Host-owned Discord AI continuations

The application may opt a successful standalone ai, grok or zen answer into plain-reply continuation. This is not sticky channel routing or a command string to replay. The adapter stores only closed routing metadata in its existing 1,000-entry command cache, alongside revision ownership and actual Discord reply IDs. There is no additional conversation-text session store.

Application contract

remember-command-continuation! accepts :command, :provider, :model and a 64-character lowercase :provider-signature, plus optional boolean :no-tools and :verified-free, bounded :toolset and integer :max-tool-calls (1–20). Unknown fields, callbacks, prompts, credentials and executable text are not accepted. It returns :accepted, :stale or :unavailable; invalid metadata throws a static IllegalArgumentException. Applications must verify successful standalone AI provenance before calling it; this is trusted host code, not an EDN/API capability or a model tool.

Admission waits for successful terminal delivery and a non-error inspection state. Working cards, partial/live answers before completion, uncertain sends and error cards do not become eligible parents. Cache eviction, source edits, message deletion and process restart invalidate ownership. An eligible reply must name a prior tracked answer in its own channel, with a normal reference. Bot authors are ignored. Explicit prefixed commands override the inherited route; embedded-looking syntax inside an admitted plain reply remains literal. Duplicate admitted gateway creates do not repeat a turn.

During a trusted active follow-up, command-followup returns its route. It returns nil for an ordinary command and :stale if follow-up ownership was lost. Applications must fail closed on :stale, never fall back to parsing the human body. Recheck after reading context and before inference. Dispatch the literal prompt through the chosen AI entrypoint, re-resolve provider identity, check the current actor/tool policy and retain existing context limits. A verified-free route must reverify the same model, never silently rotate or become paid. Explicit paid follow-ups are ordinary new paid requests, with no fallback or automatic failure retry.

Native command replies, no-ping controls, frozen public-result inspection and requester-only rerun confirmation retain their existing ownership and transport rules. Routing metadata is private host state, never an Explore data snapshot.