Exercises

Sprinter exercises are the OG Amble workshop primitive — ranking, voting, scoring, assessment, generating, answering — composed onto Sprinter's sessions, views, criteria_sets, and entity_responses substrate. No new tables, no new session_type. The discriminator is sessions.metadata.exercise_kind.

Spec & plan: documents/work/2026-05-03-exercises-module/spec.md and plan.md. PR1 ships the foundation; PR2 ships the runtime; PR5-slim ships the agent-facilitated demo.

Architecture

Facilitator (wizard surface)              Participants (rank/swipe/form surface)
        │                                          │
        ▼                                          ▼
  exercise-setup wizard ──────────► Exercise (mixed session)
   1. pick kind                       metadata.exercise_kind
   2. attach criteria                 metadata.target_entity_ids[]
   3. select target entities          metadata.invited_user_ids[]
   4. invite users                    metadata.timer
   5. set timer                       metadata.target_view_id
                                       │
                                       ▼
                                 Per-user response child sessions
                                 (session_type='response', status='draft')
                                       │
                                       ▼
                                 entity_responses (source='workshop')
                                       │
                                       ▼
                                 computeResponseAggregate / leaderboard

The diagram above is the entire module. Boxes on the right side are already-shipped Sprinter primitives — features/sessions, features/views, features/responses. The new code is the LEFT side (wizard + the kind→surface mapping + the registered tool) plus PR4's Inngest timer. Everything else is composition.

Platform-native pattern: one tool, two callers

createExercise is registered as a first-class AI tool in features/tools/exercise-tools.ts. The setup wizard's "Create" button and an in-chat agent delegation both invoke that same tool — neither uses a custom code path. The only observable downstream difference is session_events.metadata.origin:

This keeps humans + agents truly interchangeable. The PR5 demo ("Run a 10-minute scoring exercise on these 5 concepts with the team") is a tool-call away from PR1's foundation.

Six exercise kinds → view surfaces

The voting → swipe mapping delivers the OG Trinder pattern as a free byproduct.

PR1 surface

Shipped:

• features/exercises/types.ts — EXERCISE_KINDS, ExerciseTimerSchema, ExerciseMetadataSchema, parseExerciseMetadata, EXERCISE_SURFACE_BY_KIND, assertNever.
• features/exercises/server/create-exercise.ts — manual-rollback transactional insert (view + session + session.created event).
• features/exercises/server/list-exercises-for-user.ts — RLS-respecting gallery filter.
• features/tools/exercise-tools.ts — createExercise registered as an AI tool with requiredPermission: "entities.team.create".
• app/exercises/page.tsx — invitee gallery.
• app/exercises/new/page.tsx — minimal setup form (the polished 5-step wizard is the iteration-2 follow-up).

Cross-references

• content/docs/features/tool-system.mdx — tool registry and dispatch.
• content/docs/features/sessions.mdx — session lifecycle and session_events.
• content/docs/features/view-system.mdx — surface types (rank, swipe, form).
• content/docs/features/response-system.mdx — entity_responses, criteria sets, aggregation.