create-op-node 0.5.0 → 0.6.0

package/README.md

@@ -28,6 +28,52 @@ npx create-op-node bootstrap
 
 Configures macOS power settings, installs Homebrew + the CLI tool list, sets up Docker Desktop + Tailscale + Ollama, clones the node repo you created, reads the pgsodium key + Tunnel token from the Studio's Keychain (or prompts you to paste them once, then persists for re-runs), writes the LaunchAgent plist, logs into ghcr.io, pulls + warms the LLM model, and finally `docker compose --profile public pull && up -d` brings the whole stack online. Health-check loop waits until all containers are `(healthy)`.
 
+### Choosing the LLM model
+
+By default bootstrap pulls `qwen3.5:9b` (LLM) and `nomic-embed-text`
+(embeddings) — small enough to validate the inference path on first
+run. Override with flags:
+
+```bash
+# Just swap the LLM, keep the default embedding model:
+npx create-op-node bootstrap --region us-ca --llm-model llama3.3:70b
+
+# Override both:
+npx create-op-node bootstrap \
+  --region us-ca \
+  --llm-model llama3.3:70b \
+  --embedding-model mxbai-embed-large
+```
+
+The chosen models flow two places:
+
+1. **Ollama**: bootstrap pulls + warms them so the daemon has them
+   resident before the stack comes up. The embedding model pulls first
+   (small, fast feedback); the LLM pulls second (can be tens of GB).
+2. **LaunchAgent**: the plist exports `LLM_MODEL` and `EMBEDDINGS_MODEL`
+   into the launchd session, which Docker Desktop inherits — so compose
+   services read them via env, no `.env.production` edits needed.
+
+> **`--embedding-model` only takes effect when the knowledge service
+> runs with `EMBEDDINGS_PROVIDER=ollama`.** The default provider is
+> `xenova` (in-process), which bundles its own embedding model and
+> ignores both `EMBEDDINGS_MODEL` and the local Ollama model. Setting
+> `EMBEDDINGS_PROVIDER=ollama` is a separate decision (set in your
+> region repo's `.env.production`) — see `docs/provider-pattern.md`.
+
+> **Template contract**: for `--llm-model` to actually change the
+> running model, the region repo's `docker-compose-prod.yml` must use
+> `${LLM_MODEL:-qwen3.5:9b}` (or similar) on the knowledge service's
+> `environment:` block. The current `opuspopuli-node` template does;
+> a fork that hardcodes the value would ignore the flag silently.
+
+For RAM sizing, the [Docker resources doc](https://github.com/OpusPopuli/opuspopuli-node/blob/main/docs/docker-resources.md)
+has a tier table: 9B-class needs ~8 GB Ollama; 70B-class needs ~50 GB;
+frontier MoE needs ~80 GB. Allocate Docker the remainder.
+
+To switch models post-bootstrap, re-run with the new flag and
+`docker compose down && up -d` to pick up the changed env.
+
 ### Local-only mode (no Cloudflare)
 
 For local dev / testing — frontend on your laptop, backend on the Studio

package/dist/cli.js

@@ -1189,10 +1189,21 @@ async function writePgsodiumKeyFile(key, keyFile) {
     return { ok: false, reason: `writing ${keyFile} failed: ${err.message}` };
   }
 }
+var MODEL_NAME_RE = /^[A-Za-z0-9][A-Za-z0-9._:/-]*$/;
 function renderLaunchAgentPlist(input) {
   if (input.tunnelToken !== void 0 && !TUNNEL_TOKEN_RE.test(input.tunnelToken)) {
     throw new Error("Tunnel token contains characters outside the expected base64-url set");
   }
+  if (input.llmModel !== void 0 && !MODEL_NAME_RE.test(input.llmModel)) {
+    throw new Error(
+      `llmModel ${JSON.stringify(input.llmModel)} contains characters not allowed in a launchd setenv value`
+    );
+  }
+  if (input.embeddingModel !== void 0 && !MODEL_NAME_RE.test(input.embeddingModel)) {
+    throw new Error(
+      `embeddingModel ${JSON.stringify(input.embeddingModel)} contains characters not allowed in a launchd setenv value`
+    );
+  }
   if (!SAFE_PATH_RE.test(input.keyFilePath)) {
     throw new Error(
       `keyFilePath ${JSON.stringify(input.keyFilePath)} contains characters not allowed in a launchd path interpolation`
@@ -1200,7 +1211,9 @@ function renderLaunchAgentPlist(input) {
   }
   const setenvLines = [
     `launchctl setenv PGSODIUM_ROOT_KEY "$(cat ${input.keyFilePath})"`,
-    ...input.tunnelToken !== void 0 ? [`launchctl setenv TUNNEL_TOKEN "${input.tunnelToken}"`] : []
+    ...input.tunnelToken !== void 0 ? [`launchctl setenv TUNNEL_TOKEN "${input.tunnelToken}"`] : [],
+    ...input.llmModel !== void 0 ? [`launchctl setenv LLM_MODEL "${input.llmModel}"`] : [],
+    ...input.embeddingModel !== void 0 ? [`launchctl setenv EMBEDDINGS_MODEL "${input.embeddingModel}"`] : []
   ];
   const command = setenvLines.join("; ");
   return [
@@ -1255,7 +1268,9 @@ async function setupLaunchAgent(input) {
   try {
     plistContent = renderLaunchAgentPlist({
       keyFilePath: paths.keyFile,
-      ...input.tunnelToken !== void 0 ? { tunnelToken: input.tunnelToken } : {}
+      ...input.tunnelToken !== void 0 ? { tunnelToken: input.tunnelToken } : {},
+      ...input.llmModel !== void 0 ? { llmModel: input.llmModel } : {},
+      ...input.embeddingModel !== void 0 ? { embeddingModel: input.embeddingModel } : {}
     });
   } catch (err) {
     return { ok: false, paths, step: "plist", reason: err.message };
@@ -1475,7 +1490,8 @@ function assessHealth(snapshots, requireHealthy) {
 
