add cloud compiler and reset video cooldowns

frontend/src/lib/api.ts

@@ -1,6 +1,7 @@
 import { Client, handle_file } from "@gradio/client"
 import type {
   CompileResult,
+  CompilerProvider,
   DetectParams,
   LocalConfig,
   RunResult,
@@ -50,12 +51,14 @@ export async function detectAndAutomate(
 export async function compileRules(
   instruction: string,
   classes: string,
+  provider: CompilerProvider,
   baseUrl: string,
   model: string,
 ): Promise<CompileResult> {
   return call<CompileResult>("/compile_rules", {
     instruction,
     classes,
+    provider,
     base_url: baseUrl,
     model,
   })

frontend/src/lib/dashboard.tsx

@@ -12,6 +12,7 @@ import { toast } from "sonner"
 import * as api from "./api"
 import type {
   CompileResult,
+  CompilerProvider,
   DetectParams,
   RunResult,
   ValidationResult,
@@ -61,6 +62,8 @@ interface DashboardState {
   validation: ValidationResult | null
   validate: () => Promise<void>
   save: () => Promise<void>
+  compilerProvider: CompilerProvider
+  setCompilerProvider: (provider: CompilerProvider) => void
   compile: (instruction: string) => Promise<CompileResult | null>
   compiling: boolean
 }
@@ -76,6 +79,7 @@ export function DashboardProvider({ children }: { children: ReactNode }) {
   const [running, setRunning] = useState(false)
   const [error, setError] = useState<string | null>(null)
   const [validation, setValidation] = useState<ValidationResult | null>(null)
+  const [compilerProvider, setCompilerProvider] = useState<CompilerProvider>("cloud")
   const [compiling, setCompiling] = useState(false)
   const previewRef = useRef<string | null>(null)
 
@@ -93,6 +97,9 @@ export function DashboardProvider({ children }: { children: ReactNode }) {
           maxDetections: cfg.default_max_detections ?? p.maxDetections,
           webhookUrl: cfg.webhook_url ?? p.webhookUrl,
         }))
+        if (cfg.llm_provider === "cloud" || cfg.llm_provider === "local") {
+          setCompilerProvider(cfg.llm_provider)
+        }
       })
       .catch(() => void 0)
     api
@@ -178,6 +185,7 @@ export function DashboardProvider({ children }: { children: ReactNode }) {
         const c = await api.compileRules(
           instruction,
           params.classes,
+          compilerProvider,
           "http://127.0.0.1:8080/v1",
           "ggml-org/Qwen3-1.7B-GGUF:Q4_K_M",
         )
@@ -193,7 +201,7 @@ export function DashboardProvider({ children }: { children: ReactNode }) {
         setCompiling(false)
       }
     },
-    [params.classes],
+    [compilerProvider, params.classes],
   )
 
   const value = useMemo<DashboardState>(
@@ -212,6 +220,8 @@ export function DashboardProvider({ children }: { children: ReactNode }) {
       validation,
       validate,
       save,
+      compilerProvider,
+      setCompilerProvider,
       compile,
       compiling,
     }),
@@ -229,6 +239,7 @@ export function DashboardProvider({ children }: { children: ReactNode }) {
       validation,
       validate,
       save,
+      compilerProvider,
       compile,
       compiling,
     ],

frontend/src/lib/types.ts

@@ -57,6 +57,8 @@ export interface CompileResult {
   rule_count: number
 }
 
+export type CompilerProvider = "local" | "cloud"
+
 export interface LocalConfig {
   camera_url: string | null
   webhook_url: string | null
@@ -65,8 +67,11 @@ export interface LocalConfig {
   default_device: string | null
   default_image_size: number | null
   default_max_detections: number | null
+  llm_provider: CompilerProvider | null
   llamacpp_base_url: string | null
   llamacpp_model: string | null
+  replicate_model: string | null
+  replicate_reasoning_effort: string | null
 }
 
 export interface DetectParams {

frontend/src/modules/rules/RuleStudioPanel.tsx

@@ -2,6 +2,8 @@ import { useState } from "react"
 import {
   CheckCircle2,
   CircleSlash,
+  Cloud,
+  Cpu,
   Loader2,
   Save,
   ShieldCheck,
@@ -13,6 +15,8 @@ import { Button } from "@/components/ui/button"
 import { Textarea } from "@/components/ui/textarea"
 import { Badge } from "@/components/ui/badge"
 import { Tabs, TabsContent, TabsList, TabsTrigger } from "@/components/ui/tabs"
+import { Label } from "@/components/ui/label"
+import { Switch } from "@/components/ui/switch"
 import { cn } from "@/lib/utils"
 
