| 123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505 |
- """Exercise the real bundled AIAgent against a deterministic model endpoint."""
- import json
- import sys
- import os
- import ast
- import hashlib
- import sqlite3
- from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer
- from pathlib import Path
- import subprocess
- import tempfile
- import threading
- import time
- import unittest
- ROOT = Path(__file__).resolve().parents[2]
- class Model(BaseHTTPRequestHandler):
- calls = []
- def log_message(self, *args):
- pass
- def do_POST(self):
- body = json.loads(self.rfile.read(int(self.headers['Content-Length'])))
- self.calls.append(body)
- used_tool = any(m.get('role') == 'tool' for m in body['messages'])
- message = {'role': 'assistant', 'content': '执行完成:收到工具结果。'} if used_tool else {
- 'role': 'assistant', 'content': None,
- 'tool_calls': [{'id': 'call-proof', 'type': 'function', 'function': {'name': 'proof_read', 'arguments': '{"path":"proof.txt"}'}}],
- }
- finish = 'stop' if used_tool else 'tool_calls'
- self.send_response(200)
- self.send_header('Content-Type', 'text/event-stream' if body.get('stream') else 'application/json')
- self.end_headers()
- if body.get('stream'):
- delta = dict(message)
- if 'tool_calls' in delta:
- delta['tool_calls'][0]['index'] = 0
- for part in [{'id':'test','object':'chat.completion.chunk','model':'test-model','choices':[{'index':0,'delta':delta,'finish_reason':None}]},
- {'id':'test','object':'chat.completion.chunk','model':'test-model','choices':[{'index':0,'delta':{},'finish_reason':finish}],'usage':{'prompt_tokens':100,'completion_tokens':20,'total_tokens':120}}]:
- self.wfile.write(('data: '+json.dumps(part)+'\n\n').encode())
- self.wfile.write(b'data: [DONE]\n\n')
- else:
- self.wfile.write(json.dumps({'id':'test','object':'chat.completion','model':'test-model','choices':[{'index':0,'message':message,'finish_reason':finish}],'usage':{'prompt_tokens':100,'completion_tokens':20,'total_tokens':120}}).encode())
- class LegacyMemoryModel(Model):
- """Calls the retired native ``memory`` tool, then the workspace tool, then answers."""
- calls = []
- def do_POST(self):
- body = json.loads(self.rfile.read(int(self.headers['Content-Length'])))
- self.calls.append(body)
- tool_messages = [m for m in body['messages'] if m.get('role') == 'tool']
- if not tool_messages:
- call = {'id': 'call-legacy', 'type': 'function', 'function': {'name': 'memory', 'arguments': '{"action":"add","content":"x"}'}}
- elif len(tool_messages) == 1:
- call = {'id': 'call-workspace', 'type': 'function', 'function': {'name': 'workspace_memory_search', 'arguments': '{"scope":"agent"}'}}
- else:
- call = None
- message = {'role': 'assistant', 'content': None, 'tool_calls': [call]} if call else {'role': 'assistant', 'content': '记忆检查完成。'}
- finish = 'tool_calls' if call else 'stop'
- self.send_response(200)
- self.send_header('Content-Type', 'text/event-stream' if body.get('stream') else 'application/json')
- self.end_headers()
- if body.get('stream'):
- delta = json.loads(json.dumps(message))
- if 'tool_calls' in delta:
- delta['tool_calls'][0]['index'] = 0
- for part in [{'id':'test','object':'chat.completion.chunk','model':'test-model','choices':[{'index':0,'delta':delta,'finish_reason':None}]},
- {'id':'test','object':'chat.completion.chunk','model':'test-model','choices':[{'index':0,'delta':{},'finish_reason':finish}],'usage':{'prompt_tokens':100,'completion_tokens':20,'total_tokens':120}}]:
- self.wfile.write(('data: '+json.dumps(part)+'\n\n').encode())
- self.wfile.write(b'data: [DONE]\n\n')
- else:
- self.wfile.write(json.dumps({'id':'test','object':'chat.completion','model':'test-model','choices':[{'index':0,'message':message,'finish_reason':finish}],'usage':{'prompt_tokens':100,'completion_tokens':20,'total_tokens':120}}).encode())
- class KernelTest(unittest.TestCase):
- def exchange(self, request, *, interrupt_at_tool=False, approve=True, on_frame=None, steer_phase=None):
- with tempfile.TemporaryFile(mode='w+') as errors:
- entry = [str(ROOT/'tests/runtime/steer_phase_worker.py'), steer_phase] if steer_phase else [str(ROOT/'runtime/worker.py')]
- child=subprocess.Popen([str(ROOT/'.venv/bin/python'),'-u',*entry],stdin=subprocess.PIPE,stdout=subprocess.PIPE,stderr=errors,text=True)
- timeout=threading.Timer(40,child.kill);timeout.start();events=[]
- try:
- child.stdin.write(json.dumps(request)+'\n');child.stdin.flush()
- for line in child.stdout:
- frame=json.loads(line);events.append(frame)
- if on_frame:
- on_frame(frame, lambda message: (child.stdin.write(json.dumps(message)+'\n'), child.stdin.flush()))
- if frame['type']=='tool.request':
- if interrupt_at_tool == 'cancel':
- child.stdin.write(json.dumps({'type':'interrupt'})+'\n');child.stdin.flush();continue
- if interrupt_at_tool:
- child.kill();break
- result={'allow':approve} if frame['name'] in ('yaovi_mcp_approve','yaovi_native_authorize') else {'text':'random-proof-762'}
- child.stdin.write(json.dumps({'id':frame['id'],'result':result})+'\n');child.stdin.flush()
- child.wait(timeout=5)
- finally:
- timeout.cancel();child.kill();child.wait(timeout=5);child.stdin.close();child.stdout.close()
- if not interrupt_at_tool:
- errors.seek(0);self.assertEqual(child.returncode,0,errors.read()[-3000:]+str(events))
- return events
- def test_hermes_resolves_provider_from_config_and_system_env(self):
- """No baseURL/apiKey on the route: Hermes resolves custom_providers + the system-layer .env key."""
- seen = []
- class AuthModel(Model):
- def do_POST(self):
- seen.append(self.headers.get('Authorization'))
- super().do_POST()
- Model.calls=[]
- server=ThreadingHTTPServer(('127.0.0.1',0),AuthModel)
- threading.Thread(target=server.serve_forever,daemon=True).start()
- self.addCleanup(server.server_close);self.addCleanup(server.shutdown)
- with tempfile.TemporaryDirectory(prefix='yaovi-hermes-root-') as root, tempfile.TemporaryDirectory(prefix='yaovi-bot-') as directory:
- Path(root, '.env').write_text('PROOF_API_KEY=secret-proof-key\n')
- os.environ['YAOVI_HERMES_ROOT']=root
- self.addCleanup(os.environ.pop,'YAOVI_HERMES_ROOT',None)
- endpoint=f'http://127.0.0.1:{server.server_port}/v1'
- request={'workspaceId':'local','botId':'bot-one','sessionId':'session-hermes','runId':'run-hermes','dataRoot':directory,
- 'route':{'provider':'proof','model':'test-model','reasoningEffort':'none'},
- 'config':{'agent':{'max_iterations':4},'memory':{'enabled':False,'automatic':False},'delegation':{'enabled':False},'compression':{'enabled':False},
- 'model':{'provider':'proof','default':'test-model','context_length':128000},
- 'custom_providers':[{'name':'proof','base_url':endpoint,'key_env':'PROOF_API_KEY','api_mode':'chat_completions'}]},
- 'instructions':'Use proof_read once.','text':'Read proof.txt.',
- 'tools':[{'name':'proof_read','description':'Read a test artifact','parameters':{'type':'object','properties':{'path':{'type':'string'}}}}]}
- events=self.exchange(request)
- self.assertEqual(events[-1]['type'],'completed',events[-3:])
- self.assertIn('Bearer secret-proof-key',seen)
- def test_compiled_mcp_servers_expose_native_hermes_mcp_tools(self):
- """config.mcp_servers (compiled by Yaoyao) connects via Hermes; ${VAR} resolves from the system .env."""
