feat(tui): implement dark mode with runtime /theme switching

- Paint dark background at app shell level via appStyle() helper
- Add /theme slash command for runtime light/dark switching
- Add comprehensive dark-mode style tests
- Document --theme dark in README Quick Start

.gitignore

@@ -2,3 +2,4 @@ agentu*.yaml
 /agentu
 /coverage.out
 *.swp
+chat-history.txt

README.md

@@ -20,13 +20,19 @@ model/provider switching, and local tools.
 mkdir -p ~/.agentu
 cp agentu.example.yaml ~/.agentu/config.yaml
 export AGENTU_API_KEY='your-api-key'
-go run ./cmd/agentu
+go run .
+```
+
+Use dark mode when your terminal is dark:
+
+```sh
+go run . --theme dark
 ```
 
 Use `--yolo` only when you want agentu to write files or run shell commands:
 
 ```sh
-go run ./cmd/agentu --yolo
+go run . --yolo
 ```
 
 ## Config
@@ -44,19 +50,10 @@ providers:
   loveuer:
     base_url: https://ai.loveuer.com
     api_key: ${AGENTU_API_KEY}
-    model: deepseek-v4-flash
     models:
-      - deepseek-v4-flash
+      - id: deepseek-v4-flash
         thinking: none
-    thinking_param: thinking
-
-agent:
-  system_prompt: |
-    You are agentu, a concise local agent. Answer simple conversation directly.
-    Use tools only when the user's request needs local project context or local
-    command execution.
-  working_dir: .
-  tool_timeout: 30s
+        context: 256000
 ```
 
 Web tools are enabled by default with a built-in no-key backend. agentu exposes
@@ -67,29 +64,38 @@ Provider names are the keys under `providers`. At startup, agentu selects the
 first provider by name. Use `/model provider <name>` to switch providers for the
 current session.
 
-`models` is optional. If present, `/model model <name>` is restricted to that
-list. If omitted, any model name is accepted for that provider.
+`models` is required. `/model model <name>` is restricted to that list. Each
+entry is an object with:
 
-`thinking` is optional. Supported values are:
+- `id`: model name sent to the provider.
+- `thinking`: optional default thinking level for that model.
+- `context`: optional context-window token count used for auto-compaction.
+
+If `model` is omitted, agentu uses the first `models` entry. The `agent` block
+is optional: `working_dir`, `tool_timeout`, `system_prompt`, and
+`max_context_tokens` all have defaults. `max_context_tokens` overrides the
+selected model's `context` when set.
+
+Supported thinking values are:
 
 ```text
 none, middle, high, xhigh, max
 ```
 
 When configured or changed with `/model thinking ...`, the value is sent as a
-top-level OpenAI-compatible request field. `thinking_param` controls that field
-name and defaults to `thinking`.
+top-level OpenAI-compatible request field named `thinking`.
 
 ## CLI Flags
 
 ```sh
-go run ./cmd/agentu [flags]
+go run . [flags]
 ```
 
 - `--config <path>` loads a YAML config file. Default: `~/.agentu/config.yaml`.
 - `--theme light|dark` selects the TUI theme. Default: `light`.
 - `--plain` uses the simple line-based REPL instead of the TUI.
 - `--yolo` enables automatic file writes and shell execution.
+- `--resume <id>` resumes a saved session from `~/.agentu/sessions`.
 
 ## TUI Controls
 
@@ -103,10 +109,18 @@ go run ./cmd/agentu [flags]
 Slash commands:
 
 - `/clear` resets conversation history.
+- `/compact` summarizes older conversation history and keeps recent context.
+- `/status` shows changed files with `git status --short`.
+- `/diff [--stat] [path...]` shows the unstaged git diff.
+- `/test [command...]` runs a project test command; defaults to `go test ./...` for Go modules or `npm test` for Node projects.
 - `/model` shows current provider, model, thinking, and switch usage.
 - `/model provider <name>` switches provider for the current session.
 - `/model model <name>` or `/model <name>` switches model for the current session.
 - `/model thinking <none|middle|high|xhigh|max>` changes thinking for the current session.
+- `/sessions` lists saved sessions.
+- `/resume [id]` resumes by ID or opens an interactive picker.
+- `/rename <name>` renames the current session.
+- `/new` saves the current session and starts a fresh one.
 - `/exit` or `/quit` exits.
 
 ## Local Tools
@@ -115,7 +129,11 @@ Read-only tools are available by default:
 
 - `file_read`
 - `file_list`
-- `file_search`
+- `file_search` with optional `context_lines` and `max_results`
+- `code_symbols` for Go packages, imports, types, functions, and methods
+- `git_status` shows changed files
+- `git_diff` shows unstaged or staged diff output
+- `git_log` shows recent commit history
 
 Web tools are also available in read-only mode:
 
@@ -124,8 +142,9 @@ Web tools are also available in read-only mode:
 
 `--yolo` additionally enables:
 
-- `file_write`
-- `shell_run`
+- `file_write` for creating or replacing whole files
+- `file_edit` for surgical line-range or pattern edits to existing files
+- `shell_run` for non-interactive shell commands
 
