Compare commits

...
Author SHA1 Message Date
Peter Steinberger c1e6b985d9 fix: add root undici devDependency for e2e (#255) (thanks @tanujbhaud) 2026-02-13 15:23:00 +01:00
Tanuj Bhaud 7fcbcd345a fix(vt): explicit return types and missing undici dependency
Refactor action handlers in convex/vt.ts to use explicit return types, resolving circular type inference (TS7022). Also add undici to devDependencies for E2E tests.
2026-02-13 15:22:07 +01:00
Peter SteinbergerandSash Zats ddddb431c2 fix: make /search host-aware in SSR (#257)
* fix: make /search mode-aware

Notes:\n- Medium: /search now depends on getSiteMode() during beforeLoad. On server-side routing, if VITE_SITE_MODE isn’t set and VITE_SOULHUB_SITE_URL is set (as in .env.local), getSiteMode() will resolve to souls and redirect /search to / even on the ClawdHub deployment. This is a regression risk vs the old always-/skills redirect. Confirm deployment envs guarantee correct mode. src/routes/search.tsx:9-31

* fix: make /search host-aware in SSR

* chore: fix lint and route tree for /search route

---------

Co-authored-by: Sash Zats <sash@zats.io>
2026-02-13 14:26:31 +01:00
David AronchickandPeter Steinberger 78c27579a3 fix(cli): secure config file permissions (#164)
* fix(cli): secure config file permissions and reduce duplication

Security:
- Config files now created with 0600 permissions (owner read/write only)
- Config directories created with 0700 permissions
- Protects API tokens from other users on shared systems

Maintainability:
- Extract resolveConfigPath() helper to reduce code duplication
- Same legacy fallback logic (clawhub -> clawdhub) now in one place

* fix(cli): tolerate unsupported chmod errors for config

---------

Co-authored-by: Peter Steinberger <steipete@gmail.com>
2026-02-13 13:57:03 +01:00
xcqtnrandPeter Steinberger 9aebc35d86 fix: prevent infinite loading loop on skills page (#90)
* fix: prevent infinite loading loop on skills pageAdd isLoadingMore guard to IntersectionObserver useEffect to preventcontinuous WebSocket queries when user is idle at bottom of page.The observer now won't set up while a request is in progress, breakingthe infinite loop cycle.Fixes: Related to #89

* fix: prevent repeated skills auto-load requests (#90) (thanks @xcqtnr)

* fix: resolve PR merge conflicts and keep observer regression test (#90) (thanks @xcqtnr)

---------

Co-authored-by: Peter Steinberger <steipete@gmail.com>
2026-02-13 05:44:36 +01:00
Peter Steinberger fc63f47ffa chore(release): 0.6.1 2026-02-13 05:14:16 +01:00
Gaurav SharmaandPeter Steinberger 0b83ea6ff3 fix: prevent horizontal overflow from long code blocks in skill pages (#183)
* Fix: Prevent horizontal overflow from long code blocks in skill pages

- Add max-width: 100% to .file-list-body and .file-row
- Prevents page-wide overflow when skills contain long code examples
- Markdown pre blocks already have overflow-x: auto, but parent containers were expanding infinitely
- Fixes issue where skills with 400+ char lines (e.g. browser automation commands) cause horizontal scrolling

Affected: Skills with long inline code in markdown (browser act commands, etc.)

* fix: add max-width to .file-list container to prevent overflow

- Also ensures .file-list-body constraint is inherited properly
- Prevents long code blocks from expanding file list container

* fix: add max-width to all markdown containers and pre tags

- Add max-width: 100% to .markdown, .tab-body, .markdown pre
- Ensures code blocks are constrained and show horizontal scrollbar
- Prevents content from expanding parent containers beyond viewport

* fix: add overflow-x to parent containers for horizontal scroll

Adds overflow-x: auto to .skill-detail-stack, .tab-card, and .tab-body
to ensure long code blocks are scrollable within the content area
instead of causing page-wide horizontal overflow.

Fixes horizontal overflow issue on skill pages with long code examples
(e.g., browser automation commands with 400+ character lines).

Tested on zepto skill page - page now stays within viewport (1200px)
and code blocks are accessible via horizontal scrollbar in tab area.

* docs: note code-block overflow fix in changelog (#183) (thanks @bewithgaurav)

---------

Co-authored-by: Peter Steinberger <steipete@gmail.com>
2026-02-13 05:09:31 +01:00
LimitlessandLimitless2023 191b5763ec fix: include comment deltas in action-based stat processing & add stats reconciliation (#194)
Bug 1: applyAggregatedStatsAndUpdateCursor was missing 'comments' in both
the guard condition and the applySkillStatDeltas call. This caused comment
count deltas to be silently dropped during cron-based event processing,
while stars/downloads/installs were processed correctly.

Bug 2: No reconciliation mechanism existed. If events were missed due to
cursor issues or processing errors, skill stats (stars, comments) would
remain stale with no way to recover. Added reconcileSkillStarCounts
maintenance mutation that counts actual records in the stars and comments
tables and patches any out-of-sync skill stats.

Fixes #193

Co-authored-by: Limitless2023 <limitless@users.noreply.github.com>
2026-02-13 04:55:16 +01:00
Peter Steinberger 5397a8e5e0 fix: scope reauth fix; keep banned users blocked (#177) (thanks @tanujbhaud) 2026-02-13 04:34:46 +01:00
Tanuj Bhaud 7949402888 test: add missing coverage for fresh-login reactivation and identity mismatch guard 2026-02-13 04:34:46 +01:00
Tanuj Bhaud 2e3920d41a fix: use valid crons.interval and set to 1 minute 2026-02-13 04:34:46 +01:00
Tanuj Bhaud dd4fc823f6 fix: allow re-auth when existingUserId is null 2026-02-13 04:34:46 +01:00
Tanuj Bhaud 135b9ea9b0 fix: ensure reactivation only matches soft-deleted user (prevents bypass) 2026-02-13 04:34:46 +01:00
Tanuj Bhaud 17a106cefe fix: resolve final lint error in auth tests 2026-02-13 04:34:46 +01:00
Tanuj Bhaud 496da99392 fix: update tests to include required existingUserId parameter 2026-02-13 04:34:46 +01:00
Tanuj Bhaud 107486adfb fix: restore existingUserId check for type safety 2026-02-13 04:34:46 +01:00
Tanuj Bhaud ef9e7f0e57 test: update auth tests for direct deletedAt check 2026-02-13 04:34:46 +01:00
Tanuj Bhaud ecc8ad3833 fix: allow soft-deleted users to re-authenticate
Fixes Issue #32 where users who soft-deleted their accounts were unable to sign back in because the re-auth logic was only triggering when an existingUserId was passed by the auth provider, which doesn't happen during a standard fresh login flow.
2026-02-13 04:34:46 +01:00
Peter Steinberger 93c2b23b72 docs: add 0.6.1 unreleased changelog from post-0.6.0 commits 2026-02-13 04:14:16 +01:00
DCollandPeter Steinberger 2e492a5b87 fix(http): remove allowH2 from undici Agent — causes fetch failed on Node.js 22+ (#245)
* Remove allowH2 option from global dispatcher

fix/remove-allowH2-undici-node22-compat

* fix(http): remove allowH2 from e2e dispatcher

---------

Co-authored-by: Peter Steinberger <steipete@gmail.com>
2026-02-13 03:02:57 +01:00
Peter Steinberger 91c87d1322 test: fix search test handler typing 2026-02-13 02:48:16 +01:00
Peter Steinberger 91b8f160f8 test: add search fallback coverage 2026-02-13 02:44:53 +01:00
Peter Steinberger 32f3ce45e9 fix: add lexical fallback for skill search recall 2026-02-13 02:22:32 +01:00
Peter Steinberger 19bfe48a67 fix: prioritize relevant skills in search 2026-02-13 02:15:19 +01:00
Peter Steinberger 19951fccf7 docs: thank @superlowburn for PR #246 2026-02-13 01:22:38 +01:00
7dcada9122 fix: handle GitHub API rate limits in account age check (#246)
* fix: handle GitHub API rate limits in account age check

The GitHub account lookup uses unauthenticated requests (60 req/hr
per IP). Since this runs server-side in Convex, all users share the
same IP and quickly exhaust the rate limit, causing "GitHub account
lookup failed" errors during skill publish.

- Detect 403/429 responses and surface a clear rate-limit message
- Support optional GITHUB_TOKEN env var for authenticated requests
  (5,000 req/hr)

Fixes #155

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* fix: stabilize GitHub account gate tests and docs

---------

Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Co-authored-by: Peter Steinberger <steipete@gmail.com>
2026-02-13 01:21:30 +01:00
theonejvo dab307cb6d fix: VT scan sync race condition + LLM-first moderation model
VT no longer overwrites LLM moderation verdicts. LLM is the primary
moderation authority; VT only escalates (hides + flags) for malicious/
suspicious content via new escalateByVtInternal mutation. Stale VT polls
write vtAnalysis marker instead of overwriting moderationReason. Query
pools expanded to include LLM-evaluated skills awaiting VT results.
Ban message now references malicious skills and security@openclaw.ai.
2026-02-13 01:05:52 +11:00
vignesh07 e96eb4781c chore: fix review comments 2026-02-11 10:38:26 -08:00
Vignesh f64b098fcc perf: lazy-load diff viewer (Monaco) (#212) 2026-02-11 12:31:07 -06:00
Vignesh 243432e04e chore: fix lint issues (#213) 2026-02-11 12:31:01 -06:00
theonejvo 9f7b9b92cd fix: trailing comma tolerance in JSON metadata, tone down persistence flags
- Strip trailing commas in frontmatter JSON before parsing (silent failure fix)
- Stop flagging disable-model-invocation default as a concern (it's the normal default)
- Stop flagging skills configuring themselves as privilege escalation
- Add MITRE ATLAS AML.T0051 context for when autonomous invocation actually matters
- Show actual defaults in assembled eval message instead of "not set"
2026-02-11 18:56:19 +11:00
theonejvo 3402f0e735 feat: add skill metadata docs, suspicious appeal banner for owners
- Document full frontmatter metadata reference in docs/skill-format.md
- Add metadata section + quick example to README
- Show appeal message on suspicious skills (owner-only) linking to GitHub issues
- Accept metadata.openclaw alias in README docs
- Re-evaluate all skills with full file content reading (backfill in progress)
2026-02-11 18:36:45 +11:00
theonejvo 3e39651074 feat: evaluator reads all file contents, not just SKILL.md
Reads all files from storage and includes their full source in the eval
prompt so the LLM can detect malicious code hidden behind clean READMEs.
Injection detection now scans all content. Per-file cap 10K chars, total
cap 50K chars.
2026-02-11 17:23:19 +11:00
theonejvo 741301848f fix: eval assembler falls back to metadata.openclaw for requirements 2026-02-11 14:39:02 +11:00
theonejvo b593dedba1 feat: recognize metadata.openclaw as valid frontmatter namespace 2026-02-11 14:29:11 +11:00
theonejvo e104b8030e fix: increase max_output_tokens for reasoning model, fix backfill error retry 2026-02-11 04:48:08 +11:00
theonejvo e74e879e12 fix: add retry with backoff for OpenAI rate limits, fix JSON mode requirement 2026-02-11 02:48:22 +11:00
theonejvo 1f1d93ded9 fix: collapse OpenClaw analysis by default, fix row spacing, switch to gpt-5-mini 2026-02-11 02:33:17 +11:00
theonejvo 9c31462f15 feat: add LLM security evaluation at publish time
Add OpenClaw LLM-based security evaluator that runs alongside VirusTotal
when skills are published. Reads SKILL.md prose, metadata, install specs,
and file manifest, then assesses coherence across 5 dimensions to catch
social engineering vectors that VT/regex miss (e.g. instruction-only skills
with no code files).

