diff --git a/README.md b/README.md
index 9fefe68..7184105 100644
--- a/README.md
+++ b/README.md
@@ -2,7 +2,7 @@
 
 A provider-agnostic Go library for building AI agent loops with pluggable tools, memory, and middleware.
 
-Forge handles the **LLM call → tool execution → response** cycle. You supply a provider (Anthropic, OpenAI, etc.), register tools, and forge runs the loop — including error handling, iteration limits, and conversation memory.
+Forge handles the **LLM call -> tool execution -> response** cycle. You supply a provider (Anthropic, OpenAI, etc.), register tools, and forge runs the loop, including error handling, iteration limits, and conversation memory.
 
 ## Install
 
@@ -10,7 +10,7 @@ Forge handles the **LLM call → tool execution → response** cycle. You supply
 go get github.com/katasec/forge
 go get github.com/katasec/forge/provider/anthropic  # optional
 go get github.com/katasec/forge/provider/openai      # optional
-go get github.com/katasec/forge/provider/xai         # optional — xAI Responses API with web search
+go get github.com/katasec/forge/provider/xai         # optional: xAI Responses API with web search
 ```
 
 ## Quick Start
@@ -92,7 +92,7 @@ type Provider interface {
 
 ### Tools
 
-Define tools with `Func[T]` — the JSON schema for parameters is derived from the Go struct at construction time using [invopop/jsonschema](https://github.com/invopop/jsonschema):
+Define tools with `Func[T]`. The JSON schema for parameters is derived from the Go struct at construction time using [invopop/jsonschema](https://github.com/invopop/jsonschema):
 
 ```go
 type SearchInput struct {
@@ -123,9 +123,9 @@ type Tool interface {
 
 1. Load conversation history from memory
 2. Call the provider with messages + tool definitions
-3. If the provider says **stop** → return the response
-4. If the provider requests **tool use** → execute tools, feed results back, go to 2
-5. If **iteration limit** hit → return with `FinishReasonIterLimit`
+3. If the provider says **stop**, return the response
+4. If the provider requests **tool use**, execute tools, feed results back, go to 2
+5. If **iteration limit** hit, return with `FinishReasonIterLimit`
 6. Save conversation to memory
 
 For the common case, use `Ask`:
@@ -147,8 +147,8 @@ Use `Run` when you need full control over message roles, multiple messages, or a
 
 Controls what happens when a tool returns an error:
 
-- `ErrorPolicyStop` (default) — terminate the loop immediately
-- `ErrorPolicyContinue` — feed the error back to the LLM so it can adapt
+- `ErrorPolicyStop` (default): terminate the loop immediately
+- `ErrorPolicyContinue`: feed the error back to the LLM so it can adapt
 
