AI Workspace Plugin
Claude is great at long-running work: designing systems, researching decisions, planning projects. The problem is that sessions end. Context gets compacted. Next time you start fresh, you're re-explaining things you already figured out. Frontier models have memory features for this but they are opaque and stored on their infrastructure.
Threads aim to address that. A thread is a folder on disk: a README that stays current, session logs, decisions, and anything Claude generates. When you pick it back up, Claude reads what it needs and continues where you left off. Nothing lives only in a conversation window. And because threads are just Markdown files, any model can read them. No vendor lock-in, no proprietary format.
Installation
The plugin ships for both Claude Code and OpenAI Codex CLI from a single source. Threads, MCP server, and templates are shared; each CLI installs through its native plugin system.
Claude Code
/plugin marketplace add sebmartin/ai-marketplace
/plugin install ai-workspace@sebmartin
Restart Claude Code after installing.
Codex CLI (Beta)
[!WARNING]
Codex support is in beta. Core skills (init, threads, debate) work, but skill instruction-following varies by model — some commands may require natural language instead of slash commands. Please report issues.
codex plugin marketplace add sebmartin/ai-workspace-plugin
codex
/plugins
Select ai-workspace and install.
Initialize a workspace
cd ~/my-workspace
# Claude Code
/ai-workspace:init
# Codex CLI
> initialize the ai-workspace
init creates threads/, AGENTS.md (read by both), CLAUDE.md (a one-line @AGENTS.md import for Claude), and .claude/settings.json (Claude-only permission allowlist; Codex ignores). All files are vendor-safe.
Examples
Threads work for anything you'd want to revisit across sessions, not just code.
How threads work
my-workspace/
├── AGENTS.md # Workspace instructions (read by Claude and Codex)
├── CLAUDE.md # One-line "@AGENTS.md" import (Claude only)
├── threads/
│ └── {thread-name}/
│ ├── README.md # Current focus, next steps, links to everything else
│ ├── sessions/ # One file per conversation
│ ├── decisions/ # Decisions with context and rationale
│ ├── attachments/ # Files you bring in (specs, docs, data)
│ └── artifacts/ # Files the CLI generates (snapshots, reports, emails)
└── .claude/
└── settings.json # Claude-only permission allowlist
You can run Claude from your workspace or from any repo. The plugin finds your threads either way.
cd ~/my-project
claude
> resume the api-redesign thread
# → (Using threads from /Users/you/my-workspace)
Context loads on demand
Thread files are organized as a hierarchy of linked Markdown documents. When resuming a thread, Claude only reads the main thread summary and then follows links as needed, loading more context on demand rather than all at once. A thread can grow large over time and still start light. This also reduces hallucinations: instead of working from a vague summary, Claude can follow a link to the actual source when precision matters.
Start fresh, pick up where you left off
A good habit is to save the thread at a natural stopping point, then start a fresh Claude session. Resume the thread and ask "where were we?" Claude reads the thread summary and the last session log, giving you a clean starting point without the cruft that accumulates in long conversations: failed attempts, tangents, superseded ideas. The important things are saved. Everything else is gone. This keeps token usage down and the context window clean.
Debate
When you have a proposal worth stress-testing, run a debate. A proponent makes the strongest honest case for the idea and refines it under pressure. A skeptic challenges specific assumptions, surfaces blind spots, and backs off when concerns are addressed. The result is saved as a thread artifact. See it used in the architectural change example.
/ai-workspace:debate # 2 rounds (default)
/ai-workspace:debate 3 # more rounds
Both agents can call in specialist agents to validate claims, and will ask you directly when they're uncertain.
Specialist Agents
Install the tech-expert-agents plugin for a ready-made set:
