/*
** $Id: lcmd.c $
** Command execution for backtick syntax
** See Copyright Notice in lua.h
*/

#define lcmd_c
#define LUA_LIB

#include "lprefix.h"

#include <errno.h>
#include <glob.h>
#include <signal.h>
#include <stdlib.h>
#include <string.h>
#include <sys/select.h>
#include <sys/wait.h>
#include <unistd.h>

#include "lua.h"
#include "lauxlib.h"
#include "lcmd.h"
#include "lbuiltin.h"
#include "lstring.h"


/* ===== argv parser ===== */

#define PA_QUOTED 1  /* suppress glob/tilde expansion for this token */

typedef struct {
  char **argv;
  char *buf;
  int argc;
  int *flags;  /* per-token flags (PA_QUOTED etc.) */
} ParsedArgs;


static void free_argv (ParsedArgs *pa) {
  free(pa->buf);
  free(pa->argv);
  free(pa->flags);
}


/* ===== glob/tilde/brace expansion ===== */

/*
** Expand ~ at the start of a token to the home directory.
** Returns a malloc'd string (caller must free), or NULL on failure.
** If no tilde prefix, returns a strdup of the original.
*/
static char *expand_tilde (const char *token) {
  const char *home;
  size_t homelen, restlen;
  char *result;
  if (token[0] != '~')
    return strdup(token);
  /* only handle plain ~ and ~/path (not ~user for now) */
  if (token[1] != '\0' && token[1] != '/')
    return strdup(token);
  home = getenv("HOME");
  if (home == NULL)
    return strdup(token);
  homelen = strlen(home);
  restlen = strlen(token + 1);  /* everything after ~ */
  result = (char *)malloc(homelen + restlen + 1);
  if (result == NULL) return NULL;
  memcpy(result, home, homelen);
  memcpy(result + homelen, token + 1, restlen + 1);
  return result;
}


/*
** Brace expansion: expand {a,b,c} patterns into multiple strings.
** Handles prefix{a,b,c}suffix → prefixasuffix prefixbsuffix prefixcsuffix.
** Does NOT handle nested braces (kept literal).
** Writes results into out[] (caller provides), returns count.
** out[] entries are malloc'd strings. max_out limits array size.
** Returns 0 if no braces found (token unchanged).
*/
#define MAX_BRACE_RESULTS 256

static int expand_braces (const char *token, char **out, int max_out) {
  const char *open, *close, *p;
  size_t prefix_len, suffix_len;
  int depth, ncommas, count;
  const char *alts[MAX_BRACE_RESULTS];
  size_t altlens[MAX_BRACE_RESULTS];
  int nalts = 0;

  /* find first '{' */
  open = strchr(token, '{');
  if (open == NULL) return 0;

  /* find matching '}' at depth 0, counting commas */
  depth = 1; ncommas = 0;
  for (p = open + 1; *p != '\0' && depth > 0; p++) {
    if (*p == '{') depth++;
    else if (*p == '}') depth--;
    else if (*p == ',' && depth == 1) ncommas++;
  }
  if (depth != 0 || ncommas == 0)
    return 0;  /* no valid brace expression */

  close = p - 1;  /* points to '}' */
  prefix_len = (size_t)(open - token);
  suffix_len = strlen(close + 1);

  /* split alternatives on ',' at depth 0 */
  p = open + 1;
  alts[0] = p;
  for (; p < close; p++) {
    if (*p == '{') depth++;
    else if (*p == '}') depth--;
    else if (*p == ',' && depth == 0) {
      altlens[nalts] = (size_t)(p - alts[nalts]);
      nalts++;
      if (nalts >= MAX_BRACE_RESULTS) return 0;
      alts[nalts] = p + 1;
    }
  }
  altlens[nalts] = (size_t)(close - alts[nalts]);
  nalts++;

  /* generate prefix + alt + suffix for each alternative */
  count = 0;
  for (int i = 0; i < nalts && count < max_out; i++) {
    size_t total = prefix_len + altlens[i] + suffix_len + 1;
    char *s = (char *)malloc(total);
    if (s == NULL) {
      for (int j = 0; j < count; j++) free(out[j]);
      return -1;
    }
    memcpy(s, token, prefix_len);
    memcpy(s + prefix_len, alts[i], altlens[i]);
    memcpy(s + prefix_len + altlens[i], close + 1, suffix_len + 1);
    out[count++] = s;
  }
  return count;
}


/*
** Growing list of tokens used during expansion.
** Each item is a separately malloc'd string.
*/
typedef struct {
  char **argv;
  int *flags;
  int argc;
  int capacity;
} TokenList;


static int tklist_init (TokenList *tl, int capacity) {
  tl->argv = (char **)malloc((size_t)capacity * sizeof(char *));
  tl->flags = (int *)calloc((size_t)capacity, sizeof(int));
  tl->argc = 0;
  tl->capacity = capacity;
  if (tl->argv == NULL || tl->flags == NULL) {
    free(tl->argv); free(tl->flags);
    return -1;
  }
  return 0;
}


static void tklist_free (TokenList *tl) {
  int i;
  for (i = 0; i < tl->argc; i++)
    free(tl->argv[i]);
  free(tl->argv);
  free(tl->flags);
}