 ### Middleware
 
@@ -172,7 +172,7 @@ agent, _ := forge.NewAgent(forge.Config{
 })
 ```
 
-Middleware composes as decorators: given `[A, B, C]`, request flows `A → B → C → provider → C → B → A`.
+Middleware composes as decorators: given `[A, B, C]`, request flows `A -> B -> C -> provider -> C -> B -> A`.
 
 ### Memory
 
@@ -243,8 +243,9 @@ if meta, ok := forge.MetadataFromContext(ctx); ok {
 
 See the [`_examples`](./_examples) directory for runnable demos:
 
-- **[hello-world](./_examples/hello-world)** — Simplest possible example: call Claude or xAI with one flag swap
-- **[calculator](./_examples/calculator)** — Agent with math tools, middleware, and a mock provider
+- **[hello-world](./_examples/hello-world)**: simplest possible example: call Claude or xAI with one flag swap
+- **[calculator](./_examples/calculator)**: agent with math tools, middleware, and a mock provider
+- **[chat-console](./_examples/chat-console)**: interactive console app showing `forge.Config`, explicit memory, `Ask`, and `LastText`
 
 ## License
 
diff --git a/_examples/calculator/README.md b/_examples/calculator/README.md
index 7492f30..6467024 100644
--- a/_examples/calculator/README.md
+++ b/_examples/calculator/README.md
@@ -1,6 +1,6 @@
 # Calculator
 
-Demonstrates tool use with a mock provider — no API key needed.
+Demonstrates tool use with a mock provider. No API key needed.
 
 ## Run
 
@@ -31,6 +31,8 @@ Everything is in `main.go`:
 - **Tools**: `add` and `multiply` using `forge.Func[T]` with typed inputs
 - **Mock provider**: Simulates an LLM that decides to call the `add` tool, then formulates a response from the result
 - **Logging middleware**: Prints each provider call and its finish reason
-- **Agent loop**: Shows the full cycle — provider call → tool execution → provider call → stop
+- **Agent setup**: Uses `forge.Config` to wire provider, tools, middleware, and loop policy
+- **Agent loop**: Shows the full cycle: provider call, tool execution, provider call, stop
+- **Common API**: Uses `agent.Ask` and `resp.LastText` rather than manual message wiring
 
 This example is useful for understanding how the agent loop works without needing API credentials.
diff --git a/_examples/calculator/main.go b/_examples/calculator/main.go
index 11899ca..c757b0c 100644
--- a/_examples/calculator/main.go
+++ b/_examples/calculator/main.go
@@ -1,4 +1,4 @@
-// Calculator example — demonstrates forge with math tools and a mock provider.
+// Calculator demonstrates forge with math tools and a mock provider.
 //
 // This uses a mock provider that simulates an LLM deciding to call tools.
 // Replace MockProvider with a real provider (Anthropic, OpenAI, etc.) to
@@ -103,7 +103,7 @@ func main() {
 		}
 	})
 
-	// Build the agent.
+	// Build the agent runtime once: provider, tools, middleware, and loop policy.
 	agent, err := forge.NewAgent(forge.Config{
 		Provider:      &MockProvider{},
 		Tools:         []forge.Tool{addTool, mulTool},
@@ -116,7 +116,7 @@ func main() {
 		log.Fatal(err)
 	}
 
-	// Run the agent.
+	// Ask drives the full provider -> tool -> provider loop.
 	fmt.Println("User: What is 12 + 30?")
 	fmt.Println(strings.Repeat("-", 40))
 
diff --git a/_examples/chat-console/README.md b/_examples/chat-console/README.md
index 4de8d54..9bb3432 100644
--- a/_examples/chat-console/README.md
+++ b/_examples/chat-console/README.md
@@ -3,6 +3,7 @@
 An interactive console app that shows the intended forge developer experience:
 
 - configure an agent once with `forge.Config`
+- plug in an explicit memory implementation with `memory/inmem`
 - talk to it with `Ask`
 - read the latest answer with `LastText`
 - keep conversation context in memory across turns
diff --git a/_examples/hello-world/README.md b/_examples/hello-world/README.md
index 19b9567..e882d41 100644
--- a/_examples/hello-world/README.md
+++ b/_examples/hello-world/README.md
@@ -1,42 +1,32 @@
 # Hello World
 
-The simplest possible forge example — call an LLM and get a response.
+The smallest forge example: pick a provider, build an agent with `forge.Config`, ask one question, and print the latest assistant text.
 
-## Run with Claude (Anthropic)
+## Run with Claude
 
 ```bash
 export ANTHROPIC_API_KEY=sk-ant-...
 go run .
 ```
 
-## Run with Grok (xAI)
+## Run with Grok
 
 ```bash
 export XAI_API_KEY=xai-...
 go run . -provider xai
 ```
 
-## What's in here
+## Run with xAI Search
 
-| File | What it does |
-|------|-------------|
-| `main.go` | Picks a provider, builds an agent, sends "Hello!", prints the response |
-| `anthropic.go` | `AnthropicProvider` — implements `forge.Provider` using the Anthropic Messages API |
-| `openai_compat.go` | `OpenAIProvider` — implements `forge.Provider` for any OpenAI-compatible API (xAI, OpenAI, Together, Groq, etc.) |
-
-## Swapping providers
-
-The only thing that changes is one line:
-
-```go
-// Claude
-provider := NewAnthropicProvider(os.Getenv("ANTHROPIC_API_KEY"), "claude-sonnet-4-20250514")
-
-// xAI Grok
-provider := NewOpenAIProvider("https://api.x.ai/v1", os.Getenv("XAI_API_KEY"), "grok-3-mini")
-
-// OpenAI
-provider := NewOpenAIProvider("https://api.openai.com/v1", os.Getenv("OPENAI_API_KEY"), "gpt-4o")
+```bash
+export XAI_API_KEY=xai-...
+go run . -provider xai-search
 ```
 
-Everything else — agent config, tools, middleware, memory — stays the same regardless of provider.
+## What this shows
+
+- `forge.Config` as the agent setup point
+- provider swapping without changing app code
+- `agent.Ask(ctx, prompt)` for the common path
+- `resp.LastText()` for the latest assistant answer
+- provider-specific access to xAI citations when search is enabled
diff --git a/_examples/hello-world/main.go b/_examples/hello-world/main.go
index a85f490..79cfe0f 100644
--- a/_examples/hello-world/main.go
+++ b/_examples/hello-world/main.go
@@ -1,4 +1,4 @@
-// Hello World — the simplest possible forge example.
+// Hello World is the simplest possible forge example.
 //
 // Shows how to call Claude with your Anthropic API key, and how to
 // swap to xAI's Grok by changing one line.
@@ -8,7 +8,7 @@
 //	export ANTHROPIC_API_KEY=sk-ant-...
 //	go run .
 //
-//	# Or use xAI (OpenAI-compatible) instead:
+//	# Or use xAI's OpenAI-compatible endpoint instead:
 //	export XAI_API_KEY=xai-...
 //	go run . -provider xai
 //
@@ -34,7 +34,7 @@ func main() {
 	providerFlag := flag.String("provider", "anthropic", "Provider to use: anthropic, xai, or xai-search")
 	flag.Parse()
 
