langchain-ai/deepagents · error · RequestError

-32601

-32601

Error message

Method not found: session/load

What it means

`ModelSpec.parse` requires the `provider:model` format and raises `ValueError` when the input has no colon at all. The library uses this strict format everywhere a model is addressed, so a bare model name like `claude-sonnet-4-5` is rejected with a message showing the expected shape.

Source

Thrown at libs/acp/deepagents_acp/server.py:405

        # Return response with both modes (for backward compatibility) and config_options
        return NewSessionResponse(
            session_id=session_id,
            modes=self._modes if self._modes is not None else None,
            config_options=config_options,
        )

    async def load_session(
        self,
        cwd: str,
        session_id: str,
        additional_directories: list[str] | None = None,  # noqa: ARG002  # capability is not advertised
        mcp_servers: list[McpServer] | None = None,
        **kwargs: Any,  # noqa: ARG002  # ACP protocol interface parameter
    ) -> LoadSessionResponse:
        """Restore and replay a persisted ACP session."""
        if not self._load_sessions:
            method = "session/load"
            raise RequestError.method_not_found(method)

        self._session_cwds[session_id] = cwd
        agent = self._checkpointed_agent(session_id)
        metadata = (await agent.aget_state(self._session_config(session_id))).metadata or {}
        if metadata.get(_ACP_SESSION_METADATA_KEY) is not True:
            self._forget_session(session_id)
            raise RequestError.resource_not_found(session_id)
        if metadata.get("cwd") != cwd:
            self._forget_session(session_id)
            raise RequestError.invalid_params(
                {"cwd": "must match the working directory used to create the session"}
            )

        self._session_mcp_servers[session_id] = list(mcp_servers or [])
        self._initialize_session_options(session_id)
        if self._restore_session_options(session_id, metadata):
            self._reset_agent(session_id)

View on GitHub (pinned to a1af029e6e)

Solutions

1. Prefix the provider: `anthropic:claude-sonnet-4-5`
2. If you must accept bare names, resolve them to a spec yourself (e.g. via provider detection) before parsing
3. Use `ModelSpec.try_parse(...)` to validate first

Example fix

// before
spec = ModelSpec.parse("claude-sonnet-4-5")
// after
spec = ModelSpec.parse("anthropic:claude-sonnet-4-5")

def is_parseable_spec(raw: str) -> bool:
    return ModelSpec.try_parse(raw) is not None

if not is_parseable_spec(raw):
    raise ValueError(f"{raw!r} must be 'provider:model'")

def is_model_spec_string(value: object) -> bool:
    return (
        isinstance(value, str)
        and ":" in value
        and bool(value.split(":", 1)[0])
        and bool(value.split(":", 1)[1])
    )

try:
    spec = ModelSpec.parse(raw)
except ValueError:
    detected = detect_provider(raw)  # resolve bare model names if supported
    spec = ModelSpec.parse(f"{detected}:{raw}") if detected else DEFAULT_MODEL_SPEC

When it happens

Trigger: `ModelSpec.parse("claude-sonnet-4-5")` or any string without `:`; called indirectly via `parse`/`try_parse` from `_get_default_model_spec`, `_sdk_version_from_source`, or when an allowlist entry reaches spec parsing.

Common situations: Writing just the model name in `config.toml` (`model = "claude-sonnet-4-5"`) where a spec is expected; assuming auto-detection applies — only some paths (via `detect_provider`) accept bare model names; older configs written before the provider prefix became mandatory.

Related errors

• modes can only be provided when agent is a factory
• models can only be provided when agent is a factory
• -32002
• -32602
• recursion_limit must be None or a positive integer

AI-assisted analysis of langchain-ai/deepagents@a1af029e6e (2026-08-29). Data as JSON: /api/errors/a2d72974812d99dc. Report an issue: GitHub.