/*
** Add a token to the list. Takes ownership of 's' on success.
** Returns 0 on success, -1 on allocation failure (caller still owns 's').
*/
static int tklist_add (TokenList *tl, char *s, int flag) {
  if (tl->argc >= tl->capacity) {
    int newcap = tl->capacity * 2;
    char **a = (char **)realloc(tl->argv, (size_t)newcap * sizeof(char *));
    int *f = (int *)realloc(tl->flags, (size_t)newcap * sizeof(int));
    if (a == NULL || f == NULL) {
      /* if one succeeded, keep the original pointer valid */
      if (a) tl->argv = a;
      if (f) tl->flags = f;
      return -1;
    }
    tl->argv = a;
    tl->flags = f;
    tl->capacity = newcap;
  }
  tl->argv[tl->argc] = s;
  tl->flags[tl->argc] = flag;
  tl->argc++;
  return 0;
}


/*
** Glob a single token and append results to 'tl'.
** If the token has no metacharacters, it is added as-is (takes ownership).
** Otherwise, glob results are added and 'token' is freed.
** Returns 0 on success, -1 on failure (token is freed either way).
*/
static int glob_token (TokenList *tl, char *token) {
  const char *s;
  int has_meta = 0;
  for (s = token; *s != '\0'; s++) {
    if (*s == '*' || *s == '?' || *s == '[') { has_meta = 1; break; }
  }
  if (!has_meta)
    return tklist_add(tl, token, 0);  /* no glob needed */
  {
    glob_t g;
    int ret = glob(token, GLOB_NOCHECK, NULL, &g);
    free(token);
    if (ret != 0)
      return 0;  /* GLOB_NOCHECK means this shouldn't happen, but be safe */
    {
      size_t k;
      for (k = 0; k < g.gl_pathc; k++) {
        char *dup = strdup(g.gl_pathv[k]);
        if (dup == NULL || tklist_add(tl, dup, 0) != 0) {
          free(dup);
          globfree(&g);
          return -1;
        }
      }
    }
    globfree(&g);
    return 0;
  }
}


/*
** Expand unquoted tokens: tilde → brace → glob.
** Quoted tokens (PA_QUOTED) are passed through unchanged.
** Replaces pa->argv, pa->buf, and pa->flags in place.
** Returns 0 on success, -1 on allocation failure.
*/
static int expand_argv (ParsedArgs *pa) {
  int i;
  TokenList tl;
  size_t total_buflen = 0;
  char *new_buf;
  size_t bp = 0;

  if (tklist_init(&tl, pa->argc + 16) != 0)
    return -1;

  for (i = 0; i < pa->argc; i++) {
    if (pa->flags[i] & PA_QUOTED) {
      char *dup = strdup(pa->argv[i]);
      if (dup == NULL || tklist_add(&tl, dup, PA_QUOTED) != 0) {
        free(dup); goto fail;
      }
    } else {
      char *tilded;
      char *brace_results[MAX_BRACE_RESULTS];
      int nbrace;

      /* step 1: tilde expansion */
      tilded = expand_tilde(pa->argv[i]);
      if (tilded == NULL) goto fail;

      /* step 2: brace expansion */
      nbrace = expand_braces(tilded, brace_results, MAX_BRACE_RESULTS);
      if (nbrace < 0) { free(tilded); goto fail; }

      if (nbrace == 0) {
        /* no braces — glob the tilde-expanded token */
        if (glob_token(&tl, tilded) != 0) goto fail;
      } else {
        int j;
        free(tilded);
        /* glob each brace alternative */
        for (j = 0; j < nbrace; j++) {
          if (glob_token(&tl, brace_results[j]) != 0) {
            /* free remaining brace results */
            for (j++; j < nbrace; j++) free(brace_results[j]);
            goto fail;
          }
        }
      }
    }
  }

  /* pack all tokens into a single contiguous buffer */
  for (i = 0; i < tl.argc; i++)
    total_buflen += strlen(tl.argv[i]) + 1;

  new_buf = (char *)malloc(total_buflen > 0 ? total_buflen : 1);
  if (new_buf == NULL) goto fail;

  for (i = 0; i < tl.argc; i++) {
    char *old = tl.argv[i];
    size_t slen = strlen(old);
    memcpy(new_buf + bp, old, slen + 1);
    tl.argv[i] = new_buf + bp;
    bp += slen + 1;
    free(old);
  }

  /* replace old arrays */
  free(pa->argv);
  free(pa->buf);
  free(pa->flags);
  pa->argv = (char **)realloc(tl.argv, ((size_t)tl.argc + 1) * sizeof(char *));
  if (pa->argv == NULL) pa->argv = tl.argv;
  pa->argv[tl.argc] = NULL;
  pa->buf = new_buf;
  pa->flags = tl.flags;
  pa->argc = tl.argc;
  return 0;

fail:
  tklist_free(&tl);
  return -1;
}


/*
** Tokenize command string into argv array.
** Handles: whitespace splitting, single quotes (literal),
** double quotes (with \", \\, \n, \t, \$ escapes), backslash escaping.
** Returns 0 on success, -1 on error (unterminated quote).
*/
static int parse_argv (const char *cmd, ParsedArgs *result) {
  size_t len = strlen(cmd);
  char *buf = (char *)malloc(len + 1);
  int capacity = 8;
  char **argv = (char **)malloc((size_t)capacity * sizeof(char *));
  int *flags = (int *)calloc((size_t)capacity, sizeof(int));
  int argc = 0;
  size_t bp = 0;  /* position in buf */
  const char *p = cmd;

  if (buf == NULL || argv == NULL || flags == NULL) {
    free(buf);
    free(argv);
    free(flags);
    return -1;
  }

  while (*p != '\0') {
    int quoted = 0;