-	// Pick your provider — this is the only thing that changes.
+	// Pick your provider. The agent setup below stays the same.
 	var provider forge.Provider
 	var xaiProvider *xai.Provider // for citation access
 	switch *providerFlag {
@@ -61,7 +61,7 @@ func main() {
 		log.Fatalf("Unknown provider: %s (use 'anthropic', 'xai', or 'xai-search')", *providerFlag)
 	}
 
-	// Build the agent — same code regardless of provider.
+	// Build the agent: provider, prompt, and runtime behavior live in Config.
 	agent, err := forge.NewAgent(forge.Config{
 		Provider:     provider,
 		SystemPrompt: "You are a helpful assistant. Keep responses brief.",
@@ -70,7 +70,7 @@ func main() {
 		log.Fatal(err)
 	}
 
-	// Ask preserves conversation history on this agent by default.
+	// Ask is the common path: user text in, AgentResponse out.
 	resp, err := agent.Ask(context.Background(), "Hello! What are you?")
 	if err != nil {
 		log.Fatal(err)
@@ -79,12 +79,12 @@ func main() {
 	fmt.Println(resp.LastText())
 	fmt.Printf("\n[%s | tokens: %d in, %d out]\n", *providerFlag, resp.Usage.InputTokens, resp.Usage.OutputTokens)
 
-	// Show citations if using xai-search.
+	// xAI search exposes provider-specific citations after the run.
 	if xaiProvider != nil {
 		if citations := xaiProvider.LastCitations(); len(citations) > 0 {
 			fmt.Println("\nSources:")
 			for i, c := range citations {
-				fmt.Printf("  [%d] %s — %s\n", i+1, c.Title, c.URL)
+				fmt.Printf("  [%d] %s - %s\n", i+1, c.Title, c.URL)
 			}
 		}
 	}
diff --git a/agent.go b/agent.go
index d83dfde..32fb5f4 100644
--- a/agent.go
+++ b/agent.go
@@ -33,7 +33,7 @@ func (r *AgentResponse) LastText() string {
 	return ""
 }
 
-// Agent orchestrates the LLM call → tool execution → response loop.
+// Agent orchestrates the LLM call -> tool execution -> response loop.
 type Agent struct {
 	provider              Provider
 	registry              *ToolRegistry
@@ -156,7 +156,7 @@ func (a *Agent) Run(ctx context.Context, req AgentRequest) (*AgentResponse, erro
 			break
 		}
 
-		// FinishReason is tool_use — execute the tool calls.
+		// FinishReason is tool_use - execute the tool calls.
 		toolResults := a.executor.Execute(ctx, resp.Message.ToolCalls)
 
 		// Check for tool errors.
diff --git a/docs/design/design.md b/docs/design/design.md
index 7c427c1..137cd76 100644
--- a/docs/design/design.md
+++ b/docs/design/design.md
@@ -1,4 +1,4 @@
-# Forge — Type Specification & Design
+# Forge - Type Specification & Design
 
 > This is the authoritative specification for the forge Go agent framework.
 > All implementation should conform to this document.
@@ -7,7 +7,7 @@
 
 ```
 module github.com/katasec/forge
-go 1.23
+go 1.25.6
 ```
 
 Root package: `package forge` — the primary user-facing facade: `Config`, `Agent`, `Ask`, `AskIn`, core interfaces, and type aliases for common concepts.
@@ -121,10 +121,10 @@ const (
 ```
 
 **Semantics:**
-- `stop` — the provider returned a response with no tool calls. The agent loop terminates normally.
-- `tool_use` — the provider requested tool calls. The loop continues to execute them. This value appears in `ProviderResponse` but never as a final `AgentResponse.FinishReason` (the loop always processes tool calls before returning).
-- `iter_limit` — the loop hit `Config.MaxIterations`. The agent returns whatever content the last assistant message contained.
-- `error` — a provider error occurred, or a tool error occurred with `ErrorPolicyStop`.
+- `stop` - the provider returned a response with no tool calls. The agent loop terminates normally.
+- `tool_use` - the provider requested tool calls. The loop continues to execute them. This value appears in `ProviderResponse` but never as a final `AgentResponse.FinishReason` (the loop always processes tool calls before returning).
+- `iter_limit` - the loop hit `Config.MaxIterations`. The agent returns whatever content the last assistant message contained.
+- `error` - a provider error occurred, or a tool error occurred with `ErrorPolicyStop`.
 
 ### TokenUsage
 
@@ -149,8 +149,8 @@ const (
 ```
 
 Controls behavior when a tool invocation returns an error (`ToolResult.IsError == true`):
-- `stop` — the agent loop terminates immediately with `FinishReasonError`.
-- `continue` — the error is fed back to the LLM as a tool result so it can recover or try a different approach.
+- `stop` - the agent loop terminates immediately with `FinishReasonError`.
+- `continue` - the error is fed back to the LLM as a tool result so it can recover or try a different approach.
 
