Spaces:
Running
Running
Skills v2 across agent/Kype/ZyKype: true auto-inject (builtins+learned) into planner AND analysis, skill line in agent thinking stream, skills_used in Kype/ZyKype results, compact recent-message memory for Kype/ZyKype (saves after each run, no old history), ask-when-unsure prompts, relaxed over-strict safety gates (env-var keys, 16-digit numbers, educational security, general advice topics pass), new web_research + create_document skills
Browse files- agent/agent.py +33 -3
- agent/kype.py +81 -9
- agent/kype_rules.py +18 -10
- agent/prompts/system_prompt.py +28 -16
- agent/safety_gate.py +7 -5
- agent/skills_catalog.py +116 -4
- agent/zyloo_agent.py +92 -5
agent/agent.py
CHANGED
|
@@ -1,5 +1,5 @@
|
|
| 1 |
"""
|
| 2 |
-
NeuraPrompt Agent v9.
|
| 3 |
-----------------------------------------------------------------------------
|
| 4 |
Updates over v9.1:
|
| 5 |
* MongoDB persistent memory (last 5 messages + user profile) β agent_memory.py
|
|
@@ -27,6 +27,18 @@ from .prompts.system_prompt import get_system_prompt
|
|
| 27 |
from .memory.memory import get_memory_manager
|
| 28 |
from .tools.registry import register_tool, get_tool, get_tool_descriptions, list_tools
|
| 29 |
from .tools import web_tools, code_tools, file_tools, vision_tools, document_tools, media_tools, skill_tools
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 30 |
from .tools.github_tools import (
|
| 31 |
github_list_repos, github_create_repo, github_get_repo,
|
| 32 |
github_list_issues, github_create_issue, github_read_file,
|
|
@@ -629,9 +641,22 @@ async def stream_agent(
|
|
| 629 |
# Track if memory was used for the indicator
|
| 630 |
memory_used = bool(db_memory)
|
| 631 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 632 |
# Use db_memory instead of the old in-memory system
|
| 633 |
system_prompt = _build_system_prompt(db_memory)
|
| 634 |
|
|
|
|
|
|
|
|
|
|
|
|
|
| 635 |
messages = [
|
| 636 |
{"role": "system", "content": system_prompt},
|
| 637 |
{"role": "user", "content": f"user_id: {user_id}\nGoal: {goal}" + (f"\nContext: {context}" if context else "")}
|
|
@@ -824,7 +849,12 @@ async def stream_agent(
|
|
| 824 |
yield f"data: {json.dumps({'type': 'usage', 'usage': total_usage})}\n\n"
|
| 825 |
return
|
| 826 |
|
| 827 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 828 |
|
| 829 |
result, tool_elapsed = await _execute_tool_and_update_history(action, tool_input, messages, thought)
|
| 830 |
|
|
@@ -892,7 +922,7 @@ async def agent_stream(req: AgentRequest):
|
|
| 892 |
async def health_check():
|
| 893 |
return {
|
| 894 |
"status": "ok",
|
| 895 |
-
"version": "9.
|
| 896 |
"tools": list_tools(),
|
| 897 |
"groq_fallback_enabled": bool(GROQ_API_KEY),
|
| 898 |
"token_tracking_enabled": _get_token_usage() is not None,
|
|
|
|
| 1 |
"""
|
| 2 |
+
NeuraPrompt Agent v9.4 β True skills auto-inject + skill thinking line + ask-when-unsure
|
| 3 |
-----------------------------------------------------------------------------
|
| 4 |
Updates over v9.1:
|
| 5 |
* MongoDB persistent memory (last 5 messages + user profile) β agent_memory.py
|
|
|
|
| 27 |
from .memory.memory import get_memory_manager
|
| 28 |
from .tools.registry import register_tool, get_tool, get_tool_descriptions, list_tools
|
| 29 |
from .tools import web_tools, code_tools, file_tools, vision_tools, document_tools, media_tools, skill_tools
|
| 30 |
+
|
| 31 |
+
# ---- Skills auto-injection (v9.4) ----
|
| 32 |
+
try:
|
| 33 |
+
from .skills_catalog import resolve_skills, format_skill_block, skill_thinking_line
|
| 34 |
+
except ImportError:
|
| 35 |
+
try:
|
| 36 |
+
from skills_catalog import resolve_skills, format_skill_block, skill_thinking_line # type: ignore
|
| 37 |
+
except ImportError:
|
| 38 |
+
resolve_skills = lambda task, user_id="", limit=2: [] # type: ignore
|
| 39 |
+
format_skill_block = lambda skills: "" # type: ignore
|
| 40 |
+
skill_thinking_line = lambda skills: "" # type: ignore
|
| 41 |
+
log.warning("[Agent] skills_catalog not available β skill auto-inject disabled")
|
| 42 |
from .tools.github_tools import (
|
| 43 |
github_list_repos, github_create_repo, github_get_repo,
|
| 44 |
github_list_issues, github_create_issue, github_read_file,
|
|
|
|
| 641 |
# Track if memory was used for the indicator
|
| 642 |
memory_used = bool(db_memory)
|
| 643 |
|
| 644 |
+
# ββ v9.4: Skills auto-inject (builtins + learned) ββ
|
| 645 |
+
# Matched skill bodies go straight into the prompt β the model no longer
|
| 646 |
+
# wastes steps calling load_skill for a skill we already know matches.
|
| 647 |
+
matched_skills = resolve_skills(goal, user_id=user_id, limit=2)
|
| 648 |
+
skill_block = format_skill_block(matched_skills)
|
| 649 |
+
if skill_block:
|
| 650 |
+
db_memory = (db_memory + "\n\n" + skill_block) if db_memory else skill_block
|
| 651 |
+
skill_line = skill_thinking_line(matched_skills)
|
| 652 |
+
|
| 653 |
# Use db_memory instead of the old in-memory system
|
| 654 |
system_prompt = _build_system_prompt(db_memory)
|
| 655 |
|
| 656 |
+
# Show the matched skills in the thinking stream BEFORE step 1
|
| 657 |
+
if skill_line:
|
| 658 |
+
yield _make_sse("reasoning", 0, skill_line, "", 0.0, skill_line)
|
| 659 |
+
|
| 660 |
messages = [
|
| 661 |
{"role": "system", "content": system_prompt},
|
| 662 |
{"role": "user", "content": f"user_id: {user_id}\nGoal: {goal}" + (f"\nContext: {context}" if context else "")}
|
|
|
|
| 849 |
yield f"data: {json.dumps({'type': 'usage', 'usage': total_usage})}\n\n"
|
| 850 |
return
|
| 851 |
|
| 852 |
+
if action == "load_skill" and isinstance(tool_input, dict):
|
| 853 |
+
_skill_name = tool_input.get("name") or "skill"
|
| 854 |
+
yield _make_sse("reasoning", step, f"π Using {str(_skill_name).replace('_', ' ')} skill for this task", "", 0.0, f"Loading skill: {_skill_name}")
|
| 855 |
+
yield _make_sse("tool_start", step,
|
| 856 |
+
f"Loading skill: {tool_input.get('name')}..." if action == "load_skill" and isinstance(tool_input, dict) and tool_input.get("name")
|
| 857 |
+
else f"Running {action}...", action, time.time() - step_start)
|
| 858 |
|
| 859 |
result, tool_elapsed = await _execute_tool_and_update_history(action, tool_input, messages, thought)
|
| 860 |
|
|
|
|
| 922 |
async def health_check():
|
| 923 |
return {
|
| 924 |
"status": "ok",
|
| 925 |
+
"version": "9.4.0-skills", # v9.4 β true skill auto-inject + skill thinking line + ask-when-unsure
|
| 926 |
"tools": list_tools(),
|
| 927 |
"groq_fallback_enabled": bool(GROQ_API_KEY),
|
| 928 |
"token_tracking_enabled": _get_token_usage() is not None,
|
agent/kype.py
CHANGED
|
@@ -73,6 +73,32 @@ except ImportError:
|
|
| 73 |
def format_skills_hint(task):
|
| 74 |
return ""
|
| 75 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 76 |
|
| 77 |
# =================================================================
|
| 78 |
# CONFIGURATION
|
|
@@ -183,6 +209,10 @@ Output format (strict β no preamble):
|
|
| 183 |
3. RISKS: <bullet list or "none">
|
| 184 |
If DEBUGGER FEEDBACK is provided, address every flagged issue.
|
| 185 |
|
|
|
|
|
|
|
|
|
|
|
|
|
| 186 |
FORMAT: itineraries β markdown tables; code β fenced blocks; comparisons β tables. Stay on task.
|
| 187 |
{KYPE_RULES_PROMPT}"""
|
| 188 |
|
|
@@ -192,6 +222,9 @@ ANALYSIS_SYSTEM = f"""You are the ANALYSIS of Kype.
|
|
| 192 |
Execute the plan and deliver the full answer to the user's task.