    /* skip whitespace */
    while (*p == ' ' || *p == '\t' || *p == '\n' || *p == '\r')
      p++;
    if (*p == '\0')
      break;

    /* start of a token */
    if (argc >= capacity - 1) {
      capacity *= 2;
      argv = (char **)realloc(argv, (size_t)capacity * sizeof(char *));
      flags = (int *)realloc(flags, (size_t)capacity * sizeof(int));
      if (argv == NULL || flags == NULL) {
        free(buf); free(argv); free(flags);
        return -1;
      }
    }
    argv[argc] = buf + bp;

    while (*p != '\0') {
      if (*p == ' ' || *p == '\t' || *p == '\n' || *p == '\r')
        break;

      if (*p == '\'') {
        /* single quote: literal until closing quote */
        quoted = 1;
        p++;
        while (*p != '\0' && *p != '\'')
          buf[bp++] = *p++;
        if (*p == '\'') p++;
        else { free(buf); free(argv); free(flags); return -1; }
      }
      else if (*p == '"') {
        /* double quote: with escape sequences */
        quoted = 1;
        p++;
        while (*p != '\0' && *p != '"') {
          if (*p == '\\' && p[1] != '\0') {
            p++;
            switch (*p) {
              case '"':  buf[bp++] = '"';  break;
              case '\\': buf[bp++] = '\\'; break;
              case 'n':  buf[bp++] = '\n'; break;
              case 't':  buf[bp++] = '\t'; break;
              case '$':  buf[bp++] = '$';  break;
              default:   buf[bp++] = '\\'; buf[bp++] = *p; break;
            }
            p++;
          }
          else {
            buf[bp++] = *p++;
          }
        }
        if (*p == '"') p++;
        else { free(buf); free(argv); free(flags); return -1; }
      }
      else if (*p == '\\' && p[1] != '\0') {
        /* backslash escape outside quotes */
        quoted = 1;
        p++;
        buf[bp++] = *p++;
      }
      else {
        buf[bp++] = *p++;
      }
    }

    buf[bp++] = '\0';
    flags[argc] = quoted ? PA_QUOTED : 0;
    argc++;
  }

  argv[argc] = NULL;
  result->argv = argv;
  result->buf = buf;
  result->argc = argc;
  result->flags = flags;

  /* expand globs and tildes on unquoted tokens */
  if (expand_argv(result) != 0) {
    free(buf);
    free(argv);
    free(flags);
    return -1;
  }
  return 0;
}


/* ===== pipeline splitter ===== */

#define MAX_PIPELINE_STAGES 64

/*
** Split a command string on unquoted '|' into pipeline stages.
** Reuses the same quote-tracking logic as parse_argv().
** Returns 0 on success, -1 on error (empty stage, leading/trailing pipe).
** Writes stage pointers into stages[] and count into *nstages.
** Caller must free stages[0] (the backing buffer).
*/
static int split_pipeline (const char *cmd, char **stages, int *nstages) {
  size_t len = strlen(cmd);
  char *buf = (char *)malloc(len + 1);
  int count = 0;
  const char *p = cmd;
  size_t bp = 0;

  if (buf == NULL) return -1;

  stages[count++] = buf;

  while (*p != '\0') {
    if (*p == '\'' ) {
      /* single-quoted string: copy through including quotes */
      buf[bp++] = *p++;
      while (*p != '\0' && *p != '\'')
        buf[bp++] = *p++;
      if (*p == '\'') buf[bp++] = *p++;
    }
    else if (*p == '"') {
      /* double-quoted string: copy through, respecting backslash */
      buf[bp++] = *p++;
      while (*p != '\0' && *p != '"') {
        if (*p == '\\' && p[1] != '\0') {
          buf[bp++] = *p++;
          buf[bp++] = *p++;
        } else {
          buf[bp++] = *p++;
        }
      }
      if (*p == '"') buf[bp++] = *p++;
    }
    else if (*p == '\\' && p[1] != '\0') {
      /* backslash escape: copy both chars */
      buf[bp++] = *p++;
      buf[bp++] = *p++;
    }
    else if (*p == '|') {
      /* unquoted pipe — split here */
      buf[bp++] = '\0';
      if (count >= MAX_PIPELINE_STAGES) { free(buf); return -1; }
      stages[count++] = buf + bp;
      p++;
    }
    else {
      buf[bp++] = *p++;
    }
  }
  buf[bp] = '\0';

  /* validate: no empty stages (only for multi-stage pipelines) */
  if (count > 1) {
    for (int i = 0; i < count; i++) {
      const char *s = stages[i];
      /* skip whitespace */
      while (*s == ' ' || *s == '\t' || *s == '\n' || *s == '\r') s++;
      if (*s == '\0') { free(buf); return -1; }
    }
  }

