#include "control.h"
#include "cpu.h"
#include "render.h"

#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <strings.h>
#include <stdarg.h>
#include <ctype.h>
#include <errno.h>
#include <unistd.h>
#include <fcntl.h>
#include <signal.h>
#include <sys/socket.h>
#include <sys/un.h>

// ---------------------------------------------------------------------------
// Socket-based control + debug channel. See control.h for the protocol summary.
// ---------------------------------------------------------------------------

#define MAX_CLIENTS 8
#define MAX_BP      64
#define MAX_WP      32
#define LINEBUF     512
#define READ_CAP    4096   // max bytes per `read` command

typedef struct {
    int  fd;
    char line[LINEBUF];
    int  len;
} Client;

static bool  active = false;
static int   listen_fd = -1;
static char  sock_path[512];
static Client clients[MAX_CLIENTS];
static bool  quit_requested = false;

// ---- debugger execution state ----
enum { RUN, PAUSED };
static int  run_state = RUN;
static int  step_remaining = 0;     // synchronous stepping is done inline, but
                                    // a nonzero value also lets run_frame step
static bool ignore_bp_once = false; // don't re-break on the addr we resumed from

static u16  bp_addr[MAX_BP];
static bool bp_used[MAX_BP];

static u16  wp_addr[MAX_WP];
static bool wp_used[MAX_WP];
static u8   wp_last[MAX_WP];

// Last reason execution halted ("breakpoint 0x0150", "watch 0xFF44", "step",
// "pause"), so a per-request client that wasn't connected when an async event
// fired can still recover it via the `stopinfo` command.
static char last_stop[64] = "none";

// ---- frame recording ----
// Dump the RGB888 framebuffer of each *produced* frame to a flat capture file:
// a 10-byte header ("GBCV", u16 width, u16 height, u16 ms-per-frame, all LE)
// followed by raw width*height*3 byte frames. tools/gbgif.py turns it into a
// GIF. Downsample with everyN (capture 1 of every N frames) for smaller files.
static FILE *rec_fp = NULL;
static u32   rec_every = 1;
static u32   rec_phase = 0;
static u64   rec_written = 0;
static char  rec_path[512];

// ---- input replay (TAS-style movie playback) ----
// A movie is a text file of `<frames> [buttons...]` lines (see input_btn_bit);
// while playing, each emulated frame's joypad state is driven from the movie
// instead of live input, so a synthesized demo replays deterministically. Pair
// with `record` (and a reset) to produce reproducible showcase GIFs.
static FILE *movie_fp = NULL;
static int   movie_hold = 0;    // frames left to hold the current mask
static u8    movie_mask = 0;
static u64   movie_frame = 0;
static char  movie_path[512];

// ---------------------------------------------------------------------------
// small helpers
// ---------------------------------------------------------------------------

static void set_nonblock(int fd) {
    int fl = fcntl(fd, F_GETFL, 0);
    if (fl >= 0) fcntl(fd, F_SETFL, fl | O_NONBLOCK);
}

// Parse a decimal or 0x-hex number. On success returns value and *ok=true.
static long parse_num(const char *s, bool *ok) {
    if (!s || !*s) { *ok = false; return 0; }
    char *end = NULL;
    errno = 0;
    long v = strtol(s, &end, 0);   // base 0: handles 0x.. and decimal
    *ok = (errno == 0 && end && *end == '\0');
    return v;
}

static void reply(int fd, const char *fmt, ...) {
    char buf[READ_CAP * 2 + 256];
    va_list ap;
    va_start(ap, fmt);
    int n = vsnprintf(buf, sizeof(buf) - 2, fmt, ap);
    va_end(ap);
    if (n < 0) return;
    if (n > (int)sizeof(buf) - 2) n = sizeof(buf) - 2;
    buf[n++] = '\n';
    // best-effort; ignore short writes / EPIPE
    ssize_t w = write(fd, buf, n);
    (void)w;
}

static void broadcast(const char *fmt, ...) {
    char buf[512];
    va_list ap;
    va_start(ap, fmt);
    int n = vsnprintf(buf, sizeof(buf) - 2, fmt, ap);
    va_end(ap);
    if (n < 0) return;
    if (n > (int)sizeof(buf) - 2) n = sizeof(buf) - 2;
    buf[n++] = '\n';
    for (int i = 0; i < MAX_CLIENTS; i++)
        if (clients[i].fd >= 0) {
            ssize_t w = write(clients[i].fd, buf, n);
            (void)w;
        }
}

