus.resume_gap.reason': reason });\n}\n\nfunction unanswerableCounter(): Counter {\n if (!_unanswerableCounter) {\n _unanswerableCounter = meter().createCounter('semiont.bus.unanswerable', {\n description: 'Request emits that reached zero subscribers and were failed at the gateway',\n });\n }\n return _unanswerableCounter;\n}\n\n/**\n * A request-shaped emit reached no subscriber, so the gateway synthesized its\n * mapped failure (ARCHIVIST-STAYS-UP P3). By channel, this is the absence rate\n * of the service that answers it — the difference between \"it went down once\"\n * and \"it is flapping.\"\n */\nexport function recordUnanswerableRequest(channel: string): void {\n unanswerableCounter().add(1, { 'bus.channel': channel });\n}\n\n/** Claims held and reply payloads retained by a gateway's correlation registry. */\nexport interface CorrelationRegistrySnapshot {\n claims: number;\n retainedReplies: number;\n}\n\n/**\n * Register a callback returning the gateway's correlation-registry occupancy.\n *\n * COUNTS, not bytes: retention is count-budgeted today (byte-budgeting is a\n * known limit in CORRELATED-REPLY-ROUTING), so `retainedReplies` is a proxy for\n * heap, not a measure of it. It is still the closest observable to the question\n * two OOM investigations keep asking — a browse result can be 1-2 MB, and up to\n * REPLY_RETENTION_MAX of them are held at once.\n */\nexport function registerCorrelationRegistryProvider(\n provider: () => CorrelationRegistrySnapshot,\n): void {\n _correlationRegistryProvider = provider;\n if (!_correlationRegistryGauge) {\n _correlationRegistryGauge = meter().createObservableGauge('semiont.bus.correlation.size', {\n description: 'Correlation registry occupancy: live claims and retained reply payloads',\n });\n _correlationRegistryGauge.addCallback((observer) => {\n if (!_correlationRegistryProvider) return;\n const snap = _correlationRegistryProvider();\n observer.observe(snap.claims, { 'correlation.kind': 'claims' });\n observer.observe(snap.retainedReplies, { 'correlation.kind': 'retained_replies' });\n });\n }\n}\n\n/** Increment the bus-emit counter. Called at every transport `emit` site. */\nexport function recordBusEmit(channel: string, scope?: string): void {\n busEmitCounter().add(1, {\n 'bus.channel': channel,\n ...(scope ? { 'bus.scope': scope } : {}),\n });\n}\n\n/** Record an in-process actor handler's duration. */\nexport function recordHandlerDuration(actor: string, channel: string, durationMs: number): void {\n handlerDurationHistogram().record(durationMs, {\n actor,\n 'bus.channel': channel,\n });\n}\n\n/** Record a worker job's outcome and duration. */\nexport function recordJobOutcome(jobType: string, outcome: 'completed' | 'failed', durationMs: number): void {\n jobOutcomeCounter().add(1, { 'job.type': jobType, 'job.outcome': outcome });\n jobDurationHistogram().record(durationMs, { 'job.type': jobType, 'job.outcome': outcome });\n}\n\nlet _appendStageHistogram: Histogram | undefined;\nfunction appendStageHistogram(): Histogram {\n if (!_appendStageHistogram) {\n _appendStageHistogram = meter().createHistogram('semiont.record.append.duration', {\n description: 'Time spent in one stage of appending an event to the record, labeled by stage: persist (JSONL write + git), materialize (view rebuild), enrich, publish. The Archivist\\'s core write path.',\n unit: 'ms',\n });\n }\n return _appendStageHistogram;\n}\n\n/**\n * Record one stage of `EventStore.appendEvent` (ARCHIVIST-STAYS-UP P7).\n *\n * The append path is the one operation only the Archivist can perform, and it\n * was entirely dark: reads had `recordHandlerDuration` and the bus had its own\n * counters, while writes had nothing. Stage-labeled because the useful\n * question is never \"was the append slow\" but WHICH PART — and `materialize`\n * in particular does work proportional to a resource's annotation count, so it\n * degrades with history rather than with load.