SigmaFlow 0.0.78__tar.gz → 0.0.80__tar.gz

{sigmaflow-0.0.78 → sigmaflow-0.0.80}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: SigmaFlow
-Version: 0.0.78
+Version: 0.0.80
 Summary: SigmaFlow is a Python package designed to optimize the performance of task-flow related to LLMs/MLLMs or Multi-agent
 Author-email: Kangkun Mao <maokangkun@pjlab.org.cn>
 License-Expression: MIT

{sigmaflow-0.0.78 → sigmaflow-0.0.80}/src/SigmaFlow.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: SigmaFlow
-Version: 0.0.78
+Version: 0.0.80
 Summary: SigmaFlow is a Python package designed to optimize the performance of task-flow related to LLMs/MLLMs or Multi-agent
 Author-email: Kangkun Mao <maokangkun@pjlab.org.cn>
 License-Expression: MIT

{sigmaflow-0.0.78 → sigmaflow-0.0.80}/src/sigmaflow/agent/agent.py

@@ -19,7 +19,6 @@ from .conf import *
 
 SYSTEM += skill_loader.get_descriptions()
 SUBAGENT_SYSTEM += skill_loader.get_descriptions()
-# compactor = Compactor(TRANSCRIPT_DIR, client)
 console = Console()
 
 class Agent:
@@ -60,6 +59,13 @@ class Agent:
             case _:
                 raise ValueError(f"Unsupported method: {method}")
 
+        self.config = {
+            "api_method": self.client.__class__.__name__,
+            "base_url": self.client.base_url,
+            "api_key": self.client.api_key,
+            "tools": self.tools,
+        }
+
     def print_info(self, subagent=False):
         print(f"{Prompt.Baseurl}{self.client.base_url}")
         k = self.client.api_key or '*****'
@@ -69,7 +75,7 @@ class Agent:
         tools = CHILD_TOOLS if subagent else self.tools
         print(f"{Prompt.MCP}{','.join(f'{k}: {v}' for k,v in MCP.meta.items()) if MCP.meta else None}")
         print(f"{Prompt.ToolList}{' | '.join(t['name'] if 'name' in t else t['function']['name'] for t in tools)} [{len(tools)}]")
-        print(f"{Prompt.Config}API Method: {self.client.__class__.__name__}, Max Tokens: {MAX_TOKENS}, Compact {{threshhold: {THRESHOLD}, summary token: {COMPACT_TOKENS}, tool keep: {KEEP_RECENT}}}")
+        print(f"{Prompt.Config}API Method: {self.client.__class__.__name__}, Max Tokens: {MAX_TOKENS}, Compact {{threshhold: {COMPACT_THRESHOLD}, summary token: {COMPACT_TOKENS}, tool keep: {KEEP_RECENT_TOOL}}}")
         if not subagent: print(Prompt.Tips)
 
     @singledispatchmethod
@@ -88,6 +94,7 @@ class Agent:
             messages.pop(0)
 
         info = {
+            "config": self.config,
             "messages": [{"role": "system", "content": self.system if not subagent else self.sub_system}]+copy.deepcopy(messages[1:]),
             "last_response": None,
             "cost_tokens": 0,
@@ -124,8 +131,15 @@ class Agent:
         max_turn = max_turn or self.max_turn
         begin_time = time.time()
         manual_compact = False
+        ratio = 0
+        max_model_len = MAX_MODEL_LEN.get(self.model.lower(), MAX_MODEL_LEN['default'])
+        compactor = Compactor(TRANSCRIPT_DIR, self.client, max_model_len)
         while cur_turn < max_turn:
-            # messages[:] = compactor.auto_compact(messages, force=manual_compact)
+            messages[:] = compactor.auto_compact(messages, force=manual_compact, context_ratio=ratio)
+            if messages[0]["role"] == "user" and messages[0]["content"].startswith("[Conversation compressed"):
+                info["messages"].extend(messages)
+                if self.method == "openai":
+                    messages.insert(0, {"role": "system", "content": self.system if not subagent else self.sub_system})
             manual_compact = False
 
             notifs = BG.drain_notifications()
@@ -220,7 +234,8 @@ class Agent:
                                 else:
                                     print(f"{Prompt.Assistant}{block}")
 