// Format the CPU/machine context into a single line of key=value pairs.
static int fmt_cpu(GB *gb, char *buf, size_t n) {
    CPU *c = &gb->cpu;
    return snprintf(buf, n,
        "af=0x%04X bc=0x%04X de=0x%04X hl=0x%04X sp=0x%04X pc=0x%04X "
        "ime=%d halted=%d cycles=%llu",
        c->af, c->bc, c->de, c->hl, c->sp, c->pc,
        c->ime ? 1 : 0, c->halted ? 1 : 0, (unsigned long long)gb->cycles);
}

// ---------------------------------------------------------------------------
// breakpoints / watchpoints
// ---------------------------------------------------------------------------

static bool bp_hit(u16 pc) {
    for (int i = 0; i < MAX_BP; i++)
        if (bp_used[i] && bp_addr[i] == pc) return true;
    return false;
}

// Return watchpoint index whose byte changed since last check (and update its
// remembered value), or -1 if none.
static int wp_changed(GB *gb) {
    for (int i = 0; i < MAX_WP; i++) {
        if (!wp_used[i]) continue;
        u8 v = gb_read(gb, wp_addr[i]);
        if (v != wp_last[i]) { u8 old = wp_last[i]; wp_last[i] = v;
                               (void)old; return i; }
    }
    return -1;
}

static void enter_paused(void) { run_state = PAUSED; step_remaining = 0; }

static void set_stop(const char *fmt, ...) {
    va_list ap; va_start(ap, fmt);
    vsnprintf(last_stop, sizeof(last_stop), fmt, ap);
    va_end(ap);
}

// ---------------------------------------------------------------------------
// synchronous single-stepping (used by the `step` command for an immediate
// reply). Executes up to `n` instructions, stopping early on a breakpoint or
// watchpoint. Writes a reason string ("step"/"breakpoint"/"watch") and, for a
// watch hit, the watched address into *waddr.
// ---------------------------------------------------------------------------
static const char *do_steps(GB *gb, int n, u16 *waddr) {
    const char *reason = "step";
    bool first = true;
    for (int i = 0; i < n; i++) {
        // honor a breakpoint we land on, but never on the very first step off
        // the current address (otherwise we could never leave it)
        if (!first && bp_hit(gb->cpu.pc)) {
            set_stop("breakpoint 0x%04X", gb->cpu.pc); return "breakpoint";
        }
        first = false;
        cpu_step(gb);
        int w = wp_changed(gb);
        if (w >= 0) { if (waddr) *waddr = wp_addr[w];
                      set_stop("watch 0x%04X", wp_addr[w]); return "watch"; }
    }
    set_stop("step 0x%04X", gb->cpu.pc);
    return reason;
}

// ---------------------------------------------------------------------------
// command dispatch
// ---------------------------------------------------------------------------

static void cmd_break_list(int fd) {
    char buf[400]; int o = 0;
    o += snprintf(buf + o, sizeof(buf) - o, "ok breaks");
    for (int i = 0; i < MAX_BP; i++)
        if (bp_used[i])
            o += snprintf(buf + o, sizeof(buf) - o, " #%d=0x%04X", i, bp_addr[i]);
    reply(fd, "%s", buf);
}

static void cmd_watch_list(int fd) {
    char buf[400]; int o = 0;
    o += snprintf(buf + o, sizeof(buf) - o, "ok watches");
    for (int i = 0; i < MAX_WP; i++)
        if (wp_used[i])
            o += snprintf(buf + o, sizeof(buf) - o, " #%d=0x%04X:0x%02X",
                          i, wp_addr[i], wp_last[i]);
    reply(fd, "%s", buf);
}

