preppergpt 0.1.0 → 0.1.2

package/README.md

@@ -1,32 +1,45 @@
 # PrepperGPT
 
-PrepperGPT packages a local-first ChatGPT-like experience for Linux machines.
-It uses upstream OpenWebUI for the app shell and adds a hardware detector,
-model planner, Docker Compose runtime, local sidecars, and a practical
-PrepperGPT field-kit theme.
+PrepperGPT packages a local-first ChatGPT-like experience for post-apocalyptic
+or long-duration outage scenarios where hosted AI services are unavailable. It
+uses upstream OpenWebUI for the app shell and adds a hardware detector, model
+planner, Docker Compose runtime, local sidecars, and a practical PrepperGPT
+field-kit theme.
 
 The first release targets Linux with NVIDIA GPUs first, with CPU fallback where
 possible. It is an online installer: model and container downloads require a
 working network during setup.
 
+PrepperGPT optimizes for survivability over cloud-like latency. On very large
+local models, very low tokens/sec is acceptable because the alternative in the
+target scenario is no assistant at all.
+
 ## Install
 
-Until the npm package is published, install from GitHub:
+Install from npm:
 
 ```bash
-git clone https://github.com/teamslop/preppergpt.git
-cd preppergpt
-node bin/preppergpt.js install --profile balanced
-node bin/preppergpt.js start
+npx preppergpt install --profile balanced
+preppergpt start
 ```
 
-After npm publication:
+Or install globally:
 
 ```bash
-npx preppergpt install --profile balanced
+npm install -g preppergpt
+preppergpt install --profile balanced
 preppergpt start
 ```
 
+GitHub source install:
+
+```bash
+git clone https://github.com/teamslop/preppergpt.git
+cd preppergpt
+node bin/preppergpt.js install --profile balanced
+node bin/preppergpt.js start
+```
+
 Other profiles:
 
 ```bash
@@ -54,12 +67,14 @@ preppergpt stop
 preppergpt status
 preppergpt doctor
 preppergpt switch-profile --profile speed
+preppergpt bundle whisper
 ```
 
 ## Profiles
 
 - `intelligence`: chooses the strongest local reasoning route that fits the
-  machine, preferring GLM 5.2 Q4 and long-context coding routes when available.
+  machine, preferring GLM 5.2 Q8 on enterprise hardware, then GLM 5.2 Q4, then
+  long-context coding routes when available.
 - `speed`: chooses smaller GPU-friendly routes and makes low-latency chat the
   default.
 - `balanced`: uses the local auto-router as the default and keeps reasoning,
@@ -70,11 +85,20 @@ and route ordering into the generated compose override.
 
 ## Model Assets
 
-Some routes can be pulled by the runtime, while very large routes such as GLM
-5.2 Q4 and Flux weights are marked as manual or external in
+PrepperGPT installs a bundled local Whisper Base STT cache during
+`preppergpt install`. It is stored under `~/.preppergpt/data/models/whisper/base`
+by default and mounted into OpenWebUI, so speech-to-text works from local files
+after setup.
+
+Some other routes can be pulled by the runtime, while very large routes such as
+GLM 5.2 Q8/Q4 and Flux weights are marked as manual or external in
 `profiles/models.json`. `preppergpt doctor` reports which selected routes still
 need local files or endpoints.
 
+The GLM 5.2 Q8 route is intended for an enterprise/off-grid bunker-class host:
+large RAM, fast NVMe, and patience for slow local generation when no hosted
+service remains available.
+
 ## Publishing
 
 The package is designed to be published as:
@@ -84,7 +108,7 @@ npm publish --access public
 ```
 
 Publishing requires an authenticated npm account with permission to publish the
-currently unclaimed `preppergpt` package name.
+`preppergpt` package.
 
 The source repository is expected at:
 

package/compose/preppergpt.yaml

@@ -13,6 +13,7 @@ services:
         hard: 1048576
     volumes:
       - ${PREPPERGPT_DATA_DIR:?set PREPPERGPT_DATA_DIR}/openwebui:/app/backend/data
+      - ${PREPPERGPT_MODELS_DIR:?set PREPPERGPT_MODELS_DIR}:/models:ro
       - ../services/comfyui:/app/backend/data/parity-comfyui:ro
       - ../themes/preppergpt/static/favicon.svg:/app/backend/open_webui/static/favicon.svg:ro
       - ../themes/preppergpt/static/logo.svg:/app/backend/open_webui/static/logo.svg:ro
@@ -26,8 +27,8 @@ services:
       ENABLE_OLLAMA_API: "True"
       OLLAMA_BASE_URLS: "${OLLAMA_BASE_URL:-http://127.0.0.1:11434}"
       ENABLE_OPENAI_API: "True"
