Install DeepSeek RLM

DeepSeek RLM is currently a local five-package preview for one audited DeepSeek Harness release. Complete every section in order; mixing DSH versions or registry-installed host packages can split Cordis service identities.

Prerequisites

• DeepSeek Harness revision dd6322d604e00eec1ba5e0c8541159906a21094a (dsh-v0.1.2-alpha.3)
• Node.js ^22.19 or >=24
• pnpm 11.7.0
• uv, with permission to provision Python 3.11
• absolute writable directories for RLM artifacts and the managed runtime

1. Build the local packages

pnpm install --frozen-lockfile
pnpm package:bundle

The command builds, packs, installs, and imports all five packages in an isolated smoke project. The tarballs are written to artifacts/packages/.

2. Patch the pinned DSH checkout

Apply the patches in filename order to the exact revision above:

git -C /path/to/deepseek-harness apply /path/to/deepseek-rlm/patches/deepseek-harness/0002-continuable-child-deletion.patch
git -C /path/to/deepseek-harness apply /path/to/deepseek-rlm/patches/deepseek-harness/0003-public-ignorable-session-events.patch
git -C /path/to/deepseek-harness apply /path/to/deepseek-rlm/patches/deepseek-harness/0004-pi-ai-agent-session-cleanup.patch
git -C /path/to/deepseek-harness apply /path/to/deepseek-rlm/patches/deepseek-harness/0005-bounded-process-shutdown.patch

DSH alpha.3 natively provides validated, persisted per-child reasoning. Patch 2 provides public durable child deletion. Patch 3 makes independently defined rlm/* informational events safe for cold reads. Those are the two RLM capability seams. Patch 4 releases pi-ai's provider session resources with the exact owning Agent or agentless Session. Patch 5 preserves DSH's bounded CLI exit after the application tree is disposed if a transport still owns a referenced OS handle. Missing capability patches produce explicit unsupported-capability errors; the plugin does not substitute a weaker path.

Verify the series against a disposable checkout when changing either side:

DSH_SOURCE=/absolute/path/to/deepseek-harness pnpm verify:patches

On PowerShell:

$env:DSH_SOURCE = 'D:\absolute\path\to\deepseek-harness'
pnpm verify:patches

3. Initialize the DSH profile

dsh plugin --profile <profile> install

Keep the profile-created nodeLinker: hoisted and autoInstallPeers: false settings. Cordis and DSH host packages are peers and must resolve from DSH's installation-managed profiles/node_modules fallback. Do not install registry copies of those host packages into the profile.

4. Point internal edges at the local tarballs

The preview package names are not published. pnpm 11 reads overrides and build approvals from $DSH_HOME/profiles/<profile>/pnpm-workspace.yaml, not from the pnpm field in package.json. Keep the generated linker settings, approve the audited ZeroMQ native build, and add absolute file: URLs for the four internal package edges:

packages:
  - .

nodeLinker: hoisted
autoInstallPeers: false

allowBuilds:
  zeromq: true

overrides:
  '@deepseek-rlm/dsh-rlm': file:D:/deepseek-rlm/artifacts/packages/deepseek-rlm-dsh-rlm-0.1.0-preview.0.tgz
  '@deepseek-rlm/dsh-rlm-prime-runtime': file:D:/deepseek-rlm/artifacts/packages/deepseek-rlm-dsh-rlm-prime-runtime-0.1.0-preview.0.tgz
  '@deepseek-rlm/dsh-rlm-jupyter': file:D:/deepseek-rlm/artifacts/packages/deepseek-rlm-dsh-rlm-jupyter-0.1.0-preview.0.tgz
  '@deepseek-rlm/dsh-tool-ipython': file:D:/deepseek-rlm/artifacts/packages/deepseek-rlm-dsh-tool-ipython-0.1.0-preview.0.tgz

Use forward slashes in Windows file: URLs. On macOS or Linux, use an absolute URL such as file:/home/user/deepseek-rlm/artifacts/packages/...tgz.

5. Configure runtime roots

Installing the bundle activates its three Loader rows. Override the existing rlm-jupyter row by id in the profile's cordis.patch.yml; do not insert a second copy of the rows from packages/bundle/example.cordis.yml. At minimum, set these absolute paths:

- id: rlm-jupyter
  config:
    artifactRoot: D:/dsh-state/rlm-artifacts
    managedRuntimeRoot: D:/dsh-state/rlm-runtime

The example patch is for a composition that is not installing the bundle and therefore needs to insert all three rows itself.

The default composition uses rlm-spawn, maxDepth: 1, snapshots after each cell, and no Python-to-DSH tools adapter. The spawn-provider wrapper publishes its readiness service only after native DSH registration commits, so Loader row order is not used as a timing contract.

Important defaults:

Setting	Default
subagentProvider	rlm-spawn
maxDepth	1
maxConcurrentKernelBoots	4
interruptGraceMs	2000
shutdownGraceMs	5000
hostRequestDrainMs	5000
maxOutputBytes	4 MiB/cell
snapshot.policy	after-cell
snapshot.maxBytes	256 MiB/file
snapshot.maxVariableBytes	16 MiB/value
envAllowlist	empty

shellPath replaces a bare %%bash with %%script <shell>. commandPrefix is applied only to a leading %%bash cell. Keep the deferred goal and compaction adapters disabled.

6. Install and verify the bundle

dsh plugin --profile <profile> add -w /absolute/path/to/artifacts/packages/deepseek-rlm-dsh-rlm-bundle-0.1.0-preview.0.tgz
dsh --profile <profile> --dump-default-config

The dump must contain exactly one active row for each of:

• rlm-spawn-provider
• rlm-jupyter
• rlm-ipython-tool

Start the profile and verify two consecutive ipython calls share a variable. Then admit a child with await rlm(...) and require it to report through agent_message.send().

Troubleshooting

• UNSUPPORTED_DELETION: rebuild the exact DSH checkout with patch 2.
• UNSUPPORTED_IGNORABLE_SESSION_EVENTS: apply patch 3. A live-only event is intentionally not used as fallback.
• Managed runtime startup failure: confirm uv is on PATH, both runtime roots are writable, and Python 3.11 can be installed. A custom Python must be absolute and import ipykernel, dill, nest_asyncio, rlm, and dsh_rlm_runtime.
• ERR_PNPM_IGNORED_BUILDS for ZeroMQ: set allowBuilds.zeromq: true in the profile's pnpm-workspace.yaml, then rerun dsh plugin --profile <profile> install.
• %%bash on Windows: configure an absolute Git Bash, Cygwin, or compatible executable using shellPath.
• Snapshot restore diagnostic: the event/file digest pair is authoritative. Missing or corrupt files start an empty namespace; orphan files are ignored.
• Missing spawn provider: confirm the bundle row is enabled. The Jupyter row deliberately fails startup rather than racing or substituting another backend.
• Headless output completes but the process stays open: confirm patches 4 and 5 are applied. Patch 4 requests public pi-ai session cleanup; patch 5 retains the five-second post-disposal backstop for a WebSocket whose close handshake leaves a referenced transport handle.

Upgrade policy

An upstream upgrade requires updating both revision pins, regenerating Prime file digests and the Python lock, rebasing all DSH patches, and rerunning the host-protocol, kernel, event-vocabulary, Loader/package, license, and upstream patch suites. See SPEC.md and THIRD_PARTY_NOTICES.md.