mlx-code 0.0.33__tar.gz → 0.0.35__tar.gz

{mlx_code-0.0.33 → mlx_code-0.0.35}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: mlx-code
-Version: 0.0.33
+Version: 0.0.35
 Summary: Coding Agent for Mac
 Home-page: https://josefalbers.github.io/mlx-code/
 Author: J Joe
@@ -8,11 +8,12 @@ Author-email: albersj66@gmail.com
 License: Apache-2.0
 Project-URL: Source, https://github.com/JosefAlbers/mlx-code
 Project-URL: Issues, https://github.com/JosefAlbers/mlx-code/issues
-Project-URL: Documentation, https://josefalbers.github.io/mlx-code/
+Project-URL: Documentation, https://www.mlx-code.com/
 Requires-Python: >=3.12.8
 Description-Content-Type: text/markdown
 License-File: LICENSE
 Requires-Dist: mlx-lm>=0.31.3; platform_system == "Darwin"
+Requires-Dist: vllm; platform_system != "Darwin"
 Requires-Dist: httpx
 Requires-Dist: pydantic
 Provides-Extra: all
@@ -21,7 +22,6 @@ Requires-Dist: uvicorn; extra == "all"
 Requires-Dist: pygments; extra == "all"
 Requires-Dist: textual>=8.2.7; extra == "all"
 Requires-Dist: rich>=15.0.0; extra == "all"
-Requires-Dist: python-lsp-server[all]; extra == "all"
 Dynamic: author
 Dynamic: author-email
 Dynamic: description
@@ -107,7 +107,7 @@ result = await agent.run('refactor utils.py to use dataclasses')
 
 ```bash
 # ephemeral run (no installation)
-uvx --from mlx-code[all] mlc
+uvx --from "mlx-code[all]" mlc
 
 # or install into the current environment
 pip install mlx-code[all]
@@ -563,6 +563,15 @@ podman run cloudflare/cloudflared:latest tunnel --no-autoupdate run --token $JJ_
 phone http://jjoe.mlx-code.com
 ```
 
+#### VLLM
+
+```
+curl -fsSL https://raw.githubusercontent.com/vllm-project/vllm-metal/main/install.sh | bash
+source ~/.venv-vllm-metal/bin/activate # or `source ~/.venv-vllm-metal/bin/activate.fish` if fish 
+pip install mlx-code[all]
+mlc --backend vllm -m Qwen/Qwen3.5-0.8B
+```
+
 ---
 
 ## Credits

{mlx_code-0.0.33 → mlx_code-0.0.35}/README.md

@@ -70,7 +70,7 @@ result = await agent.run('refactor utils.py to use dataclasses')
 
 ```bash
 # ephemeral run (no installation)
-uvx --from mlx-code[all] mlc
+uvx --from "mlx-code[all]" mlc
 
 # or install into the current environment
 pip install mlx-code[all]
@@ -526,6 +526,15 @@ podman run cloudflare/cloudflared:latest tunnel --no-autoupdate run --token $JJ_
 phone http://jjoe.mlx-code.com
 ```
 
+#### VLLM
+
+```
+curl -fsSL https://raw.githubusercontent.com/vllm-project/vllm-metal/main/install.sh | bash
+source ~/.venv-vllm-metal/bin/activate # or `source ~/.venv-vllm-metal/bin/activate.fish` if fish 
+pip install mlx-code[all]
+mlc --backend vllm -m Qwen/Qwen3.5-0.8B
+```
+
 ---
 
 ## Credits

mlx_code-0.0.35/mlx_code/apis.py

@@ -0,0 +1,373 @@
+from __future__ import annotations
+_AN='function_call'
+_AM='Authorization'
+_AL='reasoning_content'
+_AK='output_tokens'
+_AJ='input_tokens'
+_AI='redacted_data'
+_AH='redacted_thinking'
+_AG='response'
+_AF='functionCall'
+_AE='Content-Type'
+_AD='object'
+_AC='parameters'
+_AB='description'
+_AA='tool_calls'
+_A9='tool_choice'
+_A8='stream'
+_A7='thought'
+_A6='reason'
+_A5='done'
+_A4='tool_call'
+_A3='toolcall_end'
+_A2='thinking_delta'
+_A1='index'
+_A0='data: '
+_z='start'
+_y='POST'
+_x='timestamp'
+_w='application/json'
+_v='tools'
+_u='temperature'
+_t='max_tokens'
+_s='model'
+_r='No api key'
+_q='call_id'
+_p='image_url'
+_o='length'
+_n='text_delta'
+_m='cache_write'
+_l='cache_read'
+_k=True
+_j='toolResult'
+_i='redacted'
+_h='mime_type'
