Ships supply-chain content integrity for the SkillKit ecosystem
(phase 1 of issue #90):
- New @skillkit/core integrity module computes a deterministic SHA-256
digest over every file in a skill (excluding .skillkit.json metadata,
VCS artifacts, editor noise) and emits a Subresource Integrity string
(sha256-<base64>). 50 MB safety cap.
- New `skillkit verify <path>` CLI with --expected, --against-lock,
--files, --format sri|hex, --json.
- `skillkit publish` now writes an `integrity` field per skill into
.well-known/skills/index.json.
- `skillkit install` and `skillkit update` record the full SRI digest
into ~/.skillkit/lock.json (graceful fallback to legacy short
checksum if integrity computation fails).
- Docs: fumadocs security.mdx retitled "Security & Integrity" with a
full verify section; commands.mdx + index.mdx updated; website
Commands.tsx + Features.tsx surface the new feature and comparison
matrix row.
- Root CHANGELOG.md added.
Phase 1 (content hashing) is the response to the ClawHavoc Q1 2026
supply-chain campaign and OWASP Agentic Skills Top 10 guidance.
Phase 2 (ed25519 signing + publisher key registry) is tracked
separately.
Tests: 11 unit tests for the integrity module, 5 CLI E2E tests for
`skillkit verify`. All 25 task suites green; typecheck clean.
Version bump
------------
1.22.1 -> 1.23.0 across all workspace packages. Minor bump covers the
slim-install dependency layout, safer first-time agent picker, and the
corresponding doc + branding changes.
Prune dead optional deps in @skillkit/core
------------------------------------------
@anthropic-ai/sdk, @google/genai, ollama, and openai were listed as
optionalDependencies but never imported. They were pulled through
@skillkit/cli transitively on every default install, adding ~57
packages and one deprecation warning (node-domexception via the
SDKs' own formdata-node / node-fetch transitives). Grep of
packages/core/src confirms no SDK import; providers in
src/ai/providers use direct fetch() against the provider HTTP APIs.
Drop them. better-sqlite3, sqlite-vec, and node-llama-cpp stay
because they are actually dynamic-imported by the vector store and
embedding pipeline.
Measured on a fresh npm cache (overrides-pinned file: tarballs):
before: 475 pkgs / 18s / 3 deprecation warnings
1.23.0: 418 pkgs / 18s / 2 deprecation warnings
slim: 118 pkgs / 9s / 0 deprecation warnings
The two remaining warnings are upstream:
- prebuild-install@7.1.3 from better-sqlite3 (must stay)
- glob@9.3.5 from @opentui/solid -> babel-plugin-module-resolver
README rewrite
--------------
Punchier top fold. Palette-consistent SVG tag badges under
assets/tags/ replace most shields.io badges (still keep npm/CI/etc
on shields since they need live data). Collapsible detail blocks
for every feature instead of the linear wall. Same nav anchors the
site uses. No em dashes. No AI filler words. Writing style pulls
from the ai-engineering-from-scratch and awesome-claude-design
READMEs the user already runs.
Branded SVG tags
----------------
scripts/gen-tags.mjs generates 20 SVG pills under assets/tags/ in
the SkillKit palette (#09090b base, #fafafa text, #3f3f46 border,
accent per tag). Two variants: single pill with left accent bar
(new, slim-install, v1.23.0, etc.) and split badge (agents | 45,
packages | 118, install | 9s). Matches the aesthetic used on
awesome-claude-design, tailored to SkillKit's monochrome brand.
The tags directory is fork-and-remix friendly; README points to it.
Two nits from CodeRabbit on the first-user picker:
1. detectAgent() returns 'universal' as a fallback when no concrete
adapter is detected. Passing that through to quickAgentSelect made
the picker show "Just Universal (detected)" even when no real
agent was found — misleading. Treat 'universal' as "not detected"
and fall through to the plain select/all options instead.
