dsh-otel

The DeepSeek Harness already writes a durable session log after every run, and Jaeger already knows how to draw a trace. What is missing is the translation, and that is all dsh-otel does: it reads the log, splits it into one span per turn, per step, and per tool call, links them into a single trace, and ships the result to any OTLP/HTTP collector you already run — or writes it straight to stdout. It does not observe the harness at runtime, buffer, or sample. A span carries only names, ids, timings, and the token counts the log itself reported; no prompt text, tool arguments, or tool results ever leave the process.

Every id is derived from a sha256 of the bracket's own coordinates, so exporting the same log twice produces the identical tree. If the collector was down the first time, re-running the export repairs the gap rather than duplicating it. Malformed lines are counted and skipped, unclosed brackets are stamped dsh.span.unclosed instead of dropped, and a step with no turn above it gets a synthetic parent so the tree never dangles. The OTLP/JSON document is written by hand with zero OpenTelemetry SDK as a runtime dependency, and the CLI runs against a bare checkout with nothing installed.

It is made for ML engineers and DevOps teams who already operate Jaeger, Phoenix, or Grafana Tempo: point the exporter at their existing stack and every harness run shows up in the trace view they already use. And when the harness itself is the thing that is broken, the CLI's dependency-free design makes it the lightest way to pull that run's structural timeline out for inspection.