hygate-c/server.c

#include "hygate.h"
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <unistd.h>
#include <errno.h>
#include <time.h>
#include <json-c/json.h>
#include <curl/curl.h>

/* ------------------------------------------------------------------ */
/* Streaming context                                                   */
/* ------------------------------------------------------------------ */
struct tcst {
  char *key;
  char *name;
  char *prev_args;
  bool emitted;
};

typedef struct {
  int client_fd;
  bool streaming;
  char completion_id[64];
  long created;
  const char *model;
  strbuf line;             /* SSE line buffer (used for both modes) */
  strbuf emitted_content;  /* streaming: content emitted so far */
  strbuf emitted_reasoning;/* streaming: reasoning emitted so far */
  struct tcst *tc_state;
  int tc_state_n;
  int tc_state_cap;
  bool any_toolcall;
  int include_usage;
  char *final_content;     /* non-streaming accumulation */
  char *final_reasoning;
  tool_call_t *final_tc;
  int final_n_tc;
  int status;
  bool headers_sent;
} sse_ctx_t;

/* ------------------------------------------------------------------ */
/* Low-level response writers                                          */
/* ------------------------------------------------------------------ */
static void send_sse_headers(int fd) {
  const char *h =
    "HTTP/1.1 200 OK\r\n"
    "Content-Type: text/event-stream\r\n"
    "Cache-Control: no-cache\r\n"
    "Connection: keep-alive\r\n"
    "Access-Control-Allow-Origin: *\r\n"
    "Access-Control-Allow-Methods: GET, POST, OPTIONS\r\n"
    "Access-Control-Allow-Headers: Content-Type, Authorization, api-key, X-User-Agent\r\n"
    "\r\n";
  write_all(fd, h, strlen(h));
}

static void send_options(int fd) {
  const char *h =
    "HTTP/1.1 200 OK\r\n"
    "Access-Control-Allow-Origin: *\r\n"
    "Access-Control-Allow-Methods: GET, POST, OPTIONS\r\n"
    "Access-Control-Allow-Headers: Content-Type, Authorization, api-key, X-User-Agent\r\n"
    "Connection: close\r\n"
    "Content-Length: 0\r\n\r\n";
  write_all(fd, h, strlen(h));
}

static void send_simple(int fd, int code, const char *codestr, const char *body) {
  char header[512];
  int hl = snprintf(header, sizeof header,
    "HTTP/1.1 %d %s\r\n"
    "Content-Type: text/plain; charset=utf-8\r\n"
    "Access-Control-Allow-Origin: *\r\n"
    "Access-Control-Allow-Methods: GET, POST, OPTIONS\r\n"
    "Access-Control-Allow-Headers: Content-Type, Authorization, api-key, X-User-Agent\r\n"
    "Connection: close\r\n"
    "Content-Length: %zu\r\n\r\n", code, codestr, strlen(body));
  write_all(fd, header, hl);
  write_all(fd, body, strlen(body));
}

static void send_json(int fd, const char *json) {
  char header[512];
  int hl = snprintf(header, sizeof header,
    "HTTP/1.1 200 OK\r\n"
    "Content-Type: application/json\r\n"
    "Access-Control-Allow-Origin: *\r\n"
    "Access-Control-Allow-Methods: GET, POST, OPTIONS\r\n"
    "Access-Control-Allow-Headers: Content-Type, Authorization, api-key, X-User-Agent\r\n"
    "Connection: close\r\n"
    "Content-Length: %zu\r\n\r\n", strlen(json));
  write_all(fd, header, hl);
  write_all(fd, json, strlen(json));
}

