mirror of
https://github.com/R0m1k3/Loki.git
synced 2026-10-12 01:37:06 +02:00
MichaelandClaude Opus 4.8
d19624b46b
L'auto-check séparé dans la boucle agent doublonnait _verify_written et signalait à tort les fichiers écrits par morceaux (mode append). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
394 lines
16 KiB
Python
394 lines
16 KiB
Python
"""Boucle agentique : tool-calling itératif au-dessus d'Ollama.
|
|
|
|
L'agent dialogue avec le modèle ; quand celui-ci demande un outil, on l'exécute,
|
|
on réinjecte le résultat, et on reboucle jusqu'à une réponse finale (ou la
|
|
limite d'itérations). La fonction est un générateur asynchrone d'événements
|
|
relayés tels quels au client via SSE.
|
|
|
|
Événements émis :
|
|
token {content} — fragment de texte de l'agent
|
|
tool_call {name, args} — début d'exécution d'un outil
|
|
tool_result {name, args, summary, status}
|
|
final {content, tools} — réponse complète + récap des outils
|
|
"""
|
|
from __future__ import annotations
|
|
|
|
import asyncio
|
|
import json
|
|
from typing import AsyncIterator
|
|
|
|
import httpx
|
|
|
|
from . import coder
|
|
from .ollama_client import OllamaError, ollama
|
|
from .tools import TOOL_DEFINITIONS, ToolError, run_tool
|
|
|
|
MAX_ITERATIONS = 6
|
|
MAX_TOOL_REPAIR_ATTEMPTS = 2
|
|
|
|
|
|
def _tools_not_supported(exc: OllamaError) -> bool:
|
|
"""Détecte un modèle incapable de function calling.
|
|
|
|
Deux cas :
|
|
- Ollama refuse explicitement (« does not support tools ») ;
|
|
- Ollama n'arrive pas à dériver un parseur d'appels d'outils du template
|
|
du modèle (« Unable to generate parser for this template ») — fréquent
|
|
sur des modèles exotiques dont le template Jinja lève une exception.
|
|
Dans les deux cas, on retombe sur une conversation simple, sans outils.
|
|
"""
|
|
message = str(exc).lower()
|
|
return (
|
|
"does not support tools" in message
|
|
or "does not support tool" in message
|
|
or "unable to generate parser for this template" in message
|
|
or "automatic parser generation failed" in message
|
|
)
|
|
|
|
|
|
def _thinking_not_supported(exc: OllamaError) -> bool:
|
|
"""Le modèle (ou son template) refuse le paramètre ``think``."""
|
|
message = str(exc).lower()
|
|
return "does not support thinking" in message or "thinking is not supported" in message
|
|
|
|
|
|
def _invalid_tool_arguments(exc: OllamaError) -> bool:
|
|
message = str(exc).lower()
|
|
return (
|
|
"invalid tool call arguments" in message
|
|
or "unexpected end of json" in message
|
|
or "failed to parse tool" in message
|
|
)
|
|
|
|
|
|
def _parse_args(raw) -> dict:
|
|
if isinstance(raw, dict):
|
|
return raw
|
|
if isinstance(raw, str):
|
|
try:
|
|
return json.loads(raw)
|
|
except json.JSONDecodeError:
|
|
return {}
|
|
return {}
|
|
|
|
|
|
async def run_agent(
|
|
model: str,
|
|
convo: list[dict],
|
|
*,
|
|
options: dict | None = None,
|
|
enabled_tools: list[str] | None = None,
|
|
confirm_shell: bool = True,
|
|
think: bool = True,
|
|
keep_alive: str | None = None,
|
|
mcp_tools: list[dict] | None = None,
|
|
) -> AsyncIterator[dict]:
