flowdeck/app/services/llm_client.py

"""FlowDeck — LLM client abstraction (v4.10.0).

Abstraction over multiple LLM providers so the agent never talks to the DB
directly — it emits *tool intentions* (function calls) that AgentEngine turns
into guarded internal actions.

Supported providers (OpenAI-compatible chat-completions JSON response):
  openai, anthropic*, google*, deepseek, qwencloud, nvidia, openrouter, ollama.
  (* routed through an OpenAI-compatible gateway / any configured api_base)

When no API key is configured (or provider == "offline") the client falls back
to a deterministic, dependency-free *mock planner*. This keeps the whole agent
functional — and fully testable — with zero external calls, which is what the
local deployment and the test-suite rely on.
"""
from __future__ import annotations

import asyncio
import json
import logging
import re
from dataclasses import dataclass, field

import httpx

from app.config import settings

logger = logging.getLogger(__name__)

# Provider → default model + base URL when llm_model/api_base are empty.
PROVIDERS = {
    "openai":     ("https://api.openai.com/v1",      "gpt-4o"),
    "anthropic":  ("https://api.anthropic.com/v1",   "claude-opus-4-8"),
    "google":     ("https://generativelanguage.googleapis.com/v1beta", "gemini-2.0-pro"),
    "deepseek":   ("https://api.deepseek.com/v1",    "deepseek-chat"),
    "qwencloud":  ("https://dashscope.aliyuncs.com/compatible-mode/v1", "qwen-max"),
    "nvidia":     ("https://integrate.api.nvidia.com/v1", "nvidia/llama-3.1-70b-instruct"),
    "openrouter": ("https://openrouter.ai/api/v1",   "meta-llama/llama-3.3-70b-instruct"),
    "ollama":     ("http://localhost:11434/v1",      "llama3.1"),
    "offline":    (None, None),
}

# Curated model presets surfaced by /api/agent/providers for the UI selectors.
PROVIDER_MODELS: dict[str, list[str]] = {
    "openai":     ["gpt-4o", "gpt-4o-mini", "gpt-4.1", "gpt-4.1-mini", "o3-mini", "gpt-4-turbo"],
    "anthropic":  ["claude-opus-4-8", "claude-sonnet-4-5", "claude-3-5-sonnet", "claude-haiku-4-5"],
    "google":     ["gemini-2.0-pro", "gemini-2.0-flash", "gemini-1.5-pro", "gemini-1.5-flash"],
    "deepseek":   ["deepseek-chat", "deepseek-reasoner"],
    "qwencloud":  ["qwen-max", "qwen-plus", "qwen-turbo", "qwen-long"],
    "nvidia":     ["nvidia/llama-3.1-70b-instruct", "nvidia/nemotron-4-340b-instruct"],
    "openrouter": ["meta-llama/llama-3.3-70b-instruct", "anthropic/claude-3.5-sonnet",
                   "openai/gpt-4o", "mistralai/mistral-large"],
    "ollama":     ["llama3.1", "llama3", "mistral", "qwen2.5", "gemma2", "mixtral"],
    "offline":    [],
}

# Llama-style / ChatML tool markers used by the mock planner.
_CREATE_PATTERNS = [
    (re.compile(r"cr[eéé]er\s+(?:une\s+)?collection[:\s]+[\"']?([A-Za-zÀ-ÿ0-9 _\-]+)"),
     lambda m: ("create_collection", {"name": m.group(1).strip()})),
    (re.compile(r"create\s+collection\s+[\"']?([A-Za-z0-9 _\-]+)"),
     lambda m: ("create_collection", {"name": m.group(1).strip()})),
    (re.compile(r"create\s+a\s+page\s+[\"']?([A-Za-z0-9 _\-]+)"),
     lambda m: ("create_page", {"title": m.group(1).strip()})),
]
_SEARCH_PATTERNS = [
    (re.compile(r"(?:recherche|search|trouve|find)\s+[\"']?([A-Za-z0-9 _\-]+)"),
     lambda m: ("search_workspace", {"query": m.group(1).strip()})),
]