-                            print(f"{Prompt.Tokens}inp: {response.usage.input_tokens}, out: {response.usage.output_tokens}, time: {cost_time:.2f}s, {response.usage.output_tokens / cost_time:.2f} tokens/s, tools: {sum(not i[-1].startswith('Error:') for i in info['used_tools'])}/{len(info['used_tools'])}, skills: {sum(not i[-1].startswith('Error:') for i in info['used_skills'])}/{len(info['used_skills'])}")
+                            ratio = (response.usage.input_tokens + response.usage.output_tokens) / max_model_len
+                            print(f"{Prompt.Tokens}inp: {response.usage.input_tokens}, out: {response.usage.output_tokens}, time: {cost_time:.2f}s, {response.usage.output_tokens / cost_time:.2f} tokens/s, tools: {sum(not i[-1].startswith('Error:') for i in info['used_tools'])}/{len(info['used_tools'])}, skills: {sum(not i[-1].startswith('Error:') for i in info['used_skills'])}/{len(info['used_skills'])}, context: {Prompt.RedLine*round(ratio*10)}{Prompt.BlueLine*(10 - round(ratio*10))} {ratio:.1%}")
 
                             if not BG.is_finished or BG._notification_queue:
                                 with KnightRiderStatus("run background tasks", width=11, color="#fbd367", fps=30):
@@ -245,7 +260,9 @@ class Agent:
                         else:
                             print(f"{Prompt.Assistant}{msg.content}")
                             info["last_response"] = msg.content
-                            print(f"{Prompt.Tokens}inp: {response.usage.prompt_tokens}, out: {response.usage.completion_tokens}, time: {cost_time:.2f}s, {(response.usage.completion_tokens or 0) / cost_time:.2f} tokens/s, tools: {sum(not i[-1].startswith('Error:') for i in info['used_tools'])}/{len(info['used_tools'])}, skills: {sum(not i[-1].startswith('Error:') for i in info['used_skills'])}/{len(info['used_skills'])}")
+
+                            ratio = response.usage.total_tokens / max_model_len
+                            print(f"{Prompt.Tokens}inp: {response.usage.prompt_tokens}, out: {response.usage.completion_tokens}, time: {cost_time:.2f}s, {(response.usage.completion_tokens or 0) / cost_time:.2f} tokens/s, tools: {sum(not i[-1].startswith('Error:') for i in info['used_tools'])}/{len(info['used_tools'])}, skills: {sum(not i[-1].startswith('Error:') for i in info['used_skills'])}/{len(info['used_skills'])}, context: {Prompt.RedLine*round(ratio*10)}{Prompt.BlueLine*(10 - round(ratio*10))} {ratio:.1%}")
 
                             if info["last_response"] is None and response.usage.completion_tokens == MAX_TOKENS:
                                 console.print(f"{Prompt.Error}[red]Max tokens reached[/]")
@@ -255,6 +272,15 @@ class Agent:
                     else:
                         if msg.content and msg.content.strip(): print(f"{Prompt.Assistant}{msg.content}")
 
+            match self.method:
+                case "anthropic":
+                    ratio = (response.usage.input_tokens + response.usage.output_tokens) / max_model_len
+                    print(f"{Prompt.Tokens}inp: {response.usage.input_tokens}, out: {response.usage.output_tokens}, time: {cost_time:.2f}s, {response.usage.output_tokens / cost_time:.2f} tokens/s, tools: {len([b for b in response.content if b.type == "tool_use"])}, context: {Prompt.RedLine*round(ratio*10)}{Prompt.BlueLine*(10 - round(ratio*10))} {ratio:.1%}")
+                case "openai":
+                    ratio = response.usage.total_tokens / max_model_len
+                    print(f"{Prompt.Tokens}inp: {response.usage.prompt_tokens}, out: {response.usage.completion_tokens}, time: {cost_time:.2f}s, {(response.usage.completion_tokens or 0) / cost_time:.2f} tokens/s, tool calls: {len(msg.tool_calls)}, context: {Prompt.RedLine*round(ratio*10)}{Prompt.BlueLine*(10 - round(ratio*10))} {ratio:.1%}")
+
+            # Tools calling
             results = []
             for block in response.content if self.method == 'anthropic' else msg.tool_calls:
                 match block.type:
@@ -330,7 +356,7 @@ class Agent:
                 info["messages"].extend(m)
 
         info["cost_time"] = time.time() - begin_time
-        print(f"{Prompt.Ended}total tokens: {info['cost_tokens']}, time: {info['cost_time']:.2f}s, tools: {len(info['used_tools'])}, skills: {len(info['used_skills'])}")
+        print(f"{Prompt.Ended}total tokens: {info['cost_tokens']}, time: {info['cost_time']:.2f}s, tools: {len(info['used_tools'])}, skills: {len(info['used_skills'])}, rounds: {cur_turn}")
         return info
 