|
| 193 |
No meta-commentary about the plan or your role β just the result.
|
| 194 |
If DEBUGGER FEEDBACK is provided, address each issue explicitly.
|
|
|
|
|
|
|
|
|
|
| 195 |
|
| 196 |
FORMAT: itineraries MUST use a markdown table (Day|Morning|Afternoon|Evening|Notes);
|
| 197 |
code MUST use fenced blocks with language tags; comparisons use markdown tables.
|
|
@@ -234,6 +267,7 @@ class KypeResult:
|
|
| 234 |
total_elapsed: float = 0.0
|
| 235 |
fallback_used: Optional[str] = None
|
| 236 |
tool_used: Optional[str] = None
|
|
|
|
| 237 |
usage: Dict[str, int] = field(default_factory=lambda: {"input_tokens": 0, "output_tokens": 0, "total_tokens": 0})
|
| 238 |
|
| 239 |
def to_dict(self) -> Dict[str, Any]:
|
|
@@ -244,6 +278,7 @@ class KypeResult:
|
|
| 244 |
"approved": self.approved,
|
| 245 |
"total_elapsed": round(self.total_elapsed, 2),
|
| 246 |
"fallback_used": self.fallback_used,
|
|
|
|
| 247 |
"tool_used": self.tool_used,
|
| 248 |
"usage": self.usage,
|
| 249 |
"trace": [
|
|
@@ -483,7 +518,9 @@ class KypeAgent:
|
|
| 483 |
debugger_model: str = DEFAULT_DEBUGGER_MODEL,
|
| 484 |
max_iterations: int = MAX_ITERATIONS,
|
| 485 |
allow_tools: bool = True,
|
|
|
|
| 486 |
):
|
|
|
|
| 487 |
self.api_key = api_key or os.getenv("OPENROUTE_KEY", "")
|
| 488 |
if not self.api_key and not GROQ_API_KEY:
|
| 489 |
raise EnvironmentError(
|
|
@@ -812,14 +849,13 @@ class KypeAgent:
|
|
| 812 |
# ---- stage wrappers ---------------------------------------------------
|
| 813 |
def plan(self, task: str, feedback: str = "") -> Tuple[str, str, str, str, dict]:
|
| 814 |
"""Returns (plan, tool_name, tool_result, provider, usage)."""
|
| 815 |
-
skill_hint = ""
|
| 816 |
-
try:
|
| 817 |
-
skill_hint = format_skills_hint(task) or ""
|
| 818 |
-
except Exception:
|
| 819 |
-
skill_hint = ""
|
| 820 |
user = ""
|
| 821 |
-
|
| 822 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 823 |
user += f"TASK:\n{task}\n"
|
| 824 |
if feedback:
|
| 825 |
user += f"\nDEBUGGER FEEDBACK (address every issue):\n{feedback}\n"
|
|
@@ -859,7 +895,14 @@ class KypeAgent:
|
|
| 859 |
|
| 860 |
def analyze(self, task: str, plan: str, feedback: str = "",
|
| 861 |
tool_name: str = "", tool_result: str = "") -> Tuple[str, str, dict]:
|
| 862 |
-
user =
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 863 |
if tool_name and tool_result:
|
| 864 |
user += f"\n[TOOL RESULT from {tool_name}]:\n{tool_result}\n"
|
| 865 |
if feedback:
|
|
@@ -946,6 +989,21 @@ class KypeAgent:
|
|
| 946 |
if not task or not task.strip():
|
| 947 |
raise ValueError("task must be a non-empty string")
|
| 948 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 949 |
t_start = time.time()
|
| 950 |
trace: List[AgentResponse] = []
|
| 951 |
fallback_used: Optional[str] = None
|
|
@@ -1022,8 +1080,13 @@ class KypeAgent:
|
|
| 1022 |
verdict.get("issues", []), ensure_ascii=False
|
| 1023 |
)
|
| 1024 |
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1025 |
return KypeResult(
|
| 1026 |
task=task,
|
|
|
|
| 1027 |
final_answer=draft,
|
| 1028 |
iterations=iterations_completed,
|
| 1029 |
approved=approved,
|
|
@@ -1449,6 +1512,7 @@ async def kype_run(req: KypeRequest):
|
|
| 1449 |
debugger_model=req.debugger_model or DEFAULT_DEBUGGER_MODEL,
|
| 1450 |
max_iterations=req.max_iterations,
|
| 1451 |
allow_tools=req.allow_tools,
|
|
|
|
| 1452 |
)
|
| 1453 |
result = await asyncio.to_thread(agent.run, task)
|
| 1454 |
except EnvironmentError as e:
|
|
@@ -1474,6 +1538,14 @@ async def kype_run(req: KypeRequest):
|
|
| 1474 |
except Exception as e:
|
| 1475 |
log.warning(f"[Kype] Failed to record token usage: {e}")
|
| 1476 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1477 |
# ββ Inject ad into final answer (~30% chance) ββ
|
| 1478 |
try:
|
| 1479 |
ad = maybe_get_ad(user_id=user_id, user_tier="pro" if polar_subscribed else "free", source="kype")
|
|
@@ -1501,7 +1573,7 @@ async def kype_health():
|
|
| 1501 |
return {
|
| 1502 |
"status": "maintenance" if maintenance else "ok",
|
| 1503 |
"service": "kype",
|
| 1504 |
-
"version": "1.
|
| 1505 |
"roles": ["planner", "analysis", "debugger"],
|
| 1506 |
"max_file_bytes": MAX_FILE_BYTES,
|
| 1507 |
"rate_window_seconds": KYPE_RATE_WINDOW_SECONDS,
|
|
|
|
| 73 |
def format_skills_hint(task):
|
| 74 |
return ""
|
| 75 |
|
| 76 |
+
# ---- v1.7: compact memory + unified skills resolver ----
|
| 77 |
+
try:
|
| 78 |
+
from .agent_memory import build_memory_context, save_memory
|
| 79 |
+
except ImportError:
|
| 80 |
+
try:
|
| 81 |
+
from agent_memory import build_memory_context, save_memory # type: ignore
|
| 82 |
+
except ImportError:
|
| 83 |
+
build_memory_context = lambda uid, query="": "" # type: ignore
|
| 84 |
+
save_memory = lambda uid, msg, resp: None # type: ignore
|
| 85 |
+
log.warning("[Kype] agent_memory not available β memory disabled")
|
| 86 |
+
|
| 87 |
+
try:
|
| 88 |
+
from .skills_catalog import (
|
| 89 |
+
resolve_skills as _resolve_skills, format_skill_block as _format_skill_block,
|
| 90 |
+
skills_used_names as _skills_used_names)
|
| 91 |
+
except ImportError:
|
| 92 |
+
try:
|
| 93 |
+
from skills_catalog import ( # type: ignore
|
| 94 |
+
resolve_skills as _resolve_skills, format_skill_block as _format_skill_block,
|
| 95 |
+
skills_used_names as _skills_used_names)
|
| 96 |
+
except ImportError:
|
| 97 |
+
_resolve_skills = lambda task, user_id="", limit=2: [] # type: ignore
|
| 98 |
+
_format_skill_block = lambda skills: "" # type: ignore
|
| 99 |
+
_skills_used_names = lambda skills: [] # type: ignore
|
| 100 |
+
log.warning("[Kype] skills_catalog not available β skill inject disabled")
|
| 101 |
+
|
| 102 |
|
| 103 |
# =================================================================
|
| 104 |
# CONFIGURATION
|
|
|
|
| 209 |
3. RISKS: <bullet list or "none">
|
| 210 |
If DEBUGGER FEEDBACK is provided, address every flagged issue.
|
| 211 |
|
| 212 |
+
MEMORY & SKILLS: If "USER MEMORY (recent context)" is provided, treat it as
|
| 213 |
+
conversation history β the user may refer to earlier messages. If ACTIVE
|
| 214 |
+
SKILLS are provided, build the plan around following them exactly.
|
| 215 |
+
|
| 216 |
FORMAT: itineraries β markdown tables; code β fenced blocks; comparisons β tables. Stay on task.
|
| 217 |
{KYPE_RULES_PROMPT}"""
|
| 218 |
|
|
|
|
| 222 |
Execute the plan and deliver the full answer to the user's task.
|
| 223 |
No meta-commentary about the plan or your role β just the result.
|
| 224 |
If DEBUGGER FEEDBACK is provided, address each issue explicitly.
|
| 225 |
+
If USER MEMORY is provided, use it as conversation context.
|
| 226 |
+
If ACTIVE SKILLS are provided, follow their procedures EXACTLY in the answer
|
| 227 |
+
(the itinerary skill REQUIRES a markdown table, etc).
|
| 228 |
|
| 229 |
FORMAT: itineraries MUST use a markdown table (Day|Morning|Afternoon|Evening|Notes);
|
| 230 |
code MUST use fenced blocks with language tags; comparisons use markdown tables.
