fix(cli): align termite model directory between setup and index

.changeset/fix-model-dir-mismatch.md

@@ -0,0 +1,5 @@
+---
+"@prosdevlab/dev-agent": patch
+---
+
+Fix `dev setup` reporting model ready while `dev index` fails with "model not found". The CLI's `hasModel`/`pullModel` used `~/.termite/models` but the running server looked in `~/.antfly/models`. Both now use a shared `--models-dir` pointing at the server's data directory.

README.md

@@ -24,9 +24,8 @@ dev-agent indexes your codebase and provides 6 MCP tools to AI assistants. Inste
 ```bash
 # Install
 npm install -g @prosdevlab/dev-agent
-brew install --cask antflydb/antfly/antfly
 
-# One-time setup
+# One-time setup (installs Antfly, pulls embedding model, starts server)
 dev setup
 
 # Index your repository
@@ -138,7 +137,7 @@ Server health, Antfly connectivity, and repository access.
 ## Prerequisites
 
 - Node.js 22+ (LTS)
-- [Antfly](https://antfly.io) — `brew install --cask antflydb/antfly/antfly`
+- [Antfly](https://antfly.io) — installed automatically by `dev setup`
 
 ## Development
 

packages/cli/src/utils/__tests__/antfly.test.ts

@@ -0,0 +1,92 @@
+/**
+ * Tests for antfly utility helpers.
+ *
+ * Regression for: hasModel() false positive when antfly termite list defaulted
+ * to ~/.termite/models (different from the server's ~/.antfly/models), causing
+ * "Embedding model ready" in `dev setup` but "model not found" in `dev index`.
+ */
+
+import { describe, expect, it } from 'vitest';
+
+// modelPresentInOutput is not exported — test via the exported path by extracting
+// the pure logic into a local copy that mirrors the implementation exactly.
+// This keeps the test focused on the matching logic without requiring CLI env.
+
+function modelPresentInOutput(model: string, output: string): boolean {
+  if (output.includes(model)) return true;
+
+  const shortName = model.split('/').pop() ?? model;
+  const escaped = shortName.replace(/[.*+?^${}()|[\]\\]/g, '\\$&');
+  return new RegExp(`(?<![\\w/-])${escaped}(?![\\w/-])`).test(output);
+}
+
+describe('modelPresentInOutput', () => {
+  const FULL_NAME = 'BAAI/bge-small-en-v1.5';
+  const SHORT_NAME = 'bge-small-en-v1.5';
+
+  // Simulates `antfly termite list --models-dir ~/.antfly/models` output when
+  // the model IS present (full name in NAME column, also in SOURCE column).
+  const PRESENT_OUTPUT = `Local models in /Users/dev/.antfly/models:
+
+NAME                    TYPE      SIZE      VARIANTS  SOURCE
+BAAI/bge-small-en-v1.5  embedder  127.8 MB            BAAI/bge-small-en-v1.5
+`;
+
+  // Output when NO models are installed (the bug scenario: server's models-dir
+  // is empty, but ~/.termite/models has the model — the old code read the wrong
+  // directory and would never see "No models found").
+  const EMPTY_OUTPUT = `Local models in /Users/dev/.antfly/models:
+
+NAME  TYPE  SIZE  VARIANTS  SOURCE
+No models found locally.
+
+Use 'antfly termite pull <model-name>' to download models.
+Use 'antfly termite list --remote' to see available models.
+`;
+
+  // Output with a DIFFERENT model that happens to contain the short name as a
+  // suffix — the old substring check would incorrectly return true here.
+  const OTHER_MODEL_OUTPUT = `Local models in /Users/dev/.antfly/models:
+
+NAME                          TYPE      SIZE      VARIANTS  SOURCE
+vendor/other-bge-small-en-v1.5  embedder  200.0 MB            vendor/other-bge-small-en-v1.5
+`;
+
+  it('returns true when full model name is present in output', () => {
+    expect(modelPresentInOutput(FULL_NAME, PRESENT_OUTPUT)).toBe(true);
+  });
+
+  it('returns true when only short name is present as a standalone token', () => {
+    const outputWithShortName = `Local models:\n\n${SHORT_NAME}  embedder  127 MB\n`;
+    expect(modelPresentInOutput(FULL_NAME, outputWithShortName)).toBe(true);
+  });
+
+  it('returns false when models directory is empty (server has no models)', () => {
+    // This is the core regression: old code checked ~/.termite/models which had
+    // the model, new code checks ~/.antfly/models which was empty. When empty,
+    // hasModel must return false so pullModel is invoked.
+    expect(modelPresentInOutput(FULL_NAME, EMPTY_OUTPUT)).toBe(false);
+  });
+
+  it('returns false when a different model shares the short name as a suffix', () => {
+    // Old bug: output.includes("bge-small-en-v1.5") matched
+    // "vendor/other-bge-small-en-v1.5" — false positive.
+    expect(modelPresentInOutput(FULL_NAME, OTHER_MODEL_OUTPUT)).toBe(false);
+  });
+
+  it('returns false for completely unrelated output', () => {
+    expect(modelPresentInOutput(FULL_NAME, 'No models found locally.')).toBe(false);
+  });
+
+  it('handles model names without an org prefix', () => {
+    // model = "mxbai-embed-large-v1" (no slash)
+    const bareModel = 'mxbai-embed-large-v1';
+    const output = `NAME                TYPE\nmxbai-embed-large-v1  embedder\n`;
+    expect(modelPresentInOutput(bareModel, output)).toBe(true);
+  });
+
+  it('handles bare model not present', () => {
+    const bareModel = 'mxbai-embed-large-v1';
+    expect(modelPresentInOutput(bareModel, EMPTY_OUTPUT)).toBe(false);
+  });
+});

packages/cli/src/utils/antfly.ts

@@ -5,6 +5,8 @@
  */
 
 import { execSync, spawn } from 'node:child_process';
+import { homedir } from 'node:os';
+import { join } from 'node:path';
 import { logger } from './logger.js';
 
 const DEFAULT_ANTFLY_URL = process.env.ANTFLY_URL ?? 'http://localhost:18080/api/v1';
@@ -14,6 +16,18 @@ const DOCKER_PORT = 18080;
 const STARTUP_TIMEOUT_MS = 30_000;
 const POLL_INTERVAL_MS = 500;
 
+/**
+ * The Termite models directory used by the running Antfly swarm server.
