iflytek/astron-agent · warning · CustomException

SPARK_QUICK_REPAIR_ERROR

SPARK_QUICK_REPAIR_ERROR

Error message

Sensitive content detected, LLM did not find function_call field

What it means

Raised by SparkFcLLM._recv_messages when the Spark function-call response's message sid ends with '1', which per the Spark API indicates a 'quick repair' — the server-side content auditor intercepted the request and the LLM produced no function_call. The provider converts this into a SPARK_QUICK_REPAIR_ERROR stating sensitive content was detected. It is a content-compliance rejection, not a transport or code bug.

Solutions

1. Treat as an expected user-content rejection: catch SPARK_QUICK_REPAIR_ERROR and return a polite 'content not allowed' response to the end user instead of retrying.
2. Review the offending user input for sensitive keywords and add upstream input moderation/filtering before calling the FC LLM.
3. If false positives occur, review Spark console content-audit settings/patch_id configuration and adjust the auditing parameter or escalate to iFlytek support.
4. Sanitize or truncate user content (strip flagged segments) and retry once with a compliance-safe reformulation.

Example fix

// before
name, usage, args = await async_call_spark_fc(messages, functions)
// after
try:
    name, usage, args = await async_call_spark_fc(messages, functions)
except CustomException as e:
    if e.err_code == CodeEnum.SPARK_QUICK_REPAIR_ERROR:
        return sanitize_refusal_response(messages)  # content was rejected by Spark auditing
    raise

SENSITIVE_HINTS = ("politics", "violence", "porn", ...)  # plug in your moderation list
def screen_user_input(text: str) -> bool:
    return not any(h in text.lower() for h in SENSITIVE_HINTS)
# call before async_call_spark_fc: if not screen_user_input(user_text): return refusal_response()

def is_spark_quick_repair(e: BaseException) -> bool:
    return isinstance(e, CustomException) and e.err_code == CodeEnum.SPARK_QUICK_REPAIR_ERROR

try:
    name, usage, args = await async_call_spark_fc(messages, functions, span)
except CustomException as e:
    if e.err_code == CodeEnum.SPARK_QUICK_REPAIR_ERROR:
        return FunctionCallResult(status="content_rejected", message="Your request was blocked by content moderation")
    raise  # do not retry; retrying flagged content will be rejected again

When it happens

Trigger: Calling SparkFcLLM via async_call_spark_fc with user/assistant message content that Spark's auditing flags as sensitive; the final received frame's header.sid[-1] == '1' triggers the check at spark_fc_llm.py:115 before function_call extraction.

Common situations: End users submitting politically sensitive, violent, or otherwise policy-violating prompts in a tool-calling workflow; prompts containing sensitive keywords in Chinese regulatory context; audit level 'default' in the payload's parameter.chat.auditing being strict for the deployment region.

Related errors

• SPARK_FUNCTION_NOT_CHOICE_ERROR
• SPARK_REQUEST_ERROR
• PERSONALITY_AI_GENERATE_ERROR
• MODEL_CHECK_FAILED
• MODEL_NOT_EXIST

AI-assisted analysis of iflytek/astron-agent@5e758547a8 (2026-09-12). Data as JSON: /api/errors/f011fa64e88ca46b. Report an issue: GitHub.

Appendix: source

Thrown at core/workflow/infra/providers/llm/iflytek_spark/spark_fc_llm.py:116

        :return: Tuple containing (function_name, token_usage, arguments)
        """
        while True:
            try:
                msg = json.loads(await ws_handle.recv())
                await span.add_info_events_async(
                    {"function_call_recv": json.dumps(msg, ensure_ascii=False)}
                )
                code = msg["header"]["code"]
                if code != 0:
                    raise CustomException(
                        err_code=CodeConvert.sparkCode(code),
                        cause_error=json.dumps(msg, ensure_ascii=False),
                    )
                status = msg["header"]["status"]
                llm_service_sid = msg["header"]["sid"]
                # Check if it's a quick repair: if the last character of sid is '1', it's a quick repair
                if llm_service_sid[-1] == "1":
                    raise CustomException(
                        err_code=CodeEnum.SPARK_QUICK_REPAIR_ERROR,
                        err_msg="Sensitive content detected, LLM did not find function_call field",
                        cause_error="Sensitive content detected, LLM did not find function_call field",
                    )
                if status != 2:
                    continue
                token_usage = msg["payload"]["usage"]["text"]
                if "function_call" not in msg["payload"]["choices"]["text"][0]:
                    raise CustomException(
                        err_code=CodeEnum.SPARK_FUNCTION_NOT_CHOICE_ERROR,
                        err_msg="Cannot find function_call field in LLM response",
                        cause_error="Cannot find function_call field in LLM response",
                    )

                name = msg["payload"]["choices"]["text"][0]["function_call"]["name"]
                arguments = msg["payload"]["choices"]["text"][0]["function_call"][
                    "arguments"
                ]

View on GitHub (pinned to 5e758547a8)