|
|
# enabled_tools=None -> tous les outils ; liste vide -> aucun outil.
|
|
if enabled_tools is None:
|
|
tools = list(TOOL_DEFINITIONS)
|
|
elif enabled_tools:
|
|
tools = [
|
|
t for t in TOOL_DEFINITIONS if t["function"]["name"] in enabled_tools
|
|
]
|
|
else:
|
|
tools = []
|
|
# Outils MCP des serveurs activés (déjà résolus par l'appelant).
|
|
tools.extend(mcp_tools or [])
|
|
if not tools:
|
|
tools = None
|
|
collected: list[dict] = []
|
|
text_parts: list[str] = []
|
|
thinking_parts: list[str] = []
|
|
active_tools = tools
|
|
tool_fallback_used = False
|
|
tool_repair_attempts = 0
|
|
request_options = dict(options or {})
|
|
# On n'envoie ``think`` que pour le DÉSACTIVER (False) ; laissé à None, le
|
|
# modèle garde son comportement par défaut. Repli si le modèle le refuse.
|
|
request_think: bool | None = None if think else False
|
|
|
|
# Métriques cumulées sur tous les appels Ollama du tour agentique : Ollama
|
|
# les renvoie dans le chunk final (done=true) de chaque génération.
|
|
stats = {"eval_count": 0, "eval_duration": 0, "prompt_eval_count": 0}
|
|
|
|
def _accumulate(chunk: dict) -> None:
|
|
stats["eval_count"] += chunk.get("eval_count") or 0
|
|
stats["eval_duration"] += chunk.get("eval_duration") or 0
|
|
stats["prompt_eval_count"] += chunk.get("prompt_eval_count") or 0
|
|
|
|
try:
|
|
for _ in range(MAX_ITERATIONS):
|
|
content_buf = ""
|
|
thinking_buf = ""
|
|
thinking_status_sent = False
|
|
tool_calls: list[dict] = []
|
|
|
|
# Un modèle peut savoir discuter sans supporter les outils. Ollama
|
|
# refuse alors la requête entière : on retente une fois en chat simple.
|
|
while True:
|
|
try:
|
|
async for chunk in ollama.chat(
|
|
model,
|
|
convo,
|
|
tools=active_tools,
|
|
options=request_options,
|
|
think=request_think,
|
|
keep_alive=keep_alive,
|
|
stream=True,
|
|
):
|
|
msg = chunk.get("message", {})
|
|
token = msg.get("content", "")
|
|
thinking = msg.get("thinking", "")
|
|
if thinking:
|
|
thinking_buf += thinking
|
|
# Diffuse le raisonnement en direct pour l'afficher
|
|
# dans le panneau repliable du chat.
|
|
yield {"type": "thinking", "content": thinking}
|
|
if not thinking_status_sent:
|
|
thinking_status_sent = True
|
|
yield {"type": "status", "message": "Réflexion…"}
|
|
if token:
|
|
content_buf += token
|
|
yield {"type": "token", "content": token}
|
|
if msg.get("tool_calls"):
|
|
tool_calls.extend(msg["tool_calls"])
|
|
if chunk.get("done"):
|
|
_accumulate(chunk)
|
|
break
|
|
break
|
|
except OllamaError as exc:
|
|
if (
|
|
active_tools
|
|
and tool_repair_attempts < MAX_TOOL_REPAIR_ATTEMPTS
|
|
and not content_buf
|
|
and not tool_calls
|
|
and _invalid_tool_arguments(exc)
|
|
):
|
|
tool_repair_attempts += 1
|
|
request_options["num_predict"] = max(
|
|
int(request_options.get("num_predict", 0)), 4096
|
|
)
|
|
thinking_buf = ""
|
|
# NB : on réinjecte ce rappel en `user`, pas en `system`.
|
|
# Beaucoup de templates (Gemma, Mistral…) lèvent
|
|
# « System message must be at the beginning » dès qu'un
|
|
# message system apparaît ailleurs qu'en tête, ce qui
|
|
# ferait échouer toute la requête en 400.
|
|
convo.append(
|
|
{
|
|
"role": "user",
|
|
"content": (
|
|
"L'appel d'outil précédent contenait un JSON "
|
|
"tronqué. Réessaie immédiatement avec des arguments "
|
|
"JSON valides. Pour un fichier long, utilise write_file "
|
|
"en plusieurs appels : overwrite puis append, avec des "
|
|
"morceaux courts et complets."
