@deepseek-rlm/dsh-rlm-bundle

Apply dsh.bundle.patch after a pinned DeepSeek Harness rc.7 profile. The example uses explicit absolute Windows roots; change both paths for the deployment host. The package owns the four RLM runtime packages. Cordis and every @deepseek-ai/dsh-* host package are exact peer dependencies, not ordinary dependencies: DSH profiles disable peer auto-installation and resolve them through the installation-managed profiles/node_modules fallback. This keeps the plugin on the pinned, patched host's service classes, symbols, and session seams instead of shadowing them with registry copies. The Jupyter package owns its Schemastery and ZeroMQ runtime dependencies.

Install the implementation tarballs before this bundle tarball. The DSH CLI recognizes the dsh.bundle.patch manifest and appends this layer to the selected profile. Full operation requires all three ordered patches under patches/deepseek-harness; the first two provide the Section 14 parity seams and the third makes downstream informational events cold-readable.

The rlm-spawn-provider row mounts this package's ./spawn-provider wrapper. It registers the native DSH in-process continuable provider and publishes rlmSpawnReady only after that registration commits. The unchanged rlm-jupyter row id declares inject: [rlmSpawnReady], so concurrent Loader activation cannot race provider validation. Removing or disabling the provider row leaves Jupyter pending on an explicit hard dependency; DSH's startup activation audit reports that missing service and exits nonzero. Provider and readiness registrations share one Cordis fiber, so HMR/disposal removes both and dependent rows deactivate until the same edge is ready again.