claude-dev-env 2.4.0 → 2.7.0

package/scripts/codex_compat_materializer.py

@@ -0,0 +1,1087 @@
+"""Deterministic, additive Claude-agent to Codex-agent materialization."""
+
+from __future__ import annotations
+
+import argparse
+import ast
+import hashlib
+import json
+import os
+import re
+import stat
+import tempfile
+import tomllib
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import Callable, Iterable
+
+
+ManagedContent = str | bytes
+
+path_separator = "/"
+toml_suffix = ".toml"
+reparse_point_attribute_name = "FILE_ATTRIBUTE_REPARSE_POINT"
+manifest_indentation_width = 2
+publish_plan_max_positional_arguments = 3
+publish_plan_failure_injector_position = 2
+frontmatter_required_fields = ("name", "description")
+frontmatter_unsupported_fields = ("tools", "model", "color")
+full_prune_opt_in_flag = "--allow-prune-all"
+unreadable_source_root_message = (
+    "source root is missing or is not a directory, so nothing was planned or changed; "
+    "check the source root path and re-run"
+)
+reparse_source_root_message = "source root is a reparse point, so nothing was planned or changed"
+full_prune_refusal_message = (
+    "refusing to delete every managed file: the plan is empty while the manifest still records "
+    "{count} of them, so the target root was left untouched; re-run with " + full_prune_opt_in_flag
+    + " to remove them"
+)
+unmanaged_target_message = (
+    "unmanaged file at planned target {path}: the compatibility manifest does not record it, so it "
+    "may be yours and it was not overwritten. Review it, then move or delete {path} inside the "
+    "target root and re-run. An interrupted run leaves a file whose bytes already match the plan, "
+    "and such a file is adopted automatically"
+)
+
+
+class MaterializerError(ValueError):
+    """Raised when a materialization request cannot be safely planned."""
+
+
+class ArgumentParserError(ValueError):
+    """Raised when command-line arguments cannot be parsed."""
+
+
+class MaterializerArgumentParser(argparse.ArgumentParser):
+    """Parse materializer arguments while keeping errors in the JSON contract."""
+
+    def error(self, message: str) -> None:
+        """Raise a reportable parser error instead of writing process output."""
+        raise ArgumentParserError(message)
+
+
+report_categories = (
+    "written", "unchanged", "adopted", "unmanaged_collision", "modified_managed",
+    "stale_managed", "deleted", "unsupported", "conflicted", "errors",
+)
+report_categories_public_name = "REPORT_CATEGORIES"
+frontmatter_allowed_fields = {"name", "description", "tools", "model", "color"}
+line_separator = "\n"
+comma_separator = ", "
+
+
+def __getattr__(name: str) -> tuple[str, ...]:
+    if name == report_categories_public_name:
+        return report_categories
+    raise AttributeError(name)
+
+
+@dataclass(frozen=True)
+class MaterializerConfig:
+    source_root: Path
+    target_root: Path
+    manifest_path: Path | None = None
+    should_apply: bool = False
+    should_allow_full_prune: bool = False
+
+    def __post_init__(self) -> None:
+        source = self.source_root.expanduser().resolve()
+        target = self.target_root.expanduser().resolve()
+        if source == target or source in target.parents or target in source.parents:
+            raise MaterializerError("source and target roots must not overlap")
+        object.__setattr__(self, "source_root", source)
+        object.__setattr__(self, "target_root", target)
+        manifest = (self.manifest_path or target / ".codex-compat-manifest.json").expanduser().resolve()
+        if target not in manifest.parents:
+            raise MaterializerError("manifest must be inside the target root")
+        object.__setattr__(self, "manifest_path", manifest)
+
+
+@dataclass(frozen=True)
+class ClaudeAgent:
+    source_path: Path
+    relative_source: str
+    name: str
+    description: str
+    tools: tuple[str, ...] = ()
+    model: str | None = None
+    color: str | None = None
+    unsupported: tuple[str, ...] = ()
+
+
+@dataclass(frozen=True)
+class ManifestRecord:
+    """One compatibility-manifest entry, as the publication logic reads it.
+
+    A field that failed validation is carried as ``None``, so ownership questions
+    read as attribute checks rather than repeated isinstance guards::
+
+        {"hash": "ab12..", "ownership": "codex-compat"} -> ok:   owned, refreshable
+        {"hash": "ab12.."}                              -> flag: unowned, preserved
+
+    The manifest also stores ``source`` and ``marker`` for inspection. No code path
+    reads either, so neither appears here.
+
+    Args:
+        content_hash: Hash the tool recorded when it last published the file.
+        ownership: Ownership marker the tool recorded alongside that hash.
+    """
+
+    content_hash: str | None
+    ownership: str | None
+
+    @property
+    def is_owned_by_tool(self) -> bool:
+        """Report whether this entry carries both fields that prove tool ownership."""
