Agent Workspace

agent-workspace-template is now framed as Agent Workspace: a components-style directory and starter runtime for natural-language agent harness packs, publicly hosted at agentworkspace.attrition.sh.

The repo still contains the reusable platform extracted from FloorAI, but the product story is now simpler:

• browse verified harness packs on the homepage
• open detail pages with instructions, sources, and evaluation guidance
• submit new harness packs into a lightweight review queue
• use the built-in studio routes to preview the shared chat and right-side agent rail
• adapt the underlying runtime into your own domain implementation

It keeps the shared runtime pieces that made the original project useful:

• centered shared chat plus expandable agent rail
• durable Convex message streaming via messageEvents
• file upload and evidence handling
• trace, telemetry, quality-check, and answer-packet persistence
• role-aware workspace access

It intentionally removes retail-specific domain logic. The expectation is that you will use this repo as the starting point for a new domain implementation such as:

• PropertyAI
• CourseAI
• another operations or knowledge-work agent

What this repo includes

src/app/
  /                 harness pack directory
  /packs/[slug]     pack detail pages
  /submit           submission form + review queue preview
  /chat             shared channel studio
  /workspace-a      builder preview workspace
  /workspace-b      reviewer preview workspace

src/components/
  ChatPanel         right-rail assistant UI
  GroupChat         centered shared thread
  AttachmentList    evidence rendering
  Sidebar           nav and workspace switcher
  ConvexClientProvider
  OperatorSessionProvider

convex/
  schema.ts         generic runtime tables
  users.ts          operator lookup
  workspaces.ts     workspace registry
  submissions.ts    harness pack submission queue
  access.ts         generic access helpers
  messages.ts       message + event persistence
  files.ts          upload metadata
  audit.ts          event log
  answerPackets.ts  final answer persistence
  evalRuns.ts       eval storage
  agent.ts          generic agent action shell
  seed.ts           sample workspaces and operators

What this repo does not include

• domain entities such as stores, leases, or courses
• domain tool registry for your real workflows
• domain goldens or evaluation prompts
• production retrieval corpora

Those are the first things you should add when turning the template into a real product repo.

Quick start

npm install
npx convex dev
npm run seed

Then set NEXT_PUBLIC_CONVEX_URL in .env.local using the local Convex URL printed by convex dev.

Public URL

The current public-facing catalog lives at:

• https://agentworkspace.attrition.sh

The GitHub repository stays agent-workspace-template, but the product surface is branded as Agent Workspace. The hosted URL is intentionally an Attrition subdomain rather than a separate root domain.

Adaptation sequence

Do not start by editing the large agent file. Lock these in order:

1. domain schema
2. seed data
3. brief packet shape
4. tool registry
5. answer packet shape

Only then should you rewrite the agent runtime and workspace pages.

Default runtime behavior

The template agent is intentionally lightweight:

• it writes user and assistant messages
• it emits ordered message events
• it persists a simple trace and answer packet
• it uses Gemini if GOOGLE_API_KEY is set
• otherwise it falls back to a deterministic placeholder response

That means the repo boots cleanly as a workspace scaffold before you add your own domain retrieval or evaluation logic.

Recommended next steps after cloning this repo

1. Rename workspace-a / workspace-b to real personas
2. Replace generic workspaces with domain entities
3. Add domain tool functions
4. Replace the generic prompt and renderer in convex/agent.ts
5. Add domain goldens and live eval
6. Rewrite README, slides, and demo video for the new domain

Related references

If you are coming from the FloorAI repo, the original extraction notes live there in:

• docs/DOMAIN_SPINOUT_PLAYBOOK.md

Use from Claude Code

The catalog ships two npm packages you can wire into Claude Code, Cursor, or Codex:

• attrition-mcp — stdio MCP server. Browse, resolve, and fetch packs from within your agent.
• attrition — CLI. Install a pack into your repo as a .claude/skills/ skill plus an AGENTS.md fragment.

MCP server

Add to ~/.claude.json (global) or a repo-local .claude.json:

{
  "mcpServers": {
    "attrition": {
      "command": "npx",
      "args": ["-y", "attrition-mcp"],
      "env": {
        "ATTRITION_REGISTRY_URL": "https://agentworkspace.attrition.sh"
      }
    }
  }
}

For Cursor, use the same block in ~/.cursor/mcp.json. For Codex and other MCP-capable clients, consult your client's MCP config docs — the command / args / env shape is the same.

CLI quick start

# Install a pack directly into .claude/skills/ and AGENTS.md
npx attrition-sh pack install rag-hybrid-bm25-vector

# Browse the catalog
npx attrition-sh pack list
npx attrition-sh pack search evaluator

rag-hybrid-bm25-vector is a real slug — swap it for any pack in the catalog.

For maintainers

See docs/PUBLISHING.md for the publish workflow, version bump policy, tarball smoke tests, and rollback via npm deprecate.