plugins

Portable skill definitions, and Claude Code harness plugins, for CogOS.

Most packages here are pure skill collections: each skill is a SKILL.md file — a structured prompt that gives an agent specialized knowledge for a specific domain. The format follows the Agent Skills open standard and works across Claude Code, Cursor, VS Code Copilot, Gemini CLI, Hermes, and other compatible agents.

One package, cogos-harness, is a different kind: it bundles hooks and an MCP server alongside its skills — the Claude Code plugin system's harness-integration surface, not just prompt content. It's Claude-Code-specific (hooks and MCP servers aren't portable to Hermes's skill-tap mechanism the way SKILL.md files are) and is documented separately in its own README.

Setup

Claude Code (plugin marketplace)

Install the whole collection as a marketplace tap:

claude plugin add myrgic/plugins

Or install a specific plugin package:

claude plugin add myrgic/plugins/cogos-workflow
claude plugin add myrgic/plugins/cogos-substrate
claude plugin add myrgic/plugins/cogos-research
claude plugin add myrgic/plugins/cogos-harness

cogos-harness additionally registers an MCP server (cogos-kernel) and a set of hooks — see its own README for what it does and what it needs running to be useful.

Skills are then available in any Claude Code session. Claude invokes them automatically when the conversation context matches, or you can trigger them explicitly (e.g. /plan-phases, /orchestrate).

To install a single skill manually without the plugin system:

mkdir -p .claude/skills/plan-phases
cp plugins/cogos-workflow/skills/plan-phases/SKILL.md .claude/skills/plan-phases/SKILL.md

Hermes (skills tap)

Add the repo as a Hermes skill tap:

hermes skills tap add myrgic/plugins

Then install individual skills:

hermes skills install plan-phases
hermes skills install orchestrate
hermes skills install pull-context-dispatch

Or search across all taps:

hermes skills search orchestration
hermes skills search substrate

Skills installed via tap are available in all Hermes sessions and appear in the skills_list() tool output. The same SKILL.md files work in both agents — the format is identical.

Available Skills

Plugin	Skills	Description
cogos-workflow	plan-phases, execute-plan, critical-review, dispatch-agent, council, cold-start, retrospective, local-review, orchestrate, kanban-closed-loop-supervisor, pull-context-dispatch, suggest, session-protocol	Phased planning, parallel execution, deliberation, closed-loop orchestration, pointer-based dispatch, session continuity
cogos-substrate	substrate-archaeology	Finding and trusting prior work in a CogOS/Myrgic substrate
cogos-research	literature-research, experiment-design, interdisciplinary-research, physics-validation, lab-engineering, inference-cascade	Structured research methodology
cogos-voice	voice	Voice modality via Mod³ on Apple Silicon
cogos-dev-tools	git-forensics, technical-writing, code-quality, systems-architecture, zsh-environment, macos-storage, codex, openclaw-expert	Developer tooling, code quality, shell/storage management, cross-model CLI operation
cogos-architecture	corpus-cross-reference	CogOS architecture corpus hygiene — cross-check proposed RFCs/ADRs against the substrate corpus
myrgic-org	pr-triage, issue-triage	Org-management skills for the myrgic GitHub org
cogos-harness	btw, consolidate, handoff	Session-lifecycle hooks, kernel-vitals proprioception, and the cogos-kernel MCP server, plus the three skills that lean on them. Claude-Code-only — see its own README.

Total: 35 skills across 8 plugin packages (one of which, cogos-harness, also ships hooks and an MCP server).

Cross-Compatibility

Skills in this repo are written to work in both Claude Code and Hermes without modification:

• Claude Code loads skills from .claude/skills/*/SKILL.md automatically, or via the plugin marketplace (claude plugin add).
• Hermes loads skills from ~/.hermes/skills/ and from tap-registered GitHub repos (hermes skills tap add).
• Other agents (Cursor, VS Code Copilot, Gemini CLI) load skills from .claude/skills/ following the same Agent Skills standard.

The SKILL.md frontmatter is the same format for all agents. Agent-specific metadata (Hermes toolset requirements, Claude Code allowed-tools) lives under metadata.hermes or as standard frontmatter fields and is ignored by agents that don't understand it.

Some skills include a canonical_source frontmatter field pointing to the authoritative copy in a running Hermes or cog workspace instance. In those cases the marketplace copy is a projection — update the canonical source first, then project here.

Skill Bundles (Hermes)

Hermes supports composing multiple skills into a named bundle that loads them together with a shared instruction:

# ~/.hermes/skill-bundles/cogos-operator.yaml
name: cogos-operator
description: Full CogOS operator context: substrate archaeology, architecture hygiene, pointer-based dispatch
skills:
  - substrate-archaeology
  - corpus-cross-reference
  - pull-context-dispatch
instruction: |
  You are operating in the Myrgic/CogOS substrate. Read existing maps and prior work before
  designing or writing. Update canonical sources first; project to marketplace after.

Activate with /cogos-operator in any Hermes session.

Structure

plugins/
├── cogos-workflow/skills/       # Planning + execution + review + orchestration
├── cogos-substrate/skills/      # Prior-work archaeology
├── cogos-research/skills/       # Research methodology
├── cogos-voice/skills/          # Voice modality (Mod³)
├── cogos-dev-tools/skills/      # Developer tooling
├── cogos-architecture/skills/   # Architecture corpus hygiene
├── myrgic-org/skills/           # Org management
└── cogos-harness/               # Claude Code hooks + MCP server + skills (not skills-only)
    ├── hooks/
    ├── skills/
    └── .mcp.json

Format

Each skill follows the Agent Skills open standard:

---
name: skill-name
description: "One-line description — what the skill does and when to load it."
version: 1.0.0
platforms: [linux, macos, windows]
metadata:
  hermes:
    tags: [tag1, tag2]
    related_skills: [other-skill]
---

# Skill Name

Body: when to use, method, steps, examples.

Membrane canary

canary/run_canary.py is an end-to-end check on the cogos-harness plugin as it's actually installed, not just as it reads in the repo. It:

• adds this repo as a marketplace tap and installs cogos-harness@plugins into a fresh, throwaway HOME (its own keychain, its own plugin state, no contact with a real seat's config or tokens);
• runs every hook script from the installed cache path with fixture input, once with a reachable kernel and once with COGOS_KERNEL_URL pointed at a dead port, and checks each one exits 0 with the output shape its own docstring promises (silent for the lifecycle hooks, a structured <cogos_proprioception> block for the prompt-submit hook);
• confirms nothing escaped the sandbox: the real host's files are untouched, and no hook fabricated a fake cog workspace directory;
• runs claude plugin validate --strict against the marketplace manifest;
• greps the installed package for leaked personal paths or identifiers;
• and cleans up any session it registered on a live kernel afterward.

Run it locally with:

python3 canary/run_canary.py

It prints a digest and writes a full JSON report to canary/.last_run.json (gitignored), and exits 0 only if every check passed. The kernel-reachable half of the run is skipped, not failed, when nothing answers on 127.0.0.1:6931 (or $COGOS_KERNEL_URL) at start.

Contributing

Skills follow the same PR workflow as the rest of the myrgic org — branch on the upstream repo, squash-merge to main. See CONTRIBUTING.md for details.

If a skill has a canonical source in a running workspace (indicated by canonical_source in frontmatter), update that first, then project here. Don't edit the marketplace copy directly.

License

MIT