+        return self.content_hash is not None and self.ownership is not None
+
+
+ManifestRecordByPath = dict[str, ManifestRecord | None]
+
+
+@dataclass(frozen=True)
+class PlannedFile:
+    source_identity: str
+    target_relative_path: str
+    content: ManagedContent
+    content_hash: str
+    ownership: str = "codex-compat"
+    generated_marker: str = "codex-compat-generated-v1"
+    action: str = "write"
+
+
+@dataclass
+class MaterializationReport:
+    written: int = 0
+    unchanged: int = 0
+    adopted: int = 0
+    unmanaged_collision: int = 0
+    modified_managed: int = 0
+    stale_managed: int = 0
+    deleted: int = 0
+    unsupported: int = 0
+    conflicted: int = 0
+    errors: int = 0
+    error_details: list[str] = field(default_factory=list)
+    details: dict[str, list[str]] = field(default_factory=lambda: {each_category: [] for each_category in report_categories})
+    planned_files: list[PlannedFile] = field(default_factory=list)
+    is_generation_incomplete: bool = False
+    is_reconciliation_required: bool = False
+
+    @property
+    def incomplete_generation(self) -> bool:
+        return self.is_generation_incomplete
+
+    @incomplete_generation.setter
+    def incomplete_generation(self, is_incomplete: bool) -> None:
+        self.is_generation_incomplete = is_incomplete
+
+    @property
+    def reconcile_required(self) -> bool:
+        return self.is_reconciliation_required
+
+    @reconcile_required.setter
+    def reconcile_required(self, is_required: bool) -> None:
+        self.is_reconciliation_required = is_required
+
+    @property
+    def preserved(self) -> int:
+        """Return the legacy name for unchanged planned files."""
+        return self.unchanged + self.modified_managed + self.stale_managed
+
+    def add_detail(self, category: str, relative_path: str) -> None:
+        self.details[category].append(relative_path)
+
+    def add_error(self, message: str) -> None:
+        self.errors += 1
+        self.error_details.append(message)
+        self.add_detail("errors", message)
+
+
+def _normalize_relative_path(path: str) -> str:
+    canonical_path = path.replace("\\", path_separator)
+    if not canonical_path or canonical_path.startswith(("/", "//")) or re.match(r"^[A-Za-z]:($|/)", canonical_path):
+        raise MaterializerError("rooted path is not allowed")
+    path_parts = canonical_path.split(path_separator)
+    if any(each_part in ("", ".", "..") for each_part in path_parts) or canonical_path != path_separator.join(path_parts):
+        raise MaterializerError("path is not normalized or contains traversal")
+    return canonical_path
+
+
+def _casefold_normalized_path(path: Path) -> str:
+    return path.as_posix().casefold()
+
+
+def _validate_source_identity(relative_source: str) -> str:
+    canonical_source = _normalize_relative_path(relative_source)
+    if canonical_source.startswith(".") or canonical_source.casefold().startswith(("private/", "private\\")):
+        raise MaterializerError("private source identity is not allowed")
+    return canonical_source
+
+
+def _is_reparse_point(path: Path) -> bool:
+    if path.is_symlink():
+        return True
+    try:
+        attributes = path.stat(follow_symlinks=False).st_file_attributes
+    except (AttributeError, OSError):
+        return False
+    return bool(attributes & getattr(stat, reparse_point_attribute_name, 0))
+
+
+def _validate_containment(root: Path, candidate: Path) -> Path:
+    resolved = candidate.resolve()
+    if root not in (resolved, *resolved.parents):
+        raise MaterializerError("path resolves outside target root")
+    return resolved
+
+
+def validate_target_path(target_root: Path, relative_path: str) -> Path:
+    """Resolve a relative target path inside a safe target root.
+
+    Args:
+        target_root: Directory that must contain the resolved path.
+        relative_path: Normalized path relative to ``target_root``.
+
+    Returns:
+        The resolved target path.
+
+    Raises:
+        MaterializerError: If the path is rooted, unsafe, or crosses a reparse point.
+    """
+    canonical_path = _normalize_relative_path(relative_path)
+    root = target_root.resolve()
+    candidate = root.joinpath(*canonical_path.split(path_separator))
+    for each_parent in (root, *candidate.parents):
+        if each_parent.exists() and _is_reparse_point(each_parent):
+            raise MaterializerError("target path crosses a reparse point")
+    return _validate_containment(root, candidate)
+
+
+def _parse_frontmatter_scalar(serialized_field_text: str) -> str | tuple[str, ...] | None:
+    normalized_field_text = serialized_field_text.strip()
+    if not normalized_field_text:
+        return ""
+    if normalized_field_text.startswith("["):
+        return _parse_frontmatter_list(normalized_field_text)
+    try:
+        parsed = ast.literal_eval(normalized_field_text)
+    except (SyntaxError, ValueError):
+        if normalized_field_text[:1] in {'"', "'"} or normalized_field_text[:1] == "[":
+            raise MaterializerError("malformed frontmatter value")
+        return normalized_field_text
+    if isinstance(parsed, str):
+        return parsed
+    if isinstance(parsed, list) and all(isinstance(each_entry, str) for each_entry in parsed):
+        return tuple(parsed)
+    raise MaterializerError("frontmatter value must be a string or string list")
+
+
+def _parse_frontmatter_list(serialized_list: str) -> tuple[str, ...]:
+    if not serialized_list.endswith("]"):
+        raise MaterializerError("malformed frontmatter value")
+    all_entries: list[str] = []
+    entry_start = 1
+    quote: str | None = None
+    has_escape_pending = False
+    for each_index, each_character in enumerate(serialized_list[1:-1], 1):
+        if has_escape_pending:
+            has_escape_pending = False
+            continue
+        if quote == '"' and each_character == "\\":
+            has_escape_pending = True
+            continue
+        if each_character in {'"', "'"}:
+            if quote is None:
+                quote = each_character
+            elif quote == each_character:
+                quote = None
+            continue
+        if each_character == "," and quote is None:
+            all_entries.append(_parse_frontmatter_list_entry(serialized_list[entry_start:each_index]))
+            entry_start = each_index + 1
+    if quote is not None or has_escape_pending:
+        raise MaterializerError("malformed frontmatter value")
+    all_entries.append(_parse_frontmatter_list_entry(serialized_list[entry_start:-1]))
+    return tuple(all_entries) if all_entries != [""] else ()
+
+
+def _parse_frontmatter_list_entry(raw_entry: str) -> str:
+    entry = raw_entry.strip()
+    if not entry:
+        raise MaterializerError("malformed frontmatter value")
+    try:
+        parsed = ast.literal_eval(entry)
+    except (SyntaxError, ValueError):
+        if entry[0] in {'"', "'"}:
+            raise MaterializerError("malformed frontmatter value")
+        if any(character in entry for character in "[]{}:"):
+            raise MaterializerError("malformed frontmatter value")
+        return entry
+    if not isinstance(parsed, str):
+        raise MaterializerError("frontmatter list entries must be strings")
+    return parsed
+
+
+def parse_frontmatter(source_path: Path, source_text: str, relative_source: str) -> ClaudeAgent:
+    """Parse one Claude agent's frontmatter.
