dsh-litefuse-plugin

A single user turn can fan out into multiple model calls, tool executions, and delegated subagent runs, making it hard to reconstruct true timing and cost from logs or transcripts alone. dsh-litefuse-plugin maps every DeepSeek Harness user turn directly into a complete Litefuse trace, providing observability and evals without extra collection infrastructure.\n\nThe plugin runs in-process on the harness session event stream with zero runtime dependencies — no Langfuse SDK, no file-tailing collector. Spans are sent straight to the OTLP endpoint. Each model call becomes a generation carrying true per-call latency, time-to-first-token, and a token breakdown that separates reasoning and cache tokens from the raw total. Each tool call produces its own span; run_code dispatches mount as nested spans beneath the dispatching call. Subagent runs bind to their parent delegation tool span as containers, with tokens rolling up into the parent trace automatically. A trace appears in the Litefuse UI the moment its first span completes, not at turn end.\n\nBuilt for developers and teams who orchestrate multi-model, multi-tool, subagent-bearing workflows with DeepSeek Harness and rely on Litefuse for latency analysis, cost accounting, or agent evals. All configuration is injected through environment variables — no code changes required — with optional per-field overrides in cordis.patch.yml.