Remove MAX_CHARS traffic control (All-Hands-AI#2694)

* Remove MAX_CHARS limiting

* More cleanup

agenthub/codeact_agent/codeact_agent.py

@@ -226,9 +226,6 @@ def step(self, state: State) -> Action:
             ],
             temperature=0.0,
         )
-        state.num_of_chars += sum(
-            len(message['content']) for message in messages
-        ) + len(response.choices[0].message.content)
         return self.action_parser.parse(response)
 
     def search_memory(self, query: str) -> list[str]:

agenthub/codeact_swe_agent/codeact_swe_agent.py

@@ -181,9 +181,6 @@ def step(self, state: State) -> Action:
             ],
             temperature=0.0,
         )
-        state.num_of_chars += sum(
-            len(message['content']) for message in messages
-        ) + len(response.choices[0].message.content)
 
         return self.response_parser.parse(response)
 

agenthub/micro/agent.py

@@ -66,7 +66,6 @@ def step(self, state: State) -> Action:
         messages = [{'content': prompt, 'role': 'user'}]
         resp = self.llm.completion(messages=messages)
         action_resp = resp['choices'][0]['message']['content']
-        state.num_of_chars += len(prompt) + len(action_resp)
         action = parse_response(action_resp)
         return action
 

agenthub/monologue_agent/agent.py

@@ -183,11 +183,6 @@ def step(self, state: State) -> Action:
         # format all as a single message, a monologue
         resp = self.llm.completion(messages=messages)
 
-        # keep track of max_chars fallback option
-        state.num_of_chars += len(prompt) + len(
-            resp['choices'][0]['message']['content']
-        )
-
         action = self.response_parser.parse(resp)
         self.latest_action = action
         return action

agenthub/planner_agent/agent.py

@@ -48,9 +48,6 @@ def step(self, state: State) -> Action:
         prompt = get_prompt(state)
         messages = [{'content': prompt, 'role': 'user'}]
         resp = self.llm.completion(messages=messages)
-        state.num_of_chars += len(prompt) + len(
-            resp['choices'][0]['message']['content']
-        )
         return self.response_parser.parse(resp)
 
     def search_memory(self, query: str) -> list[str]:

opendevin/controller/agent_controller.py

@@ -9,7 +9,6 @@
     LLMMalformedActionError,
     LLMNoActionError,
     LLMResponseError,
-    MaxCharsExceedError,
 )
 from opendevin.core.logger import opendevin_logger as logger
 from opendevin.core.schema import AgentState
@@ -37,7 +36,6 @@
 )
 
 MAX_ITERATIONS = config.max_iterations
-MAX_CHARS = config.llm.max_chars
 MAX_BUDGET_PER_TASK = config.max_budget_per_task
 
 
@@ -58,7 +56,6 @@ def __init__(
         event_stream: EventStream,
         sid: str = 'default',
         max_iterations: int = MAX_ITERATIONS,
-        max_chars: int = MAX_CHARS,
         max_budget_per_task: float | None = MAX_BUDGET_PER_TASK,
         initial_state: State | None = None,
         is_delegate: bool = False,
@@ -70,15 +67,13 @@ def __init__(
             event_stream: The event stream to publish events to.
             sid: The session ID of the agent.
             max_iterations: The maximum number of iterations the agent can run.
-            max_chars: The maximum number of characters the agent can output.
             max_budget_per_task: The maximum budget (in USD) allowed per task, beyond which the agent will stop.
             initial_state: The initial state of the controller.
             is_delegate: Whether this controller is a delegate.
         """
         self._step_lock = asyncio.Lock()
         self.id = sid
         self.agent = agent
-        self.max_chars = max_chars
         if initial_state is None:
             self.state = State(inputs={}, max_iterations=max_iterations)
         else:
@@ -224,7 +219,6 @@ async def start_delegate(self, action: AgentDelegateAction):
             inputs=action.inputs or {},
             iteration=0,
             max_iterations=self.state.max_iterations,
-            num_of_chars=self.state.num_of_chars,
             delegate_level=self.state.delegate_level + 1,
             # metrics should be shared between parent and child
             metrics=self.state.metrics,
@@ -235,7 +229,6 @@ async def start_delegate(self, action: AgentDelegateAction):
             agent=agent,
             event_stream=self.event_stream,
             max_iterations=self.state.max_iterations,
-            max_chars=self.max_chars,
             max_budget_per_task=self.max_budget_per_task,
             initial_state=state,
             is_delegate=True,
@@ -298,9 +291,6 @@ async def _step(self):
                 await self.event_stream.add_event(obs, EventSource.AGENT)
             return
 