 export function RuleStudioPanel() {
@@ -22,6 +26,8 @@ export function RuleStudioPanel() {
     validate,
     validation,
     save,
+    compilerProvider,
+    setCompilerProvider,
     compile,
     compiling,
   } = useDashboard()
@@ -64,10 +70,27 @@ export function RuleStudioPanel() {
       {/* Compose: NL → rules via llama.cpp */}
       <TabsContent value="compose">
         <div className="space-y-3">
-          <p className="text-xs text-muted-foreground">
-            Describe the automation in plain language. It compiles to validated rules through
-            llama.cpp — never executable code.
-          </p>
+          <div className="flex flex-col gap-3 sm:flex-row sm:items-center sm:justify-between">
+            <p className="text-xs text-muted-foreground">
+              Describe the automation in plain language. It compiles to validated rules through{" "}
+              {compilerProvider === "cloud" ? "Replicate" : "llama.cpp"} — never executable code.
+            </p>
+            <div className="flex shrink-0 items-center gap-2 rounded-md border border-border bg-black/20 px-3 py-2">
+              {compilerProvider === "cloud" ? (
+                <Cloud className="size-3.5 text-primary" />
+              ) : (
+                <Cpu className="size-3.5 text-muted-foreground" />
+              )}
+              <Label htmlFor="compiler-provider" className="text-xs text-foreground">
+                Cloud
+              </Label>
+              <Switch
+                id="compiler-provider"
+                checked={compilerProvider === "cloud"}
+                onCheckedChange={(checked) => setCompilerProvider(checked ? "cloud" : "local")}
+              />
+            </div>
+          </div>
           <Textarea
             value={instruction}
             onChange={(e) => setInstruction(e.target.value)}
@@ -76,7 +99,11 @@ export function RuleStudioPanel() {
           />
           <Button onClick={() => compile(instruction)} disabled={compiling || !instruction.trim()}>
             {compiling ? <Loader2 className="size-4 animate-spin" /> : <Wand2 className="size-4" />}
-            {compiling ? "Compiling…" : "Compile to rules"}
+            {compiling
+              ? "Compiling…"
+              : compilerProvider === "cloud"
+                ? "Compile with cloud"
+                : "Compile to rules"}
           </Button>
         </div>
       </TabsContent>

frontend/src/modules/settings/SettingsPanel.tsx

@@ -64,6 +64,10 @@ export function SettingsPanel() {
           Start with{" "}
           <span className="font-mono">llama-server -hf ggml-org/Qwen3-1.7B-GGUF:Q4_K_M</span>
         </p>
+        <p className="text-[0.6875rem] text-muted-foreground">
+          Cloud compile uses <span className="font-mono">REPLICATE_API_TOKEN</span> or{" "}
+          <span className="font-mono">.local/config.yaml</span>.
+        </p>
       </div>
     </div>
   )

poetry.lock

@@ -3433,6 +3433,24 @@ Pygments = ">=2.5.1"
 [package.extras]
 md = ["comrak (>=0.0.11)"]
 
+[[package]]
+name = "replicate"
+version = "1.0.7"
+description = "Python client for Replicate"
+optional = false
+python-versions = ">=3.8"
+groups = ["main"]
+files = [
+    {file = "replicate-1.0.7-py3-none-any.whl", hash = "sha256:667c50a9eb83be17de6278ff89483102b3b50f49a2c7fbcaa2e2b14df13816f9"},
+    {file = "replicate-1.0.7.tar.gz", hash = "sha256:d88cb2c37ba39fb370c87fc3291601c67aae64bb918a20a85b5ce399c23ee84c"},
+]
+
+[package.dependencies]
+httpx = ">=0.21.0,<1"
+packaging = "*"
+pydantic = ">1.10.7"
+typing_extensions = ">=4.5.0"
+
 [[package]]
 name = "requests"
 version = "2.34.2"
@@ -4313,4 +4331,4 @@ type = ["pytest-mypy (>=1.0.1) ; platform_python_implementation != \"PyPy\""]
 [metadata]
 lock-version = "2.1"
 python-versions = ">=3.10,!=3.14.1,<3.15"
-content-hash = "b51b0f2df8f90aff535893342ff2ff5f9001daf56e4943be201f6d3e757c0c9a"
+content-hash = "b136468f5b1d7e59f642fdb59b4cd8e9a940b6a70a13feb0b54c049f08dd656d"

