weflayr 0.24.0 → 0.26.1

package/src/instrumentation/openai/get_audio_telemetry.js

@@ -1,149 +0,0 @@
-'use strict';
-
-/**
- * Hand-written OTel instrumentor for the OpenAI audio endpoints.
- *
- * The `@traceloop/instrumentation-openai` package wraps OpenAI's chat,
- * completions, responses and image endpoints, but not the audio ones, so this
- * ships inside weflayr to cover them. It wraps the Text-to-Speech and
- * Speech-to-Text entrypoints to emit one `gen_ai.*` span per call, which flows
- * through the standard span processor -> exporter with no new transport code. It
- * is registered alongside the traceloop instrumentor whenever
- * `autoInstrument({ aiSdks: 'openai' })` runs.
- *
- * The billable unit differs by API, so each span records the one OpenAI bills:
- *   - Text-to-Speech is billed per input character -> `gen_ai.usage.input_characters`.
- *   - Speech-to-Text is billed per second of audio  -> `gen_ai.usage.input_seconds`.
- *
- * Both endpoints are request/response (unary): `audio.speech.create` resolves to
- * the full audio and `audio.transcriptions.create` to the transcription. The STT
- * audio length is only present when the call requests `response_format: 'verbose_json'`.
- */
-
-const { trace, SpanStatusCode } = require('@opentelemetry/api');
-
-const TRACER_NAME = 'weflayr.instrumentation.openai_audio';
-const TTS_OPERATION = 'text_to_speech';
-const STT_OPERATION = 'speech_to_text';
-
-// Return the gen_ai.system for this call, matching how the OpenAI chat
-// instrumentor attributes the vendor: Azure clients (reached via the same openai
-// SDK) report `azure.ai.openai`, everything else `openai`.
-function _detectSystem(instance) {
-  const baseURL = (instance && instance._client && instance._client.baseURL) || '';
-  return baseURL.toLowerCase().includes('openai.azure.com') ? 'azure.ai.openai' : 'openai';
-}
-
-function _startSpan(operation, body, system) {
-  const span = trace.getTracer(TRACER_NAME).startSpan(operation);
-  span.setAttribute('gen_ai.system', system);
-  span.setAttribute('gen_ai.operation.name', operation);
-  // `model` lives on the request body passed to both audio entrypoints.
-  const model = body && body.model;
-  if (model) {
-    span.setAttribute('gen_ai.request.model', model);
-  }
-  return span;
-}
-
-function _setInputCharacters(span, body) {
-  // `input` is the text passed to text_to_speech; the billable unit is its length.
-  span.setAttribute('gen_ai.usage.input_characters', ((body && body.input) || '').length);
-}
-
-function _setAudioDuration(span, response) {
-  // `duration` (seconds) is only present on a verbose transcription response;
-  // it is the billable unit for speech-to-text.
-  const durationSecs = response && response.duration;
-  if (durationSecs != null) {
-    span.setAttribute('gen_ai.usage.input_seconds', durationSecs);
-  }
-}
-
-function _recordError(span, error) {
-  span.setStatus({ code: SpanStatusCode.ERROR, message: error?.message ?? String(error) });
-  span.recordException(error);
-}
-
-// Wrap a request/response entrypoint: the span spans the single call, and any
-// response-derived billable unit is read off the result.
-function _consumeUnary(span, responsePromise, setResponseAttributes) {
-  return responsePromise.then(
-    (response) => {
-      if (setResponseAttributes) {
-        setResponseAttributes(span, response);
-      }
-      span.end();
-      return response;
-    },
-    (error) => {
-      _recordError(span, error);
-      span.end();
-      throw error;
-    }
-  );
-}
-
-class OpenAIAudioInstrumentor {
-  constructor() {
-    // [ownerPrototype, methodName, originalFunction] for uninstrument().
-    this._patched = [];
-  }
-
-  instrument() {
-    let Speech;
-    let Transcriptions;
-    try {
-      ({ Speech } = require('openai/resources/audio/speech'));
-      ({ Transcriptions } = require('openai/resources/audio/transcriptions'));
-    } catch (err) {
-      throw new Error(
-        'weflayr.autoInstrument(OPENAI): the openai client is not installed. ' +
-          'Install it with: npm install openai\n' +
-          `Underlying error: ${err.message}`
-      );
-    }
-
-    // TTS — request body is the 1st arg; billed per character.
-    this._wrap(Speech.prototype, {
-      operation: TTS_OPERATION,
-      setRequestAttributes: _setInputCharacters,
-      setResponseAttributes: null,
-    });
-
-    // STT — request body is the 1st arg; billed per second of audio.
-    this._wrap(Transcriptions.prototype, {
-      operation: STT_OPERATION,
-      setRequestAttributes: null,
-      setResponseAttributes: _setAudioDuration,
-    });
-  }
-
-  uninstrument() {
-    for (const [owner, methodName, original] of this._patched) {
-      owner[methodName] = original;
-    }
-    this._patched = [];
-  }
-
-  _wrap(owner, { operation, setRequestAttributes, setResponseAttributes }) {
-    const original = owner.create;
-
-    function wrapper(body, ...rest) {
-      const span = _startSpan(operation, body, _detectSystem(this));
-      if (setRequestAttributes) {
-        setRequestAttributes(span, body);
-      }
-      return _consumeUnary(
-        span,
-        Promise.resolve(original.call(this, body, ...rest)),
-        setResponseAttributes
-      );
-    }
-
-    owner.create = wrapper;
-    this._patched.push([owner, 'create', original]);
-  }
-}
-
-module.exports = { OpenAIAudioInstrumentor };