 ### ToolError
 
@@ -180,9 +180,9 @@ func WithMetadata(ctx context.Context, m Metadata) context.Context
 func MetadataFromContext(ctx context.Context) (Metadata, bool)
 ```
 
-- `WithMetadata` stores a `Metadata` in the context using the unexported `metadataKey`.
+- `WithMetadata` stores a `Metadata` in the context using the unexported metadata key.
 - `MetadataFromContext` retrieves it; returns `false` if not present.
-- `Metadata.Values` is never nil after construction — `WithMetadata` should initialize the map if nil.
+- `Metadata.Values` is never nil after construction. `WithMetadata` should initialize the map if nil.
 
 ---
 
@@ -338,7 +338,7 @@ Middleware wraps `RunFunc` in the standard decorator pattern.
 **Composition order:** middlewares are applied innermost-last. Given `[A, B, C]`:
 
 ```
-request → A → B → C → provider.Generate → C → B → A → response
+request -> A -> B -> C -> provider.Generate -> C -> B -> A -> response
 ```
 
 Applied as:
@@ -421,7 +421,7 @@ func (a *Agent) AskIn(ctx context.Context, conversationID, prompt string) (*Agen
 - Both return the full `AgentResponse`, preserving access to conversation ID, token usage, finish reason, errors, and message history.
 - `LastText` returns the latest assistant text content in the response.
 
-### Agent Loop — `Agent.Run(ctx, req)`
+### Agent Loop - `Agent.Run(ctx, req)`
 
 ```
 func (a *Agent) Run(ctx context.Context, req AgentRequest) (*AgentResponse, error)
@@ -463,7 +463,7 @@ Pseudocode:
 31.         finishReason = FinishReasonStop
 32.         break LOOP
 33.
-34.     // FinishReason is tool_use — execute the tool calls
+34.     // FinishReason is tool_use - execute the tool calls
 35.     toolResults = executor.Execute(ctx, providerResp.Message.ToolCalls)
 36.
 37.     // Check for tool errors
@@ -495,10 +495,10 @@ Pseudocode:
 ```
 
 **Key behaviors:**
-- Provider errors (line 23) are always fatal — they return an error, not an `AgentResponse`.
+- Provider errors (line 23) are always fatal. They return an error, not an `AgentResponse`.
 - Tool errors with `ErrorPolicyContinue` are collected in `errors` but the loop continues, letting the LLM see the error and adapt.
 - Tool errors with `ErrorPolicyStop` break the loop immediately but still include the tool results in the message history.
-- `FinishReasonToolUse` never appears in the final `AgentResponse` — the loop always processes tool calls.
+- `FinishReasonToolUse` never appears in the final `AgentResponse`. The loop always processes tool calls.
 - Memory is enabled by default with an in-memory store. It is saved once at the end, with the complete conversation.
 - File-backed or database-backed memory must be explicitly configured because persistence changes privacy and lifecycle expectations.
 - Context cancellation is respected: `composedRunFunc` and `tool.Invoke` should check `ctx`.
@@ -515,4 +515,3 @@ Tracked separately in `docs/design/future-refinements.md`. Not in scope for the
 - Persistent memory stores (SQLite, Redis)
 - Structured output / response format constraints
 - Token budget management (auto-truncate history)
-- Agent-to-agent delegation
diff --git a/docs/design/future-refinements.md b/docs/design/future-refinements.md
index f40a597..23739b5 100644
--- a/docs/design/future-refinements.md
+++ b/docs/design/future-refinements.md
@@ -10,4 +10,3 @@
 | Persistent memory stores | SQLite, Redis, or file-backed `MemoryStore` implementations |
 | Structured output | Response format constraints passed to providers |
 | Token budget management | Auto-truncate or summarize history when approaching limits |
-| Agent-to-agent delegation | One agent invoking another as a tool |
diff --git a/provider/xai/xai.go b/provider/xai/xai.go
index 50cf1ca..01a9083 100644
--- a/provider/xai/xai.go
+++ b/provider/xai/xai.go
@@ -293,7 +293,7 @@ func parseResponse(resp *response) (*forge.ProviderResponse, []Citation) {
 				}
 			}
 			// Server-side tool calls (web_search_call, x_search_call, etc.)
-			// are auto-executed by xAI — we don't surface them.
+			// are auto-executed by xAI, so we don't surface them.
 		}
 	}