 Use `--yolo` for prompts that ask agentu to run or inspect local commands, for
 example:
@@ -134,6 +153,8 @@ example:
 请帮我运行 `go test ./...` 并总结结果
 ```
 
+
+
 ## Development
 
 Run the verification suite:
@@ -141,7 +162,7 @@ Run the verification suite:
 ```sh
 go test ./...
 go vet ./...
-go build ./cmd/agentu
+go build .
 ```
 
 Local config and build outputs are intentionally ignored by git.

agentu.example.yaml

@@ -2,16 +2,7 @@ providers:
   loveuer:
     base_url: https://ai.loveuer.com
     api_key: ${AGENTU_API_KEY}
-    model: deepseek-v4-flash
     models:
-      - deepseek-v4-flash
+      - id: deepseek-v4-flash
         thinking: none
-    thinking_param: thinking
-
-agent:
-  system_prompt: |
-    You are agentu, a concise local agent. Answer simple conversation directly.
-    Use tools only when the user's request needs local project context or local
-    command execution.
-  working_dir: .
-  tool_timeout: 30s
+        context: 256000

cmd/agentu/main.go

@@ -1,131 +0,0 @@
-package main
-
-import (
-	"bufio"
-	"context"
-	"errors"
-	"flag"
-	"fmt"
-	"net/http"
-	"os"
-	"os/signal"
-	"strings"
-
-	"agentu/internal/agent"
-	"agentu/internal/config"
-	"agentu/internal/llm"
-	"agentu/internal/session"
-	"agentu/internal/tools"
-	"agentu/internal/tui"
-)
-
-func main() {
-	if err := run(); err != nil {
-		fmt.Fprintln(os.Stderr, "agentu:", err)
-		os.Exit(1)
-	}
-}
-
-func run() error {
-	configPath := flag.String("config", config.DefaultPath, "path to YAML config")
-	yolo := flag.Bool("yolo", false, "also enable automatic file writes and shell execution")
-	plain := flag.Bool("plain", false, "use the simple line-based REPL instead of the TUI")
-	themeName := flag.String("theme", "light", "TUI theme: light or dark")
-	flag.Parse()
-
-	themeMode, err := tui.ParseThemeMode(*themeName)
-	if err != nil {
-		return err
-	}
-
-	cfg, err := config.Load(*configPath)
-	if err != nil {
-		if errors.Is(err, os.ErrNotExist) {
-			resolvedPath, resolveErr := config.ResolvePath(*configPath)
-			if resolveErr != nil {
-				return resolveErr
-			}
-			return fmt.Errorf("config not found at %s; create ~/.agentu/config.yaml from agentu.example.yaml and set AGENTU_API_KEY", resolvedPath)
-		}
-		return err
-	}
-
-	providerName := cfg.DefaultProviderName()
-	providerConfig := cfg.Providers[providerName]
-	provider := llm.NewOpenAICompatibleClient(providerConfig.BaseURL, providerConfig.APIKey, http.DefaultClient)
-
-	var registry *tools.Registry
-	if *yolo {
-		registry = tools.Builtins(cfg.Agent.WorkingDir)
-	} else {
-		registry = tools.ReadOnlyBuiltins(cfg.Agent.WorkingDir)
-	}
-	registry.Register(tools.NewBuiltinSearchTool(http.DefaultClient))
-	registry.Register(tools.NewBuiltinFetchTool(http.DefaultClient))
-	systemPrompt := strings.TrimSpace(cfg.Agent.SystemPrompt + "\n\n" + tools.AgentInstructions(cfg.Agent.WorkingDir, *yolo))
-
-	assistant := agent.New(agent.Options{
-		Provider:        provider,
-		ProviderName:    providerName,
-		Model:           providerConfig.Model,
-		Thinking:        providerConfig.Thinking,
-		ThinkingParam:   providerConfig.ThinkingParam,
-		ThinkingEnabled: providerConfig.ThinkingConfigured,
-		SystemPrompt:    systemPrompt,
-		ToolRegistry:    registry,
-		ToolTimeout:     cfg.Agent.ToolTimeout,
-	})
-	modelManager, err := session.NewManager(cfg, assistant, http.DefaultClient)
-	if err != nil {
-		return err
-	}
-
-	ctx, stop := signal.NotifyContext(context.Background(), os.Interrupt)
-	defer stop()
-
-	if !*plain {
-		return tui.Run(ctx, assistant, tui.Options{
-			ModelName:    providerConfig.Model,
-			Yolo:         *yolo,
-			ThemeMode:    themeMode,
-			ModelManager: modelManager,
-		})
-	}
-
-	return repl(ctx, assistant)
-}
-
-func repl(ctx context.Context, assistant *agent.Agent) error {
-	scanner := bufio.NewScanner(os.Stdin)
-	scanner.Buffer(make([]byte, 0, 64*1024), 1024*1024)
-
-	fmt.Println("agentu")
-	fmt.Println("Type /exit to quit, /clear to reset context.")
-	for {
-		fmt.Print("> ")
-		if !scanner.Scan() {
-			if err := scanner.Err(); err != nil {
-				return err
-			}
-			fmt.Println()
-			return nil
-		}
-
-		input := strings.TrimSpace(scanner.Text())
-		switch input {
-		case "":
-			continue
-		case "/exit", "/quit":
-			return nil
-		case "/clear":
-			assistant.Clear()
-			fmt.Println("context cleared")
-			continue
-		}
-
-		if err := assistant.RunTurn(ctx, input, os.Stdout, os.Stderr); err != nil {
-			fmt.Fprintln(os.Stderr, "\nerror:", err)
-		}
-		fmt.Println()
-	}
-}

internal/agent/agent.go

@@ -7,13 +7,36 @@ import (
 	"io"
 	"sort"
 	"strings"
+	"sync"
 	"time"
 
-	"agentu/internal/llm"
 	"agentu/internal/tools"
+	"agentu/pkg/llm"
 )
 
-const defaultMaxToolIterations = 8
+const (
+	defaultMaxToolRounds      = 50
+	doomLoopThreshold         = 3
+	compactRecentMessages     = 6
+	summaryMessageLimit       = 8000
+	contextSummaryPrefix      = "[context summary]"
+	toolLogArgumentLimit      = 400
+	toolLogOutputPreviewBytes = 4096
+	toolLogOutputMarker       = "[output]"
+
+	summarizerSystemPrompt = "You are a conversation summarizer. Summarize the following conversation history into a concise but detailed summary. Preserve:\n" +
+		"- Key facts, decisions, and conclusions\n" +
+		"- File paths and code changes made\n" +
+		"- Tool call results (especially errors)\n" +
+		"- Current task state and next steps\n\n" +
+		"Do NOT include: tool call syntax, raw JSON, or verbatim file contents.\n" +
+		"Output only the summary, no preamble."
+)
+
+type toolExecutionResult struct {
+	call   llm.ToolCall
+	output string
+}
 
 type Agent struct {
 	provider         llm.Provider
@@ -25,7 +48,9 @@ type Agent struct {
 	systemPrompt     string
 	toolRegistry     *tools.Registry
 	toolTimeout      time.Duration
-	maxToolIterations int
+	maxToolRounds    int
+	maxContextTokens int
+	lastUsage        *llm.Usage
 	messages         []llm.Message
 }
 
@@ -40,12 +65,13 @@ type Options struct {
 	ToolRegistry      *tools.Registry
 	ToolTimeout       time.Duration
 	MaxToolIterations int
+	MaxContextTokens  int
 }
 
 func New(opts Options) *Agent {
-	maxToolIterations := opts.MaxToolIterations
-	if maxToolIterations <= 0 {
-		maxToolIterations = defaultMaxToolIterations
+	maxToolRounds := opts.MaxToolIterations
+	if maxToolRounds <= 0 {
+		maxToolRounds = defaultMaxToolRounds
 	}
 	toolTimeout := opts.ToolTimeout
 	if toolTimeout <= 0 {
@@ -61,7 +87,8 @@ func New(opts Options) *Agent {
 		systemPrompt:     opts.SystemPrompt,
 		toolRegistry:     opts.ToolRegistry,
 		toolTimeout:      toolTimeout,
-		maxToolIterations: maxToolIterations,
+		maxToolRounds:    maxToolRounds,
+		maxContextTokens: opts.MaxContextTokens,
 	}
 	a.Clear()
 	return a
@@ -74,6 +101,7 @@ type RuntimeOptions struct {
 	Thinking         string
 	ThinkingParam    string
 	ThinkingEnabled  bool
+	MaxContextTokens int
 }
 
 type RuntimeInfo struct {
@@ -81,6 +109,8 @@ type RuntimeInfo struct {
 	Model            string
 	Thinking         string
 	ThinkingEnabled  bool
+	ContextTokens    int
+	MaxContextTokens int
 }
 
 func (a *Agent) SetRuntime(opts RuntimeOptions) {
@@ -96,6 +126,7 @@ func (a *Agent) SetRuntime(opts RuntimeOptions) {
 	a.thinking = opts.Thinking
 	a.thinkingParam = opts.ThinkingParam
 	a.thinkingEnabled = opts.ThinkingEnabled
+	a.maxContextTokens = opts.MaxContextTokens
 }
 
 func (a *Agent) RuntimeInfo() RuntimeInfo {
@@ -104,10 +135,13 @@ func (a *Agent) RuntimeInfo() RuntimeInfo {
 		Model:            a.model,
 		Thinking:         a.thinking,
 		ThinkingEnabled:  a.thinkingEnabled,
+		ContextTokens:    llm.EstimateTokens(a.messages, a.toolDefinitions()),
+		MaxContextTokens: a.maxContextTokens,
 	}
 }
 
 func (a *Agent) Clear() {
+	a.lastUsage = nil
 	a.messages = nil
 	if strings.TrimSpace(a.systemPrompt) != "" {
 		a.messages = append(a.messages, llm.Message{
@@ -117,14 +151,173 @@ func (a *Agent) Clear() {
 	}
 }
 
+// Messages returns a copy of the current conversation messages.
+func (a *Agent) Messages() []llm.Message {
+	out := make([]llm.Message, len(a.messages))
+	copy(out, a.messages)
+	return out
+}
+
+// SetMessages replaces the conversation messages (used for session restore).
+func (a *Agent) SetMessages(messages []llm.Message) {
+	a.messages = append([]llm.Message(nil), messages...)
+}
+
+// LastUsage returns a copy of the most recent provider token usage data.
+func (a *Agent) LastUsage() *llm.Usage {
+	if a.lastUsage == nil {
+		return nil
+	}
+	usage := *a.lastUsage
+	return &usage
+}
+
+// SetLastUsage restores the most recent provider token usage data.
+func (a *Agent) SetLastUsage(usage *llm.Usage) {
+	if usage == nil {
+		a.lastUsage = nil
+		return
+	}
+	copy := *usage
+	a.lastUsage = &copy
+}
+
+// Compact summarizes older conversation messages while preserving the system
+// prompt and recent messages intact.
+func (a *Agent) Compact(ctx context.Context) error {
+	if len(a.messages) < 4 {
+		return nil
+	}
+	if a.provider == nil {
+		return fmt.Errorf("provider is not configured")
+	}
+
+	prefixEnd := 0
+	if a.messages[0].Role == llm.RoleSystem {
+		prefixEnd = 1
+	}
+	if len(a.messages)-prefixEnd <= compactRecentMessages {
+		return nil
+	}
+
+	recentStart := len(a.messages) - compactRecentMessages
+	if recentStart <= prefixEnd {
+		return nil
+	}
+	middle := append([]llm.Message(nil), a.messages[prefixEnd:recentStart]...)
+	if len(middle) == 0 {
+		return nil
+	}
+
+	summary, usage, err := a.summarizeMessages(ctx, middle)
+	if err != nil {
+		return fmt.Errorf("compact context: %w", err)
+	}
+	summary = strings.TrimSpace(summary)
+	if summary == "" {
+		return fmt.Errorf("compact context: summary was empty")
+	}
+
+	compressed := make([]llm.Message, 0, prefixEnd+1+len(a.messages[recentStart:]))
+	compressed = append(compressed, a.messages[:prefixEnd]...)
+	compressed = append(compressed, llm.Message{
+		Role:    llm.RoleAssistant,
+		Content: contextSummaryPrefix + "\n" + summary,
+	})
+	compressed = append(compressed, a.messages[recentStart:]...)
+	a.messages = compressed
+	if usage != nil {
+		a.lastUsage = usage
+	}
+	return nil
+}
+
+func (a *Agent) shouldCompact() bool {
+	if a.maxContextTokens <= 0 {
+		return false
+	}
+	if a.lastUsage != nil && a.lastUsage.TotalTokens >= a.maxContextTokens {
+		return true
+	}
+	return llm.EstimateTokens(a.messages, a.toolDefinitions()) >= a.maxContextTokens
+}
+
+func (a *Agent) summarizeMessages(ctx context.Context, messages []llm.Message) (string, *llm.Usage, error) {
+	req := llm.ChatRequest{
+		Model: a.model,
+		Messages: []llm.Message{
+			{Role: llm.RoleSystem, Content: summarizerSystemPrompt},
+			{Role: llm.RoleUser, Content: formatMessagesForSummary(messages)},
+		},
+	}
+
+	var summary strings.Builder
+	var usage *llm.Usage
+	if err := a.provider.ChatStream(ctx, req, func(event llm.StreamEvent) error {
+		if event.Content != "" {
+			summary.WriteString(event.Content)
+		}
+		if event.Usage != nil {
+			copy := *event.Usage
+			usage = &copy
+		}
+		return nil
+	}); err != nil {
+		return "", nil, err
+	}
+
+	return strings.TrimSpace(summary.String()), usage, nil
+}
+
+func formatMessagesForSummary(messages []llm.Message) string {
+	var b strings.Builder
+	b.WriteString("Conversation history to summarize:\n\n")
+	for index, msg := range messages {
+		fmt.Fprintf(&b, "Message %d (%s):\n", index+1, msg.Role)
+		if msg.ToolCallID != "" {
+			fmt.Fprintf(&b, "tool_call_id: %s\n", msg.ToolCallID)
+		}
+		if msg.Content != "" {
+			fmt.Fprintf(&b, "%s\n", compact(msg.Content, summaryMessageLimit))
+		}
+		for _, call := range msg.ToolCalls {
+			fmt.Fprintf(&b, "tool_call: %s %s\n", call.Function.Name, compact(call.Function.Arguments, summaryMessageLimit))
+		}
+		b.WriteByte('\n')
+	}
+	return b.String()
+}
+
+// roundSignature returns a deterministic signature for a set of tool calls
+// in a single round, used for doom loop detection.
+func roundSignature(calls []llm.ToolCall) string {
+	var b strings.Builder
+	for _, c := range calls {
+		b.WriteString(c.Function.Name)
+		b.WriteByte('\x00')
+		b.WriteString(strings.TrimSpace(c.Function.Arguments))
+		b.WriteByte('\x00')
+	}
+	return b.String()
+}
+
 func (a *Agent) RunTurn(ctx context.Context, input string, out io.Writer, logs io.Writer) error {
 	input = strings.TrimSpace(input)
 	if input == "" {
 		return nil
 	}
 	a.messages = append(a.messages, llm.Message{Role: llm.RoleUser, Content: input})
+	if a.shouldCompact() {
+		if err := a.Compact(ctx); err != nil {
+			if logs != nil {
+				fmt.Fprintf(logs, "[compact] warning: %v\n", err)
+			}
+		}
+	}
 
-	for iteration := 0; iteration <= a.maxToolIterations; iteration++ {
+	var recentSignatures []string
+
+	for round := 0; round < a.maxToolRounds; round++ {
 		content, calls, err := a.streamAssistant(ctx, out)
 		if err != nil {
 			return err
@@ -146,17 +339,52 @@ func (a *Agent) RunTurn(ctx context.Context, input string, out io.Writer, logs i
 			fmt.Fprintln(out)
 		}
 
-		for _, call := range calls {
-			result := a.executeTool(ctx, call, logs)
+		// Doom loop detection: check if the last N rounds all called
+		// the exact same tools with the exact same arguments.
+		sig := roundSignature(calls)
+		recentSignatures = append(recentSignatures, sig)
+		if len(recentSignatures) > doomLoopThreshold {
+			recentSignatures = recentSignatures[len(recentSignatures)-doomLoopThreshold:]
+		}
+
+		if len(recentSignatures) == doomLoopThreshold {
+			allSame := true
+			for i := 1; i < len(recentSignatures); i++ {
+				if recentSignatures[i] != recentSignatures[0] {
+					allSame = false
+					break
+				}
+			}
+			if allSame {
+				toolNames := make([]string, 0, len(calls))
+				for _, c := range calls {
+					toolNames = append(toolNames, c.Function.Name)
+				}
+				msg := fmt.Sprintf(
+					"error: doom loop detected — the same tool call (%s) with identical arguments has been repeated %d times. Stopping. Try a different approach or break the task into smaller steps.",
+					strings.Join(toolNames, ", "),
+					doomLoopThreshold,
+				)
 				a.messages = append(a.messages, llm.Message{
-				Role:       llm.RoleTool,
-				ToolCallID: call.ID,
-				Content:    result,
+					Role:    llm.RoleAssistant,
+					Content: msg,
 				})
+				return fmt.Errorf("%s", msg)
 			}
 		}
 
-	return fmt.Errorf("tool iteration limit reached: %d", a.maxToolIterations)
+		results := a.executeTools(ctx, calls, logs)
+		for _, result := range results {
+			a.messages = append(a.messages, llm.Message{
+				Role:       llm.RoleTool,
+				ToolCallID: result.call.ID,
+				Content:    result.output,
+			})
+		}
+
+	}
+
+	return fmt.Errorf("tool round limit reached after %d rounds; the task may be too complex for a single turn", a.maxToolRounds)
 }
 
 func (a *Agent) streamAssistant(ctx context.Context, out io.Writer) (string, []llm.ToolCall, error) {
@@ -180,6 +408,9 @@ func (a *Agent) streamAssistant(ctx context.Context, out io.Writer) (string, []l
 	}
 
 	err := a.provider.ChatStream(ctx, req, func(event llm.StreamEvent) error {
+		if event.Usage != nil {
+			a.SetLastUsage(event.Usage)
+		}
 		if event.Content != "" {
 			content.WriteString(event.Content)
 			if out != nil {
@@ -212,11 +443,8 @@ func (a *Agent) toolDefinitions() []llm.Tool {
