@gmickel/gno 0.6.0 → 0.6.1

package/src/llm/policy.ts

@@ -0,0 +1,84 @@
+/**
+ * Download policy resolution.
+ * Determines whether model downloads are allowed based on env/flags.
+ *
+ * @module src/llm/policy
+ */
+
+// ─────────────────────────────────────────────────────────────────────────────
+// Types
+// ─────────────────────────────────────────────────────────────────────────────
+
+export interface DownloadPolicy {
+  /** True if network is disabled (no HF API calls at all) */
+  offline: boolean;
+  /** True if auto-download is allowed (may still be blocked by offline) */
+  allowDownload: boolean;
+}
+
+export interface PolicyFlags {
+  /** --offline CLI flag */
+  offline?: boolean;
+}
+
+// ─────────────────────────────────────────────────────────────────────────────
+// Helpers
+// ─────────────────────────────────────────────────────────────────────────────
+
+/**
+ * Check if env var is set (non-empty and truthy).
+ * Treats "1", "true", "yes" as truthy. Empty string or "0" as falsy.
+ */
+export function envIsSet(
+  env: Record<string, string | undefined>,
+  key: string
+): boolean {
+  const val = env[key];
+  if (val === undefined || val === '') {
+    return false;
+  }
+  const lower = val.toLowerCase();
+  return lower === '1' || lower === 'true' || lower === 'yes';
+}
+
+// ─────────────────────────────────────────────────────────────────────────────
+// Main
+// ─────────────────────────────────────────────────────────────────────────────
+
+/**
+ * Resolve download policy from environment and CLI flags.
+ *
+ * Precedence (first wins):
+ * 1. --offline flag → offline=true, allowDownload=false
+ * 2. HF_HUB_OFFLINE=1 → offline=true, allowDownload=false
+ * 3. GNO_OFFLINE=1 → offline=true, allowDownload=false
+ * 4. GNO_NO_AUTO_DOWNLOAD=1 → offline=false, allowDownload=false
+ * 5. Default → offline=false, allowDownload=true
+ */
+export function resolveDownloadPolicy(
+  env: Record<string, string | undefined>,
+  flags: PolicyFlags
+): DownloadPolicy {
+  // 1. --offline flag takes highest precedence
+  if (flags.offline) {
+    return { offline: true, allowDownload: false };
+  }
+
+  // 2. HF_HUB_OFFLINE env var (standard HuggingFace offline mode)
+  if (envIsSet(env, 'HF_HUB_OFFLINE')) {
+    return { offline: true, allowDownload: false };
+  }
+
+  // 3. GNO_OFFLINE env var (GNO-specific offline mode)
+  if (envIsSet(env, 'GNO_OFFLINE')) {
+    return { offline: true, allowDownload: false };
+  }
+
+  // 4. GNO_NO_AUTO_DOWNLOAD env var (allow resolve but no download)
+  if (envIsSet(env, 'GNO_NO_AUTO_DOWNLOAD')) {
+    return { offline: false, allowDownload: false };
+  }
+
+  // 5. Default: allow downloads
+  return { offline: false, allowDownload: true };
+}

package/src/mcp/tools/query.ts

@@ -6,7 +6,9 @@
 
 import { join as pathJoin } from 'node:path';
 import { parseUri } from '../../app/constants';
+import { createNonTtyProgressRenderer } from '../../cli/progress';
 import { LlmAdapter } from '../../llm/nodeLlamaCpp/adapter';
+import { resolveDownloadPolicy } from '../../llm/policy';
 import { getActivePreset } from '../../llm/registry';
 import type {
   EmbeddingPort,
@@ -128,6 +130,12 @@ export function handleQuery(
       const preset = getActivePreset(ctx.config);
       const llm = new LlmAdapter(ctx.config);
 
+      // Resolve download policy from env (MCP has no CLI flags)
+      const policy = resolveDownloadPolicy(process.env, {});
+
+      // Non-TTY progress for MCP (periodic lines to stderr, not \r)
+      const downloadProgress = createNonTtyProgressRenderer();
+
       let embedPort: EmbeddingPort | null = null;
       let genPort: GenerationPort | null = null;
       let rerankPort: RerankPort | null = null;
@@ -135,7 +143,10 @@ export function handleQuery(
 
       try {
         // Create embedding port (for vector search) - optional
-        const embedResult = await llm.createEmbeddingPort(preset.embed);
+        const embedResult = await llm.createEmbeddingPort(preset.embed, {
+          policy,
+          onProgress: (progress) => downloadProgress('embed', progress),
+        });
         if (embedResult.ok) {
           embedPort = embedResult.value;
         }
@@ -164,7 +175,10 @@ export function handleQuery(
 
