app.py

"""Inside Out - a chat where your emotions chime in.

A warm little app, inspired by Pixar's Inside Out, where several emotion agents
react to whatever is on your mind. The goal isn't to give advice - it's to help
you notice and name what you're actually feeling.

Run:  python app.py
"""

from __future__ import annotations

import atexit
import html
import os
import re
import subprocess
import time
import urllib.request
from functools import partial
from urllib.parse import urlparse

from authlib.integrations.starlette_client import OAuth
from fastapi import FastAPI, Request
from fastapi.responses import HTMLResponse, RedirectResponse
import gradio as gr
from huggingface_hub import InferenceClient
from starlette.middleware.sessions import SessionMiddleware
import uvicorn

from agents import emotion_reply, run_turn
from emotions import EMOTIONS, EMOTION_ORDER

# Load variables from a local .env file if python-dotenv is available, so that
# HF_TOKEN, Google OAuth creds, etc. don't have to be exported by hand. Real
# shell environment variables still take precedence.
try:
    from dotenv import load_dotenv

    load_dotenv()
except ImportError:
    pass

GOOGLE_CLIENT_ID = os.environ.get("GOOGLE_CLIENT_ID")
GOOGLE_CLIENT_SECRET = os.environ.get("GOOGLE_CLIENT_SECRET")
GOOGLE_ALLOWED_DOMAIN = os.environ.get("GOOGLE_ALLOWED_DOMAIN", "").lstrip("@")
GOOGLE_AUTH_ENABLED = bool(GOOGLE_CLIENT_ID and GOOGLE_CLIENT_SECRET)
SESSION_SECRET = os.environ.get("SESSION_SECRET", "inside-out-dev-session-secret")
HF_MODEL = "google/gemma-4-26B-A4B-it:deepinfra" # "openai/gpt-oss-20b", "Qwen/Qwen3.6-27B:ovhcloud"
SHOW_LOGIN = os.environ.get("SHOW_LOGIN", "").lower() in {"1", "true", "yes"}

# When True, talk to a local llama.cpp server (OpenAI-compatible API) instead of
# the hosted Hugging Face Inference API. Defaults to False; override with the
# env var LOCAL_SERVING=true. Start the server with, e.g.:
#   llama-server -hf ggml-org/gemma-4-26b-a4b-it-GGUF:Q4_K_M \
#                --jinja --reasoning-budget 0 -ngl 99 -c 8192 --port 8080
# NOTE: --reasoning-budget 0 is required for Gemma 4 - without it the model
# stays in "thinking" mode and returns empty content (replies fall back to
# canned demo lines). If port 8080 is taken, use another port and set
# LOCAL_LLM_BASE_URL to match (e.g. http://localhost:8088/v1).
LOCAL_SERVING = os.environ.get("LOCAL_SERVING", "false").lower() in {"1", "true", "yes"}
LOCAL_LLM_BASE_URL = os.environ.get("LOCAL_LLM_BASE_URL", "http://localhost:8088/v1")
LOCAL_LLM_MODEL = os.environ.get("LOCAL_LLM_MODEL", "nemotron-3-nano-omni-30b-a3b-reasoning")  # "gemma-4-26b-a4b-it"

# Whether the served model is a reasoning/"thinking" model. When False (default)
# and we serve locally, llama.cpp is started with --reasoning-budget 0 so the
# model answers directly. When True, the model is allowed to think and the agent
# layer strips the chain-of-thought from replies. Override with USE_REASONING=true.
USE_REASONING = os.environ.get("USE_REASONING", "false").lower() in {"1", "true", "yes"}

# Where to look for local GGUF weights, and the llama.cpp server binary, used
# when LOCAL_SERVING auto-serves the model. LOCAL_MODEL_PATH pins an exact GGUF.
LOCAL_MODELS_DIR = os.environ.get("LOCAL_MODELS_DIR", os.path.expanduser("~/models"))
LOCAL_MODEL_PATH = os.environ.get("LOCAL_MODEL_PATH", "")
LLAMA_SERVER_BIN = os.environ.get(
    "LLAMA_SERVER_BIN", os.path.expanduser("~/tools/llama.cpp/build/bin/llama-server")
)