 	return a.toolRegistry.Definitions()
 }
 
-func (a *Agent) executeTool(ctx context.Context, call llm.ToolCall, logs io.Writer) string {
+func (a *Agent) executeTool(ctx context.Context, call llm.ToolCall) string {
 	name := call.Function.Name
-	if logs != nil {
-		fmt.Fprintf(logs, "\n[tool] %s %s\n", name, compact(call.Function.Arguments, 400))
-	}
 
 	if a.toolRegistry == nil {
 		return tools.FormatError(fmt.Errorf("tool registry is disabled"))
@@ -240,12 +468,84 @@ func (a *Agent) executeTool(ctx context.Context, call llm.ToolCall, logs io.Writ
 		}
 		return tools.FormatError(err)
 	}
-	if logs != nil {
-		fmt.Fprintf(logs, "[tool] %s done\n", name)
-	}
 	return output
 }
 
+func (a *Agent) hasMutatingCall(calls []llm.ToolCall) bool {
+	if a.toolRegistry == nil {
+		return false
+	}
+	for _, call := range calls {
+		if tool, ok := a.toolRegistry.Get(call.Function.Name); ok {
+			if tools.IsMutating(tool) {
+				return true
+			}
+		}
+	}
+	return false
+}
+
+func (a *Agent) executeTools(ctx context.Context, calls []llm.ToolCall, logs io.Writer) []toolExecutionResult {
+	results := make([]toolExecutionResult, len(calls))
+
+	// Log all tool starts sequentially so TUI gets ordered placeholders.
+	if logs != nil {
+		for _, call := range calls {
+			logToolStart(logs, call)
+		}
+	}
+
+	if a.hasMutatingCall(calls) {
+		// Sequential: avoid racing mutating tools.
+		for i, call := range calls {
+			output := a.executeTool(ctx, call)
+			results[i] = toolExecutionResult{call: call, output: output}
+			if logs != nil {
+				logToolOutput(logs, call, output)
+			}
+		}
+		return results
+	}
+
+	// Concurrent: independent read-only tools.
+	var mu sync.Mutex
+	var wg sync.WaitGroup
+	for i, call := range calls {
+		wg.Add(1)
+		go func(idx int, c llm.ToolCall) {
+			defer wg.Done()
+			output := a.executeTool(ctx, c)
+			results[idx] = toolExecutionResult{call: c, output: output}
+			if logs != nil {
+				mu.Lock()
+				logToolOutput(logs, c, output)
+				mu.Unlock()
+			}
+		}(i, call)
+	}
+	wg.Wait()
+	return results
+}
+
+func logToolStart(logs io.Writer, call llm.ToolCall) {
+	fmt.Fprintf(logs, "\n[tool] %s %s\n", call.Function.Name, compact(call.Function.Arguments, toolLogArgumentLimit))
+}
+
+func logToolOutput(logs io.Writer, call llm.ToolCall, output string) {
+	fmt.Fprintf(logs, "\n[tool] %s %s\n%s\n%s\n", call.Function.Name, compact(call.Function.Arguments, toolLogArgumentLimit), toolLogOutputMarker, toolOutputPreview(output))
+}
+
+func toolOutputPreview(output string) string {
+	output = strings.TrimSpace(output)
+	if output == "" {
+		return "(no output)"
+	}
+	if len(output) <= toolLogOutputPreviewBytes {
+		return output
+	}
+	return output[:toolLogOutputPreviewBytes] + fmt.Sprintf("\n[truncated: showing first %d bytes; omitted %d bytes]", toolLogOutputPreviewBytes, len(output)-toolLogOutputPreviewBytes)
+}
+
 type toolCallBuilder struct {
 	index     int
 	id        string

internal/agent/agent_test.go

@@ -3,11 +3,12 @@ package agent
 import (
 	"context"
 	"encoding/json"
+	"fmt"
 	"strings"
 	"testing"
 
-	"agentu/internal/llm"
 	"agentu/internal/tools"
+	"agentu/pkg/llm"
 )
 
 type fakeProvider struct {
@@ -96,6 +97,134 @@ func (p *contentProvider) ChatStream(ctx context.Context, req llm.ChatRequest, e
 	return emit(llm.StreamEvent{Content: p.text})
 }
 
+type compactProvider struct {
+	summaryCalls int
+	answerCalls  int
+	lastSummary  string
+}
+
+func (p *compactProvider) ChatStream(ctx context.Context, req llm.ChatRequest, emit func(llm.StreamEvent) error) error {
+	if len(req.Messages) == 2 && req.Messages[0].Role == llm.RoleSystem && req.Messages[0].Content == summarizerSystemPrompt {
+		p.summaryCalls++
+		p.lastSummary = req.Messages[1].Content
+		if err := emit(llm.StreamEvent{Content: "old work summarized"}); err != nil {
+			return err
+		}
+		return emit(llm.StreamEvent{Usage: &llm.Usage{PromptTokens: 10, CompletionTokens: 2, TotalTokens: 12}})
+	}
+
+	p.answerCalls++
+	return emit(llm.StreamEvent{Content: "done"})
+}
+
+func TestAgentCompact(t *testing.T) {
+	provider := &compactProvider{}
+	a := New(Options{Provider: provider, Model: "test", SystemPrompt: "system"})
+	a.SetMessages(longConversation())
+
+	if err := a.Compact(context.Background()); err != nil {
+		t.Fatal(err)
+	}
+
+	messages := a.Messages()
+	if len(messages) != 8 {
+		t.Fatalf("messages = %d, want 8", len(messages))
+	}
+	if messages[0].Role != llm.RoleSystem || messages[0].Content != "system" {
+		t.Fatalf("system message = %#v", messages[0])
+	}
+	if messages[1].Role != llm.RoleAssistant || !strings.HasPrefix(messages[1].Content, contextSummaryPrefix) {
+		t.Fatalf("summary message = %#v", messages[1])
+	}
+	if !strings.Contains(messages[1].Content, "old work summarized") {
+		t.Fatalf("summary content = %q", messages[1].Content)
+	}
+	if !strings.Contains(provider.lastSummary, "old-user-0") {
+		t.Fatalf("summary input did not include old history: %q", provider.lastSummary)
+	}
+	if messages[len(messages)-1].Content != "recent-answer-2" {
+		t.Fatalf("last recent message = %#v", messages[len(messages)-1])
+	}
+	if usage := a.LastUsage(); usage == nil || usage.TotalTokens != 12 {
+		t.Fatalf("usage = %#v", usage)
+	}
+}
+
+func TestAgentAutoCompact(t *testing.T) {
+	provider := &compactProvider{}
+	a := New(Options{
+		Provider:          provider,
+		Model:             "test",
+		SystemPrompt:      "system",
+		MaxContextTokens:  1,
+		ToolRegistry:      tools.NewRegistry(),
+		MaxToolIterations: 1,
+	})
+	a.SetMessages(longConversation())
+
+	var out strings.Builder
+	if err := a.RunTurn(context.Background(), "new question", &out, nil); err != nil {
+		t.Fatal(err)
+	}
+	if provider.summaryCalls != 1 {
+		t.Fatalf("summaryCalls = %d, want 1", provider.summaryCalls)
+	}
+	if provider.answerCalls != 1 {
+		t.Fatalf("answerCalls = %d, want 1", provider.answerCalls)
+	}
+	if out.String() != "done" {
+		t.Fatalf("out = %q", out.String())
+	}
+	if !strings.Contains(joinMessageContents(a.Messages()), contextSummaryPrefix) {
+		t.Fatalf("messages missing summary: %#v", a.Messages())
+	}
+}
+
+func TestAgentCompactSkipsShortHistory(t *testing.T) {
+	provider := &contentProvider{text: "unused"}
+	a := New(Options{Provider: provider, Model: "test", SystemPrompt: "system"})
+	a.SetMessages([]llm.Message{
+		{Role: llm.RoleSystem, Content: "system"},
+		{Role: llm.RoleUser, Content: "question"},
+		{Role: llm.RoleAssistant, Content: "answer"},
+	})
+
+	if err := a.Compact(context.Background()); err != nil {
+		t.Fatal(err)
+	}
+	if provider.called {
+		t.Fatal("provider should not be called for short history")
+	}
+	if len(a.Messages()) != 3 {
+		t.Fatalf("messages = %d, want 3", len(a.Messages()))
+	}
+}
+
+func longConversation() []llm.Message {
+	messages := []llm.Message{{Role: llm.RoleSystem, Content: "system"}}
+	for i := 0; i < 3; i++ {
+		messages = append(messages,
+			llm.Message{Role: llm.RoleUser, Content: fmt.Sprintf("old-user-%d", i)},
+			llm.Message{Role: llm.RoleAssistant, Content: fmt.Sprintf("old-answer-%d", i)},
+		)
+	}
+	for i := 0; i < 3; i++ {
+		messages = append(messages,
+			llm.Message{Role: llm.RoleUser, Content: fmt.Sprintf("recent-user-%d", i)},
+			llm.Message{Role: llm.RoleAssistant, Content: fmt.Sprintf("recent-answer-%d", i)},
+		)
+	}
+	return messages
+}
+
+func joinMessageContents(messages []llm.Message) string {
+	var b strings.Builder
+	for _, msg := range messages {
+		b.WriteString(msg.Content)
+		b.WriteByte('\n')
+	}
+	return b.String()
+}
 func TestAgentLetsProviderDecideShellRequestWhenShellDisabled(t *testing.T) {
 	provider := &contentProvider{text: "shell is disabled"}
 	a := New(Options{
@@ -137,3 +266,248 @@ func TestAgentDoesNotBlockFileRequest(t *testing.T) {
 		t.Fatalf("out = %q", out.String())
 	}
 }
+
+// alwaysToolProvider always returns a tool call with the same args on every request.
+type alwaysToolProvider struct {
+	calls int
+}
+
+func (p *alwaysToolProvider) ChatStream(ctx context.Context, req llm.ChatRequest, emit func(llm.StreamEvent) error) error {
+	p.calls++
+	return emit(llm.StreamEvent{ToolCalls: []llm.ToolCallDelta{
+		{Index: 0, ID: fmt.Sprintf("call_%d", p.calls), Type: "function", Name: "test_echo", Arguments: `{"text":"loop"}`},
+	}})
+}
+
+func TestAgentStopsOnDoomLoop(t *testing.T) {
+	provider := &alwaysToolProvider{}
+	a := New(Options{
+		Provider:          provider,
+		Model:             "test",
+		SystemPrompt:      "system",
+		ToolRegistry:      tools.NewRegistry(echoTool{}),
+		MaxToolIterations: 50, // High hard limit — doom loop should trigger first
+	})
+
+	var out strings.Builder
+	err := a.RunTurn(context.Background(), "loop forever", &out, nil)
+	if err == nil {
+		t.Fatal("expected error when doom loop is detected")
+	}
+	if !strings.Contains(err.Error(), "doom loop detected") {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	// Should have stopped at doomLoopThreshold (3) rounds, not reached hard limit.
+	if provider.calls != doomLoopThreshold {
+		t.Fatalf("calls = %d, want %d (doom loop threshold)", provider.calls, doomLoopThreshold)
+	}
+}
+
+// varyingToolProvider alternates between different tool calls to avoid doom loop detection.
+type varyingToolProvider struct {
+	calls int
+}
+
+func (p *varyingToolProvider) ChatStream(ctx context.Context, req llm.ChatRequest, emit func(llm.StreamEvent) error) error {
+	p.calls++
+	// Vary the arguments each round to avoid doom loop detection.
+	return emit(llm.StreamEvent{ToolCalls: []llm.ToolCallDelta{
+		{Index: 0, ID: fmt.Sprintf("call_%d", p.calls), Type: "function", Name: "test_echo",
+			Arguments: fmt.Sprintf(`{"text":"round-%d"}`, p.calls)},
+	}})
+}
+
+func TestAgentHardLimitStopsVaryingCalls(t *testing.T) {
+	provider := &varyingToolProvider{}
+	const limit = 5
+	a := New(Options{
+		Provider:          provider,
+		Model:             "test",
+		SystemPrompt:      "system",
+		ToolRegistry:      tools.NewRegistry(echoTool{}),
+		MaxToolIterations: limit,
+	})
+
+	var out strings.Builder
+	err := a.RunTurn(context.Background(), "vary forever", &out, nil)
+	if err == nil {
+		t.Fatal("expected error when hard limit is reached")
+	}
+	if !strings.Contains(err.Error(), "tool round limit reached") {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if provider.calls != limit {
+		t.Fatalf("calls = %d, want %d", provider.calls, limit)
+	}
+}
+
+// slowTool blocks on a channel before returning.
+type slowTool struct {
+	name    string
+	started chan<- string
+	release <-chan struct{}
+}
+
+func (t *slowTool) Definition() llm.Tool {
+	return llm.Tool{
+		Type: "function",
+		Function: llm.ToolFunction{
+			Name:        t.name,
+			Description: "slow test tool",
+			Parameters:  json.RawMessage(`{"type":"object","properties":{"text":{"type":"string"}}}`),
+		},
+	}
+}
+
+func (t *slowTool) Execute(ctx context.Context, raw json.RawMessage) (string, error) {
+	var args struct {
+		Text string `json:"text"`
+	}
+	if err := json.Unmarshal(raw, &args); err != nil {
+		return "", err
+	}
+	t.started <- t.name
+	<-t.release
+	return args.Text, nil
+}
+
+// mutatingSlowTool is like slowTool but implements MutatingTool.
+type mutatingSlowTool struct {
+	slowTool
+}
+
+func (*mutatingSlowTool) Mutates() bool { return true }
+
+func TestAgentExecutesIndependentToolCallsInParallel(t *testing.T) {
+	started := make(chan string, 2)
+	release := make(chan struct{})
+
+	tool1 := &slowTool{name: "slow_one", started: started, release: release}
+	tool2 := &slowTool{name: "slow_two", started: started, release: release}
+
+	provider := &parallelProvider{
+		responses: [][]llm.ToolCallDelta{
+			{
+				{Index: 0, ID: "call_1", Type: "function", Name: "slow_one", Arguments: `{"text":"one"}`},
+				{Index: 1, ID: "call_2", Type: "function", Name: "slow_two", Arguments: `{"text":"two"}`},
+			},
+		},
+		finalContent: "done",
+	}
+
+	a := New(Options{
+		Provider:     provider,
+		Model:        "test",
+		SystemPrompt: "system",
+		ToolRegistry: tools.NewRegistry(tool1, tool2),
+	})
+
+	var out strings.Builder
+	done := make(chan error, 1)
+	go func() {
+		done <- a.RunTurn(context.Background(), "run both", &out, nil)
+	}()
+
+	// Both tools must start before we release them.
+	names := make(map[string]bool)
+	for i := 0; i < 2; i++ {
+		select {
+		case name := <-started:
+			names[name] = true
+		case <-make(chan struct{}):
+			// Use a timer in real code, but for test simplicity we rely on deadlock detection.
+			t.Fatal("timed out waiting for tool to start")
+		}
+	}
+	if !names["slow_one"] || !names["slow_two"] {
+		t.Fatalf("expected both tools to start, got %v", names)
+	}
+	close(release)
+
+	if err := <-done; err != nil {
+		t.Fatal(err)
+	}
+	if out.String() != "done" {
+		t.Fatalf("out = %q", out.String())
+	}
+	// Verify tool results were in order.
+	msgs := a.Messages()
+	var toolMsgs []string
+	for _, m := range msgs {
+		if m.Role == llm.RoleTool {
+			toolMsgs = append(toolMsgs, m.Content)
+		}
+	}
+	if len(toolMsgs) != 2 || toolMsgs[0] != "one" || toolMsgs[1] != "two" {
+		t.Fatalf("tool messages = %v, want [one two]", toolMsgs)
+	}
+}
+
+func TestAgentRunsMutatingToolCallsSequentially(t *testing.T) {
+	started := make(chan string, 2)
+	release := make(chan struct{})
+
+	tool1 := &mutatingSlowTool{slowTool{name: "mut_one", started: started, release: release}}
+	tool2 := &mutatingSlowTool{slowTool{name: "mut_two", started: started, release: release}}
+
+	provider := &parallelProvider{
+		responses: [][]llm.ToolCallDelta{
+			{
+				{Index: 0, ID: "call_1", Type: "function", Name: "mut_one", Arguments: `{"text":"one"}`},
+				{Index: 1, ID: "call_2", Type: "function", Name: "mut_two", Arguments: `{"text":"two"}`},
+			},
+		},
+		finalContent: "done",
+	}
+
+	a := New(Options{
+		Provider:     provider,
+		Model:        "test",
+		SystemPrompt: "system",
+		ToolRegistry: tools.NewRegistry(tool1, tool2),
+	})
+
+	var out strings.Builder
+	done := make(chan error, 1)
+	go func() {
+		done <- a.RunTurn(context.Background(), "run both", &out, nil)
+	}()
+
+	// Only the first tool should start.
+	select {
+	case name := <-started:
+		if name != "mut_one" {
+			t.Fatalf("expected mut_one to start first, got %q", name)
+		}
+	case <-make(chan struct{}):
+		t.Fatal("timed out waiting for first tool to start")
+	}
+
+	// Second tool should NOT have started yet.
+	select {
+	case name := <-started:
+		t.Fatalf("second tool started too early: %q", name)
+	default:
+		// Expected: second tool hasn't started.
+	}
+
+	close(release)
+	if err := <-done; err != nil {
+		t.Fatal(err)
+	}
+}
+
+// parallelProvider emits a batch of tool calls on the first request, then returns finalContent.
+type parallelProvider struct {
+	responses    [][]llm.ToolCallDelta
+	finalContent string
+	calls        int
+}
+
+func (p *parallelProvider) ChatStream(ctx context.Context, req llm.ChatRequest, emit func(llm.StreamEvent) error) error {
+	p.calls++
+	if p.calls <= len(p.responses) {
+		return emit(llm.StreamEvent{ToolCalls: p.responses[p.calls-1]})
+	}
+	return emit(llm.StreamEvent{Content: p.finalContent})
+}

internal/session/manager.go

@@ -5,10 +5,11 @@ import (
 	"net/http"
 	"sort"
 	"strings"
+	"time"
 