2. The "All detected agents" label on the bottom option was wrong:
agents at the call site is getAllAdapters() (all supported
adapters), not only detected ones. Rename to "All supported
agents" and change the hint to "writes to every adapter" so
users aren't tricked into thinking this is a safe detected-only
install.
Before: first-time users (no history) saw "All detected agents"
as the first option in the agent picker. Mashing Enter would
install to all 32 agents, writing files across every adapter's
config dir. Selecting "Select specific agents" also pre-ticked
all 32 agents, requiring the user to un-tick by hand.
After:
- Default option is "Just <detected> (detected)" — one agent,
low blast radius.
- "Select specific agents" moves above "All detected agents"
and pre-ticks only the detected agent.
- "All detected agents" gets a second confirm gate
("Install to all N agents?") with initialValue: false.
- Hint text on "All" warns "use only if you know".
Returning users still see "Same as last time (Recommended)" as
the first option; their flow is unchanged.
Non-TTY path is unchanged (still auto-detects a single agent).
Wire detectAgent() through quickAgentSelect so the picker can
label and pre-select the detected agent. Empty selection from
the confirm gate is treated as cancellation.
Move @skillkit/api, @skillkit/tui, @skillkit/mesh, @skillkit/messaging
to optionalDependencies. Drop @skillkit/memory entirely — the CLI
never imported it (only @skillkit/mcp-memory consumes it).
Killing @skillkit/memory removes cozo-node (native Rust) and
@xenova/transformers (~300MB ML stack) from the critical install
path. Making @skillkit/tui optional removes @opentui/core's
prebuild-install toolchain, which was the source of the
inflight/gauge/npmlog/tar/glob deprecation warnings.
Add skipNodeModulesBundle + platform=node + target=node18 to
apps/skillkit tsup config so optional deps stay external.
All TUI/mesh/messaging/api commands already lazy-import via
dynamic import() with surrounding try/catch, so missing optional
deps surface as a graceful error rather than a crash.
Measured impact (npm i skillkit in a clean cache):
Scenario Packages Time Deprecations Vulns (crit/high)
main 544 37s 10 33 (5/5)
slim (default) 475 18s 3 25 (0/0)
slim --omit=optional 118 9s 0 0
Core commands (list, find, translate, recommend, install, --help,
--version) smoke-tested in all three scenarios. Optional commands
(ui, mesh, message, serve) confirmed to degrade with a clean
"Cannot find package" error under --omit=optional.
Add --json flag with structured JSON output to: create, enable, disable,
eval, sync, tap (add/remove/list), translate, save, read.
Add loading spinners to: audit, fix, test, plugin, team, methodology,
hook, remove. Commands that already had spinners get --json-aware no-op
spinners when --json is active to suppress visual output.
- action.yml: pass inputs via env vars, whitelist allowed commands
- All --json commands: use no-op spinner to prevent stdout pollution
- remove: run AGENTS.md update before JSON output (was skipped)
- eval: gate spinner on json format, wrap in try/catch for cleanup
- scan: wrap spinner in try/catch for cleanup on failure
- Fix "44 agents" → "45 agents" across 34 files (website, docs,
package.json, tests, extension, API, TUI)
- README: add tap, check, update, publish submit to commands section
- Commands.tsx: add Sources category with tap commands, update manage
group with remove --source/--all and check
- commands.mdx: full docs for tap, check, update, remove, publish
submit with all new flags and behaviors
- configuration.mdx: document lock.json and taps.json
- security.mdx: document trust badges system
- install: use actual installed path (handles .md files) for lock
entry and results instead of reconstructing from skillName
- update: call addSkillToLock after both local and remote updates
so lock.json stays in sync with metadata
- update: --force bypasses GitHub activity and checksum skip-checks
- update: timestamp comparison falls back to installedAt when
updatedAt is not set
- Use finally for temp cleanup so user cancels don't leak cloned dirs
- Move addSkillToLock outside inner agent loop, merge agents array
- Use source:name as lock key to avoid cross-source collisions
- Gate GitHub activity check on provider.type (skip for GitLab/Bitbucket)
- Normalize source URL to owner/repo before official check
- Include skill name in trust badge output
- Atomic lock.json writes via tmp file + renameSync
- Remove stale --federated suggestion from find no-results message
Introduce tap add/remove/list commands to manage custom skill sources
stored in ~/.skillkit/taps.json. Taps are included in find/search
results, allowing users to add private or community repositories.
