FoundationAgents/MetaGPT · error · ValueError

Cannot find the answer phrase "{response}"

Error message

Cannot find the answer phrase "{response}"

What it means

extract_step in metagpt/utils/a11y_tree.py pulls the action out of an LLM response by searching for the first pair of action_splitter fences (default '```'). If the response contains no fenced block, the regex finds nothing and this ValueError fires with the full response embedded in the message.

Source

Thrown at metagpt/utils/a11y_tree.py:174

    except Exception as e:
        raise ValueError("Element not found") from e


async def get_element_center(node_info):
    x, y, width, height = node_info["x"], node_info["y"], node_info["width"], node_info["height"]
    center_x = x + width / 2
    center_y = y + height / 2
    return center_x, center_y


def extract_step(response: str, action_splitter: str = "```") -> str:
    # find the first occurence of action
    pattern = rf"{action_splitter}((.|\n)*?){action_splitter}"
    match = re.search(pattern, response)
    if match:
        return match.group(1).strip()
    else:
        raise ValueError(f'Cannot find the answer phrase "{response}"')


async def get_bounding_rect(cdp_session, backend_node_id: str):
    try:
        remote_object = await cdp_session.send("DOM.resolveNode", {"backendNodeId": int(backend_node_id)})
        remote_object_id = remote_object["object"]["objectId"]
        response = await cdp_session.send(
            "Runtime.callFunctionOn",
            {
                "objectId": remote_object_id,
                "functionDeclaration": """
                    function() {
                        if (this.nodeType == 3) {
                            var range = document.createRange();
                            range.selectNode(this);
                            var rect = range.getBoundingClientRect().toJSON();
                            range.detach();
                            return rect;

View on GitHub (pinned to 11cdf466d0)

Solutions

1. Ensure the LLM response contains the action between triple-backtick fences, e.g. '```click [12]```'.
2. If the response was truncated, increase max_tokens or reduce prompt size so the fenced block completes.
3. Make the prompt explicitly require the fence delimiters around the action.
4. If you already have a bare action string, skip extract_step and pass it straight to execute_action.

Example fix

# before
step = extract_step('click [12]')  # raises: no ``` block

# after
step = extract_step('```click [12]```')  # -> 'click [12]'

def has_action_block(response: str, splitter: str = "```") -> bool:
    return splitter in response

try:
    step = extract_step(response)
except ValueError:
    # response malformed/truncated: re-ask the model, possibly with raised max_tokens

When it happens

Trigger: extract_step('click [12]') — action sent without ``` fences; response truncated by token limits before the closing fence; model wrapping the action in single backticks or quotes; custom action_splitter that never appears in the response.

Common situations: LLM omits the markdown fence, answers in prose ('I would click...'), or gets cut off mid-action by max_tokens. Also occurs when the caller passes a raw action string directly instead of a full model response.

Related errors

• Could not find content between [{tag}] and [/{tag}]
• Invalid python code
• Error while extracting and parsing the {data_type}: {e}
• Expecting property name enclosed in double quotes
• Expecting ':' delimiter

AI-assisted analysis of FoundationAgents/MetaGPT@11cdf466d0 (2026-08-14). Data as JSON: /api/errors/29f152730751bc9f. Report an issue: GitHub.