 	"agentu/internal/agent"
-	"agentu/internal/config"
-	"agentu/internal/llm"
+	"agentu/pkg/config"
+	"agentu/pkg/llm"
 )
 
 type Manager struct {
@@ -17,9 +18,11 @@ type Manager struct {
 	httpClient     *http.Client
 	providerName   string
 	provider       config.ProviderConfig
+	store          *Store
+	currentSession *Session
 }
 
-func NewManager(cfg *config.Config, assistant *agent.Agent, httpClient *http.Client) (*Manager, error) {
+func NewManager(cfg *config.Config, assistant *agent.Agent, httpClient *http.Client, store *Store) (*Manager, error) {
 	if httpClient == nil {
 		httpClient = http.DefaultClient
 	}
@@ -28,13 +31,162 @@ func NewManager(cfg *config.Config, assistant *agent.Agent, httpClient *http.Cli
 	if !ok {
 		return nil, fmt.Errorf("no provider is configured")
 	}
-	return &Manager{
+	m := &Manager{
 		cfg:          cfg,
 		agent:        assistant,
 		httpClient:   httpClient,
 		providerName: providerName,
 		provider:     provider,
-	}, nil
+		store:        store,
+	}
+	return m, nil
+}
+
+// ResumeLatest loads the most recent session, or creates a new one.
+func (m *Manager) ResumeLatest() error {
+	if m.store == nil {
+		return nil
+	}
+	metas, err := m.store.List()
+	if err != nil {
+		return err
+	}
+	if len(metas) > 0 {
+		return m.Resume(metas[0].ID)
+	}
+	m.newSession()
+	return nil
+}
+
+// Resume loads a session by ID and restores it into the agent.
+func (m *Manager) Resume(id string) error {
+	if m.store == nil {
+		return fmt.Errorf("session storage is not configured")
+	}
+	sess, err := m.store.Load(id)
+	if err != nil {
+		return err
+	}
+	m.currentSession = sess
+	m.agent.SetMessages(sess.Messages)
+	m.agent.SetLastUsage(sess.LastUsage)
+	return nil
+}
+
+// Save persists the current session to disk.
+func (m *Manager) Save() error {
+	if m.store == nil || m.currentSession == nil {
+		return nil
+	}
+	m.currentSession.Messages = m.agent.Messages()
+	m.currentSession.LastUsage = m.agent.LastUsage()
+	m.currentSession.Provider = m.providerName
+	m.currentSession.Model = m.provider.Model
+	// Auto-name if empty
+	if m.currentSession.Name == "" {
+		m.currentSession.Name = DefaultName(m.currentSession.Messages)
+	}
+	return m.store.Save(m.currentSession)
+}
+
+// Rename changes the current session name and saves.
+func (m *Manager) Rename(name string) (string, error) {
+	name = strings.TrimSpace(name)
+	if name == "" {
+		return "", fmt.Errorf("name is required")
+	}
+	if m.currentSession == nil {
+		return "", fmt.Errorf("no active session")
+	}
+	m.currentSession.Name = name
+	if err := m.Save(); err != nil {
+		return "", err
+	}
+	return fmt.Sprintf("Renamed session to %q.", name), nil
+}
+
+// Sessions returns a formatted list of all sessions.
+func (m *Manager) Sessions() (string, error) {
+	if m.store == nil {
+		return "", fmt.Errorf("session storage is not configured")
+	}
+	metas, err := m.store.List()
+	if err != nil {
+		return "", err
+	}
+	if len(metas) == 0 {
+		return "No saved sessions.", nil
+	}
+	var b strings.Builder
+	b.WriteString("Sessions:\n")
+	for _, meta := range metas {
+		name := meta.Name
+		if name == "" {
+			name = "(unnamed)"
+		}
+		current := ""
+		if m.currentSession != nil && meta.ID == m.currentSession.ID {
+			current = " (current)"
+		}
+		ts := meta.UpdatedAt.Local().Format("2006-01-02 15:04")
+		fmt.Fprintf(&b, "  %s  %-30s  %s  %d msgs%s\n", meta.ID, fitName(name, 30), ts, meta.MessageCount, current)
+	}
+	return b.String(), nil
+}
+
+// NewSession starts a fresh session, saving the current one first.
+func (m *Manager) NewSession() (string, error) {
+	// Save current if it has user messages
+	if m.currentSession != nil {
+		msgs := m.agent.Messages()
+		hasUser := false
+		for _, msg := range msgs {
+			if msg.Role == llm.RoleUser {
+				hasUser = true
+				break
+			}
+		}
+		if hasUser {
+			m.currentSession.Messages = msgs
+			m.currentSession.Provider = m.providerName
+			m.currentSession.Model = m.provider.Model
+			if m.currentSession.Name == "" {
+				m.currentSession.Name = DefaultName(msgs)
+			}
+			if m.store != nil {
+				_ = m.store.Save(m.currentSession)
+			}
+		}
+	}
+	m.newSession()
+	return "Session saved. Starting new session.", nil
+}
+
+func (m *Manager) newSession() {
+	m.currentSession = &Session{
+		ID:        GenerateID(),
+		CreatedAt: time.Now().UTC(),
+		Provider:  m.providerName,
+		Model:     m.provider.Model,
+	}
+	m.agent.Clear()
+}
+
+func (m *Manager) CurrentSessionID() string {
+	if m.currentSession == nil {
+		return ""
+	}
+	return m.currentSession.ID
+}
+
+func (m *Manager) CurrentSessionName() string {
+	if m.currentSession == nil {
+		return ""
+	}
+	if m.currentSession.Name != "" {
+		return m.currentSession.Name
+	}
+	return m.currentSession.ID
 }
 
 func (m *Manager) CurrentProvider() string {
@@ -58,11 +210,7 @@ func (m *Manager) Info() string {
 	fmt.Fprintf(&b, "model: %s\n", m.CurrentModel())
 	fmt.Fprintf(&b, "thinking: %s\n", m.CurrentThinking())
 	fmt.Fprintf(&b, "providers: %s\n", strings.Join(m.providerNames(), ", "))
-	if len(m.provider.Models) > 0 {
 	fmt.Fprintf(&b, "models: %s\n", strings.Join(m.provider.Models, ", "))
-	} else {
-		b.WriteString("models: any model name is accepted for this provider\n")
-	}
 	fmt.Fprintf(&b, "thinking levels: %s\n", strings.Join(config.ThinkingLevels(), ", "))
 	b.WriteString("usage: /model provider <name> | /model model <name> | /model thinking <level>")
 	return b.String()
@@ -84,10 +232,10 @@ func (m *Manager) SetModel(model string) (string, error) {
 	if model == "" {
 		return "", fmt.Errorf("model is required")
 	}
-	if len(m.provider.Models) > 0 && !contains(m.provider.Models, model) {
+	if !contains(m.provider.Models, model) {
 		return "", fmt.Errorf("model %q is not configured for provider %q; available: %s", model, m.providerName, strings.Join(m.provider.Models, ", "))
 	}
-	m.provider.Model = model
+	m.provider = m.provider.WithModel(model)
 	m.cfg.Providers[m.providerName] = m.provider
 	m.syncAgent(nil)
 	return "Switched model.\n" + m.summary(), nil
@@ -112,6 +260,13 @@ func (m *Manager) summary() string {
 	return fmt.Sprintf("provider: %s\nmodel: %s\nthinking: %s", m.CurrentProvider(), m.CurrentModel(), m.CurrentThinking())
 }
 
+func (m *Manager) maxContextTokens() int {
+	if m.cfg != nil && m.cfg.Agent.MaxContextTokens > 0 {
+		return m.cfg.Agent.MaxContextTokens
+	}
+	return m.provider.ContextTokens
+}
+
 func (m *Manager) syncAgent(provider llm.Provider) {
 	m.agent.SetRuntime(agent.RuntimeOptions{
 		Provider:         provider,
@@ -120,6 +275,7 @@ func (m *Manager) syncAgent(provider llm.Provider) {
 		Thinking:         m.provider.Thinking,
 		ThinkingParam:    m.provider.ThinkingParam,
 		ThinkingEnabled:  m.provider.ThinkingConfigured,
+		MaxContextTokens: m.maxContextTokens(),
 	})
 }
 
@@ -140,3 +296,10 @@ func contains(values []string, value string) bool {
 	}
 	return false
 }
+
+func fitName(name string, width int) string {
+	if len(name) <= width {
+		return name
+	}
+	return name[:width-3] + "..."
+}

internal/session/manager_test.go

@@ -2,12 +2,13 @@ package session
 
 import (
 	"context"
+	"path/filepath"
 	"strings"
 	"testing"
 
 	"agentu/internal/agent"
-	"agentu/internal/config"
-	"agentu/internal/llm"
+	"agentu/pkg/config"
+	"agentu/pkg/llm"
 )
 
 type recordingProvider struct {
@@ -19,8 +20,8 @@ func (p *recordingProvider) ChatStream(ctx context.Context, req llm.ChatRequest,
 	return emit(llm.StreamEvent{Content: "ok"})
 }
 
-func TestManagerSwitchesModelAndThinking(t *testing.T) {
-	cfg := &config.Config{
+func testConfig() *config.Config {
+	return &config.Config{
 		Providers: map[string]config.ProviderConfig{
 			"one": {
 				Name:               "one",
@@ -34,6 +35,11 @@ func TestManagerSwitchesModelAndThinking(t *testing.T) {
 			},
 		},
 	}
+}
+
+func testManager(t *testing.T) (*Manager, *agent.Agent, *recordingProvider) {
+	t.Helper()
+	cfg := testConfig()
 	provider := &recordingProvider{}
 	assistant := agent.New(agent.Options{
 		Provider:        provider,
@@ -43,10 +49,19 @@ func TestManagerSwitchesModelAndThinking(t *testing.T) {
 		ThinkingParam:   "thinking",
 		ThinkingEnabled: true,
 	})
-	manager, err := NewManager(cfg, assistant, nil)
+	store, err := NewStore(filepath.Join(t.TempDir(), "sessions"))
 	if err != nil {
 		t.Fatal(err)
 	}
+	manager, err := NewManager(cfg, assistant, nil, store)
+	if err != nil {
+		t.Fatal(err)
+	}
+	return manager, assistant, provider
+}
+
+func TestManagerSwitchesModelAndThinking(t *testing.T) {
+	manager, assistant, provider := testManager(t)
 
 	if _, err := manager.SetModel("model-b"); err != nil {
 		t.Fatal(err)
@@ -80,7 +95,8 @@ func TestManagerRejectsUnknownModelAndThinking(t *testing.T) {
 		},
 	}
 	assistant := agent.New(agent.Options{Provider: &recordingProvider{}, Model: "model-a"})
-	manager, err := NewManager(cfg, assistant, nil)
+	store, _ := NewStore(filepath.Join(t.TempDir(), "sessions"))
+	manager, err := NewManager(cfg, assistant, nil, store)
 	if err != nil {
 		t.Fatal(err)
 	}
@@ -92,3 +108,226 @@ func TestManagerRejectsUnknownModelAndThinking(t *testing.T) {
 		t.Fatal("expected invalid thinking error")
 	}
 }
+
+func TestManagerSaveAndResume(t *testing.T) {
+	manager, assistant, _ := testManager(t)
+
+	// Start new session and have a conversation
+	manager.newSession()
+	var out strings.Builder
+	assistant.RunTurn(context.Background(), "hello world", &out, nil)
+	assistant.SetLastUsage(&llm.Usage{PromptTokens: 10, CompletionTokens: 5, TotalTokens: 15})
+
+	// Save
+	if err := manager.Save(); err != nil {
+		t.Fatalf("save: %v", err)
+	}
+	savedID := manager.CurrentSessionID()
+	if savedID == "" {
+		t.Fatal("expected session ID after save")
+	}
+	// Name should be auto-derived from first user message
+	if manager.CurrentSessionName() != "hello world" {
+		t.Fatalf("session name = %q, want %q", manager.CurrentSessionName(), "hello world")
+	}
+
+	// Start a new session
+	manager.newSession()
+	if manager.CurrentSessionID() == savedID {
+		t.Fatal("new session should have different ID")
+	}
+
+	// Resume
+	if err := manager.Resume(savedID); err != nil {
+		t.Fatalf("resume: %v", err)
+	}
+	if manager.CurrentSessionID() != savedID {
+		t.Fatalf("after resume, ID = %q, want %q", manager.CurrentSessionID(), savedID)
+	}
+	msgs := assistant.Messages()
+	hasUser := false
+	for _, m := range msgs {
+		if m.Role == llm.RoleUser && m.Content == "hello world" {
+			hasUser = true
+		}
+	}
+	if !hasUser {
+		t.Fatal("resumed session should contain original user message")
+	}
+	if usage := assistant.LastUsage(); usage == nil || usage.TotalTokens != 15 {
+		t.Fatalf("resumed usage = %#v", usage)
+	}
+}
+
+func TestManagerRename(t *testing.T) {
+	manager, _, _ := testManager(t)
+	manager.newSession()
+
+	msg, err := manager.Rename("my cool session")
+	if err != nil {
+		t.Fatalf("rename: %v", err)
+	}
+	if !strings.Contains(msg, "my cool session") {
+		t.Fatalf("rename message = %q", msg)
+	}
+	if manager.CurrentSessionName() != "my cool session" {
+		t.Fatalf("name = %q", manager.CurrentSessionName())
+	}
+
+	// Verify persisted
+	_, err = manager.store.Load(manager.CurrentSessionID())
+	if err != nil {
+		t.Fatalf("load after rename: %v", err)
+	}
+}
+
+func TestManagerRenameEmpty(t *testing.T) {
+	manager, _, _ := testManager(t)
+	manager.newSession()
+	_, err := manager.Rename("  ")
+	if err == nil {
+		t.Fatal("expected error for empty name")
+	}
+}
+
+func TestManagerSessions(t *testing.T) {
+	manager, _, _ := testManager(t)
+
+	// No sessions yet
+	out, err := manager.Sessions()
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !strings.Contains(out, "No saved sessions") {
+		t.Fatalf("expected 'no saved sessions', got: %s", out)
+	}
+
+	// Create and save a session
+	manager.newSession()
+	manager.Save()
+
+	out, err = manager.Sessions()
+	if err != nil {
+		t.Fatal(err)
+	}
+	if strings.Contains(out, "No saved sessions") {
+		t.Fatal("expected session list")
+	}
+}
+
+func TestManagerNewSession(t *testing.T) {
+	manager, assistant, _ := testManager(t)
+	manager.newSession()
+	var out strings.Builder
+	assistant.RunTurn(context.Background(), "first message", &out, nil)
+
+	msg, err := manager.NewSession()
+	if err != nil {
+		t.Fatalf("new session: %v", err)
+	}
+	if !strings.Contains(msg, "Starting new session") {
+		t.Fatalf("message = %q", msg)
+	}
+	// Old session should be saved
+	metas, _ := manager.store.List()
+	if len(metas) != 1 {
+		t.Fatalf("expected 1 saved session, got %d", len(metas))
+	}
+}
+
+func TestManagerResumeLatest(t *testing.T) {
+	manager, assistant, _ := testManager(t)
+
+	// Create two sessions
+	manager.newSession()
+	var out strings.Builder
+	assistant.RunTurn(context.Background(), "session one", &out, nil)
+	manager.Save()
+	id1 := manager.CurrentSessionID()
+
+	manager.NewSession()
+	assistant.RunTurn(context.Background(), "session two", &out, nil)
+	manager.Save()
+	id2 := manager.CurrentSessionID()
+
+	// Clear and resume latest
+	manager.newSession()
+	if err := manager.ResumeLatest(); err != nil {
+		t.Fatalf("resume latest: %v", err)
+	}
+	// Should resume the most recently updated (id2)
+	if manager.CurrentSessionID() != id2 {
+		t.Fatalf("expected to resume %s, got %s", id2, manager.CurrentSessionID())
+	}
+
+	// Also verify id1 is resumable
+	if err := manager.Resume(id1); err != nil {
+		t.Fatalf("resume id1: %v", err)
+	}
+	if manager.CurrentSessionID() != id1 {
+		t.Fatalf("expected %s, got %s", id1, manager.CurrentSessionID())
+	}
+}
+
+type compactingProvider struct {
+	summaryCalls int
+	answerCalls  int
+}
+
+func (p *compactingProvider) ChatStream(ctx context.Context, req llm.ChatRequest, emit func(llm.StreamEvent) error) error {
+	if len(req.Messages) == 2 && req.Messages[0].Role == llm.RoleSystem {
+		p.summaryCalls++
+		return emit(llm.StreamEvent{Content: "summary"})
+	}
+	p.answerCalls++
+	return emit(llm.StreamEvent{Content: "answer"})
+}
+
+func TestManagerUsesModelContextForAgentCompaction(t *testing.T) {
+	provider := &compactingProvider{}
+	cfg := &config.Config{Providers: map[string]config.ProviderConfig{
+		"one": {
+			Name:          "one",
+			BaseURL:       "https://one.example.com",
+			APIKey:        "sk-test",
+			Model:         "model-a",
+			Models:        []string{"model-a"},
+			ModelConfigs:  map[string]config.ModelConfig{"model-a": {ID: "model-a", Thinking: "none", ContextTokens: 1}},
+			ContextTokens: 1,
+		},
+	}}
+	assistant := agent.New(agent.Options{Provider: provider, ProviderName: "one", Model: "model-a"})
+	store, err := NewStore(filepath.Join(t.TempDir(), "sessions"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	manager, err := NewManager(cfg, assistant, nil, store)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if _, err := manager.SetModel("model-a"); err != nil {
+		t.Fatal(err)
+	}
+	assistant.SetMessages([]llm.Message{
+		{Role: llm.RoleSystem, Content: "system"},
+		{Role: llm.RoleUser, Content: "old question 1"},
+		{Role: llm.RoleAssistant, Content: "old answer 1"},
+		{Role: llm.RoleUser, Content: "old question 2"},
+		{Role: llm.RoleAssistant, Content: "old answer 2"},
+		{Role: llm.RoleUser, Content: "recent question 1"},
