feat: local multimodal agent platform — Gemma 4 E4B + Flash loop

README.md

@@ -301,7 +301,7 @@ ARTEMIS supports two execution profiles tailored for different automation requir
 
 - [ ] **Android Studio Integration**: Native IDE plugin and workflow integration to enable in-editor debugging, test recording, and automated device control directly within Android Studio.
 - [ ] **iOS Platform Expansion**: Extending multimodal perception and mobile automation to iOS devices and simulators.
-- [ ] **On-Device Lightweight VLMs**: Local execution with lightweight edge vision models for low-latency, privacy-first automation.
+- [x] **On-Device Lightweight VLMs**: Local execution with lightweight edge vision models for low-latency, privacy-first automation via OpenAI-compatible local providers — see the [local models guide](./docs/local_models.md) for install/pull/run steps (`config/artemis.gemma4.jsonc` is the tested Gemma 4 E4B setup).
 - [ ] **Real-time Duplex Voice Interaction**: Voice-driven task dispatch with real-time conversational control and interruption handling.
 
 ## Community & Contributing

apps/admin_console/routers/system.py

@@ -16,6 +16,7 @@
 
 import ipaddress
 import os
+import re
 import secrets
 from urllib.parse import urlsplit
 
@@ -624,3 +625,188 @@ async def _shutdown_after_response() -> None:
         "message": "Artemis server is shutting down.",
         "pid": os.getpid(),
     }
+
+
+# --- Local model lifecycle (artemis model pull/serve/stop) ------------------
+#
+# The console surfaces the local_model_endpoint probe's guided actions; these
+# endpoints let the user run them without leaving the browser. Each action
+# spawns `python -m artemis model ...` and streams its output into a capped
+# in-memory log the frontend polls.
+
+
+class LocalModelActionRequest(BaseModel):
+    """Payload for a local-model lifecycle action."""
+
+    alias: str = Field(..., description="Catalog alias or org/repo id (e.g. gemma4-e4b)")
+    port: int | None = Field(
+        default=None,
+        ge=1,
+        le=65535,
+        description="Serve/stop port; the CLI default (8080) applies when omitted",
+    )
+
+
+_MODEL_TASKS: dict[str, dict] = {}
+_MODEL_TASK_LOG_LINES = 80
+
+
+def _model_cli_available() -> bool:
+    """Whether the ``artemis model`` command group exists in this build."""
+    import importlib.util
+
+    try:
+        return importlib.util.find_spec("artemis.interfaces.cli.commands.model") is not None
+    except Exception:
+        return False
+
+
+def _task_view(key: str, task: dict) -> dict:
+    return {
+        "key": key,
+        "action": task["action"],
+        "alias": task["alias"],
+        "cmd": task["cmd"],
+        "status": task["status"],
+        "returncode": task["returncode"],
+        "started_at": task["started_at"],
+        "log": task["log"][-12:],
+    }
+
+
+def _task_snapshot() -> dict:
+    return {k: _task_view(k, t) for k, t in _MODEL_TASKS.items()}
+
+
+async def _drain_model_task(task: dict) -> None:
+    proc = task["process"]
+    try:
+        if proc.stdout is not None:
+            async for raw in proc.stdout:
+                line = raw.decode(errors="replace").rstrip()
+                if line:
+                    task["log"].append(line)
+                    del task["log"][:-_MODEL_TASK_LOG_LINES]
+        rc = await proc.wait()
+    except Exception as e:  # pylint: disable=broad-exception-caught
+        rc = -1
+        task["log"].append(f"task monitor error: {e}")
+    task["returncode"] = rc
+    task["status"] = "done" if rc == 0 else "failed"
+
+
+async def _start_model_task(action: str, alias: str, extra_args: list[str]) -> dict:
+    """Spawn `artemis model <action>` as a subprocess with a capped log."""
+    import asyncio
+    import sys
+    import time
+
+    key = f"{action}:{alias}"
+    existing = _MODEL_TASKS.get(key)
+    if existing and existing["status"] == "running":
+        return _task_view(key, existing)
+
+    # ``stop`` is port-scoped on the CLI (no positional model arg).
+    cmd = [sys.executable, "-m", "artemis", "model", action]
+    if action != "stop":
+        cmd.append(alias)
+    cmd += extra_args
+    try:
+        proc = await asyncio.create_subprocess_exec(