package/src/instrumentation/openai/get_cache_and_reasoning_telemetry.js

@@ -1,71 +0,0 @@
-'use strict';
-
-/**
- * Cached-prompt token enrichment complement for the non-streamed OpenAI chat path.
- *
- * `@traceloop/instrumentation-openai` records `gen_ai.usage.{input,output}_tokens`
- * for non-streamed calls but never reads `usage.prompt_tokens_details.cached_tokens`,
- * so spans for prompt-cache hits carry no cached-token count. The Python
- * OpenLLMetry instrumentor sets `gen_ai.usage.cache_read_input_tokens` from that
- * field, so this brings the Node SDK to parity.
- *
- * It wraps the instrumentor instance's `_endSpan({ span, type, result })`, which
- * finalizes the span after the response resolves, letting us stamp the
- * cached-token count off `result.usage` while the span is still open. OpenAI has
- * no separate cache-write count, so only `cache_read_input_tokens` is set.
- *
- * The same hook surfaces the reasoning-token count: reasoning models report it as
- * `usage.completion_tokens_details.reasoning_tokens` (a subset of output_tokens)
- * which the Node instrumentor drops, so this stamps the flat `reasoning_tokens`
- * attribute (the form the Python instrumentor emits and the normalizer reads).
- */
-
-const CACHE_READ_INPUT_TOKENS_ATTR = 'gen_ai.usage.cache_read_input_tokens';
-const REASONING_TOKENS_ATTR = 'gen_ai.usage.reasoning_tokens';
-
-class OpenAIPromptCachingInstrumentor {
-  constructor() {
-    // [instrumentation, methodName, originalFunction] for uninstrument().
-    this._patched = [];
-  }
-
-  // No-op (and never throws) if the instrumentor's internals have changed shape —
-  // the worst case is a missing cached-token count, never a crash.
-  instrument(instrumentation) {
-    const original = instrumentation && instrumentation._endSpan;
-    if (typeof original !== 'function') return;
-
-    instrumentation._endSpan = function endSpanWithCache(args) {
-      try {
-        const span = args && args.span;
-        const usage = args && args.result && args.result.usage;
-        const recording = span && span.isRecording && span.isRecording();
-        const cachedTokens =
-          usage && usage.prompt_tokens_details && usage.prompt_tokens_details.cached_tokens;
-        if (cachedTokens && recording) {
-          span.setAttribute(CACHE_READ_INPUT_TOKENS_ATTR, cachedTokens);
-        }
-        const reasoningTokens =
-          usage &&
-          usage.completion_tokens_details &&
-          usage.completion_tokens_details.reasoning_tokens;
-        if (reasoningTokens && recording) {
-          span.setAttribute(REASONING_TOKENS_ATTR, reasoningTokens);
-        }
-      } catch {
-        // Never break span finalization over an enrichment failure.
-      }
-      return original.call(this, args);
-    };
-    this._patched.push([instrumentation, '_endSpan', original]);
-  }
-
-  uninstrument() {
-    for (const [owner, methodName, original] of this._patched) {
-      owner[methodName] = original;
-    }
-    this._patched = [];
-  }
-}
-
-module.exports = { OpenAIPromptCachingInstrumentor };