-      OPENAI_API_BASE_URLS: "${SLOCODE_BASE_URL:-http://127.0.0.1:11438/v1};${GLM52_BASE_URL:-http://127.0.0.1:11441/v1};http://127.0.0.1:18041/v1;http://127.0.0.1:18043/v1;http://127.0.0.1:18044/v1"
-      OPENAI_API_KEYS: "slopcode;glm52;deep-research;local-agent;local-vision"
+      OPENAI_API_BASE_URLS: "${GLM52_Q8_BASE_URL:-http://127.0.0.1:11446/v1};${SLOCODE_BASE_URL:-http://127.0.0.1:11438/v1};${GLM52_BASE_URL:-http://127.0.0.1:11441/v1};http://127.0.0.1:18041/v1;http://127.0.0.1:18043/v1;http://127.0.0.1:18044/v1"
+      OPENAI_API_KEYS: "glm52-q8;slopcode;glm52;deep-research;local-agent;local-vision"
       ENABLE_DIRECT_CONNECTIONS: "True"
       DEFAULT_MODELS: "${PREPPERGPT_DEFAULT_MODEL:-local-chatgpt-auto}"
       MODEL_ORDER_LIST: "${PREPPERGPT_MODEL_ORDER_LIST:-[\"local-chatgpt-auto\"]}"
@@ -69,9 +70,10 @@ services:
       CODE_INTERPRETER_JUPYTER_URL: "http://127.0.0.1:8888"
       CODE_INTERPRETER_JUPYTER_AUTH: "token"
       CODE_INTERPRETER_JUPYTER_AUTH_TOKEN: "${JUPYTER_TOKEN:?set JUPYTER_TOKEN}"
-      WHISPER_MODEL: "large-v3"
+      WHISPER_MODEL: "${PREPPERGPT_WHISPER_MODEL_PATH:-/models/whisper/base}"
+      WHISPER_MODEL_DIR: "/models/whisper"
       WHISPER_COMPUTE_TYPE: "int8"
-      WHISPER_MODEL_AUTO_UPDATE: "True"
+      WHISPER_MODEL_AUTO_UPDATE: "False"
       WHISPER_VAD_FILTER: "True"
       WHISPER_MULTILINGUAL: "True"
       ENABLE_IMAGE_GENERATION: "True"
@@ -147,8 +149,8 @@ services:
       DEEP_RESEARCH_PORT: "18041"
       DEEP_RESEARCH_PUBLIC_BASE_URL: "http://127.0.0.1:18041"
       DEEP_RESEARCH_MODEL_ID: "deep-research-glm52"
-      DEEP_RESEARCH_MODEL: "${DEEP_RESEARCH_MODEL:-glm52-q4-local}"
-      DEEP_RESEARCH_GLM_BASE_URL: "${GLM52_BASE_URL:-http://127.0.0.1:11441/v1}"
+      DEEP_RESEARCH_MODEL: "${PREPPERGPT_GLM_MODEL:-glm52-q4-local}"
+      DEEP_RESEARCH_GLM_BASE_URL: "${PREPPERGPT_GLM_BASE_URL:-http://127.0.0.1:11441/v1}"
       DEEP_RESEARCH_SEARXNG_URL: "http://127.0.0.1:18080/search"
       DEEP_RESEARCH_TIKA_URL: "http://127.0.0.1:9998/tika"
       DEEP_RESEARCH_LOCAL_APP_CONNECTOR_URL: "http://127.0.0.1:18042"
@@ -193,8 +195,8 @@ services:
       LOCAL_AGENT_PORT: "18043"
       LOCAL_AGENT_PUBLIC_BASE_URL: "http://127.0.0.1:18043"
       LOCAL_AGENT_MODEL_ID: "local-agent-glm52"
-      LOCAL_AGENT_GLM_MODEL: "glm52-q4-local"
-      LOCAL_AGENT_GLM_BASE_URL: "${GLM52_BASE_URL:-http://127.0.0.1:11441/v1}"
+      LOCAL_AGENT_GLM_MODEL: "${PREPPERGPT_GLM_MODEL:-glm52-q4-local}"
+      LOCAL_AGENT_GLM_BASE_URL: "${PREPPERGPT_GLM_BASE_URL:-http://127.0.0.1:11441/v1}"
       LOCAL_AGENT_AUTO_ROUTER_MODEL_ID: "local-auto-router"
       LOCAL_AGENT_AUTO_ROUTER_FAST_MODEL: "gemma4:12b-256k-gpu"
       LOCAL_AGENT_AUTO_ROUTER_FAST_BASE_URL: "http://127.0.0.1:11434/v1"

package/docs/bundles.md

@@ -0,0 +1,30 @@
+# Bundles
+
+PrepperGPT keeps npm lightweight but installs small always-on local assets
+during setup.
+
+## Whisper Base
+
+`preppergpt install` downloads the MIT-licensed `Systran/faster-whisper-base`
+CTranslate2 model into:
+
+```text
+~/.preppergpt/data/models/whisper/base
+```
+
+OpenWebUI receives:
+
+```text
+WHISPER_MODEL=/models/whisper/base
+WHISPER_MODEL_DIR=/models/whisper
+WHISPER_MODEL_AUTO_UPDATE=False
+```
+
+To repair or refresh the bundle:
+
+```bash
+preppergpt bundle whisper
+preppergpt bundle whisper --force
+```
+
+Source: https://huggingface.co/Systran/faster-whisper-base

package/docs/hardware.md

@@ -1,7 +1,9 @@
 # Hardware Guide
 
 PrepperGPT works best on Linux with an NVIDIA GPU and enough NVMe space for
-model weights.
+model weights. It is designed for post-apocalyptic or long-duration outage
+scenarios, so the high-end GLM tiers deliberately favor local availability and
+answer quality over hosted-service latency.
 
 Recommended starting points:
 
@@ -9,7 +11,22 @@ Recommended starting points:
 - Balanced profile: 32-64 GB RAM, 12-24 GB VRAM, 120 GB free disk.
 - Intelligence profile: 96 GB RAM or more, fast NVMe, and hundreds of GB free
   for GLM 5.2 Q4 or similar large weights.
+- Enterprise 8-bit GLM tier: 256 GB RAM or more, 48-80 GB VRAM preferred,
+  and 1.5-2 TB of fast NVMe for GLM 5.2 Q8 plus working/cache room.
 
 The installer reserves about 15-20% VRAM headroom when deciding whether a model
 fits. If a large manual model is selected, `preppergpt doctor` explains the
 endpoint or file path that must be provided.
