@lemoncat7/dsh-partner 2.5.0 → 2.7.1

package/resources/skills/drama-production/upstream/skills/short-drama-produce/scripts/production_tool.py

@@ -0,0 +1,2205 @@
+#!/usr/bin/env python3
+"""Prepare and execute explicitly confirmed short-drama media jobs.
+
+A configured adapter receives one bounded JSON job on stdin and returns local
+output files on stdout. The adapter is launched without a shell and only after
+a confirmation bound to the exact job and current project inputs. Optional
+provider adapters can ship with this skill, but credentials and adapter config
+remain outside creator projects.
+"""
+
+from __future__ import annotations
+
+import argparse
+import contextlib
+import hashlib
+import json
+import os
+import re
+import stat
+import subprocess
+import sys
+import tempfile
+import uuid
+from collections.abc import Iterator, Mapping
+from datetime import datetime, timezone
+from pathlib import Path, PurePosixPath
+from typing import Any, BinaryIO
+
+MINIMUM_PYTHON = (3, 9)
+if sys.version_info < MINIMUM_PYTHON:
+    raise SystemExit(
+        "short-drama-produce needs Python {}.{} or newer".format(*MINIMUM_PYTHON)
+    )
+
+PROJECT_FILE = "short-drama.json"
+PRODUCTION_ROOT = Path(".short-drama/production")
+JOB_SCHEMA = "1.0"
+JOB_ID_RE = re.compile(r"[A-Za-z0-9][A-Za-z0-9._-]{0,79}")
+MAX_JOB_BYTES = 256 * 1024
+MAX_RUN_RECORD_BYTES = 256 * 1024
+MAX_ADAPTER_RESPONSE_BYTES = 1024 * 1024
+MAX_OUTPUT_BYTES = 512 * 1024 * 1024
+MAX_INPUT_BYTES = 50 * 1024 * 1024
+MAX_TOTAL_INPUT_BYTES = 200 * 1024 * 1024
+MAX_TIMEOUT_SECONDS = 3600
+PUBLIC_ERROR_CATEGORIES = {
+    "authentication",
+    "configuration",
+    "contract",
+    "invalid_request",
+    "network",
+    "permission",
+    "provider_response",
+    "rate_limit",
+    "server",
+    "timeout",
+}
+PUBLIC_ERROR_TOKEN_RE = re.compile(r"[A-Za-z0-9][A-Za-z0-9._:-]{0,199}")
+ALLOWED_JOB_KEYS = {
+    "schema_version",
+    "job_id",
+    "modality",
+    "adapter",
+    "prompt",
+    "source",
+    "source_entry",
+    "references",
+    "reference_bindings",
+    "outputs",
+    "parameters",
+    "overwrite",
+}
+STORED_EXECUTION_KEYS = ALLOWED_JOB_KEYS | {"inputs"}
+STORED_JOB_KEYS = STORED_EXECUTION_KEYS | {"fingerprint", "prepared_at"}
+LEGACY_STORED_EXECUTION_KEYS = STORED_EXECUTION_KEYS - {
+    "source_entry",
+    "reference_bindings",
+}
+LEGACY_STORED_JOB_KEYS = LEGACY_STORED_EXECUTION_KEYS | {
+    "fingerprint",
+    "prepared_at",
+}
+SECRET_KEYS = {
+    "authorization",
+    "credential",
+    "credentials",
+    "password",
+    "secret",
+    "token",
+    "access_token",
+    "api_key",
+    "apikey",
+}
+MEDIA_EXTENSIONS = {
+    "image": {".png", ".jpg", ".jpeg", ".webp"},
+    "video": {".mp4", ".mov", ".webm"},
+    "tts": {".wav", ".mp3", ".m4a", ".aac", ".flac", ".opus"},
+    "music": {".wav", ".mp3", ".m4a", ".aac", ".flac", ".opus"},
+}
+MEDIA_TYPES = {
+    ".png": "image/png",
+    ".jpg": "image/jpeg",
+    ".jpeg": "image/jpeg",
+    ".webp": "image/webp",
+    ".mp4": "video/mp4",
+    ".mov": "video/quicktime",
+    ".webm": "video/webm",
+    ".wav": "audio/wav",
+    ".mp3": "audio/mpeg",
+    ".m4a": "audio/mp4",
+    ".aac": "audio/aac",
+    ".flac": "audio/flac",
+    ".opus": "audio/ogg",
+}
+CREATOR_SOURCE_NAMES = {
+    "图片提示词.md": "image",
+    "视频提示词.md": "video",
+    # A shot's frozen keyframe is a producible image and lives nowhere else, so
+    # without this the start frame a video job wants to bind has no stage that
+    # can render it.
+    "分镜.md": "image",
+}
+# Which H2 IDs a creator source may select, and where that entry keeps its
+# copyable body.
+CREATOR_SOURCE_ENTRIES = {
+    "图片提示词.md": ("IMG-", r"可复制(?:通用)?提示词", "参考"),
+    "视频提示词.md": ("MOTION-", r"可复制(?:通用)?提示词", "输入参考图"),
+    "分镜.md": ("SHOT-", r"冻结关键帧提示词", "输入参考图"),
+}
+
+REF_SLOT_RE = re.compile(r"REF-[A-Z0-9][A-Z0-9-]{0,79}")
+# A slot head, not a bare substring: a real reference file may legitimately be
+# named `plan-sheet.png`, and that is not a creator-supplied declaration.
+PLAN_SLOT_HEAD_RE = re.compile(r"PLAN-[A-Za-z0-9][A-Za-z0-9-]{0,79}（顺序：")
+PLAN_SLOT_RE = re.compile(
+    r"PLAN-[A-Za-z0-9][A-Za-z0-9-]{0,79}（顺序：[1-9]\d*）· "
+    r"(?:IMG|SHOT)-[A-Za-z0-9-]+《[^》\n]+》（[^）\n]*）"
+)
+SOURCE_ENTRY_RE = re.compile(r"[A-Z][A-Z0-9-]{1,99}")
+REFERENCE_SUFFIX_RE = r"(?:png|jpe?g|webp)"
+# 用途 is the creator-facing question "what does this picture decide here"; the
+# job's own `role` is its production translation. The segment is optional so a
+# declaration written before that field existed still prepares, and so the
+# creator-side diagnostic stays with creator_markdown_check.py.
+REFERENCE_LINE_RE = re.compile(
+    rf"(REF-[A-Z0-9][A-Z0-9-]{{0,79}})（顺序：([1-9]\d*)）· "
+    rf"([^；\n]+?\.{REFERENCE_SUFFIX_RE})《([^》\n]+)》"
+    r"（(?:用途：[^；）\n]+；)?控制：([^；）\n]+)；不得控制：([^）\n]+)）",
+    re.IGNORECASE,
+)
+
+
+class ConfirmationRequiredError(RuntimeError):
+    """The exact current job has not been explicitly confirmed."""
+
+
+class AdapterError(RuntimeError):
+    """A configured media adapter failed or broke its output contract."""
+
+    def __init__(
+        self, message: str, *, public_error: Mapping[str, Any] | None = None
+    ) -> None:
+        super().__init__(message)
+        self.public_error = dict(public_error) if public_error is not None else None
+
+
+def utc_now() -> str:
+    return datetime.now(timezone.utc).isoformat().replace("+00:00", "Z")
+
+
+def sha256_bytes(content: bytes) -> str:
+    return hashlib.sha256(content).hexdigest()
+
+
+def find_project(start: Path) -> Path:
+    candidate = start.expanduser().resolve()
+    if candidate.is_file():
+        candidate = candidate.parent
+    for directory in (candidate, *candidate.parents):
+        if (directory / PROJECT_FILE).is_file():
+            return directory
+    raise FileNotFoundError(f"no {PROJECT_FILE} found from {start}")
+
+
+def _relative_path(value: object, *, output: bool = False) -> str:
+    if not isinstance(value, str):
+        raise ValueError("project paths must be strings")
+    raw = value.replace("\\", "/")
+    pure = PurePosixPath(raw)
+    if not raw or pure.is_absolute() or any(part in {"", ".", ".."} for part in pure.parts):
+        raise ValueError(f"unsafe project-relative path: {value}")
+    if pure.parts[0].casefold() == ".short-drama" or pure.name.casefold() == PROJECT_FILE:
+        raise ValueError(f"operational project path is not allowed: {value}")
+    if output:
+        parts = pure.parts
+        top_level_production = len(parts) >= 2 and parts[0].casefold() == "production"
+        episode_production = (
+            len(parts) >= 4
+            and parts[0] in {"剧集", "episodes"}
+            and re.fullmatch(r"EP\d{3,}", parts[1], re.IGNORECASE) is not None
+            and parts[2] in {"制作成果", "production"}
+        )
+        if not top_level_production and not episode_production:
+            raise ValueError(
+                "media outputs must use top-level production/ or "
+                "剧集|episodes/<EP>/制作成果|production/"
+            )
+    return pure.as_posix()
+
+
+def _project_file(root: Path, relative: str, *, create_parent: bool = False) -> Path:
+    target = root / relative
+    current = root
+    for part in PurePosixPath(relative).parts[:-1]:
+        current /= part
+        if current.exists() and (current.is_symlink() or not current.is_dir()):
+            raise ValueError(f"unsafe project directory: {part}")
+        if create_parent and not current.exists():
+            current.mkdir()
+    if target.exists() and (target.is_symlink() or not target.is_file()):
+        raise ValueError(f"unsafe project file: {relative}")
+    if not target.parent.resolve().is_relative_to(root):
+        raise ValueError(f"path escapes project root: {relative}")
+    return target
+
+
+def _is_link_or_reparse(details: os.stat_result) -> bool:
+    attributes = getattr(details, "st_file_attributes", 0)
+    reparse_flag = getattr(stat, "FILE_ATTRIBUTE_REPARSE_POINT", 0)
+    return stat.S_ISLNK(details.st_mode) or bool(attributes & reparse_flag)
+
+
+@contextlib.contextmanager
+def _open_project_input(root: Path, relative: str) -> Iterator[BinaryIO]:
+    """Open one regular project input without following path components on POSIX."""