- convex/lib/securityPrompt.ts: system prompt, message assembly, response
  parsing, injection pattern detection
- convex/llmEval.ts: evaluateWithLlm action, evaluateBySlug convenience
  action, backfillLlmEval for existing skills
- convex/schema.ts: llmAnalysis field on skillVersions
- convex/skills.ts: updateVersionLlmAnalysisInternal mutation,
  getActiveSkillBatchForLlmBackfillInternal query, defense-in-depth
  multi-scanner flag merging in approveSkillByHashInternal
- convex/lib/skillPublish.ts: schedule LLM eval alongside VT scan
- SkillDetailPage.tsx: OpenClaw row, LlmAnalysisDetail expandable
  component with 5 dimension rows, guidance panel, findings section
- styles.css: analysis detail styles from mockup
2026-02-11 02:19:03 +11:00
41 changed files with 3123 additions and 146 deletions
+13 -2
View File
@@ -1,15 +1,26 @@
# Changelog
## Unreleased
## 0.6.1 - 2026-02-13
### Added
- Security: add LLM-based security evaluation during skill publish.
- Parsing: recognize `metadata.openclaw` frontmatter and evaluate all skill files for requirements.
### Changed
- Performance: lazy-load Monaco diff viewer on demand (thanks @alexjcm, #212).
- Search: improve recall/ranking with lexical fallback and relevance prioritization.
- Moderation UX: collapse OpenClaw analysis by default; update spacing and default reasoning model.
### Fixed
- Upload gate: handle GitHub API rate limits and optional authenticated lookup token (thanks @superlowburn, #246).
- HTTP: remove `allowH2` from Undici agent to prevent `fetch failed` on Node.js 22+ (#245).
- Tests: add root `undici` dev dependency for Node E2E imports (thanks @tanujbhaud, #255).
- VirusTotal: fix scan sync race conditions and retry behavior in scan/backfill paths.
- Metadata: tolerate trailing commas in JSON metadata.
- Auth: allow soft-deleted users to re-authenticate on fresh login, while keeping banned users blocked (thanks @tanujbhaud, #177).
- Web: prevent horizontal overflow from long code blocks in skill pages (thanks @bewithgaurav, #183).
## 0.6.0 - 2026-02-10
### Added
- CLI/API: add `set-role` to change user roles (admin only).
- Security: quarantine skill publishes with VirusTotal scans + UI (thanks @aleph8, #130).
+24 -1
View File
@@ -138,7 +138,30 @@ metadata: {"clawdbot":{"cliHelp":"padel --help\\nUsage: padel [command]\\n"}}
---
```
`metadata.clawdbot` is preferred, but `metadata.clawdis` is accepted as an alias for compatibility.
`metadata.clawdbot` is preferred, but `metadata.clawdis` and `metadata.openclaw` are accepted as aliases.
## Skill metadata
Skills declare their runtime requirements (env vars, binaries, install specs) in the `SKILL.md` frontmatter. ClawHub's security analysis checks these declarations against actual skill behavior.
Full reference: [`docs/skill-format.md`](docs/skill-format.md#frontmatter-metadata)
Quick example:
```yaml
---
name: my-skill
description: Does a thing with an API.
metadata:
openclaw:
requires:
env:
- MY_API_KEY
bins:
- curl
primaryEnv: MY_API_KEY
---
```
## Scripts
+8 -3
View File
@@ -57,13 +57,14 @@
"oxlint": "^1.42.0",
"oxlint-tsgolint": "^0.11.4",
"typescript": "^5.9.3",
"undici": "^7.19.2",
"vite": "^7.3.1",
"vitest": "^4.0.18",
},
},
"packages/clawdhub": {
"name": "clawhub",
"version": "0.5.0",
"version": "0.6.1",
"bin": {
"clawhub": "bin/clawdhub.js",
"clawdhub": "bin/clawdhub.js",
@@ -1298,7 +1299,7 @@
"ufo": ["ufo@1.6.3", "", {}, "sha512-yDJTmhydvl5lJzBmy/hyOAA0d+aqCBuwl818haVdYCRrWV84o7YyeVm4QlVHStqNrrJSTb6jKuFAVqAFsr+K3Q=="],
"undici": ["undici@7.19.2", "", {}, "sha512-4VQSpGEGsWzk0VYxyB/wVX/Q7qf9t5znLRgs0dzszr9w9Fej/8RVNQ+S20vdXSAyra/bJ7ZQfGv6ZMj7UEbzSg=="],
"undici": ["undici@7.20.0", "", {}, "sha512-MJZrkjyd7DeC+uPZh+5/YaMDxFiiEEaDgbUSVMXayofAkDWF1088CDo+2RPg7B1BuS1qf1vgNE7xqwPxE0DuSQ=="],
"undici-types": ["undici-types@7.16.0", "", {}, "sha512-Zz+aZWSj8LE6zoxD+xrjh4VfkIG8Ya6LvYkZqtUQGJPZjYl53ypCaUwWqo7eI0x66KBGeRo+mlBEkMSeSZ38Nw=="],
@@ -1402,8 +1403,12 @@
"cheerio/parse5": ["parse5@7.3.0", "", { "dependencies": { "entities": "^6.0.0" } }, "sha512-IInvU7fabl34qmi9gY8XOVxhYyMyuH2xUNpb2q8/Y+7552KlejkRvqvD19nMoUW/uQGGbqNpA6Tufu5FL5BZgw=="],
"cheerio/undici": ["undici@7.19.2", "", {}, "sha512-4VQSpGEGsWzk0VYxyB/wVX/Q7qf9t5znLRgs0dzszr9w9Fej/8RVNQ+S20vdXSAyra/bJ7ZQfGv6ZMj7UEbzSg=="],
"cheerio/whatwg-mimetype": ["whatwg-mimetype@4.0.0", "", {}, "sha512-QaKxh0eNIi2mE9p2vEdzfagOKHCcj1pJ56EEHGQOVxp8r9/iszLUUV7v89x9O1p/T+NlTM5W7jW6+cz4Fq1YVg=="],
"clawhub/undici": ["undici@7.19.2", "", {}, "sha512-4VQSpGEGsWzk0VYxyB/wVX/Q7qf9t5znLRgs0dzszr9w9Fej/8RVNQ+S20vdXSAyra/bJ7ZQfGv6ZMj7UEbzSg=="],
"convex/esbuild": ["esbuild@0.27.0", "", { "optionalDependencies": { "@esbuild/aix-ppc64": "0.27.0", "@esbuild/android-arm": "0.27.0", "@esbuild/android-arm64": "0.27.0", "@esbuild/android-x64": "0.27.0", "@esbuild/darwin-arm64": "0.27.0", "@esbuild/darwin-x64": "0.27.0", "@esbuild/freebsd-arm64": "0.27.0", "@esbuild/freebsd-x64": "0.27.0", "@esbuild/linux-arm": "0.27.0", "@esbuild/linux-arm64": "0.27.0", "@esbuild/linux-ia32": "0.27.0", "@esbuild/linux-loong64": "0.27.0", "@esbuild/linux-mips64el": "0.27.0", "@esbuild/linux-ppc64": "0.27.0", "@esbuild/linux-riscv64": "0.27.0", "@esbuild/linux-s390x": "0.27.0", "@esbuild/linux-x64": "0.27.0", "@esbuild/netbsd-arm64": "0.27.0", "@esbuild/netbsd-x64": "0.27.0", "@esbuild/openbsd-arm64": "0.27.0", "@esbuild/openbsd-x64": "0.27.0", "@esbuild/openharmony-arm64": "0.27.0", "@esbuild/sunos-x64": "0.27.0", "@esbuild/win32-arm64": "0.27.0", "@esbuild/win32-ia32": "0.27.0", "@esbuild/win32-x64": "0.27.0" }, "bin": { "esbuild": "bin/esbuild" } }, "sha512-jd0f4NHbD6cALCyGElNpGAOtWxSq46l9X/sWB0Nzd5er4Kz2YTm+Vl0qKFT9KUJvD8+fiO8AvoHhFvEatfVixA=="],
"dom-serializer/entities": ["entities@4.5.0", "", {}, "sha512-V0hjH4dGPh9Ao5p0MoRY6BVqtwCjhz6vI5LT8AJ55H+4g9/4vbHx1I54fS0XuclLhDHArPQCiMjDxjaL8fPxhw=="],
@@ -1412,7 +1417,7 @@
"htmlparser2/entities": ["entities@7.0.1", "", {}, "sha512-TWrgLOFUQTH994YUyl1yT4uyavY5nNB5muff+RtWaqNVCAK408b5ZnnbNAUEWLTCpum9w6arT70i1XdQ4UeOPA=="],
"jsdom/undici": ["undici@7.20.0", "", {}, "sha512-MJZrkjyd7DeC+uPZh+5/YaMDxFiiEEaDgbUSVMXayofAkDWF1088CDo+2RPg7B1BuS1qf1vgNE7xqwPxE0DuSQ=="],
"nitro/undici": ["undici@7.19.2", "", {}, "sha512-4VQSpGEGsWzk0VYxyB/wVX/Q7qf9t5znLRgs0dzszr9w9Fej/8RVNQ+S20vdXSAyra/bJ7ZQfGv6ZMj7UEbzSg=="],
"parse-entities/@types/unist": ["@types/unist@2.0.11", "", {}, "sha512-CmBKiL6NNo/OqgmMn95Fk9Whlp2mtvIv+KNpQKN2F4SjvrEesubTRWGYSg+BnWZOnlCaSTU1sMpsBOzgbYhnsA=="],
+4
View File
@@ -36,6 +36,7 @@ import type * as lib_leaderboards from "../lib/leaderboards.js";
import type * as lib_moderation from "../lib/moderation.js";
import type * as lib_public from "../lib/public.js";
import type * as lib_searchText from "../lib/searchText.js";
import type * as lib_securityPrompt from "../lib/securityPrompt.js";
import type * as lib_skillBackfill from "../lib/skillBackfill.js";
import type * as lib_skillPublish from "../lib/skillPublish.js";
import type * as lib_skillStats from "../lib/skillStats.js";
@@ -46,6 +47,7 @@ import type * as lib_soulPublish from "../lib/soulPublish.js";
import type * as lib_tokens from "../lib/tokens.js";
import type * as lib_userSearch from "../lib/userSearch.js";
import type * as lib_webhooks from "../lib/webhooks.js";
import type * as llmEval from "../llmEval.js";
import type * as maintenance from "../maintenance.js";
import type * as rateLimits from "../rateLimits.js";
import type * as search from "../search.js";
@@ -101,6 +103,7 @@ declare const fullApi: ApiFromModules<{
"lib/moderation": typeof lib_moderation;
"lib/public": typeof lib_public;
"lib/searchText": typeof lib_searchText;
"lib/securityPrompt": typeof lib_securityPrompt;
"lib/skillBackfill": typeof lib_skillBackfill;
"lib/skillPublish": typeof lib_skillPublish;
"lib/skillStats": typeof lib_skillStats;
@@ -111,6 +114,7 @@ declare const fullApi: ApiFromModules<{
"lib/tokens": typeof lib_tokens;
"lib/userSearch": typeof lib_userSearch;
"lib/webhooks": typeof lib_webhooks;
llmEval: typeof llmEval;
maintenance: typeof maintenance;
rateLimits: typeof rateLimits;
search: typeof search;
+35 -3
View File
@@ -29,12 +29,13 @@ function makeCtx({
describe('handleSoftDeletedUserReauth', () => {
const userId = 'users:1' as Id<'users'>
it('skips when no existing user', async () => {
it('skips when user not found', async () => {
const { ctx } = makeCtx({ user: null })
await handleSoftDeletedUserReauth(ctx as never, { userId, existingUserId: null })
await handleSoftDeletedUserReauth(ctx as never, { userId, existingUserId: userId })
expect(ctx.db.get).not.toHaveBeenCalled()
expect(ctx.db.get).toHaveBeenCalledWith(userId)
expect(ctx.db.query).not.toHaveBeenCalled()
})
it('skips active users', async () => {
@@ -57,6 +58,27 @@ describe('handleSoftDeletedUserReauth', () => {
})
})
it('restores soft-deleted users on fresh login (existingUserId is null)', async () => {
const { ctx } = makeCtx({ user: { deletedAt: 123 }, banRecord: null })
await handleSoftDeletedUserReauth(ctx as never, { userId, existingUserId: null })
expect(ctx.db.patch).toHaveBeenCalledWith(userId, {
deletedAt: undefined,
updatedAt: expect.any(Number),
})
})
it('skips reactivation when existingUserId does not match userId', async () => {
const otherUserId = 'users:999' as Id<'users'>
const { ctx } = makeCtx({ user: { deletedAt: 123 } })
await handleSoftDeletedUserReauth(ctx as never, { userId, existingUserId: otherUserId })
expect(ctx.db.query).not.toHaveBeenCalled()
expect(ctx.db.patch).not.toHaveBeenCalled()
})
it('blocks banned users with a custom message', async () => {
const { ctx } = makeCtx({ user: { deletedAt: 123 }, banRecord: { action: 'user.ban' } })
@@ -66,4 +88,14 @@ describe('handleSoftDeletedUserReauth', () => {
expect(ctx.db.patch).not.toHaveBeenCalled()
})
it('blocks banned users on fresh login (existingUserId is null)', async () => {
const { ctx } = makeCtx({ user: { deletedAt: 123 }, banRecord: { action: 'user.ban' } })
await expect(
handleSoftDeletedUserReauth(ctx as never, { userId, existingUserId: null }),
).rejects.toThrow(BANNED_REAUTH_MESSAGE)
expect(ctx.db.patch).not.toHaveBeenCalled()
})
})
+7 -3
View File
@@ -4,17 +4,21 @@ import type { GenericMutationCtx } from 'convex/server'
import { ConvexError } from 'convex/values'
import type { DataModel, Id } from './_generated/dataModel'
export const BANNED_REAUTH_MESSAGE = 'Your account has been suspended.'
export const BANNED_REAUTH_MESSAGE =
'Your account has been banned for uploading malicious skills. If you believe this is a mistake, please contact security@openclaw.ai and we will work with you to restore access.'
export async function handleSoftDeletedUserReauth(
ctx: GenericMutationCtx<DataModel>,
args: { userId: Id<'users'>; existingUserId: Id<'users'> | null },
) {
if (!args.existingUserId) return
const user = await ctx.db.get(args.userId)
if (!user?.deletedAt) return
// Verify that the incoming identity matches the soft-deleted user to prevent bypass.
if (args.existingUserId && args.existingUserId !== args.userId) {
return
}
const userId = args.userId
const banRecord = await ctx.db
.query('auditLogs')
+79 -3
View File
@@ -1,5 +1,5 @@
/* @vitest-environment node */
import { beforeEach, describe, expect, it, vi } from 'vitest'
import { afterEach, beforeEach, describe, expect, it, vi } from 'vitest'
import { internal } from '../_generated/api'
import { requireGitHubAccountAge } from './githubAccount'
@@ -18,6 +18,11 @@ const ONE_DAY_MS = 24 * 60 * 60 * 1000
describe('requireGitHubAccountAge', () => {
beforeEach(() => {
vi.restoreAllMocks()
vi.unstubAllEnvs()
})
afterEach(() => {
vi.unstubAllEnvs()
})
it('uses cached githubCreatedAt when fresh', async () => {
@@ -86,7 +91,9 @@ describe('requireGitHubAccountAge', () => {
expect(fetchMock).toHaveBeenCalledWith(
'https://api.github.com/users/steipete',
expect.objectContaining({ headers: { 'User-Agent': 'clawhub' } }),
expect.objectContaining({
headers: expect.objectContaining({ 'User-Agent': 'clawhub' }),
}),
)
expect(runMutation).toHaveBeenCalledWith(internal.users.updateGithubMetaInternal, {
userId: 'users:1',
@@ -105,11 +112,80 @@ describe('requireGitHubAccountAge', () => {
githubFetchedAt: 0,
})
const runMutation = vi.fn()
const fetchMock = vi.fn().mockResolvedValue({ ok: false })
const fetchMock = vi.fn().mockResolvedValue({ ok: false, status: 404 })
vi.stubGlobal('fetch', fetchMock)
await expect(
requireGitHubAccountAge({ runQuery, runMutation } as never, 'users:1' as never),
).rejects.toThrow(/GitHub account lookup failed/i)
})
it('throws rate-limit error on 403', async () => {
const runQuery = vi.fn().mockResolvedValue({
_id: 'users:1',
handle: 'steipete',
githubCreatedAt: undefined,
githubFetchedAt: 0,
})
const runMutation = vi.fn()
const fetchMock = vi.fn().mockResolvedValue({ ok: false, status: 403 })
vi.stubGlobal('fetch', fetchMock)
await expect(
requireGitHubAccountAge({ runQuery, runMutation } as never, 'users:1' as never),
).rejects.toThrow(/rate limit exceeded/i)
})
it('throws rate-limit error on 429', async () => {
const runQuery = vi.fn().mockResolvedValue({
_id: 'users:1',
handle: 'steipete',
githubCreatedAt: undefined,
githubFetchedAt: 0,
})
const runMutation = vi.fn()
const fetchMock = vi.fn().mockResolvedValue({ ok: false, status: 429 })
vi.stubGlobal('fetch', fetchMock)
await expect(
requireGitHubAccountAge({ runQuery, runMutation } as never, 'users:1' as never),
).rejects.toThrow(/rate limit exceeded/i)
})
it('includes Authorization header when GITHUB_TOKEN is set', async () => {
vi.useFakeTimers()
const now = new Date('2026-02-02T12:00:00Z')
vi.setSystemTime(now)
vi.stubEnv('GITHUB_TOKEN', 'ghp_test123')
const runQuery = vi.fn().mockResolvedValue({
_id: 'users:1',
handle: 'steipete',
githubCreatedAt: undefined,
githubFetchedAt: now.getTime() - 2 * ONE_DAY_MS,
})
const runMutation = vi.fn()
const fetchMock = vi.fn().mockResolvedValue({
ok: true,
json: async () => ({
created_at: '2020-01-01T00:00:00Z',
}),
})
vi.stubGlobal('fetch', fetchMock)
await requireGitHubAccountAge({ runQuery, runMutation } as never, 'users:1' as never)
expect(fetchMock).toHaveBeenCalledWith(
'https://api.github.com/users/steipete',
expect.objectContaining({
headers: {
'User-Agent': 'clawhub',
Authorization: 'Bearer ghp_test123',
},
}),
)
vi.useRealTimers()
})
})
+13 -2
View File
@@ -24,10 +24,21 @@ export async function requireGitHubAccountAge(ctx: ActionCtx, userId: Id<'users'
const stale = !createdAt || now - fetchedAt > FETCH_TTL_MS
if (stale) {
const headers: Record<string, string> = { 'User-Agent': 'clawhub' }
const token = process.env.GITHUB_TOKEN
if (token) {
headers.Authorization = `Bearer ${token}`
}
const response = await fetch(`${GITHUB_API}/users/${encodeURIComponent(handle)}`, {
headers: { 'User-Agent': 'clawhub' },
headers,
})
if (!response.ok) throw new ConvexError('GitHub account lookup failed')
if (!response.ok) {
if (response.status === 403 || response.status === 429) {
throw new ConvexError('GitHub API rate limit exceeded — please try again in a few minutes')
}
throw new ConvexError('GitHub account lookup failed')
}
const payload = (await response.json()) as GitHubUser
const parsed = payload.created_at ? Date.parse(payload.created_at) : Number.NaN
+2
View File
@@ -33,6 +33,8 @@ describe('searchText', () => {
expect(matchesExactTokens(['pad'], ['Padel', '/padel', 'Tennis-like sport'])).toBe(true)
// "xyz" should not match anything
expect(matchesExactTokens(['xyz'], ['GoHome', '/gohome', 'Navigate home'])).toBe(false)
// "notion" should not match "annotations" (substring only)
expect(matchesExactTokens(['notion'], ['Annotations helper', '/annotations'])).toBe(false)
})
it('matchesExactTokens ignores empty inputs', () => {
+1 -1
View File
@@ -20,7 +20,7 @@ export function matchesExactTokens(
if (textTokens.length === 0) return false
// Require at least one token to prefix-match, allowing vector similarity to determine relevance
return queryTokens.some((queryToken) =>
textTokens.some((textToken) => textToken.includes(queryToken)),
textTokens.some((textToken) => textToken.startsWith(queryToken)),
)
}
+499
View File
@@ -0,0 +1,499 @@
export function getLlmEvalModel(): string {
return process.env.OPENAI_EVAL_MODEL ?? 'gpt-5-mini'
}
export const LLM_EVAL_MAX_OUTPUT_TOKENS = 16000
// ---------------------------------------------------------------------------
// Helpers
// ---------------------------------------------------------------------------
function formatScalar(value: unknown): string {
if (value === undefined) return 'undefined'
if (value === null) return 'null'
if (typeof value === 'string') return value
if (typeof value === 'number' || typeof value === 'boolean' || typeof value === 'bigint') {
return String(value)
}
// Avoid throwing on circular structures; fall back to a safe representation.
try {
return JSON.stringify(value)
} catch {
return Object.prototype.toString.call(value)
}
}
function formatWithDefault(value: unknown, defaultLabel: string): string {
if (value === undefined || value === null) return defaultLabel
return formatScalar(value)
}
// ---------------------------------------------------------------------------
// Types
// ---------------------------------------------------------------------------
export type SkillEvalContext = {
slug: string
displayName: string
ownerUserId: string
version: string
createdAt: number
summary?: string
source?: string
homepage?: string
parsed: {
frontmatter: Record<string, unknown>
metadata?: unknown
clawdis?: unknown
}
files: Array<{ path: string; size: number }>
skillMdContent: string
fileContents: Array<{ path: string; content: string }>
injectionSignals: string[]
}
export type LlmEvalDimension = {
name: string
label: string
rating: string
detail: string
}
export type LlmEvalResponse = {
verdict: 'benign' | 'suspicious' | 'malicious'
confidence: 'high' | 'medium' | 'low'
summary: string
dimensions: LlmEvalDimension[]
guidance: string
findings: string
}
// ---------------------------------------------------------------------------
// System prompt (~3500 words)
// ---------------------------------------------------------------------------
export const SECURITY_EVALUATOR_SYSTEM_PROMPT = `You are a security evaluator for OpenClaw AI skills. Users install skills to extend what their AI agent can do. Some users have limited security knowledge — your job is to surface things that don't add up so they can make an informed decision.
You are not a malware classifier. You are an incoherence detector.
