captchakraken 2.3.0 → 2.6.0

package/python/src/captchakraken/prompts.py

@@ -0,0 +1,485 @@
+"""Prompt resolution: every model gets the prompts it was trained on.
+
+Why this exists
+---------------
+The prompt a captcha model is sent has to be the prompt it was trained on. When
+they drift, nothing errors — the model answers in a schema the client does not
+recognise and puzzles silently fail. That has happened: see failure-mode 3 in
+the finetune repo's `scripts/check_prompt_parity.py`, where the shipped prompt
+kept asking for the legacy `output`/`simulate_drag`/PascalCase schema long after
+the LoRA had been retrained on `action: drag`/`drags[]`.
+
+Hardcoding ONE generation of prompts makes that inevitable the moment a new
+model trains on different ones: updating the constants breaks every already-
+published model, and not updating them breaks the new one. There is no version
+of "one set of constants" that is correct.
+
+So this module ships the built-ins for EVERY generation still in service, and
+`models.json` records which generation each model was trained on. A model keeps
+working forever, whatever this client later ships.
+
+Resolution order
+----------------
+1. ``CAPTCHA_PROMPTS_FILE`` — explicit local path. Wins over everything; use it
+   for a model you built yourself and have not published.
+2. ``models.json`` — the registry shipped with this client. FIRST, not last,
+   because it is the only source that works for a private model (the production
+   adapter is gated on the Hub and 401s) or one published before prompts.json
+   existed, and because it is the source the release gate actually verifies.
+3. The served model's repo on the Hub, when it looks like ``org/name`` and
+   ``huggingface_hub`` is importable — covers models published after this
+   client was released.
+4. The built-ins for whatever generation ``latest`` is on, with a warning.
+
+A prompt lookup must never fail a solve, so every step degrades rather than
+raises.
+"""
+from __future__ import annotations
+
+import json
+import os
+import sys
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Any, Dict, Optional
+
+_MODELS_PATH = Path(__file__).with_name("models.json")
+_PROMPTS_FILE_ENV = "CAPTCHA_PROMPTS_FILE"
+_DISABLE_FETCH_ENV = "CAPTCHA_PROMPTS_NO_FETCH"
+
+# The newest generation this client ships built-ins for. NOT "the version to
+# send" — that is per model, resolved below. This exists so the finetune repo
+# can ask "do you speak the generation I am training?" without importing us.
+LATEST_PROMPT_VERSION = "2"
+
+
+# ── built-in templates, one entry per generation in service ─────────────────
+#
+# A PURE LITERAL, deliberately: the finetune repo's check_prompt_parity.py reads
+# it by AST without importing this package (which pulls in pydantic/cv2 that the
+# training venv need not have). Keep it literal — an f-string, a .join(), or a
+# reference to another constant makes it unreadable to that gate, and the gate
+# going quiet is how prompt drift got shipped last time.
+#
+#   1 — CaptchaKrakenV1_Lora, Sunlight-AWQ-4bit, Twilight-FP8, CaptchaKraken_v1.1.
+#       985-char action prompt. No slider clause. No animated ("video") family
+#       and no distorted-text ("text") family: a v1 model cannot answer either
+#       request, which is why both are null rather than a copy of v2's.
+#   2 — adds the PUZZLE PIECE SLIDER clause and the animated + text families.
+#       Byte-identical to src/synthetic/reasoning/instructions.py
+#       ::PIXEL_INSTRUCTION_TEMPLATE, ::VIDEO_INSTRUCTION_TEMPLATE and
+#       ::TEXT_INSTRUCTION in the finetune repo at PROMPT_VERSION 2.
+#
+# Do not edit a published generation's text. Ever. Those models are frozen and
+# so are their prompts; a change here is a change to what an already-shipped
+# model is asked, which is the bug this whole module exists to prevent. New
+# text means a NEW generation.