+    parts = PurePosixPath(relative).parts
+    if not parts:
+        raise ValueError("job input path is empty")
+    if os.name != "nt" and os.open in os.supports_dir_fd:
+        directory_flags = os.O_RDONLY | getattr(os, "O_DIRECTORY", 0)
+        nofollow = getattr(os, "O_NOFOLLOW", 0)
+        directory_fd = os.open(root, directory_flags)
+        file_fd: int | None = None
+        try:
+            for part in parts[:-1]:
+                next_fd = os.open(
+                    part,
+                    directory_flags | nofollow,
+                    dir_fd=directory_fd,
+                )
+                os.close(directory_fd)
+                directory_fd = next_fd
+            file_fd = os.open(
+                parts[-1], os.O_RDONLY | nofollow, dir_fd=directory_fd
+            )
+            details = os.fstat(file_fd)
+            if not stat.S_ISREG(details.st_mode):
+                raise ValueError(f"job input is not a regular file: {relative}")
+            with os.fdopen(file_fd, "rb", closefd=True) as handle:
+                file_fd = None
+                yield handle
+        except FileNotFoundError as exc:
+            raise FileNotFoundError(f"job input is missing: {relative}") from exc
+        except OSError as exc:
+            raise ValueError(f"unsafe job input path: {relative}") from exc
+        finally:
+            if file_fd is not None:
+                os.close(file_fd)
+            os.close(directory_fd)
+        return
+
+    # Windows lacks portable openat/O_NOFOLLOW support. Reject reparse/symlink
+    # components, pin the final file handle, and verify its identity before use.
+    path = root
+    for part in parts:
+        path /= part
+        try:
+            details = path.lstat()
+        except FileNotFoundError as exc:
+            raise FileNotFoundError(f"job input is missing: {relative}") from exc
+        if _is_link_or_reparse(details):
+            raise ValueError(f"unsafe job input path: {relative}")
+    if not path.resolve().is_relative_to(root):
+        raise ValueError(f"job input escapes project root: {relative}")
+    before = path.stat(follow_symlinks=False)
+    if not stat.S_ISREG(before.st_mode):
+        raise ValueError(f"job input is not a regular file: {relative}")
+    with path.open("rb") as handle:
+        opened = os.fstat(handle.fileno())
+        if (opened.st_dev, opened.st_ino) != (before.st_dev, before.st_ino):
+            raise ValueError(f"job input changed while opening: {relative}")
+        yield handle
+
+
+def _hash_project_file(root: Path, relative: str) -> str:
+    digest = hashlib.sha256()
+    size = 0
+    with _open_project_input(root, relative) as handle:
+        for chunk in iter(lambda: handle.read(1024 * 1024), b""):
+            size += len(chunk)
+            if size > MAX_INPUT_BYTES:
+                raise ValueError(f"job input exceeds the size limit: {relative}")
+            digest.update(chunk)
+    return digest.hexdigest()
+
+
+def _snapshot_inputs(
+    root: Path, job: Mapping[str, Any], snapshot_root: Path
+) -> None:
+    inputs = job.get("inputs")
+    if not isinstance(inputs, Mapping):
+        raise ConfirmationRequiredError("stored job inputs are invalid")
+    total = 0
+    for relative_value, expected_value in inputs.items():
+        relative = _relative_path(relative_value)
+        if not isinstance(expected_value, str) or re.fullmatch(
+            r"[0-9a-f]{64}", expected_value
+        ) is None:
+            raise ConfirmationRequiredError("stored job input hash is invalid")
+        target = snapshot_root.joinpath(*PurePosixPath(relative).parts)
+        target.parent.mkdir(parents=True, exist_ok=True)
+        digest = hashlib.sha256()
+        size = 0
+        try:
+            with _open_project_input(root, relative) as incoming, target.open(
+                "xb"
+            ) as outgoing:
+                for chunk in iter(lambda: incoming.read(1024 * 1024), b""):
+                    size += len(chunk)
+                    total += len(chunk)
+                    if size > MAX_INPUT_BYTES or total > MAX_TOTAL_INPUT_BYTES:
+                        raise ConfirmationRequiredError(
+                            "job inputs exceed the production size limit"
+                        )
+                    digest.update(chunk)
+                    outgoing.write(chunk)
+        except (FileNotFoundError, OSError, ValueError) as exc:
+            raise ConfirmationRequiredError(
+                "job inputs changed; prepare and confirm again"
+            ) from exc
+        if digest.hexdigest() != expected_value:
+            raise ConfirmationRequiredError(
+                "job inputs changed; prepare and confirm again"
+            )
+
+
+def _canonical(document: object) -> bytes:
+    return json.dumps(
+        document, ensure_ascii=False, sort_keys=True, separators=(",", ":")
+    ).encode("utf-8")
+
+
+def _check_metadata_parts(parts: tuple[str, ...]) -> None:
+    if any(
+        not part
+        or part in {".", ".."}
+        or "/" in part
+        or "\\" in part
+        for part in parts
+    ):
+        raise ValueError("production metadata path is invalid")
+
+
+@contextlib.contextmanager
+def _metadata_directory(
+    root: Path, parts: tuple[str, ...], *, create: bool
+) -> Iterator[tuple[Path, int | None]]:
+    """Pin one metadata directory without following project-controlled parents."""
+    root = root.resolve()
+    _check_metadata_parts(parts)
+    all_parts = (".short-drama", "production", *parts)
+    directory = root.joinpath(*all_parts)
+    if os.name != "nt" and os.open in os.supports_dir_fd:
+        flags = os.O_RDONLY | getattr(os, "O_DIRECTORY", 0)
+        nofollow = getattr(os, "O_NOFOLLOW", 0)
+        directory_fd = os.open(root, flags)
+        try:
+            try:
+                for part in all_parts:
+                    try:
+                        next_fd = os.open(
+                            part, flags | nofollow, dir_fd=directory_fd
+                        )
+                    except FileNotFoundError:
+                        if not create:
+                            raise
+                        try:
+                            os.mkdir(part, mode=0o700, dir_fd=directory_fd)
+                        except FileExistsError:
+                            pass
+                        next_fd = os.open(
+                            part, flags | nofollow, dir_fd=directory_fd
+                        )
+                    details = os.fstat(next_fd)
+                    if not stat.S_ISDIR(details.st_mode):
+                        os.close(next_fd)
+                        raise ValueError("production metadata directory is unsafe")
+                    os.close(directory_fd)
+                    directory_fd = next_fd
+            except FileNotFoundError:
+                raise
+            except OSError as exc:
+                raise ValueError("production metadata directory is unsafe") from exc
+            yield directory, directory_fd
+        finally:
+            os.close(directory_fd)
+        return
+
+    current = root
+    for part in all_parts:
+        current /= part
+        try:
+            details = current.lstat()
+        except FileNotFoundError:
+            if not create:
+                raise
+            try:
+                current.mkdir(mode=0o700)
+            except FileExistsError:
+                pass
+            details = current.lstat()
+        if _is_link_or_reparse(details) or not stat.S_ISDIR(details.st_mode):
+            raise ValueError("production metadata directory is unsafe")
+    if not current.resolve().is_relative_to(root):
+        raise ValueError("production metadata directory escapes the project")
+    yield current, None
+
+
+def _metadata_atomic_json(
+    root: Path, directory_parts: tuple[str, ...], name: str, document: Mapping[str, Any]
+) -> None:
+    _check_metadata_parts((name,))
+    content = _canonical(document) + b"\n"
+    temporary_name = f".{name}.{uuid.uuid4().hex}.tmp"
+    with _metadata_directory(root, directory_parts, create=True) as (
+        directory,
+        directory_fd,
+    ):
+        if directory_fd is not None:
+            descriptor = -1
+            try:
+                descriptor = os.open(
+                    temporary_name,
+                    os.O_WRONLY
+                    | os.O_CREAT
+                    | os.O_EXCL
+                    | getattr(os, "O_NOFOLLOW", 0),
+                    0o600,
+                    dir_fd=directory_fd,
+                )
+                with os.fdopen(descriptor, "wb", closefd=True) as handle:
+                    descriptor = -1
+                    handle.write(content)
+                    handle.flush()
+                    os.fsync(handle.fileno())
+                os.replace(
+                    temporary_name,
+                    name,
+                    src_dir_fd=directory_fd,
+                    dst_dir_fd=directory_fd,
+                )
+                os.fsync(directory_fd)
+            finally:
+                if descriptor >= 0:
+                    os.close(descriptor)
+                try:
+                    os.unlink(temporary_name, dir_fd=directory_fd)
+                except FileNotFoundError:
+                    pass
+            return
+
+        before = directory.stat(follow_symlinks=False)
+        temporary = directory / temporary_name
+        try:
+            with temporary.open("xb") as handle:
+                opened = os.fstat(handle.fileno())
+                if not stat.S_ISREG(opened.st_mode):
+                    raise ValueError("production metadata file is unsafe")
+                handle.write(content)
+                handle.flush()
+                os.fsync(handle.fileno())
+            after = directory.lstat()
+            if (
+                _is_link_or_reparse(after)
+                or not stat.S_ISDIR(after.st_mode)
+                or (after.st_dev, after.st_ino) != (before.st_dev, before.st_ino)
+            ):
+                raise ValueError("production metadata directory changed")
+            os.replace(temporary, directory / name)
+        finally:
+            try:
+                temporary.unlink()
+            except FileNotFoundError:
+                pass
+
+
+def _metadata_read_json(
+    root: Path, directory_parts: tuple[str, ...], name: str, *, maximum: int
+) -> object:
+    _check_metadata_parts((name,))
+    with _metadata_directory(root, directory_parts, create=False) as (
+        directory,
+        directory_fd,
+    ):
+        if directory_fd is not None:
+            descriptor = os.open(
+                name,
+                os.O_RDONLY | getattr(os, "O_NOFOLLOW", 0),
+                dir_fd=directory_fd,
+            )
+            with os.fdopen(descriptor, "rb", closefd=True) as handle:
+                details = os.fstat(handle.fileno())
+                if not stat.S_ISREG(details.st_mode) or details.st_size > maximum:
+                    raise ValueError("production metadata file is unsafe")
+                raw = handle.read(maximum + 1)
+        else:
+            path = directory / name
+            before = path.lstat()
+            if (
+                _is_link_or_reparse(before)
+                or not stat.S_ISREG(before.st_mode)
+                or before.st_size > maximum
+            ):
+                raise ValueError("production metadata file is unsafe")
+            with path.open("rb") as handle:
+                opened = os.fstat(handle.fileno())
+                if (opened.st_dev, opened.st_ino) != (before.st_dev, before.st_ino):
+                    raise ValueError("production metadata file changed while opening")
+                raw = handle.read(maximum + 1)
+        if len(raw) > maximum:
+            raise ValueError("production metadata file is too large")
+        return json.loads(raw.decode("utf-8"))
+
+
+def _metadata_unlink(
+    root: Path, directory_parts: tuple[str, ...], name: str
+) -> None:
+    _check_metadata_parts((name,))