     @__call__.register(str)

{sigmaflow-0.0.78 → sigmaflow-0.0.80}/src/sigmaflow/agent/conf.py

@@ -32,6 +32,8 @@ class Prompt:
     Unknown   = "[white on #444444]  UNKNOWN  [/] "
     Warning   = "[black on #F7EC4D]  WARNING  [/] "
     Ended     = "[black on #F68D2E]   ENDED   [/] "
+    RedLine   = "[red]━[/]"
+    BlueLine  = "[blue]━[/]"
 
 MAX_MODEL_LEN = {
     "qwen3.6-27b": 262144,
@@ -58,9 +60,10 @@ API_KEY = os.getenv("API_KEY")
 ANTHROPIC_BASE_URL = os.getenv("ANTHROPIC_BASE_URL")
 OPENAI_BASE_URL = os.getenv("OPENAI_BASE_URL")
 MAX_TOKENS = int(os.getenv("MAX_TOKENS", MAX_OUT_LEN["default"]))
-COMPACT_TOKENS = 2000
-KEEP_RECENT = 10000
-THRESHOLD = 5000000000
+COMPACT_TOKENS = int(os.getenv("COMPACT_TOKENS", 2048))
+COMPACT_THRESHOLD = int(os.getenv("COMPACT_THRESHOLD", 1e20))
+COMPACT_RATIO = float(os.getenv("COMPACT_RATIO", 0.95))
+KEEP_RECENT_TOOL = int(os.getenv("KEEP_RECENT_TOOL", 1e8))
 RAG_ENDPOINT = os.getenv("RAG_ENDPOINT")
 
 CMD_TIMEOUT = 600
@@ -137,7 +140,7 @@ CHILD_TOOLS = [
                 },
                 "args": {
                     "type": "string",
-                    "description": "Arguments passed after the grep-like command, for example: -n \"pattern\" ."
+                    "description": "Arguments passed after the grep-like command. Do not use the pipe symbol '|' to chain commands. Example: -n \"pattern\" <file>."
                 }
             },
             "required": ["tool", "args"]
@@ -404,6 +407,11 @@ API_ERRORS = [
     '用户额度不足',
 ]
 
+SUMMARY_PROMPT = """Summarize this conversation for continuity in the corresponding language. Include: 
+1) Core mission, 2) What was accomplished, 3) Current state, 4) Key decisions made.
+Be concise but preserve critical details (e.g. IDs, error messages, configs, paths). Keep it brief.\n
+"""
+
 @dataclass
 class MCPTool:
     name: str

{sigmaflow-0.0.78 → sigmaflow-0.0.80}/src/sigmaflow/agent/tools.py

@@ -219,36 +219,55 @@ skill_loader = SkillLoader(SKILLS_DIR)
 
 # -- Compactor: compact history --
 class Compactor:
-    def __init__(self, transript_dir: Path, client=None):
+    def __init__(self, transript_dir: Path, client=None, max_model_len=None):
         self.transript_dir = transript_dir
         self.client = client
+        self.max_model_len = max_model_len
 