+BUILTIN_PROMPTS = {
+    "1": {
+        "action_pixel": (
+            "Your task is to solve the captcha. Read the instruction at the top of the image carefully.\n\n"
+            "Look at the puzzle and decide what action solves it. All coordinates you return must be on a "
+            "normalized 0–1000 image scale (top-left = (0, 0), bottom-right = (1000, 1000)).\n\n"
+            "Name WHAT each object is with a short 1–2 word label, then give its position. "
+            "Choose ONE response:\n\n"
+            "FOR CLICK PUZZLES:\n"
+            "  Label each thing you click and give its point — subjects[i] names points[i]:\n"
+            "  → \"action\": \"click\", \"subjects\": [\"<label>\", ...], "
+            "\"points\": [[x1, y1], [x2, y2], ...]\n\n"
+            "FOR DRAG PUZZLES:\n"
+            "  Drag ONE item at a time. Label the source (the piece you pick up) and the destination "
+            "(where it belongs), each with a short 1–2 word label, and give both points:\n"
+            "  → \"action\": \"drag\", \"drags\": [{ \"source\": \"<label>\", \"from\": [x, y], "
+            "\"destination\": \"<label>\", \"to\": [x, y] }, ...]\n\n"
+            "Respond ONLY with JSON:\n"
+            "{\n"
+            "  \"action\": \"click\", \"subjects\": [ ... ], \"points\": [ ... ]\n"
+            "  // OR \"action\": \"drag\", \"drags\": [ ... ]\n"
+            "}"
+        ),
+        "grid": (
+            "Solve the captcha grid by choosing the cell numbers that match the description "
+            "from the captcha image prompt.\n\nGrid: {rows}x{cols} ({total} cells)\n{grid_hint}\n\n"
+            "If no tiles match the description (e.g., they have all been cleared or none were "
+            "present), return an empty list for target_ids: [].\n\n"
+            "Return JSON Array: [list of cell numbers (1-{total})]"
+        ),
+        "video": None,
+        "text": None,
+    },
+    "2": {
+        # UNPUBLISHED as of 2026-08-03: no entry in models.json names generation
+        # 2, `latest` is a generation-1 model, and no training run has produced
+        # a deployable one. That is the ONLY reason its text may still move —
+        # the freeze protects models that exist, and none exists here. The
+        # moment a model registers against "2", this text is frozen with it and
+        # a change means generation 3.
+        "action_pixel": (
+            "Your task is to solve the captcha. Read the instruction at the top of the image carefully.\n\n"
+            "Look at the puzzle and decide what action solves it. All coordinates you return must be on a "
+            "normalized 0–1000 image scale (top-left = (0, 0), bottom-right = (1000, 1000)).\n\n"
+            "Name WHAT each object is with a short 1–2 word label, then give its position. "
+            "Choose ONE response:\n\n"
+            "FOR CLICK PUZZLES:\n"
+            "  Label each thing you click and give its point — subjects[i] names points[i]:\n"
+            "  → \"action\": \"click\", \"subjects\": [\"<label>\", ...], "
+            "\"points\": [[x1, y1], [x2, y2], ...]\n\n"
+            "FOR DRAG PUZZLES:\n"
+            "  Drag ONE item at a time. Label the source (the piece you pick up) and the destination "
+            "(where it belongs), each with a short 1–2 word label, and give both points:\n"
+            "  → \"action\": \"drag\", \"drags\": [{ \"source\": \"<label>\", \"from\": [x, y], "
+            "\"destination\": \"<label>\", \"to\": [x, y] }, ...]\n\n"
+            "FOR PUZZLE PIECE SLIDER PUZZLES:\n"
+            "  A single jigsaw piece has to end up in the piece-shaped slot cut into the picture. "
+            "Do not pick up the piece or the slider handle — leave the source EMPTY and give only "
+            "the destination, the CENTER OF THE SLOT the piece belongs in:\n"
+            "  → \"action\": \"drag\", \"drags\": [{ \"source\": \"\", \"from\": [], "
+            "\"destination\": \"<label>\", \"to\": [x, y] }]\n\n"
+            "Respond ONLY with JSON:\n"
+            "{\n"
+            "  \"action\": \"click\", \"subjects\": [ ... ], \"points\": [ ... ]\n"
+            "  // OR \"action\": \"drag\", \"drags\": [ ... ]\n"
+            "}"
+        ),
+        "grid": (
+            "Solve the captcha grid by choosing the cell numbers that match the description "
