AI Playground — Prompt / Agent Workbench

• Area: Intelligence
• Path: services/ai/playground
• Kind: Web workbench for prompt and agent prototyping (multi-model, side-by-side, share, export)
• Status: v0.0.1 — sector bootstrapping (2026-05-09)

Role in the stack

playground flattens the prompt-iteration loop. Today prompt prototyping happens in scripts, ngrok-exposed FastAPI demos, or — worst case — directly in production. Each developer has their own setup; comparing two prompts on the same input requires running them serially and squinting at logs. Trying a new model means editing config and redeploying. None of this scales beyond one or two people.

This sector is the workbench: open the page → type a prompt → pick two models side-by-side → tweak temperature → see streaming results → save the winning version to prompt/ → export a ticket. Iteration time drops from minutes to seconds. Pre-product surface for designers, PMs, and customer-success folks who shouldn't need dev help to experiment.

It is the Koder analog of OpenAI Playground, Anthropic Workbench, and HuggingFace Spaces. Distinct from kode — kode is the IDE-class product (full project context, file editing, multi-turn agent loops); playground is the no-state workbench for "try this and tell me how it does."

Boundary vs neighbors

• services/ai/kode is IDE-class — different surface, different audience.
• services/ai/gateway, services/ai/runtime, services/ai/modelreg, services/ai/prompt are primary producers (model exec, available models, curation, prompt registry).
• services/ai/tools provides the tool palette for agent-style runs.
• services/ai/trace integrates for debug-from-trace-viewer.
• services/ai/eval receives promoted runs as eval cases.

Features (v1 target)

• Flutter Web app at playground.koder.dev (reuses koder_kit auththemei18n/safe-area)
• Multi-model side-by-side chat (1-pane or 2-pane)
• Streaming responses token-by-token; abort mid-stream
• Parameter knobs per pane: temp, topp, topk, max_tokens, stop, seed
• System prompt editor + tool palette (toggle per run)
• Model picker driven by modelreg curated tags
• Prompt registry integration: load, edit (with diff), save as new version, set active
• Variable inputs auto-rendered from {{var}} template placeholders
• Saved sessions library + 60s auto-snapshot
• Signed share links (read-only, scoped, revocable, 30d default expiry)
• One-click export-as-ticket to any accessible backlog
• Promote-to-eval action (curator role) → ships run as eval case
• Per-pane timing badges (ttft, total) + token + cost estimate
• Trace_id link for every run

Primary couplings

RFC and bootstrap

• RFC: playground-RFC-001-foundations.kmd — accepted 2026-05-09
• Bootstrap ticket: services/ai/backlog/done/137-playground-bootstrap.md
• Implementation tickets: services/ai/playground/backlog/pending/{001..005}

Self-hosted-first analysis (5 gates)