+_g='args'
+_f='parts'
+_e='properties'
+_d='tool_call_id'
+_c='tool_use'
+_b='args_buf'
+_a='message'
+_Z='stop'
+_Y='error_message'
+_X='required'
+_W='function'
+_V='toolCall'
+_U='signature'
+_T='data'
+_S='assistant'
+_R='error'
+_Q='delta'
+_P='user'
+_O='output'
+_N='input'
+_M='arguments'
+_L='partial'
+_K='stop_reason'
+_J='id'
+_I='thinking'
+_H='payload'
+_G='usage'
+_F='role'
+_E='name'
+_D='text'
+_C='content'
+_B=None
+_A='type'
+import asyncio,json,os,time,uuid,logging,httpx
+from typing import Any,Literal
+from.tools import Tool
+logger=logging.getLogger(__name__)
+class EventStream:
+	def __init__(A)->_B:(A._queue):asyncio.Queue[dict|_B]=asyncio.Queue();(A._result):dict|_B=_B;(A._task):asyncio.Task|_B=_B
+	def _attach(A,task:asyncio.Task)->_B:A._task=task
+	def push(A,event:dict)->_B:A._queue.put_nowait(event)
+	def finish(A,result:dict)->_B:A._result=result;A._queue.put_nowait(_B)
+	async def result(A)->dict:
+		if A._result is _B:
+			async for B in A:0
+		assert A._result is not _B;return A._result
+	def __aiter__(A)->'EventStream':return A
+	async def __anext__(B)->dict:
+		A=await B._queue.get()
+		if A is _B:raise StopAsyncIteration
+		return A
+_REASONING_BUDGET:dict[str,int]={'minimal':512,'low':1024,'medium':8192,'high':16000,'xhigh':32000}
+class ClaudeChat:
+	def __init__(A,*,model=_B,api_key=_B,base_url=_B,max_tokens=8192,temperature=_B,reasoning:Literal['off','minimal','low','medium','high','xhigh']='off',tool_choice=_B):
+		A.model=model or'claude-haiku-4-5';A.api_key=api_key or os.environ.get('ANTHROPIC_API_KEY');A.base_url=(base_url or'https://api.anthropic.com').rstrip('/');A.max_tokens=max_tokens;A.temperature=temperature;A.reasoning=reasoning;A.tool_choice=tool_choice
+		if not A.api_key:logger.debug(_r)
+	def _fmt_content(H,content)->str|list:
+		G='image';F=content;C='cache_control'
+		if isinstance(F,str):return F
+		B=[]
+		for A in F:
+			E=A[_A]
+			if E==_D:
+				D={_A:_D,_D:A[_D]}
+				if A.get(C):D[C]={_A:A[C]}
+				B.append(D)
+			elif E==G:
+				D={_A:G,'source':{_A:'base64','media_type':A[_h],_T:A[_T]}}
+				if A.get(C):D[C]={_A:A[C]}
+				B.append(D)
+			elif E==_I:
+				if A[_i]:B.append({_A:_AH,_T:A.get(_I,A.get(_AI,''))})
+				else:B.append({_A:_I,_I:A[_I],_U:A.get(_U)or''})
+			elif E==_V:B.append({_A:_c,_J:A[_J],_E:A[_E],_N:A[_M]})
+		return B
+	def _build_messages(E,messages:list[dict])->list[dict]:
+		H='is_error';B=messages;D=[];A=0
+		while A<len(B):
+			C=B[A]
+			if C[_F]==_P:D.append({_F:_P,_C:E._fmt_content(C[_C])});A+=1
+			elif C[_F]==_S:D.append({_F:_S,_C:E._fmt_content(C[_C])});A+=1
+			elif C[_F]==_j:
+				G=[]
+				while A<len(B)and B[A][_F]==_j:F=B[A];G.append({_A:'tool_result','tool_use_id':F[_d],_C:E._fmt_content(F[_C]),H:F[H]});A+=1
+				D.append({_F:_P,_C:G})
+			else:A+=1
+		return D
+	async def stream(C,messages:list[dict],system:str,tools:list[Tool])->EventStream:
+		E=tools;B=system;D=EventStream();G:dict[str,Any]={_s:C.model,_t:C.max_tokens,'messages':C._build_messages(messages),_A8:_k}
+		if B:G['system']=[{_A:_D,_D:B}]
+		if C.temperature:G[_u]=C.temperature
+		if C.reasoning!='off':G[_I]={_A:'enabled','budget_tokens':_REASONING_BUDGET[C.reasoning]}
+		if E:
+			G[_v]=[A.schema()for A in E];A=C.tool_choice
+			if A is not _B:G[_A9]={_A:'tool',_E:A[_E]}if isinstance(A,dict)else{_A:'any'}if A==_X else{_A:'none'}if A=='none'else{_A:'auto'}
+		W={'x-api-key':str(C.api_key),'anthropic-version':'2023-06-01','content-type':_w};logger.debug(json.dumps(G,indent=2))