+
+    Args:
+        source_path: Path used in validation errors.
+        source_text: Markdown source containing the frontmatter block.
+        relative_source: Safe source identity recorded in the manifest.
+
+    Returns:
+        The parsed Claude agent.
+
+    Raises:
+        MaterializerError: If frontmatter syntax or required fields are invalid.
+    """
+    source_identity = _validate_source_identity(relative_source)
+    lines = source_text.splitlines()
+    if len(lines) < 3 or lines[0].strip() != "---":
+        raise MaterializerError(f"malformed frontmatter: {source_path}")
+    delimiters = [each_index for each_index, line in enumerate(lines[1:], 1) if line.strip() == "---"]
+    if len(delimiters) != 1:
+        raise MaterializerError(f"malformed frontmatter delimiters: {source_path}")
+    all_fields: dict[str, str | tuple[str, ...] | None] = {}
+    for each_line in lines[1 : delimiters[0]]:
+        if not each_line.strip() or ":" not in each_line:
+            raise MaterializerError(f"malformed frontmatter: {source_path}")
+        key, serialized_field_text = each_line.split(":", 1)
+        key = key.strip()
+        if key in all_fields or not re.fullmatch(r"[A-Za-z][A-Za-z0-9_-]*", key):
+            raise MaterializerError(f"malformed frontmatter key: {source_path}")
+        all_fields[key] = _parse_frontmatter_scalar(serialized_field_text)
+    unknown = tuple(sorted(each_key for each_key in all_fields if each_key not in frontmatter_allowed_fields))
+    if unknown:
+        raise MaterializerError(f"unknown frontmatter keys: {source_path}")
+    unsupported = tuple(sorted(each_key for each_key in all_fields if each_key in frontmatter_unsupported_fields))
+    if any(not isinstance(all_fields.get(each_key), str) or not all_fields[each_key] for each_key in frontmatter_required_fields):
+        raise MaterializerError(f"name and description are required: {source_path}")
+    tools = all_fields.get("tools", ())
+    if isinstance(tools, str):
+        tools = (tools,)
+    if not isinstance(tools, tuple):
+        raise MaterializerError(f"tools must be a list: {source_path}")
+    return ClaudeAgent(source_path, source_identity, all_fields["name"], all_fields["description"], tools, all_fields.get("model"), all_fields.get("color"), unsupported)
+
+
+def convert_agent(agent: ClaudeAgent) -> str:
+    """Convert a Claude agent to validated Codex TOML content.
+
+    Args:
+        agent: Parsed Claude agent to convert.
+
+    Returns:
+        UTF-8 text containing the generated TOML document.
+
+    Raises:
+        MaterializerError: If the generated TOML cannot be validated.
+    """
+    toml_string = lambda text: json.dumps(text, ensure_ascii=False)
+    content = line_separator.join((f"name = {toml_string(agent.name)}", f"description = {toml_string(agent.description)}", f"developer_instructions = {toml_string('Claude tools: ' + comma_separator.join(agent.tools))}")) + line_separator
+    tomllib.loads(content)
+    return content
+
+
+def discover_agents(config: MaterializerConfig) -> list[ClaudeAgent]:
+    """Discover and parse Markdown agents below the source root.
+
+    Args:
+        config: Materializer paths and application settings.
+
+    An unreachable source root is an error rather than an empty discovery, so a
+    mistyped path or an offline share never reads as "this tree holds no agents"::
+
+        source root missing   -> flag: MaterializerError, nothing planned
+        source root empty     -> ok:   [] , and publication asks for prune consent
+
+    Returns:
+        Agents discovered in deterministic relative-path order.
+
+    Raises:
+        MaterializerError: If the source root is unreachable, or a source path is unsafe or malformed.
+    """
+    if not config.source_root.is_dir():
+        raise MaterializerError(f"{unreadable_source_root_message}: {config.source_root}")
+    if _is_reparse_point(config.source_root):
+        raise MaterializerError(f"{reparse_source_root_message}: {config.source_root}")
+    all_agents: list[ClaudeAgent] = []
+    for each_path in sorted(config.source_root.rglob("*.md"), key=lambda path: path.as_posix().casefold()):
+        if _is_reparse_point(each_path):
+            raise MaterializerError(f"source reparse point is not allowed: {each_path}")
+        relative_source = each_path.relative_to(config.source_root).as_posix()
+        _validate_containment(config.source_root, each_path)
+        all_agents.append(parse_frontmatter(each_path, each_path.read_text(encoding="utf-8"), relative_source))
+    return all_agents
+
+
+def _case_fold_collision_error(target_relative_path: str) -> MaterializerError:
+    """Build the error for two target names that differ only by letter case."""
+    return MaterializerError(f"case-fold collision: {target_relative_path}")
+
+
+def _validate_orphan_target_is_adoptable(
+    config: MaterializerConfig,
+    existing_path: Path | None,
+    target_relative_path: str,
+    content: ManagedContent,
+) -> None:
+    """Allow an unrecorded target file only when its bytes already match the plan.
+
+    A run interrupted between the file replacement and the manifest save leaves a
+    file the manifest does not record. Byte-identical content proves the tool wrote
+    it, so the next run adopts it instead of stopping forever::
+
+        Nova.toml bytes == planned bytes  -> ok:   adopted, publication continues
+        Nova.toml bytes != planned bytes  -> flag: MaterializerError naming the remedy
+
+    Args:
+        config: Materializer paths and application settings.
+        existing_path: Target-root path already holding the planned name, or ``None``.
+        target_relative_path: Normalized relative path the plan publishes.
+        content: Content the plan would publish at that path.
+
+    Raises:
+        MaterializerError: If the existing file differs from the plan or differs only in case.
+    """
+    if existing_path is None:
+        return
+    if existing_path.relative_to(config.target_root).as_posix() != target_relative_path:
+        raise _case_fold_collision_error(target_relative_path)
+    if not existing_path.is_file() or existing_path.read_bytes() != content_to_bytes(content):
+        raise MaterializerError(unmanaged_target_message.format(path=target_relative_path))
+
+
+def _build_plan(config: MaterializerConfig, all_agents: Iterable[ClaudeAgent]) -> tuple[list[PlannedFile], MaterializationReport]:
+    """Build planned agent publications and their report.
+
+    Args:
+        config: Materializer paths and application settings.
+        all_agents: Optional parsed agents used by callers that inject discovery results.
+
+    Returns:
+        Planned files and the report describing unsupported agent fields.
+
+    Raises:
+        MaterializerError: If a source or target path collides or is unsafe.
+    """
+    report = MaterializationReport()
+    planned: list[PlannedFile] = []
+    target_by_name: dict[str, str] = {}
+    previous_records = _manifest_record_by_path(load_manifest(config.manifest_path))
+    existing_by_name = {
+        each_path.relative_to(config.target_root).as_posix().casefold(): each_path
+        for each_path in config.target_root.rglob("*")
+        if not _is_known_managed_path(config.target_root, each_path, previous_records)
+    } if config.target_root.exists() else {}
+    for each_agent in list(all_agents):
+        source_identity = _validate_source_identity(each_agent.relative_source)
+        target_relative_path = _normalize_relative_path(each_agent.name + toml_suffix)
+        folded_path = target_relative_path.casefold()
+        if folded_path in target_by_name:
+            raise _case_fold_collision_error(target_relative_path)
+        content = convert_agent(each_agent)
+        _validate_orphan_target_is_adoptable(config, existing_by_name.get(folded_path), target_relative_path, content)
+        target_by_name[folded_path] = target_relative_path
+        target_path = validate_target_path(config.target_root, target_relative_path)
+        if _casefold_normalized_path(target_path) == _casefold_normalized_path(config.manifest_path):
+            raise MaterializerError("planned target collides with compatibility manifest")
+        planned.append(PlannedFile(source_identity, target_relative_path, content, hash_content(content)))
+        report.unsupported += len(each_agent.unsupported)
+        report.details["unsupported"].extend(f"{source_identity}:{each_key}" for each_key in each_agent.unsupported)
+    report.planned_files = planned
+    return planned, report
+
+
+def build_plan(config: MaterializerConfig, *all_arguments: object, **all_keywords: object) -> tuple[list[PlannedFile], MaterializationReport]:
+    """Build a plan while preserving the legacy optional-agent call form.
+
+    Args:
+        config: Materializer paths and application settings.
+        all_arguments: Optional positional discovered-agent iterable.
+        all_keywords: Optional ``all_agents`` keyword argument.
+
+    Returns:
+        Planned files and the report describing unsupported agent fields.
+
+    Raises:
+        TypeError: If more than one agent iterable or an unknown keyword is supplied.
+        MaterializerError: If a source or target path collides or is unsafe.
+    """
+    supplied_agents = all_keywords.pop("all_agents", None)
+    if all_keywords or len(all_arguments) > 1:
+        raise TypeError("build_plan accepts at most one agent iterable")
+    if all_arguments:
+        if supplied_agents is not None:
+            raise TypeError("build_plan received duplicate all_agents")
+        supplied_agents = all_arguments[0]
+    discovered_agents = discover_agents(config) if supplied_agents is None else supplied_agents
+    return _build_plan(config, discovered_agents)
+
+
+def content_to_bytes(content: ManagedContent) -> bytes:
+    """Encode managed text while preserving already-encoded bytes.
+
+    Args:
+        content: Text or bytes intended for publication.
+
+    Returns:
+        UTF-8 bytes for the managed content.
+    """
+    if isinstance(content, bytes):
+        return content
+    return content.encode("utf-8")
+
+
+def hash_content(content: ManagedContent) -> str:
+    return hashlib.sha256(content_to_bytes(content)).hexdigest()
+
+
+def _atomic_write(
+    target_path: Path,
+    content: ManagedContent,
+    failure_injector: Callable[[str], None] | None,
+) -> None:
+    """Write managed content through a temporary file and replace atomically.
+
+    Args:
+        target_path: Destination path for the replacement.