+    with _metadata_directory(root, directory_parts, create=False) as (
+        directory,
+        directory_fd,
+    ):
+        if directory_fd is not None:
+            os.unlink(name, dir_fd=directory_fd)
+        else:
+            path = directory / name
+            details = path.lstat()
+            if _is_link_or_reparse(details) or not stat.S_ISREG(details.st_mode):
+                raise ValueError("production metadata file is unsafe")
+            path.unlink()
+
+
+def _metadata_json_names(root: Path, directory_parts: tuple[str, ...]) -> list[str]:
+    try:
+        with _metadata_directory(root, directory_parts, create=False) as (
+            directory,
+            directory_fd,
+        ):
+            names = os.listdir(directory_fd if directory_fd is not None else directory)
+    except FileNotFoundError:
+        return []
+    return sorted(
+        name
+        for name in names
+        if isinstance(name, str)
+        and name.endswith(".json")
+        and "/" not in name
+        and "\\" not in name
+    )
+
+
+@contextlib.contextmanager
+def _project_lock(root: Path) -> Iterator[None]:
+    with _metadata_directory(root, (), create=True) as (directory, directory_fd):
+        if directory_fd is not None:
+            descriptor = os.open(
+                "lock",
+                os.O_RDWR | os.O_CREAT | getattr(os, "O_NOFOLLOW", 0),
+                0o600,
+                dir_fd=directory_fd,
+            )
+            details = os.fstat(descriptor)
+            if not stat.S_ISREG(details.st_mode):
+                os.close(descriptor)
+                raise ValueError("production lock is unsafe")
+            handle_context = os.fdopen(descriptor, "a+b", closefd=True)
+        else:
+            lock_path = directory / "lock"
+            before: os.stat_result | None
+            try:
+                before = lock_path.lstat()
+            except FileNotFoundError:
+                before = None
+            else:
+                if _is_link_or_reparse(before) or not stat.S_ISREG(before.st_mode):
+                    raise ValueError("production lock is unsafe")
+            handle_context = lock_path.open("a+b")
+            opened = os.fstat(handle_context.fileno())
+            after = lock_path.lstat()
+            if (
+                not stat.S_ISREG(opened.st_mode)
+                or _is_link_or_reparse(after)
+                or not stat.S_ISREG(after.st_mode)
+                or (opened.st_dev, opened.st_ino) != (after.st_dev, after.st_ino)
+                or (
+                    before is not None
+                    and (before.st_dev, before.st_ino)
+                    != (opened.st_dev, opened.st_ino)
+                )
+            ):
+                handle_context.close()
+                raise ValueError("production lock changed while opening")
+        with handle_context as handle:
+            if os.name == "nt":
+                import msvcrt
+
+                handle.seek(0, os.SEEK_END)
+                if handle.tell() == 0:
+                    handle.write(b"0")
+                    handle.flush()
+                handle.seek(0)
+                locking = getattr(msvcrt, "locking")
+                lock = getattr(msvcrt, "LK_LOCK")
+                unlock = getattr(msvcrt, "LK_UNLCK")
+                locking(handle.fileno(), lock, 1)
+                try:
+                    yield
+                finally:
+                    handle.seek(0)
+                    locking(handle.fileno(), unlock, 1)
+            else:
+                import fcntl
+
+                fcntl.flock(handle.fileno(), fcntl.LOCK_EX)
+                try:
+                    yield
+                finally:
+                    fcntl.flock(handle.fileno(), fcntl.LOCK_UN)
+
+
+def _job_key(job_id: str) -> str:
+    return sha256_bytes(job_id.encode("utf-8"))[:24]
+
+
+def _job_path(root: Path, job_id: str) -> Path:
+    return root / PRODUCTION_ROOT / "jobs" / f"{_job_key(job_id)}.json"
+
+
+def _confirmation_path(root: Path, job_id: str) -> Path:
+    return root / PRODUCTION_ROOT / "confirmations" / f"{_job_key(job_id)}.json"
+
+
+def _run_directory(root: Path, job_id: str) -> Path:
+    return root / PRODUCTION_ROOT / "runs" / _job_key(job_id)
+
+
+def _contains_secret_key(value: object) -> bool:
+    if isinstance(value, Mapping):
+        return any(
+            str(key).casefold() in SECRET_KEYS or _contains_secret_key(child)
+            for key, child in value.items()
+        )
+    if isinstance(value, list):
+        return any(_contains_secret_key(child) for child in value)
+    return False
+
+
+def _string_list(value: object, *, label: str, limit: int = 32) -> list[str]:
+    if not isinstance(value, list) or not all(isinstance(item, str) for item in value):
+        raise ValueError(f"{label} must be a string list")
+    if len(value) > limit:
+        raise ValueError(f"{label} has too many entries")
+    return list(value)
+
+
+def _scope_list(value: object, *, label: str) -> list[str]:
+    items = _string_list(value, label=label, limit=32)
+    if not items or any(not item.strip() or len(item) > 200 for item in items):
+        raise ValueError(f"{label} must contain non-empty bounded text")
+    normalized = [item.strip() for item in items]
+    folded = [item.casefold() for item in normalized]
+    if len(folded) != len(set(folded)):
+        raise ValueError(f"{label} must not contain duplicates")
+    return normalized
+
+
+def _canonical_creator_source_modality(source: str | None) -> str | None:
+    if source is None:
+        return None
+    path = PurePosixPath(source)
+    if (
+        len(path.parts) != 3
+        or path.parts[0] not in {"剧集", "episodes"}
+        or not path.parts[1]
+    ):
+        return None
+    return CREATOR_SOURCE_NAMES.get(path.name)
+
+
+def _normalize_reference_bindings(value: object) -> list[dict[str, Any]]:
+    if value is None:
+        return []
+    if not isinstance(value, list) or len(value) > 16:
+        raise ValueError("reference_bindings must be a list of at most sixteen entries")
+    normalized: list[dict[str, Any]] = []
+    for index, binding in enumerate(value, 1):
+        label = f"reference_bindings[{index}]"
+        if not isinstance(binding, Mapping) or set(binding) != {
+            "slot_id",
+            "order",
+            "path",
+            "label",
+            "role",
+            "may_control",
+            "must_not_control",
+        }:
+            raise ValueError(f"{label} fields are invalid")
+        slot_id = binding.get("slot_id")
+        order = binding.get("order")
+        raw_label = binding.get("label")
+        role = binding.get("role")
+        if not isinstance(slot_id, str) or REF_SLOT_RE.fullmatch(slot_id) is None:
+            raise ValueError(f"{label} slot_id is invalid")
+        if not isinstance(order, int) or isinstance(order, bool) or order < 1:
+            raise ValueError(f"{label} order is invalid")
+        if (
+            not isinstance(raw_label, str)
+            or not raw_label.strip()
+            or len(raw_label) > 200
+            or re.search(r"[\u3400-\u9fff]", raw_label) is None
+        ):
+            raise ValueError(f"{label} label must contain Chinese text")
+        if not isinstance(role, str) or not role.strip() or len(role) > 80:
+            raise ValueError(f"{label} role is invalid")
+        may_control = _scope_list(
+            binding.get("may_control"), label=f"{label}.may_control"
+        )
+        must_not_control = _scope_list(
+            binding.get("must_not_control"), label=f"{label}.must_not_control"
+        )
+        if {item.casefold() for item in may_control} & {
+            item.casefold() for item in must_not_control
+        }:
+            raise ValueError(f"{label} control scopes must not overlap")
+        normalized.append(
+            {
+                "slot_id": slot_id,
+                "order": order,
+                "path": _relative_path(binding.get("path")),
+                "label": raw_label.strip(),
+                "role": role.strip(),
+                "may_control": may_control,
+                "must_not_control": must_not_control,
+            }
+        )
+    normalized.sort(key=lambda item: int(item["order"]))
+    slots = [str(item["slot_id"]) for item in normalized]
+    orders = [int(item["order"]) for item in normalized]
+    paths = [str(item["path"]) for item in normalized]
+    if len(slots) != len(set(slots)):
+        raise ValueError("reference binding slot_ids must be unique")
+    if orders != list(range(1, len(orders) + 1)):
+        raise ValueError("reference binding order must be unique and contiguous from 1")
+    if len(paths) != len(set(paths)):
+        raise ValueError("reference binding paths must be unique")
+    return normalized
+
+
+def _markdown_section(document: str, source_entry: str) -> str:
+    heading = re.compile(
+        rf"^##\s+`?{re.escape(source_entry)}`?(?:\s+·.*)?\s*$", re.MULTILINE
+    )
+    matches = list(heading.finditer(document))
+    if not matches:
+        raise ValueError(f"source entry is missing from Markdown: {source_entry}")
+    if len(matches) != 1:
+        raise ValueError(f"source entry is duplicated in Markdown: {source_entry}")
+    match = matches[0]
+    next_heading = re.search(r"^##\s+", document[match.end() :], re.MULTILINE)
+    end = match.end() + next_heading.start() if next_heading is not None else len(document)
+    return document[match.start() : end]
+
+
+def _copyable_prompt(
+    section: str, heading: str = r"可复制(?:通用)?提示词"
+) -> str:
+    markers = list(
+        re.finditer(rf"^###\s+{heading}\s*$", section, re.MULTILINE)
+    )
+    if not markers:
+        raise ValueError("source entry has no copyable prompt")
+    if len(markers) != 1:
+        raise ValueError("source entry has duplicate copyable prompts")
+    marker = markers[0]
+    body = section[marker.end() :]
+    following = re.search(r"^###\s+|^##\s+", body, re.MULTILINE)
+    if following is not None:
+        body = body[: following.start()]
+    body_lines = [line for line in body.splitlines() if line.strip()]
+    if any(not line.startswith(">") for line in body_lines):
+        raise ValueError("source entry copyable prompt contains unquoted content")
+    lines = [line[1:].lstrip() for line in body_lines]
+    if not lines or not "\n".join(lines).strip():
+        raise ValueError("source entry copyable prompt is empty")
+    return "\n".join(lines).strip()
+
+
+def _scope_items(value: str) -> list[str]:
+    return [item.strip() for item in re.split(r"[、,，]", value) if item.strip()]
+
+
+def _contains_ref_token(value: str) -> bool:
+    return "ref-" in value.casefold()
+
+
+def _contains_plan_token(value: str) -> bool:
+    return "plan-" in value.casefold()
+
+
+def _markdown_reference_bindings(
+    section: str, *, field_name: str, creator_supplied_ok: bool = False
+) -> list[dict[str, Any]]:
+    lines = re.findall(
+        rf"^- {re.escape(field_name)}：(.+)$", section, re.MULTILINE
+    )
+    if not lines:
+        raise ValueError(f"source entry has no {field_name} declaration")
+    if len(lines) != 1:
+        raise ValueError(f"source entry has duplicate {field_name} declarations")
+    value = lines[0].strip()
+    if (
+        field_name == "输入参考图"
+        and not _contains_ref_token(value)
+        and re.fullmatch(r"无(?:（[^）\n]+）)?。?", value)
+    ):
+        return []
+    if (
+        field_name == "参考"
+        and not _contains_ref_token(value)
+        and re.fullmatch(r"无(?:外部参考)?(?:；[^\n]*)?。?", value)
+    ):
+        return []
+    # A `PLAN-...` slot is a picture the creator attaches in their own tool, so
+    # this suite has no file to send. The rendering job that draws a shot's own
+    # start frame is unaffected -- it works from the frozen keyframe text, the
+    # way it already does for 「待补参考图」 -- so only the job that would have
+    # sent these pictures as model inputs refuses. Matched as a slot head rather
+    # than a bare substring, or a real file named `plan-sheet.png` would refuse.