# Loose fallback: "collection <Name>" → create_collection (covers "crée une collection X",
# "créer la collection X", "create collection X", etc.)
_COLLECTION_LINE = re.compile(
    r"\bcollection\b[:\s]+(?:nomm[ée]e\s+)?([A-Za-zÀ-ÿ0-9_][^,.\n()]*[A-Za-zÀ-ÿ0-9_])",
    re.IGNORECASE,
)


@dataclass
class LLMResponse:
    """Normalized completion: either a final text or one or more tool calls."""
    text: str = ""
    tool_calls: list[dict] = field(default_factory=list)
    model: str = ""
    usage: dict = field(default_factory=dict)


class LLMClient:
    """Multi-provider chat client with tool-calling support and offline mock."""

    def __init__(self, provider: str | None = None, api_key: str | None = None,
                 api_base: str | None = None):
        from .llm_config import get_llm_config  # local import avoids a cycle

        cfg = get_llm_config()
        self.provider = (provider or cfg["provider"] or "offline").lower()
        self.api_key = api_key if api_key is not None else cfg["api_key"]
        self.api_base = api_base if api_base is not None else cfg["api_base"]
        base, model = PROVIDERS.get(self.provider, (None, None))
        self.api_base = self.api_base or base
        self.default_model = cfg["model"] or model or "gpt-4o"

    # ── Public API ──

    async def complete(self, messages: list[dict], *, model: str | None = None,
                       tools: list[dict] | None = None,
                       stream: bool = False) -> LLMResponse:
        """Send a chat completion. Returns text and/or tool_calls."""
        model = model or self.default_model

        if self.provider == "offline" or not self._has_credentials():
            return await self._mock_complete(messages, model, tools)

        try:
            return await asyncio.wait_for(
                self._http_complete(messages, model, tools),
                timeout=settings.agent_run_timeout_seconds,
            )
        except Exception as exc:  # noqa: BLE001 — never mask a real-provider failure
            # On NE retombe PAS silencieusement sur le mock quand un fournisseur
            # réel est configuré : l'erreur doit remonter (SSE "error") pour que
            # l'utilisateur voie pourquoi rien n'a été généré.
            logger.warning("LLM provider '%s' failed (%s)", self.provider, exc)
            raise

    async def is_available(self) -> bool:
        """True when a real provider is configured."""
        return self.provider != "offline" and self._has_credentials()

    async def ping(self, *, model: str | None = None) -> LLMResponse:
        """Reach the provider without mock fallback (used by the "Test connection"
        UI). Raises on any real error so the caller can surface it."""
        model = model or self.default_model
        if self.provider == "offline":
            return LLMResponse(text="Mode hors-ligne (mock) — aucun appel réseau nécessaire.", model=model)
        if not self._has_credentials():
            raise PermissionError(f"Clé API manquante pour le provider « {self.provider} »")
        return await asyncio.wait_for(
            self._http_complete(
                [{"role": "user", "content": "Réponds uniquement par le mot : PONG"}],
                model,
                None,
            ),
            timeout=settings.agent_run_timeout_seconds,
        )

    # ── Helpers ──

    def _has_credentials(self) -> bool:
        if self.provider == "ollama":
            return True  # local, no key required
        return bool(self.api_key)

    def _endpoint(self) -> str:
        return f"{self.api_base.rstrip('/')}/chat/completions"

    async def _http_complete(self, messages, model, tools) -> LLMResponse:
        payload: dict = {
            "model": model,
            "messages": messages,
            "temperature": 0.2,
        }
        if tools:
            payload["tools"] = [{"type": "function", "function": t} for t in tools]
            payload["tool_choice"] = "auto"
        headers = {"Content-Type": "application/json"}
        if self.api_key:
            headers["Authorization"] = f"Bearer {self.api_key}"
        async with httpx.AsyncClient(timeout=settings.agent_run_timeout_seconds) as client:
            resp = await client.post(self._endpoint(), json=payload, headers=headers)
            resp.raise_for_status()
            data = resp.json()