-        if self.state.num_of_chars > self.max_chars:
-            raise MaxCharsExceedError(self.state.num_of_chars, self.max_chars)
-
         logger.info(
             f'{self.agent.name} LEVEL {self.state.delegate_level} STEP {self.state.iteration}',
             extra={'msg_type': 'STEP'},

opendevin/controller/state/state.py

@@ -29,8 +29,6 @@ class State:
     root_task: RootTask = field(default_factory=RootTask)
     iteration: int = 0
     max_iterations: int = 100
-    # number of characters we have sent to and received from LLM so far for current task
-    num_of_chars: int = 0
     background_commands_obs: list[CmdOutputObservation] = field(default_factory=list)
     history: list[tuple[Action, Observation]] = field(default_factory=list)
     updated_info: list[tuple[Action, Observation]] = field(default_factory=list)

opendevin/core/config.py

@@ -39,7 +39,6 @@ class LLMConfig(metaclass=Singleton):
         retry_min_wait: The minimum time to wait between retries, in seconds. This is exponential backoff minimum. For models with very low limits, this can be set to 15-20.
         retry_max_wait: The maximum time to wait between retries, in seconds. This is exponential backoff maximum.
         timeout: The timeout for the API.
-        max_chars: The maximum number of characters to send to and receive from the API. This is a fallback for token counting, which doesn't work in all cases.
         temperature: The temperature for the API.
         top_p: The top p for the API.
         custom_llm_provider: The custom LLM provider to use. This is undocumented in opendevin, and normally not used. It is documented on the litellm side.
@@ -63,7 +62,6 @@ class LLMConfig(metaclass=Singleton):
     retry_min_wait: int = 3
     retry_max_wait: int = 60
     timeout: int | None = None
-    max_chars: int = 5_000_000  # fallback for token counting
     temperature: float = 0
     top_p: float = 0.5
     custom_llm_provider: str | None = None
@@ -523,13 +521,6 @@ def get_parser():
         type=float,
         help='The maximum budget allowed per task, beyond which the agent will stop.',
     )
-    parser.add_argument(
-        '-n',
-        '--max-chars',
-        default=config.llm.max_chars,
-        type=int,
-        help='The maximum number of characters to send to and receive from LLM per task',
-    )
     # --eval configs are for evaluations only
     parser.add_argument(
         '--eval-output-dir',

opendevin/core/exceptions.py

@@ -1,12 +1,3 @@
-class MaxCharsExceedError(Exception):
-    def __init__(self, num_of_chars=None, max_chars_limit=None):
-        if num_of_chars is not None and max_chars_limit is not None:
-            message = f'Number of characters {num_of_chars} exceeds MAX_CHARS limit: {max_chars_limit}'
-        else:
-            message = 'Number of characters exceeds MAX_CHARS limit'
-        super().__init__(message)
-
-
 class AgentNoInstructionError(Exception):
     def __init__(self, message='Instruction must be provided'):
         super().__init__(message)

opendevin/core/main.py

@@ -89,7 +89,6 @@ async def main(
         agent=agent,
         max_iterations=args.max_iterations,
         max_budget_per_task=args.max_budget_per_task,
-        max_chars=args.max_chars,
         event_stream=event_stream,
     )
     runtime = ServerRuntime(event_stream=event_stream, sandbox=sandbox)

opendevin/core/schema/config.py

@@ -32,7 +32,6 @@ class ConfigType(str, Enum):
     AGENT_MEMORY_MAX_THREADS = 'AGENT_MEMORY_MAX_THREADS'
     AGENT_MEMORY_ENABLED = 'AGENT_MEMORY_ENABLED'
     MAX_ITERATIONS = 'MAX_ITERATIONS'
-    MAX_CHARS = 'MAX_CHARS'
     AGENT = 'AGENT'
     E2B_API_KEY = 'E2B_API_KEY'
     SANDBOX_TYPE = 'SANDBOX_TYPE'

opendevin/server/session/agent.py

@@ -91,7 +91,6 @@ async def _create_controller(self, start_event: dict):
         api_key = args.get(ConfigType.LLM_API_KEY, config.llm.api_key)
         api_base = config.llm.base_url
         max_iterations = args.get(ConfigType.MAX_ITERATIONS, config.max_iterations)
-        max_chars = args.get(ConfigType.MAX_CHARS, config.llm.max_chars)
 
         logger.info(f'Creating agent {agent_cls} using LLM {model}')
         llm = LLM(model=model, api_key=api_key, base_url=api_base)
@@ -109,7 +108,6 @@ async def _create_controller(self, start_event: dict):
             event_stream=self.event_stream,
             agent=agent,
             max_iterations=int(max_iterations),
-            max_chars=int(max_chars),
         )
         try:
             agent_state = State.restore_from_session(self.sid)

tests/unit/test_arg_parser.py

@@ -11,7 +11,7 @@ def test_help_message(capsys):
     expected_help_message = """
 usage: pytest [-h] [-d DIRECTORY] [-t TASK] [-f FILE] [-c AGENT_CLS]
               [-m MODEL_NAME] [-i MAX_ITERATIONS] [-b MAX_BUDGET_PER_TASK]
-              [-n MAX_CHARS] [--eval-output-dir EVAL_OUTPUT_DIR]
+              [--eval-output-dir EVAL_OUTPUT_DIR]
               [--eval-n-limit EVAL_N_LIMIT]
               [--eval-num-workers EVAL_NUM_WORKERS] [--eval-note EVAL_NOTE]
               [-l LLM_CONFIG]
@@ -34,9 +34,6 @@ def test_help_message(capsys):
   -b MAX_BUDGET_PER_TASK, --max-budget-per-task MAX_BUDGET_PER_TASK
                         The maximum budget allowed per task, beyond which the
                         agent will stop.
-  -n MAX_CHARS, --max-chars MAX_CHARS
-                        The maximum number of characters to send to and
-                        receive from LLM per task
   --eval-output-dir EVAL_OUTPUT_DIR
                         The directory to save evaluation output
   --eval-n-limit EVAL_N_LIMIT