+        content: Text or bytes to write.
+        failure_injector: Test seam invoked before replacement, or ``None``.
+
+    Raises:
+        OSError: If the temporary file or replacement cannot be written.
+        RuntimeError: If the failure injector requests a failed publication.
+        ValueError: If the content cannot be encoded.
+
+    The optional injector is a test seam. When supplied, it runs after the
+    temporary file is durable and before the destination replacement.
+
+    Raises:
+        OSError: If the temporary file or replacement cannot be written.
+        RuntimeError: If the failure injector requests a failed publication.
+        ValueError: If the content cannot be encoded.
+    """
+    target_path.parent.mkdir(parents=True, exist_ok=True)
+    descriptor, temporary_name = tempfile.mkstemp(prefix=f".{target_path.name}.", dir=target_path.parent)
+    try:
+        with os.fdopen(descriptor, "wb") as temporary_file:
+            temporary_file.write(content_to_bytes(content))
+            temporary_file.flush()
+            os.fsync(temporary_file.fileno())
+        if failure_injector is not None:
+            failure_injector(str(target_path))
+        os.replace(temporary_name, target_path)
+    except (OSError, RuntimeError, ValueError):
+        Path(temporary_name).unlink(missing_ok=True)
+        raise
+
+
+def atomic_write(target_path: Path, content: ManagedContent, *all_arguments: object, **all_keywords: object) -> None:
+    """Preserve the legacy atomic-write call form while using an explicit core.
+
+    Args:
+        target_path: Destination path for the replacement.
+        content: Text or bytes to write.
+        all_arguments: Legacy positional failure-injector argument.
+        all_keywords: Legacy keyword failure-injector argument.
+
+    Raises:
+        TypeError: If more than one failure injector is supplied or it is not callable.
+        OSError, RuntimeError, ValueError: If the atomic replacement cannot be completed.
+    """
+    failure_injector = all_keywords.pop("failure_injector", None)
+    if all_keywords or len(all_arguments) > 1:
+        raise TypeError("atomic_write accepts at most one failure injector")
+    if all_arguments:
+        failure_injector = all_arguments[0]
+    if failure_injector is not None and not callable(failure_injector):
+        raise TypeError("failure injector must be callable")
+    _atomic_write(target_path, content, failure_injector)
+
+
+def load_manifest(manifest_path: Path) -> dict[str, object]:
+    """Load a compatibility manifest or return its empty schema.
+
+    Args:
+        manifest_path: Manifest file to read.
+
+    Returns:
+        A validated manifest mapping.
+
+    Raises:
+        MaterializerError: If the manifest has an unsupported shape.
+    """
+    if not manifest_path.exists():
+        return {"version": 1, "files": {}}
+    parsed = json.loads(manifest_path.read_text(encoding="utf-8"))
+    if not isinstance(parsed, dict) or parsed.get("version") != 1 or not isinstance(parsed.get("files"), dict):
+        raise MaterializerError("invalid compatibility manifest")
+    return parsed
+
+
+def save_manifest(manifest_path: Path, all_manifest: dict[str, object], failure_injector: Callable[[str], None] | None = None) -> None:
+    """Atomically save the compatibility manifest last.
+
+    Args:
+        manifest_path: Destination manifest path.
+        all_manifest: Manifest mapping to serialize.
+        failure_injector: Optional test seam invoked before replacement.
+
+    Raises:
+        OSError: If the temporary file or replacement cannot be written.
+        RuntimeError: If the failure injector requests a failed publication.
+    """
+    content = json.dumps(all_manifest, ensure_ascii=False, sort_keys=True, indent=manifest_indentation_width) + "\n"
+    manifest_path.parent.mkdir(parents=True, exist_ok=True)
+    descriptor, temporary_name = tempfile.mkstemp(prefix=f".{manifest_path.name}.", dir=manifest_path.parent)
+    try:
+        with os.fdopen(descriptor, "w", encoding="utf-8", newline="\n") as temporary_file:
+            temporary_file.write(content)
+            temporary_file.flush()
+            os.fsync(temporary_file.fileno())
+        if failure_injector is not None:
+            failure_injector("manifest_before_replace")
+        os.replace(temporary_name, manifest_path)
+    except (OSError, RuntimeError, ValueError):
+        Path(temporary_name).unlink(missing_ok=True)
+        raise
+
+
+def _parse_manifest_record(raw_record: object) -> ManifestRecord | None:
+    """Read one manifest entry into a record, or ``None`` when it is not an object."""
+    if not isinstance(raw_record, dict):
+        return None
+    expected_hash = raw_record.get("hash")
+    ownership = raw_record.get("ownership")
+    return ManifestRecord(
+        expected_hash if isinstance(expected_hash, str) else None,
+        ownership if isinstance(ownership, str) else None,
+    )
+
+
+def _manifest_record_by_path(all_previous_manifest: dict[str, object]) -> ManifestRecordByPath:
+    """Parse the manifest's file entries once, so untyped JSON stops here.
+
+    Every path stays in the mapping even when its entry is unreadable, because the
+    case-fold checks count the names the manifest claims::
+
+        {"Luna.toml": {"hash": "ab12..", "ownership": "codex-compat"}} -> ok:   record
+        {"Luna.toml": 7}                                              -> flag: None
+
+    Args:
+        all_previous_manifest: Validated manifest mapping from ``load_manifest``.
+
+    Returns:
+        Each manifest path mapped to its record, or to ``None`` when unreadable.
+
+    Raises:
+        MaterializerError: If the manifest's file entries are not a mapping.