+
+Very low tokens/sec is acceptable for the GLM 5.2 Q8 tier because that tier is
+for situations where there is no cloud model to fall back to.
+
+## Hardware Matrix
+
+| Tier | Typical specs | PrepperGPT routes |
+| --- | --- | --- |
+| Basic CPU laptop | 16 GB RAM, no GPU, 80 GB disk | `local-chatgpt-auto`, `llama3.1:8b`, `local-vision-moondream2`, bundled Whisper |
+| Mid NVIDIA | 64 GB RAM, 12 GB usable VRAM, 250 GB disk | Gemma fast lane, Qwen coder fallback, local vision, bundled Whisper |
+| High NVIDIA | 128 GB RAM, 24 GB VRAM, 750 GB NVMe | GLM 5.2 Q4 configured, Slopcode/Qwen configured, Gemma fast lane, Flux configured |
+| Full PrepperGPT rig | 128+ GB RAM, 24+ GB VRAM, 1 TB NVMe, GLM/Slopcode/Flux files present | GLM 5.2 Q4 primary, Slopcode coding, Gemma fast lane, Deep Research, Agent, Vision, Flux, Whisper |
+| Enterprise 8-bit GLM rig | 256+ GB RAM, 48-80+ GB VRAM preferred, 1.5-2 TB fast NVMe | `glm52-q8-local` primary for Max Intelligence, `glm52-q4-local` fallback, Slopcode/Qwen coding, Gemma fast lane, full sidecar stack |

package/docs/model-sources.md

@@ -3,9 +3,10 @@
 PrepperGPT separates routing from model licensing and distribution.
 
 - Ollama models are pulled by the local Ollama runtime when available.
-- OpenWebUI STT models are downloaded by OpenWebUI/faster-whisper.
+- Whisper Base STT is installer-cached from `Systran/faster-whisper-base`
+  under the local PrepperGPT model directory and mounted into OpenWebUI.
 - Hugging Face vision models are downloaded by the local vision sidecar.
-- Very large GLM, Slopcode, and Flux assets are marked as manual or external
+- Very large GLM Q8/Q4, Slopcode, and Flux assets are marked as manual or external
   until a license-compatible public download source is configured.
 
 Manual routes are still added to OpenWebUI. They become live when their local

package/docs/preppergpt-local-parity-map.md

@@ -1,10 +1,12 @@
 # PrepperGPT Local Parity Map
 
-PrepperGPT packages the local ChatGPT-like stack around OpenWebUI:
+PrepperGPT packages the local ChatGPT-like stack around OpenWebUI for resilient
+local use when hosted AI services are unavailable:
 
 - OpenWebUI UI at `http://127.0.0.1:8080`
 - Ollama fast local models at `http://127.0.0.1:11434`
-- Optional GLM 5.2 route at `http://127.0.0.1:11441/v1`
+- Optional GLM 5.2 Q8 route at `http://127.0.0.1:11446/v1`
+- Optional GLM 5.2 Q4 route at `http://127.0.0.1:11441/v1`
 - Optional Slopcode/Qwen route at `http://127.0.0.1:11438/v1`
 - Deep research sidecar at `http://127.0.0.1:18041/v1`
 - Local scheduler connector at `http://127.0.0.1:18042`

package/installer/cli.mjs

@@ -1,12 +1,13 @@
 import fs from "node:fs";
 import http from "node:http";
+import { ensureWhisperBundle, modelDirs, whisperBundleStatus } from "./lib/bundles.mjs";
 import { detectMachine } from "./lib/detect.mjs";
 import { buildPlan, normalizeProfile } from "./lib/planner.mjs";
 import { packageRoot, runtimePaths } from "./lib/paths.mjs";
 import { renderInstall } from "./lib/render.mjs";
 import { commandResult, parseArgs, readJson, shellQuote } from "./lib/util.mjs";
 
-const VERSION = "0.1.0";
+const VERSION = "0.1.2";
 
 function usage() {
   return `PrepperGPT ${VERSION}
@@ -14,11 +15,12 @@ function usage() {
 Usage:
   preppergpt detect [--json]
   preppergpt plan --profile balanced|intelligence|speed [--json]
-  preppergpt install --profile balanced|intelligence|speed [--dry-run] [--home PATH]
+  preppergpt install --profile balanced|intelligence|speed [--dry-run] [--skip-bundles] [--home PATH]
   preppergpt start [--home PATH]
   preppergpt stop [--home PATH]
   preppergpt status [--home PATH] [--json]
   preppergpt doctor [--profile balanced|intelligence|speed] [--home PATH]
+  preppergpt bundle whisper [--home PATH] [--force]
   preppergpt switch-profile --profile balanced|intelligence|speed [--home PATH]
   preppergpt version
 `;
@@ -122,6 +124,11 @@ async function commandInstall(flags) {
     return;
   }
   const paths = renderInstall(plan, detection, { home });
+  if (!flags.skip_bundles) {
+    console.log("Installing bundled Whisper base STT model...");
+    const bundle = await ensureWhisperBundle(paths.whisperHostDir, { force: Boolean(flags.force_bundle) });
+    console.log(`Whisper bundle: ${bundle.ready ? "ready" : "not ready"} at ${paths.whisperHostDir}`);
+  }
   console.log(`Wrote ${paths.envFile}`);
   console.log(`Wrote ${paths.generatedCompose}`);
   console.log(`Wrote ${paths.modelPlan}`);
@@ -187,6 +194,7 @@ async function commandStatus(flags) {
 }
 
 async function commandDoctor(flags) {
+  const paths = runtimePaths(flags.home);