+		{Role: llm.RoleAssistant, Content: "recent answer 1"},
+		{Role: llm.RoleUser, Content: "recent question 2"},
+		{Role: llm.RoleAssistant, Content: "recent answer 2"},
+	})
+
+	var out strings.Builder
+	if err := assistant.RunTurn(context.Background(), "new question", &out, nil); err != nil {
+		t.Fatal(err)
+	}
+	if provider.summaryCalls != 1 {
+		t.Fatalf("summaryCalls = %d, want 1", provider.summaryCalls)
+	}
+	if provider.answerCalls != 1 {
+		t.Fatalf("answerCalls = %d, want 1", provider.answerCalls)
+	}
+}

internal/session/store.go

@@ -0,0 +1,174 @@
+package session
+
+import (
+	"crypto/rand"
+	"encoding/hex"
+	"encoding/json"
+	"fmt"
+	"os"
+	"path/filepath"
+	"sort"
+	"strings"
+	"time"
+
+	"agentu/pkg/llm"
+)
+
+// Session represents a persisted conversation.
+type Session struct {
+	ID        string        `json:"id"`
+	Name      string        `json:"name"`
+	CreatedAt time.Time     `json:"created_at"`
+	UpdatedAt time.Time     `json:"updated_at"`
+	Provider  string        `json:"provider"`
+	Model     string        `json:"model"`
+	Messages  []llm.Message `json:"messages"`
+	LastUsage *llm.Usage    `json:"last_usage,omitempty"`
+}
+
+// SessionMeta is a lightweight summary for listing.
+type SessionMeta struct {
+	ID           string    `json:"id"`
+	Name         string    `json:"name"`
+	CreatedAt    time.Time `json:"created_at"`
+	UpdatedAt    time.Time `json:"updated_at"`
+	Provider     string    `json:"provider"`
+	Model        string    `json:"model"`
+	MessageCount int       `json:"message_count"`
+}
+
+// Store manages session files under a directory.
+type Store struct {
+	dir string
+}
+
+// NewStore creates a store, ensuring the directory exists.
+func NewStore(dir string) (*Store, error) {
+	if err := os.MkdirAll(dir, 0o755); err != nil {
+		return nil, fmt.Errorf("create session dir: %w", err)
+	}
+	return &Store{dir: dir}, nil
+}
+
+// Save writes a session to disk.
+func (s *Store) Save(sess *Session) error {
+	if sess.ID == "" {
+		return fmt.Errorf("session ID is required")
+	}
+	sess.UpdatedAt = time.Now().UTC()
+	if sess.CreatedAt.IsZero() {
+		sess.CreatedAt = sess.UpdatedAt
+	}
+	data, err := json.MarshalIndent(sess, "", "  ")
+	if err != nil {
+		return fmt.Errorf("marshal session: %w", err)
+	}
+	path := s.filePath(sess.ID)
+	if err := os.WriteFile(path, data, 0o644); err != nil {
+		return fmt.Errorf("write session: %w", err)
+	}
+	return nil
+}
+
+// Load reads a session by ID.
+func (s *Store) Load(id string) (*Session, error) {
+	path := s.filePath(id)
+	data, err := os.ReadFile(path)
+	if err != nil {
+		if os.IsNotExist(err) {
+			return nil, fmt.Errorf("session %q not found", id)
+		}
+		return nil, fmt.Errorf("read session: %w", err)
+	}
+	var sess Session
+	if err := json.Unmarshal(data, &sess); err != nil {
+		return nil, fmt.Errorf("parse session: %w", err)
+	}
+	return &sess, nil
+}
+
+// List returns metadata for all sessions, sorted by updated_at descending.
+func (s *Store) List() ([]SessionMeta, error) {
+	entries, err := os.ReadDir(s.dir)
+	if err != nil {
+		return nil, fmt.Errorf("read session dir: %w", err)
+	}
+	var metas []SessionMeta
+	for _, entry := range entries {
+		if entry.IsDir() || !strings.HasSuffix(entry.Name(), ".json") {
+			continue
+		}
+		path := filepath.Join(s.dir, entry.Name())
+		data, err := os.ReadFile(path)
+		if err != nil {
+			continue
+		}
+		var sess Session
+		if err := json.Unmarshal(data, &sess); err != nil {
+			continue
+		}
+		metas = append(metas, SessionMeta{
+			ID:           sess.ID,
+			Name:         sess.Name,
+			CreatedAt:    sess.CreatedAt,
+			UpdatedAt:    sess.UpdatedAt,
+			Provider:     sess.Provider,
+			Model:        sess.Model,
+			MessageCount: len(sess.Messages),
+		})
+	}
+	sort.Slice(metas, func(i, j int) bool {
+		return metas[i].UpdatedAt.After(metas[j].UpdatedAt)
+	})
+	return metas, nil
+}
+
+// Delete removes a session file.
+func (s *Store) Delete(id string) error {
+	path := s.filePath(id)
+	if err := os.Remove(path); err != nil {
+		if os.IsNotExist(err) {
+			return fmt.Errorf("session %q not found", id)
+		}
+		return fmt.Errorf("delete session: %w", err)
+	}
+	return nil
+}
+
+func (s *Store) filePath(id string) string {
+	return filepath.Join(s.dir, id+".json")
+}
+
+// GenerateID creates a random 8-character hex session ID.
+func GenerateID() string {
+	b := make([]byte, 4)
+	_, _ = rand.Read(b)
+	return hex.EncodeToString(b)
+}
+
+// DefaultName derives a session name from the first user message.
+func DefaultName(messages []llm.Message) string {
+	for _, msg := range messages {
+		if msg.Role != llm.RoleUser {
+			continue
+		}
+		text := strings.TrimSpace(msg.Content)
+		if text == "" {
+			continue
+		}
+		// Take first line, truncate at 40 chars at word boundary
+		if idx := strings.IndexByte(text, '\n'); idx >= 0 {
+			text = text[:idx]
+		}
+		text = strings.TrimSpace(text)
+		if len(text) > 40 {
+			text = text[:40]
+			if idx := strings.LastIndexByte(text, ' '); idx > 20 {
+				text = text[:idx]
+			}
+			text += "..."
+		}
+		return text
+	}
+	return ""
+}

internal/session/store_test.go

@@ -0,0 +1,174 @@
+package session
+
+import (
+	"os"
+	"path/filepath"
+	"testing"
+
+	"agentu/pkg/llm"
+)
+
+func tempStore(t *testing.T) *Store {
+	t.Helper()
+	dir := t.TempDir()
+	store, err := NewStore(dir)
+	if err != nil {
+		t.Fatal(err)
+	}
+	return store
+}
+
+func TestSaveAndLoad(t *testing.T) {
+	store := tempStore(t)
+	sess := &Session{
+		ID:       GenerateID(),
+		Name:     "test session",
+		Provider: "loveuer",
+		Model:    "deepseek-v4-flash",
+		Messages: []llm.Message{
+			{Role: llm.RoleSystem, Content: "you are helpful"},
+			{Role: llm.RoleUser, Content: "hello"},
+			{Role: llm.RoleAssistant, Content: "hi there"},
+		},
+		LastUsage: &llm.Usage{PromptTokens: 10, CompletionTokens: 5, TotalTokens: 15},
+	}
+	if err := store.Save(sess); err != nil {
+		t.Fatalf("save: %v", err)
+	}
+
+	loaded, err := store.Load(sess.ID)
+	if err != nil {
+		t.Fatalf("load: %v", err)
+	}
+	if loaded.ID != sess.ID {
+		t.Errorf("ID = %q, want %q", loaded.ID, sess.ID)
+	}
+	if loaded.Name != sess.Name {
+		t.Errorf("Name = %q, want %q", loaded.Name, sess.Name)
+	}
+	if len(loaded.Messages) != 3 {
+		t.Errorf("Messages len = %d, want 3", len(loaded.Messages))
+	}
+	if loaded.Messages[1].Content != "hello" {
+		t.Errorf("Messages[1].Content = %q, want %q", loaded.Messages[1].Content, "hello")
+	}
+	if loaded.LastUsage == nil || loaded.LastUsage.TotalTokens != 15 {
+		t.Errorf("LastUsage = %#v", loaded.LastUsage)
+	}
+	if loaded.UpdatedAt.IsZero() {
+		t.Error("UpdatedAt should be set")
+	}
+}
+
+func TestLoadNotFound(t *testing.T) {
+	store := tempStore(t)
+	_, err := store.Load("nonexistent")
+	if err == nil {
+		t.Fatal("expected error for nonexistent session")
+	}
+}
+
+func TestList(t *testing.T) {
+	store := tempStore(t)
+	// Empty list
+	metas, err := store.List()
+	if err != nil {
+		t.Fatalf("list: %v", err)
+	}
+	if len(metas) != 0 {
+		t.Errorf("expected empty list, got %d", len(metas))
+	}
+
+	// Save two sessions
+	store.Save(&Session{ID: "aaa", Name: "first", Messages: []llm.Message{{Role: "user", Content: "a"}}})
+	store.Save(&Session{ID: "bbb", Name: "second", Messages: []llm.Message{{Role: "user", Content: "b"}, {Role: "assistant", Content: "c"}}})
+
+	metas, err = store.List()
+	if err != nil {
+		t.Fatalf("list: %v", err)
+	}
+	if len(metas) != 2 {
+		t.Fatalf("expected 2 sessions, got %d", len(metas))
+	}
+	// Sorted by updated_at desc
+	if metas[0].ID != "bbb" && metas[0].ID != "aaa" {
+		t.Errorf("unexpected first session ID: %q", metas[0].ID)
+	}
+}
+
+func TestDelete(t *testing.T) {
+	store := tempStore(t)
+	store.Save(&Session{ID: "todelete", Name: "temp"})
+
+	if err := store.Delete("todelete"); err != nil {
+		t.Fatalf("delete: %v", err)
+	}
+	_, err := store.Load("todelete")
+	if err == nil {
+		t.Fatal("expected error after delete")
+	}
+}
+
+func TestDeleteNotFound(t *testing.T) {
+	store := tempStore(t)
+	err := store.Delete("nonexistent")
+	if err == nil {
+		t.Fatal("expected error for nonexistent delete")
+	}
+}
+
+func TestSaveRequiresID(t *testing.T) {
+	store := tempStore(t)
+	err := store.Save(&Session{Name: "no id"})
+	if err == nil {
+		t.Fatal("expected error for empty ID")
+	}
+}
+
+func TestGenerateID(t *testing.T) {
+	id := GenerateID()
+	if len(id) != 8 {
+		t.Errorf("ID length = %d, want 8", len(id))
+	}
+	id2 := GenerateID()
+	if id == id2 {
+		t.Error("expected unique IDs")
+	}
+}
+
+func TestDefaultName(t *testing.T) {
+	tests := []struct {
+		name     string
+		messages []llm.Message
+		want     string
+	}{
+		{"empty", nil, ""},
+		{"no user msg", []llm.Message{{Role: "system", Content: "hi"}}, ""},
+		{"short", []llm.Message{{Role: "user", Content: "hello world"}}, "hello world"},
+		{"long", []llm.Message{{Role: "user", Content: "this is a really long message that should be truncated at some point in the middle"}}, "this is a really long message that..."},
+		{"multiline", []llm.Message{{Role: "user", Content: "first line\nsecond line"}}, "first line"},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			got := DefaultName(tt.messages)
+			if got != tt.want {
+				t.Errorf("DefaultName = %q, want %q", got, tt.want)
+			}
+		})
+	}
+}
+
+func TestNewStoreCreatesDir(t *testing.T) {
+	dir := filepath.Join(t.TempDir(), "nested", "sessions")
+	store, err := NewStore(dir)
+	if err != nil {
+		t.Fatalf("NewStore: %v", err)
+	}
+	info, err := os.Stat(store.dir)
+	if err != nil {
+		t.Fatalf("stat: %v", err)
+	}
+	if !info.IsDir() {
+		t.Error("expected directory")
+	}
+}

internal/tools/code.go

@@ -0,0 +1,219 @@
+package tools
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"go/ast"
+	"go/parser"
+	"go/printer"
+	"go/token"
+	"io/fs"
+	"os"
+	"path/filepath"
+	"sort"
+	"strings"
+
+	"agentu/pkg/llm"
+)
+
+const defaultCodeSymbolFileLimit = 50
+
+type CodeSymbolsTool struct {
+	workingDir string
+}
+
+func NewCodeSymbolsTool(workingDir string) *CodeSymbolsTool {
+	return &CodeSymbolsTool{workingDir: workingDir}
+}
+
+func (t *CodeSymbolsTool) Definition() llm.Tool {
+	return llm.Tool{
+		Type: "function",
+		Function: llm.ToolFunction{
+			Name:        "code_symbols",
+			Description: "List Go code symbols (packages, imports, types, functions, methods) for a file or directory without reading full file contents.",
+			Parameters: JSONSchema(`{
+  "type": "object",
+  "properties": {
+    "path": {"type": "string", "description": "Go file or directory to inspect. Defaults to the configured agent working directory."},
+    "max_files": {"type": "integer", "description": "Maximum number of Go files to inspect when path is a directory. Defaults to 50."}
+  },
+  "additionalProperties": false
+}`),
+		},
+	}
+}
+
+type codeSymbolsArgs struct {
+	Path     string `json:"path"`
+	MaxFiles int    `json:"max_files"`
+}
+
+func (t *CodeSymbolsTool) Execute(ctx context.Context, raw json.RawMessage) (string, error) {
+	var args codeSymbolsArgs
+	if err := json.Unmarshal(raw, &args); err != nil {
+		return "", fmt.Errorf("parse arguments: %w", err)
+	}
+	path, err := resolvePath(t.workingDir, args.Path)
+	if err != nil {
+		return "", err
+	}
+	limit := args.MaxFiles
+	if limit <= 0 {
+		limit = defaultCodeSymbolFileLimit
+	}
+
+	files, err := goFiles(ctx, path, limit)
+	if err != nil {
+		return "", err
+	}
+	if len(files) == 0 {
+		return "no Go files found", nil
+	}
+
+	fset := token.NewFileSet()
+	var out strings.Builder
+	for i, file := range files {
+		if ctxErr := ctx.Err(); ctxErr != nil {
+			return FormatResult(out.String()), ctxErr
+		}
+		if i > 0 {
+			out.WriteString("\n")
+		}
+		if err := appendGoSymbols(&out, fset, file); err != nil {
+			fmt.Fprintf(&out, "%s\nerror: %v\n", file, err)
+		}
+		if out.Len() > MaxToolOutputBytes {
+			break
+		}
+	}
+	return FormatResult(out.String()), nil
+}
+
+func goFiles(ctx context.Context, path string, limit int) ([]string, error) {
+	info, err := os.Stat(path)
+	if err != nil {
+		return nil, err
+	}
+	if !info.IsDir() {
+		if strings.HasSuffix(path, ".go") {
+			return []string{path}, nil
+		}
+		return nil, fmt.Errorf("path is not a Go file: %s", path)
+	}
+
+	var files []string
+	stop := errors.New("file limit reached")
+	err = filepath.WalkDir(path, func(current string, d fs.DirEntry, err error) error {
+		if err != nil {
+			return nil
+		}
+		if ctxErr := ctx.Err(); ctxErr != nil {
+			return ctxErr
+		}
+		if d.IsDir() {
+			switch d.Name() {
+			case ".git", "vendor", "node_modules":
+				return filepath.SkipDir
+			}
+			return nil
+		}
+		if strings.HasSuffix(current, ".go") {
+			files = append(files, current)
+			if len(files) >= limit {
+				return stop
+			}
+		}
+		return nil
+	})
+	if err != nil && !errors.Is(err, stop) {
+		return nil, err
+	}
+	sort.Strings(files)
+	return files, nil
+}
+
+func appendGoSymbols(out *strings.Builder, fset *token.FileSet, path string) error {
+	file, err := parser.ParseFile(fset, path, nil, parser.ImportsOnly|parser.ParseComments)
+	if err != nil {
+		return err
+	}
+	full, err := parser.ParseFile(fset, path, nil, 0)
+	if err != nil {
+		return err
+	}
+
+	fmt.Fprintf(out, "%s\n", path)
+	fmt.Fprintf(out, "  package %s\n", full.Name.Name)
+	if len(file.Imports) > 0 {
+		imports := make([]string, 0, len(file.Imports))
+		for _, imp := range file.Imports {
+			imports = append(imports, strings.Trim(imp.Path.Value, "\""))
+		}
+		fmt.Fprintf(out, "  imports: %s\n", strings.Join(imports, ", "))
+	}
+
+	var types, funcs, methods []string
+	for _, decl := range full.Decls {
+		switch d := decl.(type) {
+		case *ast.GenDecl:
+			if d.Tok != token.TYPE {
+				continue
+			}
+			for _, spec := range d.Specs {
+				if ts, ok := spec.(*ast.TypeSpec); ok {
+					kind := typeKind(ts.Type)
+					types = append(types, fmt.Sprintf("%d: type %s %s", fset.Position(ts.Pos()).Line, ts.Name.Name, kind))
+				}
+			}
+		case *ast.FuncDecl:
+			line := fset.Position(d.Pos()).Line
+			if d.Recv == nil {
+				funcs = append(funcs, fmt.Sprintf("%d: func %s", line, d.Name.Name))
+				continue
+			}
+			methods = append(methods, fmt.Sprintf("%d: method %s.%s", line, receiverName(d.Recv), d.Name.Name))
+		}
+	}
+	appendSymbolGroup(out, "types", types)
+	appendSymbolGroup(out, "funcs", funcs)
+	appendSymbolGroup(out, "methods", methods)
+	return nil
+}
+
+func appendSymbolGroup(out *strings.Builder, title string, values []string) {
+	if len(values) == 0 {
+		return
+	}
+	fmt.Fprintf(out, "  %s:\n", title)
+	for _, value := range values {
+		fmt.Fprintf(out, "    - %s\n", value)
+	}
+}
+
+func typeKind(expr ast.Expr) string {
+	switch expr.(type) {
+	case *ast.StructType:
+		return "struct"
+	case *ast.InterfaceType:
+		return "interface"
+	case *ast.FuncType:
+		return "func"
+	default:
+		return "alias"
+	}
+}
+
+func receiverName(recv *ast.FieldList) string {
+	if recv == nil || len(recv.List) == 0 {
+		return "?"
+	}
+	var b bytes.Buffer
+	if err := printer.Fprint(&b, token.NewFileSet(), recv.List[0].Type); err != nil {
+		return "?"
+	}
+	return b.String()
+}

internal/tools/file.go

@@ -13,7 +13,7 @@ import (
 	"sort"
 	"strings"
 
-	"agentu/internal/llm"
+	"agentu/pkg/llm"
 )
 
 type FileReadTool struct {
