Function parse_responses_input

ds4_server.c:3153–3623

Codex /v1/responses input items have a `type` discriminator (message, * function_call, function_call_output, reasoning, custom_tool_call, * custom_tool_call_output, ...). We collapse them into chat_msgs the same way * the chat completion / Anthropic parsers do, so the rest of the engine sees a * single conversation history shape. * * Protocol contract for stateless replay: * - The client

        cursor = marker + strlen(inputs[i]->marker);
    }
    if (ok) ds4_tokenize_rendered_chat(e, cursor, &r->prompt);

done:
    for (size_t i = 0; i < count; i++)
        ds4_vision_embedding_free(&embeddings[i]);
    free(embeddings);
    free(inputs);
    if (!ok) {
        ds4_tokens_free(&r->prompt);
        for (size_t i = 0; i < r->image_count; i++)
            ds4_vision_embedding_free(&r->images[i].embedding);
        free(r->images);
        r->images = NULL;
        r->image_count = 0;
    }
    return ok;
}

/* Render only the semantic tail that must be appended to the live KV for a
 * tool-result continuation.
 *
 * In the common agent tool path, the previous assistant tool-call turn is
 * already in the model session, including hidden thinking and exact sampled
 * DSML.  The next request provides only the tool results, either as OpenAI
 * Responses tool-output items or Anthropic user content blocks.  Re-rendering
 * the assistant call here would duplicate it and destroy cache alignment, so
 * this function starts at the first new item and emits only:
 *
 *   previous EOS, tool results, and the next assistant prefix.
 *
 * This is intentionally independent from req.prompt's already-tokenized suffix:
 * suffix tokenization happens later after the cache decision, using the live
 * token prefix as the boundary.  That avoids BPE merges across the visible
 * replay/live-KV boundary. */
static char *render_deepseek_live_tool_tail(const chat_msgs *msgs, int start,
                                            ds4_think_mode think_mode) {
    const bool think = ds4_think_mode_enabled(think_mode);
    buf out = {0};
    buf_puts(&out, "<｜end▁of▁sentence｜>");

    bool pending_assistant = false;
    bool pending_tool_result = false;
    for (int i = start; msgs && i < msgs->len; i++) {
        const chat_msg *m = &msgs->v[i];
        if (role_is_system(m->role)) {
            continue;
        } else if (!strcmp(m->role, "user")) {
            buf_puts(&out, "<｜User｜>");
            buf_puts(&out, m->content ? m->content : "");
            pending_assistant = true;
            pending_tool_result = false;
        } else if (!strcmp(m->role, "tool") || !strcmp(m->role, "function")) {
            if (!pending_tool_result) buf_puts(&out, "<｜User｜>");
            buf_puts(&out, "<tool_result>");
            append_tool_result_text(&out, m->content);
            buf_puts(&out, "</tool_result>");
            pending_assistant = true;
            pending_tool_result = true;