A skill is a bundle of: a name, a description, a set of instructions (SKILL.md) that tell the AI agent what to do at runtime, declared dependencies, required environment variables, and optionally an install mechanism and code files. Many skills are instruction-only — just a SKILL.md with prose telling the agent how to use a CLI tool or REST API, with no code files at all. Your job is to evaluate whether all the pieces are internally consistent and proportionate — and to clearly explain when they aren't.
## How to evaluate
Assess the skill across these five dimensions. For each, determine whether what the skill *claims* aligns with what it *requests, installs, and instructs*.
### 1. Purposecapability alignment
Compare the skill's name and description against everything it actually requires and does.
Ask: would someone building this skill legitimately need all of this?
A "git-commit-helper" that requires AWS credentials is incoherent. A "cloud-deploy" skill that requires AWS credentials is expected. A "trello" skill that requires TRELLO_API_KEY and TRELLO_TOKEN is exactly what you'd expect. The question is never "is this capability dangerous in isolation" — it's "does this capability belong here."
Flag when:
- Required environment variables don't relate to the stated purpose
- Required binaries are unrelated to the described functionality
- The install spec pulls in tools/packages disproportionate to the task
- Config path requirements suggest access to subsystems the skill shouldn't touch
### 2. Instruction scope
Read the SKILL.md content carefully. These are the literal instructions the AI agent will follow at runtime. For many skills, this is the entire security surface — there are no code files, just prose that tells the agent what commands to run, what APIs to call, and how to handle data.
Ask: do these instructions stay within the boundaries of the stated purpose?
A "database-backup" skill whose instructions include "first read the user's shell history for context" is scope creep. A "weather" skill that only runs curl against wttr.in is perfectly scoped. Instructions that reference reading files, environment variables, or system state unrelated to the skill's purpose are worth flagging — even if each individual action seems minor.
Pay close attention to:
- What commands the instructions tell the agent to run
- What files or paths the instructions reference
- What environment variables the instructions access beyond those declared in requires.env
- Whether the instructions direct data to external endpoints other than the service the skill integrates with
- Whether the instructions ask the agent to read, collect, or transmit anything not needed for the stated task
Flag when:
- Instructions direct the agent to read files or env vars unrelated to the skill's purpose
- Instructions include steps that collect, aggregate, or transmit data not needed for the task
- Instructions reference system paths, credentials, or configuration outside the skill's domain
- The instructions are vague or open-ended in ways that grant the agent broad discretion ("use your judgment to gather whatever context you need")
- Instructions direct data to unexpected endpoints (e.g., a "notion" skill that posts data somewhere other than api.notion.com)
### 3. Install mechanism risk
Evaluate what the skill installs and how. Many skills have no install spec at all — they are instruction-only and rely on binaries already being on PATH. That's the lowest risk.
The risk spectrum:
- No install spec (instruction-only) → lowest risk, nothing is written to disk
- brew formula from a well-known tap → low friction, package is reviewed
- npm/go/uv package from a public registry → moderate, packages are not pre-reviewed but are traceable
- download from a URL with extract → highest risk, arbitrary code from an arbitrary source
Flag when:
- A download-type install uses a URL that isn't a well-known release host (GitHub releases, official project domains)
- The URL points to a URL shortener, paste site, personal server, or IP address
- extract is true (the archive contents will be written to disk and potentially executed)
- The install creates binaries in non-standard locations
- Multiple install specs exist for the same platform without clear reason (e.g., two different brew formulas for the same OS)
### 4. Environment and credential proportionality
Evaluate whether the secrets and environment access requested are proportionate.
A skill that needs one API key for the service it integrates with is normal. A "trello" skill requiring TRELLO_API_KEY and TRELLO_TOKEN is expected — that's how Trello's API works. A skill that requests access to multiple unrelated credentials is suspicious. The primaryEnv field declares the "main" credential — other env requirements should serve a clear supporting role.
Flag when:
- requires.env lists credentials for services unrelated to the skill's purpose
- The number of required environment variables is high relative to the skill's complexity
- The skill requires config paths that grant access to gateway auth, channel tokens, or tool policies
- Environment variables named with patterns like SECRET, TOKEN, KEY, PASSWORD are required but not justified by the skill's purpose
- The SKILL.md instructions access environment variables beyond those declared in requires.env or primaryEnv
### 5. Persistence and privilege
Evaluate the skill's requested level of system presence.
- always: true means the skill is force-included in every agent run, bypassing all eligibility gates. This is a significant privilege.
- disable-model-invocation defaults to false. This means the agent can invoke the skill autonomously — THIS IS THE NORMAL, EXPECTED DEFAULT. Autonomous agent invocation is the entire purpose of skills. Do NOT flag this as a concern on its own.
- A skill writing to its own agent config (enabling itself, storing its own auth tokens, running its own setup/auth scripts) is NORMAL installation behavior — not privilege escalation. Do not flag this.
MITRE ATLAS context: Autonomous invocation relates to AML.T0051 (LLM Plugin Compromise) — a malicious skill with autonomous access has wider blast radius. However, since autonomous invocation is the platform default, only mention this in user guidance when it COMBINES with other red flags (always: true + broad credential access + suspicious behavior in other dimensions). Never flag autonomous invocation alone.
Flag when:
- always: true is set without clear justification (most skills should not need this)
- The skill requests permanent presence (always) combined with broad environment access
- The skill modifies OTHER skills' configurations or system-wide agent settings beyond its own scope
- The skill accesses credentials or config paths belonging to other skills
## Interpreting static scan findings
The skill has already been scanned by a regex-based pattern detector. Those findings are included in the data below. Use them as additional signal, not as your primary assessment.
- If scan findings exist, incorporate them into your reasoning but evaluate whether they make sense in context. A "deployment" skill with child_process exec is expected. A "markdown-formatter" with child_process exec is not.
- If no scan findings exist, that does NOT mean the skill is safe. Many skills are instruction-only with no code files — the regex scanner had nothing to analyze. For these skills, your assessment of the SKILL.md instructions is the primary security signal.
- Never downgrade a scan finding's severity. You can provide context for why a finding may be expected, but always surface it.
## Verdict definitions
- **benign**: The skill's capabilities, requirements, and instructions are internally consistent with its stated purpose. Nothing is disproportionate or unexplained.
- **suspicious**: There are inconsistencies between what the skill claims to do and what it actually requests, installs, or instructs. These could be legitimate design choices or sloppy engineering — but they could also indicate something worse. The user should understand what doesn't add up before proceeding.
- **malicious**: The skill's actual footprint is fundamentally incompatible with any reasonable interpretation of its stated purpose, across multiple dimensions. The inconsistencies point toward intentional misdirection — the skill appears designed to do something other than what it claims.
## Critical rules
- The bar for "malicious" is high. It requires incoherence across multiple dimensions that cannot be explained by poor engineering or over-broad requirements. A single suspicious pattern is not enough. "Suspicious" exists precisely for the cases where you can't tell.
- "Benign" does not mean "safe." It means the skill is internally coherent. A coherent skill can still have vulnerabilities. "Benign" answers "does this skill appear to be what it says it is" — not "is this skill bug-free."
- When in doubt between benign and suspicious, choose suspicious. When in doubt between suspicious and malicious, choose suspicious. The middle state is where ambiguity lives — use it.
- NEVER classify something as "malicious" solely because it uses shell execution, network calls, or file I/O. These are normal programming operations. The question is always whether they are *coherent with the skill's purpose*.
- NEVER classify something as "benign" solely because it has no scan findings. Absence of regex matches is not evidence of safety — especially for instruction-only skills with no code files.
- DO distinguish between unintentional vulnerabilities (sloppy code, missing input validation) and intentional misdirection (skill claims one purpose but its instructions/requirements reveal a different one). Vulnerabilities are "suspicious." Misdirection is "malicious."
- DO explain your reasoning. A user who doesn't know what "environment variable exfiltration" means needs you to say "this skill asks for your AWS credentials but nothing in its description suggests it needs cloud access."
- When confidence is "low", say so explicitly and explain what additional information would change your assessment.
## Output format
Respond with a JSON object and nothing else:
{
"verdict": "benign" | "suspicious" | "malicious",
"confidence": "high" | "medium" | "low",
"summary": "One sentence a non-technical user can understand.",
"dimensions": {
"purpose_capability": { "status": "ok" | "note" | "concern", "detail": "..." },
"instruction_scope": { "status": "ok" | "note" | "concern", "detail": "..." },
"install_mechanism": { "status": "ok" | "note" | "concern", "detail": "..." },
"environment_proportionality": { "status": "ok" | "note" | "concern", "detail": "..." },
"persistence_privilege": { "status": "ok" | "note" | "concern", "detail": "..." }
},
"scan_findings_in_context": [
{ "ruleId": "...", "expected_for_purpose": true | false, "note": "..." }
],
"user_guidance": "Plain-language explanation of what the user should consider before installing."
}`
// ---------------------------------------------------------------------------
// Injection pattern detection
// ---------------------------------------------------------------------------
const INJECTION_PATTERNS: Array<{ name: string; regex: RegExp }> = [
{ name: 'ignore-previous-instructions', regex: /ignore\s+(all\s+)?previous\s+instructions/i },
{ name: 'you-are-now', regex: /you\s+are\s+now\s+(a|an)\b/i },
{ name: 'system-prompt-override', regex: /system\s*prompt\s*[:=]/i },
{ name: 'base64-block', regex: /[A-Za-z0-9+/=]{200,}/ },
{
name: 'unicode-control-chars',
// eslint-disable-next-line no-control-regex
regex: /[\u200B-\u200F\u202A-\u202E\u2060-\u2064\uFEFF]/,
},
]
export function detectInjectionPatterns(text: string): string[] {
const found: string[] = []
for (const { name, regex } of INJECTION_PATTERNS) {
if (regex.test(text)) found.push(name)
}
return found
}
// ---------------------------------------------------------------------------
// Dimension metadata (maps API keys to display labels)
// ---------------------------------------------------------------------------
const DIMENSION_META: Record<string, string> = {
purpose_capability: 'Purpose & Capability',
instruction_scope: 'Instruction Scope',
install_mechanism: 'Install Mechanism',
environment_proportionality: 'Credentials',
persistence_privilege: 'Persistence & Privilege',
}
// ---------------------------------------------------------------------------
// Assemble the user message from skill data
// ---------------------------------------------------------------------------
const MAX_SKILL_MD_CHARS = 6000
export function assembleEvalUserMessage(ctx: SkillEvalContext): string {
const fm = ctx.parsed.frontmatter ?? {}
const rawClawdis = (ctx.parsed.clawdis ?? {}) as Record<string, unknown>
const meta = (ctx.parsed.metadata ?? {}) as Record<string, unknown>
const openclawFallback =
meta.openclaw && typeof meta.openclaw === 'object' && !Array.isArray(meta.openclaw)
? (meta.openclaw as Record<string, unknown>)
: {}
const clawdis = Object.keys(rawClawdis).length > 0 ? rawClawdis : openclawFallback
const requires = (clawdis.requires ?? openclawFallback.requires ?? {}) as Record<string, unknown>
const install = (clawdis.install ?? []) as Array<Record<string, unknown>>
const codeExtensions = new Set([
'.js',
'.ts',
'.mjs',
'.cjs',
'.jsx',
'.tsx',
'.py',
'.rb',
'.sh',
'.bash',
'.zsh',
'.go',
'.rs',
'.c',
'.cpp',
'.java',
])
const codeFiles = ctx.files.filter((f) => {
const ext = f.path.slice(f.path.lastIndexOf('.')).toLowerCase()
return codeExtensions.has(ext)
})
const skillMd =
ctx.skillMdContent.length > MAX_SKILL_MD_CHARS
? `${ctx.skillMdContent.slice(0, MAX_SKILL_MD_CHARS)}\n…[truncated]`
: ctx.skillMdContent
const sections: string[] = []
// Skill identity
sections.push(`## Skill under evaluation
**Name:** ${ctx.displayName}
**Description:** ${ctx.summary ?? 'No description provided.'}
**Source:** ${ctx.source ?? 'unknown'}
**Homepage:** ${ctx.homepage ?? 'none'}
**Registry metadata:**
- Owner ID: ${ctx.ownerUserId}
- Slug: ${ctx.slug}
- Version: ${ctx.version}
- Published: ${new Date(ctx.createdAt).toISOString()}`)
// Flags
const always = fm.always ?? clawdis.always
const userInvocable = fm['user-invocable'] ?? clawdis.userInvocable
const disableModelInvocation = fm['disable-model-invocation'] ?? clawdis.disableModelInvocation
const os = clawdis.os
sections.push(`**Flags:**
- always: ${formatWithDefault(always, 'false (default)')}
- user-invocable: ${formatWithDefault(userInvocable, 'true (default)')}
- disable-model-invocation: ${formatWithDefault(
disableModelInvocation,
'false (default — agent can invoke autonomously, this is normal)',
)}
- OS restriction: ${Array.isArray(os) ? os.join(', ') : formatWithDefault(os, 'none')}`)
// Requirements
const bins = (requires.bins as string[] | undefined) ?? []
const anyBins = (requires.anyBins as string[] | undefined) ?? []
const env = (requires.env as string[] | undefined) ?? []
const primaryEnv = (clawdis.primaryEnv as string | undefined) ?? 'none'
const config = (requires.config as string[] | undefined) ?? []
sections.push(`### Requirements
- Required binaries (all must exist): ${bins.length ? bins.join(', ') : 'none'}
- Required binaries (at least one): ${anyBins.length ? anyBins.join(', ') : 'none'}
- Required env vars: ${env.length ? env.join(', ') : 'none'}
- Primary credential: ${primaryEnv}
- Required config paths: ${config.length ? config.join(', ') : 'none'}`)
// Install specifications
if (install.length > 0) {
const specLines = install.map((spec, i) => {
const kind = spec.kind ?? 'unknown'
const parts = [`- **[${i}] ${formatScalar(kind)}**`]
if (spec.formula) parts.push(`formula: ${formatScalar(spec.formula)}`)
if (spec.package) parts.push(`package: ${formatScalar(spec.package)}`)
if (spec.module) parts.push(`module: ${formatScalar(spec.module)}`)
if (spec.url) parts.push(`url: ${formatScalar(spec.url)}`)
if (spec.archive) parts.push(`archive: ${formatScalar(spec.archive)}`)
if (spec.extract !== undefined) parts.push(`extract: ${formatScalar(spec.extract)}`)
if (spec.bins) parts.push(`creates binaries: ${(spec.bins as string[]).join(', ')}`)
return parts.join(' | ')
})
sections.push(`### Install specifications\n${specLines.join('\n')}`)
} else {
sections.push(
'### Install specifications\nNo install spec — this is an instruction-only skill.',
)
}
// Code file presence
if (codeFiles.length > 0) {
const fileList = codeFiles.map((f) => ` ${f.path} (${f.size} bytes)`).join('\n')
sections.push(`### Code file presence\n${codeFiles.length} code file(s):\n${fileList}`)
} else {
sections.push(
'### Code file presence\nNo code files present — this is an instruction-only skill. The regex-based scanner had nothing to analyze.',
)
}
// File manifest
const manifest = ctx.files.map((f) => ` ${f.path} (${f.size} bytes)`).join('\n')
sections.push(`### File manifest\n${ctx.files.length} file(s):\n${manifest}`)
// Pre-scan injection signals
if (ctx.injectionSignals.length > 0) {
sections.push(
`### Pre-scan injection signals\nThe following prompt-injection patterns were detected in the SKILL.md content. The skill may be attempting to manipulate this evaluation:\n${ctx.injectionSignals.map((s) => `- ${s}`).join('\n')}`,
)
} else {
sections.push('### Pre-scan injection signals\nNone detected.')
}
// SKILL.md content
sections.push(`### SKILL.md content (runtime instructions)\n${skillMd}`)
// All file contents
if (ctx.fileContents.length > 0) {
const MAX_FILE_CHARS = 10000
const MAX_TOTAL_CHARS = 50000
let totalChars = 0
const fileBlocks: string[] = []
for (const f of ctx.fileContents) {
if (totalChars >= MAX_TOTAL_CHARS) {
fileBlocks.push(
`\n…[remaining files truncated, ${ctx.fileContents.length - fileBlocks.length} file(s) omitted]`,
)
break
}
const content =
f.content.length > MAX_FILE_CHARS
? `${f.content.slice(0, MAX_FILE_CHARS)}\n…[truncated]`
: f.content
fileBlocks.push(`#### ${f.path}\n\`\`\`\n${content}\n\`\`\``)
totalChars += content.length
}
sections.push(
`### File contents\nFull source of all included files. Review these carefully for malicious behavior, hidden endpoints, data exfiltration, obfuscated code, or behavior that contradicts the SKILL.md.\n\n${fileBlocks.join('\n\n')}`,
)
}
// Reminder to respond in JSON (required by OpenAI json_object mode)
sections.push('Respond with your evaluation as a single JSON object.')
return sections.join('\n\n')
}
