Skip to content

DEMO.md

Latest commit

 

History

History
101 lines (80 loc) · 5.25 KB

DEMO.md

File metadata and controls

101 lines (80 loc) · 5.25 KB
Raw

TagTeam — 3-minute demo script

Cast (fictional, per the brief): Ava — software engineer debugging why inference kernels underperform on the company's new AXB-200 accelerator board (host). Sam — senior hardware architect (tagged-in expert). Kai — junior engineer who can optionally shadow.

The AI backend in this demo is opencode (model-agnostic). The model name shown in the UI comes from /api/config and reflects whatever provider you configured opencode with (BigModel glm-5.2 in CI, or whatever you ran opencode auth against). In mock mode the AI label is "Assistant (mock)" and responses are canned — every multiplayer mechanic below is still fully real.

Setup (before you start the clock)

  1. npm start (or MOCK_CLAUDE=1 npm start for the zero-credential safety net — everything below works identically, responses are canned and labeled mock). Note the startup line agent backend: opencode|mock.
  2. Two browser windows side by side: left = Ava (host), right = empty (this becomes Sam). Use a normal window + a private/incognito window so the two sessions don't share storage.
  3. Left window: open http://localhost:3000, register an account, log in, type Ava, click Start session.

Fallback note: if the opencode backend is active and the configured model has tool use, you'll see tool chips (the AI actually reading files). On mock there are no chips — skip the "watch the tool chips" line and the demo still works.

The script

0:00 — The problem (say it while typing). "Ava is a software engineer. Her inference kernels are mysteriously slow on the company's new accelerator board. Normally, when she hits a hardware question, she'd screenshot the AI's answer into chat, relay the hardware team's reply back, and lose context on every hop. Watch instead."

0:15 — Host works with the AI on real files. Ava sends:

Compare the kernel profiles in the demo folder and flag anything the hardware team needs to weigh in on.

Point at the streaming response — and, on the opencode backend with tool use, at the tool chips: "The AI is doing real work here, reading actual files on the server, not roleplaying." (It reads kernel-profile.md / board-specs.md — five real kernels; four healthy, and kv_cache_gather stuck at 704 GB/s, just 22% of the board's 3,200 GB/s peak.)

0:50 — The wall. "The AI found the smoking gun — unaligned DMA reads falling back to a slow path. But is that really errata E7? And is the padding workaround safe to ship? That answer lives with the hardware team — with Sam. Today this becomes a 20-message relay thread. In TagTeam, Ava tags Sam in."

1:00 — Tag in. Ava clicks Tag in a colleague → Copy link. Mention the guardrail as you copy: "single-use link, expires in 30 minutes." Paste the URL into the right-hand window. Sam registers / logs in, types Sam, clicks Join session.

1:15 — The money shot. Sam's window shows the full transcript — the kernel-profile analysis, everything. Both rosters show Ava, Sam, and the AI. Point at the "Sam joined the session" line landing live in Ava's window.

1:25 — The expert talks to the AI directly. Sam sends:

Is the kv_cache_gather slowdown the unaligned-DMA errata, and can we pad the KV entries to 256 bytes without blowing the memory budget?

The AI answers addressing Sam by name, pulling from hw-constraints.md (errata E7 confirmed — the 160-byte KV stride breaks the 256-byte DMA alignment; padding restores the fast path but costs ~60% more HBM in that region; the gather-scatter DMA mode needs Sam's sign-off). Ava sees every token live. "No relay. No paraphrase loss. Both humans and the AI in one room, and the AI knows who's who."

2:10 — Guardrails. In Ava's roster, hover Sam's row:

  1. Click the eye icon → view-only. Sam's composer locks with a banner; have Sam try to type. "The host stays in control."
  2. Click it again → restore. Sam can send again.
  3. Click ✕ twice → kick. Sam gets a clean "removed from session" screen and the invite link is dead — single-use, remember.

(Optional, if time allows: mint a second link for Kai, a junior engineer who shadows the debugging session view-only — sessions fit two guests.)

2:40 — Close. "Single Node server, in-memory + SQLite, npm install && npm start. Deliberately deferred to a later drop-in: real identity/SSO (Entra/Google/Okta — the auth interface is already OIDC-shaped), Slack/Teams delivery of invites, and Kubernetes horizontal scale — this same protocol carries straight over. Stop playing telephone with your AI. Tag your expert in."

Recovery notes (if something goes sideways)

  • Accidental tab reload (either window): the session auto-resumes from sessionStorage and the SQLite-backed state — keep talking, don't restart.
  • AI errors mid-answer: the turn closes with an error line in the transcript; just resend. Or restart with MOCK_CLAUDE=1 — the multiplayer story needs no live model.
  • Invite pasted wrong / expired: mint a new one — takes two clicks; the guest cap counts joined guests, not minted links.
  • Second demo run: just refresh the host window and start a new session (the old one keeps living in memory + SQLite until GC'd; that's fine).