   const detection = await detectMachine();
   const plan = buildPlan(detection, profileFrom(flags));
   printPlan(plan);
@@ -200,6 +208,28 @@ async function commandDoctor(flags) {
       console.log(`  port ${port}: occupied`);
     }
   }
+  const dirs = modelDirs(paths);
+  const whisper = whisperBundleStatus(dirs.whisperHostDir);
+  console.log(`  whisper-base bundle: ${whisper.ready ? "ok" : `missing ${whisper.missing.length} files`} (${dirs.whisperHostDir})`);
+}
+
+async function commandBundle(flags, positional) {
+  const name = positional[1] || "whisper";
+  if (!["whisper", "whisper-base"].includes(name)) {
+    throw new Error(`Unknown bundle: ${name}`);
+  }
+  const paths = runtimePaths(flags.home);
+  const dirs = modelDirs(paths);
+  const bundle = await ensureWhisperBundle(dirs.whisperHostDir, {
+    force: Boolean(flags.force),
+    dryRun: Boolean(flags.dry_run)
+  });
+  console.log(`Whisper bundle ${bundle.ready ? "ready" : "not ready"} at ${dirs.whisperHostDir}`);
+  if (bundle.missing?.length) {
+    for (const file of bundle.missing) {
+      console.log(`  missing ${file}`);
+    }
+  }
 }
 
 export async function runCli(argv) {
@@ -220,6 +250,7 @@ export async function runCli(argv) {
   if (command === "stop") return commandStop(flags);
   if (command === "status") return commandStatus(flags);
   if (command === "doctor") return commandDoctor(flags);
+  if (command === "bundle") return commandBundle(flags, positional);
   if (command === "switch-profile") return commandSwitchProfile(flags);
   throw new Error(`Unknown command: ${command}\n\n${usage()}`);
 }

package/installer/lib/bundles.mjs

@@ -0,0 +1,107 @@
+import fs from "node:fs";
+import path from "node:path";
+import { Readable } from "node:stream";
+import { pipeline } from "node:stream/promises";
+import { readJson, writeJson } from "./util.mjs";
+
+export const WHISPER_BUNDLE = {
+  id: "whisper-base",
+  name: "Whisper Base STT Bundle",
+  repo: "Systran/faster-whisper-base",
+  revision: "main",
+  license: "MIT",
+  modelPathInContainer: "/models/whisper/base",
+  files: ["config.json", "model.bin", "tokenizer.json", "vocabulary.txt", "README.md"],
+  description: "CTranslate2 faster-whisper conversion of openai/whisper-base for local OpenWebUI STT."
+};
+
+function parseEnvFile(file) {
+  if (!fs.existsSync(file)) {
+    return {};
+  }
+  const entries = {};
+  for (const line of fs.readFileSync(file, "utf8").split(/\r?\n/)) {
+    if (!line || line.trim().startsWith("#") || !line.includes("=")) {
+      continue;
+    }
+    const [key, ...valueParts] = line.split("=");
+    let value = valueParts.join("=");
+    if ((value.startsWith('"') && value.endsWith('"')) || (value.startsWith("'") && value.endsWith("'"))) {
+      value = value.slice(1, -1);
+    }
+    entries[key] = value;
+  }
+  return entries;
+}
+
+export function modelDirs(paths) {
+  const env = parseEnvFile(paths.envFile);
+  const modelsDir = process.env.PREPPERGPT_MODELS_DIR || env.PREPPERGPT_MODELS_DIR || path.join(paths.dataDir, "models");
+  const whisperHostDir = path.join(modelsDir, "whisper", "base");
+  return { modelsDir, whisperHostDir };
+}
+
+export function whisperBundleStatus(targetDir) {
+  const files = WHISPER_BUNDLE.files.map((file) => path.join(targetDir, file));
+  const missing = files.filter((file) => !fs.existsSync(file));
+  let manifest = null;
+  const manifestPath = path.join(targetDir, "preppergpt-bundle.json");
+  if (fs.existsSync(manifestPath)) {
+    try {
+      manifest = readJson(manifestPath);
+    } catch {
+      manifest = null;
+    }
+  }
+  return {
+    id: WHISPER_BUNDLE.id,
+    targetDir,
+    ready: missing.length === 0,
+    missing,
+    manifest
+  };
+}
+
+async function downloadFile(url, targetFile) {
+  const response = await fetch(url, {
+    headers: {
+      "User-Agent": "preppergpt/0.1"
+    },
+    redirect: "follow"
+  });
+  if (!response.ok || !response.body) {
+    throw new Error(`Failed to download ${url}: HTTP ${response.status}`);
+  }
+  fs.mkdirSync(path.dirname(targetFile), { recursive: true });
+  const tmp = `${targetFile}.tmp-${process.pid}`;
+  await pipeline(Readable.fromWeb(response.body), fs.createWriteStream(tmp));
+  fs.renameSync(tmp, targetFile);
+}
+
+export async function ensureWhisperBundle(targetDir, options = {}) {
+  const status = whisperBundleStatus(targetDir);
+  if (status.ready && !options.force) {
+    return { ...status, changed: false };
+  }
+  fs.mkdirSync(targetDir, { recursive: true });
+  if (options.dryRun) {
+    return { ...status, changed: false, dryRun: true };
+  }
+  for (const file of WHISPER_BUNDLE.files) {