\n */\nexport function recordAppendStage(\n stage: 'persist' | 'materialize' | 'enrich' | 'publish',\n durationMs: number,\n): void {\n appendStageHistogram().record(durationMs, { 'record.stage': stage });\n}\n\nlet _gitCommandHistogram: Histogram | undefined;\nfunction gitCommandHistogram(): Histogram {\n if (!_gitCommandHistogram) {\n _gitCommandHistogram = meter().createHistogram('semiont.git.duration', {\n description: 'Wall time of a git subprocess. Async — this is latency, not event-loop blockage. Staging is deduped, so the `add` count is far below the number of appended events.',\n unit: 'ms',\n });\n }\n return _gitCommandHistogram;\n}\n\n/**\n * Record a git invocation. Read the `add` count against events appended: one\n * per event means deferred staging has stopped deduping.\n */\nexport function recordGitCommand(command: string, durationMs: number): void {\n gitCommandHistogram().record(durationMs, { 'git.command': command });\n}\n\nfunction gatherDegradeCounter(): Counter {\n if (!_gatherDegradeCounter) {\n _gatherDegradeCounter = meter().createCounter('semiont.gather.degraded', {\n description: 'Gathers that degraded because an eventually-consistent projection did not catch up within its read barrier (vectors: absent semanticContext; graph: projection-lag failure). Labeled by projection.',\n });\n }\n return _gatherDegradeCounter;\n}\n\n/**\n * Record a gather degraded by a projection read barrier: `'vectors'` — the\n * Smelter settle barrier timed out (semanticContext shipped absent);\n * `'graph'` — the Weaver applied barrier + poll floor exhausted (projection\n * lag surfaced as a distinct failure). Fleet-alertable counterpart of the\n * `[gather DEGRADED]` L4 breadcrumbs — a rising rate on either label means\n * that pipeline is not keeping up.\n */\nexport function recordGatherDegrade(projection: 'graph' | 'vectors'): void {\n gatherDegradeCounter().add(1, { projection });\n}\n\n/** Increment the SSE subscriber gauge — call on `/bus/subscribe` open. */\nexport function recordSubscriberConnect(): void {\n sseSubscribersCounter().add(1);\n}\n\n/** Decrement on disconnect. Pair with `recordSubscriberConnect`. */\nexport function recordSubscriberDisconnect(): void {\n sseSubscribersCounter().add(-1);\n}\n\n/**\n * Register a callback that returns the current job-queue snapshot.\n * Polled at the SDK's metric-collection interval. The single gauge\n * emits one observation per status (`pending`, `running`, …) tagged\n * with the `job.status` attribute. Idempotent — last registered\n * provider wins.\n */\nexport function registerJobQueueProvider(\n provider: () => Promise<JobQueueSnapshot> | JobQueueSnapshot,\n): void {\n _jobQueueProvider = provider;\n if (!_jobQueueGauge) {\n _jobQueueGauge = meter().createObservableGauge('semiont.job.queue.size', {\n description: 'Job queue size by status',\n });\n _jobQueueGauge.addCallback(async (observer) => {\n if (!_jobQueueProvider) return;\n const snap = await _jobQueueProvider();\n observer.observe(snap.pending, { 'job.status': 'pending' });\n observer.observe(snap.running, { 'job.status': 'running' });\n observer.observe(snap.complete, { 'job.status': 'complete' });\n observer.observe(snap.failed, { 'job.status': 'failed' });\n observer.observe(snap.cancelled, { 'job.status': 'cancelled' });\n });\n }\n}\n\n/**\n * Register a callback that returns the current vector-index size\n * (point count). Async to allow remote queries (Qdrant). Polled at\n * the metric-collection interval.\n */\n/**\n * Register the Archivist's fact-pump backlog — facts appended to the record\n * but not yet republished onto the bus.