// Set a named register/flag. Returns true on success.
static bool set_reg(GB *gb, const char *name, long v) {
    CPU *c = &gb->cpu;
    u16 w = (u16)v; u8 b = (u8)v;
    if (!strcasecmp(name, "a"))   { c->a = b; return true; }
    if (!strcasecmp(name, "f"))   { c->f = b & 0xF0; return true; }
    if (!strcasecmp(name, "b"))   { c->b = b; return true; }
    if (!strcasecmp(name, "c"))   { c->c = b; return true; }
    if (!strcasecmp(name, "d"))   { c->d = b; return true; }
    if (!strcasecmp(name, "e"))   { c->e = b; return true; }
    if (!strcasecmp(name, "h"))   { c->h = b; return true; }
    if (!strcasecmp(name, "l"))   { c->l = b; return true; }
    if (!strcasecmp(name, "af"))  { c->af = w & 0xFFF0; return true; }
    if (!strcasecmp(name, "bc"))  { c->bc = w; return true; }
    if (!strcasecmp(name, "de"))  { c->de = w; return true; }
    if (!strcasecmp(name, "hl"))  { c->hl = w; return true; }
    if (!strcasecmp(name, "sp"))  { c->sp = w; return true; }
    if (!strcasecmp(name, "pc"))  { c->pc = w; return true; }
    if (!strcasecmp(name, "ime")) { c->ime = (v != 0); return true; }
    return false;
}

static int hexval(int ch) {
    if (ch >= '0' && ch <= '9') return ch - '0';
    if (ch >= 'a' && ch <= 'f') return ch - 'a' + 10;
    if (ch >= 'A' && ch <= 'F') return ch - 'A' + 10;
    return -1;
}

// Parse a single byte written as hex ("de", "0xDE", "FF", "7"). Memory writes
// default to hex so they round-trip with `read` output. Returns 0..255 or -1.
static int parse_byte_hex(const char *s) {
    if (!s || !*s) return -1;
    if (s[0] == '0' && (s[1] == 'x' || s[1] == 'X')) s += 2;
    int val = 0, digits = 0;
    for (; *s; s++) {
        int h = hexval((unsigned char)*s);
        if (h < 0) return -1;
        val = (val << 4) | h;
        if (++digits > 2) return -1;
    }
    return digits ? val : -1;
}

// Dispatch one command line from client `fd`. `orig` is the untouched line
// (used for the button-command fallback); `line` is a mutable copy.
// movie button name -> our pressed-state bit (see gb.h BTN_*). Anything else
// ("-", "none", "idle", unknown) contributes nothing = released.
static u8 input_btn_bit(const char *s) {
    if (!strcasecmp(s, "a"))      return BTN_A;
    if (!strcasecmp(s, "b"))      return BTN_B;
    if (!strcasecmp(s, "start"))  return BTN_START;
    if (!strcasecmp(s, "select")) return BTN_SELECT;
    if (!strcasecmp(s, "up"))     return BTN_UP;
    if (!strcasecmp(s, "down"))   return BTN_DOWN;
    if (!strcasecmp(s, "left"))   return BTN_LEFT;
    if (!strcasecmp(s, "right"))  return BTN_RIGHT;
    return 0;
}

// Exact, repeatable power-on state so a movie replays identically every time.
// Startup is calloc(zero) -> cart_load -> gb_reset, so we reproduce it: preserve
// the loaded cart (ROM + SRAM, so the save is intact), zero *all* other emulator
// state, then the normal post-boot reset. Zeroing only some RAM (and leaving
// e.g. double-speed / timer / DMA residue from the prior session) makes the boot
// diverge run to run -- this wipes the lot.
static void replay_reset(GB *gb) {
    Cart cart = gb->cart;            // struct of pointers into the loaded ROM/SRAM
    memset(gb, 0, sizeof(*gb));
    gb->cart = cart;
    gb_reset(gb);
}

static void dispatch(GB *gb, int fd, const char *orig, char *line) {
    // tokenize (whitespace); keep argv for structured commands
    char *argv[68];
    int argc = 0;
    for (char *t = strtok(line, " \t\r\n"); t && argc < 68;
         t = strtok(NULL, " \t\r\n"))
        argv[argc++] = t;
    if (argc == 0) return;

    const char *v = argv[0];
    char cbuf[256];