@@ -81,6 +81,7 @@ type FileWriteTool struct {
 func NewFileWriteTool(workingDir string) *FileWriteTool {
 	return &FileWriteTool{workingDir: workingDir}
 }
+func (*FileWriteTool) Mutates() bool { return true }
 
 func (t *FileWriteTool) Definition() llm.Tool {
 	return llm.Tool{
@@ -223,7 +224,9 @@ func (t *FileSearchTool) Definition() llm.Tool {
   "properties": {
     "query": {"type": "string", "description": "Text or regular expression to search for."},
     "path": {"type": "string", "description": "Optional directory or file to search. Defaults to the configured agent working directory."},
-    "glob": {"type": "string", "description": "Optional file glob, for example '*.go'."}
+    "glob": {"type": "string", "description": "Optional file glob, for example '*.go'."},
+    "context_lines": {"type": "integer", "description": "Optional number of context lines around each match. Defaults to 0."},
+    "max_results": {"type": "integer", "description": "Optional maximum matches to return. Defaults to 100."}
   },
   "required": ["query"],
   "additionalProperties": false
@@ -236,6 +239,8 @@ type fileSearchArgs struct {
 	Query        string `json:"query"`
 	Path         string `json:"path"`
 	Glob         string `json:"glob"`
+	ContextLines int    `json:"context_lines"`
+	MaxResults   int    `json:"max_results"`
 }
 
 func (t *FileSearchTool) Execute(ctx context.Context, raw json.RawMessage) (string, error) {
@@ -256,8 +261,31 @@ func (t *FileSearchTool) Execute(ctx context.Context, raw json.RawMessage) (stri
 	return t.walkSearch(ctx, args, path)
 }
 
+func normalizedSearchOptions(args fileSearchArgs) (contextLines int, maxResults int) {
+	contextLines = args.ContextLines
+	if contextLines < 0 {
+		contextLines = 0
+	}
+	if contextLines > 5 {
+		contextLines = 5
+	}
+	maxResults = args.MaxResults
+	if maxResults <= 0 {
+		maxResults = 100
+	}
+	if maxResults > 1000 {
+		maxResults = 1000
+	}
+	return contextLines, maxResults
+}
+
 func (t *FileSearchTool) ripgrep(ctx context.Context, args fileSearchArgs, path string) (string, error) {
 	rgArgs := []string{"--line-number", "--color", "never"}
+	contextLines, maxResults := normalizedSearchOptions(args)
+	if contextLines > 0 {
+		rgArgs = append(rgArgs, "--context", fmt.Sprint(contextLines))
+	}
+	rgArgs = append(rgArgs, "--max-count", fmt.Sprint(maxResults))
 	if args.Glob != "" {
 		rgArgs = append(rgArgs, "--glob", args.Glob)
 	}
@@ -282,7 +310,10 @@ func (t *FileSearchTool) ripgrep(ctx context.Context, args fileSearchArgs, path
 }
 
 func (t *FileSearchTool) walkSearch(ctx context.Context, args fileSearchArgs, root string) (string, error) {
+	contextLines, maxResults := normalizedSearchOptions(args)
 	var out bytes.Buffer
+	matchCount := 0
+	stop := errors.New("result limit reached")
 	err := filepath.WalkDir(root, func(path string, d fs.DirEntry, err error) error {
 		if err != nil {
 			return nil
@@ -312,16 +343,22 @@ func (t *FileSearchTool) walkSearch(ctx context.Context, args fileSearchArgs, ro
 		}
 		lines := strings.Split(string(data), "\n")
 		for i, line := range lines {
-			if strings.Contains(line, args.Query) {
-				fmt.Fprintf(&out, "%s:%d:%s\n", path, i+1, line)
-				if out.Len() > MaxToolOutputBytes {
-					return errors.New("output limit reached")
+			if !strings.Contains(line, args.Query) {
+				continue
 			}
+			matchCount++
+			appendSearchMatch(&out, path, lines, i, contextLines)
+			if matchCount >= maxResults {
+				fmt.Fprintf(&out, "[stopped after %d matches; narrow path/query or raise max_results]\n", maxResults)
+				return stop
+			}
+			if out.Len() > MaxToolOutputBytes {
+				return stop
 			}
 		}
 		return nil
 	})
-	if err != nil && err.Error() != "output limit reached" {
+	if err != nil && !errors.Is(err, stop) {
 		return "", err
 	}
 	if out.Len() == 0 {
@@ -329,3 +366,18 @@ func (t *FileSearchTool) walkSearch(ctx context.Context, args fileSearchArgs, ro
 	}
 	return FormatResult(out.String()), nil
 }
+
+func appendSearchMatch(out *bytes.Buffer, path string, lines []string, matchIndex int, contextLines int) {
+	start := max(0, matchIndex-contextLines)
+	end := min(len(lines)-1, matchIndex+contextLines)
+	if contextLines > 0 {
+		fmt.Fprintf(out, "-- %s:%d --\n", path, matchIndex+1)
+	}
+	for i := start; i <= end; i++ {
+		marker := ":"
+		if i == matchIndex {
+			marker = "*"
+		}
+		fmt.Fprintf(out, "%s%s%d:%s\n", path, marker, i+1, lines[i])
+	}
+}

internal/tools/file_edit.go

@@ -0,0 +1,178 @@
+package tools
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"os"
+	"strings"
+
+	"agentu/pkg/llm"
+)
+
+type FileEditTool struct {
+	workingDir string
+}
+
+func NewFileEditTool(workingDir string) *FileEditTool {
+	return &FileEditTool{workingDir: workingDir}
+}
+
+func (*FileEditTool) Mutates() bool { return true }
+
+func (t *FileEditTool) Definition() llm.Tool {
+	return llm.Tool{
+		Type: "function",
+		Function: llm.ToolFunction{
+			Name:        "file_edit",
+			Description: "Edit an existing UTF-8 text file by replacing an inclusive 1-based line range or an exact old text pattern.",
+			Parameters: JSONSchema(`{
+  "type": "object",
+  "properties": {
+    "path": {"type": "string", "description": "Existing file path relative to the configured agent working directory."},
+    "start_line": {"type": "integer", "description": "1-based first line for a line-range edit."},
+    "end_line": {"type": "integer", "description": "Optional 1-based final line; defaults to start_line when omitted."},
+    "old": {"type": "string", "description": "Exact text to replace for a pattern edit; not regex."},
+    "new": {"type": "string", "description": "Replacement text; may be empty string for deletion."},
+    "replace_all": {"type": "boolean", "description": "When true, replace every exact old occurrence; otherwise require exactly one occurrence."}
+  },
+  "required": ["path", "new"],
+  "additionalProperties": false
+}`),
+		},
+	}
+}
+
+type fileEditArgs struct {
+	Path       string  `json:"path"`
+	StartLine  int     `json:"start_line"`
+	EndLine    int     `json:"end_line"`
+	Old        string  `json:"old"`
+	New        *string `json:"new"`
+	ReplaceAll bool    `json:"replace_all"`
+}
+
+func (t *FileEditTool) Execute(ctx context.Context, raw json.RawMessage) (string, error) {
+	var args fileEditArgs
+	if err := json.Unmarshal(raw, &args); err != nil {
+		return "", fmt.Errorf("parse arguments: %w", err)
+	}
+	if strings.TrimSpace(args.Path) == "" {
+		return "", errors.New("path is required")
+	}
+	if args.New == nil {
+		return "", errors.New("new is required")
+	}
+	path, err := resolvePath(t.workingDir, args.Path)
+	if err != nil {
+		return "", err
+	}
+
+	info, err := os.Stat(path)
+	if err != nil {
+		return "", fmt.Errorf("cannot access %s: %w", args.Path, err)
+	}
+	if info.IsDir() {
+		return "", fmt.Errorf("%s is a directory, not a file", args.Path)
+	}
+	perm := info.Mode().Perm()
+
+	hasLine := args.StartLine > 0 || args.EndLine > 0
+	hasOld := strings.TrimSpace(args.Old) != ""
+
+	if hasLine && hasOld {
+		return "", errors.New("provide either start_line/end_line or old, not both")
+	}
+	if !hasLine && !hasOld {
+		return "", errors.New("provide either start_line/end_line or old")
+	}
+
+	if err := ctx.Err(); err != nil {
+		return "", err
+	}
+
+	data, err := os.ReadFile(path)
+	if err != nil {
+		return "", err
+	}
+	content := string(data)
+
+	var result string
+	if hasOld {
+		result, err = editByPattern(content, args.Path, args.Old, *args.New, args.ReplaceAll)
+	} else {
+		result, err = editByLineRange(content, args.Path, args.StartLine, args.EndLine, *args.New)
+	}
+	if err != nil {
+		return "", err
+	}
+
+	if err := ctx.Err(); err != nil {
+		return "", err
+	}
+	if err := os.WriteFile(path, []byte(result), perm); err != nil {
+		return "", err
+	}
+
+	switch {
+	case hasOld:
+		count := strings.Count(content, args.Old)
+		if args.ReplaceAll {
+			return fmt.Sprintf("edited %s: replaced %d occurrences", args.Path, count), nil
+		}
+		return fmt.Sprintf("edited %s: replaced 1 occurrence", args.Path), nil
+	default:
+		endLine := args.EndLine
+		if endLine == 0 {
+			endLine = args.StartLine
+		}
+		return fmt.Sprintf("edited %s: replaced lines %d-%d", args.Path, args.StartLine, endLine), nil
+	}
+}
+
+func editByLineRange(content, displayPath string, startLine, endLine int, replacement string) (string, error) {
+	if startLine < 1 {
+		return "", errors.New("start_line must be >= 1")
+	}
+	if endLine == 0 {
+		endLine = startLine
+	}
+	if endLine < startLine {
+		return "", fmt.Errorf("end_line (%d) must be >= start_line (%d)", endLine, startLine)
+	}
+
+	lines := strings.Split(content, "\n")
+	if startLine > len(lines) {
+		return "", fmt.Errorf("line range %d-%d is outside file with %d lines", startLine, endLine, len(lines))
+	}
+	if endLine > len(lines) {
+		return "", fmt.Errorf("line range %d-%d is outside file with %d lines", startLine, endLine, len(lines))
+	}
+
+	var b strings.Builder
+	b.WriteString(strings.Join(lines[:startLine-1], "\n"))
+	if startLine > 1 {
+		b.WriteString("\n")
+	}
+	b.WriteString(replacement)
+	if endLine < len(lines) {
+		if !strings.HasSuffix(replacement, "\n") && replacement != "" {
+			b.WriteString("\n")
+		}
+		b.WriteString(strings.Join(lines[endLine:], "\n"))
+	}
+
+	return b.String(), nil
+}
+
+func editByPattern(content, displayPath, old, replacement string, replaceAll bool) (string, error) {
+	count := strings.Count(content, old)
+	if count == 0 {
+		return "", errors.New("old text not found")
+	}
+	if !replaceAll && count != 1 {
+		return "", fmt.Errorf("old text matched %d occurrences; set replace_all=true or provide a more specific old value", count)
+	}
+	return strings.ReplaceAll(content, old, replacement), nil
+}

internal/tools/file_edit_test.go

@@ -0,0 +1,136 @@
+package tools
+
+import (
+	"context"
+	"encoding/json"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+)
+
+func TestFileEditReplacesLineRange(t *testing.T) {
+	dir := t.TempDir()
+	path := filepath.Join(dir, "notes.txt")
+	if err := os.WriteFile(path, []byte("one\ntwo\nthree\n"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+
+	edit := NewFileEditTool(dir)
+	out, err := edit.Execute(context.Background(), json.RawMessage(`{"path":"notes.txt","start_line":2,"end_line":2,"new":"TWO\n"}`))
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !strings.Contains(out, "replaced lines 2-2") {
+		t.Fatalf("output = %q", out)
+	}
+	got, err := os.ReadFile(path)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if string(got) != "one\nTWO\nthree\n" {
+		t.Fatalf("file content = %q", string(got))
+	}
+}
+
+func TestFileEditReplacesExactPattern(t *testing.T) {
+	dir := t.TempDir()
+	path := filepath.Join(dir, "notes.txt")
+	if err := os.WriteFile(path, []byte("alpha beta alpha"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+
+	edit := NewFileEditTool(dir)
+	out, err := edit.Execute(context.Background(), json.RawMessage(`{"path":"notes.txt","old":"beta","new":"BETA"}`))
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !strings.Contains(out, "replaced 1 occurrence") {
+		t.Fatalf("output = %q", out)
+	}
+	got, err := os.ReadFile(path)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if string(got) != "alpha BETA alpha" {
+		t.Fatalf("file content = %q", string(got))
+	}
+}
+
+func TestFileEditRejectsAmbiguousPattern(t *testing.T) {
+	dir := t.TempDir()
+	path := filepath.Join(dir, "notes.txt")
+	if err := os.WriteFile(path, []byte("x x"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+
+	edit := NewFileEditTool(dir)
+	_, err := edit.Execute(context.Background(), json.RawMessage(`{"path":"notes.txt","old":"x","new":"y"}`))
+	if err == nil {
+		t.Fatal("expected error for ambiguous pattern")
+	}
+	if !strings.Contains(err.Error(), "matched 2 occurrences") {
+		t.Fatalf("error = %v", err)
+	}
+}
+
+func TestFileEditReplaceAll(t *testing.T) {
+	dir := t.TempDir()
+	path := filepath.Join(dir, "notes.txt")
+	if err := os.WriteFile(path, []byte("x x"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+
+	edit := NewFileEditTool(dir)
+	out, err := edit.Execute(context.Background(), json.RawMessage(`{"path":"notes.txt","old":"x","new":"y","replace_all":true}`))
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !strings.Contains(out, "replaced 2 occurrences") {
+		t.Fatalf("output = %q", out)
+	}
+	got, err := os.ReadFile(path)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if string(got) != "y y" {
+		t.Fatalf("file content = %q", string(got))
+	}
+}
+
+func TestFileEditRejectsEscapingPath(t *testing.T) {
+	dir := t.TempDir()
+	edit := NewFileEditTool(dir)
+	_, err := edit.Execute(context.Background(), json.RawMessage(`{"path":"../outside.txt","old":"x","new":"y"}`))
+	if err == nil {
+		t.Fatal("expected error for escaping path")
+	}
+	if !strings.Contains(err.Error(), "escapes") {
+		t.Fatalf("error = %v", err)
+	}
+}
+
+func TestFileEditRejectsMissingNew(t *testing.T) {
+	dir := t.TempDir()
+	path := filepath.Join(dir, "notes.txt")
+	if err := os.WriteFile(path, []byte("hello"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	edit := NewFileEditTool(dir)
+	_, err := edit.Execute(context.Background(), json.RawMessage(`{"path":"notes.txt","old":"hello"}`))
+	if err == nil {
+		t.Fatal("expected error for missing new")
+	}
+	if !strings.Contains(err.Error(), "new is required") {
+		t.Fatalf("error = %v", err)
+	}
+}
+
+func TestFileEditRejectsNonExistentFile(t *testing.T) {
+	dir := t.TempDir()
+	edit := NewFileEditTool(dir)
+	_, err := edit.Execute(context.Background(), json.RawMessage(`{"path":"missing.txt","old":"x","new":"y"}`))
+	if err == nil {
+		t.Fatal("expected error for non-existent file")
+	}
+}

internal/tools/git.go

@@ -0,0 +1,246 @@
+package tools
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"os/exec"
+	"path/filepath"
+	"strings"
+
+	"agentu/pkg/llm"
+)
+
+const (
+	defaultGitLogMaxCount = 20
+	maxGitLogMaxCount     = 100
+)
+
+func runGit(ctx context.Context, workingDir string, args ...string) (string, error) {
+	gitArgs := append([]string{"-C", workingDir}, args...)
+	cmd := exec.CommandContext(ctx, "git", gitArgs...)
+	output, err := cmd.CombinedOutput()
+	text := FormatResult(string(output))
+	if err != nil {
+		if strings.Contains(err.Error(), "context deadline exceeded") || strings.Contains(err.Error(), "signal: killed") {
+			if text != "" {
+				return text, fmt.Errorf("git command timed out: %w", err)
+			}
+			return "", fmt.Errorf("git command timed out: %w", err)
+		}
+		if text != "" {
+			return text, fmt.Errorf("git command failed: %w", err)
+		}
+		return "", fmt.Errorf("git command failed: %w", err)
+	}
+	return text, nil
+}
+
+func gitPathspec(baseDir, path string) (string, error) {
+	if strings.TrimSpace(path) == "" {
+		return "", nil
+	}
+	resolved, err := resolvePath(baseDir, path)
+	if err != nil {
+		return "", err
+	}
+	rel, err := filepath.Rel(baseDir, resolved)
+	if err != nil {
+		return "", err
+	}
+	return filepath.ToSlash(filepath.Clean(rel)), nil
+}
+
+func normalizeGitLogMaxCount(value int) int {
+	if value <= 0 {
+		return defaultGitLogMaxCount
+	}
+	if value > maxGitLogMaxCount {
+		return maxGitLogMaxCount
+	}
+	return value