- calls = []
- class McpModel(Model):
- def do_POST(self):
- body = json.loads(self.rfile.read(int(self.headers['Content-Length'])))
- calls.append(body)
- done = any(m.get('role') == 'tool' for m in body['messages'])
- message = {'role': 'assistant', 'content': 'MCP 完成'} if done else {'role': 'assistant', 'content': None,
- 'tool_calls': [{'id': 'call-mcp', 'type': 'function', 'function': {'name': 'mcp__fake__secret_tag', 'arguments': '{}'}}]}
- finish='stop' if done else 'tool_calls'
- self.send_response(200);self.send_header('Content-Type', 'text/event-stream');self.end_headers()
- if 'tool_calls' in message:
- message['tool_calls'][0]['index']=0
- for part in [{'id':'t','object':'chat.completion.chunk','model':'test-model','choices':[{'index':0,'delta':message,'finish_reason':None}]},
- {'id':'t','object':'chat.completion.chunk','model':'test-model','choices':[{'index':0,'delta':{},'finish_reason':finish}],'usage':{'prompt_tokens':1,'completion_tokens':1,'total_tokens':2}}]:
- self.wfile.write(('data: '+json.dumps(part)+'\n\n').encode())
- self.wfile.write(b'data: [DONE]\n\n')
- server=ThreadingHTTPServer(('127.0.0.1',0),McpModel)
- threading.Thread(target=server.serve_forever,daemon=True).start()
- self.addCleanup(server.server_close);self.addCleanup(server.shutdown)
- with tempfile.TemporaryDirectory(prefix='yaovi-hermes-root-') as root, tempfile.TemporaryDirectory(prefix='yaovi-bot-') as directory:
- Path(root, '.env').write_text('MCP_FAKE_SECRET=s3cret\n')
- os.environ['YAOVI_HERMES_ROOT']=root
- self.addCleanup(os.environ.pop,'YAOVI_HERMES_ROOT',None)
- request={'workspaceId':'local','botId':'bot-one','sessionId':'session-mcp','runId':'run-mcp','dataRoot':directory,
- 'route':{'provider':'openai','apiMode':'chat_completions','model':'test-model','baseURL':f'http://127.0.0.1:{server.server_port}/v1','apiKey':'test','reasoningEffort':'none'},
- 'config':{'agent':{'max_iterations':4},'memory':{'memory_enabled':False,'user_profile_enabled':False,'automatic':False},'delegation':{'enabled':False},'model':{'context_length':128000},'compression':{'enabled':False},
- 'mcp_servers':{'fake':{'command':str(ROOT/'.venv/bin/python'),'args':[str(ROOT/'tests/fixtures/fake_mcp_server.py')],'env':{'FAKE_MCP_SECRET':'${MCP_FAKE_SECRET}'},'tools':{'include':['secret_tag']}},
- 'off':{'command':'/nonexistent','enabled':False}}},
- 'instructions':'Use the MCP tool.','text':'Check the secret.','tools':[]}
- events=self.exchange(request)
- self.assertEqual(events[-1]['type'],'completed',events[-3:])
- names={t['function']['name'] for t in calls[0]['tools']}
- self.assertIn('mcp__fake__secret_tag',names)
- self.assertNotIn('mcp__fake__echo',names)
- tool_msgs=[m for m in calls[-1]['messages'] if m.get('role')=='tool']
- self.assertIn('yes',json.dumps(tool_msgs))
- self.assertEqual([e['arguments']['tool'] for e in events if e['type']=='tool.request'],['mcp__fake__secret_tag'])
- # Denied: the server is never called and the model sees the refusal.
- calls.clear()
- events=self.exchange({**request,'sessionId':'session-mcp-deny','runId':'run-mcp-deny'},approve=False)
- denied=json.dumps([m for m in calls[-1]['messages'] if m.get('role')=='tool'],ensure_ascii=False)
- self.assertIn('未批准',denied)
- self.assertNotIn('yes',denied)
- # Managed policy checks the native entry and the existing MCP guard
- # with the same model call identity, so one grant is not duplicated.
- calls.clear()
- events=self.exchange({**request,'botPolicy':{'enabled':True,'foreground':False},'sessionId':'session-mcp-managed','runId':'run-mcp-managed'})
- admissions=[e for e in events if e['type']=='tool.request']
- self.assertEqual([e['name'] for e in admissions],['yaovi_native_authorize','yaovi_mcp_approve'])
- self.assertEqual([e['modelCallId'] for e in admissions],['call-mcp','call-mcp'])
- self.assertIn('yes',json.dumps([m for m in calls[-1]['messages'] if m.get('role')=='tool']))
- def steer_request(self, directory, endpoint, session, **extra):
- return {'workspaceId':'local','botId':'bot-one','sessionId':session,'runId':'run-'+session,'dataRoot':directory,
- 'route':{'model':'test-model','baseURL':endpoint,'apiKey':'test','reasoningEffort':'none'},
- 'config':{'agent':{'max_iterations':4},'memory':{'enabled':False,'automatic':False},'delegation':{'enabled':False},'model':{'context_length':128000},'compression':{'enabled':False}},
- 'instructions':'Use proof_read once.','text':'Read proof.txt.',
- 'tools':[{'name':'proof_read','description':'Read a test artifact','parameters':{'type':'object','properties':{'path':{'type':'string'}}}}],**extra}
- def test_steer_during_tool_batch_is_inserted_as_user_row_without_interrupting(self):
- """Hermes AIAgent.steer(): accepted mid-run, lands as its own user row after the tool batch, turn completes normally."""
