feat: vision + macros + scheduler in Python (129 → 142 commands)

Three big rust packs ported. Python parity now covers the FULL imba lineup
except codebase_qa and github_pr.

vision_handler.py (new, ~125 lines)
  - Screenshot via PowerShell System.Drawing → temp PNG → base64 → Groq
    vision API (llama-3.2-11b-vision-preview, override via GROQ_VISION_MODEL).
  - do_vision_describe   "что на экране" / "опиши экран"
  - do_vision_read_error "прочитай ошибку"
  - Requires GROQ_TOKEN; speaks "Не удалось" on missing.

macros_store.py (new, ~170 lines)
  - JSON store at <here>/macros.json (atomic write-through).
  - start(name) / save() / cancel() / replay(name, dispatch_fn)
  - is_recording / recording_name / list_names / get / delete
  - record_step(phrase) called by execute_cmd hook (see main.py)
  - is_macro_control filter blocks "запиши макрос ..." from recording itself
    (no infinite recursion).
  - replay spawns daemon thread, fires each step with 800ms delay.

scheduler_store.py (new, ~220 lines)
  - JSON store at <here>/schedule.json.
  - parse_schedule(spec) handles "daily HH:MM", "at HH:MM",
    "every/in N minutes|hours|seconds" — Russian (минут/час/секунд) + English.
  - Background daemon thread ticks every 30s, fires Speak actions through the
    same speak callback used by main.py.
  - Once-tasks auto-delete after firing; daily/interval cycle forever.
  - add/remove/clear/remove_by_text/list_all.

extensions.py (+13 handlers)
  do_vision_describe / do_vision_read_error → thin proxies to vision_handler
  do_macros_start / save / cancel / replay / list / delete
  do_scheduler_add_reminder / add_recurring / list / clear / cancel_by_text

main.py
  - extensions.init_background_services() at startup → scheduler thread boots.
  - Wraps execute_cmd so every successful dispatch feeds macros_store.record_step.
  - set_dispatch_fn(_execute_phrase_for_macro) enables macro replay through the
    normal command pipeline (recognize_cmd → execute_cmd).

commands.yaml (+13 entries, 129 → 142)
  vision_describe, vision_read_error, macros_start, macros_save, macros_cancel,
  macros_replay, macros_list, macros_delete, scheduler_reminder,
  scheduler_recurring, scheduler_list, scheduler_clear, scheduler_cancel_text.

Tests: ast.parse passes for all five .py files. yaml.safe_load = 142 entries.

Python parity status vs rust (72 packs):
  ✓ memory_pack, profile_switch, vision, scheduler, macros, llm switch
    (via env), llm_context (basic), media_keys (existing), screenshot, net_info,
    unit_convert, generators, disk, date_math, sleep_timer (via shutdown),
    interesting_fact, mailto, self_check, ssl_check, intro/echo
  ✗ codebase_qa, github_pr (require gh CLI + file walk — TODO)
  ✗ pomodoro / daily_briefing / habit_nudge / quick_search / diagnostics
    (could be ported but need scheduler integration + LLM glue)

