dsh-trajectory-persistencePersist and move dsh session trajectories

An observe-only cordis plugin for the DeepSeek Harness — ship byte-faithful zstd segments of every session to S3 and restore them on another machine with sync-down, and export OpenTelemetry GenAI spans to any OTLP backend, including SigV4-signed direct delivery to AWS CloudWatch / Bedrock AgentCore Observability.

Get Started

Ship & Sync

View on GitHub

Byte-faithful ship mode

Tails the official jsonl backend's on-disk artifact read-only and uploads complete zstd frame segments plus a per-session _manifest.json watermark — no re-serialization, and the torn tail a crash leaves behind never leaves the machine.

Cross-machine sync-down

The bundled sync-down CLI restores shipped sessions into a fresh machine's session root — manifest-validated, atomically published, with no-overwrite / prefix-append / conflict-refuse semantics — then dsh resume picks up where the other machine stopped.

OpenTelemetry GenAI semantics

Turns, model calls, and tool executions map to spans following the OTel GenAI semantic conventions, with gen_ai.usage.* token attributes — ready for Jaeger, an OTel Collector, or Langfuse.

AWS SigV4 built in

Sign OTLP batches with AWS Signature Version 4 and POST straight to the CloudWatch OTLP endpoint — the ingest path of Bedrock AgentCore Observability — over the default credential chain. No collector required.

Hot-reload, zero restart

Edit $DSH_HOME/settings.yaml and the plugin rebuilds exactly the sinks whose config changed, draining the old one first so no trajectory data is lost on a switch.

Nothing lost on failure

Bounded ring buffers, exponential-backoff retry, deterministic segment keys, and a local dead-letter directory keep memory flat and trajectories recoverable when a backend stalls.