|
|
|
|
| 267 |
total_elapsed: float = 0.0
|
| 268 |
fallback_used: Optional[str] = None
|
| 269 |
tool_used: Optional[str] = None
|
| 270 |
+
skills_used: List[str] = field(default_factory=list)
|
| 271 |
usage: Dict[str, int] = field(default_factory=lambda: {"input_tokens": 0, "output_tokens": 0, "total_tokens": 0})
|
| 272 |
|
| 273 |
def to_dict(self) -> Dict[str, Any]:
|
|
|
|
| 278 |
"approved": self.approved,
|
| 279 |
"total_elapsed": round(self.total_elapsed, 2),
|
| 280 |
"fallback_used": self.fallback_used,
|
| 281 |
+
"skills_used": self.skills_used,
|
| 282 |
"tool_used": self.tool_used,
|
| 283 |
"usage": self.usage,
|
| 284 |
"trace": [
|
|
|
|
| 518 |
debugger_model: str = DEFAULT_DEBUGGER_MODEL,
|
| 519 |
max_iterations: int = MAX_ITERATIONS,
|
| 520 |
allow_tools: bool = True,
|
| 521 |
+
user_id: str = "",
|
| 522 |
):
|
| 523 |
+
self.user_id = (user_id or "").strip()
|
| 524 |
self.api_key = api_key or os.getenv("OPENROUTE_KEY", "")
|
| 525 |
if not self.api_key and not GROQ_API_KEY:
|
| 526 |
raise EnvironmentError(
|
|
|
|
| 849 |
# ---- stage wrappers ---------------------------------------------------
|
| 850 |
def plan(self, task: str, feedback: str = "") -> Tuple[str, str, str, str, dict]:
|
| 851 |
"""Returns (plan, tool_name, tool_result, provider, usage)."""
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 852 |
user = ""
|
| 853 |
+
mem_ctx = getattr(self, "_mem_context", "")
|
| 854 |
+
if mem_ctx:
|
| 855 |
+
user += f"USER MEMORY (recent context):\n{mem_ctx}\n\n"
|
| 856 |
+
skill_block = getattr(self, "_skill_block", "")
|
| 857 |
+
if skill_block:
|
| 858 |
+
user += skill_block + "\n\n"
|
| 859 |
user += f"TASK:\n{task}\n"
|
| 860 |
if feedback:
|
| 861 |
user += f"\nDEBUGGER FEEDBACK (address every issue):\n{feedback}\n"
|
|
|
|
| 895 |
|
| 896 |
def analyze(self, task: str, plan: str, feedback: str = "",
|
| 897 |
tool_name: str = "", tool_result: str = "") -> Tuple[str, str, dict]:
|
| 898 |
+
user = ""
|
| 899 |
+
mem_ctx = getattr(self, "_mem_context", "")
|
| 900 |
+
if mem_ctx:
|
| 901 |
+
user += f"USER MEMORY (recent context):\n{mem_ctx}\n\n"
|
| 902 |
+
skill_block = getattr(self, "_skill_block", "")
|
| 903 |
+
if skill_block:
|
| 904 |
+
user += skill_block + "\n\n"
|
| 905 |
+
user += f"TASK:\n{task}\n\nPLAN:\n{plan}\n"
|
| 906 |
if tool_name and tool_result:
|
| 907 |
user += f"\n[TOOL RESULT from {tool_name}]:\n{tool_result}\n"
|
| 908 |
if feedback:
|
|
|
|
| 989 |
if not task or not task.strip():
|
| 990 |
raise ValueError("task must be a non-empty string")
|
| 991 |
|
| 992 |
+
# v1.7: compact memory (profile + last 5 messages β no older history,
|
| 993 |
+
# keeps the prompt cheap) + auto-matched skills, built once per run.
|
| 994 |
+
try:
|
| 995 |
+
self._mem_context = build_memory_context(self.user_id, query="") or ""
|
| 996 |
+
except Exception:
|
| 997 |
+
self._mem_context = ""
|
| 998 |
+
try:
|
| 999 |
+
self._skills = _resolve_skills(task, user_id=self.user_id, limit=2)
|
| 1000 |
+
except Exception:
|
| 1001 |
+
self._skills = []
|
| 1002 |
+
try:
|
| 1003 |
+
self._skill_block = _format_skill_block(self._skills)
|
| 1004 |
+
except Exception:
|
| 1005 |
+
self._skill_block = ""
|
| 1006 |
+
|
| 1007 |
t_start = time.time()
|
| 1008 |
trace: List[AgentResponse] = []
|
| 1009 |
fallback_used: Optional[str] = None
|
|
|
|
| 1080 |
verdict.get("issues", []), ensure_ascii=False
|
| 1081 |
)
|
| 1082 |
|
| 1083 |
+
try:
|
| 1084 |
+
_used = _skills_used_names(self._skills)
|
| 1085 |
+
except Exception:
|
| 1086 |
+
_used = []
|
| 1087 |
return KypeResult(
|
| 1088 |
task=task,
|
| 1089 |
+
skills_used=_used,
|
| 1090 |
final_answer=draft,
|
| 1091 |
iterations=iterations_completed,
|
| 1092 |
approved=approved,
|
|
|
|
| 1512 |
debugger_model=req.debugger_model or DEFAULT_DEBUGGER_MODEL,
|
| 1513 |
max_iterations=req.max_iterations,
|
| 1514 |
allow_tools=req.allow_tools,
|
| 1515 |
+
user_id=user_id,
|
| 1516 |
)
|
| 1517 |
result = await asyncio.to_thread(agent.run, task)
|
| 1518 |
except EnvironmentError as e:
|
|
|
|
| 1538 |
except Exception as e:
|
| 1539 |
log.warning(f"[Kype] Failed to record token usage: {e}")
|
| 1540 |
|
| 1541 |
+
# v1.7: remember this exchange (compact β save_memory caps at 5 messages).
|
| 1542 |
+
# Runs BEFORE ad injection so ad text never enters memory.
|
| 1543 |
+
try:
|
| 1544 |
+
if user_id and result.final_answer:
|
| 1545 |
+
save_memory(user_id, task, result.final_answer)
|
| 1546 |
+
except Exception as e:
|
| 1547 |
+
log.warning(f"[Kype] save_memory failed: {e}")
|
| 1548 |
+
|
| 1549 |
# ββ Inject ad into final answer (~30% chance) ββ
|
| 1550 |
try:
|
| 1551 |
ad = maybe_get_ad(user_id=user_id, user_tier="pro" if polar_subscribed else "free", source="kype")
|
|
|
|
| 1573 |
return {
|
| 1574 |
"status": "maintenance" if maintenance else "ok",
|
| 1575 |
"service": "kype",
|
| 1576 |
+
"version": "1.7-skills", # v1.7 β memory + full skill inject (both stages) + ask-when-unsure
|
| 1577 |
"roles": ["planner", "analysis", "debugger"],
|
| 1578 |
"max_file_bytes": MAX_FILE_BYTES,
|
| 1579 |
"rate_window_seconds": KYPE_RATE_WINDOW_SECONDS,
|
agent/kype_rules.py
CHANGED
|
@@ -26,16 +26,26 @@ from typing import Tuple
|
|
| 26 |
# =================================================================
|
| 27 |
|
| 28 |
KYPE_RULES_PROMPT = """SCOPE: You are a task-execution agent, not a chat companion.
|
| 29 |
-
Stay
|
| 30 |
-
|
| 31 |
-
|
| 32 |
-
|
| 33 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 34 |
|
| 35 |
FORMAT RULES:
|
| 36 |
- Itineraries / travel plans β markdown TABLE (Day|Morning|Afternoon|Evening|Notes).
|
| 37 |
- Code β fenced markdown code blocks with language tags.
|
| 38 |
- Comparisons β markdown comparison tables.
|
|
|
|
| 39 |
- Stay on task; no filler small-talk."""
|
| 40 |
|
| 41 |
|
|
@@ -73,16 +83,14 @@ _BLOCK_PATTERNS: dict[str, list[str]] = {
|
|
| 73 |
r"\b(write|create|build) (a |an )?(virus|ransomware|keylogger|trojan)\b",
|
| 74 |
r"\bmalicious (code|script|payload)\b",
|
| 75 |
r"\bddos (attack|script|tool)\b",
|
| 76 |
-
r"\
|
| 77 |
-
r"\bsql injection (how to|attack|payload)\b",
|
| 78 |
r"\b(write|create) (an? )?(attack|weaponized) (script|payload)\b",
|
| 79 |
],
|
| 80 |
"professional_advice": [
|
| 81 |
-
r"\bshould i (invest|buy stock|take out a loan)\b",
|
| 82 |
r"\bdiagnose (my|this) (symptom|illness|condition)\b",
|
| 83 |
r"\bwhat (medication|dosage) should i take\b",
|
| 84 |
-
r"\
|
| 85 |
-
r"\bshould i
|
| 86 |
],
|
| 87 |
}
|
| 88 |
|
|
|
|
| 26 |
# =================================================================
|
| 27 |
|
| 28 |
KYPE_RULES_PROMPT = """SCOPE: You are a task-execution agent, not a chat companion.