  *nstages = count;
  return 0;
}


/* ===== pipe reader ===== */

#define READ_CHUNK 4096

typedef struct {
  char *data;
  size_t len;
  size_t cap;
} DynBuf;


static void dynbuf_init (DynBuf *b) {
  b->data = NULL;
  b->len = 0;
  b->cap = 0;
}


static void dynbuf_free (DynBuf *b) {
  free(b->data);
  b->data = NULL;
  b->len = b->cap = 0;
}


static int dynbuf_grow (DynBuf *b, size_t need) {
  if (b->len + need > b->cap) {
    size_t newcap = (b->cap == 0) ? READ_CHUNK : b->cap;
    char *p;
    while (newcap < b->len + need)
      newcap *= 2;
    p = (char *)realloc(b->data, newcap);
    if (p == NULL) return -1;
    b->data = p;
    b->cap = newcap;
  }
  return 0;
}


/*
** Read from two fds simultaneously into DynBufs using select().
** Avoids deadlock when child fills one pipe buffer.
*/
static void read_pipes (int fd_out, int fd_err,
                        DynBuf *buf_out, DynBuf *buf_err) {
  int nfds = (fd_out > fd_err ? fd_out : fd_err) + 1;
  int open_out = 1, open_err = 1;

  while (open_out || open_err) {
    fd_set rfds;
    int ret;

    FD_ZERO(&rfds);
    if (open_out) FD_SET(fd_out, &rfds);
    if (open_err) FD_SET(fd_err, &rfds);

    ret = select(nfds, &rfds, NULL, NULL, NULL);
    if (ret < 0) {
      if (errno == EINTR) continue;
      break;  /* unexpected error */
    }

    if (open_out && FD_ISSET(fd_out, &rfds)) {
      ssize_t n;
      if (dynbuf_grow(buf_out, READ_CHUNK) != 0) break;
      n = read(fd_out, buf_out->data + buf_out->len, READ_CHUNK);
      if (n > 0)
        buf_out->len += (size_t)n;
      else {
        close(fd_out);
        open_out = 0;
      }
    }

    if (open_err && FD_ISSET(fd_err, &rfds)) {
      ssize_t n;
      if (dynbuf_grow(buf_err, READ_CHUNK) != 0) break;
      n = read(fd_err, buf_err->data + buf_err->len, READ_CHUNK);
      if (n > 0)
        buf_err->len += (size_t)n;
      else {
        close(fd_err);
        open_err = 0;
      }
    }
  }
}


/* ===== pipeline execution ===== */

/*
** Build result table {code=N, stdout=S, stderr=S} on the Lua stack.
*/
static void push_result_table (lua_State *L, int code,
                               DynBuf *buf_out, DynBuf *buf_err) {
  lua_createtable(L, 0, 3);
  lua_pushinteger(L, code);
  lua_setfield(L, -2, "code");
  lua_pushlstring(L, buf_out->data ? buf_out->data : "", buf_out->len);
  lua_setfield(L, -2, "stdout");
  lua_pushlstring(L, buf_err->data ? buf_err->data : "", buf_err->len);
  lua_setfield(L, -2, "stderr");
}


/*
** Extract exit code from wait status.
*/
static int exit_code_from_status (int status) {
  if (WIFEXITED(status))
    return WEXITSTATUS(status);
  else if (WIFSIGNALED(status))
    return 128 + WTERMSIG(status);
  return -1;
}


/*
** Write "name: strerror(errno)\n" to stderr after exec failure.
** Only called in forked child processes.
*/
static void exec_failed (const char *name) {
  const char *err = strerror(errno);
  (void)write(STDERR_FILENO, name, strlen(name));
  (void)write(STDERR_FILENO, ": ", 2);
  (void)write(STDERR_FILENO, err, strlen(err));
  (void)write(STDERR_FILENO, "\n", 1);
}


/*
** Look up a function in lush[table][key].
** On success, pushes the function and returns 1.
** On failure, cleans the stack and returns 0.
*/
static int lookup_lush_func (lua_State *L, const char *table,
                             const char *key) {
  lua_rawgeti(L, LUA_REGISTRYINDEX, LUA_RIDX_LUSH);
  if (!lua_istable(L, -1)) {
    lua_pop(L, 1);
    return 0;
  }
  if (lua_getfield(L, -1, table) != LUA_TTABLE) {
    lua_pop(L, 2);
    return 0;
  }
  if (lua_getfield(L, -1, key) != LUA_TFUNCTION) {
    lua_pop(L, 3);
    return 0;
  }
  lua_remove(L, -2);  /* remove subtable */
  lua_remove(L, -2);  /* remove lush table */
  return 1;
}