# Startup diagnostic (boolean only - never logs the secret value) so you can
# confirm from the Space logs whether the HF_TOKEN secret reached the app.
print(
    f"[inside-out] startup: serving="
    f"{'local-llama.cpp' if LOCAL_SERVING else 'hf-inference'} "
    f"| reasoning={'on' if USE_REASONING else 'off'} "
    f"| HF_TOKEN present: {bool(os.environ.get('HF_TOKEN'))} "
    f"| model: {LOCAL_LLM_MODEL if LOCAL_SERVING else HF_MODEL}",
    flush=True,
)

oauth = OAuth()
if GOOGLE_AUTH_ENABLED:
    oauth.register(
        name="google",
        client_id=GOOGLE_CLIENT_ID,
        client_secret=GOOGLE_CLIENT_SECRET,
        server_metadata_url="https://accounts.google.com/.well-known/openid-configuration",
        client_kwargs={"scope": "openid email profile"},
    )

THEME = gr.themes.Soft(
    primary_hue="purple",
    secondary_hue="yellow",
    neutral_hue="slate",
    font=[gr.themes.GoogleFont("Quicksand"), "ui-rounded", "system-ui", "sans-serif"],
)

CSS = """
.gradio-container {
    background: radial-gradient(1200px 600px at 20% -10%, #fef3c7 0%, transparent 55%),
                radial-gradient(1000px 700px at 90% 0%, #e9d5ff 0%, transparent 50%),
                linear-gradient(160deg, #fdfbff 0%, #f3f0ff 100%) !important;
}
#title-wrap { text-align: center; margin: 6px 0 2px; }
#title-wrap h1 {
    font-size: 2.5rem; font-weight: 700; margin-bottom: 2px;
    background: linear-gradient(90deg,#f59e42,#f0533b,#9b6dd6,#5b8def,#ffd93b);
    -webkit-background-clip: text; background-clip: text; color: transparent;
}
#subtitle { text-align:center; color:#6b6480; font-size:1.02rem; margin-bottom:10px; }
#legend-hint { text-align:center; color:#9a93ad; font-size:0.82rem; margin:2px 0 6px; }
/* Emotion chips are now clickable buttons. */
#legend {
    display:flex !important; flex-wrap:wrap; gap:8px;
    justify-content:center; margin:4px 0 14px;
    background:transparent !important; border:0 !important;
}
#legend .emo-btn {
    flex:0 0 auto !important; width:auto !important; min-width:0 !important;
    padding:6px 14px !important; border-radius:999px !important;
    font-size:0.86rem !important; font-weight:600 !important;
    color:#3a3550 !important; background:#ffffffcc !important;
    border:1.5px solid #d7cdef !important;
    box-shadow:0 2px 8px rgba(120,100,180,0.10) !important;
    backdrop-filter: blur(4px);
    transition: transform .12s ease, box-shadow .12s ease;
}
#legend .emo-btn:hover {
    transform: translateY(-1px);
    box-shadow:0 5px 14px rgba(120,100,180,0.22) !important;
}
/* --- Chat surface: no grey boxes, let the bubbles breathe --- */
.gr-chatbot, #chat {
    border:none !important;
    background:transparent !important;
    box-shadow:none !important;
}
/* Neutralise the grey bubble fill at the source: whichever wrapper class
   this gradio version uses, none of them can paint the grey anymore. */
#chat, #chat * { --background-fill-secondary: transparent !important; }

#chat .bubble-wrap,
#chat .bot-row,
#chat .user-row,
#chat .message-wrap,
#chat .panel,
#chat .bubble { background: transparent !important; }
#chat .message-row { padding: 4px 0 !important; }