/* ------------------------------------------------------------------ */
/* SSE chunk emitter                                                   */
/* ------------------------------------------------------------------ */
static void emit_chunk(sse_ctx_t *ctx, const char *role, const char *content,
                       const char *reasoning, const char *toolcalls_json,
                       const char *finish_reason) {
  json_object *root = json_object_new_object();
  json_object_object_add(root, "id", json_object_new_string(ctx->completion_id));
  json_object_object_add(root, "object", json_object_new_string("chat.completion.chunk"));
  json_object_object_add(root, "created", json_object_new_int64(ctx->created));
  json_object_object_add(root, "model", json_object_new_string(ctx->model));
  json_object *choices = json_object_new_array();
  json_object *choice = json_object_new_object();
  json_object_object_add(choice, "index", json_object_new_int(0));
  json_object *delta = json_object_new_object();
  if (role) json_object_object_add(delta, "role", json_object_new_string(role));
  if (content) json_object_object_add(delta, "content", json_object_new_string(content));
  if (reasoning) json_object_object_add(delta, "reasoning_content", json_object_new_string(reasoning));
  if (toolcalls_json) {
    json_object *tc = json_tokener_parse(toolcalls_json);
    if (tc) json_object_object_add(delta, "tool_calls", tc);
  }
  json_object_object_add(choice, "delta", delta);
  if (finish_reason)
    json_object_object_add(choice, "finish_reason", json_object_new_string(finish_reason));
  json_object_array_add(choices, choice);
  json_object_object_add(root, "choices", choices);

  const char *s = json_object_to_json_string(root);
  if (s) {
    size_t need = strlen(s) + 32;
    char *buf = (char *)malloc(need);
    if (buf) {
      int bl = snprintf(buf, need, "data: %s\n\n", s);
      write_all(ctx->client_fd, buf, (size_t)bl);
      free(buf);
    }
  }
  json_object_put(root);
}

static void emit_usage_chunk(sse_ctx_t *ctx, int prompt_tokens,
                             int completion_tokens, int reasoning_tokens) {
  json_object *root = json_object_new_object();
  json_object_object_add(root, "id", json_object_new_string(ctx->completion_id));
  json_object_object_add(root, "object", json_object_new_string("chat.completion.chunk"));
  json_object_object_add(root, "created", json_object_new_int64(ctx->created));
  json_object_object_add(root, "model", json_object_new_string(ctx->model));
  json_object *choices = json_object_new_array();
  json_object_object_add(root, "choices", choices);
  json_object *usage = json_object_new_object();
  json_object_object_add(usage, "prompt_tokens", json_object_new_int(prompt_tokens));
  json_object_object_add(usage, "completion_tokens", json_object_new_int(completion_tokens));
  json_object_object_add(usage, "total_tokens",
                         json_object_new_int(prompt_tokens + completion_tokens));
  json_object *details = json_object_new_object();
  json_object_object_add(details, "reasoning_tokens", json_object_new_int(reasoning_tokens));
  json_object_object_add(usage, "completion_tokens_details", details);
  json_object_object_add(root, "usage", usage);
  const char *j = json_object_to_json_string(root);
  if (j) {
    size_t need = strlen(j) + 32;
    char *buf = (char *)malloc(need);
    if (buf) {
      int bl = snprintf(buf, need, "data: %s\n\n", j);
      write_all(ctx->client_fd, buf, (size_t)bl);
      free(buf);
    }
  }
  json_object_put(root);
}

/* ------------------------------------------------------------------ */
/* SSE line processing                                                 */
/* ------------------------------------------------------------------ */
static void handle_stream_delta(sse_ctx_t *ctx, char *c, char *r,
                               tool_call_t *tc, int ntc);

static void process_one_line(sse_ctx_t *ctx, char *line) {
  while (*line == ' ' || *line == '\t' || *line == '\r' || *line == '\n') line++;
  if (*line == 0) return;
  char *e = line + strlen(line) - 1;
  while (e > line && (*e == ' ' || *e == '\t' || *e == '\r' || *e == '\n')) *e-- = 0;
  if (strncmp(line, "data: ", 6) != 0) return;
  const char *payload = line + 6;