+		async def F()->_B:
+			A={_F:_S,_C:[],_K:_Z,_G:{_N:0,_O:0,_l:0,_m:0},_Y:_B,_x:int(time.time()*1000)}
+			try:
+				async with httpx.AsyncClient(timeout=12e1)as X:
+					async with X.stream(_y,f"{C.base_url}/v1/messages",json=G,headers=W)as N:
+						if N.status_code>=400:raise RuntimeError(f"HTTP {N.status_code}: {(await N.aread()).decode()}")
+						D.push({_A:_z,_H:{_L:A}});O:dict[int,str]={};H:dict[int,int]={}
+						async for T in N.aiter_lines():
+							if not T.startswith(_A0):continue
+							U=T[6:]
+							if U=='[DONE]':break
+							E=json.loads(U);I=E.get(_A)
+							if I=='message_start':J=E.get(_a,{}).get(_G,{});A[_G][_N]=J.get(_AJ,0);A[_G][_l]=J.get('cache_read_input_tokens',0);A[_G][_m]=J.get('cache_creation_input_tokens',0)
+							elif I=='content_block_start':
+								B=E.get(_A1,0);P=E.get('content_block',{});Q=P.get(_A)
+								if Q==_D:H[B]=len(A[_C]);A[_C].append({_A:_D,_D:''})
+								elif Q==_I:H[B]=len(A[_C]);A[_C].append({_A:_I,_I:'',_U:_B,_i:False})
+								elif Q==_AH:H[B]=len(A[_C]);A[_C].append({_A:_I,_AI:P.get(_T,''),_U:_B,_i:_k})
+								elif Q==_c:H[B]=len(A[_C]);A[_C].append({_A:_V,_J:P[_J],_E:P[_E],_M:{}});O[B]=''
+							elif I=='content_block_delta':
+								B=E.get(_A1,0);L=E.get(_Q,{});R=L.get(_A);K=H.get(B)
+								if K is _B:continue
+								F=A[_C][K]
+								if R==_n:M=L.get(_D,'');F[_D]+=M;D.push({_A:_n,_H:{_Q:M,_L:A}})
+								elif R==_A2:M=L.get(_I,'');F[_I]+=M;D.push({_A:_A2,_H:{_Q:M,_L:A}})
+								elif R=='signature_delta':F[_U]=(F.get(_U)or'')+L.get(_U,'')
+								elif R=='input_json_delta':O[B]=O.get(B,'')+L.get('partial_json','')
+							elif I=='content_block_stop':
+								B=E.get(_A1,0);K=H.get(B)
+								if K is not _B and K<len(A[_C]):
+									F=A[_C][K]
+									if F[_A]==_V:
+										V=O.pop(B,'')
+										if V:
+											try:F[_M]=json.loads(V)
+											except json.JSONDecodeError:pass
+										D.push({_A:_A3,_H:{_A4:F,_L:A}})
+							elif I=='message_delta':
+								S=E.get(_Q,{}).get(_K)
+								if S==_c:A[_K]=_c
+								elif S==_t:A[_K]=_o
+								elif S:A[_K]=_Z
+								J=E.get(_G,{})
+								if J:A[_G][_O]=J.get(_AK,A[_G][_O])
+							elif I=='message_stop':D.push({_A:_A5,_H:{_A6:A[_K],_a:A}})
+			except Exception as Y:A[_K]=_R;A[_Y]=str(Y);D.push({_A:_R,_H:{_R:A}})
+			D.finish(A)
+		D._attach(asyncio.create_task(F()));return D
+class DefaultChat:
+	def __init__(A,*,model=_B,api_key=_B,base_url=_B,max_tokens=8192,temperature=_B,tool_choice=_B):
+		C='noapi';B=api_key;A.model=model or C;A.api_key=B or C;A.base_url=(base_url or'http://127.0.0.1:8000').rstrip('/');A.max_tokens=max_tokens;A.temperature=temperature;A.tool_choice=tool_choice
+		if not B:logger.debug(_r)
+	def _build_messages(K,messages:list[dict],system:str)->list[dict]:
+		J='url';F=system;B=[]
+		if F:B.append({_F:'system',_C:F})
+		for A in messages:
+			if A[_F]==_P:
+				E=A[_C]
+				if isinstance(E,str):B.append({_F:_P,_C:E})
+				else:B.append({_F:_P,_C:[{_A:_D,_D:A[_D]}if A[_A]==_D else{_A:_p,_p:{J:f"data:{A[_h]};base64,{A[_T]}"}}for A in E]})
+			elif A[_F]==_S:
+				G=[A for A in A[_C]if A[_A]==_D];H=[A for A in A[_C]if A[_A]==_V];I=[A for A in A[_C]if A[_A]==_I];C:dict[str,Any]={_F:_S}
+				if G:C[_C]=''.join(A[_D]for A in G)
+				if I:C[_AL]=''.join(A[_I]for A in I)
+				if H:C[_AA]=[{_J:A[_J],_A:_W,_W:{_E:A[_E],_M:json.dumps(A[_M])}}for A in H]
+				B.append(C)