+    if PLAN_SLOT_HEAD_RE.search(value):
+        if not creator_supplied_ok:
+            raise ValueError(
+                "source entry declares creator-supplied references (PLAN-...); "
+                "bind the real files as REF-... before producing"
+            )
+        value = PLAN_SLOT_RE.sub("", value).strip("；;。 ")
+        if not value:
+            return []
+    matches = list(REFERENCE_LINE_RE.finditer(value))
+    if not matches:
+        raise ValueError("source entry input-reference declaration is invalid")
+    cursor = 0
+    for index, match in enumerate(matches):
+        separator = value[cursor : match.start()]
+        if separator != ("" if index == 0 else "；"):
+            raise ValueError("source entry input-reference declaration is invalid")
+        cursor = match.end()
+    if value[cursor:] not in {"", "。"}:
+        raise ValueError("source entry input-reference declaration is invalid")
+    return [
+        {
+            "slot_id": match.group(1),
+            "order": int(match.group(2)),
+            "path": _relative_path(match.group(3)),
+            "label": match.group(4).strip(),
+            "may_control": _scope_items(match.group(5)),
+            "must_not_control": _scope_items(match.group(6)),
+        }
+        for match in matches
+    ]
+
+
+def _verify_markdown_source(
+    root: Path,
+    *,
+    source: str,
+    source_entry: str,
+    prompt: str,
+    bindings: list[dict[str, Any]],
+) -> None:
+    source_path = _project_file(root, source)
+    if source_path.suffix.casefold() != ".md":
+        raise ValueError("source_entry requires a Markdown source")
+    document = source_path.read_text(encoding="utf-8")
+    section = _markdown_section(document, source_entry)
+    _, heading, field_name = CREATOR_SOURCE_ENTRIES[source_path.name]
+    if _copyable_prompt(section, heading) != prompt.strip():
+        raise ValueError("job prompt does not match the selected source entry")
+    declared = _markdown_reference_bindings(
+        section,
+        field_name=field_name,
+        # 分镜.md is the entry that renders a shot's own start frame. Its
+        # reference field records creative intent, not this job's inputs.
+        creator_supplied_ok=source_path.name == "分镜.md",
+    )
+    comparable = [
+        {key: binding[key] for key in (
+            "slot_id",
+            "order",
+            "path",
+            "label",
+            "may_control",
+            "must_not_control",
+        )}
+        for binding in bindings
+    ]
+    if declared != comparable:
+        raise ValueError("job reference bindings do not match the selected source entry")
+
+
+def _normalize_job(root: Path, raw: object) -> dict[str, Any]:
+    if not isinstance(raw, Mapping) or set(raw) - ALLOWED_JOB_KEYS:
+        raise ValueError("job contains unsupported fields")
+    if raw.get("schema_version", JOB_SCHEMA) != JOB_SCHEMA:
+        raise ValueError("unsupported job schema")
+    job_id = raw.get("job_id")
+    if not isinstance(job_id, str) or JOB_ID_RE.fullmatch(job_id) is None:
+        raise ValueError("job_id must be a portable 1-80 character identifier")
+    modality = raw.get("modality")
+    if modality not in MEDIA_EXTENSIONS:
+        raise ValueError("modality must be image, video, tts, or music")
+    adapter = raw.get("adapter")
+    if not isinstance(adapter, str) or JOB_ID_RE.fullmatch(adapter) is None:
+        raise ValueError("adapter must be a portable profile name")
+    prompt = raw.get("prompt")
+    if not isinstance(prompt, str) or not prompt.strip() or len(prompt) > 100_000:
+        raise ValueError("prompt must be non-empty and at most 100000 characters")
+    source_raw = raw.get("source")
+    source = _relative_path(source_raw) if source_raw is not None else None
+    source_entry = raw.get("source_entry")
+    if source_entry is not None and (
+        not isinstance(source_entry, str)
+        or SOURCE_ENTRY_RE.fullmatch(source_entry) is None
+    ):
+        raise ValueError("source_entry must be a visible uppercase Markdown entry ID")
+    if source_entry is not None and source is None:
+        raise ValueError("source_entry requires source")
+    creator_source_modality = _canonical_creator_source_modality(source)
+    if creator_source_modality == modality and source_entry is None:
+        raise ValueError("creator Markdown source requires source_entry")
+    if creator_source_modality is not None and modality != creator_source_modality and not (
+        creator_source_modality == "video"
+        and modality == "music"
+        and source_entry is None
+    ):
+        raise ValueError("creator Markdown source does not match the job modality")
+    expected_entry_prefix = (
+        CREATOR_SOURCE_ENTRIES[PurePosixPath(source).name][0]
+        if source is not None
+        and PurePosixPath(source).name in CREATOR_SOURCE_ENTRIES
+        and creator_source_modality == modality
+        else None
+    )
+    if source_entry is not None and creator_source_modality != modality:
+        raise ValueError("source_entry requires the canonical creator Markdown path")
+    if source_entry is not None and (
+        expected_entry_prefix is None
+        or not source_entry.startswith(expected_entry_prefix)
+    ):
+        raise ValueError("source_entry does not match the job modality")
+    references_supplied = "references" in raw
+    supplied_references = [
+        _relative_path(path)
+        for path in _string_list(raw.get("references", []), label="references", limit=16)
+    ]
+    reference_bindings = _normalize_reference_bindings(raw.get("reference_bindings"))
+    binding_references = [str(binding["path"]) for binding in reference_bindings]
+    if (
+        references_supplied
+        and (reference_bindings or source_entry is not None)
+        and supplied_references != binding_references
+    ):
+        raise ValueError("references must match reference_bindings order")
+    references = binding_references if reference_bindings else supplied_references
+    if source_entry is not None:
+        _verify_markdown_source(
+            root,
+            source=str(source),
+            source_entry=source_entry,
+            prompt=prompt,
+            bindings=reference_bindings,
+        )
+    outputs = [
+        _relative_path(path, output=True)
+        for path in _string_list(raw.get("outputs"), label="outputs", limit=16)
+    ]
+    if not outputs or len(set(outputs)) != len(outputs):
+        raise ValueError("outputs must contain unique target paths")
+    for output_path in outputs:
+        if PurePosixPath(output_path).suffix.casefold() not in MEDIA_EXTENSIONS[str(modality)]:
+            raise ValueError(f"output extension does not match {modality}: {output_path}")
+        _project_file(root, output_path)
+    parameters = raw.get("parameters", {})
+    if not isinstance(parameters, Mapping):
+        raise ValueError("parameters must be an object")
+    if _contains_secret_key(parameters):
+        raise ValueError("job parameters must not contain credentials or secrets")
+    if len(_canonical(parameters)) > 64 * 1024:
+        raise ValueError("job parameters are too large")
+    input_paths = ([source] if source is not None else []) + references
+    if len(set(input_paths)) != len(input_paths):
+        raise ValueError("source and references must be unique")
+    input_hashes = {path: _hash_project_file(root, path) for path in input_paths}
+    execution = {
+        "schema_version": JOB_SCHEMA,
+        "job_id": job_id,
+        "modality": modality,
+        "adapter": adapter,
+        "prompt": prompt,
+        "source": source,
+        "source_entry": source_entry,
+        "references": references,
+        "reference_bindings": reference_bindings,
+        "outputs": outputs,
+        "parameters": dict(parameters),
+        "overwrite": raw.get("overwrite", False),
+        "inputs": input_hashes,
+    }
+    if not isinstance(execution["overwrite"], bool):
+        raise ValueError("overwrite must be a boolean")
+    execution["fingerprint"] = sha256_bytes(_canonical(execution))
+    execution["prepared_at"] = utc_now()
+    return execution
+
+
+def prepare_job(root: Path, job_file: Path) -> dict[str, Any]:
+    root = find_project(root)
+    if job_file.stat().st_size > MAX_JOB_BYTES:
+        raise ValueError("job file is too large")
+    raw = json.loads(job_file.read_text(encoding="utf-8"))
+    job = _normalize_job(root, raw)
+    with _project_lock(root):
+        if _active_run(root, str(job["job_id"])) is not None:
+            raise RuntimeError("this job is already running")
+        job_name = f"{_job_key(str(job['job_id']))}.json"
+        _metadata_atomic_json(root, ("jobs",), job_name, job)
+        try:
+            _metadata_unlink(root, ("confirmations",), job_name)
+        except FileNotFoundError:
+            pass
+    return _preview(job)
+
+
+def _validate_stored_job(
+    root: Path, document: object, *, expected_job_id: str | None = None
+) -> dict[str, Any]:
+    root = root.resolve()
+    if not isinstance(document, dict):
+        raise ValueError("stored job fields are invalid")
+    stored_keys = set(document)
+    legacy = stored_keys == LEGACY_STORED_JOB_KEYS
+    if not legacy and stored_keys != STORED_JOB_KEYS:
+        raise ValueError("stored job fields are invalid")
+    original = document
+    document = dict(document)
+    if legacy:
+        document["source_entry"] = None
+        document["reference_bindings"] = []
+    job_id = document.get("job_id")
+    if (
+        not isinstance(job_id, str)
+        or JOB_ID_RE.fullmatch(job_id) is None
+        or (expected_job_id is not None and job_id != expected_job_id)
+    ):
+        raise ValueError("stored job id is invalid")
+    if document.get("schema_version") != JOB_SCHEMA:
+        raise ValueError("stored job schema is invalid")
+    modality = document.get("modality")
+    if modality not in MEDIA_EXTENSIONS:
+        raise ValueError("stored job modality is invalid")
+    adapter = document.get("adapter")
+    if not isinstance(adapter, str) or JOB_ID_RE.fullmatch(adapter) is None:
+        raise ValueError("stored job adapter is invalid")
+    prompt = document.get("prompt")
+    if not isinstance(prompt, str) or not prompt.strip() or len(prompt) > 100_000:
+        raise ValueError("stored job prompt is invalid")
+    source = document.get("source")
+    if source is not None and _relative_path(source) != source:
+        raise ValueError("stored job source is invalid")
+    source_entry = document.get("source_entry")
+    if source_entry is not None and (
+        not isinstance(source_entry, str)
+        or SOURCE_ENTRY_RE.fullmatch(source_entry) is None
+        or source is None
+    ):
+        raise ValueError("stored job source entry is invalid")
+    expected_entry_prefix = {"image": "IMG-", "video": "MOTION-"}.get(
+        str(modality)
+    )
+    if source_entry is not None and (