  char *c = NULL, *r = NULL;
  tool_call_t *tc = NULL;
  int ntc = 0;
  if (!parse_gradio_stream(payload, &c, &r, &tc, &ntc)) return;

  if (ctx->streaming) {
    handle_stream_delta(ctx, c, r, tc, ntc);
    free(c);
    free(r);
    free_toolcalls(tc, ntc);
  } else {
    if (!ctx->final_content || c[0] != 0) {
      free(ctx->final_content);
      ctx->final_content = c;
    } else {
      free(c);
    }
    if (!ctx->final_reasoning || r[0] != 0) {
      free(ctx->final_reasoning);
      ctx->final_reasoning = r;
    } else {
      free(r);
    }
    if (ntc > 0) {
      free_toolcalls(ctx->final_tc, ctx->final_n_tc);
      ctx->final_tc = tc;
      ctx->final_n_tc = ntc;
    } else {
      free_toolcalls(tc, ntc);
    }
  }
}

static void process_complete_lines(sse_ctx_t *ctx) {
  char *buf = ctx->line.buf;
  if (!buf) return;
  while (1) {
    char *nl = strchr(buf, '\n');
    if (!nl) break;
    *nl = 0;
    process_one_line(ctx, buf);
    size_t consumed = (size_t)(nl - buf) + 1;
    size_t rest = ctx->line.len - consumed;
    memmove(buf, nl + 1, rest + 1);
    ctx->line.len = rest;
  }
}

static void flush_remaining(sse_ctx_t *ctx) {
  if (ctx->line.len == 0 || !ctx->line.buf) return;
  process_one_line(ctx, ctx->line.buf);
  sb_reset(&ctx->line);
}

static void handle_stream_delta(sse_ctx_t *ctx, char *c, char *r,
                               tool_call_t *tc, int ntc) {
  if (strlen(r) > ctx->emitted_reasoning.len) {
    const char *delta = r + ctx->emitted_reasoning.len;
    sb_append(&ctx->emitted_reasoning, delta, strlen(delta));
    emit_chunk(ctx, NULL, NULL, delta, NULL, NULL);
  }
  if (strlen(c) > ctx->emitted_content.len) {
    const char *delta = c + ctx->emitted_content.len;
    sb_append(&ctx->emitted_content, delta, strlen(delta));
    emit_chunk(ctx, NULL, delta, NULL, NULL, NULL);
  }
  for (int i = 0; i < ntc; i++) {
    char keybuf[64];
    const char *key;
    if (tc[i].id && tc[i].id[0]) key = tc[i].id;
    else { snprintf(keybuf, sizeof keybuf, "idx_%d", i); key = keybuf; }
    struct tcst *st = NULL;
    for (int j = 0; j < ctx->tc_state_n; j++)
      if (strcmp(ctx->tc_state[j].key, key) == 0) { st = &ctx->tc_state[j]; break; }
    if (!st) {
      if (ctx->tc_state_n >= ctx->tc_state_cap) {
        int ncap = ctx->tc_state_cap ? ctx->tc_state_cap * 2 : 8;
        struct tcst *nt = (struct tcst *)realloc(ctx->tc_state,
                            (size_t)ncap * sizeof(*ctx->tc_state));
        if (!nt) continue;
        ctx->tc_state = nt;
        ctx->tc_state_cap = ncap;
      }
      char *kcopy = strdup(key);
      char *ncopy = strdup(tc[i].name ? tc[i].name : "");
      char *acopy = strdup("");
      if (!kcopy || !ncopy || !acopy) {
        free(kcopy);
        free(ncopy);
        free(acopy);
        continue;
      }
      st = &ctx->tc_state[ctx->tc_state_n++];
      st->key = kcopy;
      st->name = ncopy;
      st->prev_args = acopy;
      st->emitted = false;
    } else if (tc[i].name && tc[i].name[0] && st->name[0] == 0) {
      char *ncopy = strdup(tc[i].name);
      if (ncopy) {
        free(st->name);
        st->name = ncopy;
      }
    }
    const char *cur = tc[i].arguments ? tc[i].arguments : "";
    if (!st->emitted) {
      char *ncur = strdup(cur);
      if (!ncur) continue;
      st->emitted = true;
      free(st->prev_args);
      st->prev_args = ncur;
      ctx->any_toolcall = true;
      char *js = toolcall_delta_json(&tc[i], i);
      if (js) {
        emit_chunk(ctx, NULL, NULL, NULL, js, NULL);
        free(js);
      }
    } else {
      size_t prev = strlen(st->prev_args);
      size_t curlen = strlen(cur);
      if (curlen > prev) {
        char *ncur = strdup(cur);
        if (!ncur) continue;
        const char *delta = ncur + prev;
        free(st->prev_args);
        st->prev_args = ncur;
        char *js = toolcall_delta_argdelta(i, delta);
        if (js) {
          emit_chunk(ctx, NULL, NULL, NULL, js, NULL);
          free(js);
        }
      }
    }
  }
}

