odysseus/src/path_confinement.py
Léo 2a540f2acc fix(runtime): enforce workspace confinement in one place
"Is this path inside that root" is asked in twenty places in this tree and
answered twenty times by a locally written realpath/commonpath pair. Nine test
files exist because nine call sites each needed their own proof. Each one is
defensible alone; together they are the defect, because the boundary has no
single definition and a site that gets a detail wrong is wrong by itself.

src/path_confinement.py is that definition, and it settles the details the
copies disagreed on. Both sides get canonicalized: comparing a realpath-ed
candidate against a root that was only abspath-ed is the macOS /tmp ->
/private/tmp mismatch that has already produced a false failure here, and
canonicalizing one side is worse than canonicalizing neither. commonpath rather
than startswith, because /a/bc begins with /a/b and is not inside it. A relative
candidate joins the root rather than os.getcwd(), which is whatever directory
the server happens to be running in. NUL and newline are refused with a reason
instead of caught by a bare `except Exception` and reported as an ordinary
escape. Eighteen call sites go through it now. It deliberately does not decide
whether a path is sensitive -- that deny list answers "allowed" rather than
"inside", and it stays with src/tool_execution, which owns it. The one
commonpath left in the tree, in src/workspace_paths.py, stays: that function
translates a host path into a container path, so canonicalizing either side
would change the relative path it computes and break the mapping. It is not a
confinement check.

Two of those sites were weaker than the rest and are fixed rather than moved.
The email attachment check used abspath, which folds `..` but does not resolve
symlinks, so a symlink written into the extraction directory passed it and was
then read through. The skill-reference guard compared a realpath-ed target
against a raw dirname, so on a host where the skills tree is reached through a
symlink the two sides never matched and the guard could not fire.

The execution boundary had two separate holes.

The workspace namespace bound /home and /mnt read-write. On the one platform
where that namespace engages at all, a command inside it reaches outside the
workspace and writes to the user's home directory -- measured by running this
argv on a Linux host with working bubblewrap, not inferred from the source.
Binding the user's whole home directory into a workspace-confinement namespace
gives back most of what the namespace was for. Both are read-only now. The
workspace is also bound writable at its real host path, not only at /workspace:
BashTool's own /tmp redirect rewrites `/tmp/` to `<agent_cwd()>/.tmp/` before
the namespace is built, so the command bwrap receives already names the real
path, and those writes previously landed only because the workspace happened to
sit under the writable /home.

`namespaced or _replace_workspace_alias(...)` chose between a mount namespace
and a regex with nothing in the result saying which one ran. The fallback
rewrites the literal token /workspace in the command string, so a command that
never mentions /workspace is untouched by it and runs on the host unrestricted
-- which is every agent shell command on macOS. Both tools now ask
containment.probe() instead of each deciding for itself, and every bash and
python result carries a containment block naming the mechanism and stating
whether the filesystem dimension actually held. Under enforcing mode the
command is not run and the result says so.

That block reports the filesystem dimension only, and says so in a
reported_dimensions field. The probe knows this host could also give a process
group and a real wall clock, but these two tools still assemble their own
create_subprocess_* call and pass neither, so listing those dimensions would be
exactly the false claim src/containment.py calls worse than an honest absence.

probe() is new on src/containment.py: the same mechanism table and the same
arithmetic as acquire(), stopping before the side effects. acquire() is the
wrong shape for a decision -- it writes a durable grant record, and a record
whose pid is never filled in and whose release() never runs is an entry a
restart reaper keeps finding.

CONTAINMENT_MODE stays report_only. Flipping it refuses every agent shell
command on macOS and on any Linux host without bubblewrap, which is a product
decision rather than a code one.

Smaller things in the same area: the /tmp redirect's makedirs was unguarded, so
a read-only workspace turned a command that merely mentioned `/tmp/` into an
OSError traceback instead of a tool error; it degrades now. WORKSPACE_MOUNT
moved to src/constants.py so the namespace and the path resolvers read one
definition of the contract rather than two. The ".tmp" dirname got a constant,
since it appeared in both tool paths.

One generated artifact moved with it: website/configuration-reference.md pins
the source line where each ODYSSEUS_* variable is read, and three of those
shifted. Regenerated with scripts/generate_env_reference.py; the diff is line
numbers only.