Introduce ~/.skillkit/lock.json with LockEntry/LockFile schemas to
track all installed skills in a single file. Install writes entries,
remove deletes them. Enables future team sync and reproducible setups.
Compute SHA-256 checksum of SKILL.md on install and store in metadata.
During update, check GitHub pushedAt timestamp and compare checksums
to skip unnecessary re-downloads when content hasn't changed.
Show trust badges for each skill during install: [Official] for known
official sources (Anthropic, Vercel, Expo, Remotion, Supabase, Stripe),
or [Trusted]/[Review]/[Caution] based on TrustScorer analysis.
- Hoist cloneResult above try block so catch can clean up temp dir
on any failure path, not just the happy path
- Always include GitHub API in find search results (was behind
--federated flag, now default)
publish submit:
- Runs quality eval + security scan before submission
- Shows quality badge, stars, last push date, and warnings
- Creates issue directly via gh CLI (no browser needed)
- Falls back to manual URL if gh is not installed/authenticated
- Issue body includes quality score, stars, and push date
check:
- Fetches pushed_at from GitHub API for real update detection
- Shows quality badges next to skill names
- Shows stars and push date in verbose mode
- Displays average quality score across all skills
- findSkill: fall through to .md check when parseSkill returns null
- Use isPathInside instead of dir.includes(cwd) for location detection
- Optional chaining on meta.source in remove --source filter
- Track fence delimiter type per CommonMark (backtick vs tilde)
- Wrap statSync in try-catch so one bad source doesn't abort all installs
- Add cleanup fallback when sparse-checkout/checkout fails
Standalone .md skills (e.g. doc_explanation.md) now get metadata saved
as .doc_explanation.skillkit.json in the same directory. This allows
remove --source to find and remove them. Updated saveSkillMetadata,
loadSkillMetadata, loadMetadata, and remove command to handle both
directory and sibling metadata patterns.
- skillkit remove --all: removes all installed skills
- skillkit remove --source iii-hq/iii: removes all skills from that source
- Both check .skillkit.json metadata to match source
getSearchDirs now collects skill directories from ALL agent adapters
instead of only the configured agent. This fixes 'skill not found' when
skills were installed to .claude/skills/ but the default agent is
universal (which only searches skills/).
findSkill now also checks name.md for standalone .md skills that were
installed as plain files.
Move isStandaloneFile detection and .md suffix normalization before
the existence/force/symlink checks so all code paths use the same
effective target path. Fixes --force not detecting previously installed
standalone skills and symlinks pointing to the wrong filename.
Instead of wrapping standalone .md skills in a directory, copy them
as-is with .md extension. Skip saveSkillMetadata for standalone files
since it requires a directory. This preserves the original format that
repos like iii-hq/iii use for lightweight single-file skills.
Standalone .md skills (e.g. doc_explanation.md) were copied as plain
files, then saveSkillMetadata tried to write .skillkit.json inside
them causing ENOTDIR. Now detects file sources via statSync and
creates a directory with SKILL.md inside instead.
Replace silent execFileSync with async spawn + --progress flag so the
spinner shows real download progress (e.g. "Receiving objects 45%").
For large monorepos, try partial clone (--filter=blob:none) with sparse
checkout of SKILL_DISCOVERY_PATHS first, falling back to regular
--depth 1 clone if the server doesn't support it. Also adds
--single-branch --no-tags to all clones.
Shared cloneRepo() helper in base.ts eliminates duplicate clone logic
across all four git providers.