/* ------------------------------------------------------------------ */
/* libcurl callbacks                                                   */
/* ------------------------------------------------------------------ */
static size_t write_mem_cb(void *ptr, size_t sz, size_t nmemb, void *data) {
  strbuf *b = (strbuf *)data;
  size_t n = sz * nmemb;
  if (sz != 0 && n / sz != nmemb) return 0;
  if (sb_append(b, ptr, n) != 0) return 0;
  return n;
}

static size_t header_cb(char *buf, size_t sz, size_t nmemb, void *data) {
  sse_ctx_t *ctx = (sse_ctx_t *)data;
  size_t len = sz * nmemb;
  if (len >= 5 && strncmp(buf, "HTTP/", 5) == 0) {
    char *p = buf;
    while (*p && *p != ' ') p++;
    if (*p == ' ') ctx->status = atoi(p + 1);
    if (ctx->streaming && ctx->status == 200 && !ctx->headers_sent) {
      send_sse_headers(ctx->client_fd);
      ctx->headers_sent = true;
      emit_chunk(ctx, "assistant", NULL, NULL, NULL, NULL);
    }
  }
  return len;
}

static size_t write_sse_cb(void *ptr, size_t sz, size_t nmemb, void *data) {
  sse_ctx_t *ctx = (sse_ctx_t *)data;
  size_t n = sz * nmemb;
  if (sz != 0 && n / sz != nmemb) return 0;
  if (sb_append(&ctx->line, ptr, n) != 0) return 0;
  process_complete_lines(ctx);
  return n;
}

/* ------------------------------------------------------------------ */
/* Upstream calls (with Fibonacci backoff retry)                        */
/* ------------------------------------------------------------------ */
static int upstream_post(const char *url, const char *body, const char *ua,
                         strbuf *resp) {
  for (int attempt = 1; attempt <= MAX_RETRIES; attempt++) {
    sb_reset(resp);
    CURL *c = curl_easy_init();
    if (!c) return -1;
    struct curl_slist *hdrs = NULL;
    hdrs = curl_slist_append(hdrs, "Content-Type: application/json");
    char uah[2048];
    snprintf(uah, sizeof uah, "User-Agent: %s", ua);
    hdrs = curl_slist_append(hdrs, uah);
    curl_easy_setopt(c, CURLOPT_URL, url);
    curl_easy_setopt(c, CURLOPT_POST, 1L);
    curl_easy_setopt(c, CURLOPT_POSTFIELDS, body);
    curl_easy_setopt(c, CURLOPT_POSTFIELDSIZE, (long)strlen(body));
    curl_easy_setopt(c, CURLOPT_HTTPHEADER, hdrs);
    curl_easy_setopt(c, CURLOPT_WRITEFUNCTION, write_mem_cb);
    curl_easy_setopt(c, CURLOPT_WRITEDATA, resp);
    curl_easy_setopt(c, CURLOPT_TIMEOUT, UPSTREAM_TIMEOUT);
    CURLcode rc = curl_easy_perform(c);
    long code = 0;
    curl_easy_getinfo(c, CURLINFO_RESPONSE_CODE, &code);
    curl_slist_free_all(hdrs);
    curl_easy_cleanup(c);
    if (rc == CURLE_OK && code == 200) return 0;
    if (attempt < MAX_RETRIES) sleep((unsigned int)fib_delay(attempt));
  }
  return -1;
}