+        expected_entry_prefix is None
+        or not source_entry.startswith(expected_entry_prefix)
+    ):
+        raise ValueError("stored job source entry has the wrong modality")
+    creator_source_modality = _canonical_creator_source_modality(
+        str(source) if source is not None else None
+    )
+    if not legacy and creator_source_modality == modality and source_entry is None:
+        raise ValueError("stored creator Markdown job has no source entry")
+    if creator_source_modality is not None and modality != creator_source_modality and not (
+        creator_source_modality == "video"
+        and modality == "music"
+        and source_entry is None
+    ):
+        raise ValueError("stored creator Markdown source has the wrong modality")
+    if source_entry is not None and creator_source_modality != modality:
+        raise ValueError("stored job creator source path is invalid")
+    references = _string_list(
+        document.get("references"), label="stored references", limit=16
+    )
+    if any(_relative_path(reference) != reference for reference in references):
+        raise ValueError("stored job references are invalid")
+    reference_bindings = _normalize_reference_bindings(
+        document.get("reference_bindings")
+    )
+    binding_references = [binding["path"] for binding in reference_bindings]
+    if (
+        source_entry is not None or reference_bindings
+    ) and binding_references != references:
+        raise ValueError("stored job references do not match reference bindings")
+    outputs = _string_list(document.get("outputs"), label="stored outputs", limit=16)
+    if not outputs or len(outputs) != len(set(outputs)):
+        raise ValueError("stored job outputs are invalid")
+    for output in outputs:
+        if _relative_path(output, output=True) != output:
+            raise ValueError("stored job output is invalid")
+        if PurePosixPath(output).suffix.casefold() not in MEDIA_EXTENSIONS[str(modality)]:
+            raise ValueError("stored job output extension is invalid")
+        _project_file(root, output)
+    parameters = document.get("parameters")
+    if (
+        not isinstance(parameters, Mapping)
+        or _contains_secret_key(parameters)
+        or len(_canonical(parameters)) > 64 * 1024
+    ):
+        raise ValueError("stored job parameters are invalid")
+    if not isinstance(document.get("overwrite"), bool):
+        raise ValueError("stored job overwrite flag is invalid")
+    input_paths = ([source] if source is not None else []) + references
+    if len(input_paths) != len(set(input_paths)):
+        raise ValueError("stored job inputs are duplicated")
+    inputs = document.get("inputs")
+    if not isinstance(inputs, Mapping) or set(inputs) != set(input_paths):
+        raise ValueError("stored job input hashes are invalid")
+    if any(
+        not isinstance(digest, str) or re.fullmatch(r"[0-9a-f]{64}", digest) is None
+        for digest in inputs.values()
+    ):
+        raise ValueError("stored job input hash is invalid")
+    fingerprint = document.get("fingerprint")
+    execution_keys = (
+        LEGACY_STORED_EXECUTION_KEYS if legacy else STORED_EXECUTION_KEYS
+    )
+    execution = {key: original[key] for key in execution_keys}
+    if (
+        not isinstance(fingerprint, str)
+        or re.fullmatch(r"[0-9a-f]{64}", fingerprint) is None
+        or sha256_bytes(_canonical(execution)) != fingerprint
+    ):
+        raise ValueError("stored job fingerprint is invalid")
+    _parse_run_timestamp(document.get("prepared_at"), label="prepared_at")
+    return document
+
+
+def _read_job(root: Path, job_id: str) -> dict[str, Any]:
+    if JOB_ID_RE.fullmatch(job_id) is None:
+        raise ValueError("invalid job_id")
+    document = _metadata_read_json(
+        root,
+        ("jobs",),
+        f"{_job_key(job_id)}.json",
+        maximum=MAX_JOB_BYTES,
+    )
+    return _validate_stored_job(root, document, expected_job_id=job_id)
+
+
+def _preview(job: Mapping[str, Any]) -> dict[str, Any]:
+    confirmation = f"CONFIRM {job['job_id']} {str(job['fingerprint'])[:12]}"
+    return {
+        "job_id": job["job_id"],
+        "modality": job["modality"],
+        "adapter": job["adapter"],
+        "count": len(job["outputs"]),
+        "prompt": job["prompt"],
+        "source": job["source"],
+        "source_entry": job["source_entry"],
+        "references": job["references"],
+        "reference_bindings": job["reference_bindings"],
+        "outputs": job["outputs"],
+        "parameters": job["parameters"],
+        "overwrite": job["overwrite"],
+        "confirmation": confirmation,
+        "state": "needs_confirmation",
+    }
+
+
+def confirm_job(root: Path, *, job_id: str, confirmation: str) -> dict[str, Any]:
+    root = find_project(root)
+    with _project_lock(root):
+        if _active_run(root, job_id) is not None:
+            raise RuntimeError("this job is already running")
+        job = _read_job(root, job_id)
+        expected = _preview(job)["confirmation"]
+        if confirmation != expected:
+            raise ConfirmationRequiredError("confirmation does not match the exact current job")
+        receipt = {
+            "schema_version": JOB_SCHEMA,
+            "job_id": job_id,
+            "fingerprint": job["fingerprint"],
+            "confirmed_at": utc_now(),
+            "consumed_at": None,
+            "run_id": None,
+        }
+        _metadata_atomic_json(
+            root, ("confirmations",), f"{_job_key(job_id)}.json", receipt
+        )
+    return {"job_id": job_id, "state": "confirmed"}
+
+
+def _inputs_current(root: Path, job: Mapping[str, Any]) -> bool:
+    inputs = job.get("inputs")
+    if not isinstance(inputs, Mapping):
+        return False
+    try:
+        return all(_hash_project_file(root, str(path)) == digest for path, digest in inputs.items())
+    except (FileNotFoundError, OSError, ValueError):
+        return False
+
+
+def _load_adapter(config_path: Path, profile: str, root: Path) -> tuple[list[str], int]:
+    resolved = config_path.expanduser().resolve()
+    if resolved.is_relative_to(root):
+        raise ValueError("adapter config must live outside the project")
+    document = json.loads(resolved.read_text(encoding="utf-8"))
+    adapters = document.get("adapters") if isinstance(document, Mapping) else None
+    selected = adapters.get(profile) if isinstance(adapters, Mapping) else None
+    if not isinstance(selected, Mapping) or set(selected) - {"command", "timeout_seconds"}:
+        raise ValueError(f"adapter profile is missing or invalid: {profile}")
+    command = _string_list(selected.get("command"), label="adapter command", limit=32)
+    if not command or any(not part for part in command):
+        raise ValueError("adapter command must be a non-empty argv list")
+    timeout = selected.get("timeout_seconds", 300)
+    if not isinstance(timeout, int) or isinstance(timeout, bool) or not 1 <= timeout <= MAX_TIMEOUT_SECONDS:
+        raise ValueError(f"adapter timeout must be 1-{MAX_TIMEOUT_SECONDS} seconds")
+    return command, timeout
+
+
+def _generic_adapter_error(
+    profile: str,
+    *,
+    category: str,
+    code: str,
+    retryable: bool,
+) -> dict[str, Any]:
+    return {
+        "provider": profile,
+        "category": category,
+        "code": code,
+        "retryable": retryable,
+    }
+
+
+def _parse_public_adapter_error(
+    raw: bytes, *, profile: str, returncode: int
+) -> dict[str, Any]:
+    fallback = _generic_adapter_error(
+        profile,
+        category="provider_response",
+        code=f"adapter_exit_{returncode}",
+        retryable=False,
+    )
+    if len(raw) > MAX_ADAPTER_RESPONSE_BYTES:
+        return fallback
+    try:
+        document = json.loads(raw.decode("utf-8"))
+    except (UnicodeError, json.JSONDecodeError):
+        return fallback
+    error = document.get("error") if isinstance(document, Mapping) else None
+    if not isinstance(error, Mapping) or set(error) - {
+        "provider",
+        "category",
+        "code",
+        "http_status",
+        "request_id",
+        "retryable",
+    }:
+        return fallback
+    provider = error.get("provider")
+    category = error.get("category")
+    code = error.get("code")
+    retryable = error.get("retryable")
+    if (
+        provider != profile
+        or category not in PUBLIC_ERROR_CATEGORIES
+        or not isinstance(code, str)
+        or PUBLIC_ERROR_TOKEN_RE.fullmatch(code) is None
+        or not isinstance(retryable, bool)
+    ):
+        return fallback
+    result: dict[str, Any] = {
+        "provider": provider,
+        "category": category,
+        "code": code,
+        "retryable": retryable,
+    }
+    status = error.get("http_status")
+    if status is not None:
+        if not isinstance(status, int) or isinstance(status, bool) or not 100 <= status <= 599:
+            return fallback
+        result["http_status"] = status
+    request_id = error.get("request_id")
+    if request_id is not None:
+        if (
+            not isinstance(request_id, str)
+            or PUBLIC_ERROR_TOKEN_RE.fullmatch(request_id) is None
+        ):
+            return fallback
+        result["request_id"] = request_id
+    return result
+
+
+def _run_adapter(command: list[str], timeout: int, payload: Mapping[str, Any], root: Path) -> dict[str, Any]:
+    with tempfile.TemporaryFile() as stdout, tempfile.TemporaryFile() as stderr:
+        try:
+            completed = subprocess.run(
+                command,
+                input=_canonical(payload),
+                stdout=stdout,
+                stderr=stderr,
+                cwd=root,
+                timeout=timeout,
+                check=False,
+            )
+        except subprocess.TimeoutExpired as exc:
+            raise AdapterError(
+                "adapter timed out; confirmation was consumed",
+                public_error=_generic_adapter_error(
+                    str(payload["adapter"]),
+                    category="timeout",
+                    code="adapter_timeout",
+                    retryable=True,
+                ),
+            ) from exc
+        except OSError as exc:
+            raise AdapterError(
+                "adapter could not be started; confirmation was consumed",
+                public_error=_generic_adapter_error(
+                    str(payload["adapter"]),
+                    category="configuration",
+                    code="adapter_start_failed",
+                    retryable=False,
+                ),
+            ) from exc
+        if completed.returncode != 0:
+            size = stdout.tell()
+            stdout.seek(0)
+            raw_error = stdout.read(MAX_ADAPTER_RESPONSE_BYTES + 1)
+            public_error = _parse_public_adapter_error(
+                raw_error if size <= MAX_ADAPTER_RESPONSE_BYTES else b"",
+                profile=str(payload["adapter"]),
+                returncode=completed.returncode,