+    """
+    records = all_previous_manifest["files"]
+    if not isinstance(records, dict):
+        raise MaterializerError("invalid compatibility manifest files")
+    return {each_path: _parse_manifest_record(each_record) for each_path, each_record in records.items()}
+
+
+def _find_manifest_record(all_previous_records: ManifestRecordByPath, target_relative_path: str) -> ManifestRecord | None:
+    for each_path, each_record in all_previous_records.items():
+        if each_path.casefold() == target_relative_path.casefold():
+            return each_record
+    return None
+
+
+def _is_known_managed_path(target_root: Path, target_path: Path, all_previous_records: ManifestRecordByPath) -> bool:
+    relative_path = target_path.relative_to(target_root).as_posix()
+    previous_record = _find_manifest_record(all_previous_records, relative_path)
+    return previous_record is not None and previous_record.is_owned_by_tool
+
+
+def _is_pristine_managed(previous_record: ManifestRecord | None, current_bytes: bytes) -> bool:
+    """Report whether on-disk bytes are exactly what the tool last published there."""
+    if previous_record is None or not previous_record.is_owned_by_tool:
+        return False
+    return hash_content(current_bytes) == previous_record.content_hash
+
+
+def _record_target_conflict(report: MaterializationReport, target_relative_path: str, previous_record: ManifestRecord | None) -> None:
+    report.conflicted += 1
+    report.add_detail("conflicted", target_relative_path)
+    if previous_record is None:
+        report.unmanaged_collision += 1
+        report.add_detail("unmanaged_collision", target_relative_path)
+        return
+    report.modified_managed += 1
+    report.add_detail("modified_managed", target_relative_path)
+
+
+def _record_target_state(config: MaterializerConfig, planned_file: PlannedFile, all_previous_records: ManifestRecordByPath, report: MaterializationReport) -> tuple[Path, bytes | None, bool]:
+    """Classify one planned target and say whether publication may overwrite it.
+
+    The manifest hash decides ownership, so a file the tool wrote is refreshed and a
+    file the user edited is preserved::
+
+        target absent                    -> ok:   publish
+        on-disk bytes == planned bytes   -> ok:   unchanged, no write
+        on-disk hash == manifest hash    -> ok:   publish, the tool owns these bytes
+        on-disk hash != manifest hash    -> flag: conflicted, preserved untouched
+
+    Args:
+        config: Materializer paths and application settings.
+        planned_file: Planned publication for this target.
+        all_previous_records: Manifest records from the last successful run.
+        report: Report object to update in place.
+
+    Returns:
+        The resolved target path, its current bytes when it exists, and whether to publish.
+    """
+    target_path = validate_target_path(config.target_root, planned_file.target_relative_path)
+    if not target_path.exists():
+        return target_path, None, True
+    current_bytes = target_path.read_bytes()
+    previous_record = _find_manifest_record(all_previous_records, planned_file.target_relative_path)
+    if current_bytes == content_to_bytes(planned_file.content):
+        _record_matching_target(report, planned_file.target_relative_path, previous_record)
+        return target_path, current_bytes, False
+    if _is_pristine_managed(previous_record, current_bytes):
+        return target_path, current_bytes, True
+    _record_target_conflict(report, planned_file.target_relative_path, previous_record)
+    return target_path, current_bytes, False
+
+
+def _record_matching_target(report: MaterializationReport, target_relative_path: str, previous_record: ManifestRecord | None) -> None:
+    report.unchanged += 1
+    report.add_detail("unchanged", target_relative_path)
+    if previous_record is not None:
+        return
+    report.adopted += 1
+    report.add_detail("adopted", target_relative_path)
+
+
+def _remove_stale_files(config: MaterializerConfig, all_previous_records: ManifestRecordByPath, all_planned_files: list[PlannedFile], report: MaterializationReport, all_backups: dict[Path, bytes | None]) -> None:
+    current_names = {each_planned_file.target_relative_path.casefold() for each_planned_file in all_planned_files}
+    for each_relative_path, each_record in sorted(all_previous_records.items(), key=lambda pair: pair[0].casefold()):
+        if each_relative_path.casefold() in current_names or each_record is None:
+            continue
+        target_path = validate_target_path(config.target_root, each_relative_path)
+        if not target_path.exists():
+            report.add_error(f"missing managed path: {each_relative_path}")
+            continue
+        current_bytes = target_path.read_bytes()
+        expected_hash = each_record.content_hash
+        if expected_hash is not None and hash_content(current_bytes) == expected_hash:
+            all_backups[target_path] = current_bytes
+            target_path.unlink()
+            report.deleted += 1
+            report.add_detail("deleted", each_relative_path)
+            continue
+        if expected_hash is not None:
+            report.modified_managed += 1
+            report.add_detail("modified_managed", each_relative_path)
+
+
+def _sort_report_details(report: MaterializationReport) -> None:
+    for each_category in report_categories:
+        if each_category == "errors":
+            continue
+        report.details[each_category].sort(key=str.casefold)
+
+
+def _validate_planned_targets(
+    all_planned_files: list[PlannedFile],
+    all_previous_records: ManifestRecordByPath,
+) -> None:
+    folded_manifest_names = {each_key.casefold() for each_key in all_previous_records}
+    if len(folded_manifest_names) != len(all_previous_records):
+        raise MaterializerError("case-fold collision in compatibility manifest")
+    planned_names: set[str] = set()
+    for each_file in all_planned_files:
+        folded_target = each_file.target_relative_path.casefold()
+        if folded_target in planned_names:
+            raise _case_fold_collision_error(each_file.target_relative_path)
+        planned_names.add(folded_target)
+        has_manifest_owner = any(
