Hyperia v0.20.2 — agents get mail 📬

Agents stop shouting into each other's panes. They now have a mailbox, deliveries you approve actually go through, and every permission is tied to a real identity instead of a display name.

An inbox for every agent

Agents used to talk to each other by typing raw text into the other pane: no length limit, no record, no reply channel. There's now a local mailbox:

• msg_send leaves a durable message for an agent or a pane, up to 16 KB.
• msg_inbox previews your mail, msg_check fetches your unread mail and marks it read, msg_read acknowledges one message, and msg_search searches what you've sent and received.

The recipient gets a short, coalesced "you've got mail" notice in its pane. The notice never carries the message body, and it doesn't wait for the agent to go quiet first. Mail is addressed to a verified identity, never a display label, and an agent that runs under its own token links itself to its pane with pane_bind.

Typing a long message straight into another agent's pane is capped at 512 characters, and the error points you to the mailbox. Pass allow_long if you really need to paste it.

"Allow" actually does the thing

Previously, approving an agent's request after the ~8 second wait granted the permission but dropped the action, so the agent had to try again. Now the request is saved first and a worker runs it once you approve. Each request gets an operation ID you can look up with delivery_status, retries don't create duplicates, and if Hyperia isn't sure a write reached the pane it won't blindly send it again.

Permissions follow identities, not names

Grants, owners and consent now use a canonical identity key (agent:<name> / pane:<id>). A display label can no longer inherit someone else's access. On first launch your saved permissions are migrated once. Any grant that can't be tied to a single known agent is dropped, and that agent will ask again. Pane and tab streams check permission on every keystroke, and anonymous viewers stay read-only.

Safer ways to type into panes

• terminal_run only runs at a real shell prompt. It refuses to type a shell command into an agent or a busy pane.
• pane_send is the new way to hand an agent text, with Enter as a separate step.
• terminal_keys is for explicit control keys and never adds an Enter on its own.

Automated re-pokes now say what they are: [Hyperia auto-poke …], not a person messaging you, and how to stop it with pane_pulse_clear. The tool descriptions also warn that raw control characters sent into another agent's pane can crash it.

Heads-up for scripts and clients: refresh your tool list after upgrading. /api/type-and-collect now returns a saved shell operation instead of screen output, so read the output explicitly with terminal_screen.

Also fixed

• 2×2 quick layout: the original pane no longer gets a scrollbar with its old prompt pushed into scrollback when you apply the layout.
• Restored tabs: a tab you never renamed now gets a "(2)" suffix when you restore a copy while the original is still open, the same as renamed tabs.

Coming from further back? v0.19.7 is where saved layouts became "saved tabs".