package/src/instrumentation/openai/get_cache_config_telemetry.js

@@ -1,63 +0,0 @@
-'use strict';
-
-/**
- * Request-side prompt-caching config capture for OpenAI chat spans.
- *
- * OpenAI callers influence the (automatic, server-side) prompt cache with
- * request parameters the traceloop instrumentor never records:
- * `prompt_cache_key` (a client-chosen routing key — requests sharing a key are
- * routed to the same cache shard) and the retention tuning —
- * `prompt_cache_retention` ("in_memory" / "24h") on older models, replaced by
- * `prompt_cache_options: { ttl }` ("30m") on GPT-5.6+.
- *
- * Wrap `startSpan({ type, params })` (covering the streamed and non-streamed
- * paths) to stamp the values as `weflayr.cache.key` / `weflayr.cache.retention`
- * (raw retention/ttl string; the normalizer owns the vocabulary).
- */
-
-const CACHE_KEY_ATTR = 'weflayr.cache.key';
-const CACHE_RETENTION_ATTR = 'weflayr.cache.retention';
-
-class OpenAICacheConfigInstrumentor {
-  constructor() {
-    // [instrumentation, methodName, originalFunction] for uninstrument().
-    this._patched = [];
-  }
-
-  // No-op (and never throws) if the instrumentor's internals have changed shape —
-  // the worst case is missing cache-config capture, never a crash.
-  instrument(instrumentation) {
-    if (!instrumentation) return;
-    const startSpanOriginal = instrumentation.startSpan;
-    if (typeof startSpanOriginal !== 'function') return;
-
-    instrumentation.startSpan = function startSpanWithCacheConfig(request) {
-      const span = startSpanOriginal.call(this, request);
-      try {
-        const params = (request && request.params) || {};
-        if (params.prompt_cache_key) {
-          span.setAttribute(CACHE_KEY_ATTR, String(params.prompt_cache_key));
-        }
-        const retention =
-          params.prompt_cache_retention ||
-          (params.prompt_cache_options && params.prompt_cache_options.ttl);
-        if (retention) {
-          span.setAttribute(CACHE_RETENTION_ATTR, String(retention));
-        }
-      } catch {
-        // Never break span creation over an enrichment failure.
-      }
-      return span;
-    };
-    this._patched.push([instrumentation, 'startSpan', startSpanOriginal]);
-  }
-
-  uninstrument() {
-    for (const [owner, methodName, original] of this._patched) {
-      owner[methodName] = original;
-    }
-    this._patched = [];
-  }
-}
-
-module.exports = { OpenAICacheConfigInstrumentor };