+            *cmd,
+            stdout=asyncio.subprocess.PIPE,
+            stderr=asyncio.subprocess.STDOUT,
+        )
+    except Exception as e:
+        raise HTTPException(
+            status_code=500, detail=f"Failed to start `artemis model {action}`: {e}"
+        ) from e
+
+    task = {
+        "action": action,
+        "alias": alias,
+        "cmd": " ".join(cmd),
+        "status": "running",
+        "returncode": None,
+        "started_at": time.time(),
+        "log": [],
+        "process": proc,
+    }
+    _MODEL_TASKS[key] = task
+    task["watcher"] = asyncio.create_task(_drain_model_task(task))
+    return _task_view(key, task)
+
+
+def _require_model_cli() -> None:
+    if not _model_cli_available():
+        raise HTTPException(
+            status_code=501,
+            detail=(
+                "`artemis model` commands are not available in this build. "
+                "Run the shown command in a terminal, or manage the model with "
+                "your own OpenAI-compatible server."
+            ),
+        )
+
+
+_MODEL_REF_RE = re.compile(r"^[A-Za-z0-9][A-Za-z0-9_.-]*(/[A-Za-z0-9][A-Za-z0-9_.-]*)?$")
+
+
+def _require_model_ref(alias: str) -> None:
+    """Reject aliases the managed CLI could never resolve.
+
+    ``resolve_model_ref`` accepts catalog aliases and ``org/repo`` ids — the
+    same inputs ``artemis model pull`` takes. Anything else (bare junk, flag
+    injection like ``--port``, or argv that doesn't look like a model ref) is
+    refused before a subprocess is spawned.
+    """
+    from artemis.config.local_models import resolve_model_ref
+
+    if not _MODEL_REF_RE.match(alias) or resolve_model_ref(alias) is None:
+        raise HTTPException(
+            status_code=400,
+            detail=(
+                f"Unknown model '{alias}'. `artemis model` accepts catalog "
+                "aliases (see `artemis model list`) or a Hugging Face "
+                "'org/repo' id."
+            ),
+        )
+
+
+@router.get("/local-models/status")
+async def local_model_status():
+    """Command availability plus in-flight/finished model task states."""
+    return {
+        "commands_available": _model_cli_available(),
+        "tasks": _task_snapshot(),
+    }
+
+
+@router.post("/local-models/pull")
+async def pull_local_model(payload: LocalModelActionRequest, request: Request):
+    """Download model weights via `artemis model pull` (long-running)."""
+    _require_loopback_request(request, "Local model lifecycle is local-only.")
+    _require_model_cli()
+    alias = payload.alias.strip()
+    _require_model_ref(alias)
+    return await _start_model_task("pull", alias, [])
+
+
+@router.post("/local-models/serve")
+async def serve_local_model(payload: LocalModelActionRequest, request: Request):
+    """Start the managed model server via `artemis model serve`."""
+    _require_loopback_request(request, "Local model lifecycle is local-only.")
+    _require_model_cli()
+    alias = payload.alias.strip()
+    _require_model_ref(alias)
+    extra = ["--port", str(payload.port)] if payload.port else []
+    return await _start_model_task("serve", alias, extra)
+
+
+@router.post("/local-models/stop")
+async def stop_local_model(payload: LocalModelActionRequest, request: Request):
+    """Stop a managed model server via `artemis model stop`."""
+    _require_loopback_request(request, "Local model lifecycle is local-only.")
+    _require_model_cli()
+    extra = ["--port", str(payload.port)] if payload.port else []
+    return await _start_model_task("stop", payload.alias.strip(), extra)

apps/showcase_ui/src/app/pages/home/home.component.html

@@ -349,7 +349,7 @@ <h3 class="step-title">AI Model Setup</h3>
               <span class="choice-title">Custom Configuration</span>
             </div>
             <p class="choice-desc">Define models in <code>config/artemis.jsonc</code> & keys in <code>.env</code>
-              (OpenAI, Claude, Ollama).</p>
+              (OpenAI, Claude, Ollama — or a local model via <code>artemis model serve</code>).</p>
           </div>
         </div>
       </div>
@@ -604,6 +604,87 @@ <h3 class="step-title">AI Model Setup</h3>
           }
           </div>
 