+}
+
+// git_status
+
+type GitStatusTool struct {
+	workingDir string
+}
+
+func NewGitStatusTool(workingDir string) *GitStatusTool {
+	return &GitStatusTool{workingDir: workingDir}
+}
+
+func (t *GitStatusTool) Definition() llm.Tool {
+	return llm.Tool{
+		Type: "function",
+		Function: llm.ToolFunction{
+			Name:        "git_status",
+			Description: "Show changed files using git status --short.",
+			Parameters: JSONSchema(`{
+  "type": "object",
+  "properties": {
+    "path": {"type": "string", "description": "Optional path to limit status to. Relative paths are resolved from the configured agent working directory."}
+  },
+  "additionalProperties": false
+}`),
+		},
+	}
+}
+
+type gitStatusArgs struct {
+	Path string `json:"path"`
+}
+
+func (t *GitStatusTool) Execute(ctx context.Context, raw json.RawMessage) (string, error) {
+	var args gitStatusArgs
+	if len(raw) > 0 {
+		if err := json.Unmarshal(raw, &args); err != nil {
+			return "", fmt.Errorf("parse arguments: %w", err)
+		}
+	}
+	gitArgs := []string{"status", "--short"}
+	if args.Path != "" {
+		pathspec, err := gitPathspec(t.workingDir, args.Path)
+		if err != nil {
+			return "", err
+		}
+		gitArgs = append(gitArgs, "--", pathspec)
+	}
+	text, err := runGit(ctx, t.workingDir, gitArgs...)
+	if err != nil {
+		return text, err
+	}
+	if strings.TrimSpace(text) == "" {
+		return "Working tree clean.", nil
+	}
+	return "Changed files:\n" + text, nil
+}
+
+// git_diff
+
+type GitDiffTool struct {
+	workingDir string
+}
+
+func NewGitDiffTool(workingDir string) *GitDiffTool {
+	return &GitDiffTool{workingDir: workingDir}
+}
+
+func (t *GitDiffTool) Definition() llm.Tool {
+	return llm.Tool{
+		Type: "function",
+		Function: llm.ToolFunction{
+			Name:        "git_diff",
+			Description: "Show unstaged or staged git diff output, optionally as --stat.",
+			Parameters: JSONSchema(`{
+  "type": "object",
+  "properties": {
+    "path": {"type": "string", "description": "Optional path to limit diff to. Relative paths are resolved from the configured agent working directory."},
+    "staged": {"type": "boolean", "description": "Show staged diff instead of unstaged."},
+    "stat": {"type": "boolean", "description": "Show diffstat summary instead of full diff."}
+  },
+  "additionalProperties": false
+}`),
+		},
+	}
+}
+
+type gitDiffArgs struct {
+	Path   string `json:"path"`
+	Staged bool   `json:"staged"`
+	Stat   bool   `json:"stat"`
+}
+
+func (t *GitDiffTool) Execute(ctx context.Context, raw json.RawMessage) (string, error) {
+	var args gitDiffArgs
+	if len(raw) > 0 {
+		if err := json.Unmarshal(raw, &args); err != nil {
+			return "", fmt.Errorf("parse arguments: %w", err)
+		}
+	}
+	gitArgs := []string{"diff"}
+	if args.Staged {
+		gitArgs = append(gitArgs, "--cached")
+	}
+	if args.Stat {
+		gitArgs = append(gitArgs, "--stat")
+	}
+	if args.Path != "" {
+		pathspec, err := gitPathspec(t.workingDir, args.Path)
+		if err != nil {
+			return "", err
+		}
+		gitArgs = append(gitArgs, "--", pathspec)
+	}
+	text, err := runGit(ctx, t.workingDir, gitArgs...)
+	if err != nil {
+		return text, err
+	}
+	if strings.TrimSpace(text) == "" {
+		if args.Staged {
+			return "No staged diff.", nil
+		}
+		return "No unstaged diff.", nil
+	}
+	return text, nil
+}
+
+// git_log
+
+type GitLogTool struct {
+	workingDir string
+}
+
+func NewGitLogTool(workingDir string) *GitLogTool {
+	return &GitLogTool{workingDir: workingDir}
+}
+
+func (t *GitLogTool) Definition() llm.Tool {
+	return llm.Tool{
+		Type: "function",
+		Function: llm.ToolFunction{
+			Name:        "git_log",
+			Description: "Show recent git commits with git log --oneline --decorate.",
+			Parameters: JSONSchema(`{
+  "type": "object",
+  "properties": {
+    "path": {"type": "string", "description": "Optional path to limit log to. Relative paths are resolved from the configured agent working directory."},
+    "max_count": {"type": "integer", "description": "Maximum number of commits to show. Defaults to 20, capped at 100."}
+  },
+  "additionalProperties": false
+}`),
+		},
+	}
+}
+
+type gitLogArgs struct {
+	Path     string `json:"path"`
+	MaxCount int    `json:"max_count"`
+}
+
+func (t *GitLogTool) Execute(ctx context.Context, raw json.RawMessage) (string, error) {
+	var args gitLogArgs
+	if len(raw) > 0 {
+		if err := json.Unmarshal(raw, &args); err != nil {
+			return "", fmt.Errorf("parse arguments: %w", err)
+		}
+	}
+	maxCount := normalizeGitLogMaxCount(args.MaxCount)
+	gitArgs := []string{"log", "--oneline", "--decorate", fmt.Sprintf("-n%d", maxCount)}
+	if args.Path != "" {
+		pathspec, err := gitPathspec(t.workingDir, args.Path)
+		if err != nil {
+			return "", err
+		}
+		gitArgs = append(gitArgs, "--", pathspec)
+	}
+	text, err := runGit(ctx, t.workingDir, gitArgs...)
+	if err != nil {
+		return text, err
+	}
+	if strings.TrimSpace(text) == "" {
+		return "No git history.", nil
+	}
+	return text, nil
+}

internal/tools/git_test.go

@@ -0,0 +1,142 @@
+package tools
+
+import (
+	"context"
+	"encoding/json"
+	"os"
+	"os/exec"
+	"path/filepath"
+	"strings"
+	"testing"
+)
+
+func requireGit(t *testing.T) {
+	t.Helper()
+	if _, err := exec.LookPath("git"); err != nil {
+		t.Skip("git not installed")
+	}
+}
+
+func runGitForTest(t *testing.T, dir string, args ...string) {
+	t.Helper()
+	gitArgs := append([]string{"-C", dir}, args...)
+	cmd := exec.Command("git", gitArgs...)
+	out, err := cmd.CombinedOutput()
+	if err != nil {
+		t.Fatalf("git %v failed: %v\n%s", args, err, out)
+	}
+}
+
+func initTestRepo(t *testing.T) string {
+	t.Helper()
+	dir := t.TempDir()
+	runGitForTest(t, dir, "init")
+	runGitForTest(t, dir, "-c", "user.email=test@example.com", "-c", "user.name=Test", "commit", "--allow-empty", "-m", "initial")
+	return dir
+}
+
+func TestGitStatusToolShowsChangedFiles(t *testing.T) {
+	requireGit(t)
+	dir := initTestRepo(t)
+
+	path := filepath.Join(dir, "tracked.txt")
+	if err := os.WriteFile(path, []byte("hello"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	runGitForTest(t, dir, "add", "tracked.txt")
+	runGitForTest(t, dir, "-c", "user.email=test@example.com", "-c", "user.name=Test", "commit", "-m", "add tracked")
+
+	if err := os.WriteFile(path, []byte("world"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+
+	tool := NewGitStatusTool(dir)
+	out, err := tool.Execute(context.Background(), nil)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !strings.Contains(out, "Changed files:") {
+		t.Fatalf("output missing 'Changed files:': %q", out)
+	}
+	if !strings.Contains(out, "M tracked.txt") {
+		t.Fatalf("output missing 'M tracked.txt': %q", out)
+	}
+}
+
+func TestGitDiffToolShowsUnstagedDiff(t *testing.T) {
+	requireGit(t)
+	dir := initTestRepo(t)
+
+	path := filepath.Join(dir, "tracked.txt")
+	if err := os.WriteFile(path, []byte("hello"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	runGitForTest(t, dir, "add", "tracked.txt")
+	runGitForTest(t, dir, "-c", "user.email=test@example.com", "-c", "user.name=Test", "commit", "-m", "add tracked")
+
+	if err := os.WriteFile(path, []byte("world"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+
+	tool := NewGitDiffTool(dir)
+	out, err := tool.Execute(context.Background(), nil)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !strings.Contains(out, "diff --git") {
+		t.Fatalf("output missing 'diff --git': %q", out)
+	}
+}
+
+func TestGitDiffToolShowsStat(t *testing.T) {
+	requireGit(t)
+	dir := initTestRepo(t)
+
+	path := filepath.Join(dir, "tracked.txt")
+	if err := os.WriteFile(path, []byte("hello"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	runGitForTest(t, dir, "add", "tracked.txt")
+	runGitForTest(t, dir, "-c", "user.email=test@example.com", "-c", "user.name=Test", "commit", "-m", "add tracked")
+
+	if err := os.WriteFile(path, []byte("world"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+
+	tool := NewGitDiffTool(dir)
+	out, err := tool.Execute(context.Background(), json.RawMessage(`{"stat":true}`))
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !strings.Contains(out, "tracked.txt") {
+		t.Fatalf("output missing 'tracked.txt': %q", out)
+	}
+}
+
+func TestGitLogToolShowsCommits(t *testing.T) {
+	requireGit(t)
+	dir := initTestRepo(t)
+
+	tool := NewGitLogTool(dir)
+	out, err := tool.Execute(context.Background(), json.RawMessage(`{"max_count":1}`))
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !strings.Contains(out, "initial") {
+		t.Fatalf("output missing 'initial': %q", out)
+	}
+}
+
+func TestGitToolsRejectEscapingPath(t *testing.T) {
+	requireGit(t)
+	dir := initTestRepo(t)
+
+	tool := NewGitStatusTool(dir)
+	_, err := tool.Execute(context.Background(), json.RawMessage(`{"path":"../outside"}`))
+	if err == nil {
+		t.Fatal("expected error for escaping path")
+	}
+	if !strings.Contains(err.Error(), "escapes") {
+		t.Fatalf("error = %v", err)
+	}
+}

internal/tools/paths.go

@@ -1,7 +1,9 @@
 package tools
 
 import (
+	"fmt"
 	"path/filepath"
+	"strings"
 )
 
 func resolvePath(baseDir, path string) (string, error) {
@@ -9,7 +11,22 @@ func resolvePath(baseDir, path string) (string, error) {
 		path = "."
 	}
 	if filepath.IsAbs(path) {
-		return filepath.Clean(path), nil
+		return "", fmt.Errorf("absolute paths are not allowed; use a path relative to the project directory")
 	}
-	return filepath.Abs(filepath.Join(baseDir, path))
+	resolved := filepath.Clean(filepath.Join(baseDir, path))
+	if !isWithinDir(resolved, baseDir) {
+		return "", fmt.Errorf("path escapes the project directory: %s", path)
+	}
+	return resolved, nil
+}
+
+func isWithinDir(path, dir string) bool {
+	rel, err := filepath.Rel(dir, path)
+	if err != nil {
+		return false
+	}
+	if rel == ".." || strings.HasPrefix(rel, ".."+string(filepath.Separator)) {
+		return false
+	}
+	return true
 }

internal/tools/shell.go

@@ -9,7 +9,7 @@ import (
 	"os/exec"
 	"strings"
 
-	"agentu/internal/llm"
+	"agentu/pkg/llm"
 )
 
 type ShellRunTool struct {
@@ -19,6 +19,7 @@ type ShellRunTool struct {
 func NewShellRunTool(workingDir string) *ShellRunTool {
 	return &ShellRunTool{workingDir: workingDir}
 }
+func (*ShellRunTool) Mutates() bool { return true }
 
 func (t *ShellRunTool) Definition() llm.Tool {
 	return llm.Tool{

internal/tools/tools.go

@@ -5,17 +5,32 @@ import (
 	"encoding/json"
 	"fmt"
 	"sort"
+	"strings"
 
-	"agentu/internal/llm"
+	"agentu/pkg/llm"
 )
 
 const MaxToolOutputBytes = 64 * 1024
 
+const truncationNoticeBudget = 160
+
 type Tool interface {
 	Definition() llm.Tool
 	Execute(ctx context.Context, args json.RawMessage) (string, error)
 }
 
+// MutatingTool is optionally implemented by tools that modify external state.
+// It is used by the agent to determine whether tool calls can run in parallel.
+type MutatingTool interface {
+	Mutates() bool
+}
+
+// IsMutating reports whether tool implements MutatingTool and returns true.
+func IsMutating(tool Tool) bool {
+	mutating, ok := tool.(MutatingTool)
+	return ok && mutating.Mutates()
+}
+
 type Registry struct {
 	tools   map[string]Tool
 	primary []string
@@ -62,29 +77,40 @@ func ReadOnlyBuiltins(workingDir string) *Registry {
 	fileRead := NewFileReadTool(workingDir)
 	fileList := NewFileListTool(workingDir)
 	fileSearch := NewFileSearchTool(workingDir)
+	codeSymbols := NewCodeSymbolsTool(workingDir)
+	gitStatus := NewGitStatusTool(workingDir)
+	gitDiff := NewGitDiffTool(workingDir)
+	gitLog := NewGitLogTool(workingDir)
 
-	registry := NewRegistry(fileRead, fileList, fileSearch)
+	registry := NewRegistry(fileRead, fileList, fileSearch, codeSymbols, gitStatus, gitDiff, gitLog)
 	registry.RegisterAlias("file.read", fileRead)
 	registry.RegisterAlias("file.list", fileList)
 	registry.RegisterAlias("file.search", fileSearch)
+	registry.RegisterAlias("code.symbols", codeSymbols)
+	registry.RegisterAlias("git.status", gitStatus)
+	registry.RegisterAlias("git.diff", gitDiff)
+	registry.RegisterAlias("git.log", gitLog)
 	return registry
 }
 
 func Builtins(workingDir string) *Registry {
 	registry := ReadOnlyBuiltins(workingDir)
 	fileWrite := NewFileWriteTool(workingDir)
+	fileEdit := NewFileEditTool(workingDir)
 	shellRun := NewShellRunTool(workingDir)
 	registry.Register(fileWrite)
+	registry.Register(fileEdit)
 	registry.Register(shellRun)
 	registry.RegisterAlias("file.write", fileWrite)
+	registry.RegisterAlias("file.edit", fileEdit)
 	registry.RegisterAlias("shell.run", shellRun)
 	return registry
 }
 
 func AgentInstructions(workingDir string, yolo bool) string {
 	searchInstructions := webSearchInstructions()
-	readOnlyTools := "file_list, file_read, file_search, web_search, web_fetch"
-	yoloTools := "file_list, file_read, file_search, file_write, shell_run, web_search, web_fetch"
+	readOnlyTools := "file_list, file_read, file_search, code_symbols, git_status, git_diff, git_log, web_search, web_fetch"
+	yoloTools := "file_list, file_read, file_search, code_symbols, file_write, file_edit, shell_run, git_status, git_diff, git_log, web_search, web_fetch"
 	if yolo {
 		return fmt.Sprintf(`Local project context:
 - The project working directory is %q.
@@ -92,8 +118,11 @@ func AgentInstructions(workingDir string, yolo bool) string {
 - Do not call tools for greetings, small talk, simple Q&A, or general questions that do not require local project or command context.
 - Do not explore the project preemptively. Do not call file_list just to discover context unless the user asks about project files or the answer truly depends on local files.
 - When the user asks to inspect, list, search, or read project files, use file_list, file_search, and file_read instead of saying you cannot access local files.
+- When the user asks about code structure, symbols, functions, methods, types, or definitions in Go files, use code_symbols before reading whole files.
 - When the user asks to run, execute, test, check, debug, or inspect a shell command, use shell_run. This includes local commands written in backticks, such as go, make, git, pwd, ls, ps, and similar terminal commands.
+- When editing existing files, prefer file_edit for small or localized changes; use file_write only for creating files, replacing whole files intentionally, or appending.
 - Do not use file tools as a substitute for a command execution request.
+- When the user asks about changed files, diffs, or git history, use git_status, git_diff, or git_log instead of shell_run.
 - Do not start interactive or long-running sessions. Prefer bounded, non-interactive local commands with clear completion conditions.
 - %s
 - Tool paths should normally be relative to the project working directory.
@@ -105,7 +134,9 @@ func AgentInstructions(workingDir string, yolo bool) string {
 - Do not call tools for greetings, small talk, simple Q&A, or general questions that do not require local project context.
 - Do not explore the project preemptively. Do not call file_list just to discover context unless the user asks about project files or the answer truly depends on local files.
 - When the user asks to inspect, list, search, or read project files, use file_list, file_search, and file_read instead of saying you cannot access local files.
+- When the user asks about code structure, symbols, functions, methods, types, or definitions in Go files, use code_symbols before reading whole files.