+    const targetFile = path.join(targetDir, file);
+    if (fs.existsSync(targetFile) && !options.force) {
+      continue;
+    }
+    const url = `https://huggingface.co/${WHISPER_BUNDLE.repo}/resolve/${WHISPER_BUNDLE.revision}/${file}`;
+    if (!options.quiet) {
+      console.log(`Downloading ${WHISPER_BUNDLE.repo}/${file}`);
+    }
+    await downloadFile(url, targetFile);
+  }
+  writeJson(path.join(targetDir, "preppergpt-bundle.json"), {
+    ...WHISPER_BUNDLE,
+    installedAt: new Date().toISOString(),
+    source: `https://huggingface.co/${WHISPER_BUNDLE.repo}`
+  });
+  return { ...whisperBundleStatus(targetDir), changed: true };
+}

package/installer/lib/planner.mjs

@@ -62,7 +62,9 @@ function chooseFirst(candidates, models, detection) {
       continue;
     }
     const failures = requirementFailures(model, detection);
-    if (failures.length === 0 || model.source?.type === "manual" || model.source?.type === "external") {
+    const canUseExternalFallback =
+      ["manual", "external"].includes(model.source?.type) && !model.source?.requiresHardwareFit;
+    if (failures.length === 0 || canUseExternalFallback) {
       return { model, skipped };
     }
     skipped.push({ id, reasons: failures });
@@ -91,8 +93,9 @@ export function buildPlan(detection, requestedProfile = "balanced", catalog = lo
     }
   }
 
+  const defaultModel = selected.chat?.id || priorities.defaultModel;
   const routeIds = unique([
-    priorities.defaultModel,
+    defaultModel,
     selected.chat?.id,
     selected.fast?.id,
     selected.reasoning?.id,
@@ -136,7 +139,7 @@ export function buildPlan(detection, requestedProfile = "balanced", catalog = lo
     generatedAt: new Date().toISOString(),
     profile,
     profileLabel: priorities.label,
-    defaultModel: priorities.defaultModel,
+    defaultModel,
     routeIds,
     selected,
     skipped,
@@ -144,7 +147,7 @@ export function buildPlan(detection, requestedProfile = "balanced", catalog = lo
     estimates: estimatePlan(profile, selected),
     env: {
       PREPPERGPT_PROFILE: profile,
-      PREPPERGPT_DEFAULT_MODEL: priorities.defaultModel,
+      PREPPERGPT_DEFAULT_MODEL: defaultModel,
       PREPPERGPT_MODEL_ORDER_LIST: JSON.stringify(routeIds)
     },
     warnings

package/installer/lib/render.mjs

@@ -1,5 +1,7 @@
 import crypto from "node:crypto";
 import fs from "node:fs";
+import path from "node:path";
+import { WHISPER_BUNDLE } from "./bundles.mjs";
 import { packagedPath, runtimePaths } from "./paths.mjs";
 import { envQuote, writeJson, writeText } from "./util.mjs";
 
@@ -10,6 +12,12 @@ function secret(bytes = 24) {
 function envFile(plan, paths, detection) {
   const dataDir = process.env.PREPPERGPT_DATA_DIR || paths.dataDir;
   const modelsDir = process.env.PREPPERGPT_MODELS_DIR || `${dataDir}/models`;
+  const whisperHostDir = path.join(modelsDir, "whisper", "base");
+  const selectedReasoningModel = plan.selected?.reasoning?.id || "glm52-q4-local";
+  const selectedGlmBaseUrl =
+    selectedReasoningModel === "glm52-q8-local"
+      ? process.env.GLM52_Q8_BASE_URL || "http://127.0.0.1:11446/v1"
+      : process.env.GLM52_BASE_URL || "http://127.0.0.1:11441/v1";
   const adminPassword = process.env.PREPPERGPT_ADMIN_PASSWORD || secret(18);
   const jupyterToken = process.env.JUPYTER_TOKEN || secret(18);
   const searxngSecret = process.env.SEARXNG_SECRET_KEY || secret(24);
@@ -17,9 +25,14 @@ function envFile(plan, paths, detection) {
     PREPPERGPT_PROFILE: plan.profile,
     PREPPERGPT_DATA_DIR: dataDir,
     PREPPERGPT_MODELS_DIR: modelsDir,
+    PREPPERGPT_WHISPER_HOST_DIR: whisperHostDir,
+    PREPPERGPT_WHISPER_MODEL: WHISPER_BUNDLE.id,
+    PREPPERGPT_WHISPER_MODEL_PATH: WHISPER_BUNDLE.modelPathInContainer,
     PREPPERGPT_PORT: process.env.PREPPERGPT_PORT || "8080",
     PREPPERGPT_DEFAULT_MODEL: plan.defaultModel,
     PREPPERGPT_MODEL_ORDER_LIST: JSON.stringify(plan.routeIds),
+    PREPPERGPT_GLM_MODEL: selectedReasoningModel,
+    PREPPERGPT_GLM_BASE_URL: selectedGlmBaseUrl,
     PREPPERGPT_DOCKER_GPUS: detection.gpus?.length ? "all" : "",
     WEBUI_NAME: "PrepperGPT",
     WEBUI_ADMIN_EMAIL: process.env.WEBUI_ADMIN_EMAIL || "admin@preppergpt.local",
@@ -29,6 +42,7 @@ function envFile(plan, paths, detection) {
     JUPYTER_TOKEN: jupyterToken,
     SEARXNG_SECRET_KEY: searxngSecret,
     GLM52_BASE_URL: process.env.GLM52_BASE_URL || "http://127.0.0.1:11441/v1",
+    GLM52_Q8_BASE_URL: process.env.GLM52_Q8_BASE_URL || "http://127.0.0.1:11446/v1",
     SLOCODE_BASE_URL: process.env.SLOCODE_BASE_URL || "http://127.0.0.1:11438/v1",