- Model.calls=[]
- server=ThreadingHTTPServer(('127.0.0.1',0),Model)
- threading.Thread(target=server.serve_forever,daemon=True).start()
- self.addCleanup(server.server_close);self.addCleanup(server.shutdown)
- with tempfile.TemporaryDirectory(prefix='yaovi-steer-') as directory:
- def steer_at_tool(frame, send):
- if frame['type']=='tool.request' and frame['name']=='proof_read':
- send({'type':'steer','id':'steer-1','text':'顺便:回复用英文'})
- events=self.exchange(self.steer_request(directory,f'http://127.0.0.1:{server.server_port}/v1','session-steer'),on_frame=steer_at_tool)
- results=[e for e in events if e['type']=='steer.result']
- self.assertEqual([(r['id'],r['accepted']) for r in results],[('steer-1',True)],events)
- self.assertEqual(events[-1]['type'],'completed',events[-3:])
- self.assertFalse([e for e in events if e['type']=='steer.leftover'],'absorbed steer must not be re-delivered')
- last=Model.calls[-1]['messages']
- user_rows=[m for m in last if m.get('role')=='user']
- self.assertTrue(any('顺便:回复用英文' in json.dumps(m,ensure_ascii=False) for m in user_rows[1:]),last)
- self.assertEqual(last[-1]['role'],'user')
- def test_model_time_steer_continues_same_turn_and_preserves_completed_tools(self):
- """Real Hermes redirects before/during text, including repeated inserts and a completed tool."""
- for phase in ('before_text', 'during_text', 'after_tool', 'repeated', 'non_streaming'):
- with self.subTest(phase=phase):
- before_text = phase in ('before_text', 'non_streaming')
- started = threading.Event()
- acknowledged = [threading.Event(), threading.Event()]
- corrections = ['补一句:用中文', '再补一句:只说验证成功']
- expected = corrections if phase == 'repeated' else corrections[:1]
- class RedirectModel(Model):
- calls = []
- def do_POST(self):
- body = json.loads(self.rfile.read(int(self.headers['Content-Length'])))
- self.calls.append(body)
- rows = body['messages']
- correction_count = sum(any(text in str(row.get('content', '')) for row in rows if row['role'] == 'user') for text in expected)
- tool = phase == 'after_tool' and not any(row['role'] == 'tool' for row in rows)
- wait_for_insert = not tool and correction_count < len(expected)
- if wait_for_insert:
- started.set()
- if before_text:
- acknowledged[correction_count].wait(5)
- message = {'role': 'assistant', 'content': '验证成功'}
- if tool:
- message = {'role': 'assistant', 'content': None, 'tool_calls': [
- {'index': 0, 'id': 'proof-once', 'type': 'function', 'function': {'name': 'proof_read', 'arguments': '{}'}}]}
- elif wait_for_insert:
- message['content'] = f'正在生成旧回答{correction_count}'
- try:
- self.send_response(200)
- self.send_header('Content-Type', 'text/event-stream' if body.get('stream') else 'application/json')
- self.end_headers()
- if not body.get('stream'):
- self.wfile.write(json.dumps({'id':'t','object':'chat.completion','model':'test-model',
- 'choices':[{'index':0,'message':message,'finish_reason':'stop'}]}).encode())
- return
- def delta(value, finish=None):
- self.wfile.write(('data: '+json.dumps({'id':'t','object':'chat.completion.chunk','model':'test-model',
- 'choices':[{'index':0,'delta':value,'finish_reason':finish}]})+'\n\n').encode())
- self.wfile.flush()
- delta(message)
- if wait_for_insert and not before_text:
- acknowledged[correction_count].wait(5)
- delta({}, 'tool_calls' if tool else 'stop')
- self.wfile.write(b'data: [DONE]\n\n');self.wfile.flush()
- except (BrokenPipeError, ConnectionResetError):
- pass # The superseded model request is intentionally closed.