/* Bot / assistant messages: the colored inner cards stand on their own */
#chat .bot,
#chat .message.bot,
#chat .bubble,
#chat .bot .message-content,
#chat .message,
#chat .message-content {
    background: transparent !important;
    border: 0 !important;
    box-shadow: none !important;
    padding: 0 !important;
}

/* User messages: a soft glowing purple bubble */
#chat .user,
#chat .message.user {
    background: linear-gradient(135deg, #a78bfa 0%, #8b5cf6 100%) !important;
    color: #fff !important;
    border: 0 !important;
    border-radius: 16px 16px 4px 16px !important;
    box-shadow: 0 6px 18px rgba(139,92,246,0.30) !important;
    padding: 9px 14px !important;
    max-width: 78%;
}
#chat .user *,
#chat .message.user * { color: #fff !important; background: transparent !important; }

#chat button[aria-label*="copy" i],
#chat button[aria-label*="clear" i] { display: none !important; }

/* --- Composer: a floating rounded pill --- */
#composer {
    gap: 8px !important;
    background: #ffffffcc;
    border: 1.5px solid #e7dcff;
    border-radius: 999px;
    padding: 6px 8px 6px 18px;
    box-shadow: 0 8px 24px rgba(140,110,210,0.12);
    backdrop-filter: blur(6px);
    align-items: center;
}
#composer textarea, #composer input[type="text"] {
    background: transparent !important;
    border: 0 !important;
    box-shadow: none !important;
    font-size: 1rem !important;
    resize: none !important;
}
#composer button {
    border-radius: 999px !important;
    border: 0 !important;
    background: linear-gradient(135deg, #f59e42, #f0533b) !important;
    color: #fff !important;
    font-weight: 700 !important;
    box-shadow: 0 4px 12px rgba(240,83,59,0.30) !important;
}
footer { display:none !important; }
#reflect-note { color:#8a83a0; font-size:0.85rem; text-align:center; margin-top:6px; }
"""

# Tint each emotion chip-button with its own color.
CSS += "".join(
    f"#emo-btn-{k} {{ border-color:{EMOTIONS[k].color} !important; "
    f"color:{EMOTIONS[k].color} !important; }}\n"
    f"#emo-btn-{k}:hover {{ background:{EMOTIONS[k].color}14 !important; }}\n"
    for k in EMOTION_ORDER
)


def _safe_html(text: str) -> str:
    return html.escape(text).replace("\n", "<br>")


def bubble(emo_key: str, text: str) -> dict:
    """Render one emotion's line as a styled assistant message."""
    e = EMOTIONS[emo_key]
    safe_text = _safe_html(text)
    content = (
        f'<div style="border-left:5px solid {e.color}; background:{e.color}1A; '
        f'padding:9px 13px; border-radius:12px; margin:2px 0;">'
        f'<div style="font-weight:700; color:{e.color}; font-size:0.9rem; '
        f'margin-bottom:2px;">{e.emoji} {e.name}</div>'
        f'<div style="color:#2f2b40; line-height:1.45;">{safe_text}</div></div>'
    )
    return {"role": "assistant", "content": content}


def reflect_bubble(text: str, allow_html: bool = False) -> dict:
    rendered_text = text if allow_html else _safe_html(text)
    content = (
        '<div style="background:linear-gradient(135deg,#ffffff,#f6f1ff); '
        'border:1px dashed #c7b8f0; padding:11px 15px; border-radius:14px; '
        'margin:6px 0 2px; box-shadow:0 3px 14px rgba(150,120,220,0.12);">'
        '<div style="font-weight:700; color:#8b5cf6; font-size:0.86rem; '
        'margin-bottom:3px;">A gentle reflection</div>'
        f'<div style="color:#3a3550; line-height:1.5;">{rendered_text}</div></div>'
    )
    return {"role": "assistant", "content": content}


def find_local_gguf(model_id: str) -> str | None:
    """Find a local .gguf for a model, or None if not available locally.