     # -- Layer 1: tool_compact - replace old tool results with placeholders --
     def tool_compact(self, messages: list) -> list:
-        # Collect (msg_index, part_index, tool_result_dict) for all tool_result entries
+        # Collect tool results while tracking the tool names already seen.
         tool_results = []
-        for msg_idx, msg in enumerate(messages):
-            if msg["role"] == "user" and isinstance(msg.get("content"), list):
-                for part_idx, part in enumerate(msg["content"]):
-                    if isinstance(part, dict) and part.get("type") == "tool_result":
-                        tool_results.append((msg_idx, part_idx, part))
-        if len(tool_results) <= KEEP_RECENT:
-            return messages
-        # Find tool_name for each result by matching tool_use_id in prior assistant messages
         tool_name_map = {}
         for msg in messages:
-            if msg["role"] == "assistant":
-                content = msg.get("content", [])
-                if isinstance(content, list):
-                    for block in content:
-                        if hasattr(block, "type") and block.type == "tool_use":
-                            tool_name_map[block.id] = block.name
-        # Clear old results (keep last KEEP_RECENT)
-        to_clear = tool_results[:-KEEP_RECENT]
-        for _, _, result in to_clear:
-            if isinstance(result.get("content"), str) and len(result["content"]) < 100: continue
-            tool_id = result.get("tool_use_id", "")
-            tool_name = tool_name_map.get(tool_id, "unknown")
+            content = msg.get("content", [])
+            match msg["role"]:
+                case "assistant":
+                    if isinstance(content, list):
+                        for block in content:
+                            if hasattr(block, "type") and block.type == "tool_use":
+                                tool_name_map[block.id] = block.name
+                    for tool_call in msg.get("tool_calls", []):
+                        if isinstance(tool_call, dict):
+                            tool_id = tool_call.get("id")
+                            tool_name = tool_call.get("function", {}).get("name")
+                        else:
+                            tool_id = getattr(tool_call, "id", None)
+                            function = getattr(tool_call, "function", None)
+                            tool_name = getattr(function, "name", None)
+                        if tool_id and tool_name:
+                            tool_name_map[tool_id] = tool_name
+                case "user" if isinstance(content, list):
+                    for part in content:
+                        if isinstance(part, dict) and part.get("type") == "tool_result":
+                            tool_name = tool_name_map.get(
+                                part.get("tool_use_id", ""), "unknown"
+                            )
+                            tool_results.append((part, tool_name))
+                case "tool":
+                    tool_name = tool_name_map.get(
+                        msg.get("tool_call_id", ""), "unknown"
+                    )
+                    tool_results.append((msg, tool_name))
+                case _:
+                    continue
+        if len(tool_results) <= KEEP_RECENT_TOOL:
+            return messages
+        # Clear old results (keep last KEEP_RECENT_TOOL).
+        to_clear = tool_results[:-KEEP_RECENT_TOOL]
+        for result, tool_name in to_clear:
+            if isinstance(result.get("content"), str) and len(result["content"]) < 100:
+                continue
             result["content"] = f"[Previous: used {tool_name}]"
         return messages
 
@@ -263,20 +282,32 @@ class Compactor:
         print(f"{' '*12}transcript saved: {transcript_path}")
         # Ask LLM to summarize
         conversation_text = json.dumps(messages, default=str)
-        with KnightRiderStatus("", width=11, color="#5c9cf5", fps=30):
+        msg = [{
+            "role": "user", 
+            "content": SUMMARY_PROMPT + conversation_text
+        }]
+        # with KnightRiderStatus("", width=11, color="#5c9cf5", fps=30):
+        if self.client.__class__.__name__ == 'Anthropic':
             response = self.client.messages.create(
                 model=MODEL,
-                messages=[{"role": "user", "content":
-                    "Summarize this conversation for continuity. Include: "
-                    "1) What was accomplished, 2) Current state, 3) Key decisions made. "
-                    "Be concise but preserve critical details.\n\n" + conversation_text}],
+                messages=msg,
                 max_tokens=COMPACT_TOKENS,
             )
-        summary = response.content[0].text
-        print(f"{Prompt.Tokens}compress to: {response.usage.output_tokens}")
+            summary = response.content[0].text
+            print(f"{' '*12}summary: {summary}")
+            print(f"{Prompt.Tokens}compress to: {response.usage.output_tokens}")
+        else:
+            response = self.client.chat.completions.create(
+                model=MODEL,
+                messages=msg,
+                max_tokens=COMPACT_TOKENS
+            )
+            summary = response.choices[0].message.content
+            print(f"{' '*12}summary: {summary}")
+            print(f"{Prompt.Tokens}compress to: {response.usage.completion_tokens}")
 
         return [
-            {"role": "user", "content": f"[Conversation compressed. Transcript: {transcript_path}]\n\n{summary}"},
+            {"role": "user", "content": f"[Conversation compressed. Transcript: {transcript_path.name}]\n\n{summary}"},
             {"role": "assistant", "content": "Understood. I have the context from the summary. Continuing."},
         ]
 
@@ -284,12 +315,13 @@ class Compactor:
         """Rough token count: ~4 chars per token."""
         return len(str(messages)) // 4
 
-    def auto_compact(self, messages: list, force=False) -> list:
-        messages = self.tool_compact(messages)
-        if force or (n := self.estimate_tokens(messages)) > THRESHOLD:
-            info = f"{Prompt.Compact}manual compact" if force else f"{Prompt.Compact}estimate tokens: {n}, auto compact triggered"
+    def auto_compact(self, messages: list, force=False, context_ratio=0) -> list:
+        if force or context_ratio >= COMPACT_RATIO or (context_ratio * self.max_model_len) > COMPACT_THRESHOLD:
+            info = f"{Prompt.Compact}manual compact" if force else f"{Prompt.Compact}auto compact triggered"
             print(info)
             messages = self.summary_compact(messages)
+        else:
+            messages = self.tool_compact(messages)
         return messages
 
 # -- TaskManager: CRUD with dependency graph, persisted as JSON files --