dellweb3's picture
Release 2026.11.12-hf-deepseek.1
52107a5 verified
Raw History Blame Contribute Delete
5.8 kB
"""Server-owned Bailian capabilities and deterministic routing; never prompt-selected endpoints."""
VERSION = 'bailian-quota-burndown-v1'
PROFILES = {
'kimi-k3': {'roles':['chat','design','diagnose','code','repair'], 'json_mode':True, 'thinking':False, 'selectable':True},
'deepseek-v4-flash-0731': {'roles':['chat','design','diagnose','code','repair'], 'json_mode':True, 'thinking':False, 'selectable':True},
'qwen3.8-max-0902': {'roles':['chat','design','diagnose','code','repair'], 'json_mode':True, 'thinking':False, 'selectable':True},
'deepseek-v4.1-flash-1017': {'roles':['chat','design','diagnose','code','repair'], 'json_mode':True, 'thinking':False, 'selectable':True},
'qwen3-coder-flash': {'roles':['code','repair'], 'json_mode':False, 'thinking':None},
'qwen3.7-plus': {'roles':['design','diagnose','code','repair'], 'json_mode':True, 'thinking':False},
'qwen3.8-max': {'roles':['design','diagnose','code','repair'], 'json_mode':True, 'thinking':False},
}
# 2026-11-12: user requests DeepSeek 4.1 when Sol is unavailable.
# Models still require successful health checks; manual selections remain pinned.
PREFERENCE = {
'chat': ['deepseek-v4.1-flash-1017', 'deepseek-v4-flash-0731', 'kimi-k3', 'qwen3.8-max-0902'],
'design': ['deepseek-v4.1-flash-1017', 'kimi-k3', 'deepseek-v4-flash-0731', 'qwen3.8-max-0902'],
'diagnose': ['deepseek-v4.1-flash-1017', 'kimi-k3', 'deepseek-v4-flash-0731', 'qwen3.8-max-0902'],
'code': ['deepseek-v4.1-flash-1017', 'qwen3.8-max-0902', 'deepseek-v4-flash-0731', 'kimi-k3'],
'repair': ['deepseek-v4.1-flash-1017', 'qwen3.8-max-0902', 'deepseek-v4-flash-0731', 'kimi-k3'],
}
def selected_provider(providers, model, roles):
"""Resolve one explicit API Code; a manual choice never silently falls back."""
from src.modules.skill.domain.models import SkillError
if model in (None,'auto'):
return None
if not isinstance(model,str):
raise SkillError('MODEL_SELECTION_INVALID','请选择列表中的模型。',422)
matches=[p for p in providers if getattr(p,'model',None)==model]
if (len(matches)!=1 or not getattr(matches[0],'selectable',False)
or not set(roles).issubset(getattr(matches[0],'task_roles',[]))
or getattr(matches[0],'health',{}).get('status')!='healthy'):
raise SkillError('MODEL_SELECTION_UNAVAILABLE','所选模型当前不可用于这类任务,请重新选择。',422)
return matches[0]
def plan(providers, kind, run):
chosen=run.get('request',{}).get('selected_model')
if chosen:
from src.modules.skill.domain.models import SkillError
try:return [selected_provider(providers,chosen,[kind])],'user-selected'
except SkillError:return [],'user-selected-unavailable'
if not any(getattr(p,'routing_version',None)==VERSION for p in providers):
return list(providers), 'configured-order'
# Quota burn-down: configured, selectable and verified models, ordered by the
# per-kind PREFERENCE list so soonest-expiring quota is spent first.
# A known quota rejection may advance this list; unknown results are never replayed.
available=[p for p in providers if (getattr(p,'selectable',False) or getattr(p,'coding_only',False))
and kind in getattr(p,'task_roles',[])
and getattr(p,'health',{}).get('status')=='healthy']
order=PREFERENCE.get(kind,[])
if kind in {'chat','design','diagnose','code','repair'}:
order=['gpt-6.1-sol','gpt-6-astra','gpt-5.5',*order]
preferred={getattr(p,'automatic_preference',None) for p in providers}-{None}
if len(preferred)==1:
first=preferred.pop()
order=[first,*[model for model in order if model!=first]]
available.sort(key=lambda p: order.index(p.model) if p.model in order else len(order))
return available,('openai-medium-priority' if available and getattr(available[0],'provider_family',None)=='openai' else 'configured-verified-order')
def catalog(providers):
effective={}
for label,kind,dependencies in [('chat','chat',{}),('design','design',{}),('java-code','code',{}),
('business-code','code',{}),('dependency-code','code',{'approved':'fixed'}),('diagnose','diagnose',{}),('repair','repair',{})]:
candidates,_=plan(providers,kind,{'project_id':'erp-demo' if label=='business-code' else 'calculator-demo','snapshot':{'framework':{'dependencies':dependencies}}})
effective[label]=getattr(candidates[0],'model',candidates[0].name) if candidates else None
return {'policy_version':VERSION,'automatic':True,'effective_policy':effective,'models':[
{'model':getattr(p,'model',p.name),'roles':getattr(p,'task_roles',['configured']),
'selectable':getattr(p,'selectable',False),'json_mode':getattr(p,'json_mode',True),
'health':getattr(p,'health',{'status':'not_probed'}),
**({'provider':'openai','reasoning_effort':p.reasoning_effort} if getattr(p,'provider_family',None)=='openai' else {})}
for p in providers],
'selection_policy':'configured-verified-order',
'output_limits':{'design':8192 if effective['design'] in {'gpt-6.1-sol','gpt-6-astra','gpt-5.5'} else 2048,
'diagnose':8192 if effective['diagnose'] in {'gpt-6.1-sol','gpt-6-astra','gpt-5.5'} else 2048,
'code':16384 if effective['java-code'] in {'gpt-6.1-sol','gpt-6-astra','gpt-5.5'} else 8192,
'dependency-code':16384 if effective['dependency-code'] in {'gpt-6.1-sol','gpt-6-astra','gpt-5.5'} else 12288,
'repair':16384 if effective['repair'] in {'gpt-6.1-sol','gpt-6-astra','gpt-5.5'} else 12288,'known_output_retry_max':24576},
'unknown_outcome_replay':False}