/* ===== user command dispatch (runs in forked child) ===== */

/*
** Look up argv[0] in lush.commands. If found, call the function with
** all argv strings, extract an exit code, and _exit(). This must only
** be called inside a fork()ed child process.
** Returns 0 if not a user command (caller should fall through to execvp).
*/
static int exec_user_command (lua_State *L, ParsedArgs *pa) {
  int i, code = 0;
  if (!lookup_lush_func(L, "commands", pa->argv[0]))
    return 0;
  for (i = 0; i < pa->argc; i++)
    lua_pushstring(L, pa->argv[i]);
  if (lua_pcall(L, pa->argc, 1, 0) != LUA_OK) {
    /* write error to stderr and exit with failure */
    const char *err = lua_tostring(L, -1);
    if (err) {
      (void)write(STDERR_FILENO, err, strlen(err));
      (void)write(STDERR_FILENO, "\n", 1);
    }
    _exit(1);
  }
  /* extract exit code from return value */
  if (lua_isinteger(L, -1)) {
    code = (int)lua_tointeger(L, -1);
  } else if (lua_istable(L, -1)) {
    if (lua_getfield(L, -1, "code") == LUA_TNUMBER)
      code = (int)lua_tointeger(L, -1);
    lua_pop(L, 1);
  }
  fflush(stdout);
  fflush(stderr);
  _exit(code);
  return 1;  /* unreachable */
}


/*
** Execute a multi-stage pipeline.
** stages[0..nstages-1] are command strings. Each is parsed with parse_argv().
** Inter-stage pipes connect stdout→stdin. Only the last stage's stdout/stderr
** are captured. Middle stages' stderr goes to the parent's stderr (inherited).
** Returns 1 (one result table on the Lua stack).
*/
static int exec_pipeline (lua_State *L, char **stages, int nstages,
                          int interactive) {
  ParsedArgs *pa = NULL;
  int (*inter_pipes)[2] = NULL;  /* inter_pipes[i] connects stage i → i+1 */
  int out_pipe[2] = {-1, -1}, err_pipe[2] = {-1, -1};
  pid_t *pids = NULL;
  int i, last_code = -1;
  DynBuf buf_out, buf_err;
  struct sigaction sa_old_pipe, sa_old_int, sa_old_quit, sa_new;

  /* parse all stages */
  pa = (ParsedArgs *)calloc((size_t)nstages, sizeof(ParsedArgs));
  if (pa == NULL) return luaL_error(L, "out of memory");

  for (i = 0; i < nstages; i++) {
    if (parse_argv(stages[i], &pa[i]) != 0) {
      for (int j = 0; j < i; j++) free_argv(&pa[j]);
      free(pa);
      return luaL_error(L, "unterminated quote in pipeline stage %d", i + 1);
    }
    if (pa[i].argc == 0) {
      for (int j = 0; j <= i; j++) free_argv(&pa[j]);
      free(pa);
      return luaL_error(L, "empty command in pipeline stage %d", i + 1);
    }
  }

  /* allocate inter-stage pipes (nstages-1 of them) */
  if (nstages > 1) {
    inter_pipes = (int (*)[2])calloc((size_t)(nstages - 1), sizeof(int [2]));
    if (inter_pipes == NULL) {
      for (i = 0; i < nstages; i++) free_argv(&pa[i]);
      free(pa);
      return luaL_error(L, "out of memory");
    }
    for (i = 0; i < nstages - 1; i++) {
      if (pipe(inter_pipes[i]) != 0) {
        for (int j = 0; j < i; j++) { close(inter_pipes[j][0]); close(inter_pipes[j][1]); }
        free(inter_pipes);
        for (int j = 0; j < nstages; j++) free_argv(&pa[j]);
        free(pa);
        return luaL_error(L, "pipe() failed: %s", strerror(errno));
      }
    }
  }

  /* capture pipes for last stage (only in non-interactive mode) */
  if (!interactive) {
    if (pipe(out_pipe) != 0) {
      if (inter_pipes) {
        for (i = 0; i < nstages - 1; i++) { close(inter_pipes[i][0]); close(inter_pipes[i][1]); }
        free(inter_pipes);
      }
      for (i = 0; i < nstages; i++) free_argv(&pa[i]);
      free(pa);
      return luaL_error(L, "pipe() failed: %s", strerror(errno));
    }
    if (pipe(err_pipe) != 0) {
      close(out_pipe[0]); close(out_pipe[1]);
      if (inter_pipes) {
        for (i = 0; i < nstages - 1; i++) { close(inter_pipes[i][0]); close(inter_pipes[i][1]); }
        free(inter_pipes);
      }
      for (i = 0; i < nstages; i++) free_argv(&pa[i]);
      free(pa);
      return luaL_error(L, "pipe() failed: %s", strerror(errno));
    }
  }

  /* ignore signals in parent */
  memset(&sa_new, 0, sizeof(sa_new));
  sa_new.sa_handler = SIG_IGN;
  sigemptyset(&sa_new.sa_mask);
  sigaction(SIGPIPE, &sa_new, &sa_old_pipe);
  if (interactive) {
    sigaction(SIGINT, &sa_new, &sa_old_int);
    sigaction(SIGQUIT, &sa_new, &sa_old_quit);
  }

  /* allocate pid array */
  pids = (pid_t *)calloc((size_t)nstages, sizeof(pid_t));
  if (pids == NULL) {
    if (!interactive) {
      close(out_pipe[0]); close(out_pipe[1]);
      close(err_pipe[0]); close(err_pipe[1]);
    }
    if (inter_pipes) {
      for (i = 0; i < nstages - 1; i++) { close(inter_pipes[i][0]); close(inter_pipes[i][1]); }
      free(inter_pipes);
    }
    for (i = 0; i < nstages; i++) free_argv(&pa[i]);
    free(pa);
    sigaction(SIGPIPE, &sa_old_pipe, NULL);
    if (interactive) {
      sigaction(SIGINT, &sa_old_int, NULL);
      sigaction(SIGQUIT, &sa_old_quit, NULL);
    }
    return luaL_error(L, "out of memory");
  }