    if (!strcasecmp(v, "ping")) { reply(fd, "pong"); return; }
    if (!strcasecmp(v, "help")) {
        reply(fd, "ok commands: ping state stopinfo cpu reg read write step "
                  "continue pause break watch delete unwatch quit | button tokens "
                  "(a b start select up down left right, +x -x, name:N, release)");
        return;
    }
    if (!strcasecmp(v, "quit") || !strcasecmp(v, "exit")) {
        quit_requested = true; reply(fd, "ok bye"); return;
    }
    if (!strcasecmp(v, "state")) {
        reply(fd, "ok state %s", run_state == PAUSED ? "paused" : "running");
        return;
    }
    if (!strcasecmp(v, "stopinfo") || !strcasecmp(v, "laststop") ||
        !strcasecmp(v, "why")) {
        fmt_cpu(gb, cbuf, sizeof(cbuf));
        if (run_state == PAUSED) reply(fd, "ok stop %s %s", last_stop, cbuf);
        else reply(fd, "ok running %s", cbuf);
        return;
    }
    if (!strcasecmp(v, "cpu") || !strcasecmp(v, "regs") ||
        !strcasecmp(v, "context")) {
        fmt_cpu(gb, cbuf, sizeof(cbuf));
        reply(fd, "ok cpu %s", cbuf);
        return;
    }
    if (!strcasecmp(v, "reg")) {
        bool ok; long val;
        if (argc < 3) { reply(fd, "err reg <name> <value>"); return; }
        val = parse_num(argv[2], &ok);
        if (!ok) { reply(fd, "err bad value"); return; }
        if (!set_reg(gb, argv[1], val)) { reply(fd, "err unknown reg '%s'",
                                                 argv[1]); return; }
        reply(fd, "ok reg %s=0x%04X", argv[1], (unsigned)val);
        return;
    }
    if (!strcasecmp(v, "read") || !strcasecmp(v, "r") ||
        !strcasecmp(v, "mem")) {
        bool ok; long addr, len = 1;
        if (argc < 2) { reply(fd, "err read <addr> [len]"); return; }
        addr = parse_num(argv[1], &ok);
        if (!ok) { reply(fd, "err bad addr"); return; }
        if (argc >= 3) { len = parse_num(argv[2], &ok);
                         if (!ok) { reply(fd, "err bad len"); return; } }
        if (len < 1) len = 1;
        if (len > READ_CAP) len = READ_CAP;
        static char hex[READ_CAP * 2 + 1];
        int o = 0;
        for (long i = 0; i < len; i++) {
            u8 val = gb_read(gb, (u16)(addr + i));
            static const char *H = "0123456789abcdef";
            hex[o++] = H[val >> 4];
            hex[o++] = H[val & 0xF];
        }
        hex[o] = 0;
        reply(fd, "ok read 0x%04X %ld %s", (unsigned)addr, len, hex);
        return;
    }
    if (!strcasecmp(v, "write") || !strcasecmp(v, "w")) {
        bool ok;
        if (argc < 3) { reply(fd, "err write <addr> <byte...|hexstring>");
                        return; }
        long addr = parse_num(argv[1], &ok);
        if (!ok) { reply(fd, "err bad addr"); return; }
        int count = 0;
        // form 1: a single contiguous even-length hex string "deadbeef"
        if (argc == 3) {
            const char *s = argv[2];
            size_t sl = strlen(s);
            bool allhex = (sl >= 2 && sl % 2 == 0);
            for (size_t i = 0; i < sl && allhex; i++)
                if (hexval(s[i]) < 0) allhex = false;
            if (allhex) {
                for (size_t i = 0; i + 1 < sl; i += 2) {
                    u8 val = (u8)((hexval(s[i]) << 4) | hexval(s[i + 1]));
                    gb_write(gb, (u16)(addr + count), val);
                    count++;
                }
                reply(fd, "ok write 0x%04X %d", (unsigned)addr, count);
                return;
            }
        }
        // form 2: space-separated hex bytes ("de ad be ef", 0x-prefix ok)
        (void)ok;
        for (int i = 2; i < argc; i++) {
            int b = parse_byte_hex(argv[i]);
            if (b < 0) { reply(fd, "err bad byte '%s'", argv[i]); return; }
            gb_write(gb, (u16)(addr + count), (u8)b);
            count++;
        }
        reply(fd, "ok write 0x%04X %d", (unsigned)addr, count);
        return;
    }
    if (!strcasecmp(v, "step") || !strcasecmp(v, "s") ||
        !strcasecmp(v, "si")) {
        bool ok; long n = 1;
        if (argc >= 2) { n = parse_num(argv[1], &ok);
                         if (!ok || n < 1) n = 1; }
        ignore_bp_once = true;
        u16 waddr = 0;
        const char *reason = do_steps(gb, (int)n, &waddr);
        enter_paused();
        fmt_cpu(gb, cbuf, sizeof(cbuf));
        if (!strcmp(reason, "watch"))
            reply(fd, "ok stop watch 0x%04X %s", waddr, cbuf);
        else
            reply(fd, "ok stop %s %s", reason, cbuf);
        return;
    }
    if (!strcasecmp(v, "continue") || !strcasecmp(v, "cont") ||
        !strcasecmp(v, "c") || !strcasecmp(v, "run")) {
        run_state = RUN;
        step_remaining = 0;
        ignore_bp_once = true;   // step off current bp/pc before re-checking
        reply(fd, "ok running");
        return;
    }
    if (!strcasecmp(v, "pause") || !strcasecmp(v, "stop") ||
        !strcasecmp(v, "halt") || !strcasecmp(v, "break!")) {
        enter_paused();
        set_stop("pause 0x%04X", gb->cpu.pc);
        fmt_cpu(gb, cbuf, sizeof(cbuf));
        reply(fd, "ok paused %s", cbuf);
        return;
    }
    if (!strcasecmp(v, "break") || !strcasecmp(v, "bp") ||
        !strcasecmp(v, "b")) {
        if (argc < 2) { cmd_break_list(fd); return; }
        bool ok; long addr = parse_num(argv[1], &ok);
        if (!ok) { reply(fd, "err bad addr"); return; }
        for (int i = 0; i < MAX_BP; i++)      // dedup
            if (bp_used[i] && bp_addr[i] == (u16)addr) {
                reply(fd, "ok break #%d 0x%04X", i, (unsigned)addr); return; }
        for (int i = 0; i < MAX_BP; i++)
            if (!bp_used[i]) { bp_used[i] = true; bp_addr[i] = (u16)addr;
                reply(fd, "ok break #%d 0x%04X", i, (unsigned)addr); return; }
        reply(fd, "err breakpoint table full");
        return;
    }
    if (!strcasecmp(v, "delete") || !strcasecmp(v, "del") ||
        !strcasecmp(v, "unbreak") || !strcasecmp(v, "d")) {
        if (argc < 2) { reply(fd, "err delete <0xaddr|#idx|all>"); return; }
        if (!strcasecmp(argv[1], "all")) {
            memset(bp_used, 0, sizeof(bp_used));
            reply(fd, "ok deleted all"); return;
        }
        if (argv[1][0] == '#') {
            int idx = atoi(argv[1] + 1);
            if (idx >= 0 && idx < MAX_BP && bp_used[idx]) {
                bp_used[idx] = false; reply(fd, "ok deleted #%d", idx); return; }
            reply(fd, "err no such breakpoint"); return;
        }
        bool ok; long addr = parse_num(argv[1], &ok);
        if (!ok) { reply(fd, "err bad addr"); return; }
        for (int i = 0; i < MAX_BP; i++)
            if (bp_used[i] && bp_addr[i] == (u16)addr) {
                bp_used[i] = false;
                reply(fd, "ok deleted #%d 0x%04X", i, (unsigned)addr); return; }
        reply(fd, "err no breakpoint at 0x%04X", (unsigned)addr);
        return;
    }
    if (!strcasecmp(v, "watch") || !strcasecmp(v, "wp")) {
        if (argc < 2) { cmd_watch_list(fd); return; }
        bool ok; long addr = parse_num(argv[1], &ok);
        if (!ok) { reply(fd, "err bad addr"); return; }
        for (int i = 0; i < MAX_WP; i++)
            if (!wp_used[i]) {
                wp_used[i] = true; wp_addr[i] = (u16)addr;
                wp_last[i] = gb_read(gb, (u16)addr);
                reply(fd, "ok watch #%d 0x%04X:0x%02X", i, (unsigned)addr,
                      wp_last[i]);
                return;
            }
        reply(fd, "err watchpoint table full");
        return;
    }
    if (!strcasecmp(v, "unwatch")) {
        if (argc < 2) { reply(fd, "err unwatch <0xaddr|#idx|all>"); return; }
        if (!strcasecmp(argv[1], "all")) {
            memset(wp_used, 0, sizeof(wp_used));
            reply(fd, "ok unwatched all"); return;
        }
        if (argv[1][0] == '#') {
            int idx = atoi(argv[1] + 1);
            if (idx >= 0 && idx < MAX_WP && wp_used[idx]) {
                wp_used[idx] = false; reply(fd, "ok unwatched #%d", idx);
                return; }
            reply(fd, "err no such watchpoint"); return;
        }
        bool ok; long addr = parse_num(argv[1], &ok);
        if (!ok) { reply(fd, "err bad addr"); return; }
        for (int i = 0; i < MAX_WP; i++)
            if (wp_used[i] && wp_addr[i] == (u16)addr) {
                wp_used[i] = false;
                reply(fd, "ok unwatched 0x%04X", (unsigned)addr); return; }
        reply(fd, "err no watchpoint at 0x%04X", (unsigned)addr);
        return;
    }