|
| 29 |
+
Stay on the user's task. Do not chit-chat, roleplay, or answer unrelated
|
| 30 |
+
questions.
|
| 31 |
+
|
| 32 |
+
IF UNSURE, ASK: if the task is ambiguous or missing a critical detail
|
| 33 |
+
(format, language, target, data), do NOT guess β return ONE short
|
| 34 |
+
clarifying question as your output instead of a guessed deliverable.
|
| 35 |
+
Never invent facts to fill gaps.
|
| 36 |
+
|
| 37 |
+
ADVICE TOPICS (medical/legal/financial): explain general information
|
| 38 |
+
helpfully, then add one short line: "For your specific situation, confirm
|
| 39 |
+
with a qualified professional." These are NOT refusals.
|
| 40 |
+
|
| 41 |
+
HARD REFUSALS (brief, no how-to): violence/weapons instructions, illegal
|
| 42 |
+
activity, malware or attack code, explicit sexual content, credential dumps.
|
| 43 |
|
| 44 |
FORMAT RULES:
|
| 45 |
- Itineraries / travel plans β markdown TABLE (Day|Morning|Afternoon|Evening|Notes).
|
| 46 |
- Code β fenced markdown code blocks with language tags.
|
| 47 |
- Comparisons β markdown comparison tables.
|
| 48 |
+
- If ACTIVE SKILLS are provided for this task, follow them exactly.
|
| 49 |
- Stay on task; no filler small-talk."""
|
| 50 |
|
| 51 |
|
|
|
|
| 83 |
r"\b(write|create|build) (a |an )?(virus|ransomware|keylogger|trojan)\b",
|
| 84 |
r"\bmalicious (code|script|payload)\b",
|
| 85 |
r"\bddos (attack|script|tool)\b",
|
| 86 |
+
r"\b(write|create|build|generate) .{0,25}sql injection (payload|exploit)\b",
|
|
|
|
| 87 |
r"\b(write|create) (an? )?(attack|weaponized) (script|payload)\b",
|
| 88 |
],
|
| 89 |
"professional_advice": [
|
|
|
|
| 90 |
r"\bdiagnose (my|this) (symptom|illness|condition)\b",
|
| 91 |
r"\bwhat (medication|dosage) should i take\b",
|
| 92 |
+
r"\bhow much \d+\s*(mg|ml)\b|"
|
| 93 |
+
r"\bhow much\b.{0,30}\bshould i (take|inject)\b",
|
| 94 |
],
|
| 95 |
}
|
| 96 |
|
agent/prompts/system_prompt.py
CHANGED
|
@@ -1,14 +1,15 @@
|
|
| 1 |
"""
|
| 2 |
-
NeuraPrompt Agent β System Prompt Module (v9.
|
| 3 |
=================================================================================
|
| 4 |
|
| 5 |
-
Changes over
|
| 6 |
-
*
|
| 7 |
-
|
| 8 |
-
*
|
| 9 |
-
|
| 10 |
-
|
| 11 |
-
|
|
|
|
| 12 |
"""
|
| 13 |
|
| 14 |
def get_system_prompt(memory_context: str = "", tool_descriptions: str = "") -> str:
|
|
@@ -40,12 +41,18 @@ If the user message includes "Context:" with file content, that content was uplo
|
|
| 40 |
skill_learning = """
|
| 41 |
|
| 42 |
βββ SKILL LEARNING & USAGE
|
| 43 |
-
Skills are procedures. Memory may list "
|
| 44 |
-
-
|
| 45 |
-
-
|
|
|
|
| 46 |
- Prefer `list_skills` when unsure which skill applies, then `load_skill` for the full procedure.
|
| 47 |
- Pay attention to user preferences (languages, frameworks, style, timezone, project). The system may save them silently β do NOT announce skill saving.
|
| 48 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 49 |
βββ OUTPUT FORMATTING (frontend renders Markdown)
|
| 50 |
- Itineraries / travel plans β markdown TABLES (Day | Morning | Afternoon | Evening | Notes).
|
| 51 |
- Comparisons β markdown comparison TABLES.
|
|
@@ -53,12 +60,17 @@ Skills are procedures. Memory may list "Active skills (follow these):" and/or "L
|
|
| 53 |
- Docs / API explain β structured markdown with headings + bullets; cite snippets; do not invent APIs.
|
| 54 |
- Reports / outlines β ## headings and bullets.
|
| 55 |
|
| 56 |
-
βββ
|
| 57 |
-
|
| 58 |
-
|
| 59 |
-
illegal activity, or
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 60 |
Never dump raw tool JSON, tool names, or internal parameters in final_answer.
|
| 61 |
-
Do not write attack/exploit/malware code. Do not roleplay around
|
| 62 |
|
| 63 |
format_rules = """
|
| 64 |
|
|
|
|
| 1 |
"""
|
| 2 |
+
NeuraPrompt Agent β System Prompt Module (v9.4 β Skills + Ask-when-unsure)
|
| 3 |
=================================================================================
|
| 4 |
|
| 5 |
+
Changes over v9.3:
|
| 6 |
+
* Skills now AUTO-INJECT as "ACTIVE SKILLS" β the model must mention the
|
| 7 |
+
skill it is using in its first thought (visible in the thinking stream).
|
| 8 |
+
* New IF UNSURE, ASK rule β ask one short clarifying question instead of
|
| 9 |
+
guessing or refusing.
|
| 10 |
+
* STRICT MESSAGING relaxed: hard bans stay (weapons, malware, CSAM, illegal,
|
| 11 |
+
secrets), but medical/legal/financial topics are now answered as general
|
| 12 |
+
information with a brief professional-consult note instead of refused.
|
| 13 |
"""
|
| 14 |
|
| 15 |
def get_system_prompt(memory_context: str = "", tool_descriptions: str = "") -> str:
|
|
|
|
| 41 |
skill_learning = """
|
| 42 |
|
| 43 |
βββ SKILL LEARNING & USAGE
|
| 44 |
+
Skills are procedures. Memory may list "ACTIVE SKILLS β follow these procedures exactly" and/or "Learned skills".
|
| 45 |
+
- ACTIVE SKILLS listed for this goal were auto-matched to the user's request. FOLLOW their bodies exactly in your final_answer.
|
| 46 |
+
- IMPORTANT β SKILL TRANSPARENCY: when ACTIVE SKILLS are listed, your FIRST thought must state which skill(s) you are using and why, e.g. "Using itinerary skill (plan_itinerary) for this task β it requires a markdown table."
|
| 47 |
+
- If a Learned skill matches and is NOT already injected, call `load_skill` with the skill name.
|
| 48 |
- Prefer `list_skills` when unsure which skill applies, then `load_skill` for the full procedure.
|
| 49 |
- Pay attention to user preferences (languages, frameworks, style, timezone, project). The system may save them silently β do NOT announce skill saving.
|
| 50 |
|
| 51 |
+
βββ IF UNSURE, ASK
|
| 52 |
+
- If the goal is ambiguous or missing a critical detail (format, language, target, data), do NOT guess.
|
| 53 |
+
- Ask ONE short clarifying question as your final answer, then stop. Example: "Which format do you want the report in β PDF or Word?"
|
| 54 |
+
- Never refuse a legitimate task just because information is missing. Never invent facts, URLs, APIs, or data to fill gaps.
|
| 55 |
+
|
| 56 |
βββ OUTPUT FORMATTING (frontend renders Markdown)
|
| 57 |
- Itineraries / travel plans β markdown TABLES (Day | Morning | Afternoon | Evening | Notes).
|
| 58 |
- Comparisons β markdown comparison TABLES.
|
|
|
|
| 60 |
- Docs / API explain β structured markdown with headings + bullets; cite snippets; do not invent APIs.
|
| 61 |
- Reports / outlines β ## headings and bullets.
|
| 62 |
|
| 63 |
+
βββ SAFETY BOUNDARIES (precise, not paranoid)
|
| 64 |
+
HARD REFUSALS β refuse briefly, no tools, no elaboration:
|
| 65 |
+
weapons/explosives instructions, malware or attack code, explicit sexual content
|
| 66 |
+
(especially anything involving minors), illegal activity, or credential/key dumps.
|
| 67 |
+
HELP FREELY β these are NOT refusals:
|
| 68 |
+
- Medical, legal, or financial TOPICS: explain general information helpfully, then add
|
| 69 |
+
one short line: "For your specific situation, confirm with a qualified professional."
|
| 70 |
+
- Security topics: defensive hardening, how attacks work conceptually, safe practices.
|
| 71 |
+
- Anything the user's own systems, accounts, or code β help them.
|
| 72 |
Never dump raw tool JSON, tool names, or internal parameters in final_answer.