  /* fork all stages */
  for (i = 0; i < nstages; i++) {
    pids[i] = fork();
    if (pids[i] < 0) {
      /* fork failed — kill already-forked children */
      for (int j = 0; j < i; j++) {
        kill(pids[j], SIGTERM);
        waitpid(pids[j], NULL, 0);
      }
      if (!interactive) {
        close(out_pipe[0]); close(out_pipe[1]);
        close(err_pipe[0]); close(err_pipe[1]);
      }
      if (inter_pipes) {
        for (int j = 0; j < nstages - 1; j++) { close(inter_pipes[j][0]); close(inter_pipes[j][1]); }
        free(inter_pipes);
      }
      for (int j = 0; j < nstages; j++) free_argv(&pa[j]);
      free(pa); free(pids);
      sigaction(SIGPIPE, &sa_old_pipe, NULL);
      if (interactive) {
        sigaction(SIGINT, &sa_old_int, NULL);
        sigaction(SIGQUIT, &sa_old_quit, NULL);
      }
      return luaL_error(L, "fork() failed: %s", strerror(errno));
    }

    if (pids[i] == 0) {
      /* === child process for stage i === */
      int j;

      /* restore default signals */
      sa_new.sa_handler = SIG_DFL;
      sigaction(SIGPIPE, &sa_new, NULL);
      if (interactive) {
        sigaction(SIGINT, &sa_new, NULL);
        sigaction(SIGQUIT, &sa_new, NULL);
      }

      /* set up stdin */
      if (i > 0) {
        dup2(inter_pipes[i - 1][0], STDIN_FILENO);
      }

      /* set up stdout */
      if (i < nstages - 1) {
        dup2(inter_pipes[i][1], STDOUT_FILENO);
      } else if (!interactive) {
        /* last stage: capture stdout and stderr */
        dup2(out_pipe[1], STDOUT_FILENO);
        dup2(err_pipe[1], STDERR_FILENO);
      }
      /* else: interactive last stage inherits terminal */

      /* close all pipe fds in child */
      if (inter_pipes) {
        for (j = 0; j < nstages - 1; j++) {
          close(inter_pipes[j][0]);
          close(inter_pipes[j][1]);
        }
      }
      if (!interactive) {
        close(out_pipe[0]); close(out_pipe[1]);
        close(err_pipe[0]); close(err_pipe[1]);
      }

      exec_user_command(L, &pa[i]);
      execvp(pa[i].argv[0], pa[i].argv);
      exec_failed(pa[i].argv[0]);
      _exit(127);
    }
  }

  /* === parent: close all write ends === */
  if (inter_pipes) {
    for (i = 0; i < nstages - 1; i++) {
      close(inter_pipes[i][0]);
      close(inter_pipes[i][1]);
    }
    free(inter_pipes);
  }

  /* free parsed args (children have already exec'd) */
  for (i = 0; i < nstages; i++) free_argv(&pa[i]);
  free(pa);

  dynbuf_init(&buf_out);
  dynbuf_init(&buf_err);

  if (!interactive) {
    close(out_pipe[1]);
    close(err_pipe[1]);
    /* read captured output from last stage */
    read_pipes(out_pipe[0], err_pipe[0], &buf_out, &buf_err);
  }

  /* wait for all children, keep last stage's exit code */
  for (i = 0; i < nstages; i++) {
    int status;
    waitpid(pids[i], &status, 0);
    if (i == nstages - 1)
      last_code = exit_code_from_status(status);
  }
  free(pids);

  /* restore signals */
  sigaction(SIGPIPE, &sa_old_pipe, NULL);
  if (interactive) {
    sigaction(SIGINT, &sa_old_int, NULL);
    sigaction(SIGQUIT, &sa_old_quit, NULL);
  }

  /* push result table */
  push_result_table(L, last_code, &buf_out, &buf_err);
  dynbuf_free(&buf_out);
  dynbuf_free(&buf_err);

  return 1;
}


/* ===== builtin dispatch ===== */

/*
** Check if argv[0] is a shell builtin. If so, call it and push the
** result table. Returns 1 if handled, 0 if not a builtin.
*/
static int try_builtin (lua_State *L, ParsedArgs *pa) {
  int i;
  if (!lookup_lush_func(L, "builtins", pa->argv[0]))
    return 0;
  for (i = 0; i < pa->argc; i++)
    lua_pushstring(L, pa->argv[i]);
  lua_call(L, pa->argc, 1);
  return 1;
}


/* ===== alias expansion ===== */

/*
** Extract the first whitespace-delimited word from cmd, look it up in
** lush.aliases. If found, replace the command string on the Lua stack
** (position 1) with alias_value + remaining_text, and update *cmd.
** Expansion happens on the raw string, before pipeline splitting,
** so alias values may contain pipes.
** Returns 1 if expanded, 0 if no alias.
*/
static int expand_alias (lua_State *L, const char **cmd) {
  const char *s = *cmd;
  const char *rest;
  size_t wordlen;
  const char *alias;
  size_t alias_len, rest_len, total;
  char *expanded;