    if (!strcasecmp(v, "record") || !strcasecmp(v, "rec")) {
        const char *sub = argc >= 2 ? argv[1] : "status";
        if (!strcasecmp(sub, "start")) {
            if (argc < 3) { reply(fd, "err record start <path> [everyN]"); return; }
            if (rec_fp) { fclose(rec_fp); rec_fp = NULL; }
            u32 every = 1;
            if (argc >= 4) { bool ok; long n = parse_num(argv[3], &ok);
                             if (ok && n > 0) every = (u32)n; }
            FILE *f = fopen(argv[2], "wb");
            if (!f) { reply(fd, "err cannot open '%s'", argv[2]); return; }
            // Capture what the user sees: chrome frame + LCD recolor if enabled.
            int cw, ch; render_capture_size(&cw, &ch);
            u16 w = (u16)cw, h = (u16)ch;
            // playback interval per captured frame (the GB runs ~59.73 fps)
            u16 ms = (u16)(1000.0 * every / 59.73 + 0.5);
            u8 hdr[10] = { 'G','B','C','V',
                           (u8)(w & 0xff), (u8)(w >> 8),
                           (u8)(h & 0xff), (u8)(h >> 8),
                           (u8)(ms & 0xff), (u8)(ms >> 8) };
            fwrite(hdr, 1, sizeof(hdr), f);
            rec_fp = f; rec_every = every; rec_phase = 0; rec_written = 0;
            snprintf(rec_path, sizeof(rec_path), "%s", argv[2]);
            reply(fd, "ok record start %s every=%u", rec_path, (unsigned)every);
            return;
        }
        if (!strcasecmp(sub, "stop")) {
            if (!rec_fp) { reply(fd, "err not recording"); return; }
            fclose(rec_fp); rec_fp = NULL;
            reply(fd, "ok record stop %s frames=%llu",
                  rec_path, (unsigned long long)rec_written);
            return;
        }
        // status
        if (rec_fp) reply(fd, "ok record active %s frames=%llu every=%u",
                          rec_path, (unsigned long long)rec_written,
                          (unsigned)rec_every);
        else reply(fd, "ok record inactive");
        return;
    }

