Provider-cache economics alternatives — 2026-08-10

Verdict

INCOMPLETE. The local packet executes fak’s provider-cache economics/value fold and raw provider usage. Prometheus, OpenTelemetry, Anthropic, OpenAI, Datadog, and LangSmith retain zero measurements until their real telemetry paths process the common trace; #6164 tracks those witnesses.

Same-workload contract

Every arm receives four chronological turns across two prefix families: cold input, a 900-token cache write, a 900-token cache read, and one context-shed row. Raw totals are 1,700 input, 900 write, and 900 read tokens. The independent oracle requires exact counters, two families, four turns, and a non-zero net token-equivalent result under the same multipliers.

Complete systems report counter/value accuracy, hit rate, fold/query latency, input/write/read/context tokens, CPU/RSS/network/storage, and total cost.

Arm	Class	Local availability
fak native provider-cache economics fold	native	yes
raw provider usage without economics fold	tuned baseline	yes
fak + Prometheus	first-class integration	no
fak + OpenTelemetry	first-class integration	no
Anthropic usage and cost reporting	external	no
OpenAI usage and cost reporting	external	no
Datadog LLM Observability	external	no
LangSmith	external	no

Adapters and synthetic dashboards are not telemetry-system witnesses.

Local native witness

go test ./internal/vcacheobserve -bench BenchmarkObserveEconomics -benchmem -run '^$' -count=5

Windows/amd64, AMD Ryzen 9 9950X: 355988, 308389, 361503, 377797, 385603 ns/op. Median: 361,503 ns/four-turn fold, 217,966 B/op, 3,612 allocs/op. This includes all native report panels, not remote query latency.

Reproduce

go test ./internal/vcacheobserve -run TestCompareLocalKeepsEconomicsAlternativesExplicit
go test ./internal/vcacheobserve -bench BenchmarkObserveEconomics -benchmem -run '^$' -count=5
go test ./internal/nativebench