package/src/instrumentation/openai/get_embeddings_telemetry.js

@@ -1,107 +0,0 @@
-'use strict';
-
-/**
- * Hand-written OTel instrumentor for the OpenAI embeddings endpoint.
- *
- * `@traceloop/instrumentation-openai` (the package weflayr uses for OpenAI on
- * Node) does not wrap the embeddings endpoint, so this ships inside weflayr to
- * cover it. The Python SDK gets embeddings for free from OpenLLMetry's
- * instrumentor, so there is no Python counterpart to this file — but the span it
- * emits is shaped to normalize identically. It is registered alongside the
- * traceloop instrumentor whenever `autoInstrument({ aiSdks: 'openai' })` runs.
- *
- * Embeddings are billed per input token, so the span records the prompt tokens
- * read off the response under `gen_ai.usage.input_tokens`; there is no completion.
- */
-
-const { trace, SpanStatusCode } = require('@opentelemetry/api');
-
-const TRACER_NAME = 'weflayr.instrumentation.openai_embeddings';
-const EMBEDDINGS_OPERATION = 'embeddings';
-
-// Return the gen_ai.system for this call, matching how the OpenAI chat
-// instrumentor attributes the vendor: Azure clients (reached via the same openai
-// SDK) report `azure.ai.openai`, everything else `openai`.
-function _detectSystem(instance) {
-  const baseURL = (instance && instance._client && instance._client.baseURL) || '';
-  return baseURL.toLowerCase().includes('openai.azure.com') ? 'azure.ai.openai' : 'openai';
-}
-
-function _recordError(span, error) {
-  span.setStatus({ code: SpanStatusCode.ERROR, message: error?.message ?? String(error) });
-  span.recordException(error);
-}
-
-class OpenAIEmbeddingsInstrumentor {
-  constructor() {
-    // [ownerPrototype, methodName, originalFunction] for uninstrument().
-    this._patched = [];
-  }
-
-  instrument() {
-    let Embeddings;
-    try {
-      ({ Embeddings } = require('openai/resources/embeddings'));
-    } catch (err) {
-      throw new Error(
-        'weflayr.autoInstrument(OPENAI): the openai client is not installed. ' +
-          'Install it with: npm install openai\n' +
-          `Underlying error: ${err.message}`
-      );
-    }
-
-    const owner = Embeddings.prototype;
-    const original = owner.create;
-
-    function wrapper(body, ...rest) {
-      const span = trace.getTracer(TRACER_NAME).startSpan(EMBEDDINGS_OPERATION);
-      span.setAttribute('gen_ai.system', _detectSystem(this));
-      span.setAttribute('gen_ai.operation.name', EMBEDDINGS_OPERATION);
-      if (body && body.model) {
-        span.setAttribute('gen_ai.request.model', body.model);
-      }
-      // One user message per input element, matching the shape the chat
-      // instrumentors emit. The span processor strips it when message-content
-      // capture is disabled.
-      if (body && body.input != null) {
-        const inputs = Array.isArray(body.input) ? body.input : [body.input];
-        span.setAttribute(
-          'gen_ai.input.messages',
-          JSON.stringify(
-            inputs.map((content) => ({ role: 'user', parts: [{ type: 'text', content }] }))
-          )
-        );
-      }
-      return Promise.resolve(original.call(this, body, ...rest)).then(
-        (response) => {
-          if (response && response.model) {
-            span.setAttribute('gen_ai.response.model', response.model);
-          }
-          const promptTokens = response && response.usage && response.usage.prompt_tokens;
-          if (promptTokens != null) {
-            span.setAttribute('gen_ai.usage.input_tokens', promptTokens);
-          }
-          span.end();
-          return response;
-        },
-        (error) => {
-          _recordError(span, error);
-          span.end();
-          throw error;
-        }
-      );
-    }
-
-    owner.create = wrapper;
-    this._patched.push([owner, 'create', original]);
-  }
-
-  uninstrument() {
-    for (const [owner, methodName, original] of this._patched) {
-      owner[methodName] = original;
-    }
-    this._patched = [];
-  }
-}
-
-module.exports = { OpenAIEmbeddingsInstrumentor };