- server = ThreadingHTTPServer(('127.0.0.1', 0), RedirectModel)
- threading.Thread(target=server.serve_forever, daemon=True).start()
- try:
- with tempfile.TemporaryDirectory(prefix='yaovi-redirect-') as directory:
- sent = set()
- sender = None
- def on_frame(frame, send):
- nonlocal sender
- if before_text and frame['type'] == 'ready':
- def send_before_text():
- if started.wait(10):
- send({'type':'steer','id':'insert-0','text':expected[0]})
- sender = threading.Thread(target=send_before_text, daemon=True);sender.start()
- if not before_text and frame['type'] == 'message.delta':
- for index, correction in enumerate(expected):
- if f'正在生成旧回答{index}' in (frame.get('text') or '') and index not in sent:
- sent.add(index)
- send({'type':'steer','id':f'insert-{index}','text':correction})
- if frame['type'] == 'steer.result':
- acknowledged[int(frame['id'].split('-')[-1])].set()
- request = self.steer_request(directory, f'http://127.0.0.1:{server.server_port}/v1', 'session-'+phase,
- instructions='Read proof once then answer.' if phase == 'after_tool' else 'Answer in plain text.', text='你好')
- # A redirect reuses the current iteration, even at the one-iteration limit.
- if phase != 'after_tool':
- request['tools'] = []
- request['config']['agent']['max_iterations'] = 1
- if phase == 'non_streaming':
- request['config']['model']['streaming'] = False
- events = self.exchange(request, on_frame=on_frame)
- if sender:
- sender.join(1)
- self.assertEqual([(e['id'], e['accepted']) for e in events if e['type']=='steer.result'],
- [(f'insert-{i}', True) for i in range(len(expected))], events)
- self.assertFalse([e for e in events if e['type']=='steer.leftover'], events)
- self.assertEqual([e['text'] for e in events if e['type']=='completed'], ['验证成功'])
- self.assertEqual(len(RedirectModel.calls), len(expected) + 1 + int(phase == 'after_tool'))
- self.assertEqual(bool(RedirectModel.calls[-1].get('stream')), phase != 'non_streaming')
- last = RedirectModel.calls[-1]['messages']
- for correction in expected:
- self.assertEqual(sum(correction in str(row.get('content', '')) for row in last if row['role']=='user'), 1, last)
- tool_calls = [e for e in events if e['type']=='tool.request' and e['name']=='proof_read']
- self.assertEqual(len(tool_calls), 1 if phase == 'after_tool' else 0)
- with sqlite3.connect(Path(directory, 'state.db')) as db:
- saved = db.execute("select content from messages where role='user'").fetchall()
- for correction in expected:
- self.assertEqual(sum(correction in row[0] for row in saved), 1, saved)
- finally:
- for gate in acknowledged:
- gate.set()
- server.shutdown();server.server_close()
- def test_steer_covers_initialization_preparation_and_reply_commit(self):
- """Pause the real worker at exact lifecycle boundaries, without relying on model latency."""
- for phase in ('initializing', 'ready', 'preparing', 'settling', 'committed'):
- with self.subTest(phase=phase):
- correction = '补充要求:只回复准备阶段验证成功'
- class PreparationModel(Model):
- calls = []
- def do_POST(self):
- body = json.loads(self.rfile.read(int(self.headers['Content-Length'])))
- self.calls.append(body)
- inserted = any(correction in str(m.get('content', '')) for m in body['messages'] if m['role']=='user')
- self.send_response(200);self.send_header('Content-Type','application/json');self.end_headers()
- self.wfile.write(json.dumps({'id':'t','object':'chat.completion','model':'test-model','choices':[
- {'index':0,'message':{'role':'assistant','content':'准备阶段验证成功' if inserted else '原始回答'},'finish_reason':'stop'}]}).encode())
- server = ThreadingHTTPServer(('127.0.0.1',0),PreparationModel)
- threading.Thread(target=server.serve_forever,daemon=True).start()
- try:
- with tempfile.TemporaryDirectory(prefix='yaovi-preparing-') as directory:
- def on_frame(frame, send):
- if frame['type']=='test.steer_phase':
- send({'type':'steer','id':'phase-insert','text':correction})
- request = self.steer_request(directory,f'http://127.0.0.1:{server.server_port}/v1','session-'+phase,
- tools=[],text='原始问题',instructions='Answer in plain text.')