        choice = data["choices"][0]["message"]
        text = choice.get("content") or ""
        tool_calls = []
        for tc in choice.get("tool_calls") or []:
            fn = tc.get("function") or {}
            try:
                args = json.loads(fn.get("arguments") or "{}")
            except json.JSONDecodeError:
                args = {}
            tool_calls.append({
                "id": tc.get("id") or "",
                "name": fn.get("name"),
                "arguments": args,
                "arguments_raw": fn.get("arguments") or "",
            })

        return LLMResponse(
            text=text,
            tool_calls=tool_calls,
            model=model,
            usage=data.get("usage", {}),
        )

    # ── Offline mock planner (deterministic, no network) ──

    async def _mock_complete(self, messages, model, tools) -> LLMResponse:
        user_content = self._last_user_content(messages)
        sys_content = self._system_content(messages)
        # Only the user's objective drives the planner. The engine appends the
        # workspace/document snapshot under "# Contexte"; that text must never
        # trigger keyword heuristics (a doc mentioning "recherche"/"collection"
        # used to misroute content requests into tool calls).
        objective = user_content.split("\n# Contexte")[0]

        # Once tool results are already in the conversation, we have acted:
        # stop issuing new tool calls and conclude.
        if any(m.get("role") == "tool" for m in messages):
            return LLMResponse(
                text="Objectif traité — actions enregistrées dans le journal d'audit.",
                model=model,
            )

        # Skill-driven: if the objective names a known skill, mirror its template.
        skill_hint = self._extract_skill_hint(objective)

        if skill_hint == "sprint":
            return LLMResponse(
                tool_calls=[
                    {"name": "read_gitea_issues", "arguments": {"owner": "bruno", "repo": "flowdeck", "state": "open"}},
                    {"name": "create_collection", "arguments": {"name": "Sprint"}},
                    {"name": "add_property", "arguments": {"collection_id": 0, "name": "Status", "prop_type": "select", "options": ["Todo", "In Progress", "Done"]}},
                    {"name": "create_view", "arguments": {"collection_id": 0, "view_type": "board"}},
                ],
                text="Plan: analyze open issues, then build a sprint board.",
                model=model,
            )

        # Inline content-generation ("Ask AI" / "AI meeting note") — answered
        # before the tool-intent heuristics and scoped to the user objective
        # only, so an injected "# Contexte" that happens to mention "collection"
        # can't misroute a writing request into a create-collection action.
        draft = self._draft_reply(objective)
        if draft:
            return LLMResponse(text=draft, model=model)

        # Documents / espaces de travail (offline): unambiguous intents resolved
        # from the objective — create a document (optionally in a named
        # workspace) or list the accessible workspaces.
        doc_args = self._document_create_args(objective)
        if doc_args is not None:
            return LLMResponse(
                tool_calls=[{"name": "create_document", "arguments": doc_args}],
                text="Plan: création d'un document.",
                model=model,
            )
        if self._is_workspaces_request(objective):
            return LLMResponse(
                tool_calls=[{"name": "read_workspaces", "arguments": {}}],
                text="Plan: lister les espaces de travail.",
                model=model,
            )

        # Exact keyword → tool intent resolution (objective only).
        for regex, builder in _CREATE_PATTERNS:
            m = regex.search(objective)
            if m:
                return LLMResponse(
                    tool_calls=[dict(name=name, arguments=self._bind_placeholders(args, sys_content)) for name, args in [builder(m)]],
                    text=f"Plan: running {builder(m)[0]}.",
                    model=model,
                )

        for regex, builder in _SEARCH_PATTERNS:
            m = regex.search(objective)
            if m:
                return LLMResponse(
                    tool_calls=[dict(name=name, arguments=args) for name, args in [builder(m)]],
                    text=f"Plan: searching '{m.group(1)}'.",
                    model=model,
                )