     OLLAMA_BASE_URL: process.env.OLLAMA_BASE_URL || "http://127.0.0.1:11434"
   };
@@ -60,17 +74,23 @@ function generatedCompose(plan, detection) {
 
 export function renderInstall(plan, detection, options = {}) {
   const paths = runtimePaths(options.home);
+  const dataDir = process.env.PREPPERGPT_DATA_DIR || paths.dataDir;
+  const modelsDir = process.env.PREPPERGPT_MODELS_DIR || `${dataDir}/models`;
+  const whisperHostDir = path.join(modelsDir, "whisper", "base");
   fs.mkdirSync(paths.root, { recursive: true });
   fs.mkdirSync(paths.dataDir, { recursive: true });
   fs.mkdirSync(paths.composeDir, { recursive: true });
   fs.mkdirSync(`${paths.dataDir}/preppergpt`, { recursive: true });
-  fs.mkdirSync(`${paths.dataDir}/models`, { recursive: true });
+  fs.mkdirSync(modelsDir, { recursive: true });
+  fs.mkdirSync(whisperHostDir, { recursive: true });
   writeText(paths.envFile, envFile(plan, paths, detection), 0o600);
   writeText(paths.generatedCompose, generatedCompose(plan, detection));
   writeJson(paths.modelPlan, plan);
   writeJson(paths.detectReport, detection);
   return {
     ...paths,
+    modelsDir,
+    whisperHostDir,
     packageCompose: packagedPath("compose", "preppergpt.yaml")
   };
 }

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "preppergpt",
-  "version": "0.1.0",
-  "description": "A local-first ChatGPT-like field kit built on OpenWebUI and local models.",
+  "version": "0.1.2",
+  "description": "A post-apocalyptic local AI field kit for running a ChatGPT-like experience when hosted services are unavailable.",
   "type": "module",
   "bin": {
     "preppergpt": "bin/preppergpt.js"
@@ -19,7 +19,7 @@
   ],
   "scripts": {
     "test": "node --test",
-    "check": "npm run test && node bin/preppergpt.js plan --profile balanced --json >/dev/null && node bin/preppergpt.js install --profile balanced --dry-run",
+    "check": "npm run test && node bin/preppergpt.js plan --profile balanced --json >/dev/null && node bin/preppergpt.js install --profile balanced --dry-run && node bin/preppergpt.js bundle whisper --dry-run >/dev/null",
     "pack:dry-run": "npm pack --dry-run"
   },
   "keywords": [
@@ -28,7 +28,9 @@
     "llm",
     "ollama",
     "preppergpt",
-    "offline-ai"
+    "offline-ai",
+    "survival",
+    "post-apocalyptic"
   ],
   "homepage": "https://github.com/teamslop/preppergpt#readme",
   "bugs": {

package/profiles/models.json

@@ -5,15 +5,15 @@
       "label": "Max intelligence",
       "defaultModel": "glm52-q4-local",
       "roles": {
-        "chat": ["glm52-q4-local", "qwen3.6-35b-a3b:slopcode-cpu-64k", "qwen2.5-coder:14b", "llama3.1:8b"],
-        "reasoning": ["glm52-q4-local", "qwen3.6-35b-a3b:slopcode-cpu-64k", "qwen2.5-coder:14b"],
+        "chat": ["glm52-q8-local", "glm52-q4-local", "qwen3.6-35b-a3b:slopcode-cpu-64k", "qwen2.5-coder:14b", "llama3.1:8b"],
+        "reasoning": ["glm52-q8-local", "glm52-q4-local", "qwen3.6-35b-a3b:slopcode-cpu-64k", "qwen2.5-coder:14b"],
         "fast": ["gemma4:12b-256k-gpu", "llama3.1:8b"],
         "coding": ["qwen3.6-35b-a3b:slopcode-cpu-64k", "qwen2.5-coder:14b"],
         "research": ["deep-research-glm52"],
         "agent": ["local-agent-glm52"],
         "vision": ["local-vision-gemma4-12b", "local-vision-moondream2"],
         "image": ["flux-2-klein-9b-fp8"],
-        "stt": ["whisper-large-v3"]
+        "stt": ["whisper-base-bundled"]
       }
     },
     "balanced": {
@@ -21,14 +21,14 @@
       "defaultModel": "local-chatgpt-auto",
       "roles": {
         "chat": ["local-chatgpt-auto", "gemma4:12b-256k-gpu", "glm52-q4-local", "llama3.1:8b"],
-        "reasoning": ["glm52-q4-local", "qwen3.6-35b-a3b:slopcode-cpu-64k", "qwen2.5-coder:14b"],
+        "reasoning": ["glm52-q8-local", "glm52-q4-local", "qwen3.6-35b-a3b:slopcode-cpu-64k", "qwen2.5-coder:14b"],
         "fast": ["gemma4:12b-256k-gpu", "llama3.1:8b"],
         "coding": ["qwen3.6-35b-a3b:slopcode-cpu-64k", "qwen2.5-coder:14b"],
         "research": ["deep-research-glm52"],
         "agent": ["local-agent-glm52"],
         "vision": ["local-vision-gemma4-12b", "local-vision-moondream2"],
         "image": ["flux-2-klein-9b-fp8"],
-        "stt": ["whisper-large-v3"]
+        "stt": ["whisper-base-bundled"]
       }
     },
     "speed": {
@@ -43,7 +43,7 @@
         "agent": ["local-agent-glm52"],
         "vision": ["local-vision-moondream2", "local-vision-gemma4-12b"],
         "image": ["flux-2-klein-9b-fp8"],
-        "stt": ["whisper-large-v3"]
+        "stt": ["whisper-base-bundled"]
       }
     }
   },
@@ -66,6 +66,26 @@
         "description": "Virtual OpenAI-compatible route exposed by the local-agent sidecar."