    LOCAL_MODEL_PATH wins if set. Otherwise we walk LOCAL_MODELS_DIR and pick the
    file whose name shares the most tokens with the model id (so e.g.
    'nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16' matches a local
    'NVIDIA-Nemotron-3-Nano-Omni-30B-A3B-Reasoning-*.gguf').
    """
    if LOCAL_MODEL_PATH:
        return LOCAL_MODEL_PATH if os.path.isfile(LOCAL_MODEL_PATH) else None
    if not os.path.isdir(LOCAL_MODELS_DIR):
        return None
    name = model_id.split(":", 1)[0].rsplit("/", 1)[-1].lower()
    tokens = [t for t in re.split(r"[^a-z0-9]+", name) if t]
    best, best_score = None, 0
    for root, _dirs, files in os.walk(LOCAL_MODELS_DIR):
        for fname in files:
            low = fname.lower()
            if not low.endswith(".gguf") or "mmproj" in low:
                continue
            score = sum(1 for t in tokens if t in low)
            if score > best_score:
                best, best_score = os.path.join(root, fname), score
    # Require a meaningful overlap so we don't grab an unrelated model.
    return best if best_score >= max(2, len(tokens) // 2) else None


_local_server_proc = None


def _server_healthy() -> bool:
    health_url = LOCAL_LLM_BASE_URL.rsplit("/v1", 1)[0].rstrip("/") + "/health"
    try:
        with urllib.request.urlopen(health_url, timeout=2) as resp:
            return resp.status == 200
    except Exception:
        return False


def ensure_local_server() -> None:
    """When LOCAL_SERVING, make sure a llama.cpp server is up for the model.

    No-op if one is already responding (so uvicorn hot-reload won't double-spawn).
    Launches llama-server with the locally-discovered GGUF, disabling the
    reasoning budget unless USE_REASONING is set.
    """
    global _local_server_proc
    if not LOCAL_SERVING or _server_healthy():
        return
    gguf = find_local_gguf(HF_MODEL)
    if not gguf:
        print(
            f"[inside-out] LOCAL_SERVING on but no local GGUF for {HF_MODEL!r} "
            f"in {LOCAL_MODELS_DIR} (set LOCAL_MODEL_PATH). Falling back to demo.",
            flush=True,
        )
        return
    if not os.path.isfile(LLAMA_SERVER_BIN):
        print(f"[inside-out] llama-server not found at {LLAMA_SERVER_BIN}", flush=True)
        return
    port = str(urlparse(LOCAL_LLM_BASE_URL).port or 8088)
    cmd = [
        LLAMA_SERVER_BIN, "-m", gguf, "--jinja",
        "-ngl", "99", "-c", "8192", "--host", "127.0.0.1", "--port", port,
    ]
    if not USE_REASONING:
        cmd += ["--reasoning-budget", "0"]
    print(
        f"[inside-out] starting llama.cpp (reasoning={'on' if USE_REASONING else 'off'}): "
        f"{os.path.basename(gguf)} on :{port}",
        flush=True,
    )
    env = dict(os.environ, LD_LIBRARY_PATH=os.path.dirname(LLAMA_SERVER_BIN)
               + os.pathsep + os.environ.get("LD_LIBRARY_PATH", ""))
    _local_server_proc = subprocess.Popen(cmd, env=env)
    atexit.register(lambda: _local_server_proc and _local_server_proc.terminate())
    for _ in range(240):  # wait up to ~4 min for the weights to load
        if _server_healthy():
            print("[inside-out] llama.cpp server is ready", flush=True)
            return
        if _local_server_proc.poll() is not None:
            print("[inside-out] llama.cpp server exited during startup", flush=True)
            return
        time.sleep(1)
    print("[inside-out] timed out waiting for llama.cpp server", flush=True)


def _make_client():
    """Build the inference client.

