dsh-continuum

In multi-agent workflows built on DeepSeek Harness, every agent session is inherently ephemeral. Once the context window saturates, critical reasoning, rejected hypotheses, and uncommitted findings can vanish permanently. Successive agents end up repeating expensive codebase research or even regressing bugs that were already fixed. dsh-continuum eliminates this session-amnesia problem.\n\nAt its core the plugin uses deterministic event sourcing: every state change-task claims, findings, decisions, evidence, verifications-is recorded as an immutable, strictly monotonic event, and system state is reconstructed entirely by replaying the log. On top of that foundation it layers renewable task leases, optimistic concurrency control with typed version-conflict errors, four-eyes separation of duty (the implementer can never self-verify), a ten-tier priority-ranked token-budgeted context pack, and crash-resilient partial-line recovery after an abrupt process death. All of these capabilities are exposed through seven first-party tools that agents invoke directly.\n\nIf you are orchestrating multiple LLM agents in DeepSeek Harness for long-running software engineering projects-safe context handoffs, conflict-free concurrent writes, and lossless recovery after an unexpected process exit-dsh-continuum is the memory and coordination layer built for exactly that scenario.