       }
     },
+    {
+      "id": "glm52-q8-local",
+      "name": "GLM 5.2 Q8 Local",
+      "roles": ["chat", "reasoning"],
+      "backend": "llama.cpp",
+      "contextTokens": 65536,
+      "qualityScore": 104,
+      "speedScore": 15,
+      "tpsEstimate": "very low tokens/sec may be acceptable in disaster/off-grid use because no hosted model is available; benchmark locally",
+      "requires": {
+        "minRamGb": 192,
+        "diskGb": 1000,
+        "nvme": true
+      },
+      "source": {
+        "type": "external",
+        "requiresHardwareFit": true,
+        "description": "Run a GLM 5.2 Q8 OpenAI-compatible llama.cpp server at http://127.0.0.1:11446/v1 with weights on fast NVMe for maximum local quality when hosted services are unavailable."
+      }
+    },
     {
       "id": "glm52-q4-local",
       "name": "GLM 5.2 Q4 Local",
@@ -74,7 +94,7 @@
       "contextTokens": 65536,
       "qualityScore": 100,
       "speedScore": 25,
-      "tpsEstimate": "0.4-3 completion tokens/sec on large CPU/NVMe builds; benchmark locally",
+      "tpsEstimate": "0.4-3 completion tokens/sec on large CPU/NVMe builds; acceptable for disaster/off-grid use when no hosted service is available; benchmark locally",
       "requires": {
         "minRamGb": 96,
         "diskGb": 520,
@@ -256,21 +276,23 @@
       }
     },
     {
-      "id": "whisper-large-v3",
-      "name": "Whisper Large v3 STT",
+      "id": "whisper-base-bundled",
+      "name": "Bundled Whisper Base STT",
       "roles": ["stt"],
       "backend": "openwebui-faster-whisper",
       "contextTokens": 0,
-      "qualityScore": 82,
-      "speedScore": 65,
-      "tpsEstimate": "audio transcription speed depends on CPU/GPU and clip length",
+      "qualityScore": 72,
+      "speedScore": 82,
+      "tpsEstimate": "local speech-to-text speed depends on CPU/GPU and clip length",
       "requires": {
-        "minRamGb": 16,
-        "diskGb": 4
+        "minRamGb": 8,
+        "diskGb": 1
       },
       "source": {
-        "type": "openwebui",
-        "description": "OpenWebUI faster-whisper local STT model."
+        "type": "bundled-download",
+        "model": "Systran/faster-whisper-base",
+        "license": "MIT",
+        "description": "Installer-cached faster-whisper base model mounted into OpenWebUI for offline local STT after install."
       }
     }
   ]

package/services/local-scheduler/app.py

@@ -3780,11 +3780,11 @@ def local_parity_recommended_model(feature_family: str, primary_models: list[str
         return models[0] if models else None
 
     if "codex" in family or "software" in family or "code" in family:
-        return first_available(["slopcode-qwen-coder-local", "local-agent-glm52", "glm52-q4-local"])
+        return first_available(["slopcode-qwen-coder-local", "local-agent-glm52", "glm52-q8-local", "glm52-q4-local"])
     if "deep research" in family:
-        return first_available(["deep-research-glm52", "glm52-q4-local"])
+        return first_available(["deep-research-glm52", "glm52-q8-local", "glm52-q4-local"])
     if "developer mode" in family or "mcp" in family:
-        return first_available(["local-agent-glm52", "local-chatgpt-auto", "glm52-q4-local"])
+        return first_available(["local-agent-glm52", "local-chatgpt-auto", "glm52-q8-local", "glm52-q4-local"])
     if "image generation" in family:
         return first_available(["flux-2-klein-9b-fp8"])
     if "image editing" in family:
@@ -3792,18 +3792,18 @@ def local_parity_recommended_model(feature_family: str, primary_models: list[str
     if "image understanding" in family:
         return first_available(["local-vision-gemma4-12b", "local-vision-moondream2"])
     if "voice" in family or "record mode" in family:
-        return first_available(["whisper-large-v3", "local-agent-glm52"])
+        return first_available(["whisper-base-bundled", "whisper-large-v3", "local-agent-glm52"])
     if "shopping" in family:
-        return first_available(["glm52-shopping-research-local", "glm52-q4-local"])
+        return first_available(["glm52-shopping-research-local", "glm52-q8-local", "glm52-q4-local"])
     if "job search" in family or "resume" in family or "finance" in family:
-        return first_available(["local-agent-glm52", "local-chatgpt-auto", "glm52-q4-local"])
+        return first_available(["local-agent-glm52", "local-chatgpt-auto", "glm52-q8-local", "glm52-q4-local"])
     if "study" in family:
-        return first_available(["glm52-study-coach-local", "glm52-q4-local"])
+        return first_available(["glm52-study-coach-local", "glm52-q8-local", "glm52-q4-local"])
     if "advanced reasoning" in family or "long context" in family:
-        return first_available(["glm52-q4-local"])
+        return first_available(["glm52-q8-local", "glm52-q4-local"])
     if "data analysis" in family or "canvas" in family or "memory" in family or "agent mode" in family:
-        return first_available(["local-agent-glm52", "glm52-q4-local"])
-    return first_available(["local-chatgpt-auto", "local-auto-router", "local-instant-gemma4-12b", "glm52-q4-local"])
+        return first_available(["local-agent-glm52", "glm52-q8-local", "glm52-q4-local"])
+    return first_available(["local-chatgpt-auto", "local-auto-router", "local-instant-gemma4-12b", "glm52-q8-local", "glm52-q4-local"])
 
 
 def local_parity_route_for_model(feature_family: str, model: str | None, profiles: dict) -> dict:
@@ -3833,10 +3833,10 @@ def local_parity_route_for_model(feature_family: str, model: str | None, profile
         route_id = "slopcode_tiny"
         route_type = "benchmarked_chat_route"
         action = "Select the Slopcode/Qwen coding model in OpenWebUI for local software work."