+            each_path.casefold() == folded_target
+            and each_record is not None
+            and each_record.ownership is not None
+            for each_path, each_record in all_previous_records.items()
+        )
+        if folded_target in folded_manifest_names and not has_manifest_owner:
+            raise MaterializerError("case-fold collision in compatibility manifest")
+
+
+def _publish_planned_targets(
+    config: MaterializerConfig,
+    all_planned_files: list[PlannedFile],
+    all_previous_records: ManifestRecordByPath,
+    report: MaterializationReport,
+    all_backups: dict[Path, bytes | None],
+    failure_injector: Callable[[str], None] | None,
+) -> None:
+    for each_planned_file in all_planned_files:
+        target_path, current_bytes, is_publishable = _record_target_state(config, each_planned_file, all_previous_records, report)
+        if _casefold_normalized_path(target_path) == _casefold_normalized_path(config.manifest_path):
+            raise MaterializerError("planned target collides with compatibility manifest")
+        if not is_publishable:
+            continue
+        all_backups[target_path] = current_bytes
+        atomic_write(target_path, each_planned_file.content, failure_injector)
+        report.written += 1
+        report.add_detail("written", each_planned_file.target_relative_path)
+
+
+def _build_manifest(all_planned_files: list[PlannedFile]) -> dict[str, object]:
+    return {
+        "version": 1,
+        "files": {
+            each_file.target_relative_path: {
+                "source": each_file.source_identity,
+                "hash": each_file.content_hash,
+                "ownership": each_file.ownership,
+                "marker": each_file.generated_marker,
+            }
+            for each_file in all_planned_files
+        },
+    }
+
+
+def _rollback_publication(
+    all_backups: dict[Path, bytes | None],
+    report: MaterializationReport,
+    initial_written: int,
+    initial_deleted: int,
+) -> None:
+    for each_target_path, each_previous_content in reversed(tuple(all_backups.items())):
+        try:
+            if each_previous_content is None:
+                each_target_path.unlink(missing_ok=True)
+            else:
+                atomic_write(each_target_path, each_previous_content)
+        except OSError:
+            report.reconcile_required = True
+            report.add_error(f"rollback failed: {each_target_path}")
+    report.incomplete_generation = True
+    report.reconcile_required = True
+    report.written = initial_written
+    report.deleted = initial_deleted
+    report.details["written"] = report.details["written"][:initial_written]
+    report.details["deleted"] = report.details["deleted"][:initial_deleted]
+    report.add_error("incomplete_generation/reconcile_required")
+    _sort_report_details(report)
+
+
+def _validate_full_prune_consent(
+    config: MaterializerConfig,
+    all_planned_files: list[PlannedFile],
+    all_previous_records: ManifestRecordByPath,
+) -> None:
+    """Require an explicit opt-in before an empty plan erases every managed file.
+
+    An empty plan means every managed file is stale, so publication would delete the
+    whole set. That is a legitimate request and also what a mistyped source root
+    produces, so the caller has to ask for it by name::
+
+        empty plan, empty manifest              -> ok:   nothing to delete
+        empty plan, managed files, opt-in given -> ok:   prune proceeds
+        empty plan, managed files, no opt-in    -> flag: MaterializerError, nothing deleted
+
+    Args:
+        config: Materializer paths and application settings.
+        all_planned_files: Files the current run would publish.
+        all_previous_records: Manifest records from the last successful run.
+
+    Raises:
+        MaterializerError: If the run would delete every managed file without the opt-in.
+    """
+    if all_planned_files or not all_previous_records or config.should_allow_full_prune:
+        return
+    raise MaterializerError(full_prune_refusal_message.format(count=len(all_previous_records)))
+
+
+def _publish_plan(
+    config: MaterializerConfig,
+    all_planned_files: Iterable[PlannedFile],
+    report: MaterializationReport,
+    failure_injector: Callable[[str], None] | None,
+) -> MaterializationReport:
+    """Publish planned files with rollback and manifest-last semantics.
+
+    Args:
+        config: Materializer paths and application settings.
+        all_planned_files: Files to publish, including generic non-TOML content.
+        report: Report object to update in place.
+        failure_injector: Test seam invoked before each replacement, or ``None``.
+
+    Returns:
+        The updated materialization report.
+
+    Raises:
+        OSError, RuntimeError, ValueError: If publication fails after rollback.
+    """
+    publication = report
+    all_planned_files = list(all_planned_files)
+    publication.planned_files = all_planned_files
+    if not config.should_apply:
+        return publication
+    previous_manifest = load_manifest(config.manifest_path)
+    previous_records = _manifest_record_by_path(previous_manifest)
+    _validate_full_prune_consent(config, all_planned_files, previous_records)
+    backups: dict[Path, bytes | None] = {}
+    initial_written = publication.written
+    initial_deleted = publication.deleted
+    try:
+        _validate_planned_targets(all_planned_files, previous_records)
+        _publish_planned_targets(
+            config, all_planned_files, previous_records, publication, backups, failure_injector
+        )
+        _remove_stale_files(config, previous_records, all_planned_files, publication, backups)
+        save_manifest(config.manifest_path, _build_manifest(all_planned_files), failure_injector)
+    except (OSError, RuntimeError, ValueError) as error:
+        _rollback_publication(backups, publication, initial_written, initial_deleted)
+        raise error
+    _sort_report_details(publication)
+    return publication
+
+
+def publish_plan(config: MaterializerConfig, *all_arguments: object, **all_keywords: object) -> MaterializationReport:
+    """Publish a plan while preserving the legacy optional-argument call form.