package/src/instrumentation/openai/get_streaming_usage.js

@@ -1,81 +0,0 @@
-'use strict';
-
-/**
- * Streamed token-usage enrichment complement for the OpenAI chat path.
- *
- * `@traceloop/instrumentation-openai` does not record token usage for streamed
- * calls: its stream handler accumulates `choices[].delta` but never reads the
- * final usage chunk OpenAI sends when the request sets
- * `stream_options: { include_usage: true }`. Non-streaming calls work because
- * OpenAI returns `usage` in the response body, which the instrumentor reads
- * directly. The Python OpenLLMetry instrumentor reads the streamed usage, so
- * this gap is Node-only.
- *
- * It wraps the instrumentor instance's `_streamingWrapPromise`: that handler
- * hands us the span and ends it only after its chunk loop, so when the usage
- * chunk arrives the span is still open and we set the same `gen_ai.usage.*`
- * attributes the instrumentor sets for non-streaming calls.
- */
-
-const USAGE_INPUT_TOKENS_ATTR = 'gen_ai.usage.input_tokens';
-const USAGE_OUTPUT_TOKENS_ATTR = 'gen_ai.usage.output_tokens';
-const USAGE_TOTAL_TOKENS_ATTR = 'gen_ai.usage.total_tokens';
-const CACHE_READ_INPUT_TOKENS_ATTR = 'gen_ai.usage.cache_read_input_tokens';
-
-function _setUsageAttributes(span, usage) {
-  if (typeof usage.prompt_tokens === 'number') {
-    span.setAttribute(USAGE_INPUT_TOKENS_ATTR, usage.prompt_tokens);
-  }
-  if (typeof usage.completion_tokens === 'number') {
-    span.setAttribute(USAGE_OUTPUT_TOKENS_ATTR, usage.completion_tokens);
-  }
-  if (typeof usage.total_tokens === 'number') {
-    span.setAttribute(USAGE_TOTAL_TOKENS_ATTR, usage.total_tokens);
-  }
-  // Cached-prompt tokens ride in the same streamed usage chunk; mirror the
-  // non-streamed cache backfill in get_cache_and_reasoning_telemetry.js.
-  const cachedTokens = usage.prompt_tokens_details && usage.prompt_tokens_details.cached_tokens;
-  if (cachedTokens) {
-    span.setAttribute(CACHE_READ_INPUT_TOKENS_ATTR, cachedTokens);
-  }
-}
-
-class OpenAIStreamingUsageInstrumentor {
-  constructor() {
-    // [instrumentation, methodName, originalFunction] for uninstrument().
-    this._patched = [];
-  }
-
-  // No-op if the instrumentor's internals have changed shape — the worst case is
-  // streamed spans without token counts, never a crash.
-  instrument(instrumentation) {
-    const original = instrumentation && instrumentation._streamingWrapPromise;
-    if (typeof original !== 'function') return;
-
-    instrumentation._streamingWrapPromise = function streamingWrapWithUsage(streamArgs) {
-      const span = streamArgs && streamArgs.span;
-      const wrappedStream = original.call(this, streamArgs);
-      if (!span) return wrappedStream;
-
-      return (async function* backfillUsage() {
-        for await (const chunk of wrappedStream) {
-          const usage = chunk && chunk.usage;
-          if (usage && span.isRecording()) {
-            _setUsageAttributes(span, usage);
-          }
-          yield chunk;
-        }
-      })();
-    };
-    this._patched.push([instrumentation, '_streamingWrapPromise', original]);
-  }
-
-  uninstrument() {
-    for (const [owner, methodName, original] of this._patched) {
-      owner[methodName] = original;
-    }
-    this._patched = [];
-  }
-}
-
-module.exports = { OpenAIStreamingUsageInstrumentor };