+      <!-- Section 1b: Local Model Endpoint (shown when a custom api_base is
+           configured, or the readiness probe is failing) -->
+      @if (localModelProbe(); as lm) {
+        @if (lm.status !== 'pass' || modelConfigEnv()?.default_model?.provider === 'custom') {
+          <div class="inspector-card" [class.card-warn]="lm.status !== 'pass'">
+            <div class="inspector-card-header">
+              <div class="inspector-header-left">
+                <span class="material-symbols-outlined header-icon">dns</span>
+                <div class="inspector-header-titles">
+                  <span class="inspector-title">Local Model Endpoint</span>
+                  <span class="inspector-subtitle">{{ lm.summary }}</span>
+                </div>
+              </div>
+              <div class="inspector-header-actions">
+                @if (lm.status === 'pass') {
+                <span class="badge-env-pass">
+                  <span class="material-symbols-outlined">check_circle</span>
+                  <span>Ready</span>
+                </span>
+                } @else {
+                <span class="badge-env-unset">
+                  <span class="material-symbols-outlined">error</span>
+                  <span>Action Needed</span>
+                </span>
+                }
+              </div>
+            </div>
+
+            <p class="inspector-desc">{{ lm.description }}</p>
+
+            <!-- Guided actions: Artemis commands are clickable when the CLI
+                 exists in this build (fallback: copyable code box); hints text -->
+            @for (act of lm.actions; track act.label) {
+              @if (act.action_type === 'command') {
+                <div class="code-box small-box">
+                  <code>{{ act.payload }}</code>
+                  @if (modelActionArgs(act.payload); as args) {
+                    @if (localModelCommandsAvailable()) {
+                      @if (localModelTaskFor(args.action, args.alias)?.status === 'running') {
+                        <span class="material-symbols-outlined spinning lm-spin">progress_activity</span>
+                        <span class="lm-running-label">Running…</span>
+                      } @else {
+                        <button type="button" class="btn-run-cmd"
+                          (click)="runLocalModelAction(args)">
+                          <span class="material-symbols-outlined">play_arrow</span>
+                          <span>Run</span>
+                        </button>
+                      }
+                    }
+                  }
+                  <button type="button" class="btn-copy-code" [class.copied]="copiedId() === 'lm_' + act.label"
+                    (click)="copyToClipboard(act.payload, 'lm_' + act.label)">
+                    <span class="material-symbols-outlined">{{ copiedId() === 'lm_' + act.label ? 'check' : 'content_copy' }}</span>
+                    <span>{{ copiedId() === 'lm_' + act.label ? 'Copied!' : 'Copy' }}</span>
+                  </button>
+                </div>
+                @if (modelActionArgs(act.payload); as args) {
+                  @if (localModelTaskFor(args.action, args.alias); as task) {
+                    @if (task.status === 'failed' || task.log.length) {
+                      <div class="lm-task-log" [class.lm-task-failed]="task.status === 'failed'">
+                        @if (task.status === 'done') {
+                          <div class="lm-log-line lm-log-done">✓ {{ act.label }} finished</div>
+                        }
+                        @if (task.status === 'failed') {
+                          <div class="lm-log-line lm-log-failed">✗ {{ act.label }} failed (exit {{ task.returncode }}) — check the log below or run the command in a terminal</div>
+                        }
+                        @for (line of task.log.slice(-4); track $index) {
+                          <div class="lm-log-line">{{ line }}</div>
+                        }
+                      </div>
+                    }
+                  }
+                }
+              } @else {
+              <p class="inspector-desc"><strong>{{ act.label }}:</strong> {{ act.payload }}</p>
+              }
+            }
+          </div>
+        }
+      }
+
       <!-- Section 2: Environment API Keys (.env) -->
       <div class="inspector-card">
         <div class="inspector-card-header">