Manfred Riem 902f5431f9 Harden command registration path handling (#3088) ...

* fix: validate command 'file' field against path traversal in registrar

CommandRegistrar.register_commands() read each command body from
source_dir / cmd_file without validating the manifest 'file' field,
unlike the parallel skill and preset readers which already reject
absolute paths and '..' traversal. A malicious extension/preset/bundle
manifest with file: ../../../etc/passwd (or an absolute path) could
read arbitrary host files verbatim into a generated agent command at a
predictable path (GHSA-w5fv-7w9x-7fc5, CWE-22).

Add the same containment guard at the command read site and reject a
traversal/absolute 'file' at manifest-load time in
ExtensionManifest._validate() for defense-in-depth, plus regression
tests for both the read path and the manifest validator.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* test/fix: address review — robust absolute-path test and tolerant reads

- register_commands(): use is_file() instead of exists() and skip the
  command if read_text() raises (directory or non-UTF8 file), aligning
  with the other command/skill readers.
- Traversal tests: point the absolute-path payload at the real temp
  secret.txt (guaranteed to exist on all platforms) instead of
  /etc/passwd, so the absolute-path guard is genuinely exercised and the
  test fails if it regresses, rather than passing because the target
  happens not to exist (e.g. on Windows runners).

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* test: rename traversal fixtures to avoid CodeQL secret-storage false positive

The regression fixtures named an out-of-tree file secret.txt with
TOP-SECRET-CREDENTIAL content. CodeQL's clear-text-storage heuristic
treated that read content as sensitive and followed the static path
into the pre-existing write_text sinks in _write_registered_output,
raising false 'clear-text storage of sensitive information' alerts on
PR 3088. Rename the fixtures to neutral outside.txt / OUTSIDE-FILE-MARKER
and drop /etc/passwd payloads; the test semantics (a file outside
source_dir must never be read into a generated command) are unchanged.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* fix: reject Windows drive-relative 'file' values in traversal guards

is_absolute() is False for Windows drive-relative paths like C:outside.txt,
which contain no '..' yet resolve against the process CWD on that drive —
bypassing the containment guard on Windows. Evaluate the 'file' value under
PureWindowsPath as well so both the registrar runtime guard and the
manifest-load validator reject drive letters (and backslash '..' segments)
cross-platform. Extend the regression tests with drive-relative cases.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* fix: use anchor under both path flavors so POSIX-absolute is rejected on Windows

On a Windows runner WindowsPath('/abs/outside.md').is_absolute() is False
(no drive), so the prior native-Path check let a leading-slash 'file' value
through and the manifest validator did not raise. Evaluate the value under
both PurePosixPath and PureWindowsPath and reject any non-empty anchor —
covering POSIX-absolute, Windows drive-relative, Windows absolute, and
rooted-without-drive — in both the registrar guard and the manifest
validator. The registrar join now uses the raw 'file' string so native
separators are handled by the resolve()/relative_to() containment check.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* fix: validate command 'file' field against path traversal in registrar

CommandRegistrar.register_commands() read each command body from
source_dir / cmd_file without validating the manifest 'file' field,
unlike the parallel skill and preset readers which already reject
absolute paths and '..' traversal. A malicious extension/preset/bundle
manifest with file: ../../../etc/passwd (or an absolute path) could
read arbitrary host files verbatim into a generated agent command at a
predictable path (GHSA-w5fv-7w9x-7fc5, CWE-22).

Add the same containment guard at the command read site and reject a
traversal/absolute 'file' at manifest-load time in
ExtensionManifest._validate() for defense-in-depth, plus regression
tests for both the read path and the manifest validator.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* test/fix: address review — robust absolute-path test and tolerant reads

- register_commands(): use is_file() instead of exists() and skip the
  command if read_text() raises (directory or non-UTF8 file), aligning
  with the other command/skill readers.
- Traversal tests: point the absolute-path payload at the real temp
  secret.txt (guaranteed to exist on all platforms) instead of
  /etc/passwd, so the absolute-path guard is genuinely exercised and the
  test fails if it regresses, rather than passing because the target
  happens not to exist (e.g. on Windows runners).

