skillhub/docs/security-scanning.md

Skill Scanner Backend Runtime Guide

Overview

SkillHub now supports a backend-only security scanning chain around skill-scanner. The publish flow changes are:

1. publish request enters SkillPublishService
2. if scanner is enabled, the version moves to SCANNING
3. the backend enqueues a ScanTask
4. ScanTaskConsumer calls skill-scanner
5. the scan result is stored in security_audit
6. the version moves to PENDING_REVIEW, or to SCAN_FAILED after final retry exhaustion
7. review still happens through the existing review workflow

Frontend is intentionally out of scope here. The frontend should fetch audit details through the dedicated backend API instead of expecting scanner data inside existing review detail payloads.

Runtime Modes

Two runtime modes are supported:

• local Use POST /scan and pass a filesystem path. This only works when SkillHub and skill-scanner can see the same files.
• upload Use POST /scan-upload and upload the package archive. This is the safer default for split deployments.

Recommended usage:

• local development with shared filesystem: local
• Kubernetes or any split-service deployment: upload

Backend Configuration

Application properties:

skillhub:
  security:
    scanner:
      enabled: false
      base-url: http://localhost:8000
      health-path: /health
      scan-path: /scan-upload
      mode: local
      connect-timeout-ms: 5000
      read-timeout-ms: 900000
      retry-max-attempts: 3
    stream:
      key: skillhub:scan:requests
      group: skillhub-scanners
      reclaim-min-idle: PT16M

Important environment variables:

• SKILLHUB_SECURITY_SCANNER_ENABLED
• SKILLHUB_SECURITY_SCANNER_URL
• SKILLHUB_SECURITY_SCANNER_MODE
• SKILLHUB_SECURITY_SCANNER_READ_TIMEOUT
• SKILLHUB_SCAN_STREAM_KEY
• SKILLHUB_SCAN_STREAM_GROUP
• SKILLHUB_SCAN_STREAM_RECLAIM_MIN_IDLE

Scanner-side optional environment variables:

• SKILL_SCANNER_LLM_API_KEY
• SKILL_SCANNER_LLM_BASE_URL
• SKILL_SCANNER_LLM_MODEL
• SKILLHUB_SCANNER_MAX_CONCURRENT_SCANS (default 1)
• SKILLHUB_SCANNER_HARD_TIMEOUT_SECONDS (default 930)

If the LLM variables are absent, the scanner should still run with non-LLM analyzers. The default timeout ordering is server read timeout (900 seconds), scanner hard timeout (930 seconds), then pending-message reclaim (960 seconds). A hard timeout exits the scanner process with status 124; Compose or Kubernetes restarts it and the Redis pending task is retried.

Kubernetes Notes

Current repository manifests assume separate skillhub-server and skillhub-scanner deployments. Because these deployments do not share a writable package directory, Kubernetes should use:

SKILLHUB_SECURITY_SCANNER_MODE=upload
SKILLHUB_SECURITY_SCANNER_URL=http://skillhub-scanner:8000

Relevant manifests:

• deploy/k8s/scanner-deployment.yaml
• deploy/k8s/services.yaml
• deploy/k8s/backend-deployment.yaml
• deploy/k8s/configmap.yaml

The scanner service is internal-only by default and is consumed by the backend through cluster DNS.

Verification

Verify the scanner service itself:

sh scripts/verify-scanner.sh http://localhost:8000
sh scripts/verify-scanner.sh http://localhost:8000 /path/to/skill.zip

Recommended backend checks after enabling the feature:

1. publish a test package
2. confirm the version status becomes SCANNING
3. confirm a security_audit row is created
4. confirm the version eventually moves to PENDING_REVIEW or SCAN_FAILED
5. call GET /api/v1/skills/{skillId}/versions/{versionId}/security-audit

Audit Query API

Backend audit data is available from:

GET /api/v1/skills/{skillId}/versions/{versionId}/security-audit

Response fields include:

• scanId
• scannerType
• verdict
• isSafe
• maxSeverity
• findingsCount
• findings
• scanDurationSeconds
• scannedAt
• createdAt

Failure Semantics

• scan task retries are handled by AbstractStreamConsumer
• final failure marks the version as SCAN_FAILED
• even after scan failure, a review task is still created so the package does not get stuck forever

This keeps the existing human review path intact while making scanner failures visible.