An LLM error that states its own cause is also told to check the config file #2681

Observed

When a provider refuses a request and writes down why, the daemon reports that sentence and then appends a generic remedy anyway.

The error frame in _run_tool_loop (emrg/server/daemon.py) appends . Check config at ~/.emrg/config.toml to every LLM error except a content-filter refusal, which has its own arm. Since llm.py began extracting the provider's own sentence (provider_error_message, for a JSON body carrying error.message), refusals arrive already answered — and still collect the advice:

LLM error: LLM stream request failed: 402 — 点数余额不足. Check config at ~/.emrg/config.toml

The account is empty; the configuration is fine. One remedy is printed for every cause, so the actionable half of the message competes with a pointer that sends the host to a file that was never wrong. The same holds for an invalid key, a rate limit or an unknown model: each arrives with its own sentence, and each is currently also told to check the config.

Expected

A refusal that states its cause is printed as that sentence, with no generic remedy appended. A failure that says nothing about its cause keeps the pointer — that is exactly when a host needs to know where the configuration lives.

What a fix must satisfy

• The question is a predicate in llm.py, beside the extraction it belongs to (the is_overlong_error shape), never re-derived from the message's spelling at the daemon call site.
• Both raise paths mark it. chat() and chat_stream() each compose a refusal sentence, and the tool loop takes the streaming one while the summariser takes the other.
• The content-filter arm keeps its behaviour and its logging: its abort-run count is keyed to that case, and the new case must not fall into it.
• Two guards, one each way. With the predicate ignored, the "stated cause" arm goes red; with the advice removed wholesale, the "a silent failure keeps the pointer" arm goes red.