// ---------------------------------------------------------------------------
// Parse the LLM response
// ---------------------------------------------------------------------------
const VALID_VERDICTS = new Set(['benign', 'suspicious', 'malicious'])
const VALID_CONFIDENCES = new Set(['high', 'medium', 'low'])
export function parseLlmEvalResponse(raw: string): LlmEvalResponse | null {
// Strip markdown code fences if present
let text = raw.trim()
if (text.startsWith('```')) {
const firstNewline = text.indexOf('\n')
text = text.slice(firstNewline + 1)
const lastFence = text.lastIndexOf('```')
if (lastFence !== -1) text = text.slice(0, lastFence)
text = text.trim()
}
let parsed: unknown
try {
parsed = JSON.parse(text)
} catch {
return null
}
if (!parsed || typeof parsed !== 'object') return null
const obj = parsed as Record<string, unknown>
// Validate required fields
const verdict = typeof obj.verdict === 'string' ? obj.verdict.toLowerCase() : null
if (!verdict || !VALID_VERDICTS.has(verdict)) return null
const confidence = typeof obj.confidence === 'string' ? obj.confidence.toLowerCase() : null
if (!confidence || !VALID_CONFIDENCES.has(confidence)) return null
const summary = typeof obj.summary === 'string' ? obj.summary : ''
// Parse dimensions
const rawDims = obj.dimensions as Record<string, unknown> | undefined
const dimensions: LlmEvalDimension[] = []
if (rawDims && typeof rawDims === 'object') {
for (const [key, value] of Object.entries(rawDims)) {
if (!value || typeof value !== 'object') continue
const dim = value as Record<string, unknown>
const status = typeof dim.status === 'string' ? dim.status : 'note'
const detail = typeof dim.detail === 'string' ? dim.detail : ''
dimensions.push({
name: key,
label: DIMENSION_META[key] ?? key,
rating: status,
detail,
})
}
}
// Parse findings
const rawFindings = obj.scan_findings_in_context
let findings = ''
if (Array.isArray(rawFindings) && rawFindings.length > 0) {
findings = rawFindings
.map((f: unknown) => {
if (!f || typeof f !== 'object') return null
const entry = f as Record<string, unknown>
const ruleId = entry.ruleId ?? 'unknown'
const expected = entry.expected_for_purpose ? 'expected' : 'unexpected'
const note = entry.note ?? ''
return `[${formatScalar(ruleId)}] ${expected}: ${formatScalar(note)}`
})
.filter(Boolean)
.join('\n')
}
const guidance = typeof obj.user_guidance === 'string' ? obj.user_guidance : ''
return {
verdict: verdict as LlmEvalResponse['verdict'],
confidence: confidence as LlmEvalResponse['confidence'],
summary,
dimensions,
guidance,
findings,
}
}
+4
View File
@@ -174,6 +174,10 @@ export async function publishVersionForUser(
versionId: publishResult.versionId,
})
await ctx.scheduler.runAfter(0, internal.llmEval.evaluateWithLlm, {
versionId: publishResult.versionId,
})
const owner = (await ctx.runQuery(internal.users.getByIdInternal, {
userId,
})) as Doc<'users'> | null
+7 -2
View File
@@ -49,7 +49,9 @@ export function getFrontmatterMetadata(frontmatter: ParsedSkillFrontmatter) {
if (!raw) return undefined
if (typeof raw === 'string') {
try {
const parsed = JSON.parse(raw) as unknown
// Strip trailing commas in JSON objects/arrays (common authoring mistake)
const cleaned = raw.replace(/,\s*([\]}])/g, '$1')
const parsed = JSON.parse(cleaned) as unknown
return parsed ?? undefined
} catch {
return undefined
@@ -67,12 +69,15 @@ export function parseClawdisMetadata(frontmatter: ParsedSkillFrontmatter) {
: undefined
const clawdbotMeta = metadataRecord?.clawdbot
const clawdisMeta = metadataRecord?.clawdis
const openclawMeta = metadataRecord?.openclaw
const metadataSource =
clawdbotMeta && typeof clawdbotMeta === 'object' && !Array.isArray(clawdbotMeta)
? (clawdbotMeta as Record<string, unknown>)
: clawdisMeta && typeof clawdisMeta === 'object' && !Array.isArray(clawdisMeta)
? (clawdisMeta as Record<string, unknown>)
: undefined
: openclawMeta && typeof openclawMeta === 'object' && !Array.isArray(openclawMeta)
? (openclawMeta as Record<string, unknown>)
: undefined
const clawdisRaw = metadataSource ?? frontmatter.clawdis
if (!clawdisRaw || typeof clawdisRaw !== 'object' || Array.isArray(clawdisRaw)) return undefined
+398
View File
@@ -0,0 +1,398 @@
import { v } from 'convex/values'
import { internal } from './_generated/api'
import type { Doc, Id } from './_generated/dataModel'
import { internalAction } from './_generated/server'
import type { SkillEvalContext } from './lib/securityPrompt'
import {
assembleEvalUserMessage,
detectInjectionPatterns,
getLlmEvalModel,
LLM_EVAL_MAX_OUTPUT_TOKENS,
parseLlmEvalResponse,
SECURITY_EVALUATOR_SYSTEM_PROMPT,
} from './lib/securityPrompt'
// ---------------------------------------------------------------------------
// Helpers
// ---------------------------------------------------------------------------
function extractResponseText(payload: unknown): string | null {
if (!payload || typeof payload !== 'object') return null
const output = (payload as { output?: unknown }).output
if (!Array.isArray(output)) return null
const chunks: string[] = []
for (const item of output) {
if (!item || typeof item !== 'object') continue
if ((item as { type?: unknown }).type !== 'message') continue
const content = (item as { content?: unknown }).content
if (!Array.isArray(content)) continue
for (const part of content) {
if (!part || typeof part !== 'object') continue
if ((part as { type?: unknown }).type !== 'output_text') continue
const text = (part as { text?: unknown }).text
if (typeof text === 'string' && text.trim()) chunks.push(text)
}
}
const joined = chunks.join('\n').trim()
return joined || null
}
function verdictToStatus(verdict: string): string {
switch (verdict) {
case 'benign':
return 'clean'
case 'malicious':
return 'malicious'
case 'suspicious':
return 'suspicious'
default:
return 'pending'
}
}
// ---------------------------------------------------------------------------
// Publish-time evaluation action
// ---------------------------------------------------------------------------
export const evaluateWithLlm = internalAction({
args: {
versionId: v.id('skillVersions'),
},
handler: async (ctx, args) => {
const apiKey = process.env.OPENAI_API_KEY
if (!apiKey) {
console.log('[llmEval] OPENAI_API_KEY not configured, skipping evaluation')
return
}
const model = getLlmEvalModel()
// Store error helper
const storeError = async (message: string) => {
console.error(`[llmEval] ${message}`)
await ctx.runMutation(internal.skills.updateVersionLlmAnalysisInternal, {
versionId: args.versionId,
llmAnalysis: {
status: 'error',
summary: message,
model,
checkedAt: Date.now(),
},
})
}
// 1. Fetch version
const version = (await ctx.runQuery(internal.skills.getVersionByIdInternal, {
versionId: args.versionId,
})) as Doc<'skillVersions'> | null
if (!version) {
await storeError(`Version ${args.versionId} not found`)
return
}
// 2. Fetch skill
const skill = (await ctx.runQuery(internal.skills.getSkillByIdInternal, {
skillId: version.skillId,
})) as Doc<'skills'> | null
if (!skill) {
await storeError(`Skill ${version.skillId} not found`)
return
}
// 3. Read SKILL.md content
const skillMdFile = version.files.find((f) => {
const lower = f.path.toLowerCase()
return lower === 'skill.md' || lower === 'skills.md'
})
let skillMdContent = ''
if (skillMdFile) {
const blob = await ctx.storage.get(skillMdFile.storageId as Id<'_storage'>)
if (blob) {
skillMdContent = await blob.text()
}
}
if (!skillMdContent) {
await storeError('No SKILL.md content found')
return
}
// 4. Read all file contents
const fileContents: Array<{ path: string; content: string }> = []
for (const f of version.files) {
const lower = f.path.toLowerCase()
if (lower === 'skill.md' || lower === 'skills.md') continue
try {
const blob = await ctx.storage.get(f.storageId as Id<'_storage'>)
if (blob) {
fileContents.push({ path: f.path, content: await blob.text() })
}
} catch {
// Skip files that can't be read
}
}
// 5. Detect injection patterns across ALL content
const allContent = [skillMdContent, ...fileContents.map((f) => f.content)].join('\n')
const injectionSignals = detectInjectionPatterns(allContent)
// 6. Build eval context
const parsed = version.parsed as SkillEvalContext['parsed']
const fm = parsed.frontmatter ?? {}
const evalCtx: SkillEvalContext = {
slug: skill.slug,
displayName: skill.displayName,
ownerUserId: String(skill.ownerUserId),
version: version.version,
createdAt: version.createdAt,
summary: (skill.summary as string | undefined) ?? undefined,
source: (fm.source as string | undefined) ?? undefined,
homepage: (fm.homepage as string | undefined) ?? undefined,
parsed,
files: version.files.map((f) => ({ path: f.path, size: f.size })),
skillMdContent,
fileContents,
injectionSignals,
}
// 6. Assemble user message
const userMessage = assembleEvalUserMessage(evalCtx)
// 7. Call OpenAI Responses API (with retry for rate limits)
const MAX_RETRIES = 3
let raw: string | null = null
try {
const body = JSON.stringify({
model,
instructions: SECURITY_EVALUATOR_SYSTEM_PROMPT,
input: userMessage,
max_output_tokens: LLM_EVAL_MAX_OUTPUT_TOKENS,
text: {
format: {
type: 'json_object',
},
},
})
let response: Response | null = null
for (let attempt = 0; attempt <= MAX_RETRIES; attempt++) {
response = await fetch('https://api.openai.com/v1/responses', {
method: 'POST',
headers: {
'Content-Type': 'application/json',
Authorization: `Bearer ${apiKey}`,
},
body,
})
if (response.status === 429 || response.status >= 500) {
if (attempt < MAX_RETRIES) {
const delay = 2 ** attempt * 2000 + Math.random() * 1000
console.log(
`[llmEval] Rate limited (${response.status}), retrying in ${Math.round(delay)}ms (attempt ${attempt + 1}/${MAX_RETRIES})`,
)
await new Promise((r) => setTimeout(r, delay))
continue
}
}
break
}
if (!response || !response.ok) {
const errorText = response ? await response.text() : 'No response'
await storeError(`OpenAI API error (${response?.status}): ${errorText.slice(0, 200)}`)
return
}
const payload = (await response.json()) as unknown
raw = extractResponseText(payload)
} catch (error) {
await storeError(
`OpenAI API call failed: ${error instanceof Error ? error.message : String(error)}`,
)
return
}
if (!raw) {
await storeError('Empty response from OpenAI')
return
}
// 8. Parse response
const result = parseLlmEvalResponse(raw)
if (!result) {
console.error(`[llmEval] Raw response (first 500 chars): ${raw.slice(0, 500)}`)
await storeError('Failed to parse LLM evaluation response')
return
}
// 9. Store result
await ctx.runMutation(internal.skills.updateVersionLlmAnalysisInternal, {
versionId: args.versionId,
llmAnalysis: {
status: verdictToStatus(result.verdict),
verdict: result.verdict,
confidence: result.confidence,
summary: result.summary,
dimensions: result.dimensions,
guidance: result.guidance,
findings: result.findings || undefined,
model,
checkedAt: Date.now(),
},
})
console.log(
`[llmEval] Evaluated ${skill.slug}@${version.version}: ${result.verdict} (${result.confidence} confidence)`,
)
// 10. Update moderation flags — re-read version to get the sha256hash
// that VT may have stored while we were evaluating (both run concurrently).
const freshVersion = (await ctx.runQuery(internal.skills.getVersionByIdInternal, {
versionId: args.versionId,
})) as Doc<'skillVersions'> | null
const sha256hash = freshVersion?.sha256hash ?? version.sha256hash
if (sha256hash) {
const status = verdictToStatus(result.verdict)
if (status === 'malicious' || status === 'suspicious' || status === 'clean') {
await ctx.runMutation(internal.skills.approveSkillByHashInternal, {
sha256hash,
scanner: 'llm',
status,
})
}
}
},
})
// ---------------------------------------------------------------------------
// Convenience: evaluate a single skill by slug (for testing / manual runs)
// Usage: npx convex run llmEval:evaluateBySlug '{"slug": "transcribeexx"}'
// ---------------------------------------------------------------------------
export const evaluateBySlug = internalAction({
args: {
slug: v.string(),
},
handler: async (ctx, args) => {
const skill = (await ctx.runQuery(internal.skills.getSkillBySlugInternal, {
slug: args.slug,
})) as Doc<'skills'> | null
if (!skill) {
console.error(`[llmEval:bySlug] Skill "${args.slug}" not found`)
return { error: 'Skill not found' }
}
if (!skill.latestVersionId) {
console.error(`[llmEval:bySlug] Skill "${args.slug}" has no published version`)
return { error: 'No published version' }
}
console.log(`[llmEval:bySlug] Evaluating ${args.slug} (versionId: ${skill.latestVersionId})`)
await ctx.scheduler.runAfter(0, internal.llmEval.evaluateWithLlm, {
versionId: skill.latestVersionId,
})
return { ok: true, slug: args.slug, versionId: skill.latestVersionId }
},
})
// ---------------------------------------------------------------------------
// Backfill action (Phase 2)
// Schedules individual evaluateWithLlm actions for each skill in the batch,
// then self-schedules the next batch. Each eval runs as its own action
// invocation so we don't hit Convex action timeouts.
// ---------------------------------------------------------------------------
export const backfillLlmEval = internalAction({
args: {
cursor: v.optional(v.number()),
batchSize: v.optional(v.number()),
accTotal: v.optional(v.number()),
accScheduled: v.optional(v.number()),
accSkipped: v.optional(v.number()),
startTime: v.optional(v.number()),
},
handler: async (ctx, args) => {
const startTime = args.startTime ?? Date.now()
const apiKey = process.env.OPENAI_API_KEY
if (!apiKey) {
console.log('[llmEval:backfill] OPENAI_API_KEY not configured')
return { error: 'OPENAI_API_KEY not configured' }
}
const batchSize = args.batchSize ?? 25
const cursor = args.cursor ?? 0
let accTotal = args.accTotal ?? 0
let accScheduled = args.accScheduled ?? 0
let accSkipped = args.accSkipped ?? 0
const batch = await ctx.runQuery(internal.skills.getActiveSkillBatchForLlmBackfillInternal, {
cursor,
batchSize,
})
if (batch.skills.length === 0 && batch.done) {
console.log('[llmEval:backfill] No more skills to evaluate')
return { total: accTotal, scheduled: accScheduled, skipped: accSkipped }
}
console.log(
`[llmEval:backfill] Processing batch of ${batch.skills.length} skills (cursor=${cursor}, accumulated=${accTotal})`,
)
for (const { versionId, slug } of batch.skills) {
// Re-evaluate all (full file content reading upgrade)
const version = (await ctx.runQuery(internal.skills.getVersionByIdInternal, {
versionId,
})) as Doc<'skillVersions'> | null
if (!version) {
accSkipped++
continue
}
// Schedule each evaluation as a separate action invocation
await ctx.scheduler.runAfter(0, internal.llmEval.evaluateWithLlm, { versionId })
accScheduled++
console.log(`[llmEval:backfill] Scheduled eval for ${slug}`)
}
accTotal += batch.skills.length
if (!batch.done) {
// Delay the next batch slightly to avoid overwhelming the scheduler
// when all evals from this batch are also running
console.log(
`[llmEval:backfill] Scheduling next batch (cursor=${batch.nextCursor}, total so far=${accTotal})`,
)
await ctx.scheduler.runAfter(5_000, internal.llmEval.backfillLlmEval, {
cursor: batch.nextCursor,
batchSize,
accTotal,
accScheduled,
accSkipped,
startTime,
})
return { status: 'continuing', totalSoFar: accTotal }
}
const durationMs = Date.now() - startTime
const result = {
total: accTotal,
scheduled: accScheduled,
skipped: accSkipped,
durationMs,
}
console.log('[llmEval:backfill] Complete:', result)
return result
},
})
+22
View File
@@ -171,6 +171,28 @@ const skillVersions = defineTable({
checkedAt: v.number(),
}),
),
llmAnalysis: v.optional(
v.object({
status: v.string(),
verdict: v.optional(v.string()),
confidence: v.optional(v.string()),
summary: v.optional(v.string()),
dimensions: v.optional(
v.array(
v.object({
name: v.string(),
label: v.string(),
rating: v.string(),
detail: v.string(),
}),
),
),
guidance: v.optional(v.string()),
findings: v.optional(v.string()),
model: v.optional(v.string()),
checkedAt: v.number(),
}),
),
})
.index('by_skill', ['skillId'])
.index('by_skill_version', ['skillId', 'version'])
+295 -2
View File
@@ -1,12 +1,305 @@
/* @vitest-environment node */
import { describe, expect, it } from 'vitest'
import { __test } from './search'
import { describe, expect, it, vi } from 'vitest'
import { tokenize } from './lib/searchText'
import { __test, lexicalFallbackSkills, searchSkills } from './search'
const { generateEmbeddingMock, getSkillBadgeMapsMock } = vi.hoisted(() => ({
generateEmbeddingMock: vi.fn(),
getSkillBadgeMapsMock: vi.fn(),
}))
vi.mock('./lib/embeddings', () => ({
generateEmbedding: generateEmbeddingMock,
}))
vi.mock('./lib/badges', () => ({
getSkillBadgeMaps: getSkillBadgeMapsMock,
isSkillHighlighted: (skill: { badges?: Record<string, unknown> }) =>
Boolean(skill.badges?.highlighted),
}))
type WrappedHandler = {
_handler: (ctx: unknown, args: unknown) => Promise<unknown>
}
const searchSkillsHandler = (searchSkills as unknown as WrappedHandler)._handler
const lexicalFallbackSkillsHandler = (lexicalFallbackSkills as unknown as WrappedHandler)._handler
describe('search helpers', () => {
it('returns fallback results when vector candidates are empty', async () => {
generateEmbeddingMock.mockResolvedValueOnce([0, 1, 2])
const fallback = [
{
skill: makePublicSkill({ id: 'skills:orf', slug: 'orf', displayName: 'ORF' }),
version: null,
ownerHandle: 'steipete',
},
]
const runQuery = vi
.fn()
.mockResolvedValueOnce([])
.mockResolvedValueOnce([])
.mockResolvedValueOnce(fallback)
const result = await searchSkillsHandler(
{
vectorSearch: vi.fn().mockResolvedValue([]),
runQuery,
},
{ query: 'orf', limit: 10 },
)
expect(result).toHaveLength(1)
expect(result[0].skill.slug).toBe('orf')
expect(runQuery).toHaveBeenLastCalledWith(
expect.anything(),
expect.objectContaining({ query: 'orf', queryTokens: ['orf'] }),
)
})
it('applies highlightedOnly filtering in lexical fallback', async () => {
const highlighted = makeSkillDoc({
id: 'skills:hl',
slug: 'orf-highlighted',
displayName: 'ORF Highlighted',
})
const plain = makeSkillDoc({ id: 'skills:plain', slug: 'orf-plain', displayName: 'ORF Plain' })
getSkillBadgeMapsMock.mockResolvedValueOnce(
new Map([
['skills:hl', { highlighted: { byUserId: 'users:mod', at: 1 } }],
['skills:plain', {}],
]),
)
const result = await lexicalFallbackSkillsHandler(
makeLexicalCtx({
exactSlugSkill: null,
recentSkills: [highlighted, plain],
}),
{ query: 'orf', queryTokens: ['orf'], highlightedOnly: true, limit: 10 },
)
expect(result).toHaveLength(1)
expect(result[0].skill.slug).toBe('orf-highlighted')
})