\n *\n * At rest this is zero. A value that climbs and does not come back means the\n * pump is outrunning its transport, which is the leading hypothesis for the\n * load-correlated heap growth in `bugs/absent-archivist-wedges-browse.md`\n * (ARCHIVIST-STAYS-UP P5). The backlog is deliberately unbounded today, so\n * this number is the only thing standing between \"the pump is behind\" and an\n * OOM whose cause is inferred from RSS after the fact.\n */\nexport function registerFactPumpDepthProvider(provider: () => number): void {\n _factPumpDepthProvider = provider;\n if (!_factPumpDepthGauge) {\n _factPumpDepthGauge = meter().createObservableGauge('semiont.archivist.fact_pump.depth', {\n description: 'Facts appended to the record but not yet published to the bus. Zero at rest; a rising floor means the pump is behind its transport.',\n });\n _factPumpDepthGauge.addCallback((observer) => {\n if (_factPumpDepthProvider) observer.observe(_factPumpDepthProvider());\n });\n }\n}\n\nlet _gitStagingFailureCounter: Counter | undefined;\n\n/**\n * A staging command that could not be run. Staging the index is a CONVENIENCE\n * — the event log is the system of record — so a failure here is degraded\n * service, never a reason to exit. But degraded must be VISIBLE: this counter\n * is what stops \"the index is quietly stale\" from being invisible.\n */\nexport function recordGitStagingFailure(reason: 'index-lock' | 'other'): void {\n if (!_gitStagingFailureCounter) {\n _gitStagingFailureCounter = meter().createCounter('semiont.git.staging.failures', {\n description: 'Staging commands abandoned after retries; the index may be stale',\n });\n }\n _gitStagingFailureCounter.add(1, { reason });\n}\n\n/**\n * Process lifetime telemetry (ARCHIVIST-GIT-STAGER-CRASH).\n *\n * A supervised process that dies and comes back is INVISIBLE in logs unless\n * someone greps for boot lines, and every request in flight when it died looks\n * to its caller like a hang. On 2026-09-08 that cost a multi-hour hunt through\n * search, qdrant, neo4j and the SSE transport for a crash loop that one metric\n * would have named immediately.\n *\n * `start_time` is the diagnostic, not uptime: a CHANGE in it is unambiguous\n * proof of a restart, and uptime is derivable from it.\n */\nconst PROCESS_START_TIME_SECONDS = Math.floor(Date.now() / 1000);\nlet _processStartTimeGauge: ObservableGauge | undefined;\nlet _restartCountGauge: ObservableGauge | undefined;\nlet _restartCountProvider: (() => Promise<number> | number) | undefined;\nlet _abnormalExitCounter: Counter | undefined;\n\n/** Register `semiont.process.start_time`. Called by `initObservability*`. */\nexport function registerProcessLifetimeMetrics(): void {\n if (_processStartTimeGauge) return;\n _processStartTimeGauge = meter().createObservableGauge('semiont.process.start_time', {\n description: 'Unix seconds at which this process started; a change means it restarted',\n unit: 's',\n });\n _processStartTimeGauge.addCallback((observer) => observer.observe(PROCESS_START_TIME_SECONDS));\n}\n\n/**\n * Supply a restart count. The Archivist's supervisor is POSIX shell and cannot\n * emit OTel, but it already keeps a durable event log on the state mount — so\n * the supervised child reads it and reports the count on the supervisor's\n * behalf.\n */\nexport function registerRestartCountProvider(\n provider: () => Promise<number> | number,\n): void {\n _restartCountProvider = provider;\n if (!