+
+    Args:
+        config: Materializer paths and application settings.
+        all_arguments: Planned files, optional report, and optional injector.
+        all_keywords: ``all_planned_files``, ``report``, or ``failure_injector``.
+
+    Returns:
+        The updated materialization report.
+
+    Raises:
+        TypeError: If required data is missing or arguments are duplicated.
+        OSError, RuntimeError, ValueError: If publication fails after rollback.
+    """
+    planned_files = all_keywords.pop("all_planned_files", None)
+    report = all_keywords.pop("report", None)
+    failure_injector = all_keywords.pop("failure_injector", None)
+    if all_keywords or (planned_files is None and not all_arguments) or len(all_arguments) > publish_plan_max_positional_arguments:
+        raise TypeError("publish_plan requires planned files and accepts at most three values")
+    if planned_files is not None and all_arguments:
+        raise TypeError("publish_plan received duplicate planned files")
+    if planned_files is None:
+        planned_files = all_arguments[0]
+    if len(all_arguments) > 1:
+        if report is not None:
+            raise TypeError("publish_plan received duplicate report")
+        report = all_arguments[1]
+    if len(all_arguments) > publish_plan_failure_injector_position:
+        if failure_injector is not None:
+            raise TypeError("publish_plan received duplicate failure injector")
+        failure_injector = all_arguments[2]
+    publication = report if isinstance(report, MaterializationReport) else MaterializationReport()
+    return _publish_plan(config, planned_files, publication, failure_injector)
+
+
+def _redact_private_paths(
+    message: str,
+    config: MaterializerConfig | None,
+    all_private_paths: Iterable[Path],
+) -> str:
+    redacted_message = message
+    configured_paths = (
+        (config.source_root, config.target_root, config.manifest_path)
+        if config is not None
+        else ()
+    )
+    for each_private_path in (*configured_paths, *all_private_paths):
+        redacted_message = redacted_message.replace(str(each_private_path), "<private-path>")
+    return redacted_message
+
+
+def _build_report_payload(
+    report: MaterializationReport,
+    config: MaterializerConfig | None,
+    should_apply: bool,
+    all_private_paths: Iterable[Path],
+) -> dict[str, object]:
+    report_payload = {each_category: getattr(report, each_category) for each_category in report_categories}
+    report_payload["incomplete_generation"] = report.incomplete_generation
+    report_payload["reconcile_required"] = report.reconcile_required
+    report_payload["error_details"] = [
+        _redact_private_paths(each_message, config, all_private_paths)
+        for each_message in sorted(report.error_details, key=str.casefold)
+    ]
+    report_payload["details"] = {
+        each_category: [_redact_private_paths(each_message, config, all_private_paths) for each_message in messages]
+        for each_category, messages in report.details.items()
+    }
+    report_payload["dry_run"] = not should_apply
+    return report_payload
+
+
+def create_argument_parser() -> argparse.ArgumentParser:
+    """Create the command-line parser for compatibility materialization.
+
+    Returns:
+        A parser accepting source and target roots, the apply flag, and the prune opt-in.
+    """
+    parser = MaterializerArgumentParser(description=__doc__)
+    parser.add_argument("source_root", type=Path)
+    parser.add_argument("target_root", type=Path)
+    parser.add_argument("--apply", dest="should_apply", action="store_true")
+    parser.add_argument(full_prune_opt_in_flag, dest="should_allow_full_prune", action="store_true")
+    return parser
+
+
+def main(*all_arguments: object) -> int:
+    """Run materialization from command-line arguments.
+
+    Args:
+        all_arguments: Optional argument list; no value uses process arguments.
+
+    Returns:
+        Zero after reporting the materialization result.
+
+    Raises:
+        TypeError: If the optional argument list has an invalid shape or type.
+        MaterializerError: If the source and target configuration is unsafe.
+        OSError, RuntimeError, ValueError: If discovery, planning, or publication fails.
+    """
+    report = MaterializationReport()
+    config: MaterializerConfig | None = None
+    should_apply = False
+    source_root: Path | None = None
+    target_root: Path | None = None
+    try:
+        if len(all_arguments) > 1:
+            raise TypeError("main accepts at most one argument list")
+        cli_arguments = all_arguments[0] if all_arguments else None
+        if cli_arguments is not None and not isinstance(cli_arguments, list):
+            raise TypeError("main argument must be a list of command-line strings")
+        options = create_argument_parser().parse_args(cli_arguments)
+        should_apply = options.should_apply
+        source_root = options.source_root
+        target_root = options.target_root
+        config = MaterializerConfig(
+            source_root,
+            target_root,
+            should_apply=should_apply,
+            should_allow_full_prune=options.should_allow_full_prune,
+        )
+        discovered_agents = discover_agents(config)
+        planned, report = build_plan(config, all_agents=discovered_agents)
+        publish_plan(config, all_planned_files=planned, report=report, failure_injector=None)
+    except (MaterializerError, OSError, RuntimeError, ValueError) as error:
+        report.add_error(str(error))
+    all_private_paths = tuple(
+        each_path
+        for each_path in (source_root, target_root)
+        if each_path is not None
+    )
+    report_payload = _build_report_payload(report, config, should_apply, all_private_paths)
+    print(json.dumps(report_payload, sort_keys=True))
+    return 1 if report.errors or report.conflicted else 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())