v5.2.0

agentsys v5.2.0

The biggest release since the v5.0 rewrite. agentsys is now a true marketplace - 14 plugins live in their own repos, fetched on demand at install time. Cursor joins as the 4th supported platform. And the brand new /web-ctl plugin gives your agents a browser.

Highlights

Cursor Support

agentsys now installs to Cursor alongside Claude Code, OpenCode, and Codex CLI. Skills, commands, and rules are installed project-scoped under .cursor/. Cursor v2.4+ natively supports the Agent Skills standard - no transform needed.

agentsys --tool cursor
agentsys install next-task --tool cursor

/web-ctl - Browser Automation for Agents

New plugin for headless browser control via Playwright. Persistent encrypted sessions, human-in-the-loop auth handoff (with CAPTCHA detection), anti-bot measures, WSL detection with Windows Chrome fallback, and prompt injection defense. Includes the web-session agent and web-auth/web-browse skills.

Plugin Extraction to Standalone Repos

All 14 plugins now live in their own repos under the agent-sh org. The plugins/ directory is gone - agentsys is now a marketplace + installer that fetches plugins from GitHub at install time.

GitHub Projects as Task Source

/next-task can now pull tasks from GitHub Projects v2 boards via gh-projects source. Includes PR-linked issue exclusion, input validation, and preference caching.

Installer Fixes

This release fixes several installer issues introduced by the marketplace schema change in v5.1.0:

• Fixed crash when reading new object-format source field in marketplace.json
• Fixed .git suffix in source URLs causing 404s on GitHub tarball API
• Added fallback to main branch when version tags don't exist yet
• Fixed Windows tar extraction by converting backslash paths for MSYS2

Bug Fixes

• Windows jscpd output path mangling (#270)
• /debate 240s timeout enforcement across all platforms
• /next-task review loop stall detection (MAX_STALLS: 2 -> 1)
• /next-task Phase 12 ship invocation (was calling Task instead of Skill)
• /consult and /debate model defaults updated to latest (Gemini 3.1, Codex gpt-5.3)
• task-discoverer now excludes issues with open PRs

Install / Update

npm install -g agentsys    # Fresh install
npm update -g agentsys     # Update existing

---

Full changelog: CHANGELOG.md
14 plugins | 43 agents | 30 skills | 4 platforms

v5.1.0

Added

• /debate plugin — New plugin for structured multi-round AI dialectic. Pick two tools (e.g. codex vs gemini), set 1–5 rounds, and get a proposer/challenger debate with a synthesized verdict. Supports natural language input, effort levels (--effort=low|high|max), and context injection (--context=diff or --context=file=PATH). Available on Claude Code, OpenCode, and Codex CLI.
• /consult multi-instance support — Run N parallel consultations with the same tool using --count=N (or natural language: "ask 3 gemini about this"). Responses are numbered and a brief synthesis highlights agreements and differences.
• /consult natural language parsing — Free-form queries are now parsed automatically without requiring explicit flags. "with codex about my auth approach", "ask gemini thoroughly about this design", or "3 claude opinions on error handling" all work out of the box.

Changed

• Agent model optimization — exploration-agent and learn-agent switched from opus to sonnet, reducing cost and latency for exploration and research passes with no quality regression.

Fixed

• Debate --context=file path validation — Added path containment checks to prevent directory traversal when passing file paths as context.
• Debate prompt hardening — Context passthrough, canonical output redaction, and relaxed disagreement rules applied consistently across all debate rounds.
• Consult model/flag issues — Hardened model flag handling and non-interactive invocation across all four supported tools (Claude, Gemini, Codex, OpenCode).

v5.0.3

Fixed

• Consult: Codex command corrected to codex exec - Codex CLI uses codex exec for non-interactive mode (not -q flag). Non-interactive resume uses codex exec resume SESSION_ID "prompt" --json. All four tools (Claude, Gemini, Codex, OpenCode) now have correct native session resume support.

v5.0.2

Fixed

• Consult: Codex and OpenCode marked as continuable - Both tools support session resume but were incorrectly marked as non-continuable. Codex supports codex resume (TUI-only; non-interactive uses context prepending). OpenCode supports --session SESSION_ID and --continue flags in non-interactive mode.

v5.0.1

Fixed

• OpenCode legacy cleanup - Installer now removes legacy agent files (review.md, ship.md, workflow.md) left over from pre-rename installs
• OpenCode install validator - Now checks only the agents/commands/skills produced by discovery, preventing false positives from legacy files
• Windows compatibility - bump-version.js uses npm.cmd on win32 (fixes execFileSync PATHEXT resolution)
• Windows test fixes - Scaffold test tolerates EBUSY on temp directory cleanup; script-failure-hooks test skips bash-dependent tests on Windows
• Jest module resolution - Added moduleNameMapper for @agentsys/lib to resolve to local lib/ directory

Changed

• Workflow ship references - Updated all /ship references to ship:ship (plugin-namespaced command) across next-task command, agents, hooks, and Codex/OpenCode adapters

v5.0.0

See CHANGELOG.md for details.

v4.2.2

Fixed

• Added missing frontmatter descriptions to 3 command reference files (audit-project-agents, ship-ci-review-loop, ship-deployment) that caused Codex adapter skills to install with empty descriptions
• Added build-time validation in gen-adapters.js to error on empty Codex skill descriptions
• Added install-time guard in bin/cli.js to skip skills with missing descriptions

v4.2.1

Fixed

• Removed unused @awesome-slash/lib publish job from release workflow
• Cleaned up all references to lib as a standalone npm package (docs, scripts, tests, configs)

v4.1.1

Fixed

• Skills $ARGUMENTS parsing - Added $ARGUMENTS parsing to 13 skills that declared argument-hint but never consumed the arguments (CC-SK-012)
• agnix config - Migrated .agnix.toml disabled_rules from deprecated slug format to proper rule IDs (XP-003, AS-014)
• Memory file language - Strengthened imperative language in AGENTS.md/CLAUDE.md (PE-003, CC-MEM-006)

v4.1.0

Added

• New /agnix Plugin - Lint agent configurations before they break your workflow
  • Validates Skills, Hooks, MCP, Memory, Plugins across Claude Code, Cursor, GitHub Copilot, and Codex CLI
  • 100 validation rules from official specs, research papers, real-world testing
  • Auto-fix support with --fix flag
  • SARIF output for GitHub Code Scanning integration
  • Target-specific validation (--target claude-code|cursor|codex)
  • Requires agnix CLI (cargo install agnix-cli)

Changed

• Plugin Count - Now 11 plugins, 40 agents, 26 skills
• CLAUDE.md Rule #11 - Added rule about using [] not <> for argument hints

Fixed

• Prompt Injection - Sanitize user arguments in agnix command (validate target, strip newlines from path)
• Argument Parsing - Support both --target=value and --target value forms
• enhance-hooks/SKILL.md - Fixed path example escaping