Design notes for what needs a decision first
REMAINING.md lists items that need a design decision before they are code. These notes are proposals to decide on, not built features. Each one says what exists today (checked against the code), what the risk is, what I would build, what I would refuse to build, and the decisions only you can make, with my recommendation for each. Nothing here has been started, except Credential sources, which was decided and then built (tested against fakes; no real Vault has been tried).
| Note | The question |
|---|---|
| Browser workflows | How does an agent drive a website, with you able to take over and with a real confirm-before-submit? |
| Payments | How can an agent buy something without ever holding a card number? |
| Mail approvals for real-world mail | How do HTML and attachments get approved without approving what you cannot see? |
| Agent-proposed tools | How can an agent add a capability to itself without adding authority to itself? |
| A Windows companion | What would a Windows client be, and what must it never do? |
| Credential sources | Can a secret come from Vault or a mounted file instead of the process environment, without weakening any check? |
| Model gateway | Can agents get provider failover, a cache, and per-key spend and budgets without a key ever reaching a cell? |
| Chat approvals | Can Slack and Teams tell you about an approval, and let you decide it, without weakening the phone signature? |
| Run report | Can a finished run say what was checked, what was not, and how to roll back, without sounding more certain than the evidence? |
| OTLP export | Can the journal reach an OpenTelemetry collector without sending anything the SIEM export would not? |
| Sequence policy | Can a policy stop an agent that has already made too many writes in one session? |
| MCP gateway | Can an operator register tool servers and limit which tools an agent calls, how often, with a record, without a second policy engine? |
| Session recording | Can a person replay what a session did from a signed record that leaves out the content? |
| On behalf of | Can the journal tell a person's delegated action from the agent's own, and can a token be asked for one service? |
| Model budgets | Can an operator cap an agent's model calls and cost per day, per run and per goal, failing closed? |
| Run webhooks | Can a run's end reach another system as a signed event that carries no content? |
| UX pass | What shared motion, focus and heading rules do the console and the site follow? |
| UX plan | What comes after the shared pass to make the console and site feel like apple.com, in order, and how we will know? |
How to read a note: Decisions for you at the end of each is the part to answer. Answer in a line ("1: yes, 2: A, 3: not now") and I will turn the answers into a build plan. Anything you leave open stays open.
Rules that every note assumes (they are already how Merion works, and none of the proposals bends them):
- An approval is decided on the person's own device with a signed decision, never in the agent chat and never by the agent.
- The host, not the agent, renders what is being approved, from the real request.
- A secret (card, password, token) is never given to an agent or a cell; the host injects it and only where the descriptor allows.
- Anything that reads untrusted content marks the session tainted, and a tainted session's actions need a person.
- Nothing is public, signed with a real certificate, notarized or posted without your explicit go.