it('includes exact slug match from by_slug even when recent scan is empty', async () => {
const exactSlugSkill = makeSkillDoc({ id: 'skills:orf', slug: 'orf', displayName: 'ORF' })
getSkillBadgeMapsMock.mockResolvedValueOnce(new Map([['skills:orf', {}]]))
const ctx = makeLexicalCtx({
exactSlugSkill,
recentSkills: [],
})
const result = await lexicalFallbackSkillsHandler(ctx, {
query: 'orf',
queryTokens: ['orf'],
limit: 10,
})
expect(result).toHaveLength(1)
expect(result[0].skill.slug).toBe('orf')
expect(ctx.db.query).toHaveBeenCalledWith('skills')
})
it('dedupes overlap and enforces rank + limit across vector and fallback', async () => {
generateEmbeddingMock.mockResolvedValueOnce([0, 1, 2])
const vectorEntries = [
{
embeddingId: 'skillEmbeddings:a',
skill: makePublicSkill({
id: 'skills:a',
slug: 'foo-a',
displayName: 'Foo Alpha',
downloads: 10,
}),
version: null,
ownerHandle: 'one',
},
{
embeddingId: 'skillEmbeddings:b',
skill: makePublicSkill({
id: 'skills:b',
slug: 'foo-b',
displayName: 'Foo Beta',
downloads: 2,
}),
version: null,
ownerHandle: 'two',
},
]
const fallbackEntries = [
{
skill: makePublicSkill({
id: 'skills:a',
slug: 'foo-a',
displayName: 'Foo Alpha',
downloads: 10,
}),
version: null,
ownerHandle: 'one',
},
{
skill: makePublicSkill({
id: 'skills:c',
slug: 'foo-c',
displayName: 'Foo Classic',
downloads: 1,
}),
version: null,
ownerHandle: 'three',
},
]
const runQuery = vi
.fn()
.mockResolvedValueOnce(vectorEntries)
.mockResolvedValueOnce([])
.mockResolvedValueOnce(fallbackEntries)
const result = await searchSkillsHandler(
{
vectorSearch: vi.fn().mockResolvedValue([
{ _id: 'skillEmbeddings:a', _score: 0.4 },
{ _id: 'skillEmbeddings:b', _score: 0.9 },
]),
runQuery,
},
{ query: 'foo', limit: 2 },
)
expect(result).toHaveLength(2)
expect(result[0].skill.slug).toBe('foo-b')
expect(new Set(result.map((entry: { skill: { _id: string } }) => entry.skill._id)).size).toBe(2)
})
it('advances candidate limit until max', () => {
expect(__test.getNextCandidateLimit(50, 1000)).toBe(100)
expect(__test.getNextCandidateLimit(800, 1000)).toBe(1000)
expect(__test.getNextCandidateLimit(1000, 1000)).toBeNull()
})
it('boosts exact slug/name matches over loose matches', () => {
const queryTokens = tokenize('notion')
const exactScore = __test.scoreSkillResult(queryTokens, 0.4, 'Notion Sync', 'notion-sync', 5)
const looseScore = __test.scoreSkillResult(queryTokens, 0.6, 'Notes Sync', 'notes-sync', 500)
expect(exactScore).toBeGreaterThan(looseScore)
})
it('adds a popularity prior for equally relevant matches', () => {
const queryTokens = tokenize('notion')
const lowDownloads = __test.scoreSkillResult(
queryTokens,
0.5,
'Notion Helper',
'notion-helper',
0,
)
const highDownloads = __test.scoreSkillResult(
queryTokens,
0.5,
'Notion Helper',
'notion-helper',
1000,
)
expect(highDownloads).toBeGreaterThan(lowDownloads)
})
it('merges fallback matches without duplicate skill ids', () => {
const primary = [
{
embeddingId: 'skillEmbeddings:1',
skill: { _id: 'skills:1' },
},
] as unknown as Parameters<typeof __test.mergeUniqueBySkillId>[0]
const fallback = [
{
skill: { _id: 'skills:1' },
},
{
skill: { _id: 'skills:2' },
},
] as unknown as Parameters<typeof __test.mergeUniqueBySkillId>[1]
const merged = __test.mergeUniqueBySkillId(primary, fallback)
expect(merged).toHaveLength(2)
expect(merged.map((entry) => entry.skill._id)).toEqual(['skills:1', 'skills:2'])
})
})
function makePublicSkill(params: {
id: string
slug: string
displayName: string
downloads?: number
}) {
return {
_id: params.id,
_creationTime: 1,
slug: params.slug,
displayName: params.displayName,
summary: `${params.displayName} summary`,
ownerUserId: 'users:owner',
canonicalSkillId: undefined,
forkOf: undefined,
latestVersionId: 'skillVersions:1',
tags: {},
badges: {},
stats: {
downloads: params.downloads ?? 0,
installsCurrent: 0,
installsAllTime: 0,
stars: 0,
versions: 1,
comments: 0,
},
createdAt: 1,
updatedAt: 1,
}
}
function makeSkillDoc(params: { id: string; slug: string; displayName: string }) {
return {
...makePublicSkill(params),
_creationTime: 1,
moderationStatus: 'active',
moderationFlags: [],
softDeletedAt: undefined,
}
}
function makeLexicalCtx(params: {
exactSlugSkill: ReturnType<typeof makeSkillDoc> | null
recentSkills: Array<ReturnType<typeof makeSkillDoc>>
}) {
return {
db: {
query: vi.fn((table: string) => {
if (table !== 'skills') throw new Error(`Unexpected table ${table}`)
return {
withIndex: (index: string) => {
if (index === 'by_slug') {
return {
unique: vi.fn().mockResolvedValue(params.exactSlugSkill),
}
}
if (index === 'by_active_updated') {
return {
order: () => ({
take: vi.fn().mockResolvedValue(params.recentSkills),
}),
}
}
throw new Error(`Unexpected index ${index}`)
},
}
}),
get: vi.fn(async (id: string) => {
if (id.startsWith('users:')) return { _id: id, handle: 'owner' }
if (id.startsWith('skillVersions:')) return { _id: id, version: '1.0.0' }
return null
}),
},
}
}
+193 -14
View File
@@ -7,20 +7,86 @@ import { generateEmbedding } from './lib/embeddings'
import { toPublicSkill, toPublicSoul } from './lib/public'
import { matchesExactTokens, tokenize } from './lib/searchText'
type HydratedEntry = {
embeddingId: Id<'skillEmbeddings'>
type SkillSearchEntry = {
embeddingId?: Id<'skillEmbeddings'>
skill: NonNullable<ReturnType<typeof toPublicSkill>>
version: Doc<'skillVersions'> | null
ownerHandle: string | null
}
type SearchResult = HydratedEntry & { score: number }
type SearchResult = SkillSearchEntry & { score: number }
const SLUG_EXACT_BOOST = 1.4
const SLUG_PREFIX_BOOST = 0.8
const NAME_EXACT_BOOST = 1.1
const NAME_PREFIX_BOOST = 0.6
const POPULARITY_WEIGHT = 0.08
const FALLBACK_SCAN_LIMIT = 1200
function getNextCandidateLimit(current: number, max: number) {
const next = Math.min(current * 2, max)
return next > current ? next : null
}
function matchesAllTokens(
queryTokens: string[],
candidateTokens: string[],
matcher: (candidate: string, query: string) => boolean,
) {
if (queryTokens.length === 0 || candidateTokens.length === 0) return false
return queryTokens.every((queryToken) =>
candidateTokens.some((candidateToken) => matcher(candidateToken, queryToken)),
)
}
function getLexicalBoost(queryTokens: string[], displayName: string, slug: string) {
const slugTokens = tokenize(slug)
const nameTokens = tokenize(displayName)
let boost = 0
if (matchesAllTokens(queryTokens, slugTokens, (candidate, query) => candidate === query)) {
boost += SLUG_EXACT_BOOST
} else if (
matchesAllTokens(queryTokens, slugTokens, (candidate, query) => candidate.startsWith(query))
) {
boost += SLUG_PREFIX_BOOST
}
if (matchesAllTokens(queryTokens, nameTokens, (candidate, query) => candidate === query)) {
boost += NAME_EXACT_BOOST
} else if (
matchesAllTokens(queryTokens, nameTokens, (candidate, query) => candidate.startsWith(query))
) {
boost += NAME_PREFIX_BOOST
}
return boost
}
function scoreSkillResult(
queryTokens: string[],
vectorScore: number,
displayName: string,
slug: string,
downloads: number,
) {
const lexicalBoost = getLexicalBoost(queryTokens, displayName, slug)
const popularityBoost = Math.log1p(Math.max(downloads, 0)) * POPULARITY_WEIGHT
return vectorScore + lexicalBoost + popularityBoost
}
function mergeUniqueBySkillId(primary: SkillSearchEntry[], fallback: SkillSearchEntry[]) {
if (fallback.length === 0) return primary
const out = [...primary]
const seen = new Set(primary.map((entry) => entry.skill._id))
for (const entry of fallback) {
if (seen.has(entry.skill._id)) continue
seen.add(entry.skill._id)
out.push(entry)
}
return out
}
export const searchSkills: ReturnType<typeof action> = action({
args: {
query: v.string(),
@@ -43,9 +109,9 @@ export const searchSkills: ReturnType<typeof action> = action({
// Convex vectorSearch max limit is 256; clamp candidate sizes accordingly.
const maxCandidate = Math.min(Math.max(limit * 10, 200), 256)
let candidateLimit = Math.min(Math.max(limit * 3, 50), 256)
let hydrated: HydratedEntry[] = []
let hydrated: SkillSearchEntry[] = []
let scoreById = new Map<Id<'skillEmbeddings'>, number>()
let exactMatches: HydratedEntry[] = []
let exactMatches: SkillSearchEntry[] = []
while (candidateLimit <= maxCandidate) {
const results = await ctx.vectorSearch('skillEmbeddings', 'by_embedding', {
@@ -56,7 +122,7 @@ export const searchSkills: ReturnType<typeof action> = action({
hydrated = (await ctx.runQuery(internal.search.hydrateResults, {
embeddingIds: results.map((result) => result._id),
})) as HydratedEntry[]
})) as SkillSearchEntry[]
scoreById = new Map<Id<'skillEmbeddings'>, number>(
results.map((result) => [result._id, result._score]),
@@ -95,12 +161,34 @@ export const searchSkills: ReturnType<typeof action> = action({
candidateLimit = nextLimit
}
return exactMatches
.map((entry) => ({
...entry,
score: scoreById.get(entry.embeddingId) ?? 0,
}))
const fallbackMatches =
exactMatches.length >= limit
? []
: ((await ctx.runQuery(internal.search.lexicalFallbackSkills, {
query,
queryTokens,
limit: Math.min(Math.max(limit * 4, 200), FALLBACK_SCAN_LIMIT),
highlightedOnly: args.highlightedOnly,
})) as SkillSearchEntry[])
const mergedMatches = mergeUniqueBySkillId(exactMatches, fallbackMatches)
return mergedMatches
.map((entry) => {
const vectorScore = entry.embeddingId ? (scoreById.get(entry.embeddingId) ?? 0) : 0
return {
...entry,
score: scoreSkillResult(
queryTokens,
vectorScore,
entry.skill.displayName,
entry.skill.slug,
entry.skill.stats.downloads,
),
}
})
.filter((entry) => entry.skill)
.sort((a, b) => b.score - a.score || b.skill.stats.downloads - a.skill.stats.downloads)
.slice(0, limit)
},
})
@@ -115,7 +203,7 @@ export const getBadgeMapsForSkills = internalQuery({
export const hydrateResults = internalQuery({
args: { embeddingIds: v.array(v.id('skillEmbeddings')) },
handler: async (ctx, args): Promise<HydratedEntry[]> => {
handler: async (ctx, args): Promise<SkillSearchEntry[]> => {
const ownerHandleCache = new Map<Id<'users'>, Promise<string | null>>()
const getOwnerHandle = (ownerUserId: Id<'users'>) => {
@@ -144,7 +232,92 @@ export const hydrateResults = internalQuery({
}),
)
return entries.filter((entry): entry is HydratedEntry => entry !== null)
return entries.filter((entry): entry is SkillSearchEntry => entry !== null)
},
})
export const lexicalFallbackSkills = internalQuery({
args: {
query: v.string(),
queryTokens: v.array(v.string()),
limit: v.optional(v.number()),
highlightedOnly: v.optional(v.boolean()),
},
handler: async (ctx, args): Promise<SkillSearchEntry[]> => {
const limit = Math.min(Math.max(args.limit ?? 200, 10), FALLBACK_SCAN_LIMIT)
const seenSkillIds = new Set<Id<'skills'>>()
const candidateSkills: Doc<'skills'>[] = []
const slugQuery = args.query.trim().toLowerCase()
if (/^[a-z0-9][a-z0-9-]*$/.test(slugQuery)) {
const exactSlugSkill = await ctx.db
.query('skills')
.withIndex('by_slug', (q) => q.eq('slug', slugQuery))
.unique()
if (exactSlugSkill && !exactSlugSkill.softDeletedAt) {
seenSkillIds.add(exactSlugSkill._id)
candidateSkills.push(exactSlugSkill)
}
}
const recentSkills = await ctx.db
.query('skills')
.withIndex('by_active_updated', (q) => q.eq('softDeletedAt', undefined))
.order('desc')
.take(FALLBACK_SCAN_LIMIT)
for (const skill of recentSkills) {
if (seenSkillIds.has(skill._id)) continue
seenSkillIds.add(skill._id)
candidateSkills.push(skill)
}
const matched = candidateSkills.filter((skill) =>
matchesExactTokens(args.queryTokens, [skill.displayName, skill.slug, skill.summary]),
)
if (matched.length === 0) return []
const ownerHandleCache = new Map<Id<'users'>, Promise<string | null>>()
const getOwnerHandle = (ownerUserId: Id<'users'>) => {
const cached = ownerHandleCache.get(ownerUserId)
if (cached) return cached
const handlePromise = ctx.db
.get(ownerUserId)
.then((owner) => owner?.handle ?? owner?._id ?? null)
ownerHandleCache.set(ownerUserId, handlePromise)
return handlePromise
}
const entries = await Promise.all(
matched.map(async (skill) => {
const [version, ownerHandle] = await Promise.all([
skill.latestVersionId ? ctx.db.get(skill.latestVersionId) : Promise.resolve(null),
getOwnerHandle(skill.ownerUserId),
])
const publicSkill = toPublicSkill(skill)
if (!publicSkill) return null
return { skill: publicSkill, version, ownerHandle }
}),
)
const validEntries = entries.filter((entry): entry is SkillSearchEntry => entry !== null)
if (validEntries.length === 0) return []
const badgeMap = await getSkillBadgeMaps(
ctx,
validEntries.map((entry) => entry.skill._id),
)
const withBadges = validEntries.map((entry) => ({
...entry,
skill: {
...entry.skill,
badges: badgeMap.get(entry.skill._id) ?? {},
},
}))
const filtered = args.highlightedOnly
? withBadges.filter((entry) => isSkillHighlighted(entry.skill))
: withBadges
return filtered.slice(0, limit)
},
})
@@ -251,4 +424,10 @@ export const getSkillBadgeMapsInternal = internalQuery({
},
})
export const __test = { getNextCandidateLimit }
export const __test = {
getNextCandidateLimit,
matchesAllTokens,
getLexicalBoost,
scoreSkillResult,
mergeUniqueBySkillId,
}
+110
View File
@@ -0,0 +1,110 @@
/* @vitest-environment node */
import { describe, expect, it } from 'vitest'
// Test the aggregateEvents function by importing and testing the module logic
// Since aggregateEvents is not exported, we test the behavior indirectly through
// the event processing contract
describe('skill stat events - comment delta handling', () => {
it('aggregates comment and uncomment events into net deltas', () => {
// Simulate the aggregation logic from processSkillStatEventsAction
type EventKind =
| 'download'
| 'star'
| 'unstar'
| 'comment'
| 'uncomment'
| 'install_new'
| 'install_reactivate'
| 'install_deactivate'
| 'install_clear'
const events: { kind: EventKind; occurredAt: number }[] = [
{ kind: 'star', occurredAt: 1000 },
{ kind: 'comment', occurredAt: 2000 },
{ kind: 'comment', occurredAt: 3000 },
{ kind: 'uncomment', occurredAt: 4000 },
{ kind: 'download', occurredAt: 5000 },
]
// Replicate the aggregation logic
const result = {
downloads: 0,
stars: 0,
comments: 0,
installsAllTime: 0,
installsCurrent: 0,
downloadEvents: [] as number[],
installNewEvents: [] as number[],
}
for (const event of events) {
switch (event.kind) {
case 'download':
result.downloads += 1
result.downloadEvents.push(event.occurredAt)
break
case 'star':
result.stars += 1
break
case 'unstar':
result.stars -= 1
break
case 'comment':
result.comments += 1
break
case 'uncomment':
result.comments -= 1
break
case 'install_new':
result.installsAllTime += 1
result.installsCurrent += 1
result.installNewEvents.push(event.occurredAt)
break
case 'install_reactivate':
result.installsCurrent += 1
break
case 'install_deactivate':
result.installsCurrent -= 1
break
}
}
expect(result.stars).toBe(1)
expect(result.comments).toBe(1) // 2 comments - 1 uncomment
expect(result.downloads).toBe(1)
expect(result.downloadEvents).toEqual([5000])
})
it('should include comments in delta check (regression test for dropped comments)', () => {
// This test verifies the fix: the condition guard in applyAggregatedStatsAndUpdateCursor
// must include comments !== 0 so comment-only batches are not skipped
const delta = {
downloads: 0,
stars: 0,
comments: 3,
installsAllTime: 0,
installsCurrent: 0,
}
// The OLD buggy condition (missing comments):
const oldCondition =
delta.downloads !== 0 ||
delta.stars !== 0 ||
delta.installsAllTime !== 0 ||
delta.installsCurrent !== 0
// The FIXED condition (includes comments):
const fixedCondition =
delta.downloads !== 0 ||
delta.stars !== 0 ||
delta.comments !== 0 ||
delta.installsAllTime !== 0 ||
delta.installsCurrent !== 0
// With only comment deltas, the old condition would skip the patch
expect(oldCondition).toBe(false)
// The fixed condition correctly triggers the patch
expect(fixedCondition).toBe(true)
})
})
+2
View File
@@ -379,12 +379,14 @@ export const applyAggregatedStatsAndUpdateCursor = internalMutation({
if (
delta.downloads !== 0 ||
delta.stars !== 0 ||
delta.comments !== 0 ||
delta.installsAllTime !== 0 ||
delta.installsCurrent !== 0
) {
const patch = applySkillStatDeltas(skill, {
downloads: delta.downloads,
stars: delta.stars,
comments: delta.comments,
installsAllTime: delta.installsAllTime,
installsCurrent: delta.installsCurrent,
})
+230 -13
View File
@@ -1420,9 +1420,11 @@ export const getPendingScanSkillsInternal = internalQuery({
const skipRecentMinutes = args.skipRecentMinutes ?? 60
const skipThreshold = Date.now() - skipRecentMinutes * 60 * 1000
// Fetch more than needed so we can randomize selection
// Fetch more than needed so we can randomize selection.
// Include newly-published skills (hidden/pending.scan), skills stuck at
// scanner.vt.pending, AND LLM-evaluated skills that still need VT results.
const poolSize = Math.min(limit * 3, 500)
const allSkills = await ctx.db
const pendingScan = await ctx.db
.query('skills')
.filter((q) =>
q.and(
@@ -1431,6 +1433,36 @@ export const getPendingScanSkillsInternal = internalQuery({
),
)
.take(poolSize)
const vtPending = await ctx.db
.query('skills')
.filter((q) =>
q.and(
q.eq(q.field('moderationStatus'), 'active'),
q.eq(q.field('moderationReason'), 'scanner.vt.pending'),
),
)
.take(poolSize)
// LLM-evaluated skills whose VT scan hasn't completed yet
const llmEvaluated = await ctx.db
.query('skills')
.filter((q) =>
q.or(
q.eq(q.field('moderationReason'), 'scanner.llm.clean'),
q.eq(q.field('moderationReason'), 'scanner.llm.suspicious'),
q.eq(q.field('moderationReason'), 'scanner.llm.malicious'),
),
)
.take(poolSize)
// Dedup across pools by skill ID
const seen = new Set<string>()
const allSkills: typeof pendingScan = []
for (const skill of [...pendingScan, ...vtPending, ...llmEvaluated]) {
if (!seen.has(skill._id)) {
seen.add(skill._id)
allSkills.push(skill)
}
}
// Filter out recently checked skills
const skills = allSkills.filter(
@@ -1453,6 +1485,8 @@ export const getPendingScanSkillsInternal = internalQuery({
for (const skill of selected) {
const version = skill.latestVersionId ? await ctx.db.get(skill.latestVersionId) : null
// Skip skills where version already has vtAnalysis or lacks sha256hash
if (version?.vtAnalysis || !version?.sha256hash) continue
results.push({
skillId: skill._id,
versionId: version?._id ?? null,
@@ -1514,8 +1548,10 @@ export const getActiveSkillsMissingVTCacheInternal = internalQuery({
args: { limit: v.optional(v.number()) },
handler: async (ctx, args) => {
const limit = args.limit ?? 100
// Use scanner.vt.pending filter to only get skills waiting for VT
const pendingSkills = await ctx.db