+            )
+            raise AdapterError(
+                f"adapter exited with code {completed.returncode}; confirmation was consumed",
+                public_error=public_error,
+            )
+        size = stdout.tell()
+        if size > MAX_ADAPTER_RESPONSE_BYTES:
+            raise AdapterError("adapter response is too large; confirmation was consumed")
+        stdout.seek(0)
+        try:
+            response = json.loads(stdout.read().decode("utf-8"))
+        except (UnicodeError, json.JSONDecodeError) as exc:
+            raise AdapterError("adapter returned invalid JSON; confirmation was consumed") from exc
+    if not isinstance(response, dict):
+        raise AdapterError("adapter response must be an object; confirmation was consumed")
+    return response
+
+
+def _validate_adapter_outputs(
+    job: Mapping[str, Any], response: Mapping[str, Any], output_root: Path
+) -> list[tuple[str, Path]]:
+    entries = response.get("outputs")
+    if not isinstance(entries, list) or not all(isinstance(entry, Mapping) for entry in entries):
+        raise AdapterError("adapter outputs are invalid; confirmation was consumed")
+    result: list[tuple[str, Path]] = []
+    for entry in entries:
+        if set(entry) != {"target", "source"}:
+            raise AdapterError("adapter output fields are invalid; confirmation was consumed")
+        target = entry.get("target")
+        source = entry.get("source")
+        if not isinstance(target, str) or not isinstance(source, str):
+            raise AdapterError("adapter output paths are invalid; confirmation was consumed")
+        path = Path(source)
+        if not path.is_absolute() or path.parent != output_root or path.name in {"", ".", ".."}:
+            raise AdapterError(
+                "adapter output must use the run staging directory; confirmation was consumed"
+            )
+        result.append((target, path))
+    expected = list(job["outputs"])
+    if [target for target, _ in result] != expected:
+        raise AdapterError("adapter outputs do not match the confirmed targets; confirmation was consumed")
+    return result
+
+
+def _copy_output(
+    source: Path, target: Path, *, overwrite: bool
+) -> tuple[str, int]:
+    temporary = target.with_name(f".{target.name}.{uuid.uuid4().hex}.tmp")
+    target.parent.mkdir(parents=True, exist_ok=True)
+    digest = hashlib.sha256()
+    size = 0
+    descriptor = -1
+    try:
+        try:
+            before = source.lstat()
+        except OSError as exc:
+            raise AdapterError(
+                "adapter output file is missing; confirmation was consumed"
+            ) from exc
+        if _is_link_or_reparse(before) or not stat.S_ISREG(before.st_mode):
+            raise AdapterError("adapter output file is unsafe; confirmation was consumed")
+        descriptor = os.open(source, os.O_RDONLY | getattr(os, "O_NOFOLLOW", 0))
+        opened = os.fstat(descriptor)
+        if (
+            not stat.S_ISREG(opened.st_mode)
+            or (opened.st_dev, opened.st_ino) != (before.st_dev, before.st_ino)
+        ):
+            raise AdapterError("adapter output changed while opening; confirmation was consumed")
+        if opened.st_size > MAX_OUTPUT_BYTES:
+            raise AdapterError("adapter output file is too large; confirmation was consumed")
+        with os.fdopen(descriptor, "rb", closefd=True) as incoming, temporary.open(
+            "xb"
+        ) as outgoing:
+            descriptor = -1
+            for chunk in iter(lambda: incoming.read(1024 * 1024), b""):
+                size += len(chunk)
+                if size > MAX_OUTPUT_BYTES:
+                    raise AdapterError("adapter output exceeded the size limit")
+                digest.update(chunk)
+                outgoing.write(chunk)
+            outgoing.flush()
+            os.fsync(outgoing.fileno())
+        if overwrite:
+            os.replace(temporary, target)
+        else:
+            try:
+                os.link(temporary, target, follow_symlinks=False)
+            except FileExistsError as exc:
+                raise FileExistsError(
+                    f"output appeared while production was running: {target.name}"
+                ) from exc
+            temporary.unlink()
+    finally:
+        if descriptor >= 0:
+            os.close(descriptor)
+        try:
+            temporary.unlink()
+        except FileNotFoundError:
+            pass
+    return digest.hexdigest(), size
+
+
+def _latest_run(root: Path, job_id: str) -> dict[str, Any] | None:
+    history = _read_run_history(root, job_id)
+    return history[-1] if history else None
+
+
+def _parse_run_timestamp(value: object, *, label: str) -> datetime:
+    if not isinstance(value, str) or not value:
+        raise ValueError(f"production run {label} is invalid")
+    normalized = value[:-1] + "+00:00" if value.endswith("Z") else value
+    try:
+        parsed = datetime.fromisoformat(normalized)
+    except ValueError as exc:
+        raise ValueError(f"production run {label} is invalid") from exc
+    if parsed.tzinfo is None or parsed.utcoffset() is None:
+        raise ValueError(f"production run {label} is invalid")
+    return parsed.astimezone(timezone.utc)
+
+
+def _completed_run_order(run: Mapping[str, Any]) -> tuple[datetime, str]:
+    return (
+        _parse_run_timestamp(run.get("finished_at"), label="finished_at"),
+        str(run["run_id"]),
+    )
+
+
+def _read_run_history(root: Path, job_id: str) -> list[dict[str, Any]]:
+    directory_parts = ("runs", _job_key(job_id))
+    history: list[dict[str, Any]] = []
+    for name in _metadata_json_names(root, directory_parts):
+        document = _metadata_read_json(
+            root, directory_parts, name, maximum=MAX_RUN_RECORD_BYTES
+        )
+        if (
+            not isinstance(document, dict)
+            or document.get("job_id") != job_id
+            or not isinstance(document.get("run_id"), str)
+            or document.get("status") not in {"running", "succeeded", "failed"}
+            or not isinstance(document.get("fingerprint"), str)
+            or re.fullmatch(r"[0-9a-f]{64}", str(document.get("fingerprint"))) is None
+        ):
+            raise ValueError("production run record is invalid")
+        started_at = _parse_run_timestamp(
+            document.get("started_at"), label="started_at"
+        )
+        finished_at = document.get("finished_at")
+        if document["status"] == "running":
+            if finished_at is not None:
+                raise ValueError("running production run has finished_at")
+        else:
+            completed_at = _parse_run_timestamp(finished_at, label="finished_at")
+            if completed_at < started_at:
+                raise ValueError("production run finished before it started")
+        history.append(document)
+    # Completed runs are ordered by completion because that is when their
+    # terminal state and output claim become authoritative. Any unresolved
+    # running attempt sorts last so status cannot hide it behind a later start.
+    history.sort(
+        key=lambda run: (
+            run["status"] == "running",
+            (
+                _parse_run_timestamp(run.get("started_at"), label="started_at")
+                if run["status"] == "running"
+                else _completed_run_order(run)[0]
+            ),
+            str(run["run_id"]),
+        )
+    )
+    return history
+
+
+def _active_run(root: Path, job_id: str) -> dict[str, Any] | None:
+    running = [
+        run for run in _read_run_history(root, job_id) if run["status"] == "running"
+    ]
+    return running[-1] if running else None
+
+
+def _read_provider_handle(root: Path, job_id: str, run_id: str) -> str | None:
+    """The provider job id an adapter recorded before it started polling.
+
+    A video task is billed the moment it is submitted, not when it is collected.
+    If the adapter is then interrupted — killed, disconnected, the machine
+    sleeps — the run record would otherwise say `running` with no way back to a
+    task that is alive and already paid for, and the only move left is to submit
+    (and pay) again. The adapter writes the id here as soon as it has one, into
+    project metadata rather than the attempt's temporary directory, so it
+    outlives the process that created it.
+    """
+
+    try:
+        document = _metadata_read_json(
+            root, ("handles", _job_key(job_id)), f"{run_id}.json",
+            maximum=MAX_RUN_RECORD_BYTES,
+        )
+    except (FileNotFoundError, ValueError, json.JSONDecodeError):
+        return None
+    if not isinstance(document, Mapping):
+        return None
+    value = document.get("provider_job_id")
+    if isinstance(value, str) and 0 < len(value) <= 200:
+        return value
+    return None
+
+
+def _handle_metadata_path(root: Path, job_id: str, run_id: str) -> Path:
+    return root / PRODUCTION_ROOT / "handles" / _job_key(job_id) / f"{run_id}.json"
+
+
+def _write_run(root: Path, job_id: str, run: Mapping[str, Any]) -> None:
+    _metadata_atomic_json(
+        root,
+        ("runs", _job_key(job_id)),
+        f"{run['run_id']}.json",
+        run,
+    )
+
+
+def run_job(root: Path, *, job_id: str, adapter_config: Path) -> dict[str, Any]:
+    root = find_project(root)
+    with tempfile.TemporaryDirectory(prefix="short-drama-inputs-") as directory:
+        attempt_root = Path(directory)
+        snapshot_root = attempt_root / "inputs"
+        output_root = attempt_root / "outputs"
+        snapshot_root.mkdir()
+        output_root.mkdir()
+        with _project_lock(root):
+            if _active_run(root, job_id) is not None:
+                raise RuntimeError("this job is already running")
+            job = _read_job(root, job_id)
+            command, timeout = _load_adapter(adapter_config, str(job["adapter"]), root)
+            try:
+                receipt = _metadata_read_json(
+                    root,
+                    ("confirmations",),
+                    f"{_job_key(job_id)}.json",
+                    maximum=MAX_RUN_RECORD_BYTES,
+                )
+            except FileNotFoundError as exc:
+                raise ConfirmationRequiredError("job needs explicit confirmation") from exc
+            if (
+                not isinstance(receipt, dict)
+                or receipt.get("fingerprint") != job.get("fingerprint")
+                or receipt.get("consumed_at") is not None
+            ):
+                raise ConfirmationRequiredError("job needs a new explicit confirmation")
+            for output in job["outputs"]:
+                target = _project_file(root, output)
+                if target.exists() and not job["overwrite"]:
+                    raise FileExistsError(f"output exists and overwrite is false: {output}")
+            # Pin every confirmed input into a private immutable snapshot before
+            # consuming confirmation. Provider adapters never reopen live project paths.
+            _snapshot_inputs(root, job, snapshot_root)
+            run_id = f"{datetime.now(timezone.utc).strftime('%Y%m%dT%H%M%S')}-{uuid.uuid4().hex[:8]}"
+            receipt["consumed_at"] = utc_now()
+            receipt["run_id"] = run_id
+            _metadata_atomic_json(
+                root,
+                ("confirmations",),
+                f"{_job_key(job_id)}.json",
+                receipt,
+            )
+            run = {
+                "schema_version": JOB_SCHEMA,
+                "run_id": run_id,
+                "job_id": job_id,
+                "fingerprint": job["fingerprint"],
+                "modality": job["modality"],
+                "adapter": job["adapter"],
+                "status": "running",
+                "started_at": utc_now(),
+                "finished_at": None,
+                "outputs": [],
+            }
+            _write_run(root, job_id, run)
+
+        payload = {key: job[key] for key in ALLOWED_JOB_KEYS if key in job}
+        payload.update(
+            {
+                "run_id": run_id,
+                "project_root": str(snapshot_root),
+                "output_root": str(output_root),
+                # Where to record the provider job id the instant it exists, so
+                # an interrupted attempt can be collected instead of re-billed.
+                "handle_path": str(_handle_metadata_path(root, job_id, run_id)),
+            }
+        )
+        try:
+            response = _run_adapter(command, timeout, payload, root)
+            adapter_outputs = _validate_adapter_outputs(job, response, output_root)
+            written: list[dict[str, Any]] = []
+            with _project_lock(root):
+                for target_name, source in adapter_outputs:
+                    target = _project_file(root, target_name, create_parent=True)
+                    digest, size = _copy_output(
+                        source, target, overwrite=bool(job["overwrite"])
+                    )
+                    written.append(
+                        {
+                            "path": target_name,
+                            "media_type": MEDIA_TYPES[PurePosixPath(target_name).suffix.casefold()],
+                            "bytes": size,
+                            "sha256": digest,
+                        }
+                    )
+                if _run_status(root, job_id, run_id) == "succeeded":
+                    # `collect` reached the same provider task first and its
+                    # outputs are already in the project. Leave that record
+                    # alone rather than replacing a collected attempt.
+                    raise RuntimeError(
+                        "这次尝试已经被 collect 取回并标为成功；本次运行不覆盖它"
+                    )
+                run["status"] = "succeeded"
+                run["finished_at"] = utc_now()
+                run["outputs"] = written
+                provider_job_id = response.get("provider_job_id") or _read_provider_handle(
+                    root, job_id, run_id
+                )
+                if isinstance(provider_job_id, str) and 0 < len(provider_job_id) <= 200:
+                    run["provider_job_id"] = provider_job_id
+                _write_run(root, job_id, run)
+        except Exception as exc:
+            with _project_lock(root):
+                run["status"] = "failed"
+                run["finished_at"] = utc_now()
+                if isinstance(exc, AdapterError) and exc.public_error is not None:
+                    run["error"] = exc.public_error
+                # A timeout or a killed adapter does not cancel the provider's
+                # task. Carry the handle onto the failed record so `audit` can
+                # say "collect this" instead of the creator paying twice.
+                orphan = _read_provider_handle(root, job_id, run_id)
+                if orphan is not None:
+                    run["provider_job_id"] = orphan
+                _write_run(root, job_id, run)
+            raise
+    return {
+        "job_id": job_id,
+        "run_id": run_id,
+        "state": "succeeded",
+        "outputs": run["outputs"],
+    }
+
+
+def _run_status(root: Path, job_id: str, run_id: str) -> str | None:
+    """The status one attempt currently has on disk, re-read rather than cached."""
+
+    for run in _read_run_history(root, job_id):
+        if str(run.get("run_id")) == run_id:
+            status = run.get("status")
+            return status if isinstance(status, str) else None
+    return None
+
+
+def collect_job(root: Path, *, job_id: str, adapter_config: Path) -> dict[str, Any]:
+    """Collect an already-submitted task's result without submitting a new one.
+
+    This is deliberately outside the confirmation gate. The gate exists to stop
+    an unintended charge; collecting costs nothing, because the charge already
+    happened when the task was submitted. Requiring a fresh confirmation here
+    would mean the cheapest way out of an interrupted attempt is to pay again,
+    which is exactly the outcome the gate is meant to prevent.
+    """
+
+    root = find_project(root)
+    with tempfile.TemporaryDirectory(prefix="short-drama-collect-") as directory:
+        output_root = Path(directory) / "outputs"
+        output_root.mkdir(parents=True)
+        with _project_lock(root):
+            job = _read_job(root, job_id)
+            command, timeout = _load_adapter(adapter_config, str(job["adapter"]), root)
+            target_run = None
+            for run in reversed(_read_run_history(root, job_id)):
+                if run.get("status") == "succeeded":
+                    continue
+                handle = run.get("provider_job_id") or _read_provider_handle(
+                    root, job_id, str(run["run_id"])
+                )
+                if handle:
+                    target_run = (run, handle)
+                    break
+            if target_run is None:
+                raise RuntimeError(
+                    "no unfinished attempt of this job carries a provider job id; "
+                    "there is nothing to collect"
+                )
+            run, provider_job_id = target_run
+
+        payload = {key: job[key] for key in ALLOWED_JOB_KEYS if key in job}
+        payload.update(
+            {
+                "run_id": str(run["run_id"]),
+                "output_root": str(output_root),
+                "collect_provider_job_id": provider_job_id,
+            }
+        )
+        response = _run_adapter(command, timeout, payload, root)
+        adapter_outputs = _validate_adapter_outputs(job, response, output_root)
+        written: list[dict[str, Any]] = []
+        with _project_lock(root):
+            # A live `run_job` on the same provider task is indistinguishable
+            # from a dead one — both sit at `running`. Re-read before writing,
+            # or two writers race onto the same bytes and one record is lost.
+            if _run_status(root, job_id, str(run["run_id"])) == "succeeded":
+                raise RuntimeError(
+                    "这次尝试在取回期间已经自己完成了，产物已经落进项目；"
+                    "没有覆盖它——先看 status，确认无误就不需要再做别的"
+                )
+            for target_name, source in adapter_outputs:
+                target = _project_file(root, target_name, create_parent=True)
+                digest, size = _copy_output(
+                    source, target, overwrite=bool(job["overwrite"])
+                )
+                written.append(
+                    {
+                        "path": target_name,
+                        "media_type": MEDIA_TYPES[PurePosixPath(target_name).suffix.casefold()],
+                        "bytes": size,
+                        "sha256": digest,
+                    }
+                )
+            collected = dict(run)
+            collected["status"] = "succeeded"
+            collected["finished_at"] = utc_now()
+            collected["outputs"] = written
+            collected["provider_job_id"] = provider_job_id
+            collected["collected"] = True
+            collected.pop("error", None)
+            _write_run(root, job_id, collected)
+    return {
+        "job_id": job_id,
+        "run_id": str(run["run_id"]),
+        "state": "succeeded",
+        "collected": True,
+        "provider_job_id": provider_job_id,
+        "outputs": written,
+    }
+
+
+def job_status(root: Path, *, job_id: str) -> dict[str, Any]:
+    root = find_project(root)
+    job = _read_job(root, job_id)
+    latest = _latest_run(root, job_id)
+    if not _inputs_current(root, job):
+        state = "needs_reconfirmation"
+    else:
+        try:
+            receipt = _metadata_read_json(
+                root,
+                ("confirmations",),
+                f"{_job_key(job_id)}.json",
+                maximum=MAX_RUN_RECORD_BYTES,
+            )
+        except FileNotFoundError:
+            receipt = None
+        if latest and latest.get("status") == "running":
+            state = "running"
+        elif latest and latest.get("status") in {"succeeded", "failed"}:
+            state = str(latest["status"])
+        elif isinstance(receipt, Mapping) and receipt.get("consumed_at") is None:
+            state = "confirmed"
+        else:
+            state = "needs_confirmation"
+    return {
+        "job_id": job_id,
+        "modality": job["modality"],
+        "adapter": job["adapter"],
+        "outputs": job["outputs"],
+        "state": state,
+        "latest_run": latest,
+    }
+
+
+def _hash_production_output(root: Path, relative: str) -> tuple[str, int]:
+    digest = hashlib.sha256()
+    size = 0
+    with _open_project_input(root, relative) as handle:
+        for chunk in iter(lambda: handle.read(1024 * 1024), b""):
+            size += len(chunk)
+            if size > MAX_OUTPUT_BYTES:
+                raise ValueError("production output exceeds the size limit")
+            digest.update(chunk)
+    return digest.hexdigest(), size
+
+
+def _validated_succeeded_outputs(
+    job: Mapping[str, Any], run: Mapping[str, Any]
+) -> list[tuple[str, str, int]]:
+    expected = job.get("outputs")
+    entries = run.get("outputs")
+    if (
+        not isinstance(expected, list)
+        or not isinstance(entries, list)
+        or len(entries) != len(expected)
+    ):
+        raise ValueError("succeeded outputs do not match the current job")
+    validated: list[tuple[str, str, int]] = []
+    # ``zip(strict=True)`` needs 3.10; the suite's floor is 3.9, so the length
+    # agreement is asserted directly rather than by the zip flag.
