server: route unterminated reasoning to reasoning_content, not content

[datanerdie]

Problem

When generation hits the token cap before </think> arrives, the unterminated reasoning buffer is emitted as content. A truncated chain-of-thought therefore reaches clients looking exactly like a finished answer.

The server already detects the condition and logs it:

ds4-server: thinking not closed, ignoring DSML in reasoning

…but the message it then builds is indistinguishable from a normal reply. Observed on DeepSeek-V4-Flash-0731:

response	finish_reason	content	reasoning_content
completed	stop	2,491 B (the answer)	27,631 B
truncated	length	40,786 B of raw reasoning	0 B

Any client that does not inspect finish_reason will treat 40 KB of abandoned deliberation as the assistant's answer. That matters most for agent loops, where a delegate may consume the result without a human ever seeing it.

Why not fix split_reasoning_content()

It cannot make the distinction on its own. An unterminated reasoning buffer and a legitimate non-thinking answer both lack </think>, and the stored text does not retain the opening <think> either — so there is no discriminator in the text. It also has other call sites where the current behaviour is correct; changing it there would route ordinary non-thinking answers into reasoning_content.

The information exists only at the call sites that already know thinking was expected, via require_thinking_closed. That is where this patch acts.

Change

A small helper, used at the two require_thinking_closed && !think_end sites — one in parse_deepseek_generated_message_ex, one in parse_glm_generated_message_ex, so the two parsers stay symmetric:

static void ds4_local_unterminated_reasoning(const char *text,
                                             char **content_out,
                                             char **reasoning_out) {
    const char *body = text ? text : "";
    if (!strncmp(body, "<think>", 7)) body += 7;
    *reasoning_out = xstrdup(body);
    *content_out = xstrdup("");
}

Truncation becomes self-describing on the wire for every consumer, rather than only for those that check finish_reason.

On "" vs NULL

content is set to "" rather than NULL deliberately. json_escape() dereferences its argument without a NULL check; the response writer guards with text ? text : "", but I did not audit every consumer. An empty answer is already unmistakably not an answer, and finish_reason plus a populated reasoning_content carry the rest of the signal. Happy to switch to a literal JSON null if you would prefer that — it is one further line in the response writer.

The helper name carries a ds4_local_ prefix because it began as a local patch; rename as you see fit.

Verification

Apple M4 Max, Metal, DeepSeek-V4-Flash-0731 q2-q4-imatrix:

case	finish	content	reasoning_content
truncated thinking	length	""	333 B
completed thinking	stop	"12 x 12 = 144."	186 B
non-thinking	stop	"12 x 12 = 144."	""

The latter two are unchanged from before the patch — including the non-thinking case, which is the one that would break if split_reasoning_content() were changed directly.

ds4_test (which #includes ds4_server.c, so it compiles the change) reports an identical set of assertion failures with and without this patch on my machine — a diff of the sorted failure lists is empty. Those failures appear to be pre-existing in my environment rather than related to this change. test_q4k_dot (4/4), ds4_agent_test, and ds4-eval --self-test-extractors all pass.

🤖 Generated with Claude Code