|
|
),
|
|
}
|
|
)
|
|
yield {
|
|
"type": "notice",
|
|
"message": (
|
|
"Appel d'outil tronqué : nouvelle tentative "
|
|
f"{tool_repair_attempts}/{MAX_TOOL_REPAIR_ATTEMPTS}."
|
|
),
|
|
}
|
|
continue
|
|
if (
|
|
active_tools
|
|
and not tool_fallback_used
|
|
and not content_buf
|
|
and not tool_calls
|
|
and _tools_not_supported(exc)
|
|
):
|
|
active_tools = None
|
|
tool_fallback_used = True
|
|
yield {
|
|
"type": "notice",
|
|
"message": (
|
|
"Ce modèle ne supporte pas les outils ; "
|
|
"réponse en mode conversation simple."
|
|
),
|
|
}
|
|
continue
|
|
if request_think is not None and _thinking_not_supported(exc):
|
|
# Le modèle n'accepte pas qu'on désactive sa réflexion :
|
|
# on retire le paramètre et on relance.
|
|
request_think = None
|
|
continue
|
|
raise
|
|
|
|
assistant_turn: dict = {"role": "assistant", "content": content_buf}
|
|
if thinking_buf:
|
|
assistant_turn["thinking"] = thinking_buf
|
|
thinking_parts.append(thinking_buf)
|
|
if tool_calls:
|
|
assistant_turn["tool_calls"] = tool_calls
|
|
convo.append(assistant_turn)
|
|
if content_buf.strip():
|
|
text_parts.append(content_buf.strip())
|
|
|
|
if not tool_calls:
|
|
break
|
|
|
|
# Exécution des outils demandés, puis réinjection des résultats.
|
|
awaiting_confirmation = False
|
|
for tc in tool_calls:
|
|
fn = tc.get("function", {})
|
|
name = fn.get("name", "")
|
|
args = _parse_args(fn.get("arguments"))
|
|
|
|
yield {"type": "tool_call", "name": name, "args": args}
|
|
|
|
# run_shell est sensible : on demande validation au lieu d'exécuter.
|
|
if name == "run_shell" and confirm_shell:
|
|
command = args.get("command", "")
|
|
record = {
|
|
"name": name,
|
|
"args": args,
|
|
"summary": "validation requise",
|
|
"status": "pending",
|
|
}
|
|
collected.append(record)
|
|
yield {"type": "tool_confirm", "name": name, "command": command}
|
|
convo.append(
|
|
{
|
|
"role": "tool",
|
|
"tool_name": name,
|
|
"content": json.dumps(
|
|
{
|
|
"ok": False,
|
|
"status": "pending",
|
|
"message": "Commande en attente de validation "
|
|
"de l'utilisateur. N'exécute rien d'autre.",
|
|
},
|
|
ensure_ascii=False,
|
|
),
|
|
}
|
|
)
|
|
awaiting_confirmation = True
|
|
continue
|
|
|
|
# code_task : délégué au moteur code (Aider), long -> thread.
|
|
if name == "code_task":
|
|
code_model = await coder.pick_code_model(model)
|
|
result = await asyncio.to_thread(
|
|
coder.run_code_task,
|
|
args.get("instruction", ""),
|
|
code_model,
|
|
args.get("files") or [],
|
|
)
|
|
summary = result.get("summary", "terminé")
|
|
status = "ok" if result.get("ok") else "error"
|
|
record = {"name": name, "args": {"instruction": args.get("instruction", "")},
|
|
"summary": summary, "status": status}
|
|
collected.append(record)
|
|
yield {"type": "tool_result", **record}
|
|
convo.append({
|
|
"role": "tool",
|
|
"tool_name": name,
|
|
"content": json.dumps(
|
|
{k: result.get(k) for k in ("ok", "summary", "files", "text")},
|
|
ensure_ascii=False,
|
|
),
|
|
})