+    if len(expected) != len(entries):
+        raise ValueError("expected %d outputs, got %d" % (len(expected), len(entries)))
+    for expected_path, output in zip(expected, entries):
+        if not isinstance(output, Mapping) or set(output) != {
+            "path",
+            "media_type",
+            "bytes",
+            "sha256",
+        }:
+            raise ValueError("succeeded output record fields are invalid")
+        output_path = output.get("path")
+        digest = output.get("sha256")
+        size = output.get("bytes")
+        media_type = output.get("media_type")
+        if (
+            not isinstance(expected_path, str)
+            or output_path != expected_path
+            or not isinstance(digest, str)
+            or re.fullmatch(r"[0-9a-f]{64}", digest) is None
+            or not isinstance(size, int)
+            or isinstance(size, bool)
+            or size < 0
+            or media_type
+            != MEDIA_TYPES.get(PurePosixPath(expected_path).suffix.casefold())
+        ):
+            raise ValueError("succeeded output record is invalid")
+        validated.append((output_path, digest, size))
+    return validated
+
+
+def audit_project(root: Path) -> dict[str, Any]:
+    """Reconcile local attempt history and current output bytes, not media quality."""
+    root = find_project(root)
+    jobs: list[dict[str, Any]] = []
+    problems: list[dict[str, Any]] = []
+    for name in _metadata_json_names(root, ("jobs",)):
+        try:
+            document = _metadata_read_json(
+                root, ("jobs",), name, maximum=MAX_JOB_BYTES
+            )
+            job = _validate_stored_job(root, document)
+            job_id = job["job_id"]
+            if not isinstance(job_id, str) or name != f"{_job_key(job_id)}.json":
+                raise ValueError("stored job is invalid")
+            jobs.append(job)
+        except (OSError, ValueError, json.JSONDecodeError):
+            problems.append(
+                {
+                    "code": "invalid_job_record",
+                    "record": name,
+                    "action": "repair_production_metadata",
+                }
+            )
+
+    attempts_total = 0
+    attempts_succeeded = 0
+    attempts_failed = 0
+    attempts_running = 0
+    attempts_superseded = 0
+    repeated_content = 0
+    recovered_jobs = 0
+    running_jobs = 0
+    terminal_failed_jobs = 0
+    retryable_terminal_failed_jobs = 0
+    current_output_claims: dict[
+        str, tuple[tuple[datetime, str], str, int, str, str]
+    ] = {}
+
+    for job in jobs:
+        job_id = str(job["job_id"])
+        try:
+            history = _read_run_history(root, job_id)
+        except (OSError, ValueError, json.JSONDecodeError):
+            problems.append(
+                {
+                    "code": "invalid_run_history",
+                    "job_id": job_id,
+                    "action": "repair_production_metadata",
+                }
+            )
+            continue
+        attempts_total += len(history)
+        completed = [run for run in history if run["status"] != "running"]
+        running = [run for run in history if run["status"] == "running"]
+        current = [
+            run for run in history if run["fingerprint"] == job["fingerprint"]
+        ]
+        current_completed = [run for run in current if run["status"] != "running"]
+        current_running = [run for run in current if run["status"] == "running"]
+        attempts_succeeded += sum(run["status"] == "succeeded" for run in completed)
+        attempts_failed += sum(run["status"] == "failed" for run in completed)
+        attempts_running += len(running)
+        attempts_superseded += len(history) - len(current)
+        for run in current:
+            if run.get("status") == "succeeded":
+                continue
+            handle = run.get("provider_job_id") or _read_provider_handle(
+                root, job_id, str(run["run_id"])
+            )
+            if not handle:
+                continue
+            problems.append(
+                {
+                    "code": "orphaned_provider_job",
+                    "job_id": job_id,
+                    "run_id": run["run_id"],
+                    "provider_job_id": handle,
+                    "action": "collect_before_retry",
+                    # A live attempt and a dead one both sit at `running`, so
+                    # this finding cannot tell them apart. Collecting is still
+                    # the right first move — it is the only one that costs
+                    # nothing — and it refuses safely if the attempt finished.
+                    "note": "status 为 running 时它可能仍在正常轮询；collect 会先确认再写",
+                }
+            )
+        if running:
+            running_jobs += 1
+            for run in running:
+                problems.append(
+                    {
+                        "code": "running_attempt",
+                        "job_id": job_id,
+                        "run_id": str(run["run_id"]),
+                        "action": "wait_or_investigate_running_attempt",
+                    }
+                )
+        fingerprints = [
+            str(run.get("fingerprint"))
+            for run in history
+            if isinstance(run.get("fingerprint"), str)
+        ]
+        repeated_content += len(fingerprints) - len(set(fingerprints))
+        succeeded_indexes = [
+            index
+            for index, run in enumerate(current_completed)
+            if run["status"] == "succeeded"
+        ]
+        failed_indexes = [
+            index
+            for index, run in enumerate(current_completed)
+            if run["status"] == "failed"
+        ]
+        if succeeded_indexes and failed_indexes and min(failed_indexes) < max(succeeded_indexes):
+            recovered_jobs += 1
+        if (
+            not current_running
+            and current_completed
+            and current_completed[-1]["status"] == "failed"
+        ):
+            terminal_failed_jobs += 1
+            error = current_completed[-1].get("error")
+            retryable = isinstance(error, Mapping) and error.get("retryable") is True
+            if retryable:
+                retryable_terminal_failed_jobs += 1
+            problems.append(
+                {
+                    "code": (
+                        "terminal_retryable_failure" if retryable else "terminal_failure"
+                    ),
+                    "job_id": job_id,
+                    "run_id": current_completed[-1]["run_id"],
+                    "action": (
+                        "inspect_then_reconfirm_retry" if retryable else "inspect_failure"
+                    ),
+                }
+            )
+        for run in current:
+            if run["status"] != "succeeded":
+                continue
+            try:
+                outputs = _validated_succeeded_outputs(job, run)
+            except ValueError:
+                problems.append(
+                    {
+                        "code": "invalid_succeeded_output_record",
+                        "job_id": job_id,
+                        "run_id": str(run["run_id"]),
+                        "action": "repair_production_metadata",
+                    }
+                )
+                continue
+            for output_path, digest, size in outputs:
+                run_id = str(run["run_id"])
+                claim_order = _completed_run_order(run)
+                claim = current_output_claims.get(output_path)
+                if claim is None or claim_order > claim[0]:
+                    current_output_claims[output_path] = (
+                        claim_order,
+                        digest,
+                        size,
+                        job_id,
+                        run_id,
+                    )
+
+        if not _inputs_current(root, job):
+            problems.append(
+                {
+                    "code": "job_inputs_changed",
+                    "job_id": job_id,
+                    "action": "prepare_and_confirm_again",
+                }
+            )
+
+    output_verified = 0
+    output_missing = 0
+    output_modified = 0
+    for relative, (
+        _claim_order,
+        expected_digest,
+        expected_size,
+        job_id,
+        run_id,
+    ) in sorted(current_output_claims.items()):
+        try:
+            safe_relative = _relative_path(relative, output=True)
+            digest, size = _hash_production_output(root, safe_relative)
+        except (FileNotFoundError, OSError, ValueError):
+            output_missing += 1
+            problems.append(
+                {
+                    "code": "output_missing_or_unsafe",
+                    "job_id": job_id,
+                    "run_id": run_id,
+                    "path": relative,
+                    "action": "restore_or_reproduce_output",
+                }
+            )
+            continue
+        if digest != expected_digest or size != expected_size:
+            output_modified += 1
+            problems.append(
+                {
+                    "code": "output_digest_mismatch",
+                    "job_id": job_id,
+                    "run_id": run_id,
+                    "path": relative,
+                    "action": "review_current_bytes_or_reproduce_output",
+                }
+            )
+        else:
+            output_verified += 1
+
+    return {
+        "status": "attention" if problems else "pass",
+        "scope": "operational_evidence_only",
+        "quality_verdict": "not_assessed",
+        "jobs": {
+            "total": len(jobs),
+            "running": running_jobs,
+            "recovered": recovered_jobs,
+            "terminal_failed": terminal_failed_jobs,
+            "retryable_terminal_failed": retryable_terminal_failed_jobs,
+        },
+        "attempts": {
+            "total": attempts_total,
+            "succeeded": attempts_succeeded,
+            "failed": attempts_failed,
+            "running": attempts_running,
+            "superseded": attempts_superseded,
+            "repeated_content": repeated_content,
+        },
+        "outputs": {
+            "claimed_current": len(current_output_claims),
+            "verified": output_verified,
+            "missing": output_missing,
+            "modified": output_modified,
+        },
+        "problems": problems,
+    }
+
+
+def build_parser() -> argparse.ArgumentParser:
+    parser = argparse.ArgumentParser(description="Run confirmed short-drama media jobs.")
+    commands = parser.add_subparsers(dest="command", required=True)
+    prepare = commands.add_parser("prepare", help="Validate and preview a media job.")
+    prepare.add_argument("project")
+    prepare.add_argument("--job", required=True)
+    confirm = commands.add_parser("confirm", help="Confirm the exact prepared job.")
+    confirm.add_argument("project")
+    confirm.add_argument("--job-id", required=True)
+    confirm.add_argument("--confirmation", required=True)
+    run = commands.add_parser("run", help="Execute a confirmed job through an adapter.")
+    run.add_argument("project")
+    run.add_argument("--job-id", required=True)
+    run.add_argument("--adapter-config", required=True)
+    collect = commands.add_parser(
+        "collect",
+        help="Fetch the result of an already-submitted provider task without paying again.",
+    )
+    collect.add_argument("project")
+    collect.add_argument("--job-id", required=True)
+    collect.add_argument("--adapter-config", required=True)
+    status = commands.add_parser("status", help="Show one media job state.")
+    status.add_argument("project")
+    status.add_argument("--job-id", required=True)
+    audit = commands.add_parser(
+        "audit", help="Reconcile attempt history and current output bytes."
+    )
+    audit.add_argument("project")
+    return parser
+
+
+def main(argv: list[str] | None = None) -> int:
+    args = build_parser().parse_args(argv)
+    try:
+        if args.command == "prepare":
+            result = prepare_job(Path(args.project), Path(args.job))
+        elif args.command == "confirm":
+            result = confirm_job(
+                Path(args.project), job_id=args.job_id, confirmation=args.confirmation
+            )
+        elif args.command == "run":
+            result = run_job(
+                Path(args.project),
+                job_id=args.job_id,
+                adapter_config=Path(args.adapter_config),
+            )
+        elif args.command == "collect":
+            result = collect_job(
+                Path(args.project),
+                job_id=args.job_id,
+                adapter_config=Path(args.adapter_config),
+            )
+        elif args.command == "status":
+            result = job_status(Path(args.project), job_id=args.job_id)
+        else:
+            result = audit_project(Path(args.project))
+        print(json.dumps(result, ensure_ascii=True, sort_keys=True))
+        return 0
+    except Exception as exc:
+        print(str(exc), file=sys.stderr)
+        return 2
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())