"""OpenAI wire format, role isolation and ledger protections; no live calls.""" import json import time import unittest from pathlib import Path from types import SimpleNamespace import httpx from src.adapters.llm.openai_code import OpenAICodeProvider,MODELS,ENDPOINT,read_token,providers_from_connection from src.modules.llm_gateway.routing import VERSION,plan,catalog from src.modules.llm_gateway.service import LLMGateway from src.modules.skill.domain.models import SkillError from src.tests.runtime.fixtures import setup,workspace,USER,SOURCE class OpenAICodeTests(unittest.TestCase): def provider(self,handler=None,model=MODELS[0]): p=OpenAICodeProvider(model=model,token='sk-synthetic-openai-test-only',transport=httpx.MockTransport(handler) if handler else None) p.health={'status':'healthy'} return p def test_medium_wire_format_counts_all_completion_tokens(self): def reply(req): body=json.loads(req.content) self.assertEqual(str(req.url),ENDPOINT) self.assertEqual(body['reasoning_effort'],'medium') self.assertEqual(body['max_completion_tokens'],16384) self.assertFalse(body['store']) self.assertNotIn('max_tokens',body);self.assertNotIn('enable_thinking',body) self.assertNotIn('temperature',body);self.assertNotIn('tools',body) return httpx.Response(200,json={'id':'chatcmpl-openai-test', 'choices':[{'finish_reason':'stop','message':{'content':json.dumps({'files':SOURCE})}}], 'usage':{'prompt_tokens':100,'completion_tokens':80,'completion_tokens_details':{'reasoning_tokens':60}}}) with workspace() as folder: rt,request,_,_=setup(folder);r=rt.create(USER,request,'openai-ledger-test') rt.store.mutate(r['id'],lambda r:r.update(status='running')) body=LLMGateway(rt.store,[self.provider(reply)]).generate(r['id'],'code',{},lambda:False) self.assertEqual(body['files'],SOURCE) call=rt.store.calls(r['id'])[0] self.assertEqual(call['charged'],180) # Reasoning tokens already included in 80. self.assertEqual(call['metadata']['reasoning_effort'],'medium') self.assertNotIn('sk-synthetic-openai-test-only',json.dumps(call)) def test_priority_is_61_astra_55_in_all_requested_roles_and_picker(self): legacy=SimpleNamespace(name='legacy',model='kimi-k3',routing_version=VERSION,selectable=True, task_roles=['chat','design','diagnose','code','repair'],health={'status':'healthy'}) providers=[legacy,*[self.provider(model=m) for m in reversed(MODELS)]] for stage in ('chat','design','diagnose','code','repair'): self.assertEqual([p.model for p in plan(providers,stage,{})[0]],list(MODELS)+['kimi-k3']) self.assertEqual(catalog(providers)['effective_policy']['java-code'],MODELS[0]) self.assertEqual(catalog(providers)['effective_policy']['design'],MODELS[0]) self.assertEqual(catalog(providers)['output_limits']['design'],8192) self.assertEqual(catalog(providers)['output_limits']['code'],16384) self.assertTrue(all(m['selectable'] for m in catalog(providers)['models'])) for model in MODELS: self.assertEqual(plan(providers,'design',{'request':{'selected_model':model}})[0][0].model,model) def test_unhealthy_model_is_excluded_and_manual_choice_is_preserved(self): p=self.provider();p.health={'status':'unknown'} self.assertEqual(plan([p],'code',{})[0],[]) legacy=SimpleNamespace(name='legacy',model='kimi-k3',routing_version=VERSION,selectable=True, task_roles=['code'],health={'status':'healthy'}) self.assertEqual(plan([self.provider(),legacy],'code',{'request':{'selected_model':'kimi-k3'}})[0],[legacy]) def test_design_medium_has_reasoning_space_and_retry_respects_unlimited_policy(self): limits=[] def