|
|
continue
|
|
|
|
# Outils MCP : dispatch asynchrone vers le serveur concerné.
|
|
if name.startswith("mcp_"):
|
|
from .mcp_client import manager as mcp_manager
|
|
mcp_res = await mcp_manager.call_tool(name, args)
|
|
status = "ok" if mcp_res["ok"] else "error"
|
|
record = {"name": name, "args": args,
|
|
"summary": mcp_res["summary"], "status": status}
|
|
collected.append(record)
|
|
yield {"type": "tool_result", **record}
|
|
convo.append({
|
|
"role": "tool",
|
|
"tool_name": name,
|
|
"content": json.dumps(
|
|
{"ok": mcp_res["ok"], "content": mcp_res["content"]},
|
|
ensure_ascii=False,
|
|
),
|
|
})
|
|
continue
|
|
|
|
try:
|
|
result = run_tool(name, args)
|
|
summary = result.get("summary", "terminé")
|
|
status = result.get("_status", "ok")
|
|
except ToolError as exc:
|
|
result = {"ok": False, "error": str(exc)}
|
|
summary = str(exc)
|
|
status = "error"
|
|
|
|
record = {"name": name, "args": args, "summary": summary, "status": status}
|
|
collected.append(record)
|
|
yield {"type": "tool_result", **record}
|
|
|
|
convo.append(
|
|
{
|
|
"role": "tool",
|
|
"tool_name": name,
|
|
"content": json.dumps(result, ensure_ascii=False),
|
|
}
|
|
)
|
|
|
|
# Une commande shell attend une validation : on interrompt la boucle.
|
|
if awaiting_confirmation:
|
|
# Laisse le modèle conclure son tour (message d'attente).
|
|
final_chunk = ""
|
|
async for chunk in ollama.chat(
|
|
model, convo, options=request_options,
|
|
think=request_think, keep_alive=keep_alive, stream=True
|
|
):
|
|
tok = chunk.get("message", {}).get("content", "")
|
|
if tok:
|
|
final_chunk += tok
|
|
yield {"type": "token", "content": tok}
|
|
if chunk.get("done"):
|
|
_accumulate(chunk)
|
|
break
|
|
if final_chunk.strip():
|
|
text_parts.append(final_chunk.strip())
|
|
break
|
|
except OllamaError as exc:
|
|
# Échec signalé par Ollama (HTTP ou ligne d'erreur dans le flux) :
|
|
# souvent un débordement mémoire / contexte trop grand. On l'expose.
|
|
yield {"type": "error", "message": f"Ollama : {exc}"}
|
|
return
|
|
except (httpx.HTTPError, OSError) as exc:
|
|
yield {
|
|
"type": "error",
|
|
"message": f"Impossible de joindre Ollama ({ollama.host}) : {exc}",
|
|
}
|
|
return
|
|
|
|
final_content = "\n\n".join(text_parts).strip()
|
|
if not final_content and not collected:
|
|
yield {
|
|
"type": "error",
|
|
"message": (
|
|
"Le modèle a terminé sans renvoyer de texte (il n'a produit que "
|
|
"du raisonnement). Désactive « Mode réflexion » dans les Réglages, "
|
|
"ou essaie un modèle de chat plus récent."
|
|
),
|
|
}
|
|
return
|
|
|
|
eval_secs = stats["eval_duration"] / 1e9
|
|
final_stats = {
|
|
"eval_count": stats["eval_count"],
|
|
"prompt_eval_count": stats["prompt_eval_count"],
|
|
"tokens_per_sec": (
|
|
round(stats["eval_count"] / eval_secs, 1) if eval_secs > 0 else None
|
|
),
|
|
}
|
|
yield {
|
|
"type": "final",
|
|
"content": final_content,
|
|
"tools": collected,
|
|
"stats": final_stats,
|
|
"thinking": "\n\n".join(thinking_parts).strip(),
|
|
}
|