Glossary

Activity chip. The small line shown while a tool runs, such as Reading the policy…. From a tool's activity() and the activity event.

AI Gateway. Cloudflare's proxy for model traffic. Gives request logs, cost, retries and rate limiting. On when both gateway ids are set.

AiProvider. The model port in api/src/ai/provider.ts. Gemini and the scripted model are its adapters.

Brand. The names the assistant goes by: claimpilot (Claimpilot Chat in ClaimPilot, the default and what hosts get) or clonepilot (the demo).

Capability. Something the assistant can do for a claim: a playbook prompt plus optional tools. Public half in the shared catalogue, server half in api/src/capabilities/.

Claim brief. The compact summary of the claim in the system prompt (buildClaimBrief). Detail comes from tools.

Claim contract. The Claim zod schema in packages/shared/src/claim.ts. Eight required fields; everything else defaulted.

Claimpanion. frontend/public/claimpanion.html: a pretend host claims system that embeds the element.

ClaimRepository. The claims store port. D1, Neon and "none" are its adapters.

ClaimView. Everything a claim screen displays, derived once by deriveClaimView. Both product layouts render it.

Client action. Something the model asks the browser to do on screen, such as highlight_panel. Forwarded as a client_action event.

Clonepilot. The demo app's brand name.

D1. Cloudflare's SQLite database. The default store; a local SQLite file under wrangler dev.

data-panel. The attribute that marks a region of a claim screen the assistant can highlight.

Demo hint. The floating tip on a claim saying what it is good for showing, from metadata.demo.

DGPH-34. The Davies claims-processing ontology the Claim document and the generator follow.

Dock placement. Where a docked chat goes: layout (a column in the app), overlay (pinned to the viewport, embedded) or none.

Embed key. EMBED_KEY: a shared key on the model-facing endpoints. A budget guard, not access control.

Fingerprint. A short hash naming the set of Prompt Lab overrides that answered a turn, or default.

Fixture. One of the three hand-written claims in packages/shared/src/fixtures/. The generator's examples and the tests' data.

Gotcha. A trap planted in a generated claim's record for the assistant to find.

Locale block. system.locale_uk or system.locale_us: the spelling, date style, terminology and accent prompt.

Lookups. The expandable trace under an answer listing each tool call and its result.

MatterHome. UK ClaimPilot's claim screen.

Neon. Serverless Postgres. The alternative store, selected with DATABASE_PROVIDER=neon.

Playbook. A capability's prompt, capability.<id>, describing what good looks like for that task.

Prompt Lab. The page for editing, versioning and trying prompts. Versions live in the browser.

providerState. An adapter's opaque replay data for one assistant turn. For Gemini, the model's own parts with their thought signatures.

Quick action. Running a capability from a chip, the More menu or a demo hint.

Region. UK or US ClaimPilot: chrome, palette, formats and default currency, from regions.ts.

Safety prompt. system.safety: the nine boundaries, always assembled before any claim data.

Scenario. A brief the generator writes a claim around, with the capabilities it showcases and the gotchas to plant.

Screen action. A tool that acts on the page rather than the claim. Offered only when the caller sets screenActions.

Scripted model. api/src/ai/scripted.ts, the deterministic offline adapter the tests run against (AI_PROVIDER=scripted).

Shared tool. One of the five general tools every capability can use: get_claim_section, search_claim, get_document, financial_summary, highlight_panel.

Thought signature. Gemini 3's signature on function-call parts, which must be replayed verbatim.

Turn. One user message and the assistant's answer to it: one POST /api/chat, up to eight model round-trips.

Turn id. The id minted for each answer, in the done event, that ties a feedback vote to it.