- request['config']['model']['streaming'] = False
- request['config']['agent']['max_iterations'] = 1
- events = self.exchange(request,on_frame=on_frame,steer_phase=phase)
- accepted = phase != 'committed'
- self.assertEqual([e['accepted'] for e in events if e['type']=='steer.result'],[accepted],events)
- self.assertFalse([e for e in events if e['type']=='steer.leftover'],events)
- self.assertEqual([e['text'] for e in events if e['type']=='completed'],
- ['准备阶段验证成功' if accepted else '原始回答'])
- self.assertEqual(len(PreparationModel.calls),2 if phase=='settling' else 1)
- last = PreparationModel.calls[-1]['messages']
- self.assertEqual(sum(correction in str(m.get('content','')) for m in last if m['role']=='user'),int(accepted),last)
- with sqlite3.connect(Path(directory,'state.db')) as db:
- saved = db.execute("select content from messages where role='user'").fetchall()
- self.assertEqual(sum(correction in row[0] for row in saved),int(accepted),saved)
- finally:
- server.shutdown();server.server_close()
- def test_restart_restores_history_without_replaying_completed_tool(self):
- Model.calls=[]
- server=ThreadingHTTPServer(('127.0.0.1',0),Model)
- threading.Thread(target=server.serve_forever,daemon=True).start()
- self.addCleanup(server.server_close);self.addCleanup(server.shutdown)
- with tempfile.TemporaryDirectory(prefix='yaovi-resume-') as directory:
- request={'workspaceId':'local','botId':'bot-one','sessionId':'session-resume','runId':'run-one','dataRoot':directory,
- 'route':{'model':'test-model','baseURL':f'http://127.0.0.1:{server.server_port}/v1','apiKey':'test','reasoningEffort':'none'},
- 'config':{'agent':{'max_iterations':4},'memory':{'enabled':False,'automatic':False},'delegation':{'enabled':False},'model':{'context_length':128000},'compression':{'enabled':False}},
- 'instructions':'Use proof_read once.','text':'Read proof.txt.',
- 'tools':[{'name':'proof_read','description':'Read a test artifact','parameters':{'type':'object','properties':{'path':{'type':'string'}}}}]}
- first=self.exchange(request)
- # Resume an actual pre-rename persisted prompt. Only the two built-in
- # identity/help blocks may change; tools, history and quoted names stay.
- constants={}
- source=ast.parse((ROOT/'runtime/hermes/agent/prompt_builder.py').read_text())
- for node in source.body:
- if isinstance(node,ast.Assign) and isinstance(node.targets[0],ast.Name) and node.targets[0].id in ('DEFAULT_AGENT_IDENTITY','HERMES_AGENT_HELP_GUIDANCE_NO_SKILLS'):
- constants[node.targets[0].id]=ast.literal_eval(node.value)
- with sqlite3.connect(Path(directory)/'state.db') as db:
- prompt=db.execute('SELECT p.prompt FROM sessions s JOIN system_prompts p ON p.hash=s.system_prompt_hash WHERE s.id=?',(request['sessionId'],)).fetchone()[0]
- self.assertTrue(prompt.startswith('You are Yaoyao (夭夭 AI).'))
- legacy=constants['DEFAULT_AGENT_IDENTITY']+'\n\n'+constants['HERMES_AGENT_HELP_GUIDANCE_NO_SKILLS']+'\n\n'+prompt.split('\n\n',2)[2]+'\n\nHistorical note: Hermes is a source dependency.'
- digest=hashlib.sha256(legacy.encode()).hexdigest()
- db.execute('INSERT INTO system_prompts(hash,prompt) VALUES(?,?)',(digest,legacy))
- db.execute('UPDATE sessions SET system_prompt_hash=? WHERE id=?',(digest,request['sessionId']))
- request.update(runId='run-two',text='Recall the previous result.')
- second=self.exchange(request)
- self.assertEqual(sum(e['type']=='tool.request' for e in first),1)
- self.assertEqual(sum(e['type']=='tool.request' for e in second),0)
- self.assertIn('random-proof-762',json.dumps(Model.calls[-1]))
- users=[m for m in Model.calls[-1]['messages'] if m['role']=='user']
- self.assertEqual(sum(m.get('content')=='Recall the previous result.' for m in users),1)
- system='\n'.join(m['content'] for m in Model.calls[-1]['messages'] if m['role'] in ('system','developer'))
- self.assertIn('You are Yaoyao (夭夭 AI).',system)
- self.assertNotIn('You are Hermes Agent',system)
- self.assertNotIn('You run on Hermes Agent',system)
- self.assertIn('Historical note: Hermes is a source dependency.',system)
- def test_restart_marks_unconfirmed_tool_unknown_without_replay(self):
- Model.calls=[]
- server=ThreadingHTTPServer(('127.0.0.1',0),Model)
- threading.Thread(target=server.serve_forever,daemon=True).start()
- self.addCleanup(server.server_close);self.addCleanup(server.shutdown)
- with tempfile.TemporaryDirectory(prefix='yaovi-unknown-') as directory:
- request={'workspaceId':'local','botId':'bot-one','sessionId':'session-unknown','runId':'run-one','dataRoot':directory,
- 'route':{'model':'test-model','baseURL':f'http://127.0.0.1:{server.server_port}/v1','apiKey':'test','reasoningEffort':'none'},
- 'config':{'agent':{'max_iterations':4},'memory':{'enabled':False,'automatic':False},'delegation':{'enabled':False},'model':{'context_length':128000},'compression':{'enabled':False}},
- 'instructions':'Use proof_read once.','text':'Read proof.txt.',
- 'tools':[{'name':'proof_read','description':'Read a test artifact','parameters':{'type':'object','properties':{'path':{'type':'string'}}}}]}
- first=self.exchange(request,interrupt_at_tool=True)
- self.assertTrue(any(e['type']=='tool.request' for e in first))
- request.update(runId='run-two',text='Inspect the previous outcome before proceeding.')
