b8yg7vjstj-ctrl/dsh-llamacpp-bridge

dsh-llamacpp-bridge is a native double-bundle plugin built specifically for DSH's web profile, designed to integrate a local llama.cpp llama-server runtime as a first-class model provider called llamacpp inside DSH's model selector. The package is split into a Node ESM host that registers the llamacpp route, manages llama-server child processes, synchronizes the models directory, and exposes a same-origin HTTP/SSE data plane, while the browser-side client bundle renders a sidebar entry row, a self-contained terminal-output monitoring panel, and a graphical settings page. Installation is performed via dsh plugin --profile web add ./dsh-llamacpp-bridge-<version>.tgz from the Releases page, and the command automatically maintains ~/.dsh/profiles/web/package.json under dsh.profile.bundles; after installation the host must be restarted and the browser hard-refreshed with Cmd/Ctrl+Shift+R.

The typical workflow starts when a user picks a local GGUF model under the llamacpp group in the selector and sends a message: the plugin's ensure(model) step first stops the previously running model, then launches the target llama-server in either single-model -m <model> mode or --models-dir router mode, waits until the OpenAI-compatible port is actually ready, and only then routes the request. Vision projector files (mmproj-*.gguf) are paired with their base models via a three-tier mechanism that combines manual binding, a <model>-mmproj.gguf suffix convention, and a normalized-name prefix fuzzy match, after which the projector is passed automatically through --mmproj. The sidebar terminal panel streams SSE output in real time, supports keyword filtering, clearing, model switching, and a true terminal Ctrl+C → Y graceful stop. The plugin targets developers who want to power DSH conversations with local llama.cpp models, particularly on Apple Silicon Macs.

Dependencies include Node ≥ 20, an installed DSH web profile that can load the plugin's client half-bundle, and a locally built llama-server; macOS and Linux are supported, with Apple Silicon explicitly verified and Windows untested (the gracefulStop routine falls back to a pipe mode there). The models directory stays in sync through a startup full rescan, fingerprint-based lazy rescans, fs.watch, plus a 30-second polling fallback, and display names longer than 30 characters are truncated to the first 30 characters followed by ... while the internal id remains intact. First-run caveats include killing the host listening on port 3080 via lsof -nP -iTCP:3080 -sTCP:LISTEN then relaunching with dsh web, force-refreshing the browser to clear stale bundles, and avoiding ERR_PNPM_IGNORED_BUILDS by adding allowBuilds: { fsevents: true, sharp: true } to the profile's pnpm-workspace.yaml. Source builds require npm install first and then npm run link:dsh to link the @deepseek-ai/* type packages shipped alongside DSH, and the type layer uses minimal structural typing to keep the plugin compiling across DSH revisions. The license is MIT.