+			elif A[_F]==_j:D=A[_C];B.append({_F:'tool',_d:A[_d],_C:D[0][_D]if len(D)==1 and D[0][_A]==_D else[{_A:_D,_D:A[_D]}if A[_A]==_D else{_A:_p,_p:{J:f"data:{A[_h]};base64,{A[_T]}"}}for A in D]})
+		return B
+	async def stream(C,messages:list[dict],system:str,tools:list[Tool])->EventStream:
+		D=tools;B=EventStream();E:dict[str,Any]={_s:C.model,_t:C.max_tokens,'messages':C._build_messages(messages,system),_A8:_k}
+		if C.temperature:E[_u]=C.temperature
+		if D:
+			E[_v]=[{_A:_W,_W:{_E:A.name,_AB:A.description,_AC:{_A:_AD,_e:A.parameters.model_json_schema().get(_e,{}),_X:A.parameters.model_json_schema().get(_X,[])}}}for A in D];A=C.tool_choice
+			if A is not _B:E[_A9]={_A:_W,_W:{_E:A[_E]}}if isinstance(A,dict)else A
+		X={_AM:f"Bearer {C.api_key}",_AE:_w};logger.debug(json.dumps(E,indent=2))
+		async def F()->_B:
+			A={_F:_S,_C:[],_K:_Z,_G:{_N:0,_O:0,_l:0,_m:0},_Y:_B,_x:int(time.time()*1000)}
+			try:
+				async with httpx.AsyncClient(timeout=_B)as Y:
+					async with Y.stream(_y,f"{C.base_url}/v1/chat/completions",json=E,headers=X)as G:
+						if G.status_code>=400:raise RuntimeError(f"HTTP {G.status_code}: {(await G.aread()).decode()}")
+						B.push({_A:_z,_H:{_L:A}});F:dict[int,dict]={};L=M='';H=_B
+						async for P in G.aiter_lines():
+							if not P.startswith(_A0):continue
+							N=P[6:].strip()
+							if N=='[DONE]'or not N:continue
+							Q=json.loads(N);R=(Q.get('choices')or[{}])[0];I=R.get(_Q,{});H=R.get('finish_reason')or H
+							if(S:=I.get(_AL)or I.get(_I)):M+=S;B.push({_A:_A2,_H:{_Q:S,_L:A}})
+							if(T:=I.get(_C)):L+=T;B.push({_A:_n,_H:{_Q:T,_L:A}})
+							for O in I.get(_AA)or[]:
+								J=O.get(_A1,0);D=O.get(_W,{})
+								if J not in F:F[J]={_J:O.get(_J,str(uuid.uuid4())),_E:D.get(_E,''),_b:D.get(_M,'')}
+								else:
+									if _E in D:F[J][_E]+=D[_E]
+									if _M in D:F[J][_b]+=D[_M]
+							if(U:=Q.get(_G)):A[_G][_N]=U.get('prompt_tokens',A[_G][_N]);A[_G][_O]=U.get('completion_tokens',A[_G][_O])
+						if M:A[_C].append({_A:_I,_I:M,_U:_B,_i:False})
+						if L:A[_C].append({_A:_D,_D:L})
+						for(a,K)in sorted(F.items()):
+							try:V=json.loads(K[_b])if K[_b]else{}
+							except json.JSONDecodeError:V={}
+							W={_A:_V,_J:K[_J],_E:K[_E],_M:V};A[_C].append(W);B.push({_A:_A3,_H:{_A4:W,_L:A}})
+						A[_K]=_c if H==_AA else _o if H in(_o,_t)else _Z;B.push({_A:_A5,_H:{_A6:A[_K],_a:A}})
+			except Exception as Z:A[_K]=_R;A[_Y]=str(Z);B.push({_A:_R,_H:{_R:A}})
+			B.finish(A)
+		B._attach(asyncio.create_task(F()));return B
+class GeminiChat:
+	def __init__(A,*,model=_B,api_key=_B,base_url=_B,max_tokens=8192,temperature=_B,thinking=False,thinking_budget=8192,tool_choice=_B):
+		A.model=model or'gemini-3.1-flash-lite-preview';A.api_key=api_key or os.environ.get('GEMINI_API_KEY');A.base_url=(base_url or'https://generativelanguage.googleapis.com').rstrip('/');A.max_tokens=max_tokens;A.temperature=temperature;A.thinking=thinking;A.thinking_budget=thinking_budget;A.tool_choice=tool_choice
+		if not A.api_key:logger.debug(_r)
+	def _build_contents(M,messages:list[dict])->list[dict]:
+		L='tool_name';C=[];E:list[dict]=[]
+		def G():
+			if E:C.append({_F:_P,_f:list(E)});E.clear()
+		for A in messages:
+			if A[_F]==_P:
+				G();H=A[_C]
+				if isinstance(H,str):C.append({_F:_P,_f:[{_D:H}]})
+				else:C.append({_F:_P,_f:[{_D:A[_D]}if A[_A]==_D else{'inlineData':{'mimeType':A[_h],_T:A[_T]}}for A in H]})
+			elif A[_F]==_S:
+				G();D=[]
+				for B in A[_C]:
+					if B[_A]==_I:D.append({_A7:B[_I]})