pyproject.toml

@@ -14,6 +14,7 @@ dependencies = [
     "pydantic>=2.8",
     "PyYAML>=6.0",
     "requests>=2.32",
+    "replicate>=1.0",
     "numpy>=1.26",
     "imageio-ffmpeg>=0.6.0",
     "static-ffmpeg>=2.13",

requirements.txt

@@ -4,4 +4,5 @@ opencv-python-headless>=4.10
 pydantic>=2.8
 PyYAML>=6.0
 requests>=2.32
+replicate>=1.0
 numpy>=1.26

server.py

@@ -13,6 +13,7 @@ does not require torch/ultralytics.
 
 from __future__ import annotations
 
+import os
 from pathlib import Path
 from typing import Any
 
@@ -21,6 +22,7 @@ from fastapi.responses import FileResponse, HTMLResponse, JSONResponse
 
 from tiny_trigger import (
     compile_automation_with_llamacpp,
+    compile_automation_with_replicate,
     evaluate_video_detections,
     load_automation_text,
     parse_class_prompt,
@@ -29,10 +31,8 @@ from tiny_trigger import (
 from tiny_trigger.actions import dispatch_events
 from tiny_trigger.store import (
     load_local_config,
-    load_runtime_state,
     load_saved_automations,
     save_automations,
-    save_runtime_state,
     append_events,
 )
 from tiny_trigger.video import render_automation_video
@@ -112,7 +112,6 @@ def detect_and_automate(
         raise ValueError("A video file is required.")
 
     rules = load_automation_text(rules_text)
-    state = load_runtime_state()
 
     result = process_video(
         video_path=video_path,
@@ -127,14 +126,16 @@ def detect_and_automate(
         output_dir=str(RENDERS),
     )
 
-    events, last_fired = evaluate_video_detections(
-        rules.rules, result.detections, last_fired=state.last_fired
+    events, _last_fired = evaluate_video_detections(
+        rules.rules,
+        result.detections,
+        # Uploaded videos use clip-relative timestamps, so cooldowns reset for
+        # each run. A live camera mode can persist wall-clock cooldowns later.
+        last_fired=None,
     )
     dispatched = dispatch_events(
         events, enable_webhooks=enable_webhooks, webhook_url=webhook_url or None
     )
-    state.last_fired = last_fired
-    save_runtime_state(state)
     append_events(dispatched)
 
     automation_path = render_automation_video(
@@ -169,17 +170,35 @@ def detect_and_automate(
 def compile_rules(
     instruction: str,
     classes: str = "",
+    provider: str = "local",
     base_url: str = "http://127.0.0.1:8080/v1",
     model: str = "ggml-org/Qwen3-1.7B-GGUF:Q4_K_M",
+    replicate_model: str = "openai/gpt-5.2",
+    replicate_reasoning_effort: str = "medium",
 ) -> dict:
     """Compile a natural-language request into validated automation rules."""
     class_names = parse_class_prompt(classes) if classes else []
-    compiled = compile_automation_with_llamacpp(
-        instruction=instruction,
-        class_names=class_names,
-        base_url=base_url,
-        model=model,
-    )
+    cfg = load_local_config()
+    if provider == "cloud":
+        api_token = os.environ.get("REPLICATE_API_TOKEN") or cfg.replicate_api_token
+        if not api_token:
+            raise ValueError("Set REPLICATE_API_TOKEN or replicate_api_token in .local/config.yaml.")
+        compiled = compile_automation_with_replicate(
+            instruction=instruction,
+            class_names=class_names,
+            api_token=api_token,
+            model=replicate_model or cfg.replicate_model or "openai/gpt-5.2",
+            reasoning_effort=(
+                replicate_reasoning_effort or cfg.replicate_reasoning_effort or "medium"
+            ),
+        )
+    else:
+        compiled = compile_automation_with_llamacpp(
+            instruction=instruction,
+            class_names=class_names,
+            base_url=base_url or cfg.llamacpp_base_url or "http://127.0.0.1:8080/v1",
+            model=model or cfg.llamacpp_model or "ggml-org/Qwen3-1.7B-GGUF:Q4_K_M",
+        )
     return {
         "rules_text": compiled.document.model_dump_json(by_alias=True, indent=2),
         "raw_text": compiled.raw_text,
@@ -227,7 +246,7 @@ def load_rules() -> dict:
 @app.api(name="get_config")
 def get_config() -> dict:
     cfg = load_local_config()
-    return cfg.model_dump()
+    return cfg.model_dump(exclude={"replicate_api_token"})
 
