Your exception filter never sees cron, and that is where the model spend goes

A dev.to write-up on NestJS error capture makes a narrow but load-bearing argument: an HTTP exception filter cannot own observability for an agent workload, because scheduled route refreshes and queue-driven shipment checks never enter the HTTP pipeline at all ^[2]. That matters because the unattended paths are the ones that quietly consume model budget, and they are precisely the ones a filter never sees ^{[2] [4]}.

The prescription is a single failure envelope used at every execution boundary, with cost and latency attached before the boundary reports or rethrows ^[1]. Framework adapters stay thin, one recorder owns normalization and redaction, and each boundary keeps its native retry semantics ^[3]. In practice that means three adapters feeding one recorder: HTTP, cron, and queue ^[6]. The HTTP interceptor establishes timing and correlation around controller execution while the exception filter translates an uncaught failure into the shared envelope and preserves the response contract ^[7]. A cron wrapper has to mint a fresh execution context because there is no inbound request ^[8]. A queue wrapper starts from message metadata, records the current delivery attempt, and rethrows, so the queue system rather than the tracking layer decides retry versus dead-letter ^[9].

The author frames this as a cost-attribution problem as much as an exception-capture problem. If an agent loop calls a model three times, queries a carrier twice, then fails while persisting a route, a stack trace without tenant_id, workflow_id, attempt, elapsed time and accumulated usage tells on-call where execution stopped but not which customer or workflow spent the money ^[4]. That is five external calls already paid for before the failing step ^[1], which is why recording only the failed call is insufficient: the cost belongs to the whole attempt, successful steps included ^[5].

Cardinality discipline follows. Boundary, operation, outcome and a low-cardinality error class are metric dimensions; shipment IDs, workflow IDs, carrier responses and stack traces belong in event detail ^[10]. A tenant identifier may be needed for chargeback, but the recommendation is an approved internal identifier with redaction applied before any exporter sees the event, not raw customer data fanned out to every sink ^[11].

The design test offered is the useful part: can the same logical workflow be reconstructed when its initial HTTP request returns before the queue work begins ^[12]. If correlation depends on request-local state, it cannot. The fix is an explicit workflow identifier carried in the job payload and a new span established in the consumer ^[13]. On buy versus build, the author's position is that a managed tracker reduces exporter maintenance but cannot infer business attribution keys or decide which queue attempts count against an SLO ^[14], while self-hosting buys retention and routing control at the price of capacity planning, upgrades and an on-call surface; neither removes the need for a stable application envelope ^[15].

The reference contract is deliberately small Go rather than NestJS: boundary constants for http, cron and queue, an Event carrying workflow, tenant, attempt, duration, usage and error, a Recorder interface, and an Observe wrapper, with TypeScript adapters assumed in the actual codebase ^{[16] [19]}. Usage is two counters, input and output units ^[18].

Watch the seam in that code. Observe returns early when work succeeds and only stamps duration, usage and error on the failure path ^[17], so a fully successful attempt emits no cost record at all ^[2] - which contradicts the whole-attempt accounting the same piece argues for ^[5]. The supplied text also ends by flagging a deliberate limitation without stating what it is ^[20].