    - LOCAL_SERVING: ensure a local llama.cpp server is running and point at it
      (OpenAI-compatible), no HF token needed.
    - otherwise: the hosted HF Inference API, or None when no token is set
      (which makes the app fall back to offline demo replies).
    """
    if LOCAL_SERVING:
        ensure_local_server()
        return InferenceClient(
            base_url=LOCAL_LLM_BASE_URL,
            api_key=os.environ.get("LOCAL_LLM_API_KEY", "sk-no-key-needed"),
        )
    token = os.environ.get("HF_TOKEN")
    return InferenceClient(token=token, model=HF_MODEL) if token else None


def _last_user_message(chat_history: list[dict]) -> str:
    """Most recent thing the person actually typed, for context."""
    for turn in reversed(chat_history):
        if turn.get("role") == "user":
            return str(turn.get("content", "")).strip()
    return ""


def respond(
    message: str,
    chat_history: list[dict] | None,
):
    message = (message or "").strip()
    chat_history = list(chat_history or [])
    if not message:
        return "", chat_history

    client = _make_client()

    chat_history = chat_history + [{"role": "user", "content": message}]

    replies, reflection = run_turn(message, chat_history[:-1], client=client)
    for key, text in replies:
        chat_history.append(bubble(key, text))
    chat_history.append(reflect_bubble(reflection))
    return "", chat_history


def chime(emo_key: str, chat_history: list[dict] | None):
    """The user tapped an emotion chip - let that emotion speak up directly.

    It reacts to the conversation so far (the most recent thing the person
    said, plus recent history). No orchestrator, no reflection - just this one
    emotion chiming in.
    """
    chat_history = list(chat_history or [])
    client = _make_client()

    context = _last_user_message(chat_history)
    if not context:
        context = "They haven't said anything yet - gently invite them to share."

    text = emotion_reply(emo_key, context, chat_history, client)
    chat_history.append(bubble(emo_key, text))
    return chat_history


def greeting() -> list[dict]:
    text = (
        "Hi there. This is a safe little space inside your head.<br>"
        "Tell me what's on your mind - a thought, a moment, anything - and your "
        "emotions will each chime in. There are no wrong feelings here."
    )
    return [reflect_bubble(text, allow_html=True)]


def create_demo(show_login: bool = False) -> gr.Blocks:
    with gr.Blocks(title="Inside Out - Chat with your emotions") as blocks:
        if show_login:
            with gr.Sidebar():
                if GOOGLE_AUTH_ENABLED:
                    gr.HTML(
                        '<a href="/logout" style="display:inline-block; padding:8px 12px; '
                        'border-radius:999px; border:1px solid #d8c7ff; '
                        'background:#ffffffcc; color:#6d28d9; font-weight:700; '
                        'text-decoration:none;">Sign out</a>'
                    )
                else:
                    gr.Markdown(
                        "Google login is off locally. Set `GOOGLE_CLIENT_ID` and "
                        "`GOOGLE_CLIENT_SECRET` to require Google sign-in."
                    )
                gr.Markdown(
                    f"Set `HF_TOKEN` to let `{HF_MODEL}` generate emotion responses. "
                    "Without it, the app uses demo replies."
                )

        gr.HTML(
            '<div id="title-wrap"><h1>Inside Out</h1></div>'
            '<div id="subtitle">Share what\'s on your mind, and let your emotions '
            'chime in to help you discover how you really feel.</div>'
        )
        gr.HTML(
            '<div id="legend-hint">Tap an emotion to let it speak up</div>'
        )
        emo_buttons: dict[str, gr.Button] = {}
        with gr.Row(elem_id="legend"):
            for k in EMOTION_ORDER:
                e = EMOTIONS[k]
                emo_buttons[k] = gr.Button(
                    f"{e.emoji} {e.name}",
                    elem_id=f"emo-btn-{k}",
                    elem_classes="emo-btn",
                    size="sm",
                    variant="secondary",
                )

        clear = gr.Button(
            "Start fresh", size="sm", variant="secondary", elem_id="start-fresh"
        )

        chatbot = gr.Chatbot(
            value=greeting(),
            elem_id="chat",
            height=460,
            show_label=False,
            sanitize_html=False,
            buttons=[],
            group_consecutive_messages=False,
        )

        with gr.Row(elem_id="composer"):
            msg = gr.Textbox(
                placeholder="What's on your mind today?",
                show_label=False,
                scale=8,
                autofocus=True,
                container=False,
            )
            send = gr.Button("Share", variant="primary", scale=1)

        gr.HTML('<div id="reflect-note">Made with love.</div>')

        send.click(respond, [msg, chatbot], [msg, chatbot])
        msg.submit(respond, [msg, chatbot], [msg, chatbot])
        clear.click(lambda: greeting(), None, chatbot)
        for k, btn in emo_buttons.items():
            btn.click(partial(chime, k), chatbot, chatbot)

    return blocks


demo = create_demo(show_login=SHOW_LOGIN)


def _google_user(request: Request) -> str | None:
    user = request.session.get("google_user")
    if not isinstance(user, dict):
        return None
    email = user.get("email")
    return str(email) if email else None


server = FastAPI()


@server.middleware("http")
async def redirect_unauthenticated_root(request: Request, call_next):
    if GOOGLE_AUTH_ENABLED and request.method == "GET" and request.url.path == "/":
        user = request.session.get("google_user")
        if not user:
            return RedirectResponse(url="/login")
    return await call_next(request)


server.add_middleware(
    SessionMiddleware,
    secret_key=SESSION_SECRET,
    same_site="lax",
    https_only=os.environ.get("COOKIE_SECURE", "").lower() in {"1", "true", "yes"},
)


@server.get("/login")
async def google_login(request: Request):
    if not GOOGLE_AUTH_ENABLED:
        return HTMLResponse(
            "<h1>Google login is not configured</h1>"
            "<p>Set GOOGLE_CLIENT_ID and GOOGLE_CLIENT_SECRET, then restart the app.</p>"
            '<p><a href="/">Continue to local demo</a></p>'
        )
    redirect_uri = request.url_for("google_auth_callback")
    return await oauth.google.authorize_redirect(request, redirect_uri)


@server.get("/auth/callback")
async def google_auth_callback(request: Request):
    token = await oauth.google.authorize_access_token(request)
    userinfo = token.get("userinfo")
    if userinfo is None:
        userinfo = await oauth.google.userinfo(token=token)

    email = userinfo.get("email")
    if not email:
        return HTMLResponse("<h1>Google login did not return an email address.</h1>", status_code=400)

    domain = email.rsplit("@", 1)[-1]
    if GOOGLE_ALLOWED_DOMAIN and domain != GOOGLE_ALLOWED_DOMAIN:
        return HTMLResponse(
            f"<h1>Access denied</h1><p>{html.escape(email)} is not in the allowed domain.</p>",
            status_code=403,
        )

    request.session["google_user"] = {
        "email": email,
        "name": userinfo.get("name") or email,
        "picture": userinfo.get("picture"),
    }
    return RedirectResponse(url="/")


@server.get("/logout")
async def google_logout(request: Request):
    request.session.clear()
    return RedirectResponse(url="/login" if GOOGLE_AUTH_ENABLED else "/")


app = gr.mount_gradio_app(
    server,
    demo,
    path="/",
    theme=THEME,
    css=CSS,
    auth_dependency=_google_user if GOOGLE_AUTH_ENABLED else None,
    ssr_mode=False,
)


if __name__ == "__main__":
    # Set DEV=1 to hot-reload on file changes (uvicorn watches the source and
    # re-imports the module, so CSS, callbacks, and the FastAPI mount all refresh).
    dev = os.environ.get("DEV", "").lower() in {"1", "true", "yes"}
    uvicorn.run(
        "app:app" if dev else app,
        host="0.0.0.0",
        port=int(os.environ.get("PORT", "7860")),
        reload=dev,
    )