        # Loose "collection <X>" detection → treat as a create intent.
        low = objective.lower()
        if "collection" in low:
            m = _COLLECTION_LINE.search(objective)
            if m:
                name = m.group(1).strip()
                return LLMResponse(
                    tool_calls=[{"name": "create_collection",
                                 "arguments": self._bind_placeholders({"name": name}, sys_content)}],
                    text=f"Plan: create collection '{name}'.",
                    model=model,
                )

        # Plain conversational objective → final answer (no tool).
        return LLMResponse(
            text=self._summarize(objective),
            model=model,
        )

    def _bind_placeholders(self, args: dict, sys_content: str) -> dict:
        """Inject a collection id from the context when the planner left it as 0."""
        args = dict(args)
        if args.get("collection_id") == 0:
            match = re.search(r"Collection IDs?:\s*([0-9,\s]+)", sys_content)
            if match:
                ids = [int(x) for x in re.split(r"[,\s]+", match.group(1).strip()) if x.isdigit()]
                if ids:
                    args["collection_id"] = ids[0]
        return args

    @staticmethod
    def _document_create_args(content: str) -> dict | None:
        """Deterministic `create_document` intent for the offline mock.

        Only fires when the user clearly asks to *create* a document (a content
        rewrite such as « résume / traduis ce document » is left to `_draft_reply`).
        Returns None when the message is not a create-document intent.
        """
        low = content.lower()
        if "document" not in low:
            return None
        if not any(k in low for k in ("création", "créer", "crée", "crées", "create",
                                      "nouveau document", "nouvelle page", "faire un")):
            return None
        quotes = re.findall(r'[«"]([^«»"]{1,80})[»"]', content)
        title = quotes[0].strip() if quotes else None
        if not title:
            m = re.search(
                r"\bdocument\b\s*(?:nomm[ée]e?\s+|intitul[ée]e?\s+|appel[ée]e?\s+)?"
                r'[«"]?\s*([A-Za-zÀ-ÿ0-9][A-Za-zÀ-ÿ0-9_ \-]{1,60})',
                content, re.IGNORECASE,
            )
            if m:
                title = m.group(1).strip()
        if not title:
            return None
        args = {"title": title}
        if len(quotes) > 1 and re.search(r"\b(workspace|espace de travail)\b", low):
            args["workspace_name"] = quotes[-1].strip()
        return args

    @staticmethod
    def _is_workspaces_request(content: str) -> bool:
        """True when the user asks to list / locate the workspaces."""
        low = content.lower()
        has_ws = any(w in low for w in ("workspace", "espace de travail", "espaces de travail"))
        has_verb = any(v in low for v in ("liste", "lister", "list", "quels", "montre",
                                          "affiche", "mes espaces", "ou sont", "où sont"))
        return has_ws and has_verb

    @staticmethod
    def _system_content(messages) -> str:
        return "\n".join(m.get("content", "") for m in messages if m.get("role") == "system")

    @staticmethod
    def _last_user_content(messages) -> str:
        for m in reversed(messages):
            if m.get("role") == "user":
                c = m.get("content", "")
                if isinstance(c, list):
                    return " ".join(p.get("text", "") for p in c if isinstance(p, dict))
                return str(c)
        return ""

    @staticmethod
    def _extract_skill_hint(content: str) -> str | None:
        low = content.lower()
        if "sprint" in low or "préparation de sprint" in low:
            return "sprint"
        return None