- selectSkills/selectAgents now return { result, exitCode } to
distinguish errors (exit 1) from cancellations (exit 0)
- Fixes: --skill with nonexistent name returned 0 instead of 1
- Fixes: --agent with invalid name returned 0 instead of 1
- ai.ts: s.stop('msg', 1) → s.stop(colors.error('msg'))
(clack spinner.stop only accepts message, not exit code)
generate.ts and manifest.ts now warn when --agents (deprecated) is
used instead of --agent. Message: "Warning: --agents is deprecated,
use --agent instead (--agents will be removed in v2.0)"
Findings 3 (scan.ts stderr) and 4 (translate.ts stderr) skipped:
error() from onboarding goes to stdout via clack — this is the
standard pattern across all migrated commands. Changing individual
commands would break consistency.
* fix: version display, install UX — audit findings
- Fix version showing hardcoded 1.7.6 instead of actual 1.19.2
- Export setVersion/setAgentCount from @skillkit/cli
- Call setVersion(version) and setAgentCount(getAllAdapters().length) in cli.ts
- Change fallback from '1.7.6' to 'dev' so stale values are obvious
- Fix install pre-selecting all skills by default (bad UX)
- Add "Install all / Select specific" prompt before multiselect
- When selecting manually, start with nothing pre-selected
- Apply same fix to both InstallCommand and runInstallFlow()
- Matches existing quickAgentSelect pattern
Fixes found during comprehensive codebase audit.
Related issues: #87, #88, #89, #90, #91, #92, #93, #94
* refactor: extract quickSkillSelect, use getAdapterCount
- Extract duplicated "Install all / Select specific" pattern into
quickSkillSelect() in prompts.ts (mirrors existing quickAgentSelect)
- Both install.ts and runInstallFlow() now use the shared helper
- Add getAdapterCount() to @skillkit/agents to avoid allocating 44
spread-copied objects just to call .length
- Use getAdapterCount() in cli.ts instead of getAllAdapters().length
* feat: add multi-tier eval engine (skillkit eval)
6-tier evaluation system for comprehensive skill assessment:
- Tier 1: LLM quality scoring (G-Eval pattern, 6 dimensions)
- Tier 2: Contradiction detection (formal rules + LLM semantic)
- Tier 3: Behavioral security (code analysis + taint tracking + LLM)
- Tier 4: Sandbox execution testing (Docker, graceful skip)
- Tier 5: Dynamic marketplace benchmarks (percentile ranking)
- Tier 6: Community signals (freshness, metadata, content health)
Works without API keys (heuristic fallback for Tier 1, Tiers 5-6).
Full LLM evaluation with --provider flag.
* fix: address Devin review — input validation for eval CLI
- --min-score NaN bypass: validate parseInt result before comparison
- --timeout NaN propagation: skip NaN timeout instead of passing to Docker
- --tier invalid input: graceful error message instead of unhandled throw
- scoreContentSize formula: cap sub-optimal at 99 to distinguish from optimal
* fix: address CodeRabbit review findings on eval system
- Harden contradiction/security prompts against prompt injection by
wrapping untrusted content in <skill_content> XML tags and adding
explicit "treat as untrusted data" instructions
- Use head+tail content sampling in security prompt instead of head-only
truncation to catch malicious payloads at end of content
- Add missing Tier 4 sandbox formatter to reporter.ts
- Fix misleading sampleSize fallback in dynamic-benchmark.ts
- Move --min-score NaN validation before engine.evaluate() to avoid
wasting LLM calls on invalid input
- Add truncation indicator to quality-cot.ts prompts
- Add term length limits (>100 chars) in contradiction.ts to mitigate
ReDoS risk on dynamically constructed regex
* fix: address round 2 review findings on eval system
- Wrap quality-cot prompts with <skill_content> XML tags and untrusted
data instructions, consistent with contradiction/security prompts
- Change truncateContent to head+tail sampling (4000+2000) instead of
head-only, so completeness/token-efficiency evaluators see both
beginning and end of long skills
- Fix escapeXmlTags regex to handle whitespace variants before closing
