| Index | index by Group | index by Distribution | index by Vendor | index by creation date | index by Name | Mirrors | Help | Search |
| Name: skillspector-mcp | Distribution: openSUSE Tumbleweed |
| Version: 2.13.3 | Vendor: openSUSE |
| Release: 1.1 | Build date: Mon Nov 2 14:22:36 2026 |
| Group: Unspecified | Build host: reproducible |
| Size: 0 | Source RPM: skillspector-2.13.3-1.1.src.rpm |
| Packager: https://bugs.opensuse.org | |
| Url: https://github.com/NVIDIA/skillspector | |
| Summary: MCP server mode for SkillSpector | |
This subpackage enables the Model Context Protocol (MCP) server mode of SkillSpector, exposed through the "skillspector mcp" subcommand. It pulls in the optional MCP runtime dependency so the FastMCP-based server can be started for local CLI agents or over HTTP.
Apache-2.0
* Mon Nov 02 2026 Martin Pluskal <mpluskal@suse.com>
- Update to version 2.13.3:
* Scan integrity: baseline and recursive reports can no longer
overwrite the primary input or a child skill input, missing
custom YARA rule directories report incomplete inspection,
and artifact inspection continues after a matcher timeout
* Detection coverage: npx commands in JSON MCP configs, Pipenv
dependencies and locked versions, urllib-header credential
flows, agent-memory writes from test carriers, CRLF-shebang
launchers, and restored PowerShell SC3/P2 checks
* Deselect two retry-after deadline tests: the Factory SDK names
the rate-limit failure OpenAIRateLimitError, not RateLimitError
(deterministic mismatch; deadline behavior itself still passes
and stays covered by the build_context deadline tests)
* Fewer false incomplete scans: bounded shell syntax, generated
browser-template text, and Markdown code-fence line handling
* Large scans: configurable tool-description batch limit and a
longer default workflow deadline
* Thu Oct 29 2026 Martin Pluskal <mpluskal@suse.com>
- Update to version 2.13.2:
* No more false parser-limit reports on bare command
wrappers (timeout, sudo, nice, xargs) at clause
boundaries in tables and prose references
* sudo -n parsed as flag; descriptor redirections
(e.g. 2>&1) count as complete; MCP batch cap
aligned with the 128-file limit
* Direct template messages on verified built-in Node
assertion imports no longer trip shell-substitution
limits; .mjs/.cjs classified consistently
- Deselect two wall-clock-bound tests on aarch64 (pre-existing,
untouched by this update): the 10 s subprocess and 1 s worker
bounds overrun slow interpreter teardown, with all detection
assertions passing.
* Mon Oct 26 2026 Martin Pluskal <mpluskal@suse.com>
- Update to version 2.13.1:
* Fewer false incomplete scans: quote boundaries in JSON
examples, documented dotted config fields, quoted YAML
versions, corroborated passive images, and possessive prose
no longer exhaust the shell parser
* Fail closed more often: obfuscated curl via declared markers,
report-shaped strings reaching shell execution, and the
bounded-helper proof keep incomplete inspection
* Hardening backports: atomic report/baseline writes refusing
symlinked destinations, per-file static-analysis time limits,
and failed gap-fill passes reported incomplete
* Release sdist now excludes the uv build cache
- Add skillspector-typer-parametersource.patch: compare the
transitive-depth parameter source by member name, not enum
identity. Factory ships typer 0.27, which vendors click, while
upstream pins typer < 0.24, so the check rejected every scan
- -mcp-registry invocation with exit 2.
- Deselect eleven test_cli_subprocess.py tests: they re-exec
the CLI against the sourceless src/ tree and die with
PackageNotFoundError before the CLI runs; the same guard
behavior is covered in-process.
* Sat Oct 17 2026 Martin Pluskal <mpluskal@suse.com>
- Update to version 2.13.0:
* Security-hardening backport: decode-and-execute chain detection,
constructed sensitive-path and secret-flow detection, BiDi control
checks in bundled scripts, executable-Python and TypeScript
declaration inspection, risky bundled-settings/MCP flags
* Fail closed more often: malformed LLM responses, unloadable YARA
rules and skipped linked skills now report incomplete coverage;
archive-member/disk-path collisions keep the real content
* MCP HTTP listeners restricted to loopback, Git credentials isolated
for HTTP scans, Codex CLI provider disabled (unsafe sandbox),
release workflow pinned, report-output overwrite guards
* NVIDIA Build default moves to GLM-5.3, agent CLI environment
allowlisted, per-skill batch timeouts, scan-hang fixes
* Many more fixes and improvements; see upstream release notes
for the full list
- Require git-core for %check (new credential-isolation test shells
out to git) and deselect one MCP re-exec test that needs an
installed dist-info.