Three existing tests changed. test_workspace_artifact_tool_floor asserted that
an unsafe interpreter prefix produces no `--ro-bind <prefix> <prefix>`, which
now fires on /home because /home is legitimately a read-only base mount.
Asserting the absence of a literal flag string cannot distinguish "the prefix
was rejected" from "the argv mounted that root itself", so it compares the argv
against the no-prefix baseline instead: an unsafe prefix must add nothing.

The Windows bash test asserted dict equality on the
whole result, which makes adding a field to every bash result impossible without
touching a test about tmux; it asserts the shape now. The personal-dir symlink
test grepped the resolver's source for the literal "os.path.realpath", which is
gone because the resolution moved into the shared boundary -- it keeps the
negative assertion that the closure must not grow its own abspath check again,
and the behavioural half now runs against the boundary, where it covers every
call site instead of one closure.

Not verified: the bubblewrap argv is asserted, not executed. There is no bwrap
on macOS, and in Docker it needs --privileged to work at all -- default and
seccomp=unconfined both fail with "Creating new namespace failed", and
--cap-add=SYS_ADMIN fails at pivot_root. The Python tool's
needs_virtual_namespace gate means ordinary Python code gets no namespace even
on a Linux host that could provide one; that is reported now but deliberately
not changed, because it alters the Linux Python path on every call and cannot be
checked from here.
2026-10-01 19:45:59 +02:00