+					elif B[_A]==_D:D.append({_D:B[_D]})
+					elif B[_A]==_V:
+						I:dict={_E:B[_E],_g:B[_M]}
+						if B[_J]!=B[_E]:I[_J]=B[_J]
+						D.append({_AF:I})
+				if D:C.append({_F:_s,_f:D})
+			elif A[_F]==_j:
+				F=A[_C]
+				try:J=json.loads(F[0][_D])if F else{}
+				except(json.JSONDecodeError,KeyError):J={'result':F[0][_D]if F else''}
+				K:dict={_E:A[L],_AG:J}
+				if A[_d]!=A[L]:K[_J]=A[_d]
+				E.append({'functionResponse':K})
+		G();return C
+	async def stream(A,messages:list[dict],system:str,tools:list[Tool])->EventStream:
+		I='ANY';H=tools;G=system;F='generationConfig';E='mode';B=EventStream();D:dict[str,Any]={'contents':A._build_contents(messages),F:{'maxOutputTokens':A.max_tokens}}
+		if G:D['systemInstruction']={_f:[{_D:G}]}
+		if A.temperature:D[F][_u]=A.temperature
+		if A.thinking:D[F]['thinkingConfig']={'includeThoughts':_k,'thinkingBudget':A.thinking_budget}
+		if H:
+			D[_v]=[{'functionDeclarations':[{_E:A.name,_AB:A.description,_AC:{_A:_AD,_e:A.parameters.model_json_schema().get(_e,{}),_X:A.parameters.model_json_schema().get(_X,[])}}for A in H]}];C=A.tool_choice
+			if C is not _B:D['toolConfig']={'functionCallingConfig':{E:I,'allowedFunctionNames':[C[_E]]}if isinstance(C,dict)else{E:I}if C==_X else{E:'NONE'}if C=='none'else{E:'AUTO'}}
+		S=f"{A.base_url}/v1beta/models/{A.model}:streamGenerateContent?alt=sse&key={A.api_key}";logger.debug(json.dumps(D,indent=2))
+		async def J()->_B:
+			A={_F:_S,_C:[],_K:_Z,_G:{_N:0,_O:0,_l:0,_m:0},_Y:_B,_x:int(time.time()*1000)}
+			try:
+				async with httpx.AsyncClient(timeout=12e1)as T:
+					async with T.stream(_y,S,json=D,headers={_AE:_w})as H:
+						if H.status_code>=400:raise RuntimeError(f"HTTP {H.status_code}: {(await H.aread()).decode()}")
+						B.push({_A:_z,_H:{_L:A}});I=J='';E:dict[str,dict]={};K=_B
+						async for L in H.aiter_lines():
+							if not L.startswith(_A0):continue
+							M=L[6:].strip()
+							if not M:continue
+							N=json.loads(M);O=(N.get('candidates')or[{}])[0];K=O.get('finishReason')or K
+							for C in O.get(_C,{}).get(_f)or[]:
+								if _A7 in C:J+=C[_A7];B.push({_A:_A2,_H:{_Q:C[_A7],_L:A}})
+								elif _D in C:I+=C[_D];B.push({_A:_n,_H:{_Q:C[_D],_L:A}})
+								elif _AF in C:
+									F=C[_AF];G=F.get(_J)or F[_E]
+									if G not in E:E[G]={_E:F[_E],_g:F.get(_g,{})}
+									else:E[G][_g].update(F.get(_g,{}))
+							if(P:=N.get('usageMetadata')):A[_G][_N]=P.get('promptTokenCount',A[_G][_N]);A[_G][_O]=P.get('candidatesTokenCount',A[_G][_O])
+						if J:A[_C].append({_A:_I,_I:J,_U:_B,_i:False})
+						if I:A[_C].append({_A:_D,_D:I})
+						for(G,Q)in E.items():R={_A:_V,_J:G,_E:Q[_E],_M:Q[_g]};A[_C].append(R);B.push({_A:_A3,_H:{_A4:R,_L:A}})
+						A[_K]=_o if K=='MAX_TOKENS'else _c if E else _Z;B.push({_A:_A5,_H:{_A6:A[_K],_a:A}});B.finish(A);return
+			except Exception as U:A[_K]=_R;A[_Y]=str(U);B.push({_A:_R,_H:{_R:A}})
+			B.finish(A)
+		B._attach(asyncio.create_task(J()));return B
+class CodexChat:
+	def __init__(A,*,model=_B,api_key=_B,base_url=_B,max_tokens=8192,temperature=_B,tool_choice=_B):
+		A.model=model or'gpt-5.4-mini';A.api_key=api_key or os.environ.get('OPENAI_API_KEY');A.base_url=(base_url or'https://api.openai.com').rstrip('/');A.max_tokens=max_tokens;A.temperature=temperature;A.tool_choice=tool_choice
+		if not A.api_key:logger.debug(_r)
+	def _build_input(F,messages:list[dict],system:str)->list[dict]:
+		E=system;B=[]
+		if E:B.append({_A:_a,_F:'developer',_C:E})
