Marina's hard engineering rules covering code documentation style, git commit discipline, and test-driven development. Use whenever writing code, editing files, planning changes in a git repository, or preparing commits.
Marina's worktree policy. Use whenever creating, listing, or removing git worktrees, or when starting work that warrants an isolated checkout (Linear ticket, feature branch, experimental spike).
Marina's personal Claude Code workspace plugin.
The plugin assumes the following binaries are on the host's PATH. The SessionStart hook checks them and tells Claude to ask the user to install whatever is missing before continuing.
jq — parses JSON for the PreToolUse Bash policy hook. Without it, every Bash tool call errors out at the hook.tree (optional) — used by the SessionStart git-context hook for a tidy directory listing. Falls back to find -maxdepth 2 if missing.On a fresh session (source: startup) inside a git work tree, hooks/git-context.sh injects a short briefing so Claude doesn't have to rediscover the project: the last 5 commits, a depth-2 directory tree, the language manifest if one is present (package.json, pyproject.toml, go.mod, Cargo.toml, Makefile), and the first 60 lines of README.md. The hook stays silent for resume/compact sessions and outside git repos.
Alongside it, hooks/standards.sh injects the full engineering-standards skill content (self-documenting code, atomic commits with detailed bodies, TDD) so the rules are loaded into context from turn one instead of waiting for the skill to be discovered. Same guard as above — startup-only, git-repo-only.
This plugin depends on two other Claude Code plugins. When installed through a marketplace that allows cross-marketplace dependencies (see below), Claude Code resolves and installs them automatically.
exa@claude-plugins-official) — web search & researchUsed for any task that needs fresh information from the open web: deep research, lead generation, competitive analysis, literature reviews, and follow-up "find me everything about X" queries. Without it, this plugin falls back to whatever the model already knows, which is stale by definition.
colgrep@lightonai-colgrep) — semantic code searchReplaces plain grep for code search. Supports natural-language queries (colgrep "<query>"), literal/regex pre-filter (colgrep -e "<pattern>"), and hybrid mode that combines both. Without it, code search falls back to noisy grep/rg runs that miss intent-based matches.
Claude Code (v2.1.110+) supports a dependencies array in plugin.json. When a user installs this plugin, listed dependencies are resolved and installed automatically.
Because the required plugins live in different marketplaces than this one, the marketplace that hosts workspace-plugin must opt into cross-marketplace dependencies. Add this to that marketplace's .claude-plugin/marketplace.json:
{
"name": "amphetamarina",
"allowCrossMarketplaceDependenciesOn": [
"claude-plugins-official",
"lightonai-colgrep"
],
"plugins": [ /* … */ ]
}
Then declare the dependencies in .claude-plugin/plugin.json:
{
"name": "workspace-plugin",
"version": "0.1.0",
"dependencies": [
{ "name": "exa", "marketplace": "claude-plugins-official" },
{ "name": "colgrep", "marketplace": "lightonai-colgrep" }
]
}
Users still need each upstream marketplace added (claude plugin marketplace add …) before install — Claude Code refuses to pull from a marketplace the user has not configured. Once those are in place, installing workspace-plugin brings the dependencies with it.
If you are loading this plugin directly via --plugin-dir or your marketplace setup does not allow cross-marketplace deps, install each requirement yourself:
claude plugin marketplace add anthropics/claude-plugins-official # exa lives here by default
claude plugin install exa@claude-plugins-official
claude plugin install colgrep@lightonai-colgrep
Load the plugin directly from this directory without publishing or installing:
claude --plugin-dir /home/amphetamarina/workspace/workspace-plugin
Inside the session, verify the pieces are wired up:
/plugin — confirm workspace-plugin is listed and enabled/help — look for the skill /workspace-plugin:engineering-standardsgrep foo bar → blocked, pointed at colgrepgit commit -m one-liner → blockedgit commit with a Co-authored-by: line → blockedWebSearch → blocked with a pointer to the Exa MCP toolscolgrep foo without --json → blocked, pointed at --jsoncat package.json → blocked, pointed at jqgh pr list without --json → blocked, pointed at --json <fields>git-context.sh injects the briefing; resumed/compacted sessions stay silentIf jq is missing on the host, the SessionStart hook surfaces a notice describing what to install.
Iterate on hooks or the skill, then run /reload-plugins to pick up changes without restarting the session.
Executes bash commands
Hook triggers when Bash tool is used
Own this plugin?
Verify ownership to unlock analytics, metadata editing, and a verified badge. GitHub access is read-only (username + org membership).
Sign in to claimOwn this plugin?
Verify ownership to unlock analytics, metadata editing, and a verified badge. GitHub access is read-only (username + org membership).
Sign in to claimBased on adoption, maintenance, documentation, and repository signals. Not a security audit or endorsement.
npx claudepluginhub amphetamarina/workspace-plugin --plugin workspace-pluginPersistent, public, tamper-evident memory for a human-plus-agent team, stored as dust coins on eCash (XEC). Exposes a verb per capability through the bj CLI and as skills (load, capture, remember, forget, private, consensus, consolidate, pin, unpin), so it works across agent harnesses.
Ultra-compressed communication mode. Cuts ~75% of tokens while keeping full technical accuracy by speaking like a caveman.
Memory compression system for Claude Code - persist context across sessions
Multi-model consensus engine integrating OpenAI Codex CLI, Gemini CLI, and Claude CLI for collaborative code review and problem-solving.
Curate auto-memory, promote learnings to CLAUDE.md and rules, extract proven patterns into reusable skills.