+            "from the captcha image prompt.\n\nGrid: {rows}x{cols} ({total} cells)\n{grid_hint}\n\n"
+            "A cell that is already selected — small checkmark badge, border or highlight — "
+            "still counts. Include it if it matches.\n\n"
+            "A cell being REPLACED does not: a large checkmark over the middle of the picture, "
+            "a picture fading to white, or a new picture fading in. That cell is on its way to "
+            "showing something else, so leave it out however well it matches.\n\n"
+            "If no cells match the description, return an empty list for target_ids: [].\n\n"
+            "Return JSON Array: [list of cell numbers (1-{total})]"
+        ),
+        "video": (
+            "Your task is to solve the captcha. This challenge is animated, so instead of one "
+            "picture you are given {n} still keyframes cut from a short recording of it, in "
+            "order: {listing}.\n\n"
+            "Every keyframe shows the SAME puzzle at a different moment. Read the instruction at "
+            "the top of the keyframes carefully. What you need to act on may be visible in only "
+            "some of the frames — sprites fade in and out, boards cycle their contents — so pick "
+            "the ONE frame in which your target is clearest and report its number as \"frame\".\n\n"
+            "The frame number is there so the solver knows WHEN to press the mouse: it waits for the "
+            "widget to look like that frame before clicking. If your answer does not depend on the "
+            "frame — the target is in the same place in every one of them — then there is nothing to "
+            "wait for, and you may leave \"frame\" out entirely.\n\n"
+            "Read your coordinates off THAT frame. All coordinates must be on a normalized 0–1000 "
+            "image scale (top-left = (0, 0), bottom-right = (1000, 1000)).\n\n"
+            "Name WHAT each object is with a short 1–2 word label, then give its position. "
+            "Choose ONE response:\n\n"
+            "FOR CLICK PUZZLES:\n"
+            "  Label each thing you click and give its point — subjects[i] names points[i]:\n"
+            "  → \"frame\": <1-{n}>, \"action\": \"click\", \"subjects\": [\"<label>\", ...], "
+            "\"points\": [[x1, y1], [x2, y2], ...]\n\n"
+            "FOR DRAG PUZZLES:\n"
+            "  Drag ONE item at a time. Label the source (the piece you pick up) and the "
+            "destination (where it belongs), each with a short 1–2 word label, and give both "
+            "points:\n"
+            "  → \"frame\": <1-{n}>, \"action\": \"drag\", \"drags\": [{{ \"source\": \"<label>\", "
+            "\"from\": [x, y], \"destination\": \"<label>\", \"to\": [x, y] }}, ...]\n\n"
+            "Respond ONLY with JSON:\n"
+            "{{\n"
+            "  \"frame\": <1-{n}>,   // omit if your answer holds in every frame\n"
+            "  \"action\": \"click\", \"subjects\": [ ... ], \"points\": [ ... ]\n"
+            "  // OR \"action\": \"drag\", \"drags\": [ ... ]\n"
+            "}}"
+        ),
+        "text": (
+            "Your task is to solve the captcha. The image shows a short code drawn in distorted, "
+            "overlapping or warped characters, sometimes over a busy background.\n\n"
+            "Read the code exactly as printed. Preserve letter case when the characters clearly "
+            "show it, and do not add spaces the image does not show. Decoration — strike-through "
+            "lines, dots, blobs, background texture — is not part of the code.\n\n"
+            "Respond ONLY with JSON:\n"
+            "{\n"
+            "  \"action\": \"type\", \"text\": \"<the code>\"\n"
+            "}"
+        ),
+    },
+}
+
+
+# ── the registry ────────────────────────────────────────────────────────────
+
+def _load_registry() -> Dict[str, Any]:
+    try:
+        with _MODELS_PATH.open(encoding="utf-8") as fh:
+            return json.load(fh)
+    except (OSError, ValueError):  # a broken registry must not fail a solve
+        return {}
+
+
+_REGISTRY: Optional[Dict[str, Any]] = None
+
+
+def registry() -> Dict[str, Any]:
+    global _REGISTRY
+    if _REGISTRY is None:
+        _REGISTRY = _load_registry()
+    return _REGISTRY
+
+
+def registered_models() -> Dict[str, Dict[str, Any]]:
+    """Every model this client knows about, repo id -> entry."""
+    return {k: v for k, v in (registry().get("models") or {}).items()
+            if isinstance(v, dict)}
+
+
+def latest_model() -> Optional[str]:
+    """The model you get when you do not pin one."""
+    return registry().get("latest")
+
+
+def canonical_model_id(model: Optional[str]) -> Optional[str]:
+    """Repo id for `model`, resolving a served alias.
+
+    A hosted endpoint serves a BARE NAME (`captcha`), not a repo id, so the
+    string the client sends as `model` is often not something the registry can
+    be keyed by. Without this indirection every hosted solve would fall through
+    to the default-version branch and lose the whole point of the map.
+    """
+    if not model:
+        return None
+    if model in registered_models():
+        return model
+    alias = (registry().get("served_aliases") or {}).get(model)
+    return alias if isinstance(alias, str) else None
+
+
+# ── prompt sets ─────────────────────────────────────────────────────────────
+
+@dataclass
+class PromptSet:
+    """The prompts to send one particular model, and where they came from."""
+
+    version: str
+    action_prompt: str
+    grid_template: str
+    video_template: Optional[str]
+    text_template: Optional[str]
+    grid_by_type: Dict[str, str]
+    source: str
+
+    def grid_prompt(self, *, rows: int, cols: int, grid_hint: str = "",
+                    puzzle_type: Optional[str] = None) -> str:
+        """A published prompts.json carries grid prompts already rendered per
+        puzzle type; prefer that exact string when the type is known, since it
+        is literally what the model trained on. Otherwise render the template,
+        which is how this client has always built them."""
+        if puzzle_type and puzzle_type in self.grid_by_type:
+            return self.grid_by_type[puzzle_type]
+        return self.grid_template.format(rows=rows, cols=cols,
+                                         total=rows * cols, grid_hint=grid_hint)
+
+    def video_prompt(self, n_keyframes: int) -> str:
+        """The prompt for a challenge served as `n_keyframes` stills.
+
+        The count is in the text because the model has no other way to know how
+        many images arrived — frame identities live in the prompt, not in the
+        image payload.
+        """
+        n = int(n_keyframes)
+        if n < 1:
+            raise ValueError(f"a keyframe request needs at least one frame, got {n_keyframes}")
+        if self.video_template is None:
+            raise ValueError(
+                f"prompt generation {self.version} has no animated-puzzle prompt — "
+                f"model was trained before the video family existed. Use a model on "
+                f"generation 2 or later for animated captchas."
+            )
+        listing = ", ".join(f"frame {i}" for i in range(1, n + 1))
+        return self.video_template.format(n=n, listing=listing)
+
+    def text_prompt(self) -> str:
+        """The prompt for a DISTORTED-TEXT captcha — one where the answer is a
+        string typed into the widget's box, not a point on the picture.