_restartCountGauge) {\n _restartCountGauge = meter().createObservableGauge('semiont.process.restarts', {\n description: 'Times the supervisor has restarted this service',\n });\n _restartCountGauge.addCallback(async (observer) => {\n if (_restartCountProvider) observer.observe(await _restartCountProvider());\n });\n }\n}\n\n/**\n * Record that this process is dying abnormally, and mark the active span so a\n * trace shows a span that ENDED IN DEATH rather than one that simply never\n * ends. Never swallows: callers re-raise, so Node's own semantics are intact.\n */\nexport function recordAbnormalTermination(reason: string, detail?: string): void {\n if (!_abnormalExitCounter) {\n _abnormalExitCounter = meter().createCounter('semiont.process.abnormal_exit', {\n description: 'Process terminations that were not a clean shutdown',\n });\n }\n _abnormalExitCounter.add(1, { reason });\n const active = trace.getActiveSpan();\n if (active) {\n active.setStatus({ code: SpanStatusCode.ERROR, message: `${reason}: ${detail ?? ''}`.trim() });\n active.setAttribute('semiont.process.abnormal_exit', reason);\n active.end();\n }\n}\n\nexport function registerVectorIndexSizeProvider(\n provider: () => Promise<number> | number,\n): void {\n _vectorIndexSizeProvider = provider;\n if (!_vectorIndexSizeGauge) {\n _vectorIndexSizeGauge = meter().createObservableGauge('semiont.vector.index.size', {\n description: 'Vector store point count',\n });\n _vectorIndexSizeGauge.addCallback(async (observer) => {\n if (_vectorIndexSizeProvider) {\n const value = await _vectorIndexSizeProvider();\n observer.observe(value);\n }\n });\n }\n}\n\n/**\n * Record an inference call. Token counts are optional — providers that\n * don't expose them (or fail before generating) record only call count\n * and duration.\n */\nexport function recordInferenceUsage(opts: {\n provider: string;\n model: string;\n durationMs: number;\n outcome: 'success' | 'error';\n inputTokens?: number;\n outputTokens?: number;\n}): void {\n const baseAttrs = {\n 'inference.provider': opts.provider,\n 'inference.model': opts.model,\n 'inference.outcome': opts.outcome,\n };\n inferenceCallsCounter().add(1, baseAttrs);\n inferenceDurationHistogram().record(opts.durationMs, baseAttrs);\n if (opts.inputTokens != null && opts.inputTokens > 0) {\n inferenceTokensCounter().add(opts.inputTokens, {\n 'inference.provider': opts.provider,\n 'inference.model': opts.model,\n 'inference.direction': 'input',\n });\n }\n if (opts.outputTokens != null && opts.outputTokens > 0) {\n inferenceTokensCounter().add(opts.outputTokens, {\n 'inference.provider': opts.provider,\n 'inference.model': opts.model,\n 'inference.direction': 'output',\n });\n }\n}\n\nlet _detectionCallCounter: Counter | undefined;\nlet _detectionDurationHistogram: Histogram | undefined;\nlet _detectionItemsHistogram: Histogram | undefined;\nlet _detectionTokensHistogram: Histogram | undefined;\n\nfunction detectionCallCounter(): Counter {\n if (!_detectionCallCounter) {\n _detectionCallCounter = meter().createCounter('semiont.detection.calls', {\n description: 'Detection model calls, labeled by motivation, outcome, subdivision depth and whether this was the floor re-roll.',\n });\n }\n return _detectionCallCounter;\n}\n\nfunction detectionDurationHistogram(): Histogram {\n if (!_detectionDurationHistogram) {\n _detectionDurationHistogram = meter().createHistogram('semiont.detection.call.duration', {\n description: 'Wall time of one detection model call, including the attempts that failed and were retried smaller.',\n unit: 'ms',\n });\n }\n return _detectionDurationHistogram;\n}\n\nfunction detectionItemsHistogram(): Histogram {\n if (!_detectionItemsHistogram) {\n _detectionItemsHistogram = meter().createHistogram('semiont.detection.call.items', {\n description: 'Annotations returned by one detection call. Against the input size on the same record, this is yield.',\n });\n }\n return _detectionItemsHistogram;\n}\n\nfunction detectionTokensHistogram(): Histogram {\n if (!