Decisions

Why a harness program meets Link as it does, where the interface leaves a choice.

Switching harnesses

The two harnesses' contracts say which state a switch carries over the Agent State Protocol (ASP) and which capabilities it leaves behind. Native file formats belong to each adapter, so the switch asks the contract rather than reading them.

Why does a switch say what stays behind, rather than carry it?

An export does not imply that another harness can import it. The contract declares both directions of ASP state separately, so a switch carries conversations where the source exports them and the destination imports them. It names the unsupported parts before changing anything; the source harness's own state stays available for a switch back.

Why does a switch ask before it leaves something?

A switch you didn't expect to lose something is the one surprise worth asking about, so lnk agent harness set asks before it switches, and --yes agrees ahead for a script.

Why is what stays behind computed from the two contracts?

Each contract declares state.export and state.import for its conversations, native_memory, identity and notes for memory, and features for skills and persona. Contract::left_behind computes the answer from both contracts, tested without a harness running. An outside adapter gets the same behavior by declaring the same fields; absent state capabilities promise no conversation transfer.

Why does a switch carry conversations whole?

ASP represents a session as its Agent Client Protocol (ACP) updates, so an export need not replace a conversation with a summary. Each adapter's import determines which updates it retains: the text projection keeps user and assistant messages and omits other content. A sessions capability declares the supported direction; it does not certify native-state fidelity.

Why is portable context independent of connection transport?

A native harness can keep its gateways while reading and contributing the same durable records. The typed protocol uses the existing authenticated bridge, store protocol, agent identity, leases and file CAS. The local file driver stays rooted in protected collections, so access needs neither a chat endpoint nor an unrestricted folder mount. Automatic agent startup grants storage only to harnesses declaring a restricted collection layout; their local file collections remain there.

Portable views

Portable views give common meaning to exposed event fields, sourced memory documents and finite cursors. Original JSON remains the archive. Native bindings and resume/context policy stay in adapter state, with no new session, memory or handoff service. Transcript roles and correlation establish no authority. Portable context documents the client.

Why may a namespace hold a call no model is offered?

A harness sometimes calls an agent's tools itself. Link Harness keeps a quiet conversation's compaction with the scheduler as each turn ends, and the model must never see or make that call. Since the harness is handed the scheduler only through its namespaces, the call has to live in one. It is cron's compact, declared with harness = true.

Unlisted, still checked

A harness action is left out of the namespace's tools/list, so no model is offered it, and Link Harness refuses a call to any tool it didn't list. When the harness calls it, the namespace server checks it as every other call. The provider's declared schema refuses an argument it doesn't take, and its approval policy still gates the call. The scheduler declares compact for that reason, rather than accepting an undeclared name, and never asks for it, since no one is there to ask as a turn ends.

One way on every machine

The same declaration builds the namespace on the owner's machine and on a Link Cloud box, so compaction is kept through cron on both. A separate variable for the scheduler, as before, made the box and the machine hand different tools.