+
+        Asking for this when the generation has no such family is a hard error
+        rather than a silent fall back to the click/drag prompt: that prompt
+        asks for coordinates the puzzle has no use for, so a v1 model handed a
+        BotDetect image answers with a point, the driver clicks a random spot on
+        the letters, and the box stays empty. Failing loudly here names the
+        actual problem — the model predates the family.
+        """
+        if self.text_template is None:
+            raise ValueError(
+                f"prompt generation {self.version} has no distorted-text prompt — "
+                f"model was trained before the text family existed. Use a model on "
+                f"generation 2 or later for BotDetect/MTCaptcha/Yandex text captchas."
+            )
+        return self.text_template
+
+
+def builtin(version: str) -> Optional[PromptSet]:
+    """The built-in PromptSet for a generation, or None if we don't ship it."""
+    spec = BUILTIN_PROMPTS.get(str(version))
+    if spec is None:
+        return None
+    return PromptSet(
+        version=str(version),
+        action_prompt=spec["action_pixel"],
+        grid_template=spec["grid"],
+        video_template=spec["video"],
+        text_template=spec["text"],
+        grid_by_type={},
+        source=f"built-in v{version}",
+    )
+
+
+def _from_doc(doc: Dict[str, Any], source: str) -> PromptSet:
+    """A published prompts.json. Anything it omits falls back to the built-ins
+    for the version it declares — a partial file should narrow what we take from
+    it, not silently drop a whole prompt family."""
+    version = str(doc.get("prompt_version", LATEST_PROMPT_VERSION))
+    base = builtin(version) or builtin(LATEST_PROMPT_VERSION)
+    templates = doc.get("templates") or {}
+    grid_by_type = {k.split("grid::", 1)[1]: v for k, v in templates.items()
+                    if k.startswith("grid::") and isinstance(v, str)}
+    return PromptSet(
+        version=version,
+        action_prompt=templates.get("action_pixel") or base.action_prompt,
+        grid_template=templates.get("grid") or base.grid_template,
+        video_template=templates.get("video") or base.video_template,
+        text_template=templates.get("text") or base.text_template,
+        grid_by_type=grid_by_type,
+        source=source,
+    )
+
+
+_cache: Dict[str, PromptSet] = {}
+
+
+def _warn(msg: str) -> None:
+    print(f"[captchakraken] {msg}", file=sys.stderr)
+
+
+def resolve(model: Optional[str]) -> PromptSet:
+    """The PromptSet for `model`. Never raises — see the module docstring for
+    the resolution order and why the registry comes before the Hub."""
+    key = model or ""
+    if key in _cache:
+        return _cache[key]
+
+    ps: Optional[PromptSet] = None
+
+    explicit = os.getenv(_PROMPTS_FILE_ENV)
+    if explicit:
+        try:
+            with open(explicit, encoding="utf-8") as fh:
+                ps = _from_doc(json.load(fh), f"file:{explicit}")
+        except Exception as exc:  # noqa: BLE001 — never fail a solve over this
+            _warn(f"could not read {_PROMPTS_FILE_ENV}={explicit}: {exc}; "
+                  "falling back to the registry")
+
+    repo_id = canonical_model_id(model)
+    if ps is None and repo_id:
+        entry = registered_models().get(repo_id) or {}
+        version = str(entry.get("prompt_version", "")) or None
+        if version:
+            ps = builtin(version)
+            if ps is None:
+                # Registered against a generation we ship no text for. The
+                # release gate exists to make this unreachable; if it happens
+                # anyway, say so loudly rather than guessing.
+                _warn(f"{repo_id} is registered as prompt_version {version}, which this "
+                      f"client ships no prompts for — using v{LATEST_PROMPT_VERSION}. "
+                      "Upgrade captchakraken.")