         // Create generation port (for expansion) - optional
         if (!noExpand) {
-          const genResult = await llm.createGenerationPort(preset.gen);
+          const genResult = await llm.createGenerationPort(preset.gen, {
+            policy,
+            onProgress: (progress) => downloadProgress('gen', progress),
+          });
           if (genResult.ok) {
             genPort = genResult.value;
           }
@@ -172,7 +186,10 @@ export function handleQuery(
 
         // Create rerank port - optional
         if (!noRerank) {
-          const rerankResult = await llm.createRerankPort(preset.rerank);
+          const rerankResult = await llm.createRerankPort(preset.rerank, {
+            policy,
+            onProgress: (progress) => downloadProgress('rerank', progress),
+          });
           if (rerankResult.ok) {
             rerankPort = rerankResult.value;
           }

package/src/mcp/tools/vsearch.ts

@@ -6,7 +6,9 @@
 
 import { join as pathJoin } from 'node:path';
 import { parseUri } from '../../app/constants';
+import { createNonTtyProgressRenderer } from '../../cli/progress';
 import { LlmAdapter } from '../../llm/nodeLlamaCpp/adapter';
+import { resolveDownloadPolicy } from '../../llm/policy';
 import { getActivePreset } from '../../llm/registry';
 import { formatQueryForEmbedding } from '../../pipeline/contextual';
 import type { SearchResult, SearchResults } from '../../pipeline/types';
@@ -109,9 +111,18 @@ export function handleVsearch(
       const preset = getActivePreset(ctx.config);
       const modelUri = preset.embed;
 
+      // Resolve download policy from env (MCP has no CLI flags)
+      const policy = resolveDownloadPolicy(process.env, {});
+
+      // Non-TTY progress for MCP (periodic lines to stderr, not \r)
+      const downloadProgress = createNonTtyProgressRenderer();
+
       // Create LLM adapter for embeddings
       const llm = new LlmAdapter(ctx.config);
-      const embedResult = await llm.createEmbeddingPort(modelUri);
+      const embedResult = await llm.createEmbeddingPort(modelUri, {
+        policy,
+        onProgress: (progress) => downloadProgress('embed', progress),
+      });
       if (!embedResult.ok) {
         throw new Error(
           `Failed to load embedding model: ${embedResult.error.message}. ` +

package/src/serve/context.ts

@@ -6,7 +6,9 @@
  */
 
 import type { Config } from '../config/types';
+import type { CreatePortOptions } from '../llm/nodeLlamaCpp/adapter';
 import { LlmAdapter } from '../llm/nodeLlamaCpp/adapter';
+import { resolveDownloadPolicy } from '../llm/policy';
 import { getActivePreset } from '../llm/registry';
 import type {
   DownloadProgress,
@@ -87,8 +89,27 @@ export async function createServerContext(
     const preset = getActivePreset(config);
     const llm = new LlmAdapter(config);
 
+    // Resolve download policy from env (serve has no CLI flags)
+    const policy = resolveDownloadPolicy(process.env, {});
+
+    // Progress callback updates downloadState for WebUI polling
+    const createPortOptions = (type: ModelType): CreatePortOptions => ({
+      policy,
+      onProgress: (progress) => {
+        downloadState.active = true;
+        downloadState.currentType = type;
+        downloadState.progress = progress;
+        if (progress.percent >= 100) {
+          downloadState.completed.push(type);
+        }
+      },
+    });
+
     // Try to create embedding port
-    const embedResult = await llm.createEmbeddingPort(preset.embed);
+    const embedResult = await llm.createEmbeddingPort(
+      preset.embed,
+      createPortOptions('embed')
+    );
     if (embedResult.ok) {
       embedPort = embedResult.value;
       const initResult = await embedPort.init();
@@ -108,18 +129,30 @@ export async function createServerContext(
     }
 
     // Try to create generation port
-    const genResult = await llm.createGenerationPort(preset.gen);
+    const genResult = await llm.createGenerationPort(
+      preset.gen,
+      createPortOptions('gen')
+    );
     if (genResult.ok) {
       genPort = genResult.value;
       console.log('AI answer generation enabled');
     }
 
     // Try to create rerank port
-    const rerankResult = await llm.createRerankPort(preset.rerank);
+    const rerankResult = await llm.createRerankPort(
+      preset.rerank,
+      createPortOptions('rerank')
+    );
     if (rerankResult.ok) {
       rerankPort = rerankResult.value;
       console.log('Reranking enabled');
     }
+
+    // Reset download state after initialization
+    if (downloadState.active) {
+      downloadState.active = false;
+      downloadState.currentType = null;
+    }
   } catch (e) {
     // Log but don't fail - models are optional
     console.log(