package/src/instrumentation/openai/set_telemetry_attributes.js

@@ -1,41 +0,0 @@
-'use strict';
-
-/**
- * OpenAI telemetry complement — the registered orchestrator.
- *
- * Wires every weflayr-supplied patch the traceloop OpenAI instrumentor misses:
- *   - the audio (TTS/STT) and embeddings endpoints it doesn't cover, and
- *   - usage it drops on the chat endpoint it does cover: streamed token usage and
- *     non-streamed cached-prompt tokens (these enrich the traceloop instrumentor
- *     instance, hence the `instrumentation` argument).
- */
-
-const { OpenAIAudioInstrumentor } = require('./get_audio_telemetry');
-const { OpenAICacheConfigInstrumentor } = require('./get_cache_config_telemetry');
-const { OpenAIEmbeddingsInstrumentor } = require('./get_embeddings_telemetry');
-const { OpenAIStreamingUsageInstrumentor } = require('./get_streaming_usage');
-const { OpenAIPromptCachingInstrumentor } = require('./get_cache_and_reasoning_telemetry');
-
-class OpenAITelemetryInstrumentor {
-  constructor() {
-    this._subInstrumentors = [
-      new OpenAIAudioInstrumentor(),
-      new OpenAIEmbeddingsInstrumentor(),
-      new OpenAIStreamingUsageInstrumentor(),
-      new OpenAIPromptCachingInstrumentor(),
-      new OpenAICacheConfigInstrumentor(),
-    ];
-  }
-
-  // The endpoint complements ignore `instrumentation`; the streaming/cache ones
-  // patch its internal handlers.
-  instrument(instrumentation) {
-    for (const sub of this._subInstrumentors) sub.instrument(instrumentation);
-  }
-
-  uninstrument() {
-    for (const sub of this._subInstrumentors) sub.uninstrument();
-  }
-}
-
-module.exports = { OpenAITelemetryInstrumentor };

package/src/instrumentation/vercel/full_instrumentation.js

@@ -1,52 +0,0 @@
-'use strict';
-
-/**
- * Instrumentor for the Vercel AI SDK (`ai` package, v7+).
- *
- * The AI SDK ships its own OTel integration in `@ai-sdk/otel`; instrumenting
- * is registering it globally so every generateText/streamText/embed call emits
- * GenAI semconv spans through the tracer provider weflayr's setup() installed.
- * `usage: true` adds the `ai.usage.*` token details (reasoning tokens, uncached
- * input) that the semconv attributes do not carry; `providerMetadata: true`
- * opts into `ai.response.providerMetadata` on the spans that carry it.
- *
- * The aggregate spans the integration emits around each model call
- * (invoke_agent, agent_step, duplicated embeddings spans) are filtered out
- * before export by otel/vercel-ai-sdk-span-filter.js.
- */
-class VercelAiSdkInstrumentor {
-  instrument() {
-    let registerTelemetry;
-    let OpenTelemetry;
-    try {
-      ({ registerTelemetry } = require('ai'));
-      ({ OpenTelemetry } = require('@ai-sdk/otel'));
-    } catch (err) {
-      throw new Error(
-        'weflayr.autoInstrument("vercel_ai_gateway"): failed to load "ai" / "@ai-sdk/otel". ' +
-          `Install them with: npm install ai @ai-sdk/otel\nUnderlying error: ${err.message}`
-      );
-    }
-    // registerTelemetry only exists since v7.
-    if (typeof registerTelemetry !== 'function') {
-      throw new Error(
-        'weflayr.autoInstrument("vercel_ai_gateway") requires version 7+ of the Vercel AI SDK ' +
-          '(the "ai" package), which added global telemetry registration. ' +
-          'Upgrade with: npm install ai@latest @ai-sdk/otel.\n' +
-          'On ai <= 6 telemetry can only be enabled per call: remove "vercel_ai_gateway" from ' +
-          'aiSdks (pass an empty list if no other sdks need to be instrumented). Instead, make sure ' +
-          'experimental_telemetry: { isEnabled: true } is passed on each generateText/streamText call.'
-      );
-    }
-    const dispatcher = new OpenTelemetry({ usage: true, providerMetadata: true });
-    // The integration never emits the raw provider usage, the only place the
-    // written cache tokens per TTL (Anthropic's `cache_creation`) are
-    // reported; this complement stamps it on the model-call span (see
-    // get_cache_creation_telemetry).
-    const { VercelCacheCreationInstrumentor } = require('./get_cache_creation_telemetry');
-    new VercelCacheCreationInstrumentor().instrument(dispatcher);
-    registerTelemetry(dispatcher);
-  }
-}
-
-module.exports = { VercelAiSdkInstrumentor };