+ *
+ * `antfly swarm` uses `--data-dir` (default: ~/.antfly) as its root for all
+ * storage, including Termite models at {data-dir}/models.
+ * `antfly termite list/pull` defaults to --models-dir ~/.termite/models, which
+ * is a DIFFERENT path. We must always pass --models-dir explicitly so that
+ * `pullModel` and `hasModel` operate on the same directory the server uses.
+ */
+const ANTFLY_DATA_DIR = process.env.ANTFLY_DATA_DIR ?? join(homedir(), '.antfly');
+const TERMITE_MODELS_DIR = join(ANTFLY_DATA_DIR, 'models');
+
 /**
  * Ensure antfly is running. Auto-starts if needed.
  *
@@ -32,10 +46,14 @@ export async function ensureAntfly(options?: { quiet?: boolean }): Promise<strin
     if (!options?.quiet) logger.info('Starting Antfly server...');
     // Use custom ports to avoid 8080 conflicts (Docker, other services).
     // metadata-api on 18080 (our default), store-api on 18381, raft on 19017/19021.
+    // --data-dir is passed explicitly so the server's embedded Termite node stores
+    // models in the same directory that pullModel/hasModel use (TERMITE_MODELS_DIR).
     const child = spawn(
       'antfly',
       [
         'swarm',
+        '--data-dir',
+        ANTFLY_DATA_DIR,
         '--metadata-api',
         'http://0.0.0.0:18080',
         '--store-api',
@@ -187,9 +205,14 @@ export function getNativeVersion(): string | null {
 
 /**
  * Pull a Termite embedding model (native binary).
+ *
+ * Always targets TERMITE_MODELS_DIR so the model ends up in the same directory
+ * the running Antfly swarm server uses for its embedded Termite node.
  */
 export function pullModel(model: string): void {
-  execSync(`antfly termite pull ${model}`, { stdio: 'inherit' });
+  execSync(`antfly termite pull --models-dir ${TERMITE_MODELS_DIR} ${model}`, {
+    stdio: 'inherit',
+  });
 }
 
 /**
@@ -201,33 +224,63 @@ export function pullModelDocker(model: string): void {
 }
 
 /**
- * Check if a Termite model is available locally (native binary).
+ * Check if a Termite model is available in the directory used by the running
+ * Antfly swarm server (TERMITE_MODELS_DIR = ~/.antfly/models by default).
+ *
+ * Checks for the full model name first (e.g. "BAAI/bge-small-en-v1.5"), then
+ * the short name as a whole word (e.g. "bge-small-en-v1.5"). Previously used
+ * a simple substring match on the short name, which caused false positives when
+ * `antfly termite list` defaulted to ~/.termite/models — a different directory
+ * from the one the server reads, so the model appeared present but was not
+ * available to the server during embedding.
  */
 export function hasModel(model: string): boolean {
   try {
-    const output = execSync('antfly termite list', {
+    const output = execSync(`antfly termite list --models-dir ${TERMITE_MODELS_DIR}`, {
       encoding: 'utf-8',
       stdio: ['pipe', 'pipe', 'pipe'],
     });
-    const shortName = model.split('/').pop() ?? model;
-    return output.includes(shortName);
+    return modelPresentInOutput(model, output);
   } catch {
     return false;
   }
 }
 
 /**
  * Check if a Termite model is available inside the Docker container.
+ *
+ * Checks for the full model name first (e.g. "BAAI/bge-small-en-v1.5"), then
+ * the short name as a whole word (e.g. "bge-small-en-v1.5"). Simple substring
+ * matching on the short name was causing false positives when other models or
+ * partial download records shared the suffix.
  */
 export function hasModelDocker(model: string): boolean {
   try {
     const output = execSync(`docker exec ${CONTAINER_NAME} /antfly termite list`, {
       encoding: 'utf-8',
       stdio: ['pipe', 'pipe', 'pipe'],
     });
-    const shortName = model.split('/').pop() ?? model;
-    return output.includes(shortName);
+    return modelPresentInOutput(model, output);
   } catch {
     return false;
   }
 }
+
+/**
+ * Return true when the model name is present in `antfly termite list` output.
+ *
+ * Strategy (most-specific first):
+ *   1. Full name exact match  — "BAAI/bge-small-en-v1.5" appears verbatim.
+ *   2. Short name word-boundary — "bge-small-en-v1.5" appears as a whole token
+ *      (not as a suffix of a different model name).
+ */
+function modelPresentInOutput(model: string, output: string): boolean {
+  // Full name check (covers "BAAI/bge-small-en-v1.5" style output)
+  if (output.includes(model)) return true;
+
+  // Short name check with word-boundary anchors so "bge-small-en-v1.5" does not
+  // match inside "other-bge-small-en-v1.5" or a partial download entry.
+  const shortName = model.split('/').pop() ?? model;
+  const escaped = shortName.replace(/[.*+?^${}()|[\]\\]/g, '\\$&');
+  return new RegExp(`(?<![\\w/-])${escaped}(?![\\w/-])`).test(output);
+}

website/content/docs/install.mdx

@@ -3,7 +3,7 @@
 ## Requirements
 
 - **Node.js 22+** (LTS recommended)
-- **[Antfly](https://antfly.io)** — search backend (`brew install --cask antflydb/antfly/antfly`)
+- **[Antfly](https://antfly.io)** — search backend (installed automatically by `dev setup`)
 - **Cursor** or **Claude Code** (for MCP integration)
 
 ## Install dev-agent

website/content/docs/quickstart.mdx

@@ -5,7 +5,7 @@ Get from zero to semantic search in 5 minutes.
 ## Prerequisites
 
 - Node.js 22+ installed
-- [Antfly](https://antfly.io) — search backend (`brew install --cask antflydb/antfly/antfly`)
+- [Antfly](https://antfly.io) — search backend (installed automatically by `dev setup`)
 - Cursor IDE (or Claude Code)
 - A code repository to index
 

website/content/index.mdx

@@ -80,7 +80,6 @@ The key difference: semantic search finds code by **meaning**, not text matching
 