* Fri Sep 25 2026 Martin Pluskal <mpluskal@suse.com>
- Add skillspector-fifo-timeout.patch: raise the fifo-swap
regression test child timeout 5s -> 30s. The cold import of
the langchain cone in a fresh interpreter alone takes ~8s
on aarch64 builders, so the test killed its own child
during import (returncode -9); the FIFO guard itself is
correct. Upstream-test-only change.
* Fri Sep 25 2026 Martin Pluskal <mpluskal@suse.com>
- Update to version 2.12.0:
* Opt-in CLI gate for any active finding,
configurable static-analysis allowance
* OpenCode integrations, Gemini 3.5 Flash registry
guidance, interactive scan progress
* TP4 analysis of executable Markdown fences
* Stronger local input and report handling, wider
detection of reflective Python access and shipped
bytecode, retries for transient provider failures
* Missing references distinguished from ambiguous
MCP installation blockers
* Thu Sep 10 2026 Martin Pluskal <mpluskal@suse.com>
- Update to version 2.11.2:
* Deduplicate reference-coverage records so duplicate Markdown
references no longer abort with fatal unaccounted_work errors
* Account distinct reference targets sharing one source line
without conflicting completion records
* Stop false static-parser limits on simple runtime parameters,
inline skill invocations, PowerShell member access and long
quoted prose
* Keep runtime-selected printf/command/builtin/env paths marked
partially inspected instead of claiming exact reconstruction
- Deselect two more aarch64-speed tests, both green in upstream CI:
test_cross_window_separator_pair_across_public_surfaces (only the
is_complete projection fails at the 30 s per-artifact budget; every
detection assertion passes) and
test_pe3_repeated_nouns_have_bounded_qualifier_cost (1.28 s against
a 1.0 s wall-clock bound)
* Tue Sep 08 2026 Martin Pluskal <mpluskal@suse.com>
- Update to version 2.11.1:
* Raise the default aggregate scan deadline from 60 s to 600 s
and add SKILLSPECTOR_MAX_WORKFLOW_SECONDS to override it
* Enforce SKILLSPECTOR_MAX_LLM_CONCURRENCY across all concurrent
LLM analyzers rather than separately in each one
* Keep safe and unsafe findings distinct when they share a rule
fingerprint so deduplication cannot drop security evidence
* Detect instructions concealed with ignorable characters or
inter-character separators, and fail closed when obfuscation
prevents complete interpretation
* Parse whitespace-separated allowed-tools without
least-privilege false positives
* Sun Aug 30 2026 Martin Pluskal <mpluskal@suse.com>
- Update to version 2.11.0:
* Analyse bundled execution surfaces: hooks/hooks.json and
.claude/settings*.json now yield BH1 (hooks can execute), BH2 (proven
remote transfer of sensitive content) and BH3 (broad or ignored
permission modes) findings
* Resolve exact npm versions from package-lock.json and
npm-shrinkwrap.json (formats 1-3) instead of manifest ranges
* Add SKILLSPECTOR_TEMPERATURE and SKILLSPECTOR_SEED sampling controls,
keeping provider defaults when unset
* Traverse path ancestors with O_PATH so restricted sandboxes need only
search permission on them
* Drop the nominal MP3/P6 false positives without weakening directive
detection
* Resolve the model default from the provider that actually builds the
chat model on the OpenAI fallback path
- New dependencies python-pywhatwgurl and python-regex for the
bundled-execution-surface analyzer; upstream pins both exactly, floors
only here because Factory already ships newer
- Re-select test_dense_directory_discovery_and_cache_complete_with_...:
the O_PATH traversal took it from 19.7 s to 1.7 s, back inside its own
5 s ceiling
- Deselect the four test_bundled_execution_surface_acceptance.py tests
that shell out to /usr/bin/skillspector, which only an editable dev
install provides, and the malformed_schema case of
test_discovery_parser_bounds_and_ledger_table, where Factory's
pywhatwgurl 0.1.2 parses two invalid punycode hosts that upstream's
pinned 0.1.1 rejects
* Thu Aug 27 2026 Martin Pluskal <mpluskal@suse.com>
- Update to version 2.10.0:
* Inspect hidden files and nested ZIP, DOCX, XLSX and PPTX
artifacts under cumulative bounds, with HIGH SC9 findings for
concealed executables
* Add opt-in transitive reference scanning (--transitive and
its depth/allow/deny options), bounded and fail-closed
* Recognise AISOP/AISP structured skill bundles, report-only
* Add EA5 detection for external model or provider selection
* Add SKILLSPECTOR_OUTPUT_LANGUAGE for localised LLM finding
text and risk_assessment.max_issue_severity for policy gates
* Floor an otherwise SAFE recommendation to CAUTION when