    if (!strcasecmp(v, "input") || !strcasecmp(v, "movie") ||
        !strcasecmp(v, "play")) {
        const char *sub = argc >= 2 ? argv[1] : "status";
        if (!strcasecmp(sub, "play") || !strcasecmp(sub, "load")) {
            if (argc < 3) { reply(fd, "err input play <path> [reset]"); return; }
            if (movie_fp) { fclose(movie_fp); movie_fp = NULL; }
            FILE *f = fopen(argv[2], "r");
            if (!f) { reply(fd, "err cannot open '%s'", argv[2]); return; }
            bool do_reset = (argc >= 4 && !strcasecmp(argv[3], "reset"));
            if (do_reset) replay_reset(gb);
            movie_fp = f; movie_hold = 0; movie_mask = 0; movie_frame = 0;
            snprintf(movie_path, sizeof movie_path, "%s", argv[2]);
            reply(fd, "ok input play %s%s", movie_path,
                  do_reset ? " reset" : "");
            return;
        }
        if (!strcasecmp(sub, "stop")) {
            if (!movie_fp) { reply(fd, "err not playing"); return; }
            fclose(movie_fp); movie_fp = NULL; movie_hold = 0;
            gb->buttons = 0;
            reply(fd, "ok input stop %s frame=%llu", movie_path,
                  (unsigned long long)movie_frame);
            return;
        }
        if (movie_fp) reply(fd, "ok input playing %s frame=%llu", movie_path,
                            (unsigned long long)movie_frame);
        else reply(fd, "ok input idle");
        return;
    }