|
| 73 |
+
Do not write attack/exploit/malware code. Do not roleplay around the hard refusals."""
|
| 74 |
|
| 75 |
format_rules = """
|
| 76 |
|
agent/safety_gate.py
CHANGED
|
@@ -19,7 +19,7 @@ _GATE_RULES: list[tuple[str, str, str]] = [
|
|
| 19 |
"secrets",
|
| 20 |
r"(-----BEGIN (?:RSA |EC |OPENSSH |DSA )?PRIVATE KEY-----|"
|
| 21 |
r"\b(?:sk-[A-Za-z0-9_-]{20,}|sk-or-v1-[A-Za-z0-9_-]{20,}|gsk_[A-Za-z0-9_-]{20,})\b|"
|
| 22 |
-
r"\b(?:aws_secret_access_key|api[_-]?key|secret[_-]?key)\s*[:=]\s*\
|
| 23 |
r"\bAKIA[0-9A-Z]{16}\b)",
|
| 24 |
"I can't continue β please never share private keys, API secrets, or credentials. "
|
| 25 |
"Rotate anything you pasted and try again without secrets.",
|
|
@@ -28,7 +28,8 @@ _GATE_RULES: list[tuple[str, str, str]] = [
|
|
| 28 |
"bank_details",
|
| 29 |
r"\b(?:account\s*(?:number|no\.?)|iban|swift\s*code|routing\s*number|"
|
| 30 |
r"sort\s*code)\s*[:=#]?\s*[A-Z0-9\s-]{8,}|"
|
| 31 |
-
r"\b\d{4}[-\s]?\d{4}[-\s]?\d{4}[-\s]?\d{4}\b"
|
|
|
|
| 32 |
"I can't process bank or payment card details. Remove them and continue without financial numbers.",
|
| 33 |
),
|
| 34 |
(
|
|
@@ -48,10 +49,11 @@ _GATE_RULES: list[tuple[str, str, str]] = [
|
|
| 48 |
# Malware / offensive cyber
|
| 49 |
(
|
| 50 |
"malware",
|
| 51 |
-
r"\b(?:write|create|build|generate)\s+(?:
|
| 52 |
r"\b(?:malicious\s+(?:code|script|payload)|ddos\s+(?:attack|script|tool))\b|"
|
| 53 |
r"\b(?:sql\s*injection\s+how\s*to|exploit\s+(?:code|poc)\s+for)\b|"
|
| 54 |
-
r"\bhow to (?:hack|break into|
|
|
|
|
| 55 |
"I can't help with malware, exploits, or attack tooling.",
|
| 56 |
),
|
| 57 |
# Explicit / CSAM
|
|
@@ -73,7 +75,7 @@ _GATE_RULES: list[tuple[str, str, str]] = [
|
|
| 73 |
(
|
| 74 |
"legal_financial",
|
| 75 |
r"\b(?:should i sue|is this contract legally binding|"
|
| 76 |
-
r"
|
| 77 |
r"guaranteed\s+(?:returns|profit))\b",
|
| 78 |
"I can't give legal or financial advice. Please consult a qualified professional.",
|
| 79 |
),
|
|
|
|
| 19 |
"secrets",
|
| 20 |
r"(-----BEGIN (?:RSA |EC |OPENSSH |DSA )?PRIVATE KEY-----|"
|
| 21 |
r"\b(?:sk-[A-Za-z0-9_-]{20,}|sk-or-v1-[A-Za-z0-9_-]{20,}|gsk_[A-Za-z0-9_-]{20,})\b|"
|
| 22 |
+
r"\b(?:aws_secret_access_key|api[_-]?key|secret[_-]?key)\s*[:=]\s*[\"\x27]?[A-Za-z0-9_-]{25,}|"
|
| 23 |
r"\bAKIA[0-9A-Z]{16}\b)",
|
| 24 |
"I can't continue β please never share private keys, API secrets, or credentials. "
|
| 25 |
"Rotate anything you pasted and try again without secrets.",
|
|
|
|
| 28 |
"bank_details",
|
| 29 |
r"\b(?:account\s*(?:number|no\.?)|iban|swift\s*code|routing\s*number|"
|
| 30 |
r"sort\s*code)\s*[:=#]?\s*[A-Z0-9\s-]{8,}|"
|
| 31 |
+
r"\b(?:card|credit\s*card|debit\s*card|visa|mastercard)\b[^.]{0,40}\b\d{4}[-\s]?\d{4}[-\s]?\d{4}[-\s]?\d{4}\b|"
|
| 32 |
+
r"\b\d{4}[-\s]?\d{4}[-\s]?\d{4}[-\s]?\d{4}\b\s*(?:card|exp|expiry|cvv)\b",
|
| 33 |
"I can't process bank or payment card details. Remove them and continue without financial numbers.",
|
| 34 |
),
|
| 35 |
(
|
|
|
|
| 49 |
# Malware / offensive cyber
|
| 50 |
(
|
| 51 |
"malware",
|
| 52 |
+
r"\b(?:write|create|build|generate)\s+(?:\w+\s+){0,2}(?:virus|ransomware|keylogger|trojan|rootkit|botnet)\b|"
|
| 53 |
r"\b(?:malicious\s+(?:code|script|payload)|ddos\s+(?:attack|script|tool))\b|"
|
| 54 |
r"\b(?:sql\s*injection\s+how\s*to|exploit\s+(?:code|poc)\s+for)\b|"
|
| 55 |
+
r"\bhow to (?:hack|break) into\s+(?:someone|someone\x27?s|an?other|other people|his|her|their|a\s+(?:company|school|bank|email|phone|account|system|computer|wifi))\b|"
|
| 56 |
+
r"\bhow to (?:hack|steal|bypass)\s+(?:someone\x27?s|another|other|his|her|their)\s+(?:account|email|password|phone|wifi|system|computer)\b",
|
| 57 |
"I can't help with malware, exploits, or attack tooling.",
|
| 58 |
),
|
| 59 |
# Explicit / CSAM
|
|
|
|
| 75 |
(
|
| 76 |
"legal_financial",
|
| 77 |
r"\b(?:should i sue|is this contract legally binding|"
|
| 78 |
+
r"guarantee\s+\d+%?\s+(?:returns|profit)|"
|
| 79 |
r"guaranteed\s+(?:returns|profit))\b",
|
| 80 |
"I can't give legal or financial advice. Please consult a qualified professional.",
|
| 81 |
),
|
agent/skills_catalog.py
CHANGED
|
@@ -12,6 +12,39 @@ from typing import Any, Dict, List, Optional
|
|
| 12 |
|
| 13 |
|
| 14 |
BUILTIN_SKILLS: List[Dict[str, str]] = [
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 15 |
{
|
| 16 |
"id": "read_docs",
|
| 17 |
"name": "read_docs",
|
|
@@ -122,14 +155,21 @@ def _score_skill(skill: Dict[str, str], query: str) -> int:
|
|
| 122 |
if name and name in q:
|
| 123 |
score += 6
|
| 124 |
|
| 125 |
-
# Token overlap with when_to_use
|
| 126 |
-
|
| 127 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 128 |
overlap = when_tokens & q_tokens
|
| 129 |
score += min(10, len(overlap) * 2)
|
| 130 |
|
| 131 |
# Domain keyword hints
|
| 132 |
hints = {
|
|
|
|
|
|
|
| 133 |
"read_docs": ("docs", "documentation", "api", "endpoint", "sdk", "reference", "explain this"),
|
| 134 |
"format_code": ("format", "refactor", "rewrite", "clean up", "beautify", "lint", "code"),
|
| 135 |
"plan_itinerary": ("itinerary", "trip", "travel", "vacation", "day trip", "sightseeing", "packing"),
|
|
@@ -137,7 +177,9 @@ def _score_skill(skill: Dict[str, str], query: str) -> int:
|
|
| 137 |
"write_report_outline": ("report", "outline", "briefing", "write-up", "structured document"),
|
| 138 |
}
|
| 139 |
for hint in hints.get(name, ()):
|
| 140 |
-
|
|
|
|
|
|
|
| 141 |
score += 5
|
| 142 |
|
| 143 |
# Light body keyword hit (weak)
|
|
@@ -172,9 +214,79 @@ def format_matched_bodies(query: str, limit: int = 2) -> str:
|
|
| 172 |
return "\n\n".join(parts)
|
| 173 |
|
| 174 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 175 |
__all__ = [
|
| 176 |
"BUILTIN_SKILLS",
|
| 177 |
"get_builtin",
|
| 178 |
"match_builtins",
|
| 179 |
"format_matched_bodies",
|
|
|
|
|
|
|
|
|
|
|
|
|
| 180 |
]
|
|
|
|
| 12 |
|
| 13 |
|
| 14 |
BUILTIN_SKILLS: List[Dict[str, str]] = [
|
| 15 |
+
{
|
| 16 |
+
"id": "web_research",
|
| 17 |
+
"name": "web_research",
|
| 18 |
+
"when_to_use": (
|
| 19 |
+
"When the user needs current facts, news, prices, recent events, "
|
| 20 |
+
"or anything that requires searching the web."
