ys/GenericAgent
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat: refine vision processing logic and update SOP

Browse Source
This commit is contained in:
Liang Jiaqing
2026-01-31 15:47:23 +08:00
parent 01c477c050
commit 669095699e
4 changed files with 40 additions and 21 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
agent_loop.py
View File

@@ -48,8 +48,8 @@ def agent_runner_loop(client, system_prompt, user_input, handler, tools_schema,
response = client.chat(messages=messages, tools=tools_schema)
if response.thinking: yield '<thinking>' + response.thinking + '</thinking>\n\n'
if '</summary>```' in response.content: response.content = response.content.replace('</summary>```', '</summary>\n```')
showcontent = response.content
if '</summary>' in showcontent: showcontent = showcontent.replace('</summary>', '</summary>\n\n')
if '</file_content>' in showcontent:
showcontent = re.sub(r'<file_content>\s*(.*?)\s*</file_content>', r'\n````\n<file_content>\n\1\n</file_content>\n````', showcontent, flags=re.DOTALL)
yield showcontent + '\n\n'

15
agentapp.py
View File

@@ -18,8 +18,11 @@ from agent_loop import agent_runner_loop, StepOutcome, BaseHandler
@st.cache_resource
def init():
if not os.path.exists('temp'): os.makedirs('temp')
mainllm = SiderLLMSession(multiturns=6)
llmclient = ToolClient(mainllm.ask, auto_save_tokens=True)
llm_sessions = [SiderLLMSession(default_model="gemini-3.0-flash"),
SiderLLMSession(default_model="gpt-5-mini"),
SiderLLMSession(default_model="claude-4.5-haiku"),
LLMSession()]
llmclient = ToolClient([x.ask for x in llm_sessions], auto_save_tokens=True)
return llmclient
llmclient = init()
@@ -47,6 +50,7 @@ def agent_backend_stream(raw_query):
sys_prompt = get_system_prompt()
handler = GenericAgentHandler(None, history, './temp')
llmclient.last_tools = ''
llmclient.raw_api = llmclient.raw_apis[st.session_state.get("llm_no", 0)]
ret = yield from agent_runner_loop(llmclient,
sys_prompt, raw_query, handler,
TOOLS_SCHEMA, max_turns=25)
@@ -62,6 +66,13 @@ for msg in st.session_state.messages:
with st.chat_message(msg["role"]):
st.markdown(msg["content"])
with st.sidebar:
current_idx = st.session_state.get("llm_no", 0)
st.caption(f"LLM Core: {current_idx}")
if st.button("切换备用链路"):
st.session_state.llm_no = (st.session_state.get("llm_no", 0) + 1) % len(llmclient.raw_apis)
st.rerun()
if prompt := st.chat_input("请输入指令"):
st.session_state.messages.append({"role": "user", "content": prompt})
with st.chat_message("user"):

