Explore three connected projects in the real LoopX Personal Workspace: a 120-person community event, a home-energy buying guide, and a neighborhood website release. Each has four work roles, 18 delivery tasks across multiple phases, two owner decisions and two scheduled watches.
From a source checkout with Python 3.11+ and Node 22.22.3+:
python -m demo.workspace serveOpen the printed loopback URL. The command creates a disposable directory and serves the existing packaged frontend through the standard Chat backend. For a repeatable location:
python -m demo.workspace prepare --root /tmp/workspace-stories
python -m demo.workspace serve --root /tmp/workspace-stories --port 8791Only a new empty directory or this demo's matching manifest is accepted. Prepare preserves an existing replay; choose a new directory to start again. Older manifest versions require a new directory.
| Project | Work and constraints | Decisions |
|---|---|---|
| Riverside Community Day | Three-venue comparison, $6000 limit, $5400 plan, $600 contingency, revised catering quote, access review, 18 volunteer shifts and two first-aid gaps | Venue reservation; invitation approval |
| Home Energy Buying Guide | 12 source cards, three household profiles, 27 tariff/efficiency combinations, conflicting assumptions, five-section editorial plan | Cost assumptions; publication approval |
| Riverside Neighborhood Website | Six pages, 18-route inventory, 24 accessibility criteria, navigation/form review findings, content permissions, rollback and handoff | Content freeze; deployment approval |
Each project has seven replayed completion checkpoints, five ready tasks, four blocked tasks and two deferred follow-ups. Dependency notes retain predecessor IDs; deferred tasks use real todo_done resume conditions. Two watch-only monitors have cadence and next-due metadata. No scheduler or live Agent is started by the demo.
Switch Board/List, filter by Agent, expand completed history, inspect the owner decisions and scheduled watches. BRIEF.md, working-table.csv, calculations.json and DELIVERY-PLAN.md preserve the planning inputs and dependencies. The energy sensitivity table and event contingency are calculated when preparing the workspace.
Replay one decision:
python -m demo.workspace advance --root /tmp/workspace-stories \
--story research-brief --decision assumptionsRefresh the UI afterward. Only that decision and its direct blocked successor are advanced through the canonical User completion owner. The App's Approve, Reject and Withdraw controls use the same decision relationship; reject and withdraw leave the dependent blocked. Other owner decisions and downstream blockers remain intact. The command does not purchase, publish, deploy or start an Agent. Use a fresh directory after updating from demo manifest v2; existing replays are preserved rather than migrated.
These are authored scenario replays using real LoopX APIs and state transitions, not customer case studies or receipts of live Agent execution. Natural project titles keep the interface readable; the manifest and completion evidence retain provenance. Source-card inventories and website checklists are planning inputs, not claims of external research or executed website tests. Event money is a project budget, not model spending.
Each newly prepared story Goal selects the existing soft_claim handoff mode before tasks are seeded. The stories have named claim owners but no live managed worker or execution identity, so the replay does not acquire task leases. This setting applies only to the isolated demo Goals and does not change LoopX defaults.
The demo does not import personal registries, session history or credentials, and does not sync into the global registry. Prepare, advance and serve run in a separate HOME/CODEX_HOME with a minimal environment; even preparation never discovers personal default registries. The loopback server uses unavailable Agent/Lark binaries. Chat and Lark connection errors are intentional isolation and do not qualify live IM behavior. Stop with Ctrl-C.
This remains a source-checkout demo under demo/, outside the installed wheel and capability catalog. Screenshots and recordings belong in ignored output/playwright/. Keep real operating statistics separately timestamped with their counting scope.
The repository README hero is rendered from the packaged App on this demo. With the dashboard dependencies installed, build the App and start a fresh isolated demo:
uv run --extra test python scripts/chat_bundle.py build --install
uv run --extra test python -m demo.workspace serve --root /tmp/workspace-hero --port 8791In another terminal:
node demo/workspace/readme-hero/render.mjs --url http://127.0.0.1:8791 \
--demo-root /tmp/workspace-hero --out docs/assets/personal-workspaceThe hero shows the real packaged Work map: recorded task relations, responsible agents, progress and pending owner decisions. These come from the authored scenario replay; the capture runs no live Agent task and adds no mocked conversation or team record. The renderer requires a complete map and a fresh v3 demo, prepares a pending venue approval through the existing typed action preview API, and verifies its confirmation button in both languages without applying it. All projected Todo states must remain unchanged after capture. It fails on older Apps, incomplete maps or advanced decisions. Intermediate frames stay in ignored output/playwright/readme-hero/frames/.
Validation: python -m pytest tests/test_workspace_story_demo.py -q exercises real state, directory isolation, repeatable prepare, computed artifacts and decision-scoped transitions. The normal Workspace browser smoke covers the shared Board/List and completed-history behavior, including #3961.