static int upstream_get_stream(const char *url, const char *ua, sse_ctx_t *ctx) {
  for (int attempt = 1; attempt <= MAX_RETRIES; attempt++) {
    sb_reset(&ctx->line);
    ctx->status = 0;
    free(ctx->final_content);
    ctx->final_content = NULL;
    free(ctx->final_reasoning);
    ctx->final_reasoning = NULL;
    free_toolcalls(ctx->final_tc, ctx->final_n_tc);
    ctx->final_tc = NULL;
    ctx->final_n_tc = 0;
    CURL *c = curl_easy_init();
    if (!c) return -1;
    struct curl_slist *hdrs = NULL;
    hdrs = curl_slist_append(hdrs, "Accept: text/event-stream");
    char uah[2048];
    snprintf(uah, sizeof uah, "User-Agent: %s", ua);
    hdrs = curl_slist_append(hdrs, uah);
    curl_easy_setopt(c, CURLOPT_URL, url);
    curl_easy_setopt(c, CURLOPT_HTTPGET, 1L);
    curl_easy_setopt(c, CURLOPT_HTTPHEADER, hdrs);
    curl_easy_setopt(c, CURLOPT_HEADERFUNCTION, header_cb);
    curl_easy_setopt(c, CURLOPT_HEADERDATA, ctx);
    curl_easy_setopt(c, CURLOPT_WRITEFUNCTION, write_sse_cb);
    curl_easy_setopt(c, CURLOPT_WRITEDATA, ctx);
    curl_easy_setopt(c, CURLOPT_TIMEOUT, UPSTREAM_TIMEOUT);
    CURLcode rc = curl_easy_perform(c);
    long code = 0;
    curl_easy_getinfo(c, CURLINFO_RESPONSE_CODE, &code);
    curl_slist_free_all(hdrs);
    curl_easy_cleanup(c);
    if (rc == CURLE_OK && code == 200) return 0;
    if (ctx->headers_sent) break;
    if (attempt < MAX_RETRIES) sleep((unsigned int)fib_delay(attempt));
  }
  return -1;
}

/* ------------------------------------------------------------------ */
/* Endpoint handlers                                                   */
/* ------------------------------------------------------------------ */
static void handle_models(int fd, const config_t *cfg) {
  json_object *root = json_object_new_object();
  json_object_object_add(root, "object", json_object_new_string("list"));
  json_object *data = json_object_new_array();
  json_object *m = json_object_new_object();
  json_object_object_add(m, "id", json_object_new_string(cfg->model));
  json_object_object_add(m, "object", json_object_new_string("model"));
  json_object_object_add(m, "created", json_object_new_int64((long)time(NULL)));
  json_object_object_add(m, "owned_by", json_object_new_string("hunyuan"));
  json_object_array_add(data, m);
  json_object_object_add(root, "data", data);
  const char *s = json_object_to_json_string(root);
  if (s) send_json(fd, s);
  json_object_put(root);
}

