Arena

Where this fits

Arena is the during-turn, operator-initiated comparison surface in the conversation quality lifecycle. You run the same prompt against multiple configured models, inspect bounded metadata, then accept, dismiss, stitch, or merge a winner into the session.

Shadow Evals measure change across sessions; arena compares models inside a session before you commit to one answer. That makes Arena a decision-forum concept, adjacent to the Sessions continuity family rather than a session restore/checkpoint primitive.

Changes live output?

Only after you accept (or auto-select accepts for you). Until then, status surfaces are metadata-only — no candidate bodies in /arena or bmo arena status.

Reach for it when

You want two or more model voices on the same prompt — for example before merging a docs paragraph, recipe tweak, or verified worktree change — without guessing which model answered best.

Live BMO TUI: `/arena` after a text-mode comparison round — shared status shows enablement, eligibility, `candidate_summary`, and a metadata-only `latest_round` with per-model keys (no candidate bodies).

Inspect

Job	Route	What it shows
Maintainer posture	`bmo arena status` [`--session-id`] [`--json`]	Detached/live posture, enablement, eligibility, candidate totals, latest round summary
TUI round launcher	`/arena`	Same snapshot plus create-round form
Remote / agent	`GET /v1/sessions/{id}/arena/status`, `get_arena_status`	Parity read surface (mutations stay on create/accept tools)

The snapshot never includes raw prompt text, candidate bodies, validation stdout, diffs, or worktree paths.

Walkthrough: docs paragraph shootout

You are refactoring this feature page and want two model voices on the opening copy before merge.

Step	Action	Question answered
1	Enable `[options.arena]` with `mode = "text"` and two models	Is arena armed for this workspace?
2	Open `/arena` in an active session	What is eligibility and session-bound posture?
3	Enter a prompt and create a round	How many candidates ran and in what state?
4	Review metadata in `/arena` or chat; accept one candidate	Which answer becomes the next assistant message?

Optional: pair with Shadow Evals when auto_select_policy = "eval_shadow" — a separate, opt-in policy that picks a winner from shadow scorecards for the same session and user message.

Creating and resolving rounds

TUI: /arena — inspect status, enter a prompt, optionally override models.
Agent / MCP: create_arena, arena_accept, arena_dismiss, stitch, merge.
HTTP: session/message arena routes for remote clients.

Verified mode

Set options.arena.mode = "verified" for worktree-backed candidates with an optional validation_command. Arena status reports whether the current working directory is eligible (valid Git checkout with at least one commit).

Conversation quality — lifecycle map
Concept Map — where Arena sits in the docs taxonomy
Sessions — session lifecycle, branches, and checkpoints
Shadow Evals — across-session measurement; eval_shadow auto-select
Slash commands
CLI commands
Configuration reference
Implementation notes: arena topic