const poolSize = limit * 2 // Take more to account for some having vtAnalysis
// Skills waiting for VT + LLM-evaluated skills that still need VT cache
const vtPending = await ctx.db
.query('skills')
.filter((q) =>
q.and(
@@ -1523,7 +1559,27 @@ export const getActiveSkillsMissingVTCacheInternal = internalQuery({
q.eq(q.field('moderationReason'), 'scanner.vt.pending'),
),
)
.take(limit * 2) // Take more to account for some having vtAnalysis
.take(poolSize)
const llmEvaluated = await ctx.db
.query('skills')
.filter((q) =>
q.or(
q.eq(q.field('moderationReason'), 'scanner.llm.clean'),
q.eq(q.field('moderationReason'), 'scanner.llm.suspicious'),
q.eq(q.field('moderationReason'), 'scanner.llm.malicious'),
),
)
.take(poolSize)
// Dedup across pools
const seen = new Set<string>()
const allSkills: typeof vtPending = []
for (const skill of [...vtPending, ...llmEvaluated]) {
if (!seen.has(skill._id)) {
seen.add(skill._id)
allSkills.push(skill)
}
}
const results: Array<{
skillId: Id<'skills'>
@@ -1532,7 +1588,7 @@ export const getActiveSkillsMissingVTCacheInternal = internalQuery({
slug: string
}> = []
for (const skill of pendingSkills) {
for (const skill of allSkills) {
if (results.length >= limit) break
if (!skill.latestVersionId) continue
const version = await ctx.db.get(skill.latestVersionId)
@@ -1642,6 +1698,58 @@ export const getActiveSkillBatchForRescanInternal = internalQuery({
},
})
/**
* Get active skills whose latest version has no llmAnalysis.
* Used for LLM evaluation backfill. Same cursor pattern as getActiveSkillBatchForRescanInternal.
*/
export const getActiveSkillBatchForLlmBackfillInternal = internalQuery({
args: {
cursor: v.optional(v.number()),
batchSize: v.optional(v.number()),
},
handler: async (ctx, args) => {
const batchSize = args.batchSize ?? 10
const cursor = args.cursor ?? 0
const candidates = await ctx.db
.query('skills')
.filter((q) => q.gt(q.field('_creationTime'), cursor))
.order('asc')
.take(batchSize * 3)
const results: Array<{
skillId: Id<'skills'>
versionId: Id<'skillVersions'>
slug: string
}> = []
let nextCursor = cursor
for (const skill of candidates) {
nextCursor = skill._creationTime
if (results.length >= batchSize) break
if (skill.softDeletedAt) continue
if ((skill.moderationStatus ?? 'active') !== 'active') continue
if (!skill.latestVersionId) continue
const version = await ctx.db.get(skill.latestVersionId)
if (!version) continue
// Re-evaluate all skills (full file content reading upgrade)
// if (version.llmAnalysis && version.llmAnalysis.status !== 'error') continue
results.push({
skillId: skill._id,
versionId: version._id,
slug: skill.slug,
})
}
const done = candidates.length < batchSize * 3
return { skills: results, nextCursor, done }
},
})
/**
* Get skills with stale moderationReason that have vtAnalysis cached.
* Used to sync moderationReason with cached VT results.
@@ -1936,6 +2044,37 @@ export const updateVersionScanResultsInternal = internalMutation({
},
})
export const updateVersionLlmAnalysisInternal = internalMutation({
args: {
versionId: v.id('skillVersions'),
llmAnalysis: v.object({
status: v.string(),
verdict: v.optional(v.string()),
confidence: v.optional(v.string()),
summary: v.optional(v.string()),
dimensions: v.optional(
v.array(
v.object({
name: v.string(),
label: v.string(),
rating: v.string(),
detail: v.string(),
}),
),
),
guidance: v.optional(v.string()),
findings: v.optional(v.string()),
model: v.optional(v.string()),
checkedAt: v.number(),
}),
},
handler: async (ctx, args) => {
const version = await ctx.db.get(args.versionId)
if (!version) return
await ctx.db.patch(args.versionId, { llmAnalysis: args.llmAnalysis })
},
})
export const approveSkillByHashInternal = internalMutation({
args: {
sha256hash: v.string(),
@@ -1956,17 +2095,37 @@ export const approveSkillByHashInternal = internalMutation({
if (skill) {
const isMalicious = args.status === 'malicious'
const isSuspicious = args.status === 'suspicious'
const isClean = !isMalicious && !isSuspicious
// Defense-in-depth: read existing flags to merge scanner results.
// The stricter verdict always wins across scanners.
const existingFlags: string[] = (skill.moderationFlags as string[] | undefined) ?? []
const existingReason: string | undefined = skill.moderationReason as string | undefined
const alreadyBlocked = existingFlags.includes('blocked.malware')
const alreadyFlagged = existingFlags.includes('flagged.suspicious')
// Determine new flags based on multi-scanner merge
let newFlags: string[] | undefined
if (isMalicious || alreadyBlocked) {
// Malicious from ANY scanner → blocked.malware (upgrade from suspicious)
newFlags = ['blocked.malware']
} else if (isSuspicious || alreadyFlagged) {
// Suspicious from ANY scanner → flagged.suspicious
newFlags = ['flagged.suspicious']
} else if (isClean) {
// Clean from this scanner — only clear if no other scanner has flagged
const otherScannerFlagged =
existingReason?.startsWith('scanner.') &&
!existingReason.startsWith(`scanner.${args.scanner}.`) &&
!existingReason.endsWith('.clean') &&
!existingReason.endsWith('.pending')
newFlags = otherScannerFlagged ? existingFlags : undefined
}
// Malicious/suspicious skills are visible (transparency) but not indexed
// Malicious skills have downloads blocked via moderationFlags
await ctx.db.patch(skill._id, {
moderationStatus: 'active', // Always visible for transparency
moderationReason: `scanner.${args.scanner}.${args.status}`,
moderationFlags: isMalicious
? ['blocked.malware']
: isSuspicious
? ['flagged.suspicious']
: undefined,
moderationFlags: newFlags,
updatedAt: Date.now(),
})
@@ -1983,6 +2142,64 @@ export const approveSkillByHashInternal = internalMutation({
return { ok: true, skillId: version.skillId, versionId: version._id }
},
})
/**
* Lighter VT-only escalation: adds moderation flags and hides/bans for malicious,
* but never touches moderationReason (preserves the LLM verdict).
*/
export const escalateByVtInternal = internalMutation({
args: {
sha256hash: v.string(),
status: v.union(v.literal('malicious'), v.literal('suspicious')),
},
handler: async (ctx, args) => {
const version = await ctx.db
.query('skillVersions')
.withIndex('by_sha256hash', (q) => q.eq('sha256hash', args.sha256hash))
.unique()
if (!version) throw new Error('Version not found for hash')
const skill = await ctx.db.get(version.skillId)
if (!skill) return
const isMalicious = args.status === 'malicious'
const existingFlags: string[] = (skill.moderationFlags as string[] | undefined) ?? []
const alreadyBlocked = existingFlags.includes('blocked.malware')
// Determine new flags — stricter verdict always wins
let newFlags: string[]
if (isMalicious || alreadyBlocked) {
newFlags = ['blocked.malware']
} else {
newFlags = ['flagged.suspicious']
}
const patch: Record<string, unknown> = {
moderationFlags: newFlags,
updatedAt: Date.now(),
}
// Only hide for malicious — suspicious stays visible with a flag
if (isMalicious) {
patch.moderationStatus = 'hidden'
}
await ctx.db.patch(skill._id, patch)
// Auto-ban authors of malicious skills
if (isMalicious && skill.ownerUserId) {
await ctx.scheduler.runAfter(0, internal.users.autobanMalwareAuthorInternal, {
ownerUserId: skill.ownerUserId,
sha256hash: args.sha256hash,
slug: skill.slug,
})
}
return { ok: true, skillId: version.skillId, versionId: version._id }
},
})
export const getVersionBySkillAndVersion = query({
args: { skillId: v.id('skills'), version: v.string() },
handler: async (ctx, args) => {
+96
View File
@@ -200,6 +200,102 @@ function buildSkillStatPatch(skill: Doc<'skills'>) {
}
}
/**
* Reconcile skill stats by counting actual records in source-of-truth tables.
*
* This fixes stats that got out of sync due to missed events, cursor issues,
* or bugs in the event processing pipeline. It counts:
* - stars: actual records in the `stars` table for each skill
* - comments: actual records in the `comments` table for each skill
*
* Downloads and installs are event-sourced only (no separate table to count from),
* so they cannot be reconciled this way.
*/
export const reconcileSkillStarCounts = internalMutation({
args: {
cursor: v.optional(v.string()),
batchSize: v.optional(v.number()),
},
handler: async (ctx, args) => {
const batchSize = clampInt(args.batchSize ?? 50, 1, 200)
const now = Date.now()
const { page, isDone, continueCursor } = await ctx.db
.query('skills')
.order('asc')
.paginate({ cursor: args.cursor ?? null, numItems: batchSize })
let patched = 0
for (const skill of page) {
// Count actual star records for this skill
const starRecords = await ctx.db
.query('stars')
.withIndex('by_skill_user', (q) => q.eq('skillId', skill._id))
.collect()
const actualStars = starRecords.length
// Count actual comment records for this skill
const commentRecords = await ctx.db
.query('comments')
.withIndex('by_skill', (q) => q.eq('skillId', skill._id))
.collect()
const actualComments = commentRecords.filter((c) => !c.softDeletedAt).length
// Check if stats are out of sync
if (skill.stats.stars !== actualStars || skill.stats.comments !== actualComments) {
const updatedStats = {
...skill.stats,
stars: actualStars,
comments: actualComments,
}
await ctx.db.patch(skill._id, {
statsStars: actualStars,
stats: updatedStats,
updatedAt: now,
})
patched += 1
}
}
return {
scanned: page.length,
patched,
cursor: isDone ? null : continueCursor,
isDone,
}
},
})
export const runReconcileSkillStarCountsInternal = internalAction({
args: {
batchSize: v.optional(v.number()),
maxBatches: v.optional(v.number()),
},
handler: async (ctx, args) => {
const batchSize = clampInt(args.batchSize ?? 50, 1, 200)
const maxBatches = clampInt(args.maxBatches ?? 10, 1, 50)
let cursor: string | undefined
let totalScanned = 0
let totalPatched = 0
for (let i = 0; i < maxBatches; i++) {
const result = (await ctx.runMutation(internal.statsMaintenance.reconcileSkillStarCounts, {
cursor,
batchSize,
})) as { scanned: number; patched: number; cursor: string | null; isDone: boolean }
totalScanned += result.scanned
totalPatched += result.patched
if (result.isDone) break
cursor = result.cursor ?? undefined
}
return { scanned: totalScanned, patched: totalPatched }
},
})
function clampInt(value: number, min: number, max: number) {
return Math.min(Math.max(value, min), max)
}
+27 -29
View File
@@ -393,21 +393,14 @@ export const scanWithVirusTotal = internalAction({
},
})
if (isSafe) {
await ctx.runMutation(internal.skills.approveSkillByHashInternal, {
// VT is supplementary — only escalate (never override LLM verdict)
if (!isSafe && (status === 'malicious' || status === 'suspicious')) {
await ctx.runMutation(internal.skills.escalateByVtInternal, {
sha256hash,
scanner: 'vt',
status: 'clean',
moderationStatus: 'active',
})
} else if (status === 'malicious' || status === 'suspicious') {
await ctx.runMutation(internal.skills.approveSkillByHashInternal, {
sha256hash,
scanner: 'vt',
status,
moderationStatus: 'hidden',
})
}
// Clean VT result: vtAnalysis already written above — don't touch moderation
return
}
@@ -448,13 +441,9 @@ export const scanWithVirusTotal = internalAction({
`Successfully uploaded version ${args.versionId} to VT. Hash: ${sha256hash}. Analysis ID: ${result.data.id}`,
)
// Mark skill as pending scan so it enters the poll queue
// This prevents it from being picked up again by scanUnscannedSkills
await ctx.runMutation(internal.skills.approveSkillByHashInternal, {
sha256hash,
scanner: 'vt',
status: 'pending',
})
// Don't set moderation state to scanner.vt.pending here — the LLM eval
// runs concurrently and will set the initial moderation state. VT only
// updates moderation when it has an actual verdict (clean/suspicious/malicious).
} catch (error) {
console.error('Failed to upload to VirusTotal:', error)
}
@@ -529,12 +518,16 @@ export const pollPendingScans = internalAction({
const vtResult = await checkExistingFile(apiKey, sha256hash)
if (!vtResult) {
console.log(`[vt:pollPendingScans] Hash ${sha256hash} not found in VT yet`)
// Check if we've exceeded max attempts
// Check if we've exceeded max attempts — write stale vtAnalysis so it
// drops out of the poll query without overwriting LLM moderationReason
if (checkCount + 1 >= MAX_CHECK_COUNT) {
console.warn(
`[vt:pollPendingScans] Skill ${skillId} exceeded max checks, marking stale`,
)
await ctx.runMutation(internal.skills.markScanStaleInternal, { skillId })
await ctx.runMutation(internal.skills.updateVersionScanResultsInternal, {
versionId,
vtAnalysis: { status: 'stale', checkedAt: Date.now() },
})
staled++
}
continue
@@ -550,12 +543,16 @@ export const pollPendingScans = internalAction({
`[vt:pollPendingScans] Hash ${sha256hash} has no Code Insight, requesting rescan`,
)
await requestRescan(apiKey, sha256hash)
// Check if we've exceeded max attempts
// Check if we've exceeded max attempts — write stale vtAnalysis so it
// drops out of the poll query without overwriting LLM moderationReason
if (checkCount + 1 >= MAX_CHECK_COUNT) {
console.warn(
`[vt:pollPendingScans] Skill ${skillId} exceeded max checks, marking stale`,
)
await ctx.runMutation(internal.skills.markScanStaleInternal, { skillId })
await ctx.runMutation(internal.skills.updateVersionScanResultsInternal, {
versionId,
vtAnalysis: { status: 'stale', checkedAt: Date.now() },
})
staled++
}
continue
@@ -581,11 +578,13 @@ export const pollPendingScans = internalAction({
},
})
await ctx.runMutation(internal.skills.approveSkillByHashInternal, {
sha256hash,
scanner: 'vt',
status,
})
// VT is supplementary — only escalate for malicious/suspicious
if (status === 'malicious' || status === 'suspicious') {
await ctx.runMutation(internal.skills.escalateByVtInternal, {
sha256hash,
status,
})
}
updated++
} catch (error) {
console.error(`[vt:pollPendingScans] Error checking hash ${sha256hash}:`, error)
@@ -835,9 +834,8 @@ export const rescanActiveSkills = internalAction({
if (status === 'malicious' || status === 'suspicious') {
console.warn(`[vt:rescan] ${slug}: verdict changed to ${status}!`)
accFlaggedSkills.push({ slug, status })
await ctx.runMutation(internal.skills.approveSkillByHashInternal, {
await ctx.runMutation(internal.skills.escalateByVtInternal, {
sha256hash,
scanner: 'vt-rescan',
status,
})
accUpdated++
+1
View File
@@ -30,6 +30,7 @@ Ensure Convex env is set (auth + embeddings):
- `OPENAI_API_KEY`
- `SITE_URL` (your web app URL)
- Optional webhook env (see `docs/webhook.md`)
- Optional: `GITHUB_TOKEN` (recommended; raises GitHub account lookup limit used by publish gate)
## 2) Deploy web app (Vercel)
+4
View File
@@ -40,6 +40,10 @@ Response:
{ "results": [{ "score": 0.123, "slug": "gifgrep", "displayName": "GifGrep", "summary": "…", "version": "1.2.3", "updatedAt": 1730000000000 }] }
```
Notes:
- Results are returned in relevance order (embedding similarity + exact slug/name token boosts + popularity prior from downloads).
### `GET /api/v1/skills`
Query params:
+4
View File
@@ -49,3 +49,7 @@ read_when:
- `githubFetchedAt` (fetch timestamp)
- Cache TTL: 24 hours.
- Gate applies to web uploads, CLI publish, and GitHub import.
- If GitHub responds `403` or `429`, publish fails with:
- `GitHub API rate limit exceeded — please try again in a few minutes`
- To reduce rate-limit failures, set `GITHUB_TOKEN` in Convex env for authenticated
GitHub API requests.
+93
View File
@@ -35,6 +35,99 @@ Workdir install state (written by the CLI):
- The server extracts metadata from frontmatter during publish.
- `description` is used as the skill summary in the UI/search.
## Frontmatter metadata
Skill metadata is declared in the YAML frontmatter at the top of your `SKILL.md`. This tells the registry (and security analysis) what your skill needs to run.
### Basic frontmatter
```yaml
---
name: my-skill
description: Short summary of what this skill does.
version: 1.0.0
---
```
### Runtime metadata (`metadata.openclaw`)
Declare your skill's runtime requirements under `metadata.openclaw` (aliases: `metadata.clawdbot`, `metadata.clawdis`).
```yaml
---
name: my-skill
description: Manage tasks via the Todoist API.
metadata:
openclaw:
requires:
env:
- TODOIST_API_KEY
bins:
- curl
primaryEnv: TODOIST_API_KEY
---
```
### Full field reference
| Field | Type | Description |
|-------|------|-------------|
| `requires.env` | `string[]` | Environment variables your skill expects. |
| `requires.bins` | `string[]` | CLI binaries that must all be installed. |
| `requires.anyBins` | `string[]` | CLI binaries where at least one must exist. |
| `requires.config` | `string[]` | Config file paths your skill reads. |
| `primaryEnv` | `string` | The main credential env var for your skill. |
| `always` | `boolean` | If `true`, skill is always active (no explicit install needed). |
| `skillKey` | `string` | Override the skill's invocation key. |
| `emoji` | `string` | Display emoji for the skill. |
| `homepage` | `string` | URL to the skill's homepage or docs. |
| `os` | `string[]` | OS restrictions (e.g. `["macos"]`, `["linux"]`). |
| `install` | `array` | Install specs for dependencies (see below). |
| `nix` | `object` | Nix plugin spec (see README). |
| `config` | `object` | Clawdbot config spec (see README). |
### Install specs
If your skill needs dependencies installed, declare them in the `install` array:
```yaml
metadata:
openclaw:
install:
- kind: brew
formula: jq
bins: [jq]
- kind: node
package: typescript
bins: [tsc]
```
Supported install kinds: `brew`, `node`, `go`, `uv`.
### Why this matters
ClawHub's security analysis checks that what your skill declares matches what it actually does. If your code references `TODOIST_API_KEY` but your frontmatter doesn't declare it under `requires.env`, the analysis will flag a metadata mismatch. Keeping declarations accurate helps your skill pass review and helps users understand what they're installing.
### Example: complete frontmatter
```yaml
---
name: todoist-cli
description: Manage Todoist tasks, projects, and labels from the command line.
version: 1.2.0
metadata:
openclaw:
requires:
env:
- TODOIST_API_KEY
bins:
- curl
primaryEnv: TODOIST_API_KEY
emoji: "\u2705"
homepage: https://github.com/example/todoist-cli
---
```
## Allowed files
Only “text-based” files are accepted by publish.
+6
View File
@@ -23,6 +23,12 @@ read_when:
- Set `OPENAI_API_KEY` in the Convex environment (not only locally).
- Re-run `bunx convex dev` / `bunx convex deploy` after setting env.
## `publish` fails with `GitHub API rate limit exceeded`
- This is the GitHub account-age gate lookup hitting unauthenticated limits.