26
ga.py
View File

@@ -66,12 +66,12 @@ def code_run(code: str, code_type: str = "python", timeout: int = 60, cwd: str =
status = "success" if exit_code == 0 else "error"
status_icon = "✅" if exit_code == 0 else "❌"
if exit_code is None: status_icon = "⏳"
output_snippet = (stdout_str[:100] + '...' + stdout_str[-100:]) if len(stdout_str) > 300 else stdout_str
output_snippet = smart_format(stdout_str, max_str_len=600, omit_str='\n[omitted long output]\n')
yield f"[Status] {status_icon} Exit Code: {exit_code}\n[Stdout]\n{output_snippet}\n"
if process.stdout: process.stdout.close()
if process.stdout: threading.Thread(target=process.stdout.close, daemon=True).start()
return {
"status": status,
"stdout": stdout_str[-2000:],
"stdout": smart_format(stdout_str, max_str_len=4000, omit_str='\n[omitted long output]\n'),
"exit_code": exit_code
}
except Exception as e:
@@ -198,7 +198,7 @@ def file_read(path, start=1, keyword=None, count=100, show_linenos=True):
try:
with open(path, 'r', encoding='utf-8', errors='replace') as f:
stream = (
(i, (l[:L_MAX].rstrip() + TAG if len(l) > L_MAX else l.rstrip()))
(i, (l[:L_MAX].rstrip() + TAG if len(l) > L_MAX else l.rstrip('\r\n')))
for i, l in enumerate(f, 1)
)
stream = itertools.dropwhile(lambda x: x[0] < start, stream)
@@ -211,15 +211,15 @@ def file_read(path, start=1, keyword=None, count=100, show_linenos=True):
before.append((i, l))
else: return f"Keyword '{keyword}' not found after line {start}."
else: res = itertools.islice(stream, count)
return "\n".join(f"{i}| {l}" if show_linenos else l for i, l in res)
return "\n".join(f"{i}|{l}" if show_linenos else l for i, l in res)
except Exception as e:
return f"Error: {str(e)}"
def smart_format(data, max_depth=2, max_str_len=100):
def smart_format(data, max_depth=2, max_str_len=100, omit_str=' ... '):
def truncate(obj, depth):
if isinstance(obj, str):
if len(obj) > max_str_len: return f"{obj[:max_str_len//2]} ... {obj[-max_str_len//2:]}"
return obj
if len(obj) < max_str_len+len(omit_str)*2: return obj
return f"{obj[:max_str_len//2]}{omit_str}{obj[-max_str_len//2:]}"
if depth >= max_depth: return truncate(str(obj), depth + 1)
if isinstance(obj, dict): return {k: truncate(v, depth + 1) for k, v in obj.items()}
if isinstance(obj, list): return [truncate(i, depth + 1) for i in obj]
@@ -259,15 +259,17 @@ class GenericAgentHandler(BaseHandler):
# 从 response.content 中提取代码块, 匹配 ```python ... ``` 或 ```powershell ... ```
pattern = rf"```{code_type}\n(.*?)\n```"
matches = re.findall(pattern, response.content, re.DOTALL)
warning = ""
if not matches:
return StepOutcome(None, next_prompt=f"【系统错误】：你调用了 code_run，但未在回复中提供 ```{code_type} 代码块。请重新输出代码并附带工具调用。")
# 提取最后一个代码块（通常是模型修正后的最终逻辑）
code = matches[-1].strip()
code = args.get("code")
if not code: return StepOutcome(None, next_prompt=f"【系统错误】：你调用了 code_run，但未在回复中提供 ```{code_type} 代码块。请重新输出代码并附带工具调用。")
warning = "\n下次要记得在回复中提供代码块，而不是放在参数中"
else: code = matches[-1].strip() # 提取最后一个代码块（通常是模型修正后的最终逻辑）
timeout = args.get("timeout", 60)
raw_path = os.path.join(self.cwd, args.get("cwd", './'))
cwd = os.path.normpath(os.path.abspath(raw_path))
result = yield from code_run(code, code_type, timeout, cwd)
next_prompt = self._get_anchor_prompt()
next_prompt = self._get_anchor_prompt() + warning
return StepOutcome(result, next_prompt=next_prompt)
def do_ask_user(self, args, response):

18
sidercall.py
View File

@@ -8,9 +8,11 @@ except ImportError:
capikey = ""
class SiderLLMSession:
def __init__(self, multiturns=6):
def __init__(self, multiturns=6, default_model="gemini-3.0-flash"):
self._core = Session(cookie=sider_cookie, proxies={'https':'127.0.0.1:2082'})
def ask(self, prompt, model="gemini-3.0-flash"):
self.default_model = default_model
def ask(self, prompt, model=None):
if model is None: model = self.default_model
if len(prompt) > 29000:
print(f"[Warn] Prompt too long ({len(prompt)} chars), truncating.")
prompt = prompt[-29000:]
@@ -44,7 +46,7 @@ class LLMSession:
elif not omit_images and msg['image']:
messages.append({"role": msg['role'], "content": [
{"type": "image_url", "image_url": {"url": f"data:image/png;base64,{msg['image']}"}},
{"type": "text", "text": msg['prompt']} ]})
{"type": "text", "text": msg['prompt']} ]})
else:
messages.append({"role": msg['role'], "content": msg['prompt']})
return messages
@@ -56,8 +58,10 @@ class LLMSession:
p = "Summarize prev summary and prev conversations into compact memory (facts/decisions/constraints/open questions). Do NOT restate long schemas. The new summary should less than 1000 tokens.\n"
messages = self.make_messages(old, omit_images=True)
messages += [{"role":"user", "content":p}]
self.summary = self.raw_ask(messages, model, temperature=0.1)
self.raw_msgs.insert(0, {"role":"system", "prompt":"Prev summary:\n"+self.summary, "image":None})
summary = self.raw_ask(messages, model, temperature=0.1)
if not summary.startswith("Error:"):
self.raw_msgs.insert(0, {"role":"system", "prompt":"Prev summary:\n"+summary, "image":None})
else: self.raw_msgs = old + self.raw_msgs # 不做了，下次再做
def ask(self, prompt, model="openai/gpt-5.1", image_base64=None):
self.raw_msgs.append({"role": "user", "prompt": prompt, "image": image_base64})
@@ -92,7 +96,9 @@ class MockResponse:
class ToolClient:
def __init__(self, raw_api_func, auto_save_tokens=False):
self.raw_api = raw_api_func
if isinstance(raw_api_func, list): self.raw_apis = raw_api_func
else: self.raw_apis = [raw_api_func]
self.raw_api = self.raw_apis[0]
self.auto_save_tokens = auto_save_tokens
self.last_tools = ''
self.total_cd_tokens = 0