Minke Harness overlay

This package is Minke's product-owned extension layer for DeepSeek Harness. It is installed into the generated desktop runtime and composed through the public --patch bundle seam. Nothing in this package is copied into or applied over vendor/deepseek-harness.

The host composition also exposes two optional capabilities:

• subagent_codex delegates a self-contained task to the native codex app-server --stdio process found on PATH. Codex keeps ownership of its login, model, sandbox, and workspace behavior.
• model-runtime is a DSH plugin that owns local model discovery and optional service lifecycle for exactly two product runtimes: LM Studio and Ollama. LM Studio uses lms server status --json / lms server start and enriches its OpenAI-compatible catalog with LM Studio's v1 loaded-instance metadata. Before dispatch, it verifies that the selected instance has the configured context window. An externally started service is never reconfigured: Minke reports the current and required context lengths and asks the user to reload the model in LM Studio. If Minke started the service itself, it may load or reload the default model instance with the configured context while preserving its supported load parameters. Ollama uses its OpenAI-compatible /v1/models endpoint and starts through ollama serve. A generic openAICompatible adapter remains available for manually configured loopback servers; it does not gain command discovery or process management.

The model runtime executes CLIs through ctx.subprocess, resolves credential references through ctx.credentials, and mounts the upstream @deepseek-ai/dsh-llm-pi-ai plugin after service preparation. Discovered provider metadata is only the composition base layer; it is never serialized to settings.yaml, and user model settings continue to override it. Secrets are resolved for discovery but never copied into provider profiles.

Service policy is explicit:

• external only discovers an already-running service;
• ensure-running starts a missing service. LM Studio's one-shot CLI leaves the shared service running; an ollama serve process started by Minke is owned by the Harness process and stopped with it;
• managed stops the service on plugin disposal only when this plugin proved it started that service.

Both auto-start preferences default to false under modelRuntime.{lmStudio,ollama}.enabled in ~/.minke/desktop/minke.config.json. Electron checks known installation paths and PATH without executing either CLI. The Models page always keeps both provider rows available for configuration, while an auto-start switch appears on a row only when its command was found. Changes take effect after restarting Minke. LM_STUDIO_BASE_URL and OLLAMA_BASE_URL can select explicit loopback endpoints and are also applied to services Minke starts. Without an override, Minke follows the runtimes' official defaults: LM Studio uses http://127.0.0.1:1234/v1 and Ollama uses http://127.0.0.1:11434/v1. Port 0 is rejected because a client Base URL must contain the service's resolved, connectable port. LM_API_TOKEN is used for LM Studio when configured.

The browser half owns Minke's product policy, configurable keyboard shortcuts, and post-boot desktop surface adaptation. It uses:

• settings.onboarding slot shadowing to bypass Harness's developer-only internal-testing notice without changing the upstream plugin;
• settings.section to render its settings page;
• ctx.workspaces.startSession() for the New Session action;
• the Settings trigger's accessible DOM contract for the Settings action;
• Harness's ctx.locale registry and revision source for synchronized zh/en copy;
• Harness's locale snapshot and locale/change event for native desktop copy;
• Harness's ctx.theme snapshot and theme/change event for native window synchronization;
• a lifecycle-managed DOM adapter for the macOS sidebar, titlebar, and translucent surfaces; the adapter is capability-gated by the isolated preload and removes its observer, markers, and stylesheet on disposal;
• the isolated Minke preload bridge for durable desktop-owned preferences.

The separate document-start extension remains CSS-only. It exists solely because first-paint transparency and Electron drag regions must be present before Harness initializes; it does not traverse or modify the Harness DOM.