Skip to content

v15.10.0 "Portable Agent Plugins and Governed Workflows"

Choose a tag to compare

View all tags
@sickn33 sickn33 released this 07 Aug 06:41
· 3 commits to main since this release
v15.10.0
3d56e84
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

[15.10.0] - 2026-08-07 - "Portable Agent Plugins and Governed Workflows"

Added portable Agent Plugins 1.0 exports, a video-production router, consent-gated outreach, and stronger trust boundaries across the catalog.

This release gives Claude Code, Cursor, Codex CLI, Gemini CLI, Antigravity, and related AI coding assistants a safer, more portable 2,005-skill catalog. It packages 58 eligible editorial bundles in the shared Agent Plugins format alongside the existing Codex and Claude distributions, adds reply-aware outreach with explicit approval before real sends, introduces an upstream-pinned video routing guide, and hardens web scraping against hostile page content.

Start here:

Added

  • Added outreachagent for inbox, contact, template, workflow, webhook, and delivery-metric operations through OutreachAgent's public REST API, with separate draft and final-send approvals, exact rendered-payload review, reply-aware exits, and untrusted-inbound handling.
  • Added video-router from the official OrkasVideoStudio source for locking generation, deterministic composition, supplied-footage editing, or an automatic cross-modal production plan before execution, with explicit runtime and produced-vs-planned boundaries.

Changed

  • Added Agent Plugins 1.0 as a portable packaging target for eligible editorial bundles, with schema-pinned root manifests, collision-safe flattened skill IDs, Agent Skills-compatible packaged frontmatter, and preserved AAS provenance and risk metadata. Broad host-specific roots remain on their existing Codex and Claude surfaces rather than making a false portability claim.
  • Regenerated the canonical catalog, offline AAS Core data, marketplaces, bundles, Agent Plugins exports, and Codex/Claude plugin distributions for 2,005 skills.

Security

  • Hardened web-scraper against prompt injection from page content, constrained downloads to an exact user-confirmed output path, required consent before browser escalation, and clarified stop conditions for authenticated or sensitive pages.
  • Hardened coordinated-disclosure trust boundaries across NotebookLM, YouTube ingestion, Instagram OAuth and CSV export, Telegram HTML replies, Loki task export, Vercel claim verification, macOS packaging templates, and generated TSX content.
  • Added focused regression coverage for path containment, private temporary storage, inert rendering, OAuth callback binding, data-only version parsing, and untrusted browser or model output.

Who should care

  • Agent clients and plugin authors adopting the shared Agent Plugins format across compatible hosts.
  • Video teams that need an explicit, reviewable decision between generation, composition, editing, and a cross-modal production plan before execution.
  • Teams building AI-assisted outbound workflows that need explicit recipient, sender, payload, schedule, and real-send approval boundaries.
  • Agents extracting public web content where hostile page instructions must remain untrusted data.
  • Security-conscious users relying on NotebookLM, YouTube, Instagram, Telegram, Loki, Vercel, or macOS packaging skills.
  • Claude Code, Cursor, Codex CLI, Gemini CLI, and Antigravity users installing from canonical or plugin distributions.

Validation

  • Passed canonical skill and reference validation, documentation security, warning-budget enforcement, changed-skill evidence, the complete repository suite, web-app coverage, package dry-run checks, and dependency audits.
  • Validated every generated portable manifest against the pinned official Agent Plugins 1.0 schema, every eligible package with the Codex plugin validator, and each packaged skill instance against the Agent Skills reference validator.
  • Confirmed protected main CI and CodeQL are green and canonical regeneration is idempotent before release preparation.

Credits

Assets 2
Loading