-    elif model_text == "glm52-q4-local" or "advanced reasoning" in family or "long context" in family:
+    elif model_text in {"glm52-q8-local", "glm52-q4-local"} or "advanced reasoning" in family or "long context" in family:
         route_id = "glm_tiny"
         route_type = "benchmarked_chat_route"
-        action = "Select GLM 5.2 Q4 in OpenWebUI for private long-context reasoning."
+        action = "Select the best available local GLM 5.2 route in OpenWebUI for private long-context reasoning."
     elif "shopping" in model_text:
         route_id = "glm52_shopping_research_preset"
         route_type = "chat_preset"
@@ -4462,9 +4462,9 @@ WORKFLOW_RECIPE_BLUEPRINTS = [
         "id": "private-long-context-workflow",
         "task_id": "private-long-context-reasoning",
         "title": "Private long-context reasoning with GLM 5.2",
-        "openwebui_entrypoint": "Model picker -> glm52-q4-local",
+        "openwebui_entrypoint": "Model picker -> glm52-q8-local on enterprise rigs, otherwise glm52-q4-local",
         "steps": [
-            "Select glm52-q4-local when privacy and context length matter more than latency.",
+            "Select glm52-q8-local on enterprise hardware when maximum local quality matters; otherwise select glm52-q4-local.",
             "Keep the prompt bounded when possible; use files/projects for reusable context.",
             "Use fast local routes for quick follow-ups when GLM latency is not needed.",
         ],
@@ -5293,6 +5293,7 @@ def local_parity_dashboard() -> dict:
         },
         "urls": {
             "openwebui": "http://127.0.0.1:8080",
+            "glm52_q8_openai": "http://127.0.0.1:11446/v1",
             "glm52_openai": "http://127.0.0.1:11441/v1",
             "slopcode_openai": "http://127.0.0.1:11438/v1",
             "deep_research_openai": "http://127.0.0.1:18041/v1",
@@ -5306,6 +5307,7 @@ def local_parity_dashboard() -> dict:
         },
         "primary_models": [
             {"id": "local-chatgpt-auto", "route": "fast_router", "best_for": "default local ChatGPT-like routing"},
+            {"id": "glm52-q8-local", "route": "glm_tiny", "context_tokens": 65536, "best_for": "enterprise 8-bit private long-context reasoning"},
             {"id": "glm52-q4-local", "route": "glm_tiny", "context_tokens": 65536, "best_for": "private long-context reasoning"},
             {
                 "id": "qwen3.6-35b-a3b:slopcode-cpu-64k",
@@ -5317,7 +5319,7 @@ def local_parity_dashboard() -> dict:
             {"id": "local-agent-glm52", "route": "local_agent", "best_for": "tool and agent workflows"},
             {"id": "local-vision-gemma4-12b", "route": "local_vision", "best_for": "image understanding"},
             {"id": "flux-2-klein-9b-fp8", "route": "comfyui_flux", "best_for": "image generation"},
-            {"id": "whisper-large-v3", "route": "local_whisper_stt", "best_for": "speech-to-text"},
+            {"id": "whisper-base-bundled", "route": "local_whisper_stt", "best_for": "speech-to-text"},
         ],
         "route_profiles": {
             key: {
@@ -6660,7 +6662,7 @@ def local_parity_audit_html() -> str:
       {metric("Starter prompts", f"{starter_summary.get('ready_starter_prompts')}/{starter_summary.get('starter_prompts')}", "prompt-library items")}
       {metric("Current release", f"{source_freshness_summary.get('current_release_covered_families')}/{source_freshness_summary.get('current_release_expected_families')}", "families")}
       {metric("Release evidence", f"{source_freshness_summary.get('current_release_covered_evidence_terms')}/{source_freshness_summary.get('current_release_expected_evidence_terms')}", "terms")}
-      {metric("Primary GLM route", "glm52-q4-local", "local long-context model")}
+      {metric("Primary GLM route", "glm52-q8-local / glm52-q4-local", "local long-context model")}
       {metric("Scope exclusions", f"{frontier_summary.get('excluded_from_local_goal_items')}/{frontier_summary.get('boundary_items')}", "hosted capabilities")}
       {metric("Evidence artifacts", f"{evidence_summary.get('ready_artifacts')}/{evidence_summary.get('artifacts')}", "privacy-safe proof")}
       {metric("Quality evals", scorecard_summary.get('quality_evals'), "executable")}
@@ -9653,7 +9655,7 @@ def local_parity_gap_report_html() -> str:
       {metric("Quality evals", summary.get('quality_evals'), "executable catalog")}
       {metric("Continuity", summary.get('continuity_status'), "fallback status")}
       {metric("Sources", summary.get('source_entries'), "source snapshot")}
-      {metric("Primary GLM route", "glm52-q4-local", "local long-context model")}
+      {metric("Primary GLM route", "glm52-q8-local / glm52-q4-local", "local long-context model")}
       {metric("GLM context", "65,536", "tokens")}
       {metric("Scope exclusions", f"{frontier_summary.get('excluded_from_local_goal_items')}/{frontier_summary.get('boundary_items')}", "hosted capabilities")}
     </section>
@@ -12797,7 +12799,7 @@ def local_model_route_recommendations() -> dict:
         "glm_tiny": {
             "title": "Private GLM 5.2 reasoning route",
             "benchmark_suite": "glm_tiny",
-            "default_model": "glm52-q4-local",
+            "default_model": "glm52-q8-local or glm52-q4-local",
             "target_tps": 0.1,
             "best_for": [
                 "private long-context reasoning",