User companion

DelphAI xLM

Standalone user-facing runtime for chat, saved prompts, schedule-driven briefings, and league-grounded sports assistance.

Conversation canvas

Session opening-drive

This route models the immersive chat canvas: recent turns, context disclosures, tool expectations, and the handoff points for saving reports or scheduling a follow-up run.

Session contract

The frontend owns the conversation canvas and user intent. The Foundry runtime owns prompt assembly, tool invocation, and persistence handoffs.

Mode

Interactive briefing

The chat path is responsible for live answers and follow-up control.

Grounding

Persisted session

Conversation history loaded from companion.inference_run.

Trace posture

Live traces

The conversation is being reconstructed from persisted orchestration turns.

Open route

Open session history

Move from an active conversation to saved reports and prior runs.

Open route

Promote to prompt asset

Reusable briefings belong in the prompt library, not only in chat history.

Open route

Schedule follow-up run

Queue a later lineup or readiness check from the current conversation.

Transcript

The answer should show its work

The transcript surface is where the user sees assistant reasoning boundaries, saved context, and the difference between real capability and blocked capability.

Companion runtime

Recovered 2 persisted turns for this conversation.

Conversation history loaded from companion.inference_run.

User

Build me a quick MLB briefing for today and call out any lineup readiness risk.

Sent 2026-05-09 03:14:41.501112+00.

DelphAI Companion

I assembled this MLB response with the DelphAI Companion prompt stack. If you need a tighter lineup answer, send a specific gameId so the runtime can switch from schedule lookup to game-level readiness detail.

Model local-preview/deterministic-preview-v1.

User

Build me a quick MLB briefing for today and call out any lineup readiness risk.

Sent 2026-05-09 03:14:42.073562+00.

DelphAI Companion

I assembled this MLB response with the DelphAI Companion prompt stack. If you need a tighter lineup answer, send a specific gameId so the runtime can switch from schedule lookup to game-level readiness detail.

Model local-preview/deterministic-preview-v1.

Open conversation history

Context and tools

The surrounding surfaces are already defined

These supporting panels make the next integration work local: PostgreSQL prompt assets, Cosmos-backed turns, and stateless MCP tool calls.

Context bundle

Recent preference snapshot

Favorite league: MLB. Preferred output: concise briefings. Citation style: tool and table disclosure.

Prompt asset

Saturday slate briefing

A reusable prompt asset should pre-wire schedule ranges, readiness checks, and the preferred report shape.

Notification intent

Late lineup follow-up

A schedule handoff can queue a notification when lineups post or readiness shifts materially.

MLB tool

mlb_schedule_lookup

Returns normalized schedule slices with season, date, and team filters and explicit live or stale state.

MLB tool

mlb_game_lookup

Returns one grounded game shell, event summary, and readiness posture without mutating session state.

Shared tool

league_toolkit_capabilities

Lets the companion disclose blocked NBA capability rather than pretending the data exists.