dsh-local-diagnostics

When running DeepSeek Harness locally, one of the hardest problems is figuring out where an Agent is actually stalling: slow dispatch, frequent tool failures, or a repeated error class. dsh-local-diagnostics is built for exactly this scenario. It is a purely in-memory, zero-egress diagnostics plugin that passively observes runtime state through Harness public Cordis extension points. It never modifies the Agent Loop, never collects prompts, messages, tool arguments, session IDs, or credentials, never writes to disk, and never makes network requests.

The plugin plants lightweight probes on five extension points: tools/execute measures each dispatch duration and always forwards the call chain; tools/result tallies successes, failures, and structured error codes; session/event and agent/error count only event types and error class names respectively; diagnostics_report returns a de-identified aggregate report for the current Session with a single Agent query. All data covers only the process lifetime after plugin load and clears on restart, making it ideal for local debugging and short-lived troubleshooting.

It is best suited for developers, operators, or security-sensitive teams running DeepSeek Harness locally who want a structured overview of performance and errors without introducing OpenTelemetry, standing up a remote monitoring stack, or uploading any content off-machine.