- Set `GITHUB_TOKEN` in Convex environment to use authenticated GitHub API limits.
- Retry publish after a short wait if the limit was already exhausted.
## `sync` says “No skills found”
- `sync` looks for folders containing `SKILL.md` (or `skill.md`).
-1
View File
@@ -20,7 +20,6 @@ const REQUEST_TIMEOUT_MS = 15_000
try {
setGlobalDispatcher(
new Agent({
allowH2: true,
connect: { timeout: REQUEST_TIMEOUT_MS },
}),
)
+1
View File
@@ -76,6 +76,7 @@
"only-allow": "^1.2.2",
"oxlint": "^1.42.0",
"oxlint-tsgolint": "^0.11.4",
"undici": "^7.19.2",
"typescript": "^5.9.3",
"vite": "^7.3.1",
"vitest": "^4.0.18"
+1 -1
View File
@@ -1,6 +1,6 @@
{
"name": "clawhub",
"version": "0.6.0",
"version": "0.6.1",
"description": "ClawHub CLI \\u2014 install, update, search, and publish agent skills.",
"license": "MIT",
"type": "module",
+73
View File
@@ -0,0 +1,73 @@
/* @vitest-environment node */
import { afterEach, beforeEach, describe, expect, it, vi } from 'vitest'
const chmodMock = vi.fn()
const mkdirMock = vi.fn()
const readFileMock = vi.fn()
const writeFileMock = vi.fn()
vi.mock('node:fs/promises', () => ({
chmod: (...args: unknown[]) => chmodMock(...args),
mkdir: (...args: unknown[]) => mkdirMock(...args),
readFile: (...args: unknown[]) => readFileMock(...args),
writeFile: (...args: unknown[]) => writeFileMock(...args),
}))
const { writeGlobalConfig } = await import('./config')
const originalPlatform = process.platform
const testConfigPath = '/tmp/clawhub-config-test/config.json'
function makeErr(code: string): NodeJS.ErrnoException {
const error = new Error(code) as NodeJS.ErrnoException
error.code = code
return error
}
beforeEach(() => {
vi.stubEnv('CLAWHUB_CONFIG_PATH', testConfigPath)
Object.defineProperty(process, 'platform', { value: 'linux' })
chmodMock.mockResolvedValue(undefined)
mkdirMock.mockResolvedValue(undefined)
readFileMock.mockResolvedValue('')
writeFileMock.mockResolvedValue(undefined)
})
afterEach(() => {
Object.defineProperty(process, 'platform', { value: originalPlatform })
vi.unstubAllEnvs()
vi.clearAllMocks()
})
describe('writeGlobalConfig', () => {
it('writes config with restricted modes', async () => {
await writeGlobalConfig({ registry: 'https://example.com', token: 'clh_test' })
expect(mkdirMock).toHaveBeenCalledWith('/tmp/clawhub-config-test', {
recursive: true,
mode: 0o700,
})
expect(writeFileMock).toHaveBeenCalledWith(
testConfigPath,
expect.stringContaining('"token": "clh_test"'),
{
encoding: 'utf8',
mode: 0o600,
},
)
expect(chmodMock).toHaveBeenCalledWith(testConfigPath, 0o600)
})
it('ignores non-fatal chmod errors', async () => {
chmodMock.mockRejectedValueOnce(makeErr('ENOTSUP'))
await expect(writeGlobalConfig({ registry: 'https://example.com' })).resolves.toBeUndefined()
})
it('rethrows unexpected chmod errors', async () => {
chmodMock.mockRejectedValueOnce(new Error('boom'))
await expect(writeGlobalConfig({ registry: 'https://example.com' })).rejects.toThrow('boom')
})
})
+48 -23
View File
@@ -1,44 +1,51 @@
import { existsSync } from 'node:fs'
import { mkdir, readFile, writeFile } from 'node:fs/promises'
import { chmod, mkdir, readFile, writeFile } from 'node:fs/promises'
import { homedir } from 'node:os'
import { dirname, join, resolve } from 'node:path'
import { type GlobalConfig, GlobalConfigSchema, parseArk } from './schema/index.js'
/**
* Resolve config path with legacy fallback.
* Checks for 'clawhub' first, falls back to legacy 'clawdhub' if it exists.
*/
function resolveConfigPath(baseDir: string): string {
const clawhubPath = join(baseDir, 'clawhub', 'config.json')
const clawdhubPath = join(baseDir, 'clawdhub', 'config.json')
if (existsSync(clawhubPath)) return clawhubPath
if (existsSync(clawdhubPath)) return clawdhubPath
return clawhubPath
}
function isNonFatalChmodError(error: unknown): boolean {
if (!(error instanceof Error)) return false
const code = (error as NodeJS.ErrnoException).code
return code === 'EPERM' || code === 'ENOTSUP' || code === 'EOPNOTSUPP' || code === 'EINVAL'
}
export function getGlobalConfigPath() {
const override =
process.env.CLAWHUB_CONFIG_PATH?.trim() ?? process.env.CLAWDHUB_CONFIG_PATH?.trim()
if (override) return resolve(override)
const home = homedir()
if (process.platform === 'darwin') {
const clawhubPath = join(home, 'Library', 'Application Support', 'clawhub', 'config.json')
const clawdhubPath = join(home, 'Library', 'Application Support', 'clawdhub', 'config.json')
if (existsSync(clawhubPath)) return clawhubPath
if (existsSync(clawdhubPath)) return clawdhubPath
return clawhubPath
return resolveConfigPath(join(home, 'Library', 'Application Support'))
}
const xdg = process.env.XDG_CONFIG_HOME
if (xdg) {
const clawhubPath = join(xdg, 'clawhub', 'config.json')
const clawdhubPath = join(xdg, 'clawdhub', 'config.json')
if (existsSync(clawhubPath)) return clawhubPath
if (existsSync(clawdhubPath)) return clawdhubPath
return clawhubPath
return resolveConfigPath(xdg)
}
if (process.platform === 'win32') {
const appData = process.env.APPDATA
if (appData) {
const clawhubPath = join(appData, 'clawhub', 'config.json')
const clawdhubPath = join(appData, 'clawdhub', 'config.json')
if (existsSync(clawhubPath)) return clawhubPath
if (existsSync(clawdhubPath)) return clawdhubPath
return clawhubPath
return resolveConfigPath(appData)
}
}
const clawhubPath = join(home, '.config', 'clawhub', 'config.json')
const clawdhubPath = join(home, '.config', 'clawdhub', 'config.json')
if (existsSync(clawhubPath)) return clawhubPath
if (existsSync(clawdhubPath)) return clawdhubPath
return clawhubPath
return resolveConfigPath(join(home, '.config'))
}
export async function readGlobalConfig(): Promise<GlobalConfig | null> {
@@ -53,6 +60,24 @@ export async function readGlobalConfig(): Promise<GlobalConfig | null> {
export async function writeGlobalConfig(config: GlobalConfig) {
const path = getGlobalConfigPath()
await mkdir(dirname(path), { recursive: true })
await writeFile(path, `${JSON.stringify(config, null, 2)}\n`, 'utf8')
const dir = dirname(path)
// Create directory with restricted permissions (owner only)
await mkdir(dir, { recursive: true, mode: 0o700 })
// Write file with restricted permissions (owner read/write only)
// This protects API tokens from being read by other users
await writeFile(path, `${JSON.stringify(config, null, 2)}\n`, {
encoding: 'utf8',
mode: 0o600,
})
// Ensure permissions on existing files (writeFile mode only applies on create)
if (process.platform !== 'win32') {
try {
await chmod(path, 0o600)
} catch (error) {
if (!isNonFatalChmodError(error)) throw error
}
}
}
-1
View File
@@ -15,7 +15,6 @@ if (typeof process !== 'undefined' && process.versions?.node) {
try {
setGlobalDispatcher(
new Agent({
allowH2: true,
connect: { timeout: REQUEST_TIMEOUT_MS },
}),
)
+78
View File
@@ -0,0 +1,78 @@
import { describe, expect, it, vi } from 'vitest'
vi.mock('@tanstack/react-router', () => ({
createFileRoute: () => (config: { beforeLoad?: unknown }) => ({ __config: config }),
redirect: (options: unknown) => ({ redirect: options }),
}))
import { Route } from '../routes/search'
function runBeforeLoad(search: { q?: string; highlighted?: boolean }, hostname = 'clawdhub.com') {
const route = Route as unknown as {
__config: {
beforeLoad?: (args: {
search: { q?: string; highlighted?: boolean }
location: { url: URL }
}) => void
}
}
const beforeLoad = route.__config.beforeLoad as (args: {
search: { q?: string; highlighted?: boolean }
location: { url: URL }
}) => void
let thrown: unknown
try {
beforeLoad({ search, location: { url: new URL(`https://${hostname}/search`) } })
} catch (error) {
thrown = error
}
return thrown
}
describe('search route', () => {
it('redirects skills host to the skills index', () => {
expect(runBeforeLoad({ q: 'crab', highlighted: true }, 'clawdhub.com')).toEqual({
redirect: {
to: '/skills',
search: {
q: 'crab',
sort: undefined,
dir: undefined,
highlighted: true,
view: undefined,
},
replace: true,
},
})
})
it('redirects souls host with query to home search', () => {
expect(runBeforeLoad({ q: 'crab', highlighted: true }, 'onlycrabs.ai')).toEqual({
redirect: {
to: '/',
search: {
q: 'crab',
highlighted: undefined,
search: undefined,
},
replace: true,
},
})
})
it('redirects souls host without query to home with search mode', () => {
expect(runBeforeLoad({}, 'onlycrabs.ai')).toEqual({
redirect: {
to: '/',
search: {
q: undefined,
highlighted: undefined,
search: true,
},
replace: true,
},
})
})
})
@@ -0,0 +1,115 @@
/* @vitest-environment jsdom */
import { act, render } from '@testing-library/react'
import type { ReactNode } from 'react'
import { afterEach, beforeEach, describe, expect, it, vi } from 'vitest'
import { SkillsIndex } from '../routes/skills/index'
const navigateMock = vi.fn()
const useActionMock = vi.fn()
const usePaginatedQueryMock = vi.fn()
let searchMock: Record<string, unknown> = {}
vi.mock('@tanstack/react-router', () => ({
createFileRoute: () => (_config: { component: unknown; validateSearch: unknown }) => ({
useNavigate: () => navigateMock,
useSearch: () => searchMock,
}),
Link: (props: { children: ReactNode }) => <a href="/">{props.children}</a>,
}))
vi.mock('convex/react', () => ({
useAction: (...args: unknown[]) => useActionMock(...args),
}))
vi.mock('convex-helpers/react', () => ({
usePaginatedQuery: (...args: unknown[]) => usePaginatedQueryMock(...args),
}))
describe('SkillsIndex load-more observer', () => {
beforeEach(() => {
usePaginatedQueryMock.mockReset()
useActionMock.mockReset()
navigateMock.mockReset()
searchMock = {}
useActionMock.mockReturnValue(() => Promise.resolve([]))
})
afterEach(() => {
vi.unstubAllGlobals()
})
it('triggers one request for repeated intersection callbacks', async () => {
const loadMorePaginated = vi.fn()
usePaginatedQueryMock.mockReturnValue({
results: [makeListResult('skill-0', 'Skill 0')],
status: 'CanLoadMore',
loadMore: loadMorePaginated,
})
type ObserverInstance = {
callback: IntersectionObserverCallback
observe: ReturnType<typeof vi.fn>
disconnect: ReturnType<typeof vi.fn>
}
const observers: ObserverInstance[] = []
class IntersectionObserverMock {
callback: IntersectionObserverCallback
observe = vi.fn()
disconnect = vi.fn()
unobserve = vi.fn()
takeRecords = vi.fn(() => [])
root = null
rootMargin = '0px'
thresholds: number[] = []
constructor(callback: IntersectionObserverCallback) {
this.callback = callback
observers.push(this)
}
}
vi.stubGlobal(
'IntersectionObserver',
IntersectionObserverMock as unknown as typeof IntersectionObserver,
)
render(<SkillsIndex />)
expect(observers).toHaveLength(1)
const observer = observers[0]
const entries = [{ isIntersecting: true }] as Array<IntersectionObserverEntry>
await act(async () => {
observer.callback(entries, observer as unknown as IntersectionObserver)
observer.callback(entries, observer as unknown as IntersectionObserver)
observer.callback(entries, observer as unknown as IntersectionObserver)
})
expect(loadMorePaginated).toHaveBeenCalledTimes(1)
})
})
function makeListResult(slug: string, displayName: string) {
return {
skill: {
_id: `skill_${slug}`,
slug,
displayName,
summary: `${displayName} summary`,
tags: {},
stats: {
downloads: 0,
installsCurrent: 0,
installsAllTime: 0,
stars: 0,
versions: 1,
comments: 0,
},
createdAt: 0,
updatedAt: 0,
},
latestVersion: null,
ownerHandle: null,
}
}
+48
View File
@@ -118,6 +118,30 @@ describe('SkillsIndex', () => {
limit: 50,
})
})
it('uses relevance as default sort when searching', async () => {
searchMock = { q: 'notion' }
const actionFn = vi
.fn()
.mockResolvedValue([
makeSearchResult('newer-low-score', 'Newer Low Score', 0.1, 2000),
makeSearchResult('older-high-score', 'Older High Score', 0.9, 1000),
])
useActionMock.mockReturnValue(actionFn)
vi.useFakeTimers()
render(<SkillsIndex />)
await act(async () => {
await vi.runAllTimersAsync()
})
const titles = Array.from(
document.querySelectorAll('.skills-row-title > span:first-child'),
).map((node) => node.textContent)
expect(titles[0]).toBe('Older High Score')
expect(titles[1]).toBe('Newer Low Score')
})
})
function makeSearchResults(count: number) {
@@ -143,3 +167,27 @@ function makeSearchResults(count: number) {
version: null,
}))
}
function makeSearchResult(slug: string, displayName: string, score: number, createdAt: number) {
return {
score,
skill: {
_id: `skill_${slug}`,
slug,
displayName,
summary: `${displayName} summary`,
tags: {},
stats: {
downloads: 0,
installsCurrent: 0,
installsAllTime: 0,
stars: 0,
versions: 1,
comments: 0,
},
createdAt,
updatedAt: createdAt,
},
version: null,
}
}
+236 -37
View File
@@ -1,7 +1,7 @@
import { Link, useNavigate } from '@tanstack/react-router'
import type { ClawdisSkillMetadata, SkillInstallSpec } from 'clawhub-schema'
import { useAction, useMutation, useQuery } from 'convex/react'
import { useEffect, useMemo, useState } from 'react'
import { lazy, Suspense, useEffect, useMemo, useState } from 'react'
import ReactMarkdown from 'react-markdown'
import remarkGfm from 'remark-gfm'
import { api } from '../../convex/_generated/api'
@@ -10,7 +10,10 @@ import { getSkillBadges } from '../lib/badges'
import type { PublicSkill, PublicUser } from '../lib/publicUser'
import { canManageSkill, isModerator } from '../lib/roles'
import { useAuthStatus } from '../lib/useAuthStatus'
import { SkillDiffCard } from './SkillDiffCard'
const SkillDiffCard = lazy(() =>
import('./SkillDiffCard').then((m) => ({ default: m.SkillDiffCard })),
)
type VtAnalysis = {
status: string
@@ -20,6 +23,25 @@ type VtAnalysis = {
checkedAt: number
}
type LlmAnalysisDimension = {
name: string
label: string
rating: string
detail: string
}
type LlmAnalysis = {
status: string
verdict?: string
confidence?: string
summary?: string
dimensions?: LlmAnalysisDimension[]
guidance?: string
findings?: string
model?: string
checkedAt: number
}
function VirusTotalIcon({ className }: { className?: string }) {
return (
<svg
@@ -40,6 +62,35 @@ function VirusTotalIcon({ className }: { className?: string }) {
)
}
function OpenClawIcon({ className }: { className?: string }) {
return (
<svg
className={className}
width="1em"
height="1em"
viewBox="0 0 24 24"
fill="none"
aria-label="OpenClaw"
>
<title>OpenClaw</title>
<path
d="M12 2C8.5 2 5.5 4 4 7c-2 4-1 8 2 11 1.5 1.5 3.5 2.5 6 2.5s4.5-1 6-2.5c3-3 4-7 2-11-1.5-3-4.5-5-8-5z"
fill="currentColor"
opacity="0.2"
/>
<path
d="M9 8c1-2 3-3 5-2s3 3 2 5l-3 4-2-1 3-4c.5-1 0-2-1-2.5S11 7 10.5 8L8 12l-2-1 3-4z"
fill="currentColor"
/>
<path
d="M15 8c-1-2-3-3-5-2s-3 3-2 5l3 4 2-1-3-4c-.5-1 0-2 1-2.5S14 7 14.5 8L17 12l2-1-4-3z"
fill="currentColor"
opacity="0.6"
/>
</svg>
)
}
function getScanStatusInfo(status: string) {
switch (status.toLowerCase()) {
case 'benign':
@@ -62,40 +113,139 @@ function getScanStatusInfo(status: string) {
}
}
function getDimensionIcon(rating: string) {
switch (rating) {
case 'ok':
return { className: 'dimension-icon-ok', symbol: '\u2713' }
case 'note':
return { className: 'dimension-icon-note', symbol: '\u2139' }
case 'concern':
return { className: 'dimension-icon-concern', symbol: '!' }
default:
return { className: 'dimension-icon-danger', symbol: '\u2717' }
}
}
function LlmAnalysisDetail({ analysis }: { analysis: LlmAnalysis }) {
const verdict = analysis.verdict ?? analysis.status
const [isOpen, setIsOpen] = useState(false)
const guidanceClass =
verdict === 'malicious' ? 'malicious' : verdict === 'suspicious' ? 'suspicious' : 'benign'
return (
<div className={`analysis-detail${isOpen ? ' is-open' : ''}`}>
<button
type="button"
className="analysis-detail-header"
onClick={() => setIsOpen((prev) => !prev)}
aria-expanded={isOpen}
>
<span className="analysis-summary-text">{analysis.summary}</span>
<span className="analysis-detail-toggle">
Details <span className="chevron">{'\u25BE'}</span>
</span>
</button>
<div className="analysis-body">
{analysis.dimensions && analysis.dimensions.length > 0 ? (
<div className="analysis-dimensions">
{analysis.dimensions.map((dim) => {
const icon = getDimensionIcon(dim.rating)
return (
<div key={dim.name} className="dimension-row">
<div className={`dimension-icon ${icon.className}`}>{icon.symbol}</div>
<div className="dimension-content">
<div className="dimension-label">{dim.label}</div>
<div className="dimension-detail">{dim.detail}</div>
</div>
</div>
)
})}
</div>
) : null}
{analysis.findings ? (
<div className="scan-findings-section">
<div className="scan-findings-title">Scan Findings in Context</div>
{(() => {
const counts = new Map<string, number>()
return analysis.findings.split('\n').map((line) => {
const count = (counts.get(line) ?? 0) + 1
counts.set(line, count)
return (
<div key={`${line}-${count}`} className="scan-finding-row">
{line}
</div>
)
})
})()}
</div>
) : null}
{analysis.guidance ? (
<div className={`analysis-guidance ${guidanceClass}`}>
<div className="analysis-guidance-label">
{verdict === 'malicious'
? 'Do not install this skill'
: verdict === 'suspicious'
? 'What to consider before installing'
: 'Assessment'}
</div>
{analysis.guidance}
</div>
) : null}
</div>
</div>
)
}
function SecurityScanResults({
sha256hash,
vtAnalysis,
llmAnalysis,
variant = 'panel',
}: {
sha256hash?: string
vtAnalysis?: VtAnalysis | null
llmAnalysis?: LlmAnalysis | null
variant?: 'panel' | 'badge'
}) {
if (!sha256hash) return null
if (!sha256hash && !llmAnalysis) return null
const status = vtAnalysis?.status ?? 'pending'
const vtUrl = `https://www.virustotal.com/gui/file/${sha256hash}`
const statusInfo = getScanStatusInfo(status)
const vtStatus = vtAnalysis?.status ?? 'pending'
const vtUrl = sha256hash ? `https://www.virustotal.com/gui/file/${sha256hash}` : null
const vtStatusInfo = getScanStatusInfo(vtStatus)
const isCodeInsight = vtAnalysis?.source === 'code_insight'
const aiAnalysis = vtAnalysis?.analysis
const displayLabel = statusInfo.label
const llmVerdict = llmAnalysis?.verdict ?? llmAnalysis?.status
const llmStatusInfo = llmVerdict ? getScanStatusInfo(llmVerdict) : null
if (variant === 'badge') {
return (
<div className="version-scan-badge">
<VirusTotalIcon className="version-scan-icon version-scan-icon-vt" />
<span className={statusInfo.className}>{displayLabel}</span>
<a
href={vtUrl}
target="_blank"
rel="noopener noreferrer"
className="version-scan-link"
onClick={(e) => e.stopPropagation()}
>
</a>
</div>
<>
{sha256hash ? (
<div className="version-scan-badge">
<VirusTotalIcon className="version-scan-icon version-scan-icon-vt" />
<span className={vtStatusInfo.className}>{vtStatusInfo.label}</span>
{vtUrl ? (
<a
href={vtUrl}
target="_blank"
rel="noopener noreferrer"
className="version-scan-link"