    // fallback: treat the whole line as button-command tokens (legacy FIFO
    // vocabulary). Reply ok so scripted callers can synchronize.
    input_handle_command(orig);
    reply(fd, "ok");
}

// ---------------------------------------------------------------------------
// connection handling
// ---------------------------------------------------------------------------

static void drop_client(int i) {
    if (clients[i].fd >= 0) { close(clients[i].fd); clients[i].fd = -1; }
    clients[i].len = 0;
}

static void accept_new(void) {
    for (;;) {
        int cf = accept(listen_fd, NULL, NULL);
        if (cf < 0) break;                 // EAGAIN / no pending
        set_nonblock(cf);
        int slot = -1;
        for (int i = 0; i < MAX_CLIENTS; i++)
            if (clients[i].fd < 0) { slot = i; break; }
        if (slot < 0) { close(cf); continue; }   // table full
        clients[slot].fd = cf;
        clients[slot].len = 0;
        reply(cf, "ok sl0pboy control channel; 'help' for commands");
    }
}

static void service_client(GB *gb, int i) {
    char buf[512];
    for (;;) {
        int n = (int)read(clients[i].fd, buf, sizeof(buf));
        if (n == 0) { drop_client(i); return; }       // peer closed
        if (n < 0) break;                              // EAGAIN
        for (int k = 0; k < n; k++) {
            char c = buf[k];
            if (c == '\n' || c == '\r') {
                clients[i].line[clients[i].len] = 0;
                if (clients[i].len) {
                    char orig[LINEBUF];
                    memcpy(orig, clients[i].line, clients[i].len + 1);
                    dispatch(gb, clients[i].fd, orig, clients[i].line);
                }
                clients[i].len = 0;
            } else if (clients[i].len < LINEBUF - 1) {
                clients[i].line[clients[i].len++] = c;
            }
        }
    }
}

// ---------------------------------------------------------------------------
// public API
// ---------------------------------------------------------------------------

void control_open(const char *path) {
    // A client can connect and vanish before we finish writing (e.g. a liveness
    // probe that connects and closes without reading the greeting). Ignore
    // SIGPIPE so those writes fail with EPIPE instead of killing the emulator.
    signal(SIGPIPE, SIG_IGN);
    for (int i = 0; i < MAX_CLIENTS; i++) clients[i].fd = -1;

    listen_fd = socket(AF_UNIX, SOCK_STREAM, 0);
    if (listen_fd < 0) { perror("socket"); return; }

    struct sockaddr_un sa;
    memset(&sa, 0, sizeof(sa));
    sa.sun_family = AF_UNIX;
    snprintf(sa.sun_path, sizeof(sa.sun_path), "%s", path);
    snprintf(sock_path, sizeof(sock_path), "%s", path);

    unlink(sock_path);                     // clear a stale socket
    if (bind(listen_fd, (struct sockaddr *)&sa, sizeof(sa)) < 0) {
        perror("bind"); close(listen_fd); listen_fd = -1; return;
    }
    if (listen(listen_fd, 4) < 0) {
        perror("listen"); close(listen_fd); listen_fd = -1;
        unlink(sock_path); return;
    }
    set_nonblock(listen_fd);
    active = true;
}

void control_close(void) {
    if (rec_fp) { fclose(rec_fp); rec_fp = NULL; }
    if (movie_fp) { fclose(movie_fp); movie_fp = NULL; }
    if (!active) return;
    for (int i = 0; i < MAX_CLIENTS; i++) drop_client(i);
    if (listen_fd >= 0) { close(listen_fd); listen_fd = -1; }
    unlink(sock_path);
    active = false;
}