angle bracket (</skill_content >, </skill_content\n>)
- Type DIMENSION_PROMPTS with Record<EvalDimension, PromptBuilder> for
compile-time key validation
- Clamp estimatePercentile return to [0,100] range defensively
- Make buildStatsFromMarketplace async with periodic yields (every 50
skills) to avoid blocking the event loop on first-run cache miss
- Extract parsedTimeout local variable in eval.ts to avoid double
parseInt parsing
* fix: address round 3 Devin + CodeRabbit findings on eval system
- Scan both code blocks AND full content for dangerous patterns in
behavioral-security.ts, with deduplication by description+snippet
to avoid double-reporting (Devin: security scanner skipped prose)
- Use non-greedy regex /\[[\s\S]*?\]/ in parseSemanticFindings to
match first complete JSON array instead of greedily spanning to
last bracket in LLM response (Devin: greedy regex corrupted JSON)
- Fix formatSandboxDetails to use r.testCase matching SandboxResult
type instead of r.name which doesn't exist (Devin: property mismatch)
- Make MarketplaceSkill.description optional to match marketplace
schema where only id/name/source/tags are required (CodeRabbit)
- Rename sandbox test case from "code example execution" to "code
block content validation" since it validates content not execution
(CodeRabbit)
* fix: address round 4 Devin + CodeRabbit findings on eval system
Inline fixes:
- Replace greedy JSON regex in behavioral-security.ts runLLMAnalysis
with balanced bracket extraction that correctly handles ] inside
JSON string values
- Replace non-greedy regex in contradiction.ts parseSemanticFindings
with same balanced bracket extraction (non-greedy also breaks on
inner brackets)
- Fix crossValidatedCount in behavioral-security.ts to accumulate
(+= uniqueEngines.size) instead of replacing (Math.max)
- Fix gradeDeterministic: cleanExit && hasOutput without matching
expectedOutcome now returns passed:false score:50 instead of
passed:true score:75
- Docker unavailable returns score:-1 sentinel instead of 0/F;
engine.ts filters score<0 tiers from overall average so
Docker-less machines aren't penalized
Duplicate fix:
- Add score field to SandboxResult type; store gradeResult.score
in each result; compute tier score from average of numeric scores
instead of boolean pass rate
Nitpick fixes:
- Distinguish "Docker unavailable" from "No tests executed" in
formatSandboxDetails
- Tighten base64 obfuscation regex to require 4-char groups
- Only push 'llm' to engines when LLM actually produced findings
- Bound \s+ to \s{1,10} in dynamically constructed regexes in
contradiction.ts to prevent catastrophic backtracking
- Validate cache timestamp with Number.isFinite() in loadCache to
reject corrupted cache files with NaN timestamps
* feat: integrate skills.sh as first-class registry
Add SkillsShRegistry and SkillsShProvider so users can install skills
directly from skills.sh (e.g. `skillkit install skills.sh/owner/repo/skill`).
The find command now searches skills.sh by default alongside the local
marketplace. Post-install shows skills.sh popularity stats and pro tips
highlighting skillkit's unique features (translate, recommend, evaluate).
* fix: address CodeRabbit and Devin review findings for skills.sh integration
- Fix path traversal vulnerability in SkillsShProvider.clone (validate subpath containment)
- Fix federated search skipped for interactive queries (use searchTerm instead of this.query)
- Guard showSkillsShStats to skip network call for local providers
- Replace deprecated npm --production with --omit=dev
- Fix loose includes() match in getSkillStats to use strict endsWith()
- Fix FederatedSearch resilience: don't re-throw RateLimitError, return partial results
- Forward timeoutMs through FederatedSearch to registries
- Extract shared formatCount utility to helpers.ts (deduplicate find.ts/install.ts)
- Fix install hint to use skill.source instead of fragile skill.description
- Fix timer leak in fetchLeaderboard using finally block
- Deduplicate URL parsing via shared parseSkillsShParts helper
- Validate --agent flag values against known agent types