(WIP) feat: pause/resume runtime, run store, useChat

[marslavish]

Pause/Resume Agent Runtime + React Bindings

Implements the test infrastructure and pause/resume runtime defined in ROADMAP.md. The single architectural change: the agent loop becomes checkpoint-able. Tools may declare a decision JSON Schema; when the loop hits one, it persists run state, emits a tool_decision_pending event, and waits for agent.resume(runId, decision) before continuing.

Phase 0 — Test Infrastructure (must)

• 0.1 Test conventions — deterministic-by-default unit tests; live tests gated behind *.live.test.ts; one env per package (jsdom only in @agentic-kit/react)
• 0.2 Shared test helpers — tools/test/ with createScriptedProvider, createScriptedSSEResponse, parseSSEStream, fixtures; existing tests migrated
• 0.4 SSE wire-format tests — packages/agent/__tests__/sse.test.ts covers split chunks, multi-line data:, comments, event-type framing, mid-event abort
• 0.3 Integration test lane — tests/integration/ scaffolded; first tests land with 1.1 / 1.3 over real HTTP
• 0.2 (cont.) runRunStoreContractTests — portable contract suite for any RunStore implementation

Phase 1 — Pause/Resume + React Bindings (must)

• 1.1 Pausable tools — decision schema on AgentTool; tool_decision_pending event; agent.resume(runId, decision); reuses validateSchema for the decision payload
• 1.2 RunStore — RunStore interface + MemoryRunStore default; consumer supplies Redis / KV / DB backend in production
• 1.3 Run serialization helpers — AgentRunHandle with events(), toReadableStream(), toResponse(); canonical SSE wire format on Content-Type: text/event-stream
• 1.4 @agentic-kit/react — new package; useChat hook with send, respondWithDecision, abort, lifecycle callbacks (onMessage, onFinish, onDecisionPending)

Phase 2 — Production Polish (should)

• 2.1 Prompt caching API — cache?: 'short' | 'long' flag, per-provider translation (Anthropic cache_control, OpenAI native hint, Ollama silent ignore)
• 2.2 Telemetry / middleware hooks — before/after provider call, before/after tool call, stream-event tap; transient-vs-terminal error type

Phase 3 — Optional Extensions (could)

• 3.1 Full Ollama tool support — streaming tool-call parsing on the Ollama adapter
• 3.2 Retry / backoff — opt-in transient retry layer with jittered exponential backoff
• 3.3 Stream resume on disconnect — reconnect via runId, Last-Event-ID semantics
• 3.4 Client-side tool execution — runs: 'client' flag for browser-only capabilities (deferred until a real use case appears)

Out of scope

See ROADMAP.md § Non-Goals: conversation-history persistence, generateObject analog, schema-library coupling, framework-specific helpers, UI components, embeddings as primary capability, prompt-construction utilities, conversation modes, production storage backends.

Open questions

• Run record schema versioning (compatibility surface once RunStore ships)
• runRunStoreContractTests coverage for discriminated-union / oneOf decision schemas
• agent_resume event vs. redocumenting agent_start as "loop entry"
• SSE vs. NDJSON default
• useChat auto-routing on respondWithDecision
• Live test policy for paid providers (Anthropic / OpenAI credits)

See ROADMAP.md § Open Questions.