 // src/lib/ollama.ts
 var OLLAMA_URL = "http://localhost:11434";
-var DEFAULT_MODELS = ["qwen3.5:9b", "nomic-embed-text"];
+var DEFAULT_LLM_MODEL = "qwen3.5:9b";
+var DEFAULT_EMBEDDING_MODEL = "nomic-embed-text";
 var PROBE_ALPINE_TAG = "3.20";
 async function startOllamaService() {
   const res = await safeExeca("brew", ["services", "start", "ollama"]);
@@ -1646,7 +1662,17 @@ var bootstrapCommand = new Command("bootstrap").description(
     "--skip-launch-agent",
     "Skip the LaunchAgent setup (assumes one is already in place)"
   ).default(false)
-).addOption(new Option("--skip-ollama", "Skip the Ollama model pull + warm").default(false)).addOption(new Option("--skip-stack", "Stop before `docker compose pull && up`").default(false)).addOption(
+).addOption(new Option("--skip-ollama", "Skip the Ollama model pull + warm").default(false)).addOption(
+  new Option(
+    "--llm-model <model>",
+    `Ollama LLM model to pull and warm. Default: ${DEFAULT_LLM_MODEL}. Examples: \`llama3.3:70b\`, \`qwen2.5:72b\`. Memory sizing table: docs/docker-resources.md in the opuspopuli-node template (or your region repo's checkout).`
+  )
+).addOption(
+  new Option(
+    "--embedding-model <model>",
+    `Ollama embedding model. Default: ${DEFAULT_EMBEDDING_MODEL}. Only takes effect when the knowledge service runs with EMBEDDINGS_PROVIDER=ollama (otherwise embeddings are computed in-process via xenova).`
+  )
+).addOption(new Option("--skip-stack", "Stop before `docker compose pull && up`").default(false)).addOption(
   new Option(
     "--local-only",
     "Run for local dev / testing: no Tunnel token required, cloudflared stays down. Auto-generates the pgsodium key if not in Keychain (init unnecessary)."
@@ -1664,6 +1690,7 @@ var bootstrapCommand = new Command("bootstrap").description(
   const repoName = `opuspopuli-node-${region}`;
   const composeFileDefault = opts.localOnly ? ["docker-compose-prod.yml"] : ["docker-compose-prod.yml", "docker-compose-backup.yml"];
   const composeFile = opts.composeFile ?? composeFileDefault;
+  const [embeddingModel, llmModel] = resolveModels(opts);
   const sysSpin = p3.spinner();
   sysSpin.start("Inspecting macOS\u2026");
   const snap = await inspectSystem();
@@ -1799,7 +1826,9 @@ var bootstrapCommand = new Command("bootstrap").description(
     laSpin.start("Writing pgsodium key file + LaunchAgent plist\u2026");
     const la = await setupLaunchAgent({
       pgsodiumKey,
-      ...tunnelToken !== void 0 ? { tunnelToken } : {}
+      ...tunnelToken !== void 0 ? { tunnelToken } : {},
+      llmModel,
+      embeddingModel
     });
     if (!la.ok) {
       laSpin.stop(pc2.red(`\u2717 LaunchAgent step ${la.step} failed.`));
@@ -1847,7 +1876,7 @@ var bootstrapCommand = new Command("bootstrap").description(
         process.exit(1);
       }
     }
-    const modelReport = await setupModels(DEFAULT_MODELS, (model, status) => {
+    const modelReport = await setupModels([embeddingModel, llmModel], (model, status) => {
       olSpin.message(`${status}: ${model}`);
     });
     olSpin.stop(
@@ -1966,6 +1995,12 @@ function resolveComposeFiles(repoPath, composeFile) {
   const inputs = composeFile ?? ["docker-compose-prod.yml"];
   return inputs.map((f) => f.startsWith("/") ? f : join(repoPath, f));
 }
+function resolveModels(opts) {
+  return [
+    opts.embeddingModel ?? DEFAULT_EMBEDDING_MODEL,
+    opts.llmModel ?? DEFAULT_LLM_MODEL
+  ];
+}
 function kvBool(b) {
   return b ? pc2.green("on") : pc2.yellow("off");
 }
@@ -3808,7 +3843,7 @@ async function looksLikeRegionsRepo(dir) {
 }
 
 // src/cli.ts
-var VERSION = "0.5.0";
+var VERSION = "0.6.0";
 var program = new Command();
 program.name("create-op-node").description(
   "Interactive bootstrap for an Opus Populi federation node.\nCloudflare infrastructure \u2192 Mac Studio \u2192 live public API."