 
 # ── static frontend + media (custom routes take priority over gradio's) ──────

tests/test_automation.py

@@ -155,6 +155,31 @@ def test_evaluate_video_detections_returns_updated_last_fired() -> None:
     assert unchanged_last_fired == last_fired
 
 
+def test_video_cooldowns_are_fresh_without_persisted_last_fired() -> None:
+    document = load_automation_text(
+        json.dumps(
+            {
+                "rules": [
+                    {
+                        "name": "turn-on-pc",
+                        "when": {"all": [{"present": {"label": "person"}}]},
+                        "gate": {"cooldown": {"key": "turn-on-pc", "seconds": 60}},
+                        "then": [{"type": "simulate", "name": "turn on pc"}],
+                    }
+                ]
+            }
+        )
+    )
+    detections = [detection("person", (0.1, 0.1, 0.2, 0.2), frame=2, time=12.0)]
+
+    first_events, first_last_fired = evaluate_video_detections(document.rules, detections)
+    second_events, second_last_fired = evaluate_video_detections(document.rules, detections)
+
+    assert len(first_events) == 1
+    assert len(second_events) == 1
+    assert first_last_fired == second_last_fired == {"turn-on-pc": 12.0}
+
+
 def test_invalid_rule_rejected() -> None:
     with pytest.raises(ValidationError):
         load_automation_text(

tests/test_llm.py

@@ -8,6 +8,7 @@ from tiny_trigger.llm import (
     SYSTEM_PROMPT,
     _build_user_prompt,
     _chat_payload,
+    compile_automation_with_replicate,
     compile_automation_with_llamacpp,
     _post_chat_completion,
     _validate_compile_result,
@@ -29,6 +30,7 @@ def test_llamacpp_payload_uses_json_object() -> None:
     payload = _chat_payload(model="qwen", user_prompt="compile this", response_format="json_object")
 
     assert payload["response_format"] == {"type": "json_object"}
+    assert payload["max_tokens"] == 512
 
 
 def test_prompt_includes_validation_schema() -> None:
@@ -124,3 +126,31 @@ def test_compile_retries_after_invalid_response(monkeypatch) -> None:
 
     assert Requests.calls == 2
     assert result.document.rules[0].name == "notify"
+
+
+def test_replicate_compile_uses_stream_api(monkeypatch) -> None:
+    class Replicate:
+        class Client:
+            def __init__(self, api_token):
+                assert api_token == "test-token"
+
+            def stream(self, model, input):
+                assert model == "openai/gpt-5.2"
+                assert input["messages"] == []
+                assert input["verbosity"] == "medium"
+                assert input["reasoning_effort"] == "low"
+                assert "Return only the JSON object." in input["prompt"]
+                yield '{"rules":[{"name":"notify","when":{"all":[{"present":{"label":"person"}}]},'
+                yield '"then":[{"type":"simulate","name":"notify"}]}]}'
+
+    monkeypatch.setitem(sys.modules, "replicate", Replicate)
+
+    result = compile_automation_with_replicate(
+        instruction="if person present notify",
+        class_names=["person"],
+        api_token="test-token",
+        model="openai/gpt-5.2",
+        reasoning_effort="low",
+    )
+
+    assert result.document.rules[0].name == "notify"

tests/test_store.py

@@ -17,12 +17,26 @@ from tiny_trigger.store import (
 
 def test_local_config_yaml(tmp_path) -> None:
     path = tmp_path / "config.yaml"
-    path.write_text("default_device: cuda:0\nwebhook_url: http://example.test\n", encoding="utf-8")
+    path.write_text(
+        "\n".join(
+            [
+                "default_device: cuda:0",
+                "webhook_url: http://example.test",
+                "llm_provider: cloud",
+                "replicate_model: openai/gpt-5.2",
+                "replicate_reasoning_effort: medium",
+            ]
+        ),
+        encoding="utf-8",
+    )
 
     config = load_local_config(path)
 
     assert config.default_device == "cuda:0"
     assert config.webhook_url == "http://example.test"
+    assert config.llm_provider == "cloud"
+    assert config.replicate_model == "openai/gpt-5.2"
+    assert config.replicate_reasoning_effort == "medium"
 
 
 def test_save_and_load_automations(tmp_path) -> None:

tiny_trigger/__init__.py

@@ -2,13 +2,14 @@
 
 from .automation import RuleEngine, evaluate_video_detections, load_automation_text
 from .detector import UltralyticsYOLOEDetector, parse_class_prompt
-from .llm import compile_automation_with_llamacpp
+from .llm import compile_automation_with_llamacpp, compile_automation_with_replicate
 from .video import process_video
 
