ADR-0001 demotes Langfuse to a projection of Kept's in-process trace

Consider the crash halfway through a tool call. In Kept, ending a trace sweeps every span still marked `in_progress` and stamps it `undetermined` rather than `error`, at `trace.ts#L82-L91` and `span.ts#L92-L105`, so the crash arrives as a span with an unknown result instead of a gap in the tree ^[17]. That distinction is what the type sketches were built to satisfy. Each round was driven by one self-test: can the types express a model call inside turn 2 followed by a tool span whose result state was unknown ^[10]. The answer required dropping "every model call is a span" in favour of three span kinds with a parent link, the turn span sitting above model calls and tool executions as siblings, with causality carried by a `callId` ^[11].

The stated rationale behind the ordering is that logs and metrics matter more for reliability than an ideal architecture or an industry-leading framework ^[3]. The research phase earned it. The original plan had Langfuse holding all spans and traces and serving them wherever they were needed ^[5]. Two readers broke that: the human wants a structured UI for analysis, search and filtering, while the evals want fast access to a reliable record, and routing the evals through Langfuse would have put API calls and latency in the eval path along with concurrency cases to account for ^[5][6].

So ADR-0001 casts Langfuse as a sink for human readers, leaving the machine-facing record to the in-process trace ^[7], and the in-process trace model became the source of truth that evals read, with Langfuse as a projection of it ^[8]. The enforcement is the part worth copying: the core package carries neither the Langfuse SDK nor the OTel SDK, and `check-eval-boundary.mjs` holds the line ^[9]. An ADR that only humans read is a suggestion.

Stamps were picked by asking who reads each one later, guided by the OTel GenAI semantic conventions, and defined at `types.ts#L129-L144` ^[12]. `promptHash` hashes the template rather than the rendered prompt, so the value means prompt version, which is the grouping an eval can actually compare across runs ^[13]. `backendKind` exists to localize a failure to adapter or core, and `traceId` stays distinct from `sessionId` so evals can replay a conversation without merging traces into one garbled tree ^[14][15]. The `faultToggles` array records which faults were injected on that run, so a red cell in the planned CI matrix points back at the toggle that caused it ^[16].

For the shape to transfer, the evals have to run where the trace object lives. In-process truth is only cheap if the reader is in the process; an eval service that reads history after the fact cannot reach that object and is back to querying a store. The ADR relocates that cost; it does not remove it.

The price of the ordering is not in evidence. The post promises three benefits and two drawbacks, plus advice for applying the approach without a fresh codebase ^[18], and the available text stops mid-sentence on why `undetermined` is deliberately not `error` ^[19], leaving all five of those items unpublished ^[20]. What can be checked here is a design and its rationale. Whether front-loading the tracing layer costs more than it saved is a claim the author has flagged and not yet made.