  /* skip leading whitespace */
  while (*s == ' ' || *s == '\t') s++;
  if (*s == '\0') return 0;

  /* find end of first word */
  rest = s;
  while (*rest && *rest != ' ' && *rest != '\t') rest++;
  wordlen = (size_t)(rest - s);

  /* look up in lush.aliases */
  lua_rawgeti(L, LUA_REGISTRYINDEX, LUA_RIDX_LUSH);
  lua_getfield(L, -1, "aliases");
  lua_pushlstring(L, s, wordlen);
  lua_gettable(L, -2);

  if (!lua_isstring(L, -1)) {
    lua_pop(L, 3);
    return 0;
  }

  alias = lua_tolstring(L, -1, &alias_len);
  rest_len = strlen(rest);  /* includes leading space if any */

  total = alias_len + rest_len;
  expanded = (char *)malloc(total + 1);
  if (expanded == NULL) {
    lua_pop(L, 3);
    return 0;
  }
  memcpy(expanded, alias, alias_len);
  memcpy(expanded + alias_len, rest, rest_len);
  expanded[total] = '\0';
  lua_pop(L, 3);

  /* replace command on Lua stack */
  lua_pushstring(L, expanded);
  lua_replace(L, 1);
  free(expanded);
  *cmd = lua_tostring(L, 1);
  return 1;
}


/* ===== single-command fork/exec/wait ===== */

/*
** Fork, exec a single command, wait, and push a result table.
** If capture is true, stdout/stderr are captured via pipes.
** If capture is false, the child inherits the terminal.
*/
static void fork_exec_single (lua_State *L, ParsedArgs *pa, int capture) {
  int out_pipe[2], err_pipe[2];
  pid_t pid;
  int status;
  DynBuf buf_out, buf_err;
  struct sigaction sa_old_pipe, sa_old_int, sa_old_quit, sa_new;

  if (capture) {
    if (pipe(out_pipe) != 0)
      luaL_error(L, "pipe() failed: %s", strerror(errno));
    if (pipe(err_pipe) != 0) {
      close(out_pipe[0]); close(out_pipe[1]);
      luaL_error(L, "pipe() failed: %s", strerror(errno));
    }
  }

  memset(&sa_new, 0, sizeof(sa_new));
  sa_new.sa_handler = SIG_IGN;
  sigemptyset(&sa_new.sa_mask);
  sigaction(SIGPIPE, &sa_new, &sa_old_pipe);
  if (!capture) {
    sigaction(SIGINT, &sa_new, &sa_old_int);
    sigaction(SIGQUIT, &sa_new, &sa_old_quit);
  }

  pid = fork();
  if (pid < 0) {
    if (capture) {
      close(out_pipe[0]); close(out_pipe[1]);
      close(err_pipe[0]); close(err_pipe[1]);
    }
    sigaction(SIGPIPE, &sa_old_pipe, NULL);
    if (!capture) {
      sigaction(SIGINT, &sa_old_int, NULL);
      sigaction(SIGQUIT, &sa_old_quit, NULL);
    }
    luaL_error(L, "fork() failed: %s", strerror(errno));
  }

  if (pid == 0) {
    /* child */
    sa_new.sa_handler = SIG_DFL;
    sigaction(SIGPIPE, &sa_new, NULL);
    if (!capture) {
      sigaction(SIGINT, &sa_new, NULL);
      sigaction(SIGQUIT, &sa_new, NULL);
    }
    if (capture) {
      close(out_pipe[0]);
      close(err_pipe[0]);
      dup2(out_pipe[1], STDOUT_FILENO);
      dup2(err_pipe[1], STDERR_FILENO);
      close(out_pipe[1]);
      close(err_pipe[1]);
    }
    exec_user_command(L, pa);
    execvp(pa->argv[0], pa->argv);
    exec_failed(pa->argv[0]);
    _exit(127);
  }

  /* parent */
  dynbuf_init(&buf_out);
  dynbuf_init(&buf_err);
  if (capture) {
    close(out_pipe[1]);
    close(err_pipe[1]);
    read_pipes(out_pipe[0], err_pipe[0], &buf_out, &buf_err);
  }

  waitpid(pid, &status, 0);

  sigaction(SIGPIPE, &sa_old_pipe, NULL);
  if (!capture) {
    sigaction(SIGINT, &sa_old_int, NULL);
    sigaction(SIGQUIT, &sa_old_quit, NULL);
  }

  push_result_table(L, exit_code_from_status(status), &buf_out, &buf_err);
  dynbuf_free(&buf_out);
  dynbuf_free(&buf_err);
}


/* ===== lushCmd_command ===== */

int lushCmd_command (lua_State *L) {
  const char *cmd = luaL_checkstring(L, 1);
  char *stages[MAX_PIPELINE_STAGES];
  int nstages;
  ParsedArgs pa;
  DynBuf empty_out, empty_err;

  /* expand alias before pipeline splitting */
  expand_alias(L, &cmd);

  /* try to split into pipeline stages */
  if (split_pipeline(cmd, stages, &nstages) != 0)
    return luaL_error(L, "invalid pipeline syntax in command");

  /* multi-stage pipeline: delegate to exec_pipeline */
  if (nstages > 1) {
    int result = exec_pipeline(L, stages, nstages, 0);
    free(stages[0]);  /* free backing buffer */
    return result;
  }