+            else:
+                ps = PromptSet(**{**ps.__dict__, "source": f"registry:{repo_id}"})
+
+    if ps is None and model and "/" in model and os.getenv(_DISABLE_FETCH_ENV, "0") != "1":
+        try:
+            from huggingface_hub import hf_hub_download  # lazy: optional dependency
+
+            path = hf_hub_download(repo_id=model, filename="prompts.json")
+            with open(path, encoding="utf-8") as fh:
+                ps = _from_doc(json.load(fh), f"hub:{model}")
+        except Exception:
+            # No prompts.json, no hub, offline, or gated. All expected — the
+            # registry above is the supported path; this is a bonus for models
+            # published after this client shipped.
+            ps = None
+
+    if ps is None:
+        default_version = str(
+            (registered_models().get(latest_model() or "") or {})
+            .get("prompt_version", LATEST_PROMPT_VERSION))
+        ps = builtin(default_version) or builtin(LATEST_PROMPT_VERSION)
+        if model:
+            _warn(f"model {model!r} is not in models.json and published no prompts.json; "
+                  f"using generation {ps.version}. If it was trained on different prompts, "
+                  "its answers will be quietly worse — register it.")
+
+    _cache[key] = ps
+    return ps
+
+
+# ─── training pixel budget ──────────────────────────────────────────────────
+# Same registry, same reason, different axis. See models.json's header comment:
+# an adapter learns to read a puzzle at whatever MIN_PIXELS/MAX_PIXELS band its
+# training exported, and serving it under a different band is the prompt bug
+# wearing a different hat — silent, and wrong on every puzzle.
+
+MIN_PIXELS_ENV = "CAPTCHA_MIN_PIXELS"
+MAX_PIXELS_ENV = "CAPTCHA_MAX_PIXELS"
+
+
+@dataclass(frozen=True)
+class PixelBudget:
+    """The image-area band a model was trained under, in pixels.
+
+    `minimum` is a floor: smaller images are upscaled to it. `maximum` is a
+    ceiling, or None for "send it as-is and let the server decide" — which is
+    what every model published before this field existed did.
+    """
+    minimum: int
+    maximum: Optional[int]
+    source: str
+
+
+#: What the client did before models.json carried a budget. 448² is Qwen's own
+#: ViT floor; there was never a ceiling. Anything unregistered keeps this, so
+#: adding the field cannot change an existing deployment.
+DEFAULT_PIXEL_BUDGET = PixelBudget(minimum=448 * 448, maximum=None,
+                                   source="client-default")
+
+
+def _env_int(name: str) -> Optional[int]:
+    try:
+        value = int(os.environ.get(name) or 0)
+    except ValueError:
+        return None
+    return value if value > 0 else None
+
+
+def pixel_budget(model: Optional[str]) -> PixelBudget:
+    """The budget to send `model` images at. Never raises.
+
+    Order mirrors `resolve`: an explicit env pin wins, then the registry, then
+    the historical default. There is deliberately no Hub fetch — a wrong budget
+    degrades quality silently, so it may only come from a source we can verify
+    offline.
+    """
+    env_min, env_max = _env_int(MIN_PIXELS_ENV), _env_int(MAX_PIXELS_ENV)
+    if env_min or env_max:
+        return PixelBudget(
+            minimum=env_min or DEFAULT_PIXEL_BUDGET.minimum,
+            maximum=env_max if env_max else DEFAULT_PIXEL_BUDGET.maximum,
+            source="env")
+
+    repo_id = canonical_model_id(model)
+    entry = (registered_models().get(repo_id) or {}) if repo_id else {}
+    budget = entry.get("pixel_budget") or {}
+    if budget:
+        return PixelBudget(
+            minimum=int(budget.get("min") or DEFAULT_PIXEL_BUDGET.minimum),
+            maximum=int(budget["max"]) if budget.get("max") else None,
+            source=f"registry:{repo_id}")
+    return DEFAULT_PIXEL_BUDGET
+
+
+def clear_cache() -> None:
+    _cache.clear()
+    global _REGISTRY
+    _REGISTRY = None