// Called once per main-loop iteration, right after control_run_frame(). Appends
// the just-produced frame to the capture file (honoring the everyN downsample).
// Gated on frame_ready so a paused/stepping machine doesn't re-record a still.
void control_record_frame(GB *gb) {
    if (!rec_fp || !gb->ppu.frame_ready) return;
    gb->ppu.frame_ready = false;   // consume: don't re-capture while paused
    if ((rec_phase++ % rec_every) != 0) return;
    int w, h;
    const u8 *frame = render_capture_frame(gb, &w, &h);   // matches the display
    fwrite(frame, 1, (size_t)w * h * 3, rec_fp);
    rec_written++;
}

// Called once per main-loop iteration BEFORE control_run_frame(): if a movie is
// playing, drive this frame's joypad state from it (overriding live input).
// Frame-locked, so a synthesized demo replays deterministically. No-op (live
// input passes through) when idle or while the debugger holds execution.
void control_input_frame(GB *gb) {
    if (!movie_fp || run_state == PAUSED) return;
    while (movie_hold <= 0) {
        char buf[256];
        if (!fgets(buf, sizeof buf, movie_fp)) {   // end of movie
            fclose(movie_fp); movie_fp = NULL;
            gb->buttons = 0;                       // release everything
            return;
        }
        char *hash = strchr(buf, '#');
        if (hash) *hash = 0;
        char *tok = strtok(buf, " \t\r\n");
        if (!tok) continue;                        // blank / comment-only line
        bool ok; long n = parse_num(tok, &ok);
        if (!ok || n <= 0) continue;               // need a positive frame count
        u8 mask = 0;
        for (char *t = strtok(NULL, " \t\r\n"); t; t = strtok(NULL, " \t\r\n"))
            mask |= input_btn_bit(t);
        movie_hold = (int)n;
        movie_mask = mask;
    }
    gb->buttons = movie_mask;
    movie_hold--;
    movie_frame++;
}

bool control_active(void) { return active; }
bool control_paused(void) { return active && run_state == PAUSED; }

bool control_poll(GB *gb) {
    if (!active) return true;
    accept_new();
    for (int i = 0; i < MAX_CLIENTS; i++)
        if (clients[i].fd >= 0) service_client(gb, i);
    return !quit_requested;
}

void control_run_frame(GB *gb) {
    // Plain full-frame emulation when there's no active debug channel.
    if (!active) {
        gb->ppu.frame_ready = false;
        u64 budget = gb->cycles + 70224 * 2;
        while (!gb->ppu.frame_ready && gb->cycles < budget) cpu_step(gb);
        return;
    }

    // Fully paused (not stepping): freeze the machine entirely.
    if (run_state == PAUSED) return;

    gb->ppu.frame_ready = false;
    u64 budget = gb->cycles + 70224 * 2;
    while (gb->cycles < budget) {
        // breakpoint check before executing the next instruction
        if (!ignore_bp_once && bp_hit(gb->cpu.pc)) {
            enter_paused();
            set_stop("breakpoint 0x%04X", gb->cpu.pc);
            char cbuf[256]; fmt_cpu(gb, cbuf, sizeof(cbuf));
            broadcast("event stop breakpoint %s", cbuf);
            return;
        }
        ignore_bp_once = false;

        cpu_step(gb);

        int w = wp_changed(gb);
        if (w >= 0) {
            enter_paused();
            set_stop("watch 0x%04X", wp_addr[w]);
            char cbuf[256]; fmt_cpu(gb, cbuf, sizeof(cbuf));
            broadcast("event stop watch 0x%04X %s", wp_addr[w], cbuf);
            return;
        }

        if (gb->ppu.frame_ready) return;
    }
}