Add the missing Seatbelt interpreter permission through an exact-version, fail-closed compatibility shim so shell and env shebang commands work without weakening filesystem visibility. Align the Linux credential-directory test with its disposable tmpfs behavior and document when the temporary upstream compatibility layer can be removed.
Allow bounded parent directories for symlinked interpreters required by Bubblewrap, expose the macOS shell selector, and reject executable paths that overlap credential directories even through filesystem aliases. Add unit and real containment regressions for these platform-specific cases.
Resolve and expose the pinned seccomp helper before Linux execution, make doctor validate the same platform resources, filter invalid host environment keys from backend declarations, and canonicalize cross-platform path assertions so real containment checks behave consistently on GitHub runners.
Pin Anthropic's sandbox runtime and translate trusted skill profiles into lightweight, on-demand filesystem, network, credential, and process-tree restrictions. Add readiness diagnostics, secure command and Git workspace discovery, signal-safe cleanup, Linux and macOS integration coverage, packaged-runtime checks, and practical compatibility guidance for Codex, Claude Code, OpenCode, and Pi.
Define versioned capability profiles for every bundled skill, add policy resolution and diagnostic commands, preserve sandbox metadata during selective installation, and cover the contract and packaged artifacts with tests.
- add native Claude Code plugin and marketplace manifests
- document Claude Code installation, invocation, updates, and removal
- reduce routing metadata and enforce a catalog budget
- validate Agent Skills and Claude manifests in CI
- remove wooyun-legacy from the package and provenance
- update package metadata and tests for the 54-skill catalog
Adds a design rubric for detecting generic AI output, with full
UPSTREAMS.json provenance. Updates skill count from 11 to 12 across
package metadata, README, and tests.
- write reports and snapshots atomically, cleaning temp files on failure
- isolate CLI run tests from repository artifacts via env overrides
- cover version-detection failures, unsupported modes, manifest bounds,
redaction and grading edges, and malformed mutations with tests
- document path overrides and completed plan steps
feature-dev, code-architect, and mcp-builder gain a shared "Working
discipline" section (think before acting, simplicity first, surgical changes,
goal-driven). code-review and code-reviewer get a lighter touch — a "scope
creep" criterion added to their review checklists. security-review and
skill-creator already embody this discipline, so they are not modified, and
frontend-design is left alone because "simplicity first" would clash with the
skill's deliberate bold/maximalist direction.
The five command files are regenerated from their skills (in sync). A new
tests/working-discipline.test.mjs guards the additions so the rule cannot
silently regress.
The skill told the model to "edit whichever rules file the project already
uses." Under OpenCode's first-match-wins resolution, a CLAUDE.md co-located
with an AGENTS.md is never read, so learnings captured there would be silently
lost. The skill now describes first-match-wins and directs capture to
AGENTS.md when both exist (never the shadowed CLAUDE.md), reinforced in the
"where each addition belongs" step. A test guards it; command kept in sync.
The skill already warned against converging on common choices, but the new
font and layout lists risk becoming the next default the way Space Grotesk
did. This adds an explicit "vary across generations" rule — treat the lists
as seeds to rotate, deliberately diverge when a brief resembles a prior one,
and never ship two different products as the same template. A matching
self-critique check rejects output that reads as a default pattern or a
previous generation. A test guards the rule; the command is regenerated from
the skill.
The skill was pure aesthetic prose with no self-review. This adds three
sections, reworded from cited sources and kept under 500 lines:
- Concrete anti-default moves: font-by-vibe picks, layout archetypes, type
contrast, semantic CSS-variable tokens, one orchestrated page-load.
- An objective self-critique rubric the model runs before returning (WCAG 2.2
AA contrast/semantics/alt/focus/reflow plus craft checks), framed honestly
since models tend to over-praise their own output.
- An optional, cross-platform validation note (html-validate static; pa11y or
axe for rendered DOM), including an opt-in vision-only Playwright screenshot
loop that text-only models skip.
Adds tests/frontend-design.test.mjs to guard the additions. The command is
regenerated from the skill and stays in sync.
The "Critical: only HIGH-SIGNAL issues should be flagged" block appeared
twice, back to back, in code-review's SKILL.md and command. The second copy
was a strict subset of the first (it dropped the "Concrete edge cases"
bullet), so it was pure redundancy. The duplicate is removed; the complete
first copy stays.
The skill-lint suite gains a per-skill "no duplicated block" check (no
non-empty line >= 50 chars repeats within a SKILL.md or its command), which
catches this class of copy-paste error. SKILL.md and command stay in sync.
A generation step had truncated nine command descriptions at ~200
characters, cutting them mid-sentence. Each commands/*.md description is
restored to match its SKILL.md source, so the slash-command menu shows the
full text and skill and command stay in sync.
Adds a zero-dependency node:test suite (tests/skill-lint.test.mjs) that
enforces the invariants going forward: directory name matches the
frontmatter name and slug regex, description length within 1..1024 and not
truncated, SKILL.md body under 500 lines, and each command inlines its
SKILL.md body, shares its description, and ends with $ARGUMENTS. A GitHub
Actions workflow runs the suite on every push and pull request.