_detectionTokensHistogram) {\n _detectionTokensHistogram = meter().createHistogram('semiont.detection.call.tokens', {\n description: \"Provider-reported tokens for one detection call, by direction. Deliberately separate from semiont.inference.tokens: that series is the authoritative total but carries no subdivision depth, and 'what does a depth-2 call cost' is the question every sizing decision asks.\",\n });\n }\n return _detectionTokensHistogram;\n}\n\n/**\n * Record one detection model call (DETECTION-QUALITY-THROUGHPUT P1).\n *\n * The adapters already record provider/model/duration/tokens for every\n * inference call. What they cannot know is the detection shape around it:\n * which motivation asked, how big the piece was, how many annotations came\n * back, how deep subdivision had descended, and whether this was the floor\n * re-roll. Those are the facts that distinguish a healthy call from a\n * expensive descent, and without them a slow detection run is one\n * undifferentiated number.\n *\n * FAILED attempts are recorded too, and that is the point: the calls paid for\n * and thrown away during a descent are exactly the cost later phases exist to\n * avoid, so a record only of successes would hide the thing being optimized.\n *\n * Tokens are the PROVIDER's counts, passed through — never estimated. Absent\n * means the provider did not report them.\n */\nexport function recordDetectionCall(opts: {\n label: string;\n pieceChars: number;\n durationMs: number;\n items: number;\n depth: number;\n reroll: boolean;\n outcome: 'success' | 'truncated' | 'timeout' | 'collapsed' | 'error';\n inputTokens?: number;\n outputTokens?: number;\n}): void {\n const attrs = {\n 'detection.label': opts.label,\n 'detection.outcome': opts.outcome,\n 'detection.depth': opts.depth,\n 'detection.reroll': opts.reroll,\n };\n detectionCallCounter().add(1, attrs);\n detectionDurationHistogram().record(opts.durationMs, attrs);\n detectionItemsHistogram().record(opts.items, attrs);\n if (opts.inputTokens !== undefined) {\n detectionTokensHistogram().record(opts.inputTokens, { ...attrs, 'detection.direction': 'input' });\n }\n if (opts.outputTokens !== undefined) {\n detectionTokensHistogram().record(opts.outputTokens, { ...attrs, 'detection.direction': 'output' });\n }\n}\n\nlet _anchorOutcomeCounter: Counter | undefined;\nfunction anchorOutcomeCounter(): Counter {\n if (!_anchorOutcomeCounter) {\n _anchorOutcomeCounter = meter().createCounter('semiont.detection.anchors', {\n description: 'Every annotation anchoring, labeled by the method that resolved it. EVERY outcome is counted, not just the risky ones, because a bare count of degraded anchors has no denominator — the rate is the precision signal.',\n });\n }\n return _anchorOutcomeCounter;\n}\n\n/**\n * Record how one annotation got anchored (DETECTION-QUALITY-THROUGHPUT P5).\n *\n * The selector-vs-source check is already a WRITE-TIME INVARIANT — both\n * `buildTextAnnotation` and `buildPdfAnnotation` throw on a selector that does\n * not match its source — so mechanical correctness is guaranteed rather than\n * sampled, and auditing it would measure a constant.\n *\n * What is genuinely uncertain is which anchoring METHOD got there. An `exact`\n * the model quoted verbatim and that appears once is certain; one resolved by\n * `first-of-many` (several occurrences, no usable context) or `fuzzy-match`\n * picked a plausible occurrence and may have picked wrong. Those were visible\n * only as log warnings — countable by a human reading worker output, which is\n * how 47 of them went unreviewed. As a rate they are the precision number that\n * sits beside the yield numbers.