skillspector-mcp-2.11.2-1.1 RPM for noarch

From OpenSuSE Tumbleweed for noarch

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.

Provides

• skillspector-mcp

Requires

• python313-mcp >= 1.2.0
• rpmlib(CompressedFileNames) <= 3.0.4-1
• rpmlib(FileDigests) <= 4.6.0-1
• rpmlib(PayloadFilesHavePrefix) <= 4.0-1
• rpmlib(PayloadIsZstd) <= 5.4.18-1
• skillspector = 2.11.2

License

Apache-2.0

Changelog

* 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)

Files

Generated by rpm2html 1.8.1

Fabrice Bellet, Tue Sep 22 22:25:44 2026