[ryannzander]

chat.send and chat.history, the messages table, a mock layer that runs the whole thing with nothing real behind it, and two screens. Plus the discovery that next build has never worked in this repo.

Screens. Wallet is the main one, navbar leads to the chat, per Ryan. The wallet leads with a panel drawn as the Flipper's amber LCD, and when a proposal is waiting that panel becomes the request, so the screen and the device in your hand show the same words at the same time. Statuses are worded for a person: "Waiting on you", not PENDING_HUMAN.

AGENT_MODE=mock stands in for the chain, the device and the model at once, since all three are missing and none of them are our code. ScriptedResponses matches keywords and calls the real tools through the real loop, so only the choice of tool is fake. Opt-in, because an invented balance looks exactly like a real one, and both screens say so when it is on.

Design calls in the chat layer

• The transcript and the model's history are one table. Split apart they could disagree, and a model that saw a different conversation from the one on screen produces symptoms that all read as the model being stupid.
• The user's row is written before the turn runs. A turn can fail after the model has already written a proposal, and a proposal that nothing in the transcript asked for is worse than a dangling user row.

Two bugs found on the way

• The scripted router read an amount of 0 out of the digits inside a recipient address, so "send some eth to 0x7099…" would have proposed sending nothing. Caught by a test.
• next build has never succeeded, on any commit. Confirmed against the tree before these changes, not assumed. @tappy/protocol imports its own modules as ./types.js, which Node and tsc resolve and webpack does not. And lint had never run, so errors had piled up in loop.ts, approvals.ts and schema.ts. Turbopack serves dev and is more forgiving, which is why nobody hit either.

CI ran neither the web tests, nor lint, nor the build. 52 tests only ever ran on somebody's laptop. All three are in the workflow now. The build is the only check that resolves modules the way Vercel will.

Checked: 52 web tests, 8 protocol tests, typecheck, lint, and a production build all pass locally.

Not checked, and this is what to do before merging. Nobody has seen either screen, verify:chat has never run, and migration 0001 has never been applied. This machine had no Postgres, no Docker and no .env. All three need a database and nothing else:

pnpm --filter @tappy/web db:migrate
pnpm --filter @tappy/web verify:chat        # scripted model, no API key needed
AGENT_MODE=mock pnpm --filter @tappy/web dev

Still open: chat.send and chat.history are unauthenticated. See the comment at the top of chat.ts and the earlier comment on this PR. Blocks deploying, not merging.

[ryannzander]

A background security review flagged this branch, and it is right on all three counts. They are one root cause: chat.send and chat.history are publicProcedure and there is no middleware.ts anywhere in the app.

No money moves without the human's press. What does break on a public URL:

• a stranger spends the OpenAI budget and makes the Flipper buzz mid-demo;
• one stranger proposal sits at PENDING_HUMAN and insertProposal refuses a second, so the real one is locked out for ten minutes;
• the transcript is one global conversation, so their text is in the history the model reads on the next turn.

Not covered by DECISIONS #10, which is about the agent's tools. trpc.ts already draws the same line for the bridge.

Documented in the router and in docs/HANDOFF.md rather than fixed. Picking a scheme decides how the UI talks to the server, which is not mine to choose. This needs an answer before the app is deployed anywhere, not before this PR merges.