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* test: rename traversal fixtures to avoid CodeQL secret-storage false positive

The regression fixtures named an out-of-tree file secret.txt with
TOP-SECRET-CREDENTIAL content. CodeQL's clear-text-storage heuristic
treated that read content as sensitive and followed the static path
into the pre-existing write_text sinks in _write_registered_output,
raising false 'clear-text storage of sensitive information' alerts on
PR 3088. Rename the fixtures to neutral outside.txt / OUTSIDE-FILE-MARKER
and drop /etc/passwd payloads; the test semantics (a file outside
source_dir must never be read into a generated command) are unchanged.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* fix: reject Windows drive-relative 'file' values in traversal guards

is_absolute() is False for Windows drive-relative paths like C:outside.txt,
which contain no '..' yet resolve against the process CWD on that drive —
bypassing the containment guard on Windows. Evaluate the 'file' value under
PureWindowsPath as well so both the registrar runtime guard and the
manifest-load validator reject drive letters (and backslash '..' segments)
cross-platform. Extend the regression tests with drive-relative cases.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* fix: use anchor under both path flavors so POSIX-absolute is rejected on Windows

On a Windows runner WindowsPath('/abs/outside.md').is_absolute() is False
(no drive), so the prior native-Path check let a leading-slash 'file' value
through and the manifest validator did not raise. Evaluate the value under
both PurePosixPath and PureWindowsPath and reject any non-empty anchor —
covering POSIX-absolute, Windows drive-relative, Windows absolute, and
rooted-without-drive — in both the registrar guard and the manifest
validator. The registrar join now uses the raw 'file' string so native
separators are handled by the resolve()/relative_to() containment check.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* refactor: harden register_commands inputs and tighten manifest 'file' validation

Address review feedback on #3088:
- register_commands(): skip non-string/empty 'file' values instead of
  raising TypeError, and hoist source_dir.resolve() out of the per-command
  loop.
- ExtensionManifest._validate(): reject 'file' values with leading/trailing
  whitespace with a clear ValidationError instead of a confusing
  missing-file failure later.
- tests: add non-string 'file' and whitespace cases; use yaml.safe_dump
  with explicit utf-8 encoding in the manifest validation test.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* refactor: align runtime '..' policy, correct comment, dedupe test helper

Address review feedback on #3088:
- register_commands(): also reject '..' segments under both POSIX and
  Windows semantics, keeping runtime policy consistent with
  ExtensionManifest._validate() and the skill/preset readers (not just
  relying on the resolve()/relative_to() containment backstop).
- Replace the version-dependent is_absolute() claim in the extensions.py
  comment with the actual portability rationale (native Path is OS-
  dependent; C:foo is anchored but not absolute).
- Extract the duplicated leak-detection assertion into
  _assert_no_marker_leak() and add an in-bounds '..' payload that
  exercises the new runtime '..' rejection.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Extract shared path-safety policy and warn on unreadable command files

Introduce relative_extension_path_violation() in _utils.py as the single
source of truth for the extension-relative `file` path-safety policy, and
use it from both the runtime registrar guard (agents.py) and the
manifest-load validator (extensions.py) so the two cannot drift.

Warn (instead of silently skipping) when an in-bounds command file exists
but cannot be read/decoded, surfacing misconfigured extensions.

Add unit tests for the shared helper, a read-skip warning test, and make
the in-bounds `..` test create its target file so the skip is attributable
to the `..` rejection rather than file absence.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Retrigger CI

Empty commit to re-trigger code scanning / CodeQL analysis on the PR
merge ref.

Assisted-by: GitHub Copilot CLI (model: Claude Opus 4.8, autonomous)
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

---------

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

2026-06-22 10:25:29 -05:00

13 KiB

Raw Permalink Blame History

View Raw