requested LLM analysis fails or the provider is unavailable
* Custom CLI providers keep the existing complete() contract,
but one used on a deadline-bounded path may now also accept
an optional timeout keyword
* Reduce false positives, and report and baseline only the
findings that drove the risk score; see upstream's release
notes for the full list
- Drop the langgraph-cli dependency, build-time and runtime:
upstream moved langgraph-cli[inmem] into a new langgraph-dev
extra, and neither the shipped code nor the tests import it
- Deselect four tests new in this release that fail on aarch64
because this release analyses artifacts more slowly than the
ceilings they assume, confirmed undeselected on an idle
machine; three hit the scanner's own 30 s per-artifact budget
so the ledger fail-closes to partial, and one allows 5 s to
walk and cache 256 one-byte files but takes 19.7 s. The spec
records the timings and the coverage that remains
* Thu Aug 20 2026 Martin Pluskal <mpluskal@suse.com>
- Update to version 2.9.6:
* MCP registry posture scanning
* Ollama, Azure OpenAI, and generic OpenAI-compatible providers
* Opt-in discovery of an author-shipped baseline
(.skillspector-baseline.yaml)
* Detect insecure deserialization (AST10, TT6, DS1-DS4)
* Detect whitespace-padding prompt injection (P9)
* HIGH SC8 when a skill ships __pycache__ or .pyc
* Bound URL, zip and git ingest to limit resource exhaustion
* SC4 reports only verified OSV hits; use lockfile versions
* HTTP MCP rejects local filesystem targets and local YARA dirs
* Reject symlink traversal in skill content; disable git
symlinks on clone
* PE3 no longer flags ordinary OAuth "access token"
documentation
* LLM analyzers retry or isolate malformed responses and
connection failures; incomplete analysis is recorded as
skipped
* Reduce false positives (instructional prose, Markdown
tables, JS RegExp.exec, negated safety constraints, OMS
signatures)
* Map YARA matches via byte offsets; scope the locality guard
to the built-in rule namespace
- Switch source to the official GitHub release tarball
- Drop _service, skillspector-2.5.1.obscpio and
skillspector.obsinfo
- Add packaging >= 24.0 (requirement parsing for OSV lookups)
* Tue Aug 04 2026 Martin Pluskal <mpluskal@suse.com>
- Update to version 2.5.1 (first tagged releases; switches the package
from a pinned git snapshot to the upstream release tags):
* Canonical inspection-ledger reporting: JSON and SARIF reports now
carry execution-completeness and analyzer status, so a consumer
can distinguish a clean zero-finding scan from one that did not
execute reliably
* SKILLSPECTOR_MAX_LLM_CONCURRENCY serializes or bounds asynchronous
LLM analyzer batches for rate-limited providers
* Read exact versions from Python lockfiles for OSV lookups
* Exclude valid OMS signatures from content analysis
- Ship the wheel metadata under the release version again: upstream's
pyproject version now matches the tag, so the dist-info glob follows
the package version
* Fri Jul 03 2026 Martin Pluskal <mpluskal@suse.com>
- Relax over-strict dependency floors set by upstream at initial
release (pin-inflation, not API requirements; the code only uses
long-stable APIs of all three):
* pydantic >= 2.12.0 -> 2.11.7
* rich >= 14.3.0 -> 14.0.0
* typer >= 0.23.0 -> 0.16.0
- Verified against Leap 16.0's shipped versions (pydantic 2.11.7,
rich 14.0.0, typer 0.16.0, click 8.2.1, Python 3.13): full
offline unit suite passes (1249 passed, 12 skipped, 6 xfailed)
and the CLI entry point works.
* Wed Jul 01 2026 Martin Pluskal <mpluskal@suse.com>
- Switch source acquisition to an obs_scm git snapshot of upstream
main: NVIDIA/skillspector publishes no git tags, no GitHub releases
and nothing on PyPI, so no downloadable release tarball exists
- Version changed from 2.3.10 (a locally-built sdist that exists
nowhere upstream) to the snapshot 2.3.9~git20260701.326a2b48
(upstream main HEAD, commit 326a2b48)
- Add _service; drop the hand-built skillspector-2.3.10.tar.gz
* Mon Jun 29 2026 Martin Pluskal <mpluskal@suse.com>
- Use %{primary_python} instead of a hardwired python flavor
* Sun Jun 28 2026 Martin Pluskal <mpluskal@suse.com>
- Initial package skillspector 2.3.10 (CLI scanner; -mcp subpackage
for the MCP server)
- Enable the pytest test suite now that the full langchain/langgraph
runtime stack is packaged; the offline unit tests run at build time
(integration/provider markers and four SSRF allowed-host tests that
require live DNS are deselected)
Generated by rpm2html 1.8.1
Fabrice Bellet, Mon Nov 9 23:33:15 2026