|
| 21 |
+
),
|
| 22 |
+
"body": """## Skill: web_research
|
| 23 |
+
|
| 24 |
+
1. Search BEFORE answering β do not rely on memory for anything current
|
| 25 |
+
(news, prices, releases, schedules, scores).
|
| 26 |
+
2. Run at most 2-3 searches; refine the query once if the first is weak.
|
| 27 |
+
3. Fetch the most relevant source when details matter.
|
| 28 |
+
4. Synthesize the findings in your own words β cite the source name or URL
|
| 29 |
+
inline for key claims.
|
| 30 |
+
5. State clearly when sources disagree or info could not be verified.""",
|
| 31 |
+
},
|
| 32 |
+
{
|
| 33 |
+
"id": "create_document",
|
| 34 |
+
"name": "create_document",
|
| 35 |
+
"when_to_use": (
|
| 36 |
+
"When the user asks for a file, PDF, DOCX, spreadsheet, CSV, "
|
| 37 |
+
"report file, or says 'create/download/generate a document'."
|
| 38 |
+
),
|
| 39 |
+
"body": """## Skill: create_document
|
| 40 |
+
|
| 41 |
+
1. Confirm the deliverable type (PDF / DOCX / XLSX / CSV) and filename.
|
| 42 |
+
2. Draft the content fully FIRST, then create the file.
|
| 43 |
+
3. Call the correct file tool with a clear snake_case filename.
|
| 44 |
+
4. After creating, give the download link and a 1-2 line summary of contents.
|
| 45 |
+
5. If the request is ambiguous (e.g. 'make a report' with no format),
|
| 46 |
+
ask ONE short clarifying question before generating.""",
|
| 47 |
+
},
|
| 48 |
{
|
| 49 |
"id": "read_docs",
|
| 50 |
"name": "read_docs",
|
|
|
|
| 155 |
if name and name in q:
|
| 156 |
score += 6
|
| 157 |
|
| 158 |
+
# Token overlap with when_to_use (stopwords removed β common words like
|
| 159 |
+
# "for/the/when" caused false matches, e.g. create_document on news queries)
|
| 160 |
+
_STOP = {"the", "and", "for", "with", "you", "your", "when", "user",
|
| 161 |
+
"asks", "from", "about", "are", "this", "that", "have", "not",
|
| 162 |
+
"one", "any", "all", "who", "what", "how", "why", "should",
|
| 163 |
+
"their", "will", "can", "was", "out", "get", "need", "needs"}
|
| 164 |
+
when_tokens = {t for t in re.findall(r"[a-z0-9_]{3,}", when) if t not in _STOP}
|
| 165 |
+
q_tokens = {t for t in re.findall(r"[a-z0-9_]{3,}", q) if t not in _STOP}
|
| 166 |
overlap = when_tokens & q_tokens
|
| 167 |
score += min(10, len(overlap) * 2)
|
| 168 |
|
| 169 |
# Domain keyword hints
|
| 170 |
hints = {
|
| 171 |
+
"web_research": ("search", "latest", "news", "current", "today", "price", "recent", "who won", "update"),
|
| 172 |
+
"create_document": ("pdf", "docx", "word document", "excel", "spreadsheet", "csv", "file", "download", "document"),
|
| 173 |
"read_docs": ("docs", "documentation", "api", "endpoint", "sdk", "reference", "explain this"),
|
| 174 |
"format_code": ("format", "refactor", "rewrite", "clean up", "beautify", "lint", "code"),
|
| 175 |
"plan_itinerary": ("itinerary", "trip", "travel", "vacation", "day trip", "sightseeing", "packing"),
|
|
|
|
| 177 |
"write_report_outline": ("report", "outline", "briefing", "write-up", "structured document"),
|
| 178 |
}
|
| 179 |
for hint in hints.get(name, ()):
|
| 180 |
+
# Word-boundary match β plain substring matched "document" inside
|
| 181 |
+
# "documentation", wrongly triggering create_document for docs queries.
|
| 182 |
+
if re.search(rf"\b{re.escape(hint)}\b", q):
|
| 183 |
score += 5
|
| 184 |
|
| 185 |
# Light body keyword hit (weak)
|
|
|
|
| 214 |
return "\n\n".join(parts)
|
| 215 |
|
| 216 |
|
| 217 |
+
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 218 |
+
# UNIFIED RESOLVER (v2.0) β used by agent.py, kype.py, zyloo_agent.py
|
| 219 |
+
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 220 |
+
|
| 221 |
+
def resolve_skills(task: str, user_id: str = "", limit: int = 2) -> List[Dict[str, Any]]:
|
| 222 |
+
"""Match builtin + learned skills against a task (best first).
|
| 223 |
+
Returns list of dicts: {name, when_to_use, body, source}.
|
| 224 |
+
Prefers agent_memory.match_skills_for_query (covers both sources);
|
| 225 |
+
falls back to builtin-only matching when memory is unavailable."""
|
| 226 |
+
if not task or not task.strip():
|
| 227 |
+
return []
|
| 228 |
+
try:
|
| 229 |
+
from .agent_memory import match_skills_for_query
|
| 230 |
+
except ImportError:
|
| 231 |
+
try:
|
| 232 |
+
from agent_memory import match_skills_for_query # type: ignore
|
| 233 |
+
except ImportError:
|
| 234 |
+
match_skills_for_query = None # type: ignore
|
| 235 |
+
if match_skills_for_query is not None:
|
| 236 |
+
try:
|
| 237 |
+
matched = match_skills_for_query(user_id or "", task, limit=limit)
|
| 238 |
+
if matched:
|
| 239 |
+
return matched[:limit]
|
| 240 |
+
except Exception:
|
| 241 |
+
pass
|
| 242 |
+
return match_builtins(task, limit=limit)
|
| 243 |
+
|
| 244 |
+
|
| 245 |
+
def format_skill_block(skills: List[Dict[str, Any]]) -> str:
|
| 246 |
+
"""Prompt-injection block for matched skills ("" if none)."""
|
| 247 |
+
if not skills:
|
| 248 |
+
return ""
|
| 249 |
+
lines = ["ACTIVE SKILLS β follow these procedures exactly for this task:"]
|
| 250 |
+
for sk in skills:
|
| 251 |
+
lines.append(f"### {sk.get('name', '?')}")
|
| 252 |
+
if sk.get("when_to_use"):
|
| 253 |
+
lines.append(f"**When to use:** {sk['when_to_use']}")
|
| 254 |
+
lines.append(sk.get("body") or "")
|
| 255 |
+
return "\n".join(lines)
|
| 256 |
+
|
| 257 |
+
|
| 258 |
+
def skill_thinking_line(skills: List[Dict[str, Any]]) -> str:
|
| 259 |
+
"""One-liner for the thinking/reasoning stream, e.g.
|
| 260 |
+
'π Using doc skill (read_docs) + web research skill (web_research) for this task'."""
|
| 261 |
+
if not skills:
|
| 262 |
+
return ""
|
| 263 |
+
friendly = {
|
| 264 |
+
"read_docs": "doc skill", "format_code": "code skill",
|
| 265 |
+
"plan_itinerary": "itinerary skill", "compare_options_table": "comparison skill",
|
| 266 |
+
"write_report_outline": "report skill", "web_research": "web research skill",
|
| 267 |
+
"create_document": "document skill",
|
| 268 |
+
}
|
| 269 |
+
parts = []
|
| 270 |
+
for sk in skills:
|
| 271 |
+
name = sk.get("name", "?")
|
| 272 |
+
parts.append(f"{friendly.get(name, name.replace('_', ' '))} ({name})")
|
| 273 |
+
joined = " + ".join(parts)
|
| 274 |
+
return f"π Using {joined} for this task"
|
| 275 |
+
|
| 276 |
+
|
| 277 |
+
def skills_used_names(skills: List[Dict[str, Any]]) -> List[str]:
|
| 278 |
+
"""Plain skill names for the result API (skills_used field)."""
|
| 279 |
+
return [sk.get("name", "") for sk in skills if sk.get("name")]
|
| 280 |
+
|
| 281 |
+
|
| 282 |
+
|
| 283 |
__all__ = [
|
| 284 |
"BUILTIN_SKILLS",
|
| 285 |
"get_builtin",
|
| 286 |
"match_builtins",
|
| 287 |
"format_matched_bodies",
|
| 288 |
+
"resolve_skills",
|
| 289 |
+
"format_skill_block",
|
| 290 |
+
"skill_thinking_line",
|
| 291 |
+
"skills_used_names",
|
| 292 |
]
|
agent/zyloo_agent.py
CHANGED
|
@@ -57,6 +57,32 @@ from pydantic import BaseModel
|
|
| 57 |
|
| 58 |
log = logging.getLogger("zyloo_agent")
|
| 59 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 60 |
|
| 61 |
# =================================================================
|
| 62 |
# CONFIGURATION
|
|
@@ -164,7 +190,15 @@ Output format (strict -- no preamble):
|
|
| 164 |
- Step 1: ...