    @staticmethod
    def _summarize(content: str) -> str:
        """Produce a terse final summary from a conversational objective."""
        content = content.split("\n# Contexte")[0]
        return (content[:600] + "…" if len(content) > 600 else content)

    def _draft_reply(self, content: str) -> str | None:
        """Generate usable structured copy for content/meeting requests when no
        real LLM is configured (offline mock). Returns None when the message is
        not a clear content-generation intent so other paths keep their behavior.
        """
        from datetime import date

        low = content.lower()
        title = None
        m = re.search(r"intitul[ée]e\s*[«\"']([^»\"']+)[»\"']", content)
        if m:
            title = m.group(1).strip()
        today = date.today().isoformat()

        # ── AI meeting note template ──
        if any(k in low for k in ("ai meeting note", "meeting note",
                                  "compte-rendu", "compte rendu",
                                  "notes de réunion", "réunion")):
            return (
                "📅 AI Meeting Note\n"
                f"Date : {today} · Participants : (à renseigner)\n"
                "\n"
                "## Résumé\n"
                "Point central de la discussion et contexte (à compléter).\n"
                "\n"
                "## Décisions\n"
                "• Décision 1 — valider le périmètre et les responsables.\n"
                "• Décision 2 — définir la prochaine échéance.\n"
                "\n"
                "## Action items\n"
                "☐ Action 1 — responsable : …, échéance : …\n"
                "☐ Action 2 — responsable : …, échéance : …\n"
                "\n"
                "## Prochaines étapes\n"
                "• Planifier le suivi et archiver ce compte-rendu.\n"
            )

        # ── Traduction / analyse du document (hors-ligne) ──
        if re.search(r"traduis|traduit|translate", low):
            return (
                "⚠️ **Traduction non disponible en mode hors-ligne** (aucun modèle d'IA "
                "connecté).\n\n"
                "Connectez un fournisseur dans **Paramètres → Agent & IA**, puis relancez "
                "« Traduire cette page » : le document traduit apparaîtra ici, avec un aperçu "
                "à approuver ou à rejeter avant application."
            )
        if re.search(r"r[ée]sum|am[ée]lior|sugg[èe]re des|propose des", low):
            return (
                "⚠️ **Cette action nécessite un modèle d'IA connecté** pour analyser le "
                "document.\n\n"
                "Configurez une clé API dans **Paramètres → Agent & IA**, puis relancez "
                "l'action : l'agent générera la proposition ici, avec un aperçu à approuver "
                "ou à rejeter avant application."
            )

        # ── Page / document draft (contextual "Ask AI") ──
        if title:
            t = title[:80]
            return (
                f"{t}\n"
                "\n"
                f"Présentation générale du sujet « {t} » : objectif, contexte et "
                "public visé en quelques phrases. (Document généré hors-ligne — "
                "connectez une clé API pour une rédaction complète.)\n"
                "\n"
                "## Objectif\n"
                "• Clarifier le besoin couvert par ce document.\n"
                "• Lister les livrables attendus.\n"
                "\n"
                "## Points clés\n"
                "• Idée principale 1 avec les arguments associés.\n"
                "• Idée principale 2 et les exemples concrets.\n"
                "\n"
                "## Prochaines étapes\n"
                "• Relire, compléter et mettre en forme ce contenu.\n"
            )

        # ── Generic drafting verb, no title (typing directly in the chat) ──
        if re.search(r"^(r[ée]dige|[ée]cris|[ée]crire|g[ée]n[èe]re|produis|d[ée]veloppe|"
                     r"[ée]cris\s+un|g[ée]n[èe]re\s+un|r[ée]dige\s+un)\b", low):
            return (
                "## Introduction\n"
                "Contexte et objectif de ce texte, en une à deux phrases.\n"
                "\n"
                "## Développement\n"
                "• Premier argument structuré avec un exemple.\n"
                "• Deuxième argument appuyé par une donnée ou un fait.\n"
                "\n"
                "## Conclusion\n"
                "Synthèse et prochaine étape recommandée.\n"
            )

        return None