+		for A in messages:
+			if A[_F]==_P:D=A[_C];B.append({_A:_a,_F:_P,_C:D if isinstance(D,str)else[{_A:'input_text',_D:A[_D]}if A[_A]==_D else{_A:'input_image',_p:f"data:{A[_h]};base64,{A[_T]}"}for A in D]})
+			elif A[_F]==_S:
+				for C in A[_C]:
+					if C[_A]==_D:B.append({_A:_a,_F:_S,_C:[{_A:'output_text',_D:C[_D]}]})
+					elif C[_A]==_V:B.append({_A:_AN,_q:C[_J],_E:C[_E],_M:json.dumps(C[_M])})
+			elif A[_F]==_j:B.append({_A:'function_call_output',_q:A[_d],_O:A[_C][0][_D]if A[_C]else''})
+		return B
+	async def stream(B,messages:list[dict],system:str,tools:list[Tool])->EventStream:
+		D=tools;C=EventStream();E:dict[str,Any]={_s:B.model,'max_output_tokens':B.max_tokens,_N:B._build_input(messages,system),_A8:_k}
+		if B.temperature:E[_u]=B.temperature
+		if D:
+			E[_v]=[{_A:_W,_E:A.name,_AB:A.description,_AC:{_A:_AD,_e:A.parameters.model_json_schema().get(_e,{}),_X:A.parameters.model_json_schema().get(_X,[])}}for A in D];A=B.tool_choice
+			if A is not _B:E[_A9]={_A:_W,_E:A[_E]}if isinstance(A,dict)else A
+		T={_AM:f"Bearer {B.api_key}",_AE:_w};logger.debug(json.dumps(E,indent=2))
+		async def F()->_B:
+			A={_F:_S,_C:[],_K:_Z,_G:{_N:0,_O:0,_l:0,_m:0},_Y:_B,_x:int(time.time()*1000)}
+			try:
+				async with httpx.AsyncClient(timeout=12e1)as U:
+					async with U.stream(_y,f"{B.base_url}/v1/responses",json=E,headers=T)as H:
+						if H.status_code>=400:raise RuntimeError(f"HTTP {H.status_code}: {(await H.aread()).decode()}")
+						C.push({_A:_z,_H:{_L:A}});L='';F:dict[str,dict]={};M=_B
+						async for N in H.aiter_lines():
+							if not N.startswith(_A0):continue
+							O=N[6:].strip()
+							if not O:continue
+							D=json.loads(O);I=D.get(_A,'')
+							if I=='response.output_text.delta':P=D.get(_Q,'');L+=P;C.push({_A:_n,_H:{_Q:P,_L:A}})
+							elif I=='response.output_item.added':
+								J=D.get('item',{})
+								if J.get(_A)==_AN:G=J[_J];F[G]={_E:J.get(_E,''),_q:J.get(_q,G),_b:''}
+							elif I=='response.function_call_arguments.delta':
+								G=D.get('item_id','')
+								if G in F:F[G][_b]+=D.get(_Q,'')
+							elif I=='response.completed':
+								M=D.get(_AG,{}).get('status')
+								if(Q:=D.get(_AG,{}).get(_G)):A[_G][_N]=Q.get(_AJ,A[_G][_N]);A[_G][_O]=Q.get(_AK,A[_G][_O])
+						if L:A[_C].append({_A:_D,_D:L})
+						for K in F.values():
+							try:R=json.loads(K[_b])if K[_b]else{}
+							except json.JSONDecodeError:R={}
+							S={_A:_V,_J:K[_q],_E:K[_E],_M:R};A[_C].append(S);C.push({_A:_A3,_H:{_A4:S,_L:A}})
+						A[_K]=_c if F else _o if M=='incomplete'else _Z;C.push({_A:_A5,_H:{_A6:A[_K],_a:A}})
+			except Exception as V:A[_K]=_R;A[_Y]=str(V);C.push({_A:_R,_H:{_R:A}})
+			C.finish(A)
+		C._attach(asyncio.create_task(F()));return C
+def resolve_api(api,*,model,api_key,base_url):
+	D=base_url;C=api_key;B=model;A=api
+	if not isinstance(A,str):return A
+	if A=='claude':return ClaudeChat(model=B,api_key=C,base_url=D)
+	if A=='gemini':return GeminiChat(model=B,api_key=C,base_url=D)
+	if A=='codex':return CodexChat(model=B,api_key=C,base_url=D)
+	return DefaultChat(model=B,api_key=C,base_url=D)

mlx_code-0.0.35/mlx_code/bare.py

@@ -0,0 +1,177 @@
+from __future__ import annotations
+_J='commit'
+_I='is_error'
+_H='content'
+_G='thinking_delta'
+_F='type'
+_E='text'
+_D='\n'
+_C=False
+_B=True
+_A=None
+import asyncio,json,logging,select,signal,sys,threading
+from typing import Callable
+from.repl import Agent,TabModel,CommandEngine,UIAdapter,HELP_TEXT,format_exit_summary