|
| 165 |
- Step 2: ...
|
| 166 |
3. RISKS: <bullet list or "none">
|
| 167 |
-
If DEBUGGER FEEDBACK is provided, address every flagged issue.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 168 |
|
| 169 |
ANALYSIS_SYSTEM = f"""You are the ANALYSIS of ZyKype Agent.
|
| 170 |
{ZYLOO_IDENTITY}
|
|
@@ -182,6 +216,13 @@ STRICT RULES:
|
|
| 182 |
|
| 183 |
Return only the user-facing response.
|
| 184 |
If DEBUGGER FEEDBACK is provided, address each issue explicitly.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 185 |
"""
|
| 186 |
|
| 187 |
DEBUGGER_SYSTEM = f"""You are the DEBUGGER of ZyKype Agent.
|
|
@@ -219,6 +260,7 @@ class ZylooResult:
|
|
| 219 |
total_elapsed: float = 0.0
|
| 220 |
fallback_used: Optional[str] = None
|
| 221 |
tool_used: Optional[str] = None
|
|
|
|
| 222 |
usage: Dict[str, int] = field(default_factory=lambda: {"input_tokens": 0, "output_tokens": 0, "total_tokens": 0})
|
| 223 |
|
| 224 |
def to_dict(self) -> Dict[str, Any]:
|
|
@@ -229,6 +271,7 @@ class ZylooResult:
|
|
| 229 |
"approved": self.approved,
|
| 230 |
"total_elapsed": round(self.total_elapsed, 2),
|
| 231 |
"fallback_used": self.fallback_used,
|
|
|
|
| 232 |
"tool_used": self.tool_used,
|
| 233 |
"usage": self.usage,
|
| 234 |
"trace": [
|
|
@@ -439,7 +482,9 @@ class ZylooAgent:
|
|
| 439 |
debugger_model: str = DEFAULT_DEBUGGER_MODEL,
|
| 440 |
max_iterations: int = MAX_ITERATIONS,
|
| 441 |
allow_tools: bool = True,
|
|
|
|
| 442 |
):
|
|
|
|
| 443 |
self.api_key = api_key or ZYLOO_API_KEY
|
| 444 |
if not self.api_key and not GROQ_API_KEY:
|
| 445 |
raise EnvironmentError(
|
|
@@ -801,7 +846,14 @@ class ZylooAgent:
|
|
| 801 |
# ---- stage wrappers -- identical to Kype (with zyloo_doc_tools hook) ----
|
| 802 |
def plan(self, task: str, feedback: str = "") -> Tuple[str, str, str, str, dict]:
|
| 803 |
"""Returns (plan, tool_name, tool_result, provider, usage)."""
|
| 804 |
-
user =
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 805 |
if feedback:
|
| 806 |
user += f"\nDEBUGGER FEEDBACK (address every issue):\n{feedback}\n"
|
| 807 |
if self.allow_tools:
|
|
@@ -839,7 +891,14 @@ class ZylooAgent:
|
|
| 839 |
|
| 840 |
def analyze(self, task: str, plan: str, feedback: str = "",
|
| 841 |
tool_name: str = "", tool_result: str = "") -> Tuple[str, str, dict]:
|
| 842 |
-
user =
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 843 |
if tool_name and tool_result:
|
| 844 |
user += f"\n[TOOL RESULT from {tool_name}]:\n{tool_result}\n"
|
| 845 |
if feedback:
|
|
@@ -919,6 +978,21 @@ class ZylooAgent:
|
|
| 919 |
if not task or not task.strip():
|
| 920 |
raise ValueError("task must be a non-empty string")
|
| 921 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 922 |
t_start = time.time()
|
| 923 |
trace: List[AgentResponse] = []
|
| 924 |
fallback_used: Optional[str] = None
|
|
@@ -992,8 +1066,13 @@ class ZylooAgent:
|
|
| 992 |
verdict.get("issues", []), ensure_ascii=False
|
| 993 |
)
|
| 994 |
|
|
|
|
|
|
|
|
|
|
|
|
|
| 995 |
return ZylooResult(
|
| 996 |
task=task,
|
|
|
|
| 997 |
final_answer=_clean_final_answer(draft),
|
| 998 |
iterations=iterations_completed,
|
| 999 |
approved=approved,
|
|
@@ -1187,7 +1266,7 @@ async def zyloo_run(req: ZylooRequest):
|
|
| 1187 |
subscription. Uses the FIXED polar_subscription module which fails
|
| 1188 |
CLOSED on auth errors (401/403).
|
| 1189 |
2. Rate limit -- 1 task per 30s per user (returns 429)
|
| 1190 |
-
3. Token budget -- Pro 1M/month, Free
|
| 1191 |
4. Context cap -- 5000 tokens max per task (returns 413)
|
| 1192 |
|
| 1193 |
On Zyloo failure: automatic Groq fallback (same prompts, reduced tokens).
|
|
@@ -1347,6 +1426,7 @@ async def zyloo_run(req: ZylooRequest):
|
|
| 1347 |
debugger_model=req.debugger_model or DEFAULT_DEBUGGER_MODEL,
|
| 1348 |
max_iterations=req.max_iterations,
|
| 1349 |
allow_tools=req.allow_tools,
|
|
|
|
| 1350 |
)
|
| 1351 |
result = await asyncio.to_thread(agent.run, task)
|
| 1352 |
except EnvironmentError as e:
|
|
@@ -1372,6 +1452,13 @@ async def zyloo_run(req: ZylooRequest):
|
|
| 1372 |
except Exception as e:
|
| 1373 |
log.warning(f"[Zyloo] Failed to record token usage: {e}")
|
| 1374 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1375 |
out = result.to_dict()
|
| 1376 |
if req.privacy:
|
| 1377 |
out = _apply_privacy(out)
|
|
@@ -1389,7 +1476,7 @@ async def zyloo_health():
|
|
| 1389 |
return {
|
| 1390 |
"status": "maintenance" if maintenance else "ok",
|
| 1391 |
"service": "zyloo_agent",
|
| 1392 |
-
"version": "1.
|
| 1393 |
"roles": ["planner", "analysis", "debugger"],
|
| 1394 |
"primary_provider": "zyloo",
|
| 1395 |
"zyloo_base_url": "https://api.zyloo.io/v1",
|
|
|
|
| 57 |
|
| 58 |
log = logging.getLogger("zyloo_agent")
|
| 59 |
|
| 60 |
+
# ---- v1.1: compact memory + unified skills resolver ----
|
| 61 |
+
try:
|
| 62 |
+
from .agent_memory import build_memory_context, save_memory
|
| 63 |
+
except ImportError:
|
| 64 |
+
try:
|
| 65 |
+
from agent_memory import build_memory_context, save_memory # type: ignore
|
| 66 |
+
except ImportError:
|
| 67 |
+
build_memory_context = lambda uid, query="": "" # type: ignore
|
| 68 |
+
save_memory = lambda uid, msg, resp: None # type: ignore
|
| 69 |
+
log.warning("[Zyloo] agent_memory not available β memory disabled")
|
| 70 |
+
|
| 71 |
+
try:
|
| 72 |
+
from .skills_catalog import (
|
| 73 |
+
resolve_skills as _resolve_skills, format_skill_block as _format_skill_block,
|
| 74 |
+
skills_used_names as _skills_used_names)
|
| 75 |
+
except ImportError:
|
| 76 |
+
try:
|
| 77 |
+
from skills_catalog import ( # type: ignore
|
| 78 |
+
resolve_skills as _resolve_skills, format_skill_block as _format_skill_block,
|
| 79 |
+
skills_used_names as _skills_used_names)
|
| 80 |
+
except ImportError:
|
| 81 |
+
_resolve_skills = lambda task, user_id="", limit=2: [] # type: ignore
|
| 82 |
+
_format_skill_block = lambda skills: "" # type: ignore
|
| 83 |
+
_skills_used_names = lambda skills: [] # type: ignore
|
| 84 |
+
log.warning("[Zyloo] skills_catalog not available β skill inject disabled")
|
| 85 |
+
|
| 86 |
|
| 87 |
# =================================================================
|
| 88 |
# CONFIGURATION
|
|
|
|
| 190 |
- Step 1: ...
|
| 191 |
- Step 2: ...
|
| 192 |
3. RISKS: <bullet list or "none">
|
| 193 |
+
If DEBUGGER FEEDBACK is provided, address every flagged issue.
|
| 194 |
+
|
| 195 |
+
IF UNSURE, ASK: if the task is ambiguous or missing a critical detail
|
| 196 |
+
(format, language, target, data), do NOT guess β plan around ONE short
|
| 197 |
+
clarifying question as the deliverable. Never invent facts to fill gaps.