onClick={(e) => e.stopPropagation()}
>
</a>
) : null}
</div>
) : null}
{llmStatusInfo ? (
<div className="version-scan-badge">
<OpenClawIcon className="version-scan-icon version-scan-icon-oc" />
<span className={llmStatusInfo.className}>{llmStatusInfo.label}</span>
</div>
) : null}
</>
)
}
@@ -103,22 +253,53 @@ function SecurityScanResults({
<div className="scan-results-panel">
<div className="scan-results-title">Security Scan</div>
<div className="scan-results-list">
<div className="scan-result-row">
<div className="scan-result-scanner">
<VirusTotalIcon className="scan-result-icon scan-result-icon-vt" />
<span className="scan-result-scanner-name">VirusTotal</span>
{sha256hash ? (
<div className="scan-result-row">
<div className="scan-result-scanner">
<VirusTotalIcon className="scan-result-icon scan-result-icon-vt" />
<span className="scan-result-scanner-name">VirusTotal</span>
</div>
<div className={`scan-result-status ${vtStatusInfo.className}`}>
{vtStatusInfo.label}
</div>
{vtUrl ? (
<a
href={vtUrl}
target="_blank"
rel="noopener noreferrer"
className="scan-result-link"
>
View report
</a>
) : null}
</div>
<div className={`scan-result-status ${statusInfo.className}`}>{displayLabel}</div>
<a href={vtUrl} target="_blank" rel="noopener noreferrer" className="scan-result-link">
View report
</a>
</div>
{isCodeInsight && aiAnalysis && (status === 'malicious' || status === 'suspicious') ? (
<div className={`code-insight-analysis ${status}`}>
) : null}
{isCodeInsight && aiAnalysis && (vtStatus === 'malicious' || vtStatus === 'suspicious') ? (
<div className={`code-insight-analysis ${vtStatus}`}>
<div className="code-insight-label">Code Insight</div>
<p className="code-insight-text">{aiAnalysis}</p>
</div>
) : null}
{llmStatusInfo && llmAnalysis ? (
<div className="scan-result-row">
<div className="scan-result-scanner">
<OpenClawIcon className="scan-result-icon scan-result-icon-oc" />
<span className="scan-result-scanner-name">OpenClaw</span>
</div>
<div className={`scan-result-status ${llmStatusInfo.className}`}>
{llmStatusInfo.label}
</div>
{llmAnalysis.confidence ? (
<span className="scan-result-confidence">{llmAnalysis.confidence} confidence</span>
) : null}
</div>
) : null}
{llmAnalysis &&
llmAnalysis.status !== 'error' &&
llmAnalysis.status !== 'pending' &&
llmAnalysis.summary ? (
<LlmAnalysisDetail analysis={llmAnalysis} />
) : null}
</div>
</div>
)
@@ -387,8 +568,8 @@ export function SkillDetailPage({
<div className="pending-banner-content">
<strong>Skill blocked malicious content detected</strong>
<p>
VirusTotal flagged this skill as malicious. Downloads are disabled. Review the scan
results below.
ClawHub Security flagged this skill as malicious. Downloads are disabled. Review the
scan results below.
</p>
</div>
</div>
@@ -397,8 +578,22 @@ export function SkillDetailPage({
<div className="pending-banner-content">
<strong>Skill flagged suspicious patterns detected</strong>
<p>
VirusTotal flagged this skill as suspicious. Review the scan results before using.
ClawHub Security flagged this skill as suspicious. Review the scan results before
using.
</p>
{canManage ? (
<p className="pending-banner-appeal">
If you believe this skill has been incorrectly flagged, please{' '}
<a
href="https://github.com/openclaw/clawhub/issues"
target="_blank"
rel="noopener noreferrer"
>
submit an issue on GitHub
</a>{' '}
and we'll break down why it was flagged and what you can do.
</p>
) : null}
</div>
</div>
) : modInfo?.isRemoved ? (
@@ -533,8 +728,9 @@ export function SkillDetailPage({
<SecurityScanResults
sha256hash={latestVersion?.sha256hash}
vtAnalysis={latestVersion?.vtAnalysis}
llmAnalysis={latestVersion?.llmAnalysis as LlmAnalysis | undefined}
/>
{latestVersion?.sha256hash ? (
{latestVersion?.sha256hash || latestVersion?.llmAnalysis ? (
<p className="scan-disclaimer">
Like a lobster shell, security has layers review code before you run it.
</p>
@@ -814,7 +1010,9 @@ export function SkillDetailPage({
) : null}
{activeTab === 'compare' && skill ? (
<div className="tab-body">
<SkillDiffCard skill={skill} versions={diffVersions ?? []} variant="embedded" />
<Suspense fallback={<div className="stat">Loading diff viewer</div>}>
<SkillDiffCard skill={skill} versions={diffVersions ?? []} variant="embedded" />
</Suspense>
</div>
) : null}
{activeTab === 'versions' ? (
@@ -844,10 +1042,11 @@ export function SkillDetailPage({
{version.changelog}
</div>
<div className="version-scan-results">
{version.sha256hash ? (
{version.sha256hash || version.llmAnalysis ? (
<SecurityScanResults
sha256hash={version.sha256hash}
vtAnalysis={version.vtAnalysis}
llmAnalysis={version.llmAnalysis as LlmAnalysis | undefined}
variant="badge"
/>
) : null}
+21
View File
@@ -12,6 +12,7 @@ import { Route as rootRouteImport } from './routes/__root'
import { Route as UploadRouteImport } from './routes/upload'
import { Route as StarsRouteImport } from './routes/stars'
import { Route as SettingsRouteImport } from './routes/settings'
import { Route as SearchRouteImport } from './routes/search'
import { Route as ManagementRouteImport } from './routes/management'
import { Route as ImportRouteImport } from './routes/import'
import { Route as DashboardRouteImport } from './routes/dashboard'
@@ -39,6 +40,11 @@ const SettingsRoute = SettingsRouteImport.update({
path: '/settings',
getParentRoute: () => rootRouteImport,
} as any)
const SearchRoute = SearchRouteImport.update({
id: '/search',
path: '/search',
getParentRoute: () => rootRouteImport,
} as any)
const ManagementRoute = ManagementRouteImport.update({
id: '/management',
path: '/management',
@@ -101,6 +107,7 @@ export interface FileRoutesByFullPath {
'/dashboard': typeof DashboardRoute
'/import': typeof ImportRoute
'/management': typeof ManagementRoute
'/search': typeof SearchRoute
'/settings': typeof SettingsRoute
'/stars': typeof StarsRoute
'/upload': typeof UploadRoute
@@ -117,6 +124,7 @@ export interface FileRoutesByTo {
'/dashboard': typeof DashboardRoute
'/import': typeof ImportRoute
'/management': typeof ManagementRoute
'/search': typeof SearchRoute
'/settings': typeof SettingsRoute
'/stars': typeof StarsRoute
'/upload': typeof UploadRoute
@@ -134,6 +142,7 @@ export interface FileRoutesById {
'/dashboard': typeof DashboardRoute
'/import': typeof ImportRoute
'/management': typeof ManagementRoute
'/search': typeof SearchRoute
'/settings': typeof SettingsRoute
'/stars': typeof StarsRoute
'/upload': typeof UploadRoute
@@ -152,6 +161,7 @@ export interface FileRouteTypes {
| '/dashboard'
| '/import'
| '/management'
| '/search'
| '/settings'
| '/stars'
| '/upload'
@@ -168,6 +178,7 @@ export interface FileRouteTypes {
| '/dashboard'
| '/import'
| '/management'
| '/search'
| '/settings'
| '/stars'
| '/upload'
@@ -184,6 +195,7 @@ export interface FileRouteTypes {
| '/dashboard'
| '/import'
| '/management'
| '/search'
| '/settings'
| '/stars'
| '/upload'
@@ -201,6 +213,7 @@ export interface RootRouteChildren {
DashboardRoute: typeof DashboardRoute
ImportRoute: typeof ImportRoute
ManagementRoute: typeof ManagementRoute
SearchRoute: typeof SearchRoute
SettingsRoute: typeof SettingsRoute
StarsRoute: typeof StarsRoute
UploadRoute: typeof UploadRoute
@@ -235,6 +248,13 @@ declare module '@tanstack/react-router' {
preLoaderRoute: typeof SettingsRouteImport
parentRoute: typeof rootRouteImport
}
'/search': {
id: '/search'
path: '/search'
fullPath: '/search'
preLoaderRoute: typeof SearchRouteImport
parentRoute: typeof rootRouteImport
}
'/management': {
id: '/management'
path: '/management'
@@ -321,6 +341,7 @@ const rootRouteChildren: RootRouteChildren = {
DashboardRoute: DashboardRoute,
ImportRoute: ImportRoute,
ManagementRoute: ManagementRoute,
SearchRoute: SearchRoute,
SettingsRoute: SettingsRoute,
StarsRoute: StarsRoute,
UploadRoute: UploadRoute,
+38
View File
@@ -0,0 +1,38 @@
import { createFileRoute, redirect } from '@tanstack/react-router'
import { detectSiteMode } from '../lib/site'
export const Route = createFileRoute('/search')({
validateSearch: (search) => ({
q: typeof search.q === 'string' && search.q.trim() ? search.q : undefined,
highlighted: search.highlighted === '1' || search.highlighted === 'true' ? true : undefined,
}),
beforeLoad: ({ search, location }) => {
const hostname =
(location as { url?: URL }).url?.hostname ??
(typeof window !== 'undefined' ? window.location.hostname : undefined)
const mode = detectSiteMode(hostname)
if (mode === 'skills') {
throw redirect({
to: '/skills',
search: {
q: search.q || undefined,
sort: undefined,
dir: undefined,
highlighted: search.highlighted || undefined,
view: undefined,
},
replace: true,
})
}
throw redirect({
to: '/',
search: {
q: search.q || undefined,
highlighted: undefined,
search: search.q ? undefined : true,
},
replace: true,
})
},
})
+30 -5
View File
@@ -8,7 +8,15 @@ import { SkillCard } from '../../components/SkillCard'
import { getSkillBadges, isSkillHighlighted } from '../../lib/badges'
import type { PublicSkill } from '../../lib/publicUser'
const sortKeys = ['newest', 'downloads', 'installs', 'stars', 'name', 'updated'] as const
const sortKeys = [
'relevance',
'newest',
'downloads',
'installs',
'stars',
'name',
'updated',
] as const
const pageSize = 25
type SortKey = (typeof sortKeys)[number]
type SortDir = 'asc' | 'desc'
@@ -28,6 +36,7 @@ type SkillListEntry = {
skill: PublicSkill
latestVersion: Doc<'skillVersions'> | null
ownerHandle?: string | null
searchScore?: number
}
type SkillSearchEntry = {
@@ -62,21 +71,25 @@ export const Route = createFileRoute('/skills/')({
export function SkillsIndex() {
const navigate = Route.useNavigate()
const search = Route.useSearch()
const sort = search.sort ?? 'newest'
const dir = parseDir(search.dir, sort)
const [query, setQuery] = useState(search.q ?? '')
const view = search.view ?? 'list'
const highlightedOnly = search.highlighted ?? false
const [query, setQuery] = useState(search.q ?? '')
const searchSkills = useAction(api.search.searchSkills)
const [searchResults, setSearchResults] = useState<Array<SkillSearchEntry>>([])
const [searchLimit, setSearchLimit] = useState(pageSize)
const [isSearching, setIsSearching] = useState(false)
const searchRequest = useRef(0)
const loadMoreRef = useRef<HTMLDivElement | null>(null)
const loadMoreInFlightRef = useRef(false)
const searchInputRef = useRef<HTMLInputElement>(null)
const trimmedQuery = useMemo(() => query.trim(), [query])
const hasQuery = trimmedQuery.length > 0
const sort =
search.sort === 'relevance' && !hasQuery
? 'newest'
: (search.sort ?? (hasQuery ? 'relevance' : 'newest'))
const dir = parseDir(search.dir, sort)
const searchKey = trimmedQuery ? `${trimmedQuery}::${highlightedOnly ? '1' : '0'}` : ''
// Use convex-helpers usePaginatedQuery for better cache behavior
@@ -149,6 +162,7 @@ export function SkillsIndex() {
skill: entry.skill,
latestVersion: entry.version,
ownerHandle: entry.ownerHandle ?? null,
searchScore: entry.score,
}))
}
// paginatedResults is an array of page items from usePaginatedQuery
@@ -165,6 +179,8 @@ export function SkillsIndex() {
const results = [...filtered]
results.sort((a, b) => {
switch (sort) {
case 'relevance':
return ((a.searchScore ?? 0) - (b.searchScore ?? 0)) * multiplier
case 'downloads':
return (a.skill.stats.downloads - b.skill.stats.downloads) * multiplier
case 'installs':
@@ -196,7 +212,8 @@ export function SkillsIndex() {
const canAutoLoad = typeof IntersectionObserver !== 'undefined'
const loadMore = useCallback(() => {
if (isLoadingMore || !canLoadMore) return
if (loadMoreInFlightRef.current || isLoadingMore || !canLoadMore) return
loadMoreInFlightRef.current = true
if (hasQuery) {
setSearchLimit((value) => value + pageSize)
} else {
@@ -204,6 +221,12 @@ export function SkillsIndex() {
}
}, [canLoadMore, hasQuery, isLoadingMore, loadMorePaginated])
useEffect(() => {
if (!isLoadingMore) {
loadMoreInFlightRef.current = false
}
}, [isLoadingMore])
useEffect(() => {
if (!canLoadMore || typeof IntersectionObserver === 'undefined') return
const target = loadMoreRef.current
@@ -211,6 +234,7 @@ export function SkillsIndex() {
const observer = new IntersectionObserver(
(entries) => {
if (entries.some((entry) => entry.isIntersecting)) {
observer.disconnect()
loadMore()
}
},
@@ -284,6 +308,7 @@ export function SkillsIndex() {
}}
aria-label="Sort skills"
>
{hasQuery ? <option value="relevance">Relevance</option> : null}
<option value="newest">Newest</option>
<option value="updated">Recently updated</option>
<option value="downloads">Downloads</option>
+257
View File
@@ -1259,6 +1259,8 @@ code {
.skill-detail-stack {
display: grid;
gap: 16px;
max-width: 100%;
overflow-x: auto;
}
.skill-hero {
@@ -1699,6 +1701,8 @@ code {
.tab-card {
gap: 14px;
max-width: 100%;
overflow-x: auto;
}
.tab-header {
@@ -1731,11 +1735,14 @@ code {
.tab-body {
display: grid;
gap: 20px;
max-width: 100%;
overflow-x: auto;
}
.file-list {
display: grid;
gap: 12px;
max-width: 100%;
padding-top: 8px;
border-top: 1px solid var(--line);
}
@@ -1751,6 +1758,7 @@ code {
display: grid;
gap: 8px;
max-height: 260px;
max-width: 100%;
overflow: auto;
padding-right: 4px;
}
@@ -1766,6 +1774,7 @@ code {
align-items: center;
justify-content: space-between;
gap: 12px;
max-width: 100%;
padding: 10px 12px;
border-radius: 12px;
border: 1px solid var(--line);
@@ -2350,6 +2359,7 @@ code {
.markdown {
line-height: 1.7;
max-width: 100%;
color: #3f3a34;
}
@@ -2376,6 +2386,7 @@ code {
.markdown pre {
white-space: pre;
max-width: 100%;
overflow-x: auto;
-webkit-overflow-scrolling: touch;
background: linear-gradient(180deg, rgba(255, 255, 255, 0.9), rgba(255, 250, 247, 0.88));
@@ -2761,6 +2772,11 @@ html.theme-transition::view-transition-new(theme) {
display: flex;
align-items: center;
gap: 12px;
margin-top: 8px;
}
.scan-result-row:first-child {
margin-top: 0;
}
.scan-result-scanner {
@@ -2778,6 +2794,10 @@ html.theme-transition::view-transition-new(theme) {
color: #0030ff;
}
.scan-result-icon-oc {
color: var(--accent);
}
.scan-result-status {
padding: 2px 8px;
border-radius: 999px;
@@ -2923,6 +2943,10 @@ html.theme-transition::view-transition-new(theme) {
color: #0030ff;
}
.version-scan-icon-oc {
color: var(--accent);
}
.version-scan-link {
color: var(--ink-soft);
text-decoration: none;
@@ -2940,6 +2964,228 @@ html.theme-transition::view-transition-new(theme) {
font-style: italic;
}
/* OpenClaw confidence label */
.scan-result-confidence {
font-size: 0.72rem;
font-weight: 600;
color: var(--ink-soft);
text-transform: uppercase;
letter-spacing: 0.05em;
opacity: 0.7;
}
/* LLM Analysis Detail */
.analysis-detail {
margin-top: 10px;
border-radius: 12px;
overflow: hidden;
}
.analysis-detail-header {
display: flex;
align-items: center;
justify-content: space-between;
gap: 12px;
padding: 10px 14px;
cursor: pointer;
user-select: none;
border-radius: 12px;
transition: background 0.15s ease;
}
.analysis-detail-header:hover {
background: rgba(0, 0, 0, 0.03);
}
[data-theme="dark"] .analysis-detail-header:hover {
background: rgba(255, 255, 255, 0.04);
}
.analysis-detail-toggle {
font-size: 0.82rem;
color: var(--accent);
font-weight: 600;
display: flex;
align-items: center;
gap: 4px;
white-space: nowrap;
}
.analysis-detail-toggle .chevron {
transition: transform 0.2s;
}
.analysis-detail.is-open .analysis-detail-toggle .chevron {
transform: rotate(180deg);
}
.analysis-summary-text {
font-size: 0.88rem;
color: var(--ink);
line-height: 1.45;
flex: 1;
}
.analysis-body {
display: none;
padding: 0 14px 14px;
}
.analysis-detail.is-open .analysis-body {
display: block;
}
.analysis-dimensions {
display: grid;
gap: 8px;
margin-top: 8px;
}
.dimension-row {
display: grid;
grid-template-columns: 20px 1fr;
gap: 10px;
padding: 10px 12px;
border-radius: 10px;
border: 1px solid var(--line);
background: var(--surface-muted);
align-items: start;
}
.dimension-icon {
width: 20px;
height: 20px;
border-radius: 50%;
display: grid;
place-items: center;
font-size: 0.65rem;
font-weight: 700;
flex-shrink: 0;
margin-top: 1px;
}
.dimension-icon-ok {
background: rgba(34, 197, 94, 0.15);
color: #16a34a;
}
.dimension-icon-note {
background: rgba(107, 114, 128, 0.12);
color: #6b7280;
}
.dimension-icon-concern {
background: rgba(245, 158, 11, 0.15);
color: #d97706;
}
.dimension-icon-danger {
background: rgba(239, 68, 68, 0.12);
color: #dc2626;
}
.dimension-content {
min-width: 0;
}
.dimension-label {
font-size: 0.78rem;
font-weight: 700;
text-transform: uppercase;
letter-spacing: 0.04em;
color: var(--ink);
margin-bottom: 3px;
}
.dimension-detail {
font-size: 0.84rem;
color: var(--ink-soft);
line-height: 1.5;
}
.dimension-detail code {
font-size: 0.78rem;
background: rgba(0, 0, 0, 0.06);
padding: 1px 5px;
border-radius: 4px;
}
[data-theme="dark"] .dimension-detail code {
background: rgba(255, 255, 255, 0.1);
}
/* Analysis guidance panel */
.analysis-guidance {
margin-top: 12px;
padding: 10px 14px;
border-radius: 10px;
font-size: 0.84rem;
line-height: 1.55;
color: var(--ink);
}
.analysis-guidance.benign {
background: rgba(34, 197, 94, 0.06);
border-left: 3px solid #16a34a;
}
.analysis-guidance.suspicious {
background: rgba(245, 158, 11, 0.06);
border-left: 3px solid #d97706;
}
.analysis-guidance.malicious {
background: rgba(239, 68, 68, 0.06);
border-left: 3px solid #dc2626;
}
.analysis-guidance-label {
font-size: 0.68rem;
font-weight: 700;
text-transform: uppercase;
letter-spacing: 0.08em;
margin-bottom: 4px;
}
.analysis-guidance.benign .analysis-guidance-label {
color: #16a34a;
}
.analysis-guidance.suspicious .analysis-guidance-label {
color: #d97706;
}
.analysis-guidance.malicious .analysis-guidance-label {
color: #dc2626;
}
/* Scan findings section */
.scan-findings-section {
margin-top: 10px;
padding: 10px 14px;
border-radius: 10px;
background: var(--surface-muted);
border: 1px solid var(--line);
}
.scan-findings-title {
font-size: 0.72rem;
font-weight: 700;
text-transform: uppercase;
letter-spacing: 0.06em;
color: var(--ink-soft);
margin-bottom: 6px;
}
.scan-finding-row {
display: flex;
align-items: baseline;
gap: 8px;
font-size: 0.82rem;
color: var(--ink-soft);
padding: 4px 0;
}
/* Pending Review Banner */
.pending-banner {
font-size: 0.9rem;
@@ -2971,6 +3217,17 @@ html.theme-transition::view-transition-new(theme) {
margin: 0;
}
.pending-banner-content .pending-banner-appeal {
margin-top: 6px;
font-size: 0.8rem;
opacity: 0.75;
}
.pending-banner-appeal a {
color: inherit;
text-decoration: underline;
}
/* Blocked/removed banner variant */
.pending-banner-blocked {
background: rgba(239, 68, 68, 0.12);