+logger=logging.getLogger(__name__)
+_INTERRUPTED=object()
+class BareAdapter(UIAdapter):
+	def __init__(A,repl:'BareRepl'):A.repl=repl
+	def show_error(A,text:str)->_A:print(f"\n✗ {text}",flush=_B)
+	def show_command_result(B,cmd:str,content:str|object)->_A:
+		A=content
+		if isinstance(A,str):print(A)
+		else:print(str(A))
+	def show_diff(A,diff_text:str,ref1_label:str,ref2_label:str)->_A:print(diff_text)
+	def show_history_list(A,lines:list[str])->_A:print(_D.join(lines))
+	def show_history_raw(A,json_text:str)->_A:print(json_text)
+	async def add_tab(A,tab:TabModel)->_A:0
+	def remove_tab(A,removed_index:int)->_A:
+		if A.repl.engine.active_index>=len(A.repl.engine.tabs):A.repl.engine.active_index=len(A.repl.engine.tabs)-1
+	def switch_to_tab(A,index:int)->_A:A.repl._render_tab_delimiter();A.repl._print_history_for_tab(A.repl.engine.tabs[index])
+	def refresh_chrome(A)->_A:0
+	def clear_tab_display(A,tab:TabModel)->_A:0
+	def clear_ephemeral(A,tab:TabModel)->_A:0
+	def on_agent_event(A,event:dict,tab:TabModel)->_A:
+		if tab is not A.repl.engine.active_tab:return
+		A.repl._handle_event(event,tab)
+	async def run_interactive_shell(B,command:str,cwd:str,env:dict|_A)->int:A=await asyncio.create_subprocess_shell(command,cwd=cwd,stdin=_A,stdout=_A,stderr=_A,env=env);await A.wait();return A.returncode or 0
+	def exit_app(B,summary:list[dict])->_A:
+		A=format_exit_summary(summary)
+		if A:print(A)
+		B.repl._should_exit=_B
+	def peek_tree_text(A,tab:TabModel)->str:return'Peek is not supported in bare mode.'
+	async def peek_open(A,tab:TabModel,path:list[Agent])->_A:0
+	def peek_close(A,tab:TabModel,agent_id:str|_A)->bool:return _C
+class BareRepl:
+	def __init__(A,engine:CommandEngine,init_prompt:str|_A=_A):A.engine=engine;A.adapter=BareAdapter(A);A.engine.bind(A.adapter);A.engine.attach_agent(A.engine.tabs[0]);A.init_prompt=init_prompt;(A._pending_nls):int=0;(A._awaiting_content):bool=_C;(A._has_output):bool=_C;(A._last_stream_type):str|_A=_A;A._should_exit=_C;A._interrupted=threading.Event()
+	async def run(A)->_A:
+		C=asyncio.get_running_loop()
+		if A.init_prompt:D,A.init_prompt=A.init_prompt,_A;await A.engine.handle_input(D);await A._await_active_task(C)
+		while _B:
+			if A._should_exit:break
+			B=await A._read_input_interruptible(C)
+			if B is _INTERRUPTED:continue
+			if B is _A:print();break
+			B=B.strip()
+			if not B:continue
+			if B.lower()in{'exit','quit'}:break
+			await A.engine.handle_input(B)
+			if A._should_exit:break
+			await A._await_active_task(C)
+	async def _read_input_interruptible(B,loop:asyncio.AbstractEventLoop):
+		A=loop
+		def D():B._interrupted.set()
+		C=_C
+		try:A.add_signal_handler(signal.SIGINT,D);C=_B
+		except(NotImplementedError,RuntimeError):pass
+		try:return await A.run_in_executor(_A,B._read_input)
+		finally:
+			if C:A.remove_signal_handler(signal.SIGINT)
+	async def _await_active_task(C,loop:asyncio.AbstractEventLoop)->_A:
+		A=C.engine.active_tab
+		if A.running_task is _A:return
+		D=A.agent
+		def E()->_A:D.abort();print('\n(aborting…)')
+		B=_C
+		try:loop.add_signal_handler(signal.SIGINT,E);B=_B
+		except(NotImplementedError,RuntimeError):pass
+		try:await A.running_task
+		except asyncio.CancelledError:pass
+		finally:
+			if B:loop.remove_signal_handler(signal.SIGINT)
+	def _read_input(A):
+		G=A.engine.active_tab;D=f"[{G.title}] ≫ ";B:list[str]=[]
+		while _B:
+			sys.stdout.write(D);sys.stdout.flush()
+			while _B:
+				if A._interrupted.is_set():A._interrupted.clear();print();return _INTERRUPTED
+				try:E,H,H=select.select([sys.stdin],[],[],.1)
+				except(OSError,ValueError):E=[sys.stdin]
+				if E:break
+			F=sys.stdin.readline()
+			if F=='':return
+			C=F.rstrip(_D);B.append(C)
+			if C.endswith('\\'):B[-1]=C[:-1];D='... '
+			else:break
+		return _D.join(B)
+	def _handle_event(A,event:dict,tab:TabModel)->_A:
+		K='error';H=event;B,C=H[_F],H.get('payload',{})
+		if B in('text_delta',_G):
+			I=C.get('delta','')
+			if I:A._write_delta(I,B)
+		elif B=='tool_start':A._pending_nls=0;A._awaiting_content=_C;A._has_output=_B;A._last_stream_type=B
+		elif B=='tool_end':
+			L=C.get('result',{});E=L.get(_H);M=C.get(_I,_C);D=''
+			if E:
+				F:list[str]=[]
+				if isinstance(E,str):F.append(E)
+				elif isinstance(E,list):
+					for G in E:
+						if isinstance(G,dict)and G.get(_F)==_E:F.append(G.get(_E,''))
+				D=_D.join(F).strip(_D)
+			if M:
+				J='✗ '
+				if not D:D=f"{C.get("name","?")} failed"
+			else:J='→ 'if D else''
+			if D:A._write_delta(J+D,'tool_result')
+			A._last_stream_type=B;print()
+		elif B==_J:A._pending_nls=0;A._awaiting_content=_C;A._has_output=_B;print(f"\n◇ [{C.get("sha","")}] committed",flush=_B);A._last_stream_type=B
+		elif B==K:A._pending_nls=0;A._awaiting_content=_C;A._has_output=_B;N=str(C.get(K,C));print(f"\n✗ {N}",flush=_B);A._last_stream_type=B
+		elif B in('agent_start','turn_start'):A._pending_nls=0;A._awaiting_content=_C;A._has_output=_C;A._last_stream_type=_A
+		elif B=='agent_end':
+			A._pending_nls=0
+			if A._has_output:print()
+			A._last_stream_type=_A;A._has_output=_C;A._awaiting_content=_C
+	def _write_delta(A,text:str,delta_type:str)->_A:
+		D=delta_type;B=text
+		if D!=A._last_stream_type:A._pending_nls=0;A._awaiting_content=_B;A._last_stream_type=D
+		if A._awaiting_content:
+			B=B.lstrip(_D)
+			if not B:return
+		if A._awaiting_content:
+			if A._has_output:print()
+			A._awaiting_content=_C
+		if not A._awaiting_content and A._pending_nls>0:print(_D*A._pending_nls,end='',flush=_B);A._pending_nls=0
+		C=B.rstrip(_D)
+		if C:
+			if D==_G:print(f"{C}",end='',flush=_B)
+			else:print(C,end='',flush=_B)
+			A._has_output=_B
+		A._pending_nls=len(B)-len(C)
+	def _render_tab_delimiter(C)->_A:
+		A:list[str]=[]
+		for(B,D)in enumerate(C.engine.tabs):
+			if B==C.engine.active_index:A.append(f"▶ {B+1}. {D.title}")
+			else:A.append(f"▷ {B+1}. {D.title}")
+		print(_D+'┗━━┫ '+' ┃ '.join(A)+' ┃')
+	def _print_history_for_tab(B,tab:TabModel)->_A:
+		A=tab
+		for C in A.agent.messages:B._print_message(C)
+		if A.agent.stream is not _A:B._print_message(A.agent.stream)
+	def _print_message(O,msg:dict)->_A:
+		M='thinking';L='toolResult';F=msg;C=F.get('role','');D=F.get(_H,'');I=F.get(_I,_C)
+		if isinstance(D,list):G=D
+		elif isinstance(D,str):G=[{_F:_E,_E:D}]
+		else:return
+		if C==L:
+			H:list[str]=[]
+			for A in G:
+				if isinstance(A,dict)and A.get(_F)==_E:
+					J=A.get(_E,'').strip(_D)
+					if J:H.append(J)
+			if H:N='✗ 'if I else'→ ';print(N+_D.join(H))
+			return
+		for A in G:
+			K=A.get(_F,_E)
+			if K=='toolCall':
+				E=A.get('arguments',{})
+				if isinstance(E,dict):E=json.dumps(E,ensure_ascii=_C)
+				print(f"⚙ {A.get("name","")} {E}");continue
+			B=A.get(_E,'')or A.get(M,'')or'';B=B.strip(_D)
+			if not B:continue
+			if K==M:print(f"{B}")
+			elif I:print(f"✗ {B}")
+			elif C=='user':print(f"≫ {B}")
+			elif C==_J:print(f"◇ {B}")
+			elif C==L:print(f"→ {B}")
+			else:print(B)