- second=self.exchange(request)
- self.assertFalse(any(e['type']=='tool.request' for e in second))
- self.assertTrue(any('outcomeUnknown' in str(m.get('content')) for m in Model.calls[-1]['messages'] if m['role']=='tool'))
- def test_cancel_releases_pending_rpc_and_saves_interrupted_checkpoint(self):
- Model.calls=[]
- server=ThreadingHTTPServer(('127.0.0.1',0),Model)
- threading.Thread(target=server.serve_forever,daemon=True).start()
- self.addCleanup(server.server_close);self.addCleanup(server.shutdown)
- with tempfile.TemporaryDirectory(prefix='yaovi-cancel-') as directory:
- request={'workspaceId':'local','botId':'bot-one','sessionId':'session-cancel','runId':'run-one','dataRoot':directory,
- 'route':{'model':'test-model','baseURL':f'http://127.0.0.1:{server.server_port}/v1','apiKey':'test','reasoningEffort':'none'},
- 'config':{'agent':{'max_iterations':4},'memory':{'enabled':False,'automatic':False},'delegation':{'enabled':False},'model':{'context_length':128000},'compression':{'enabled':False}},
- 'instructions':'Use proof_read once.','text':'Read proof.txt.',
- 'tools':[{'name':'proof_read','description':'Read a test artifact','parameters':{'type':'object','properties':{'path':{'type':'string'}}}}]}
- started=time.monotonic();events=self.exchange(request,interrupt_at_tool='cancel')
- self.assertLess(time.monotonic()-started,10)
- self.assertTrue(any(e['type']=='failed' for e in events))
- checkpoint=json.loads((Path(directory)/'contexts/session-cancel.json').read_text())
- self.assertEqual(checkpoint['status'],'interrupted')
- def test_real_loop_routes_tool_and_persists_context(self):
- Model.calls=[]
- server=ThreadingHTTPServer(('127.0.0.1',0),Model)
- threading.Thread(target=server.serve_forever,daemon=True).start()
- self.addCleanup(server.server_close)
- self.addCleanup(server.shutdown)
- with tempfile.TemporaryDirectory(prefix='yaovi-kernel-') as directory:
- request={'workspaceId':'local','botId':'bot-one','sessionId':'session-one','runId':'run-one','dataRoot':directory,
- 'route':{'provider':'openai','apiMode':'chat_completions','model':'test-model','baseURL':f'http://127.0.0.1:{server.server_port}/v1','apiKey':'test','reasoningEffort':'none'},
- 'config':{'agent':{'max_iterations':4},'memory':{'memory_enabled':False,'user_profile_enabled':False,'automatic':False},'delegation':{'enabled':False},'model':{'context_length':128000},'compression':{'enabled':False}},
- 'instructions':'Use the proof tool once.','text':'Read proof.txt.',
- 'tools':[{'name':'proof_read','description':'Read a test artifact','parameters':{'type':'object','properties':{'path':{'type':'string'}},'required':['path']}}]}
- with tempfile.TemporaryFile(mode='w+') as errors:
- child=subprocess.Popen([str(ROOT/'.venv/bin/python'),'-u',str(ROOT/'runtime/worker.py')],stdin=subprocess.PIPE,stdout=subprocess.PIPE,stderr=errors,text=True)
- timeout=threading.Timer(40,child.kill);timeout.start()
- events=[]
- try:
- child.stdin.write(json.dumps(request)+'\n');child.stdin.flush()
- for line in child.stdout:
- frame=json.loads(line);events.append(frame)
- if frame['type']=='tool.request':
- self.assertEqual(frame['name'],'proof_read')
- child.stdin.write(json.dumps({'id':frame['id'],'result':{'text':'random-proof-762'}})+'\n');child.stdin.flush()
- child.wait(timeout=5)
- finally:
- timeout.cancel();child.kill();child.stdin.close();child.stdout.close()
- errors.seek(0)
- self.assertEqual(child.returncode,0,errors.read()[-6000:]+'\n'+str(events))
- self.assertEqual([e['type'] for e in events].count('tool.request'),1)
- self.assertTrue(any(e['type']=='completed' and '执行完成' in e['text'] for e in events))
