Plugin install hoists the @deepseek-ai runtime into the profile's node_modules; every agent-preset mount fails on the next cold boot

[Futuri-Risk]

Summary

Installing @deepseek-ai/dsh-subagent-codex into the desktop profile (via the plugin manager) copies the whole @deepseek-ai/* runtime stack into the profile's node_modules (pnpm nodeLinker: hoisted + autoInstallPeers: true, as configured by the app's pnpm-workspace.yaml). After the next full app restart, every agent preset that contains a persona row and a tool-subagent row with standing modelSelectionSettings — i.e. the shipped standard and cordis presets — fails to mount. Resuming any session recorded with those presets then fails, and the failure persists across restarts.

Environment

• DeepSeek Harness desktop 0.2.1-rc.3 (nightly update channel), Windows x64
• Desktop profile plugins: @deepseek-ai/dsh-subagent-codex@0.2.1-rc.3, dsh-codex-connect@0.2.0-alpha.10, dsh-mobile@0.6.16
• Bundles include the experimental agent-team profile stack

Steps to reproduce

1. Fresh install of desktop 0.2.1-rc.3 with a clean profile; presets mount fine.
2. Install @deepseek-ai/dsh-subagent-codex through the plugin manager. Its manifest declares hard dependencies on @deepseek-ai/dsh-tool-subagent, dsh-brand, dsh-sdk-protocol and peers on @deepseek-ai/cordis, dsh-llm, dsh-session, dsh-subagent, dsh-subprocess, dsh-timeout. With the profile's pnpm config, the entire @deepseek-ai/* runtime (~48 packages: cordis, dsh-scope, dsh-system-prompt, dsh-tool-subagent, …) is materialized under ~/.dsh/profiles/desktop/node_modules/@deepseek-ai/.
3. Fully restart the app.
4. Resume any session whose log records agentPreset: "standard" (or create a new session).

Actual behavior

Resume fails for every such session:

resume failed for session "<id>":
RemoteError: persona (@deepseek-ai/dsh-persona): prompt section "deployment:persona-prefix" is already registered (for a per-agent override, register through that agent's `agent.ctx` instead)
tool-subagent (@deepseek-ai/dsh-tool-subagent): tool-subagent: standing `modelSelectionSettings` requires a scoped preset Context (gateway/internal)

Both rows fail activation in auditRows() during mountPreset(); the mount failure is final for that boot, so it persists across restarts until the hoisted packages are removed.

Root cause (from code inspection; workaround confirmed)

The profile-layer resolver intends @deepseek-ai/* runtime packages to resolve to the installation entries embedded in the app, with the profile layer supplying a package only on an installation miss (the dsh-app-boot interception layer / localPackageNames gating; bounded native misses fall through to installation entries). The hoisted copies preempt the miss: any package name physically present under profile/node_modules/@deepseek-ai resolves to the profile copy, while names absent there (dsh-persona, dsh-agent-preset-registry, …) resolve to the app-embedded copy. The process then loads two instances of the runtime.

Scope identity lives in module-level state (@deepseek-ai/dsh-scope: createScope / scopeOf). The preset registry (app-embedded instance) creates the preset scope with one instance; rows and services resolved from the hoisted copies consult the other. scopeOf() therefore returns undefined for the rows' activation contexts:

• the persona row's section registration lands in the global prompt layer and collides with dsh-system-prompt's own deployment deployment:persona-prefix registration ("already registered" without a scope);
• the tool-subagent row's standing modelSelectionSettings finds no composition scope and throws (the compositionScope === void 0 branch in dsh-tool-subagent's activation).

An app process that was already running when the plugin was installed keeps working (modules cached, preset generation already mounted), which is why the breakage only appears on the first cold boot after the install.

Workaround (verified)

Move the hoisted runtime packages out of the profile's resolution path (keep schemastery and dsh-subagent-codex itself, which legitimately live there):

cd ~/.dsh/profiles/desktop/node_modules/@deepseek-ai
mkdir -p ~/.dsh/profiles/desktop/.hoist-shadow-backup
for d in *; do
  [ "$d" = "schemastery" ] && continue
  [ "$d" = "dsh-subagent-codex" ] && continue
  mv "$d" ~/.dsh/profiles/desktop/.hoist-shadow-backup/
done

After a restart, every @deepseek-ai/* import from the profile layer misses natively, falls through to the embedded installation copies, and the presets mount again; sessions resume normally. Note: any later plugin-manager pnpm operation re-materializes the hoisted stack, so the workaround must be re-applied after plugin installs/updates — which is why this feels worth fixing at the packaging or resolution layer.

Suggested fix directions

Either of:

1. Packaging — dsh-subagent-codex (and similar profile plugins) should not declare host-provided runtime packages as hard deps/peers that pnpm materializes into the profile layer; or the plugin manager's pnpm invocation should exclude @deepseek-ai/* from hoisting (hoistPattern / publicHoistPattern exclusions).
2. Resolution — the profile-layer resolver should prefer installation entries over files physically present under profile/node_modules for installation-registered names, so hoisted copies cannot shadow the embedded runtime.

Happy to provide more detail (audit lines, the hoisted package list, plugin-manager install logs) if useful.