straylight: install OpenCode config for local llama.cpp models

Managed /etc/opencode/opencode.json points at straylight:11434 with the router model ids, Laguna as default, Ornith 1.5 9B as small_model, and a vision subagent on Qwen3-VL-8B abliterated.
2026-09-16 06:34:41 -07:00
parent 56b233afb6
commit 52859524d7
2 changed files with 68 additions and 0 deletions
@@ -0,0 +1,67 @@
+# OpenCode pointed at straylight's llama.cpp OpenAI-compatible API.
+# Managed config lives in /etc/opencode/ (highest OpenCode precedence on Linux).
+# llama-server is on straylight:11434; MagicDNS resolves that to loopback here.
+
+{ pkgs, lib, ... }:
+let
+  llamaBase = "http://straylight:11434/v1";
+  limit = { context = 131072; output = 32768; };
+  text = name: {
+    inherit name limit;
+    modalities = { input = [ "text" ]; output = [ "text" ]; };
+  };
+  vision = name: {
+    inherit name limit;
+    modalities = { input = [ "text" "image" ]; output = [ "text" ]; };
+  };
+  cfg = {
+    "$schema" = "https://opencode.ai/config.json";
+    enabled_providers = [ "llama.cpp" ];
+    model = "llama.cpp/laguna-s-2.1";
+    small_model = "llama.cpp/ornith-1.5-9b-uncensored";
+    provider = {
+      "llama.cpp" = {
+        npm = "@ai-sdk/openai-compatible";
+        name = "llama.cpp (straylight)";
+        options = {
+          baseURL = llamaBase;
+          apiKey = "none";
+        };
+        models = {
+          "laguna-s-2.1" = text "Laguna S 2.1";
+          "qwen3.6-35b-a3b-abliterated" = text "Qwen3.6 35B A3B";
+          "gemma4-26b-a4b-abliterated" = text "Gemma 4 26B A4B";
+          "ornith-1.0-35b" = text "Ornith 1.0 35B";
+          "ornith-1.5-9b-uncensored" = text "Ornith 1.5 9B";
+          "ornith-1.5-35b-a3b" = text "Ornith 1.5 35B MoE";
+          "qwen3.8-27b-uncensored" = text "Qwen3.8 27B";
+          "qwen3-vl-8b-abliterated" = vision "Qwen3-VL 8B";
+        };
+      };
+    };
+    agent = {
+      vision = {
+        description = "Look at images, screenshots, diagrams, UI captures, and other visual artifacts. Use whenever the task needs to see a picture rather than guess from a filename or OCR text.";
+        mode = "subagent";
+        model = "llama.cpp/qwen3-vl-8b-abliterated";
+        color = "accent";
+        permission = {
+          edit = "deny";
+        };
+        prompt = ''
+          You are a vision specialist running on Qwen3-VL-8B (abliterated) via llama.cpp.
+
+          You receive image paths and/or attached images. Describe what is actually in the image.
+          Transcribe visible text. Answer the caller's question about the image. Do not invent UI
+          that is not there. If the image cannot be read, say so and what you would need instead.
+
+          You may read files (including images) and run read-only shell commands. Do not edit the
+          workspace; report findings back to the parent agent.
+        '';
+      };
+    };
+  };
+in
+{
+  environment.etc."opencode/opencode.json".text = builtins.toJSON cfg;
+}