superzip-security

name	superzip-security
description	Security review and hardening workflow for SuperZip archive extraction, path handling, Defender integration, GitHub Actions scanning, secrets, and third-party dependency updates.

SuperZip Security Skill

Before editing security-sensitive code, identify the boundary:

• Archive metadata parser.
• Destination path join and canonicalization.
• ZIP compatibility extraction.
• LHA/LZH compatibility extraction.
• SUZIP block metadata validation.
• Microsoft Defender opt-in scan.
• GitHub Actions scanner integration.

Required selection step:

tools/verification_plan.ps1 -IncludeUntracked
tools/verify_changes.ps1 -IncludeUntracked

Single-agent and Codex Security worker policy:

• Run SuperZip security work serially in the current agent. Do not spawn, fork, delegate to, or fan out subagents or worker agents for discovery, validation, attack-path analysis, workflow review, release-artifact review, or remediation unless the maintainer explicitly reverses this policy for a specific task.
• Codex Security scanner subagents are on-demand only. Use them only when the maintainer explicitly invokes a Codex Security scan phase that genuinely requires delegated workers or explicitly authorizes them for that scan. Prefer the one-worker or serial Codex Security path whenever the installed plugin can run it honestly, and record why each worker was necessary.
• If any other external security plugin requires delegated workers for a named mode, do not claim that mode ran. Run the closest serial phase sequence instead, save the evidence, and record the delegated mode as unavailable by maintainer policy.
• Parallel shell commands are allowed only when they do not create model-agent workers and do not reduce audit quality; prefer serial execution when the maintainer asks for efficient subscription usage.

Smart security-scan triggers:

• Trigger tools/security_scan.ps1 in addition to the selected verification plan when a change is extensive or touches dangerous functions, archive parser state machines, destination path validation, verified publication, overwrite decisions, process creation, dynamic library loading, Defender scanning, SHA-256 integrity, installer/update logic, release packaging, workflow permissions, dependency provenance, or agent/MCP tooling.
• Use the narrowest targeted scan that proves the changed boundary. Escalate to tools/verify_changes.ps1 -IncludeUntracked -Full when the boundary is broad, the changed paths are unknown, a targeted scan fails, or a security regression is plausible but not localized.
• Do not run heavyweight scans merely because unrelated text changed. The trigger is risk and boundary driven.

Use the full profile automatically when the plan escalates, a targeted check fails, the changed path set is broad or unknown, or a wider security regression is suspected:

tools/verify_changes.ps1 -IncludeUntracked -Full

After push, wait only for relevant workflows selected by the classifier:

tools/wait_relevant_workflows.ps1 -Commit <sha> -Mode final

For non-critical intermediate commits, -Mode opportunistic may be used only when workflowWaitPolicy.deferAllowed=true. Security, workflow, verifier, MCP, skill, and full-escalation changes must not be reported complete until final mode and any required post-push audit pass. The waiter resolves local refs to full commit SHAs and performs a GitHub CLI preflight before polling. Authentication failures, authorization failures, and gh run list errors are hard blockers; do not keep waiting on a state where every selected workflow appears missing. Fuzzing is intentionally long-running. It is observed separately from normal blocking waits; check it opportunistically during iteration and include it in the blocking wait only for final handoff or release.

Add or update tests for:

• .., absolute paths, drive paths, UNC paths.
• Reserved Windows names.
• Existing destination junctions, symlinks, mount points, and other reparse parents that could redirect extraction publication outside the selected root.
• Source-tree junctions, symlinks, mount points, and other reparse entries that could make archive creation include files outside selected inputs.
• Decoded or extracted output totals that exceed SuperZip's resource policy, especially compatibility extractors with independent byte-accounting loops.
• Existing-file overwrite refusal.
• Corrupt payload and CRC mismatch.
• Oversized or malformed archive metadata.
• Symbolic links and other unsupported special-file entries in compatibility formats.

Do not store credentials, PATs, org IDs, scan targets, or Defender results in tracked files.

For broad hardening or modernization requests, map the request to concrete secure-by-design controls: fail-closed parsing, bounded allocation, least privilege, provenance, explicit opt-ins, and scanner-visible remediation. Do not substitute process text, exclusions, or skipped checks for root-cause fixes.

Workflow and release hardening rules:

• Do not add GitHub Actions environment: blocks, deployment: keys, or any workflow mechanism that creates deployment records.
• Greenbone/OpenVAS workflow-dispatch target text is only a request to the external OIDC broker. The workflow must use the broker-returned greenbone_target as the effective scan target and must not let repository inputs bypass broker authorization.
• Do not hide scanner findings with broad exclusions, placeholder tests, or event-specific jobs that normally show as skipped. A narrow generated-output skip is acceptable only when the skipped path is documented and not product source.
• Keep release artifacts HIP-enabled, Windows x64-only, and equivalent between MSI and portable ZIP. CPU-only artifacts are validation-only and must not be published.
• Do not track extracted runtime binaries. Keep upstream packages under third_party/upstream/**, record checksums and license notes, and extract runtime DLLs into the build tree only after checksum verification.
• LHA/LZH support must stay extraction-only, in-process, and backed by the vendored Lhasa 0.5.0 decoder. Keep SuperZip's two-pass path/payload validation around Lhasa and preserve the unmodified upstream archive under third_party/upstream/lhasa/0.5.0/.
• WIM support must stay extraction-only for standalone WIMs, in-process, and backed by the bundled app-local wimlib 1.14.5 DLL. Keep split WIMs rejected until multipart handling is deliberately implemented, reject reparse points, hard links, alternate data streams, device entries, encrypted/offline/virtual files, and publish only rechecked staged regular files through SuperZip's verified temporary-file path.
• LZMA support must stay extraction-only for legacy .lzma LZMA-Alone streams, in-process, and backed by the vendored LZMA SDK 26.01 decoder. Keep it single-file, reject oversized dictionaries and decoded output, and publish only through SuperZip's verified temporary-file path.
• Lzip support must stay extraction-only for .lz and TAR-stream-only for .tar.lz/.tlz, in-process, and backed by SuperZip's lzip wrapper over the vendored LZMA SDK 26.01 decoder. Keep version, dictionary, EOS, CRC32, data-size, and member-size checks mandatory for every member, reject trailing non-member data, and publish only through verified temporary paths.
• CPIO.GZ support must stay in-process as a Gzip-filtered SVR4 new ASCII CPIO stream. Keep the two-pass validation model, reject unsafe inner paths, unsupported special files, overwrite attempts, malformed Gzip trailers, and 070702 checksum mismatches before publishing output.
• The only acceptable open code-scanning findings are Scorecard , , , and .