-- logtail.wo — LogTail.hx, file for file: bounded tail reads over `fs`.
-- One poll = read at most CHUNK new tail bytes, judge only complete lines
-- (plan 02: never scan a file front to back; torn final line held back).
--
-- The three Haxe imports (sys.FileSystem, sys.io.File, sys.io.FileSeek)
-- collapse into one `use fs`: stat carries the inode, read_at takes the
-- offset as a parameter (no seek state), and the file handle lives and dies
-- inside the builtin — RAII instead of the open/try/close dance.

use fs

-- LogTail.hx:6-11. Field defaults replace LogTail.newState(): construction
-- is the brace literal `TailState {}`, the defaults fill in.
typedef TailState = {
  offset:          Int  = 0     -- next read position (past last complete line)
  ino:             Int  = -1    -- inode at last poll, -1 before first sight
  last_level:      Text = ""    -- level of the last complete entry, "" if none
  last_newline_at: Int  = 0     -- ms clock when complete lines last arrived
}

typedef PollResult = {
  exists:     Bool
  new_lines:  Int
  bytes_read: Int
}

const CHUNK = 65536

-- `now` is injected (ms) so tests can drive synthetic time. LogTail.hx:28-75.
pub fn poll(path: Text, mut st: TailState, now: Int) -> PollResult {
  let stat = fs.stat(path);
  if stat == nil { return PollResult { exists: false, new_lines: 0, bytes_read: 0 }; }

  if st.ino == -1 {
    -- first sight: start at most CHUNK before EOF; a partial first line is
    -- read as a continuation, which is acceptable
    st.ino = stat.inode;
    st.offset = 0;
    if stat.size > CHUNK { st.offset = stat.size - CHUNK; }
  } else if stat.inode != st.ino or stat.size < st.offset {
    -- rotation (rename/recreate or truncate): restart from the top
    st.ino = stat.inode;
    st.offset = 0;
    st.last_level = "";
  }
  if stat.size - st.offset > CHUNK { st.offset = stat.size - CHUNK; }  -- burst: jump to tail
  if stat.size <= st.offset { return PollResult { exists: true, new_lines: 0, bytes_read: 0 }; }

  -- Shrunk between stat and read (the Haxe Eof catch): read_at returns what
  -- is actually there; the next poll re-syncs.
  let chunk = fs.read_at(path, st.offset, stat.size - st.offset);
  if len(chunk) == 0 { return PollResult { exists: true, new_lines: 0, bytes_read: 0 }; }

  -- only complete lines count: cut at the last newline, hold the rest
  let nl = last_index_of(chunk, "\n");
  if nl == -1 { return PollResult { exists: true, new_lines: 0, bytes_read: len(chunk) }; }

  let lines = split(substr(chunk, 0, nl + 1), "\n");
  pop(lines);                                  -- empty piece after the final newline
  st.offset = st.offset + nl + 1;
  for line in lines {
    -- relaxed rule (service-health): timestamped app logs must classify
    -- too, or their errors never trip the alert rule
    let lv = classify_loose(sanitize(line));
    if lv != nil { st.last_level = lv; }       -- else continuation: inherits
  }
  if len(lines) > 0 { st.last_newline_at = now; }
  return PollResult { exists: true, new_lines: len(lines), bytes_read: len(chunk) };
}

-- Strict prefix rule for demo/test logs. LogTail.hx:77-82.
pub fn classify(line: Text) -> ?Text {
  if starts_with(line, "info")  { return "info"; }
  if starts_with(line, "warn")  { return "warn"; }
  if starts_with(line, "error") { return "error"; }
  return nil;
}

-- Real app logs put a timestamp first ("2026-07-22T15:40:00 - error: …");
-- the relaxed rule also accepts the level after a leading token. Haxe used
-- an EReg (LogTail.hx:87, ~/^\S+\s+-\s+(info|warn|error)\b/); with no regex
-- in the language the same rule is spelled out: token, lone dash, level
-- with a word boundary. Callers pass a sanitized line — ANSI codes hide
-- the prefix.
pub fn classify_loose(line: Text) -> ?Text {
  let lv = classify(line);
  if lv != nil { return lv; }
  let tokens = split_ws(line);
  if len(tokens) < 3 { return nil; }
  if tokens[1] != "-" { return nil; }
  return level_bounded(tokens[2]);
}

-- The regex's \b: "error:" and "error," carry the level; "errors" does not.
fn level_bounded(tok: Text) -> ?Text {
  for lv in ["info", "warn", "error"] {
    if starts_with(tok, lv) {
      if len(tok) == len(lv) { return lv; }
      if is_word_byte(byte_at(tok, len(lv))) { return nil; }
      return lv;
    }
  }
  return nil;
}

fn is_word_byte(c: Int) -> Bool {
  if c >= 48 and c <= 57 { return true; }      -- 0-9
  if c >= 65 and c <= 90 { return true; }      -- A-Z
  if c >= 97 and c <= 122 { return true; }     -- a-z
  return c == 95;                              -- _
}

-- Tools.hx:24-34, moved here beside its heaviest caller (poll). Log lines
-- can carry ANSI color sequences and stray control bytes; they would break
-- classification and produce invalid JSON at the client. Strip ESC[…letter
-- sequences, drop other C0 bytes (tab stays). The Haxe EReg becomes an
-- explicit byte scan.
pub fn sanitize(line: Text) -> Text {
  let out = "";
  let i = 0;
  while i < len(line) {
    let c = byte_at(line, i);
    if c == 27 and i + 1 < len(line) and byte_at(line, i + 1) == 91 {
      i = i + 2;                               -- skip ESC [
      while i < len(line) {
        let f = byte_at(line, i);
        i = i + 1;
        if (f >= 65 and f <= 90) or (f >= 97 and f <= 122) { break; }
      }
      continue;
    }
    if c >= 32 or c == 9 { out = out + char_of(c); }
    i = i + 1;
  }
  return out;
}

-- Last <= n complete lines from the final CHUNK bytes of the file. A
-- cut-off first line is acceptable (same as first-sight poll); an
-- unterminated final line is dropped. nil when the file is missing.
-- LogTail.hx:98-120.
pub fn last_lines(path: Text, n: Int) -> ?multi Text {
  let stat = fs.stat(path);
  if stat == nil { return nil; }
  let start = 0;
  if stat.size > CHUNK { start = stat.size - CHUNK; }
  if stat.size - start <= 0 { return []; }
  let chunk = fs.read_at(path, start, stat.size - start);
  if len(chunk) == 0 { return []; }
  let nl = last_index_of(chunk, "\n");
  if nl == -1 { return []; }
  let lines = split(substr(chunk, 0, nl + 1), "\n");
  pop(lines);                                  -- empty piece after the final newline
  if start > 0 and len(lines) > 0 { shift(lines); }   -- cut-off first line
  if len(lines) > n { return slice(lines, len(lines) - n, len(lines)); }
  return lines;
}