 ```bash
 npm install -g @prosdevlab/dev-agent
-brew install --cask antflydb/antfly/antfly
 ```
 
 ### Setup and index

website/content/latest-version.ts

@@ -4,10 +4,10 @@
  */
 
 export const latestVersion = {
-  version: '0.10.2',
-  title: 'MCP Install Fix & Dependency Cleanup',
+  version: '0.10.3',
+  title: 'Fix Setup/Index Model Directory Mismatch',
   date: 'March 30, 2026',
   summary:
-    'Fixed dev mcp install check, removed dead metrics module and better-sqlite3 dependency.',
-  link: '/updates#v0102--mcp-install-fix--dependency-cleanup',
+    'Fixed dev setup reporting model ready while dev index fails with "model not found" due to mismatched model directories.',
+  link: '/updates#v0103--fix-setupindex-model-directory-mismatch',
 } as const;

website/content/updates/index.mdx

@@ -9,6 +9,18 @@ What's new in dev-agent. We ship improvements regularly to help AI assistants un
 
 ---
 
+## v0.10.3 — Fix Setup/Index Model Directory Mismatch
+
+*March 30, 2026*
+
+**Fixed `dev setup` reporting model ready while `dev index` fails with "model not found".**
+
+- `hasModel`/`pullModel` used `~/.termite/models` but the running Antfly server looked in `~/.antfly/models` — both now use a shared `--models-dir` pointing at the server's data directory
+- Improved model name matching to avoid false positives from substring collisions
+- Added unit tests for model detection logic
+
+---
+
 ## v0.10.2 — MCP Install Fix & Dependency Cleanup
 
 *March 30, 2026*