\n */\nexport function recordAnchorOutcome(label: string, method: string): void {\n anchorOutcomeCounter().add(1, { 'detection.label': label, 'anchor.method': method });\n}\n\n// ── Re-exports from @opentelemetry/api ─────────────────────────────────\n\nexport { SpanKind, SpanStatusCode, type Attributes, type Span } from '@opentelemetry/api';\n","/**\n * Process-level structured logger for Node entry points.\n *\n * Used by long-lived Node processes (gateway, workers, smelter) that\n * want JSON-structured stdout with active-span trace correlation. The\n * `trace_id` / `span_id` fields are populated from the current OTel\n * span context via `getLogTraceContext` — this is the same Tier 3\n * correlation that lets a grep through stdout line up with the trace\n * UI without manual stitching.\n *\n * Reads `LOG_LEVEL` (default `info`) and `LOG_FORMAT` (`json` default,\n * `simple` for human-friendly dev output).\n *\n * Co-located with `getLogTraceContext` deliberately: this is the\n * only reasonably-shaped consumer of that helper, and putting them in\n * the same package keeps the trace-id wiring in one place.\n */\n\nimport winston from 'winston';\nimport type { Logger } from '@semiont/core';\nimport { getLogTraceContext } from './index.js';\n\nconst traceContextFormat = winston.format((info) => {\n const trace = getLogTraceContext();\n if (trace) {\n info.trace_id = trace.trace_id;\n info.span_id = trace.span_id;\n }\n return info;\n})();\n\nexport function createProcessLogger(component: string): Logger {\n const level = process.env.LOG_LEVEL ?? 'info';\n const format = process.env.LOG_FORMAT === 'simple'\n ? winston.format.combine(\n winston.format.timestamp({ format: 'YYYY-MM-DD HH:mm:ss' }),\n winston.format.errors({ stack: true }),\n traceContextFormat,\n winston.format.printf(({ level: lvl, message, timestamp, ...meta }) => {\n const metaStr = Object.keys(meta).length > 0 ? ` ${JSON.stringify(meta)}` : '';\n return `${timestamp} [${lvl.toUpperCase()}] [${component}] ${message}${metaStr}`;\n }),\n )\n : winston.format.combine(\n winston.format.timestamp(),\n winston.format.errors({ stack: true }),\n traceContextFormat,\n winston.format.json(),\n );\n\n const logger = winston.createLogger({\n level,\n defaultMeta: { component },\n format,\n transports: [new winston.transports.Console()],\n });\n\n return logger;\n}\n"]}
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "@semiont/observability",
3
- "version": "0.5.30",
3
+ "version": "0.5.32",
4
4
  "description": "OpenTelemetry-based tracing for Semiont — Tier 2 of OBSERVABILITY.md. Process-init helpers (Node + Web), withSpan helper, W3C traceparent inject/extract for bus payloads. No-op when no exporter is configured.",
5
5
  "type": "module",
6
6
  "main": "dist/index.js",
@@ -63,16 +63,16 @@
63
63
  },
64
64
  "dependencies": {
65
65
  "@opentelemetry/api": "^1.9.1",
66
- "@opentelemetry/context-async-hooks": "^2.10.0",
67
- "@opentelemetry/core": "^2.10.0",
68
- "@opentelemetry/exporter-metrics-otlp-http": "^0.221.0",
69
- "@opentelemetry/exporter-trace-otlp-http": "^0.221.0",
70
- "@opentelemetry/resources": "^2.10.0",
66
+ "@opentelemetry/context-async-hooks": "^2.11.0",
67
+ "@opentelemetry/core": "^2.11.0",
68
+ "@opentelemetry/exporter-metrics-otlp-http": "^0.222.0",
69
+ "@opentelemetry/exporter-trace-otlp-http": "^0.222.0",
70
+ "@opentelemetry/resources": "^2.11.0",
71
71
  "@opentelemetry/sdk-metrics": "^2.10.0",
72
- "@opentelemetry/sdk-trace-base": "^2.10.0",
73
- "@opentelemetry/sdk-trace-web": "^2.10.0",
72
+ "@opentelemetry/sdk-trace-base": "^2.11.0",
73
+ "@opentelemetry/sdk-trace-web": "^2.11.0",
74
74
  "@opentelemetry/semantic-conventions": "^1.43.0",
75
- "@semiont/core": "0.5.30",
75
+ "@semiont/core": "0.5.32",
76
76
  "winston": "^3.17.0"
77
77
  }
78
78
  }