 __all__ = [
     "RuleEngine",
     "UltralyticsYOLOEDetector",
     "compile_automation_with_llamacpp",
+    "compile_automation_with_replicate",
     "evaluate_video_detections",
     "load_automation_text",
     "parse_class_prompt",

tiny_trigger/llm.py

@@ -73,6 +73,92 @@ def compile_automation_with_llamacpp(
             ) from second_error
 
 
+def compile_automation_with_replicate(
+    *,
+    instruction: str,
+    class_names: list[str],
+    api_token: str,
+    model: str = "openai/gpt-5.2",
+    reasoning_effort: str = "medium",
+    timeout: float = 600.0,
+) -> LLMCompileResult:
+    """Compile natural language into validated rules through Replicate."""
+    try:
+        import replicate
+    except ImportError as exc:  # pragma: no cover - dependency guard
+        raise RuntimeError("Install replicate to use the Replicate compiler.") from exc
+
+    user_prompt = _build_user_prompt(instruction=instruction, class_names=class_names)
+    raw_text = _stream_replicate_completion(
+        model=model,
+        prompt=_provider_prompt(user_prompt),
+        api_token=api_token,
+        reasoning_effort=reasoning_effort,
+        timeout=timeout,
+        replicate_module=replicate,
+    )
+    try:
+        return _validate_compile_result(raw_text)
+    except (json.JSONDecodeError, ValidationError, ValueError) as first_error:
+        repair_prompt = _build_repair_prompt(
+            original_prompt=user_prompt,
+            bad_response=raw_text,
+            error=str(first_error),
+        )
+        repaired_text = _stream_replicate_completion(
+            model=model,
+            prompt=_provider_prompt(repair_prompt),
+            api_token=api_token,
+            reasoning_effort=reasoning_effort,
+            timeout=timeout,
+            replicate_module=replicate,
+        )
+        try:
+            return _validate_compile_result(repaired_text)
+        except (json.JSONDecodeError, ValidationError, ValueError) as second_error:
+            raise ValueError(
+                "Replicate returned invalid automation JSON after a repair attempt. "
+                f"Last validation error: {second_error}. Last response: {repaired_text}"
+            ) from second_error
+
+
+def _provider_prompt(user_prompt: str) -> str:
+    return f"{SYSTEM_PROMPT}\n\n{user_prompt}\n\nReturn only the JSON object."
+
+
+def _stream_replicate_completion(
+    *,
+    model: str,
+    prompt: str,
+    api_token: str,
+    reasoning_effort: str,
+    timeout: float,
+    replicate_module: Any,
+) -> str:
+    _split_replicate_model(model)
+    payload = {
+        "prompt": prompt,
+        "messages": [],
+        "verbosity": "medium",
+        "reasoning_effort": reasoning_effort,
+    }
+    client = replicate_module.Client(api_token=api_token)
+    chunks: list[str] = []
+    for event in client.stream(model, input=payload):
+        chunks.append(str(event))
+    text = "".join(chunks).strip()
+    if not text:
+        raise ValueError("Replicate stream returned no output.")
+    return text
+
+
+def _split_replicate_model(model: str) -> tuple[str, str]:
+    parts = model.strip().split("/", 1)
+    if len(parts) != 2 or not all(parts):
+        raise ValueError("Replicate model must be in owner/model format, for example openai/gpt-5.2.")
+    return parts[0], parts[1]
+
+
 def _post_chat_completion(
     *,
     endpoint: str,
@@ -191,6 +277,7 @@ def _chat_payload(*, model: str, user_prompt: str, response_format: str) -> dict
             {"role": "system", "content": SYSTEM_PROMPT},
             {"role": "user", "content": user_prompt},
         ],
+        "max_tokens": 512,
         "temperature": 0,
         "stream": False,
     }

tiny_trigger/store.py

@@ -28,6 +28,10 @@ class LocalConfig(BaseModel):
     default_max_detections: int | None = None
     llamacpp_base_url: str | None = None
     llamacpp_model: str | None = None
+    llm_provider: str | None = None
+    replicate_api_token: str | None = None
+    replicate_model: str | None = None
+    replicate_reasoning_effort: str | None = None
 
 
 class RuntimeState(BaseModel):