static void handle_chat(int fd, http_req_t *req, const config_t *cfg) {
  json_object *jreq = json_tokener_parse(req->body);
  if (!jreq || json_object_get_type(jreq) != json_type_object) {
    send_simple(fd, 400, "Bad Request", "Invalid JSON payload");
    if (jreq) json_object_put(jreq);
    return;
  }
  bool stream = false;
  json_object *o = NULL;
  if (json_object_object_get_ex(jreq, "stream", &o) &&
      json_object_get_type(o) == json_type_boolean)
    stream = json_object_get_boolean(o);
  char model_buf[256];
  const char *model = cfg->model;
  if (json_object_object_get_ex(jreq, "model", &o) &&
      json_object_get_type(o) == json_type_string &&
      json_object_get_string(o)[0]) {
    snprintf(model_buf, sizeof model_buf, "%s", json_object_get_string(o));
    model = model_buf;
  }

  char *payload = build_gradio_payload(jreq);

  int prompt_tokens = count_messages_tokens(jreq);
  bool include_usage = false;
  json_object *so = NULL;
  if (json_object_object_get_ex(jreq, "stream_options", &so) &&
      json_object_get_type(so) == json_type_object) {
    json_object *iu = NULL;
    if (json_object_object_get_ex(so, "include_usage", &iu) &&
        json_object_get_type(iu) == json_type_boolean)
      include_usage = json_object_get_boolean(iu);
  }

  json_object_put(jreq);
  if (!payload) {
    send_simple(fd, 400, "Bad Request", "Invalid request: missing messages");
    return;
  }

  const char *ua = req->has_x_ua ? req->ua_override
                 : (cfg->user_agent && cfg->user_agent[0] ? cfg->user_agent : DEFAULT_UA);

  /* endpoint without trailing slash (mirrors TrimRight(endpoint, "/")) */
  const char *ep = cfg->endpoint;
  size_t eplen = strlen(ep);
  char epclean[2048];
  if (eplen > 0 && ep[eplen - 1] == '/') {
    size_t clen = eplen - 1 < sizeof epclean - 1 ? eplen - 1 : sizeof epclean - 1;
    memcpy(epclean, ep, clen);
    epclean[clen] = 0;
  } else {
    strncpy(epclean, ep, sizeof epclean - 1);
    epclean[sizeof epclean - 1] = 0;
  }

  char call_url[4096];
  snprintf(call_url, sizeof call_url, "%s/gradio_api/call/chat", epclean);

  strbuf post_resp;
  sb_init(&post_resp);
  if (!post_resp.buf || upstream_post(call_url, payload, ua, &post_resp) != 0) {
    send_simple(fd, 502, "Bad Gateway",
                "Upstream error: failed to reach Gradio call endpoint");
    sb_free(&post_resp);
    free(payload);
    return;
  }

  char event_id[256];
  event_id[0] = 0;
  json_object *jr = json_tokener_parse(post_resp.buf);
  if (jr) {
    json_object *ev = NULL;
    if (json_object_object_get_ex(jr, "event_id", &ev) &&
        json_object_get_type(ev) == json_type_string) {
      strncpy(event_id, json_object_get_string(ev), sizeof event_id - 1);
      event_id[sizeof event_id - 1] = 0;
    }
    json_object_put(jr);
  }
  sb_free(&post_resp);
  if (event_id[0] == 0) {
    send_simple(fd, 502, "Bad Gateway",
                "Upstream error: failed to parse Gradio event ID");
    free(payload);
    return;
  }

  char stream_url[8192];
  snprintf(stream_url, sizeof stream_url, "%s/gradio_api/call/chat/%s", epclean, event_id);

  sse_ctx_t ctx;
  memset(&ctx, 0, sizeof ctx);
  ctx.client_fd = fd;
  ctx.streaming = stream;
  ctx.model = model;
  ctx.include_usage = include_usage;
  ctx.created = (long)time(NULL);
  char uuid[37];
  generate_uuid(uuid);
  snprintf(ctx.completion_id, sizeof ctx.completion_id, "chatcmpl-%s", uuid);
  sb_init(&ctx.line);
  if (stream) {
    sb_init(&ctx.emitted_content);
    sb_init(&ctx.emitted_reasoning);
  }

