Ask Playbook on AgentsKit Chat

Ask Playbook is a production dogfood host for @agentskit/chat. The host keeps the playbook corpus, endpoint, brand, CTA, and CSS. AgentsKit Chat owns the definition, session, ordered content, standard components, and React shell; AgentsKit owns messages, streaming, memory, cancellation, retry, editing, and regeneration.

Host recipe

1. Pin the stable npm package @agentskit/chat at exact version 0.3.0 and import renderer and protocol APIs from its /react and /protocol subpaths. Its published manifest resolves the internal graph without host overrides.
2. Build one defineChat definition with the public standard component manifest, an AgentsKit AdapterFactory, and ChatMemory.
3. Load and verify the Playbook-owned deterministic artifact on first intent; exact patterns, prompts, templates, scripts, phases, and navigation answer locally before the fallback adapter is considered.
4. Project the Ask service's validated NDJSON records into ordered text and the standard source-list component. Unknown records stay inert.
5. Render AgentChat and use native React slots only for Playbook branding, linked prose, the composer, loading copy, and citations.
6. Keep corpus="playbook" and NEXT_PUBLIC_ASK_ENDPOINT in this host.

The host shell lives in components/ask-widget.tsx; deterministic composition lives in lib/discovery.ts; artifact generation lives in scripts/lib/playbook-discovery.mjs. Do not add a reducer, stream loop, message store, cancellation state, matcher, or second component protocol to the host.

Migration and parity evidence

The former widget managed messages, streaming, AbortController, NDJSON, sessionStorage, message rendering, and scrolling itself. It was removed only after the AgentsKit Chat integration covered:

• text, citations, errors, done, malformed and unknown records;
• cancellation through the public adapter source;
• migration from ak:ask-thread-v2:playbook to canonical serialized memory;
• safe, bounded citations and ordered text + source-list rendering;
• verified local answers, ambiguity choices, semantic escalation, artifact failure/retry, and truthful local/backend provenance;
• empty/loading states, close, clear, send, stop, retry, edit, and regenerate;
• production build plus responsive evidence at 375, 768, 1280, and 1440 px, with no sub-44 px panel targets.

See AgentsKit Chat issue #27 for the shared Registry/Playbook evidence and upstream-adoption record.