188 lines
8.2 KiB
Python

"""The filesystem confinement boundary. One implementation, every call site.
"Is this path inside that root" is asked in twenty places in this tree, and
twenty times it is answered by a locally written ``realpath`` +
``os.path.commonpath`` pair. Each one is defensible on its own. Together they
are the problem: the boundary has no single definition, so a site that gets a
detail wrong is wrong *alone*, and a site added tomorrow starts from whichever
neighbour its author happened to copy.
The details that differ between those copies, and what this module settles:
**Both sides get canonicalized.** Comparing a ``realpath``-ed candidate against
a root that was only ``abspath``-ed is the bug class that has already cost this
project real time: on macOS ``/tmp`` is a symlink to ``/private/tmp``, so the
two sides disagree about a path neither of them is wrong about. It reads as an
escape and refuses a legitimate access. Canonicalizing one side is worse than
canonicalizing neither.
**``commonpath``, never ``startswith``.** ``/a/bc`` begins with ``/a/b`` and is
not inside it.
**Case folding is the filesystem's business, not the comparison's.**
``os.path.normcase`` lowercases on Windows and is the identity everywhere else
— including macOS, whose default filesystem is case-insensitive while its
``realpath`` preserves case. So normcase alone does not make the comparison
agree with the filesystem on macOS, and :func:`is_inside` does not pretend
otherwise: it answers about the canonical path, which is the question a
confinement check should be asking. Where a caller needs to match the
filesystem's own folding it must compare real paths of real files, not strings.
**A relative candidate joins the root, never the process cwd.** ``abspath`` of a
relative path silently uses ``os.getcwd()``, which is whatever the server
happens to be running in. A confinement helper that does that is resolving
against the wrong base before it even starts comparing.
**NUL and newline are rejected, not caught.** Several of the copies wrap the
whole comparison in ``except Exception: return False``, which turns a malformed
path into "outside" — the safe answer, reached by accident. Here it is a
``ValueError`` with a reason.
**``commonpath`` raising means outside.** It raises across Windows drive letters
and for mixed absolute/relative inputs. Both mean the candidate is not under the
root, so the refusal is deliberate rather than incidental.
What this module does *not* do: decide whether a path is sensitive (``.ssh``,
``id_rsa``, …). That is a separate deny list applied inside an allowed root, and
it lives with the callers that own it — ``src/tool_execution`` for the agent
tools. Confinement answers "inside the root"; it does not answer "allowed".
Relationship to :mod:`src.containment`: that module is the boundary for *where a
process runs*; this one is the boundary for *which paths a path check accepts*.
A contained process is restricted by a mount namespace, which this module cannot
express and does not try to; an in-process read of a model-supplied path is
restricted by this module, which a namespace does not see.
"""
from __future__ import annotations
import os
__all__ = [
"PathEscape",
"canonical_root",
"confine",
"is_inside",
]
class PathEscape(ValueError):
"""A candidate path does not resolve inside the root it was checked against.
A subclass of :class:`ValueError` so the call sites this replaces — which
raise ``ValueError`` and are caught as such by their callers and their
tests — keep behaving the way they did.
"""
def __init__(self, root: str, candidate: str, reason: str = "") -> None:
self.root = str(root)
self.candidate = str(candidate)
self.reason = str(reason or "outside the allowed root")
super().__init__(
f"path {self.candidate!r} is {self.reason} ({self.root})"
)
def _reject_unusable(value: str, *, label: str) -> str:
"""Normalize a path argument to ``str``, refusing the unusable shapes.
``\\x00`` is refused here because the OS layer raises on it much later and
from somewhere unhelpful, and because a broad ``except Exception`` around
the comparison would otherwise record it as an ordinary escape. Newlines
are refused for the same reason the workspace-mount parser refuses them:
a path carrying one has been built by splitting something that was not a
path list.
"""
if value is None:
raise ValueError(f"{label} is required")
if isinstance(value, os.PathLike):
value = os.fspath(value)
if not isinstance(value, str):
raise ValueError(f"{label} must be a path, got {type(value).__name__}")
text = value.strip()
if not text:
raise ValueError(f"{label} is required")
if "\x00" in text:
raise ValueError(f"{label} must not contain NUL")
if "\n" in text or "\r" in text:
raise ValueError(f"{label} must not contain a newline")
return text
def canonical_root(root) -> str:
"""The canonical form of a confinement root.
Exposed because a caller that holds a root across several checks should
canonicalize it once, and because a caller comparing two paths itself needs
the same canonical form this module compares against — a realpath-ed value
tested against a raw one is the asymmetry this module exists to remove.
"""
text = _reject_unusable(root, label="root")
return os.path.realpath(os.path.expanduser(text))
def _canonical_candidate(root: str, candidate) -> str:
"""Canonicalize ``candidate``, resolving a relative path under ``root``.
``realpath`` is deliberately the non-strict kind: a final component that
does not exist yet is normalized rather than refused, because a write target
is a legitimate thing to confine. Everything that *does* exist is resolved,
so a symlink anywhere in the chain — including the final component — is
followed before the comparison rather than after the open.
"""
text = _reject_unusable(candidate, label="path")
expanded = os.path.expanduser(text)
if not os.path.isabs(expanded):
expanded = os.path.join(root, expanded)
return os.path.realpath(expanded)
def is_inside(root, candidate, *, allow_root: bool = True) -> bool:
"""True when ``candidate`` resolves inside ``root``.
The boolean form, for call sites whose contract is a predicate. A malformed
argument is ``False`` here rather than a raise, because a predicate that
raises is the reason those call sites wrapped themselves in
``except Exception`` in the first place. Use :func:`confine` where the
caller wants the resolved path and a reason for the refusal.
``allow_root=False`` excludes the root itself, for a caller whose operation
is only meaningful on something *under* the root — deleting a file, say,
where the root is the directory it must not be.
"""
try:
confine(root, candidate, allow_root=allow_root)
return True
except (ValueError, OSError):
return False
def confine(root, candidate, *, allow_root: bool = True) -> str:
"""Resolve ``candidate`` inside ``root``, or raise.
Returns the canonical absolute path, which is what the caller should then
open: resolving and then opening the *original* string re-introduces the
symlink race the resolution just closed.
:raises ValueError: either argument is unusable as a path.
:raises PathEscape: the candidate resolves outside the root.
"""
base = canonical_root(root)
resolved = _canonical_candidate(base, candidate)
if resolved == base:
if allow_root:
return resolved
raise PathEscape(base, candidate, "the root itself, not a path inside it")
# normcase folds case on Windows and is the identity elsewhere; it is
# applied to both sides or to neither, which is the whole point.
try:
common = os.path.commonpath([os.path.normcase(resolved), os.path.normcase(base)])
except ValueError:
# Different Windows drives, or mixed absolute/relative. Both mean the
# candidate is not under the root.
raise PathEscape(base, candidate) from None
if common != os.path.normcase(base):
raise PathEscape(base, candidate)
return resolved