package/src/instrumentation/vercel/get_cache_creation_telemetry.js

@@ -1,70 +0,0 @@
-'use strict';
-
-/**
- * Cache-write capture for Vercel AI SDK (ai >= 7) model-call spans.
- *
- * Anthropic reports the written cache tokens per TTL in its raw usage
- * (`usage.cache_creation`, ephemeral_5m/1h buckets — billed at different
- * rates). ai >= 7 carries that verbatim on the model-call telemetry event as
- * `event.usage.raw` (a documented field of LanguageModelUsage), but the
- * `@ai-sdk/otel` integration never emits it, so the span loses it.
- *
- * Wrap the telemetry dispatcher's `onLanguageModelCallEnd(event)` (a public
- * callback of the telemetry interface) to read `usage.cache_creation` off the
- * raw usage and stamp it verbatim as `weflayr.cache.creation` on the
- * still-open model-call span (reached through the integration's call state)
- * before the original handler ends it; the normalizer owns the mapping to TTL
- * tiers.
- */
-
-const CACHE_CREATION_ATTR = 'weflayr.cache.creation';
-
-// The raw `usage.cache_creation` of a provider usage, or undefined when the
-// provider reports none (or reports only zeros — non-caching calls carry no
-// attribute).
-function cacheCreationFromRawUsage(rawUsage) {
-  const cacheCreation = rawUsage && rawUsage.cache_creation;
-  if (!cacheCreation) return undefined;
-  return Object.values(cacheCreation).some(Boolean) ? cacheCreation : undefined;
-}
-
-class VercelCacheCreationInstrumentor {
-  constructor() {
-    // [owner, methodName, originalFunction] for uninstrument().
-    this._patched = [];
-  }
-
-  // No-op (and never throws) if the interface is missing — the worst case is
-  // the cache-write capture staying empty, exactly the pre-complement
-  // behaviour.
-  instrument(dispatcher) {
-    if (!dispatcher || typeof dispatcher.onLanguageModelCallEnd !== 'function') return;
-    const original = dispatcher.onLanguageModelCallEnd;
-
-    dispatcher.onLanguageModelCallEnd = function onLanguageModelCallEndWithCacheCreation(event) {
-      try {
-        const cacheCreation = cacheCreationFromRawUsage(event && event.usage && event.usage.raw);
-        if (cacheCreation) {
-          const callState = this.getCallState && this.getCallState(event.callId);
-          const span = callState && callState.inferenceSpan;
-          if (span && typeof span.setAttribute === 'function') {
-            span.setAttribute(CACHE_CREATION_ATTR, JSON.stringify(cacheCreation));
-          }
-        }
-      } catch {
-        // Never break span finalization over an enrichment failure.
-      }
-      return original.call(this, event);
-    };
-    this._patched.push([dispatcher, 'onLanguageModelCallEnd', original]);
-  }
-
-  uninstrument() {
-    for (const [owner, methodName, original] of this._patched) {
-      owner[methodName] = original;
-    }
-    this._patched = [];
-  }
-}
-
-module.exports = { VercelCacheCreationInstrumentor, cacheCreationFromRawUsage };

package/src/logger.js

@@ -1,8 +0,0 @@
-'use strict';
-
-const { debuglog } = require('node:util');
-
-// Enable with NODE_DEBUG=weflayr
-const debug = debuglog('weflayr');
-
-module.exports = { debug };