  int rc = upstream_get_stream(stream_url, ua, &ctx);

  if (rc != 0 || ctx.status != 200) {
    if (!ctx.headers_sent)
      send_simple(fd, 502, "Bad Gateway", "Upstream stream error");
    goto cleanup;
  }

  flush_remaining(&ctx);

  if (stream) {
    emit_chunk(&ctx, NULL, NULL, NULL, NULL, ctx.any_toolcall ? "tool_calls" : "stop");
    if (ctx.include_usage) {
      int reasoning_tokens = tok_count_str(ctx.emitted_reasoning.buf);
      int completion_tokens = tok_count_str(ctx.emitted_content.buf) + reasoning_tokens;
      for (int i = 0; i < ctx.tc_state_n; i++) {
        completion_tokens += tok_count_str(ctx.tc_state[i].name);
        completion_tokens += tok_count_str(ctx.tc_state[i].prev_args);
      }
      emit_usage_chunk(&ctx, prompt_tokens, completion_tokens, reasoning_tokens);
    }
    const char *done = "data: [DONE]\n\n";
    write_all(fd, done, strlen(done));
  } else {
    char *resp = build_completion_response(ctx.completion_id, ctx.created, ctx.model,
                                          ctx.final_content, ctx.final_reasoning,
                                          ctx.final_tc, ctx.final_n_tc, prompt_tokens);
    if (resp) {
      send_json(fd, resp);
      free(resp);
    } else {
      send_simple(fd, 500, "Internal Server Error", "Failed to build response");
    }
  }

cleanup:
  sb_free(&ctx.line);
  sb_free(&ctx.emitted_content);
  sb_free(&ctx.emitted_reasoning);
  for (int i = 0; i < ctx.tc_state_n; i++) {
    free(ctx.tc_state[i].key);
    free(ctx.tc_state[i].name);
    free(ctx.tc_state[i].prev_args);
  }
  free(ctx.tc_state);
  free(ctx.final_content);
  free(ctx.final_reasoning);
  free_toolcalls(ctx.final_tc, ctx.final_n_tc);
  free(payload);
}

/* ------------------------------------------------------------------ */
/* Request parsing / connection dispatch                               */
/* ------------------------------------------------------------------ */
int read_headers(int fd, strbuf *buf) {
  if (!buf->buf) return -1;
  char tmp[4096];
  while (1) {
    if (strstr(buf->buf, "\r\n\r\n")) break;
    ssize_t n = read(fd, tmp, sizeof tmp);
    if (n <= 0) return -1;
    if (sb_append(buf, tmp, (size_t)n) != 0) return -1;
    if (buf->len > (1 << 20)) return -1;
  }
  return 0;
}

int parse_http_request(strbuf *buf, int fd, http_req_t *req, long max_body_bytes) {
  if (!buf->buf) return -1;
  char *p = buf->buf;
  char *line_end = strstr(p, "\r\n");
  if (!line_end) return -1;
  *line_end = 0;
  char *sp1 = strchr(p, ' ');
  if (!sp1) return -1;
  *sp1 = 0;
  strncpy(req->method, p, sizeof req->method - 1);
  req->method[sizeof req->method - 1] = 0;
  char *rest = sp1 + 1;
  char *sp2 = strchr(rest, ' ');
  if (sp2) *sp2 = 0;
  strncpy(req->path, rest, sizeof req->path - 1);
  req->path[sizeof req->path - 1] = 0;