- self.assertTrue((Path(directory)/'contexts/session-one.json').exists())
- self.assertIn('random-proof-762',json.dumps(Model.calls[-1]))
- self.assertEqual({t['function']['name'] for t in Model.calls[0]['tools']},{'proof_read','clarify'})
- def test_legacy_memory_executor_is_retired_without_a_bridge(self):
- code=("import sys,json;sys.path.insert(0,'runtime/hermes');sys.path.insert(0,'runtime')\n"
- "from agent import runtime_context as rc;import agent.inline_tool_executors as e\n"
- "calls=[];rc._context={'tools':[]};rc._dispatch=lambda *a,**k:calls.append(a)\n"
- "class A:session_id='s';_memory_store=None;_memory_manager=None\n"
- "r=json.loads(e._memory(A(),{'action':'add','content':'x'},e.InlineToolContext(effective_task_id='t',tool_call_id='c')))\n"
- "print(json.dumps({'r':r,'calls':len(calls)}))")
- with tempfile.TemporaryDirectory() as home:
- out=subprocess.run([sys.executable,'-c',code],cwd=ROOT,capture_output=True,text=True,env={**os.environ,'HERMES_HOME':home,'YAOVI_KERNEL':'1'},check=True)
- result=json.loads(out.stdout.strip().splitlines()[-1])
- self.assertEqual(result['calls'],0)
- self.assertFalse(result['r']['success'])
- self.assertIn("tool 'memory' is retired",result['r']['error'])
- def test_only_workspace_memory_tools_and_legacy_memory_call_is_rejected(self):
- LegacyMemoryModel.calls=[]
- server=ThreadingHTTPServer(('127.0.0.1',0),LegacyMemoryModel)
- threading.Thread(target=server.serve_forever,daemon=True).start()
- self.addCleanup(server.server_close)
- self.addCleanup(server.shutdown)
- workspace_tools=[{'name':n,'description':n,'parameters':{'type':'object','properties':{'scope':{'type':'string'}}}} for n in ('workspace_memory_search','workspace_memory_write','workspace_memory_forget')]
- with tempfile.TemporaryDirectory(prefix='yaovi-kernel-') as directory:
- request={'workspaceId':'local','botId':'bot-one','sessionId':'session-memory','runId':'run-memory','dataRoot':directory,
- 'route':{'provider':'openai','apiMode':'chat_completions','model':'test-model','baseURL':f'http://127.0.0.1:{server.server_port}/v1','apiKey':'test','reasoningEffort':'none'},
- 'config':{'agent':{'max_iterations':6},'memory':{'enabled':True,'automatic':True},'delegation':{'enabled':False},'model':{'context_length':128000},'compression':{'enabled':False}},
- 'instructions':'Check memory.','text':'Remember something.','tools':workspace_tools}
- events=self.exchange(request)
- offered={t['function']['name'] for t in LegacyMemoryModel.calls[0]['tools']}
- self.assertNotIn('memory',offered)
- self.assertTrue({'workspace_memory_search','workspace_memory_write','workspace_memory_forget'} <= offered)
- # The retired tool never reaches the control plane; only the workspace call does.
- self.assertEqual([e['name'] for e in events if e['type']=='tool.request'],['workspace_memory_search'])
- legacy=[m for m in LegacyMemoryModel.calls[1]['messages'] if m.get('role')=='tool' and m.get('tool_call_id')=='call-legacy']
- self.assertEqual(len(legacy),1)
- self.assertRegex(str(legacy[0]['content']).lower(),"tool 'memory' does not exist")
- self.assertIn('workspace_memory_write',str(legacy[0]['content']))
- self.assertTrue(any(e['type']=='completed' for e in events))
- system='\n'.join(str(m['content']) for m in LegacyMemoryModel.calls[0]['messages'] if m['role'] in ('system','developer'))
- self.assertIn('workspace_memory_search',system)
- self.assertFalse((Path(directory)/'memories').exists())
- self.assertFalse(any(Path(directory).rglob('MEMORY.md')) or any(Path(directory).rglob('USER.md')))
- self.assertNotIn('MEMORY.md (memory tool',system)
- if __name__=='__main__':
- unittest.main()
|