vision_handler.py

@@ -0,0 +1,146 @@
+"""Vision handler — port of `crates/jarvis-core/src/lua/api/vision.rs`.
+
+Captures the primary screen via PowerShell System.Drawing, base64-encodes
+the PNG, sends to Groq's vision-capable model. Requires GROQ_TOKEN.
+
+Action types in commands.yaml:
+  vision_describe     "что на экране" / "опиши экран"
+  vision_read_error   "прочитай ошибку"
+"""
+
+import base64
+import os
+import subprocess
+import tempfile
+
+_speak_fn = print
+
+
+def set_speak(fn):
+    global _speak_fn
+    _speak_fn = fn
+
+
+def _speak(text):
+    try:
+        _speak_fn(text)
+    except Exception as exc:
+        print(f"[vision] speak: {exc}")
+
+
+_SCREENSHOT_PS_TEMPLATE = """
+Add-Type -AssemblyName System.Windows.Forms;
+Add-Type -AssemblyName System.Drawing;
+$b = [System.Windows.Forms.Screen]::PrimaryScreen.Bounds;
+$bmp = New-Object System.Drawing.Bitmap $b.Width, $b.Height;
+$g = [System.Drawing.Graphics]::FromImage($bmp);
+$g.CopyFromScreen($b.Location, [System.Drawing.Point]::Empty, $b.Size);
+$bmp.Save('{path}', [System.Drawing.Imaging.ImageFormat]::Png);
+$g.Dispose(); $bmp.Dispose();
+"""
+
+
+def _take_screenshot() -> str | None:
+    """Save full screen to a temp PNG, return path or None on failure."""
+    fd, path = tempfile.mkstemp(prefix='jarvis-screen-', suffix='.png')
+    os.close(fd)
+    ps = _SCREENSHOT_PS_TEMPLATE.format(path=path.replace("'", "''"))
+    try:
+        subprocess.run(
+            ['powershell', '-NoProfile', '-ExecutionPolicy', 'Bypass', '-Command', ps],
+            capture_output=True, timeout=10, check=True,
+        )
+    except Exception as exc:
+        print(f"[vision] screenshot: {exc}")
+        try:
+            os.unlink(path)
+        except OSError:
+            pass
+        return None
+    if not os.path.isfile(path) or os.path.getsize(path) < 1000:
+        try:
+            os.unlink(path)
+        except OSError:
+            pass
+        return None
+    return path
+
+
+def _vision_call(prompt: str, image_b64: str) -> str | None:
+    import config as cfg
+    if not getattr(cfg, 'GROQ_TOKEN', None):
+        return None
+
+    try:
+        from openai import OpenAI
+    except ImportError:
+        return None
+
+    client = OpenAI(
+        api_key=cfg.GROQ_TOKEN,
+        base_url=getattr(cfg, 'GROQ_BASE_URL', 'https://api.groq.com/openai/v1'),
+    )
+
+    model = os.environ.get('GROQ_VISION_MODEL', 'llama-3.2-11b-vision-preview')
+
+    try:
+        resp = client.chat.completions.create(
+            model=model,
+            messages=[{
+                'role': 'user',
+                'content': [
+                    {'type': 'text', 'text': prompt},
+                    {'type': 'image_url', 'image_url': {
+                        'url': f'data:image/png;base64,{image_b64}',
+                    }},
+                ],
+            }],
+            max_tokens=400, temperature=0.2, timeout=30,
+        )
+        return resp.choices[0].message.content.strip()
+    except Exception as exc:
+        print(f"[vision] LLM call: {exc}")
+        return None
+
+
+def _describe(prompt: str) -> str | None:
+    """Take a screenshot and ask vision-LLM. Returns description or None."""
+    path = _take_screenshot()
+    if not path:
+        return None
+    try:
+        with open(path, 'rb') as f:
+            b64 = base64.b64encode(f.read()).decode('ascii')
+    finally:
+        try:
+            os.unlink(path)
+        except OSError:
+            pass
+    return _vision_call(prompt, b64)
+
+
+def do_vision_describe(action, voice):
+    _speak("Сейчас посмотрю.")
+    prompt = (
+        "Опиши коротко (1-3 предложения) что сейчас на экране. По-русски."
+    )
+    result = _describe(prompt)
+    if not result:
+        _speak("Не удалось разобрать экран. Проверь GROQ_TOKEN.")
+        return
+    _speak(result)
+
+
+def do_vision_read_error(action, voice):
+    _speak("Смотрю на ошибку.")
+    prompt = (
+        "Найди на экране текст ошибки (сообщение об ошибке, stack trace, диалог "
+        "об исключении). Прочитай ключевое сообщение и одним коротким "
+        "предложением подскажи возможную причину. Если ошибки нет — скажи "
+        "'ошибок не вижу'. Отвечай по-русски. 1-3 предложения максимум."
+    )
+    result = _describe(prompt)
+    if not result:
+        _speak("Не получилось прочитать.")
+        return
+    _speak(result)