v15.14.0 "Agent QA and Durable Knowledge"

[15.14.0] - 2026-08-16 - "Agent QA and Durable Knowledge"

Added an evidence-first Agent QA workflow and a durable knowledge-compilation skill, while refreshing multilingual video guidance and the catalog's dependency baseline. The published catalog contains 2,013 skills.

This installable library update gives Claude Code, Cursor, Codex CLI, Gemini CLI,
Antigravity, and related AI coding assistants clearer paths for authoring tests,
triaging failures, making scoped repairs, and preserving hard-won findings
across sessions.

Start here:

• Install: npx agentic-awesome-skills
• Choose your tool
• Best skills by tool
• Bundles
• Workflows

Added

• Added agent-qa-authoring for creating,
  validating, and running schema-safe Agent QA tests, suites, and hooks while
  preserving canonical IDs and explicit approval boundaries.
• Added agent-qa-result-triage for classifying
  failed Agent QA runs from recorded steps, artifacts, and logs with a fixed
  category, confidence, likely ownership, and an evidence-backed next action.
• Added agent-qa-debug-fix for making the smallest
  justified repair after triage and verifying it narrowly without rewriting a
  test merely to hide a product or infrastructure defect.
• Added compile-knowledge for preserving durable,
  non-obvious findings as atomic, interlinked Markdown notes with a maintained
  index instead of accumulating routine or duplicate memory.

Changed

• Expanded riffkit guidance from English and Spanish to all
  nine currently supported native output languages: English, Spanish,
  Portuguese, Indonesian, German, French, Italian, Japanese, and Simplified
  Chinese (#1155).
• Refreshed the cohesivity description to present its
  headless hosting, database, storage, model, and third-party API surface more
  clearly while preserving its existing consent and safety boundaries
  (#1157).
• Regenerated the canonical catalog, offline AAS Core data, web assets,
  marketplaces, editorial bundles, and Codex/Claude plugin distributions for
  2,013 skills.

Security

• Updated the web app's transitive nanoid dependency from 3.3.17 to
  3.3.18, resolving the zero-length custom-generator denial-of-service
  advisory without changing the direct dependency set
  (#1150).
• Kept Agent QA authoring and repair operations behind explicit workspace,
  scope, side-effect, and production-action approval boundaries; triage remains
  evidence-first and read-only until a separate repair step is chosen.
• Kept durable knowledge stores free of secrets and personal identifiers, and
  confined writes and deletions to the user-approved store.

Who should care

• QA and product teams using Agent QA with Claude Code, Cursor, Codex CLI, or
  Gemini CLI who need a reviewable path from test definition to failure triage
  and minimal repair.
• Researchers and agents that repeatedly rediscover the same project facts and
  need a deliberately selective, interlinked memory store.
• Riffkit users producing native-language short-form video beyond English and
  Spanish.
• Maintainers and web-catalog users who need the patched nanoid dependency and
  a version-aligned 2,013-skill distribution.

Validation

• Passed repository validation, reference validation, documentation-security
  checks, warning-budget enforcement, the complete 110-group repository test
  suite, plugin-compatibility and bundle checks, web-app install/build/prerender,
  and the npm package dry run on the protected release base.
• Reviewed the Agent QA schemas, bundled references, fixed triage categories,
  mutation approvals, provenance, risk labels, and license declarations; also
  reviewed compile-knowledge store boundaries, provenance, and limitations.

Limitations

• The Agent QA skills document Agent QA's MCP and CLI contracts; they do not
  bundle the Agent QA service, credentials, a configured workspace, or a test
  environment. Mutating or production-facing runs still require explicit user
  approval.
• compile-knowledge requires a user-approved Markdown store and disciplined
  curation; it intentionally skips routine, derivable, secret, personal, and
  conversation-only facts.
• Riffkit remains a hosted, metered service that requires an account session and
  explicit approval before billable rendering.

Credits

• @pranshuchittora and the official
  vostride/agent-qa source for the
  three Agent QA skills in PR #1158.
• @lodar and
  5dive-ai/skills for
  compile-knowledge in
  PR #1159.
• @owengu-ai for the nine-language
  riffkit refresh in
  PR #1155.
• @shouryamaanjain for the
  cohesivity description refresh in
  PR #1157.