"""Native vocabulary, seven-row simultaneous GUI fill contract v1. Coordinates are three separately encoded digits. No tokenizer mutation, action head, autoregressive decoding, or executable GUI action is provided here. """ import math import string from .gui_fill_table import KEYS as NAMED_KEYS from .gui_fill_table import MASK, MODIFIERS, NA from .visual_nanojev_data import validate_step SCHEMA = ( ("Click", ("coord", "coord", "button", "count")), ("TypeText", ("yes",)), ("PressKeys", ("key",) * 4), ("Scroll", ("direction", "level")), ("Wait", ("duration",)), ("Drag", ("coord",) * 4), ("Done", ("yes",)), ) WIDTHS = {"coord": 3, "key": 3} # Symbols are logical source keys, not a claim of cross-layout physical mapping. KEYS = NAMED_KEYS | frozenset(string.punctuation) DOMAINS = { "coord": tuple(f"{i:03d}" for i in range(1000)), "key": tuple(sorted(KEYS)), "button": ("LEFT", "RIGHT"), "count": ("0", "1", "2"), "yes": ("YES",), "direction": ("UP", "DOWN"), "level": ("1", "2", "3"), "duration": ("1", "2", "4"), } def coordinate(value): if type(value) not in (int, float) or not math.isfinite(value) or not 0 <= value <= 1: raise ValueError("source_coordinate_out_of_range") return f"{math.floor(value * 999 + 0.5):03d}" def canonical_keys(keys): aliases = { "RETURN": "ENTER", "ESCAPE": "ESC", "CONTROL": "CTRL", "COMMAND": "CMD", " ": "SPACE", } keys = [aliases.get(k.upper(), k.upper()) for k in keys] if not 1 <= len(keys) <= 4 or len(set(keys)) != len(keys) or any(k not in KEYS for k in keys): raise ValueError("invalid_keys") return [k for k in MODIFIERS if k in keys] + sorted(k for k in keys if k not in MODIFIERS) def validate_action(operation, values): kinds = dict(SCHEMA).get(operation) if kinds is None or len(values) != len(kinds): raise ValueError("invalid_operation_or_field_count") for index, (kind, value) in enumerate(zip(kinds, values)): optional = kind == "key" or (operation == "Click" and index == 2 and values[3] == "0") if value not in DOMAINS[kind] and not (optional and value == NA): raise ValueError("missing_or_invalid_parameter") if operation == "Click": if values[3] == "0" and values[2] != NA: raise ValueError("move_button_must_be_NA") if values[3] == "2" and values[2] != "LEFT": raise ValueError("nonleft_double_click") if operation == "PressKeys": keys = [k for k in values if k != NA] if values != canonical_keys(keys) + [NA] * (4 - len(keys)): raise ValueError("noncanonical_key_order_or_padding") if operation == "Drag" and values[:2] == values[2:]: raise ValueError("zero_distance_drag") return {"operation": operation, "values": list(values)} def convert_target(target): actions = validate_step(target)["actions"] if len(actions) == 2 and [a.get("skill") for a in actions] == ["move_to", "drag_to"]: start, end = [a["args"] for a in actions] if end["button"] != "left": raise ValueError("nonleft_drag") return validate_action( "Drag", [coordinate(v) for v in (start["x"], start["y"], end["x"], end["y"])] ) if len(actions) != 1: raise ValueError("compound_without_intermediate_observation") action = actions[0] if action.get("done"): raise ValueError("unverified_completion") skill, args = action["skill"], action["args"] if skill in {"click", "double_click", "move_to"}: values = [coordinate(args[k]) for k in ("x", "y")] values += [ NA if skill == "move_to" else args["button"].upper(), {"click": "1", "double_click": "2", "move_to": "0"}[skill], ] return validate_action("Click", values) if skill == "type_text": return validate_action("TypeText", ["YES"]) if skill in {"press_key", "hotkey"}: keys = canonical_keys([args["key"]] if skill == "press_key" else args["keys"]) return validate_action("PressKeys", keys + [NA] * (4 - len(keys))) raise ValueError(f"unsupported_or_unverified_source_skill:{skill}") class NativeCodec: """Finite legal encodings assembled in ID space, never retokenized.""" anchor_positions = (7, 8, 9, 21, 23, 24, 36) def __init__(self, tokenizer): self.tokenizer = tokenizer self.vocab_size = tokenizer.get_vocab_size() def one(text): ids = tokenizer.encode(text, add_special_tokens=False).ids if len(ids) != 1: raise ValueError(f"native single token required: {text}") return ids[0] self.na_id, self.mask_id = one(NA), one(MASK) self.encodings, self.inverse = {}, {} for kind, domain in DOMAINS.items(): width = WIDTHS.get(kind, 1) mapping = {NA: (self.na_id,) * width} for value in domain: ids = ( [one(d) for d in value] if kind == "coord" else tokenizer.encode(value, add_special_tokens=False).ids ) if not 1 <= len(ids) <= width or self.na_id in ids or self.mask_id in ids: raise ValueError(f"native encoding exceeds contract: {kind}/{value}") mapping[value] = tuple(ids + [self.na_id] * (width - len(ids))) inverse = {ids: value for value, ids in mapping.items()} if len(inverse) != len(mapping): raise ValueError("ambiguous_native_encoding") self.encodings[kind], self.inverse[kind] = mapping, inverse self.input_ids, self.slot_positions, self.field_slices = [], [], [] self.row_slices = [] def literal(text): self.input_ids.extend(tokenizer.encode(text, add_special_tokens=False).ids) for row_index, (name, kinds) in enumerate(SCHEMA): if row_index: literal("\n") literal(name + "(") row_start = len(self.slot_positions) for field_index, kind in enumerate(kinds): if field_index: literal(",") width = WIDTHS.get(kind, 1) start = len(self.slot_positions) for _ in range(width): self.slot_positions.append(len(self.input_ids)) self.input_ids.append(self.mask_id) self.field_slices.append(slice(start, start + width)) literal(")") self.row_slices.append(slice(row_start, len(self.slot_positions))) if len(self.slot_positions) != 37 or tokenizer.get_vocab_size() != self.vocab_size: raise ValueError("native_contract_mutated") def encode(self, operation, values): validate_action(operation, values) result = [] for name, kinds in SCHEMA: fields = values if name == operation else [NA] * len(kinds) for kind, value in zip(kinds, fields): result.extend(self.encodings[kind][value]) return result def decode(self, ids): if len(ids) != 37 or any(type(i) is not int for i in ids): raise ValueError("expected_37_integer_ids") active, offset = [], 0 for name, kinds in SCHEMA: values = [] for kind in kinds: width = WIDTHS.get(kind, 1) tokens = tuple(ids[offset : offset + width]) if tokens not in self.inverse[kind]: raise ValueError("invalid_field_token_sequence") values.append(self.inverse[kind][tokens]) offset += width if any(v != NA for v in values): active.append(validate_action(name, values)) if len(active) != 1: raise ValueError("expected_exactly_one_active_row") return active[0]