  char *h = line_end + 2;
  /* buf->buf was NUL-terminated at the first CRLF above, so searching from
     buf->buf would only see the request line. Search from h (the intact
     header block) for the end-of-headers marker. */
  char *hend = strstr(h, "\r\n\r\n");
  if (!hend) return -1;
  char *hlimit = hend;
  while (h < hlimit) {
    char *le = strstr(h, "\r\n");
    if (!le) break;
    *le = 0;
    char *colon = strchr(h, ':');
    if (colon) {
      *colon = 0;
      char *k = h;
      char *v = colon + 1;
      if (*v == ' ') v++;
      if (iequals(k, "content-length")) {
      errno = 0;
      char *endp = NULL;
      long long cl = strtoll(v, &endp, 10);
      while (endp && (*endp == ' ' || *endp == '\t' || *endp == '\r' || *endp == '\n')) endp++;
      if (errno != 0 || endp == v || (endp && *endp != '\0') || cl < 0) return -1;
      req->content_length = (long)cl;
    }
      else if (iequals(k, "content-type")) {
        strncpy(req->content_type, v, sizeof req->content_type - 1);
        req->content_type[sizeof req->content_type - 1] = 0;
      } else if (iequals(k, "x-user-agent")) {
        strncpy(req->ua_override, v, sizeof req->ua_override - 1);
        req->ua_override[sizeof req->ua_override - 1] = 0;
        req->has_x_ua = true;
      }
    }
    h = le + 2;
  }

  char *body_start = hlimit + 4;
  if ((size_t)(body_start - buf->buf) > buf->len) return -1;
  size_t body_avail = buf->len - (size_t)(body_start - buf->buf);
  if (req->content_length < 0) return -1;
  if (max_body_bytes > 0 && req->content_length > max_body_bytes) return -2;
  if (req->content_length > 0) {
    req->body = (char *)malloc((size_t)req->content_length + 1);
    if (!req->body) return -1;
    size_t got = body_avail < (size_t)req->content_length
                   ? body_avail
                   : (size_t)req->content_length;
    if (got > 0) memcpy(req->body, body_start, got);
    while (got < (size_t)req->content_length) {
      ssize_t n = read(fd, req->body + got, (size_t)req->content_length - got);
      if (n <= 0) { free(req->body); req->body = NULL; return -1; }
      got += (size_t)n;
    }
    req->body[got] = 0;
    req->body_len = got;
  } else {
    req->body = strdup("");
    if (!req->body) return -1;
    req->body_len = 0;
  }
  return 0;
}

void http_req_free(http_req_t *r) {
  free(r->body);
  r->body = NULL;
}

void handle_connection(int fd, const config_t *cfg) {
  strbuf reqbuf;
  sb_init(&reqbuf);
  if (read_headers(fd, &reqbuf) != 0) {
    sb_free(&reqbuf);
    return;
  }
  http_req_t req;
  memset(&req, 0, sizeof req);
  int prc = parse_http_request(&reqbuf, fd, &req, cfg->max_body_bytes);
  if (prc != 0) {
    if (prc == -2)
      send_simple(fd, 413, "Payload Too Large",
                  "Request body exceeds the maximum allowed size");
    else
      send_simple(fd, 400, "Bad Request", "Malformed HTTP request");
    http_req_free(&req);
    sb_free(&reqbuf);
    return;
  }

  char *q = strchr(req.path, '?');
  if (q) *q = 0;

  bool is_models = (strcmp(req.path, "/models") == 0 ||
                    strcmp(req.path, "/v1/models") == 0);
  bool is_chat = (strcmp(req.path, "/chat/completions") == 0 ||
                  strcmp(req.path, "/v1/chat/completions") == 0);

  if (strcmp(req.method, "OPTIONS") == 0) {
    send_options(fd);
  } else if (is_models) {
    if (strcmp(req.method, "GET") == 0) handle_models(fd, cfg);
    else send_simple(fd, 405, "Method Not Allowed", "Method not allowed");
  } else if (is_chat) {
    if (strcmp(req.method, "POST") == 0) handle_chat(fd, &req, cfg);
    else send_simple(fd, 405, "Method Not Allowed", "Method not allowed");
  } else {
    send_simple(fd, 404, "Not Found", "Not Found");
  }

  http_req_free(&req);
  sb_free(&reqbuf);
}