  /* single stage: free pipeline buffer and use original path */
  free(stages[0]);

  /* parse command into argv */
  if (parse_argv(cmd, &pa) != 0)
    return luaL_error(L, "unterminated quote in command");

  /* empty command: return {code=0, stdout="", stderr=""} */
  if (pa.argc == 0) {
    free_argv(&pa);
    dynbuf_init(&empty_out);
    dynbuf_init(&empty_err);
    push_result_table(L, 0, &empty_out, &empty_err);
    return 1;
  }

  /* check for shell builtin */
  if (try_builtin(L, &pa)) {
    free_argv(&pa);
    return 1;
  }

  fork_exec_single(L, &pa, 1);
  free_argv(&pa);
  return 1;
}


/* ===== lushCmd_subcmd ===== */

/*
** Subcommand substitution: run a command, return stdout with trailing
** newlines stripped.  Used for $(cmd) inside command strings.
*/
int lushCmd_subcmd (lua_State *L) {
  size_t len;
  const char *s;
  /* run the command — pushes {code, stdout, stderr} table */
  lushCmd_command(L);
  /* extract .stdout from result table */
  lua_getfield(L, -1, "stdout");
  lua_remove(L, -2);  /* remove the table, keep stdout string */
  /* strip trailing newlines */
  s = lua_tolstring(L, -1, &len);
  while (len > 0 && s[len - 1] == '\n')
    len--;
  lua_pushlstring(L, s, len);
  lua_remove(L, -2);  /* remove original stdout string */
  return 1;
}


/* ===== lushCmd_interactive ===== */

int lushCmd_interactive (lua_State *L) {
  const char *cmd = luaL_checkstring(L, 1);
  char *stages[MAX_PIPELINE_STAGES];
  int nstages;
  ParsedArgs pa;
  DynBuf empty_out, empty_err;

  /* expand alias before pipeline splitting */
  expand_alias(L, &cmd);

  /* try to split into pipeline stages */
  if (split_pipeline(cmd, stages, &nstages) != 0)
    return luaL_error(L, "invalid pipeline syntax in command");

  /* multi-stage pipeline: delegate to exec_pipeline with interactive=1 */
  if (nstages > 1) {
    exec_pipeline(L, stages, nstages, 1);
    free(stages[0]);
    goto set_and_return;
  }

  /* single stage */
  free(stages[0]);

  if (parse_argv(cmd, &pa) != 0)
    return luaL_error(L, "unterminated quote in command");

  if (pa.argc == 0) {
    free_argv(&pa);
    dynbuf_init(&empty_out);
    dynbuf_init(&empty_err);
    push_result_table(L, 0, &empty_out, &empty_err);
    goto set_and_return;
  }

  /* check for shell builtin */
  if (try_builtin(L, &pa)) {
    free_argv(&pa);
    goto set_and_return;
  }

  fork_exec_single(L, &pa, 0);
  free_argv(&pa);

set_and_return:
  lua_pushvalue(L, -1);    /* duplicate result table */
  lua_setglobal(L, "_");   /* pop one copy into _ */
  return 1;                 /* return the other */
}


/* ===== lushCmd_getenv ===== */

int lushCmd_getenv (lua_State *L) {
  const char *name = luaL_checkstring(L, 1);
  const char *val = getenv(name);
  if (val == NULL)
    lua_pushnil(L);
  else
    lua_pushstring(L, val);
  return 1;
}


/* ===== lushCmd_setenv ===== */

int lushCmd_setenv (lua_State *L) {
  const char *name = luaL_checkstring(L, 1);
  if (lua_isnoneornil(L, 2))
    unsetenv(name);
  else {
    const char *val = luaL_tolstring(L, 2, NULL);
    setenv(name, val, 1);
  }
  return 0;
}


/* ===== lush standard library ===== */

TString *lushname[LUSH_OP_COUNT];

static const luaL_Reg lushlib[] = {
  {"command", lushCmd_command},
  {"interactive", lushCmd_interactive},
  {"getenv", lushCmd_getenv},
  {"setenv", lushCmd_setenv},
  {"subcmd", lushCmd_subcmd},
  {NULL, NULL}
};

LUAMOD_API int luaopen_lush (lua_State *L) {
  luaL_newlib(L, lushlib);
  /* create aliases subtable */
  lua_createtable(L, 0, 4);
  lua_setfield(L, -2, "aliases");
  /* create commands subtable for user-defined commands */
  lua_createtable(L, 0, 4);
  lua_setfield(L, -2, "commands");
  /* intern function name strings for OP_LUSH VM access */
  lushname[LUSH_OP_COMMAND]     = luaS_new(L, "command");
  lushname[LUSH_OP_INTERACTIVE] = luaS_new(L, "interactive");
  lushname[LUSH_OP_GETENV]      = luaS_new(L, "getenv");
  lushname[LUSH_OP_SETENV]      = luaS_new(L, "setenv");
  lushname[LUSH_OP_SUBCMD]      = luaS_new(L, "subcmd");
  /* store in registry for OP_LUSH access */
  lua_pushvalue(L, -1);
  lua_rawseti(L, LUA_REGISTRYINDEX, LUA_RIDX_LUSH);
  luaopen_builtins(L);  /* register builtins in lush table */
  return 1;
}