(WIP) Redesign agentic-kit core around structured streams and add minimal agent runtime

README.md

@@ -11,38 +11,66 @@
    <a href="https://github.com/constructive-io/agentic-kit/blob/main/LICENSE"><img height="20" src="https://img.shields.io/badge/license-MIT-blue.svg"/></a>
 </p>
 
-A unified, streaming-capable interface for multiple LLM providers.
+A provider-portable LLM toolkit with structured streaming, model registries,
+cross-provider message normalization, and an optional stateful agent runtime.
 
 ## Packages
 
-- **agentic-kit** — core library with provider abstraction and `AgentKit` manager
+- **agentic-kit** — low-level portability layer with model descriptors, registries, structured event streams, and compatibility helpers
+- **@agentic-kit/agent** — minimal stateful runtime with sequential tool execution and lifecycle events
 - **@agentic-kit/ollama** — adapter for local Ollama inference
 - **@agentic-kit/anthropic** — adapter for Anthropic Claude models
-- **@agentic-kit/openai** — adapter for OpenAI and OpenAI-compatible APIs
+- **@agentic-kit/openai** — generalized adapter for OpenAI-compatible chat completion APIs
 
 ## Getting Started
 
 ```bash
 git clone git@github.com:constructive-io/agentic-kit.git
 cd agentic-kit
-yarn install
-yarn build
-yarn test
+pnpm install
+pnpm build
+pnpm test
 ```
 
 ## Usage
 
 ```typescript
-import { createOllamaKit, createMultiProviderKit, OllamaAdapter } from 'agentic-kit';
+import { complete, getModel } from "agentic-kit";
 
-const kit = createOllamaKit('http://localhost:11434');
-const text = await kit.generate({ model: 'mistral', prompt: 'Hello' });
+const model = getModel("openai", "gpt-4o-mini");
+const message = await complete(model!, {
+  messages: [{ role: "user", content: "Hello", timestamp: Date.now() }],
+});
 
-// Multi-provider
-const multi = createMultiProviderKit();
-multi.addProvider(new OllamaAdapter('http://localhost:11434'));
+console.log(message.content);
 ```
 
 ## Contributing
 
 See individual package READMEs for docs and local dev instructions.
+
+## Testing
+
+Default tests stay deterministic and local:
+
+```bash
+pnpm test
+```
+
+There is also a local-only Ollama live lane that does not hit hosted
+providers. The default root command runs the fast smoke tier:
+
+```bash
+OLLAMA_LIVE_MODEL=qwen3.5:4b pnpm test:live:ollama
+```
+
+Run the broader lane explicitly when you want slower behavioral coverage:
+
+```bash
+OLLAMA_LIVE_MODEL=qwen3.5:4b pnpm test:live:ollama:extended
+```
+
+The Ollama live script performs a preflight against `OLLAMA_BASE_URL` and exits
+cleanly if the local server or requested model is unavailable. If
+`nomic-embed-text:latest` is installed, the lane also exercises local embedding
+generation.

REDESIGN_DECISIONS.md

@@ -0,0 +1,76 @@
+# Agentic Kit Redesign Decisions
+
+Date: 2026-04-18
+
+This document records the redesign decisions made while evaluating `agentic-kit`
+against the comparable `pi-mono` architecture, especially `packages/ai` and
+`packages/agent`.
+
+## Scope and Package Boundaries
+
+1. `agentic-kit` remains the low-level provider portability layer.
+2. Stateful orchestration moves into a separate `@agentic-kit/agent` package.
+3. Tool execution stays out of `agentic-kit` core; the core only models tools,
+   tool calls, and tool results.
+4. `@agentic-kit/agent` v1 should be intentionally minimal, shipping only the
+   sequential tool loop, lifecycle events, abort/continue, and pluggable context
+   transforms. Steering/follow-up queues and richer interruption policies are
+   deferred to phase 2.
+
+## Core Type System
+
+5. Core tool definitions use plain JSON Schema.
+6. TypeBox/Zod support stays as helper adapters, not the core contract.
+7. Core models are represented by a provider-independent `ModelDescriptor`
+   registry with capability metadata.
+8. The model registry must support both built-in descriptors and runtime
+   registration of custom models/providers from day one.
+9. The core message model treats `image` input and `thinking` output as
+   first-class content blocks in v1.
+10. `usage`, `cost`, `stopReason`, and abort-driven partial-result semantics are
+    mandatory parts of the core contract in v1.
+
+## Streaming and Conversation Semantics
+
+11. Structured event streams become the primary streaming primitive; text-only
+    chunk callbacks remain as convenience wrappers.
+12. Cross-provider replay and handoff is a hard requirement for v1, including
+    normalization for reasoning blocks, tool-call IDs, aborted turns, and
+    orphaned tool results.
+
+## Provider Strategy
+
+13. OpenAI-compatible backends should be handled by one generalized adapter path
+    with compatibility flags, not many first-class provider packages.
+14. Embeddings stay out of the primary conversational core and live behind a
+    separate optional capability interface or companion package.
+
+## Migration Strategy
+
+15. `agentic-kit` should ship a backward-compatibility layer for the current
+    `generate({ model, prompt }, { onChunk })` API for one transition release.
+
+## Architectural Implications
+
+These decisions imply the following target architecture:
+
+- `agentic-kit`
+  Low-level portability layer. Owns message/content types, model descriptors,
+  provider registry, streaming event protocol, compatibility transforms, usage,
+  and provider adapters.
+- `@agentic-kit/agent`
+  Optional stateful runtime. Owns tool execution, sequential loop semantics,
+  lifecycle events, context transforms, and abort/continue behavior.
+- Separate optional capabilities or companion packages
+  For non-conversational workloads such as embeddings, and optional schema
+  helpers such as TypeBox/Zod integration.
+
+## Design Principles Confirmed
+
+- Keep the protocol portable and runtime-agnostic.
+- Normalize provider differences in the core instead of leaking them upward.
+- Treat OpenAI-compatible APIs as a compatibility class, not a brand-specific
+  architecture.
+- Avoid coupling the low-level layer to any single schema library or vendor SDK.
+- Preserve a migration path from the existing text-only API while moving the
+  real architecture to structured messages and events.

apps/tanstack-chat-demo/.cta.json

@@ -0,0 +1,16 @@
+{
+  "projectName": "tanstack-chat-demo",
+  "mode": "file-router",
+  "typescript": true,
+  "tailwind": true,
+  "packageManager": "pnpm",
+  "git": false,
+  "install": false,
+  "addOnOptions": {},
+  "includeExamples": false,
+  "envVarValues": {},
+  "routerOnly": false,
+  "version": 1,
+  "framework": "react",
+  "chosenAddOns": []
+}

apps/tanstack-chat-demo/.gitignore

@@ -0,0 +1,13 @@
+node_modules
+.DS_Store
+dist
+dist-ssr
+*.local
+.env
+.nitro
+.tanstack
+.wrangler
+.output
+.vinxi
+__unconfig*
+todos.json

apps/tanstack-chat-demo/.vscode/settings.json

@@ -0,0 +1,11 @@
+{
+  "files.watcherExclude": {
+    "**/routeTree.gen.ts": true
+  },
+  "search.exclude": {
+    "**/routeTree.gen.ts": true
+  },
+  "files.readonlyInclude": {
+    "**/routeTree.gen.ts": true
+  }
+}