|
| 198 |
+
|
| 199 |
+
MEMORY & SKILLS: If "USER MEMORY (recent context)" is provided, treat it as
|
| 200 |
+
conversation history β the user may refer to earlier messages. If ACTIVE
|
| 201 |
+
SKILLS are provided, build the plan around following them exactly."""
|
| 202 |
|
| 203 |
ANALYSIS_SYSTEM = f"""You are the ANALYSIS of ZyKype Agent.
|
| 204 |
{ZYLOO_IDENTITY}
|
|
|
|
| 216 |
|
| 217 |
Return only the user-facing response.
|
| 218 |
If DEBUGGER FEEDBACK is provided, address each issue explicitly.
|
| 219 |
+
|
| 220 |
+
IF UNSURE, ASK: if a critical detail is still ambiguous, return ONE short
|
| 221 |
+
clarifying question as the answer instead of guessing.
|
| 222 |
+
|
| 223 |
+
If USER MEMORY is provided, use it as conversation context.
|
| 224 |
+
If ACTIVE SKILLS are provided, follow their procedures EXACTLY in the
|
| 225 |
+
answer (the itinerary skill REQUIRES a markdown table, etc.).
|
| 226 |
"""
|
| 227 |
|
| 228 |
DEBUGGER_SYSTEM = f"""You are the DEBUGGER of ZyKype Agent.
|
|
|
|
| 260 |
total_elapsed: float = 0.0
|
| 261 |
fallback_used: Optional[str] = None
|
| 262 |
tool_used: Optional[str] = None
|
| 263 |
+
skills_used: List[str] = field(default_factory=list)
|
| 264 |
usage: Dict[str, int] = field(default_factory=lambda: {"input_tokens": 0, "output_tokens": 0, "total_tokens": 0})
|
| 265 |
|
| 266 |
def to_dict(self) -> Dict[str, Any]:
|
|
|
|
| 271 |
"approved": self.approved,
|
| 272 |
"total_elapsed": round(self.total_elapsed, 2),
|
| 273 |
"fallback_used": self.fallback_used,
|
| 274 |
+
"skills_used": self.skills_used,
|
| 275 |
"tool_used": self.tool_used,
|
| 276 |
"usage": self.usage,
|
| 277 |
"trace": [
|
|
|
|
| 482 |
debugger_model: str = DEFAULT_DEBUGGER_MODEL,
|
| 483 |
max_iterations: int = MAX_ITERATIONS,
|
| 484 |
allow_tools: bool = True,
|
| 485 |
+
user_id: str = "",
|
| 486 |
):
|
| 487 |
+
self.user_id = (user_id or "").strip()
|
| 488 |
self.api_key = api_key or ZYLOO_API_KEY
|
| 489 |
if not self.api_key and not GROQ_API_KEY:
|
| 490 |
raise EnvironmentError(
|
|
|
|
| 846 |
# ---- stage wrappers -- identical to Kype (with zyloo_doc_tools hook) ----
|
| 847 |
def plan(self, task: str, feedback: str = "") -> Tuple[str, str, str, str, dict]:
|
| 848 |
"""Returns (plan, tool_name, tool_result, provider, usage)."""
|
| 849 |
+
user = ""
|
| 850 |
+
mem_ctx = getattr(self, "_mem_context", "")
|
| 851 |
+
if mem_ctx:
|
| 852 |
+
user += f"USER MEMORY (recent context):\n{mem_ctx}\n\n"
|
| 853 |
+
skill_block = getattr(self, "_skill_block", "")
|
| 854 |
+
if skill_block:
|
| 855 |
+
user += skill_block + "\n\n"
|
| 856 |
+
user += f"TASK:\n{task}\n"
|
| 857 |
if feedback:
|
| 858 |
user += f"\nDEBUGGER FEEDBACK (address every issue):\n{feedback}\n"
|
| 859 |
if self.allow_tools:
|
|
|
|
| 891 |
|
| 892 |
def analyze(self, task: str, plan: str, feedback: str = "",
|
| 893 |
tool_name: str = "", tool_result: str = "") -> Tuple[str, str, dict]:
|
| 894 |
+
user = ""
|
| 895 |
+
mem_ctx = getattr(self, "_mem_context", "")
|
| 896 |
+
if mem_ctx:
|
| 897 |
+
user += f"USER MEMORY (recent context):\n{mem_ctx}\n\n"
|
| 898 |
+
skill_block = getattr(self, "_skill_block", "")
|
| 899 |
+
if skill_block:
|
| 900 |
+
user += skill_block + "\n\n"
|
| 901 |
+
user += f"TASK:\n{task}\n\nPLAN:\n{plan}\n"
|
| 902 |
if tool_name and tool_result:
|
| 903 |
user += f"\n[TOOL RESULT from {tool_name}]:\n{tool_result}\n"
|
| 904 |
if feedback:
|
|
|
|
| 978 |
if not task or not task.strip():
|
| 979 |
raise ValueError("task must be a non-empty string")
|
| 980 |
|
| 981 |
+
# v1.1: compact memory (profile + last 5 messages β no older history,
|
| 982 |
+
# keeps the prompt cheap) + auto-matched skills, built once per run.
|
| 983 |
+
try:
|
| 984 |
+
self._mem_context = build_memory_context(self.user_id, query="") or ""
|
| 985 |
+
except Exception:
|
| 986 |
+
self._mem_context = ""
|
| 987 |
+
try:
|
| 988 |
+
self._skills = _resolve_skills(task, user_id=self.user_id, limit=2)
|
| 989 |
+
except Exception:
|
| 990 |
+
self._skills = []
|
| 991 |
+
try:
|
| 992 |
+
self._skill_block = _format_skill_block(self._skills)
|
| 993 |
+
except Exception:
|
| 994 |
+
self._skill_block = ""
|
| 995 |
+
|
| 996 |
t_start = time.time()
|
| 997 |
trace: List[AgentResponse] = []
|
| 998 |
fallback_used: Optional[str] = None
|
|
|
|
| 1066 |
verdict.get("issues", []), ensure_ascii=False
|
| 1067 |
)
|
| 1068 |
|
| 1069 |
+
try:
|
| 1070 |
+
_used = _skills_used_names(self._skills)
|
| 1071 |
+
except Exception:
|
| 1072 |
+
_used = []
|
| 1073 |
return ZylooResult(
|
| 1074 |
task=task,
|
| 1075 |
+
skills_used=_used,
|
| 1076 |
final_answer=_clean_final_answer(draft),
|
| 1077 |
iterations=iterations_completed,
|
| 1078 |
approved=approved,
|
|
|
|
| 1266 |
subscription. Uses the FIXED polar_subscription module which fails
|
| 1267 |
CLOSED on auth errors (401/403).
|
| 1268 |
2. Rate limit -- 1 task per 30s per user (returns 429)
|
| 1269 |
+
3. Token budget -- Pro 1M/month, Max 1.5M/month, Free 200K/month + 5 daily limits (returns 402/429)
|
| 1270 |
4. Context cap -- 5000 tokens max per task (returns 413)
|
| 1271 |
|
| 1272 |
On Zyloo failure: automatic Groq fallback (same prompts, reduced tokens).
|
|
|
|
| 1426 |
debugger_model=req.debugger_model or DEFAULT_DEBUGGER_MODEL,
|
| 1427 |
max_iterations=req.max_iterations,
|
| 1428 |
allow_tools=req.allow_tools,
|
| 1429 |
+
user_id=user_id,
|
| 1430 |
)
|
| 1431 |
result = await asyncio.to_thread(agent.run, task)
|
| 1432 |
except EnvironmentError as e:
|
|
|
|
| 1452 |
except Exception as e:
|
| 1453 |
log.warning(f"[Zyloo] Failed to record token usage: {e}")
|
| 1454 |
|
| 1455 |
+
# v1.1: remember this exchange (compact β save_memory caps at 5 messages)
|
| 1456 |
+
try:
|
| 1457 |
+
if user_id and result.final_answer:
|
| 1458 |
+
save_memory(user_id, task, result.final_answer)
|
| 1459 |
+
except Exception as e:
|
| 1460 |
+
log.warning(f"[Zyloo] save_memory failed: {e}")
|
| 1461 |
+
|
| 1462 |
out = result.to_dict()
|
| 1463 |
if req.privacy:
|
| 1464 |
out = _apply_privacy(out)
|
|
|
|
| 1476 |
return {
|
| 1477 |
"status": "maintenance" if maintenance else "ok",
|
| 1478 |
"service": "zyloo_agent",
|
| 1479 |
+
"version": "1.1", # v1.1 β memory + skill inject (both stages) + ask-when-unsure
|
| 1480 |
"roles": ["planner", "analysis", "debugger"],
|
| 1481 |
"primary_provider": "zyloo",
|
| 1482 |
"zyloo_base_url": "https://api.zyloo.io/v1",
|