Display local-agent tool cards
Negotiate tool.cards explicitly to receive structured live and saved tool cards. The backend owns built-in/MCP execution, approvals, cancellation, and history; separately negotiated client-provided tools execute in the client. Cards are display data, not instructions or proof that an operation succeeded.
Consume live cards
- Key cards by child, session, run, and activity identity.
- Apply increasing revisions as replacement snapshots, not patches.
- Use explicit parent/task IDs. Do not guess hierarchy from labels or tool names.
- Keep missing facts unknown. A stopped connection does not turn an unfinished card into success.
- Preserve safety notices and omission counts even when your layout hides optional detail.
Base-only clients keep their original event and history shapes. Rich clients receive the advertised schema descriptor, tool.card events, and saved activity attachments.
Reopen saved cards
Use the attachment on history part 0. Its ID belongs to that immutable history revision. Do not reconstruct a card by reading current files or pairing adjacent events yourself. Missing detail is not_recorded, not an empty successful result.
See the wire schema, shared projection, resource facts, execution facts, and web/image/control facts.
Archived implementation and acceptance evidence records source ownership, pinned process observations, and verification limits.