reply(req): body=json.loads(req.content);limits.append(body['max_completion_tokens']) self.assertEqual(body['reasoning_effort'],'medium') return httpx.Response(200,json={'id':'chatcmpl-design-test', 'choices':[{'finish_reason':'length' if len(limits)==1 else 'stop', 'message':{'content':'{"summary":"concise design"}'}}], 'usage':{'prompt_tokens':100,'completion_tokens':8192 if len(limits)==1 else 100}}) with workspace() as folder: rt,request,_,_=setup(folder) request['budget_policy']={'mode':'unlimited-usage-v1'} run=rt.create(USER,request,'openai-design-space') rt.store.mutate(run['id'],lambda r:r.update(status='running')) result=LLMGateway(rt.store,[self.provider(reply)]).generate(run['id'],'design',{},lambda:False,token_limit=1000) self.assertEqual(result,{'summary':'concise design'}) self.assertEqual(limits,[8192,16384]) self.assertTrue(all(c['state']=='settled' for c in rt.store.calls(run['id']))) def test_unknown_transport_keeps_reservation_and_never_tries_backup(self): seen=[] def timeout(req):seen.append('first');raise httpx.ReadTimeout('synthetic') def backup(req):seen.append('backup');raise AssertionError('must not replay') with workspace() as folder: rt,request,_,_=setup(folder);r=rt.create(USER,request,'openai-unknown-test') rt.store.mutate(r['id'],lambda r:r.update(status='running')) gateway=LLMGateway(rt.store,[self.provider(timeout),self.provider(backup,MODELS[1])]) with self.assertRaises(SkillError):gateway.generate(r['id'],'code',{},lambda:False) self.assertEqual(seen,['first']) call=rt.store.calls(r['id'])[0] self.assertEqual(call['state'],'unknown');self.assertIsNone(call['charged']) def test_connection_requires_fresh_matching_credential_validation(self): with workspace() as folder: root=Path(folder);key=root/'key.txt';key.write_text('sk-synthetic-openai-test-only',encoding='utf8') config={'endpoint':ENDPOINT,'credential_source':str(key),'reasoning_effort':'medium'} health={**config,'credential_mtime_ns':key.stat().st_mtime_ns,'checked_at':time.time(), 'models':{MODELS[0]:{'status':'healthy'},MODELS[1]:{'status':'not_probed'}}} path=root/'connection.json';path.write_text(json.dumps(config),encoding='utf8') h=root/'health.json';h.write_text(json.dumps(health),encoding='utf8') self.assertEqual([p.model for p in providers_from_connection(path)],[MODELS[0]]) health['checked_at']-=86401;h.write_text(json.dumps(health),encoding='utf8') self.assertEqual(providers_from_connection(path),[]) health['checked_at']=time.time();health['credential_mtime_ns']+=1 h.write_text(json.dumps(health),encoding='utf8') self.assertEqual(providers_from_connection(path),[]) config['endpoint']='https://unapproved.example/v1/chat/completions' path.write_text(json.dumps(config),encoding='utf8') with self.assertRaises(ValueError):providers_from_connection(path) def test_token_reader_rejects_ambiguous_keys_without_echo(self): with workspace() as folder: p=Path(folder)/'key.txt';p.write_text('sk-first-secret sk-second-secret',encoding='utf8') with self.assertRaisesRegex(ValueError,'^OPENAI_CREDENTIAL_CONFIGURATION_INVALID$'):read_token(p) with self.assertRaises(ValueError):OpenAICodeProvider(model='gpt-unapproved',token='sk-synthetic') def test_duplicate_listener_is_rejected_before_workbench_initialization(self): import socket from scripts.serve_local import require_available_ports with socket.socket() as listener: listener.bind(('127.0.0.1',0));listener.listen() port=listener.getsockname()[1] with self.assertRaisesRegex(RuntimeError,'LOCAL_LISTENER_ALREADY_IN_USE'): require_available_ports(port)