 - Shell command execution is disabled in this mode. When the user asks to run, execute, test, check, debug, or inspect a shell command, do not use file tools as a substitute; explain that they need to restart agentu with --yolo to enable shell_run.
+- When the user asks about changed files, diffs, or git history, use git_status, git_diff, or git_log instead of shell_run.
 - %s
 - Tool paths should normally be relative to the project working directory.
 - Available tools: %s.`, workingDir, searchInstructions, readOnlyTools)
@@ -120,10 +151,40 @@ func JSONSchema(schema string) json.RawMessage {
 }
 
 func FormatResult(output string) string {
-	if len(output) <= MaxToolOutputBytes {
+	return FormatResultLimit(output, MaxToolOutputBytes)
+}
+
+func FormatResultLimit(output string, limit int) string {
+	if limit <= 0 || len(output) <= limit {
 		return output
 	}
-	return output[:MaxToolOutputBytes] + fmt.Sprintf("\n\n[truncated after %d bytes]", MaxToolOutputBytes)
+	if limit <= truncationNoticeBudget {
+		return output[:limit]
+	}
+	notice := fmt.Sprintf("\n\n[truncated: showing head and tail; omitted %d bytes]\n\n", len(output)-limit)
+	keep := limit - len(notice)
+	if keep <= 0 {
+		return output[:limit]
+	}
+	headLen := keep / 2
+	tailLen := keep - headLen
+	head := trimHeadAtLine(output[:headLen])
+	tail := trimTailAtLine(output[len(output)-tailLen:])
+	return head + notice + tail
+}
+
+func trimHeadAtLine(value string) string {
+	if idx := strings.LastIndexByte(value, '\n'); idx > 0 {
+		return value[:idx+1]
+	}
+	return value
+}
+
+func trimTailAtLine(value string) string {
+	if idx := strings.IndexByte(value, '\n'); idx >= 0 && idx+1 < len(value) {
+		return value[idx+1:]
+	}
+	return value
 }
 
 func FormatError(err error) string {

internal/tools/tools_test.go

@@ -43,6 +43,41 @@ func TestFileTools(t *testing.T) {
 	}
 }
 
+func TestFileSearchSupportsContextAndMaxResults(t *testing.T) {
+	dir := t.TempDir()
+	path := filepath.Join(dir, "notes.txt")
+	content := strings.Join([]string{"before", "needle one", "middle", "needle two", "after"}, "\n")
+	if err := os.WriteFile(path, []byte(content), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	search := NewFileSearchTool(dir)
+	out, err := search.walkSearch(context.Background(), fileSearchArgs{Query: "needle", ContextLines: 1, MaxResults: 1}, dir)
+	if err != nil {
+		t.Fatal(err)
+	}
+	for _, want := range []string{"-- ", "before", "needle one", "middle", "stopped after 1 matches"} {
+		if !strings.Contains(out, want) {
+			t.Fatalf("search output missing %q:\n%s", want, out)
+		}
+	}
+	if strings.Contains(out, "needle two") {
+		t.Fatalf("search should stop after max_results:\n%s", out)
+	}
+}
+
+func TestFormatResultKeepsHeadAndTail(t *testing.T) {
+	input := strings.Join([]string{"head-1", "head-2", strings.Repeat("x", 220), "tail-1", "tail-2"}, "\n")
+	out := FormatResultLimit(input, 220)
+	for _, want := range []string{"head-1", "truncated: showing head and tail", "tail-2"} {
+		if !strings.Contains(out, want) {
+			t.Fatalf("formatted output missing %q:\n%s", want, out)
+		}
+	}
+	if len(out) > 220 {
+		t.Fatalf("formatted output len = %d, want <= 220", len(out))
+	}
+}
+
 func TestShellRunTool(t *testing.T) {
 	dir := t.TempDir()
 	if err := os.WriteFile(filepath.Join(dir, "x.txt"), []byte("x"), 0o644); err != nil {
@@ -69,12 +104,12 @@ func TestReadOnlyBuiltinsExposeOpenAICompatibleNamesAndAliases(t *testing.T) {
 		}
 	}
 
-	for _, want := range []string{"file_list", "file_read", "file_search"} {
+	for _, want := range []string{"file_list", "file_read", "file_search", "code_symbols", "git_status", "git_diff", "git_log"} {
 		if !slices.Contains(names, want) {
 			t.Fatalf("definitions missing %s: %#v", want, names)
 		}
 	}
-	for _, alias := range []string{"file.list", "file.read", "file.search"} {
+	for _, alias := range []string{"file.list", "file.read", "file.search", "code.symbols", "git.status", "git.diff", "git.log"} {
 		if _, ok := registry.Get(alias); !ok {
 			t.Fatalf("alias missing: %s", alias)
 		}
@@ -82,6 +117,37 @@ func TestReadOnlyBuiltinsExposeOpenAICompatibleNamesAndAliases(t *testing.T) {
 	if _, ok := registry.Get("file_write"); ok {
 		t.Fatal("read-only registry should not expose file_write")
 	}
+	if _, ok := registry.Get("file_edit"); ok {
+		t.Fatal("read-only registry should not expose file_edit")
+	}
+}
+
+func TestCodeSymbolsToolListsGoSymbols(t *testing.T) {
+	dir := t.TempDir()
+	path := filepath.Join(dir, "sample.go")
+	if err := os.WriteFile(path, []byte(`package sample
+
+import "context"
+
+type Worker struct{}
+type Runner interface{ Run(context.Context) error }
+
+func NewWorker() *Worker { return &Worker{} }
+func (w *Worker) Run(ctx context.Context) error { return nil }
+`), 0o644); err != nil {
+		t.Fatal(err)
+	}
+
+	tool := NewCodeSymbolsTool(dir)
+	out, err := tool.Execute(context.Background(), json.RawMessage(`{"path":"sample.go"}`))
+	if err != nil {
+		t.Fatal(err)
+	}
+	for _, want := range []string{"package sample", "imports: context", "type Worker struct", "type Runner interface", "func NewWorker", "method *Worker.Run"} {
+		if !strings.Contains(out, want) {
+			t.Fatalf("symbols output missing %q:\n%s", want, out)
+		}
+	}
 }
 
 func TestRegistryCanExposeWebSearch(t *testing.T) {
@@ -107,7 +173,7 @@ func TestRegistryCanExposeWebSearch(t *testing.T) {
 
 func TestBuiltinsExposeYoloTools(t *testing.T) {
 	registry := Builtins(t.TempDir())
-	for _, name := range []string{"file_write", "shell_run", "file.write", "shell.run"} {
+	for _, name := range []string{"file_write", "file_edit", "shell_run", "file.write", "file.edit", "shell.run"} {
 		if _, ok := registry.Get(name); !ok {
 			t.Fatalf("tool missing: %s", name)
 		}
@@ -116,14 +182,14 @@ func TestBuiltinsExposeYoloTools(t *testing.T) {
 
 func TestAgentInstructionsExplainCommandModes(t *testing.T) {
 	readOnly := AgentInstructions("/tmp/project", false)
-	for _, want := range []string{"Shell command execution is disabled", "--yolo", "Do not call tools for greetings", "Do not explore the project preemptively", "Do not call file_list just to discover context"} {
+	for _, want := range []string{"Shell command execution is disabled", "--yolo", "Do not call tools for greetings", "Do not explore the project preemptively", "Do not call file_list just to discover context", "git_status, git_diff, git_log"} {
 		if !strings.Contains(readOnly, want) {
 			t.Fatalf("read-only instructions missing %q:\n%s", want, readOnly)
 		}
 	}
 
 	yolo := AgentInstructions("/tmp/project", true)
-	for _, want := range []string{"use shell_run", "commands written in backticks", "Do not start interactive", "non-interactive local commands", "Do not call tools for greetings"} {
+	for _, want := range []string{"use shell_run", "commands written in backticks", "Do not start interactive", "non-interactive local commands", "Do not call tools for greetings", "prefer file_edit"} {
 		if !strings.Contains(yolo, want) {
 			t.Fatalf("yolo instructions missing %q:\n%s", want, yolo)
 		}
@@ -132,9 +198,78 @@ func TestAgentInstructionsExplainCommandModes(t *testing.T) {
 
 func TestAgentInstructionsExplainWebSearch(t *testing.T) {
 	withSearch := AgentInstructions("/tmp/project", false)
-	for _, want := range []string{"use web_search", "Use web_fetch", "latest information", "source URLs", "Available tools: file_list, file_read, file_search, web_search, web_fetch"} {
+	for _, want := range []string{"use web_search", "Use web_fetch", "latest information", "source URLs", "Available tools: file_list, file_read, file_search, code_symbols, git_status, git_diff, git_log, web_search, web_fetch"} {
 		if !strings.Contains(withSearch, want) {
 			t.Fatalf("web search instructions missing %q:\n%s", want, withSearch)
 		}
 	}
 }
+
+func TestResolvePathRejectsAbsolutePaths(t *testing.T) {
+	base := t.TempDir()
+	_, err := resolvePath(base, "/etc/passwd")
+	if err == nil {
+		t.Fatal("expected error for absolute path")
+	}
+	if !strings.Contains(err.Error(), "absolute paths are not allowed") {
+		t.Fatalf("unexpected error: %v", err)
+	}
+}
+
+func TestResolvePathRejectsEscapingBaseDir(t *testing.T) {
+	base := filepath.Join(t.TempDir(), "project")
+	if err := os.MkdirAll(base, 0o755); err != nil {
+		t.Fatal(err)
+	}
+	_, err := resolvePath(base, "../../../etc/passwd")
+	if err == nil {
+		t.Fatal("expected error for path escaping base dir")
+	}
+	if !strings.Contains(err.Error(), "escapes the project directory") {
+		t.Fatalf("unexpected error: %v", err)
+	}
+}
+
+func TestResolvePathAllowsSubdirTraversal(t *testing.T) {
+	base := t.TempDir()
+	if err := os.MkdirAll(filepath.Join(base, "a", "b"), 0o755); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(filepath.Join(base, "a", "target.txt"), []byte("ok"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	// "a/b/../target.txt" resolves to "a/target.txt" which is within base.
+	resolved, err := resolvePath(base, "a/b/../target.txt")
+	if err != nil {
+		t.Fatal(err)
+	}
+	want := filepath.Join(base, "a", "target.txt")
+	if resolved != want {
+		t.Fatalf("resolved = %q, want %q", resolved, want)
+	}
+}
+
+func TestResolvePathRejectsDotDotThatEscapes(t *testing.T) {
+	base := t.TempDir()
+	sub := filepath.Join(base, "a")
+	if err := os.MkdirAll(sub, 0o755); err != nil {
+		t.Fatal(err)
+	}
+	_, err := resolvePath(sub, "..")
+	if err == nil {
+		t.Fatal("expected error for path escaping via ..")
+	}
+}
+
+func TestIsMutating(t *testing.T) {
+	dir := t.TempDir()
+	if IsMutating(NewFileReadTool(dir)) {
+		t.Fatal("FileReadTool should not be mutating")
+	}
+	if !IsMutating(NewFileWriteTool(dir)) {
+		t.Fatal("FileWriteTool should be mutating")
+	}
+	if !IsMutating(NewShellRunTool(dir)) {
+		t.Fatal("ShellRunTool should be mutating")
+	}
+}

internal/tools/web_builtin.go

@@ -10,7 +10,7 @@ import (
 	"net/url"
 	"strings"
 
-	"agentu/internal/llm"
+	"agentu/pkg/llm"
 
 	"golang.org/x/net/html"
 )

internal/tui/app_test.go

@@ -2,22 +2,75 @@ package tui
 
 import (
 	"context"
+	"os"
+	"os/exec"
+	"path/filepath"
 	"strings"
 	"testing"
 
+	"agentu/internal/agent"
+	"agentu/pkg/llm"
+
 	tea "github.com/charmbracelet/bubbletea"
+	"github.com/charmbracelet/lipgloss"
 )
 
 func TestModelRendersChatShell(t *testing.T) {
 	m := newModel(context.Background(), nil, Options{ModelName: "test-model", Yolo: true})
 	next, _ := m.Update(tea.WindowSizeMsg{Width: 100, Height: 30})
-	rendered := next.(model).View()
+	m = next.(model)
+	rendered := m.View()
 
-	for _, want := range []string{"agentu", "test-model", "yolo tools", "light", "ctrl+j", "alt+enter"} {
+	for _, want := range []string{"test-model", "yolo tools", "light", "ctrl+j", "alt+enter", "ctrl+p/ctrl+n history"} {
 		if !strings.Contains(rendered, want) {
 			t.Fatalf("rendered view missing %q:\n%s", want, rendered)
 		}
 	}
+	if strings.Contains(rendered, "local agent") {
+		t.Fatalf("top header should not render:\n%s", rendered)
+	}
+	if _, ok := m.styles.App.GetBackground().(lipgloss.NoColor); !ok {
+		t.Fatalf("app should not paint a global background")
+	}
+	if _, ok := m.styles.Viewport.GetBackground().(lipgloss.NoColor); !ok {
+		t.Fatalf("viewport should not paint an assistant background")
+	}
+}
+
+func TestInputTextDoesNotPaintBackground(t *testing.T) {
+	m := newModel(context.Background(), nil, Options{ModelName: "test-model"})
+	for name, color := range map[string]any{
+		"input box":           m.styles.InputBox.GetBackground(),
+		"focused base":        m.input.FocusedStyle.Base.GetBackground(),
+		"focused cursor line": m.input.FocusedStyle.CursorLine.GetBackground(),
+		"blurred base":        m.input.BlurredStyle.Base.GetBackground(),
+		"blurred cursor line": m.input.BlurredStyle.CursorLine.GetBackground(),
+	} {
+		if _, ok := color.(lipgloss.NoColor); !ok {
+			t.Fatalf("%s should not paint input text background", name)
+		}
+	}
+}
+
+func TestInputBoxIsRoomier(t *testing.T) {
+	m := newModel(context.Background(), nil, Options{ModelName: "test-model"})
+	next, _ := m.Update(tea.WindowSizeMsg{Width: 100, Height: 30})
+	m = next.(model)
+
+	if got := m.styles.InputBox.GetVerticalFrameSize(); got != 2 {
+		t.Fatalf("input box vertical frame should stay compact, got %d", got)
+	}
+	if got := m.styles.InputBox.GetHorizontalFrameSize(); got < 6 {
+		t.Fatalf("input box horizontal frame should include roomier padding, got %d", got)
+	}
+	if got, want := m.input.Width(), 100-m.styles.InputBox.GetHorizontalFrameSize(); got != want {
+		t.Fatalf("input content width = %d, want %d", got, want)
+	}
+	rendered := m.inputView()
+	line := strings.Split(rendered, "\n")[0]
+	if got := lipgloss.Width(line); got != 100 {
+		t.Fatalf("input box rendered width = %d, want 100:\n%s", got, rendered)
+	}
 }
 
 func TestModelMetaRendersBelowInput(t *testing.T) {
@@ -35,6 +88,20 @@ func TestModelMetaRendersBelowInput(t *testing.T) {
 	}
 }
 
+func TestModelMetaRendersContextUsage(t *testing.T) {
+	assistant := agent.New(agent.Options{Model: "test-model", MaxContextTokens: 1000})
+	assistant.SetMessages([]llm.Message{{Role: llm.RoleUser, Content: strings.Repeat("x", 800)}})
+	m := newModel(context.Background(), assistant, Options{ModelName: "test-model"})
+	next, _ := m.Update(tea.WindowSizeMsg{Width: 120, Height: 30})
+	rendered := next.(model).modelMetaView()
+
+	for _, want := range []string{"ctx ~", "/1k", "%"} {
+		if !strings.Contains(rendered, want) {
+			t.Fatalf("model meta missing %q:\n%s", want, rendered)
+		}
+	}
+}
+
 func TestInitialViewDoesNotRenderTopNote(t *testing.T) {
 	m := newModel(context.Background(), nil, Options{ModelName: "test-model"})
 	next, _ := m.Update(tea.WindowSizeMsg{Width: 100, Height: 30})
@@ -56,43 +123,97 @@ func TestRunningStateRendersActivityIndicator(t *testing.T) {
 	m.resize()
 
 	rendered := m.View()
-	for _, want := range []string{"Working", "thinking...", "esc cancels"} {
+	for _, want := range []string{iconWorking, "Working", "thinking...", "esc cancels"} {
 		if !strings.Contains(rendered, want) {
 			t.Fatalf("running view missing %q:\n%s", want, rendered)
 		}
 	}
+	if backgroundPattern.MatchString(m.activityView()) {
+		t.Fatalf("activity view should not paint a background:\n%q", m.activityView())
+	}
 }
 
-func TestUserMessageRendersOnLeft(t *testing.T) {
+func TestUserMessageRendersFullWidthBubble(t *testing.T) {
 	m := newModel(context.Background(), nil, Options{ModelName: "test-model"})
 	next, _ := m.Update(tea.WindowSizeMsg{Width: 100, Height: 30})
 	m = next.(model)
 
-	rendered := m.messageView(roleUser, "hello", 80)
-	if strings.HasPrefix(rendered, " ") {
-		t.Fatalf("user message should not be right-aligned:\n%q", rendered)
-	}
-	if !strings.Contains(rendered, "You") || !strings.Contains(rendered, "hello") {
+	rendered := m.messageView(roleUser, "hi", 80)
+	if !strings.Contains(rendered, "hi") {
 		t.Fatalf("user message missing content:\n%s", rendered)
 	}
+	if strings.Contains(rendered, "You") {
+		t.Fatalf("user message should not render label:\n%s", rendered)
+	}
+	if !strings.Contains(rendered, " hi") || !strings.Contains(rendered, "hi ") {
+		t.Fatalf("user message should have horizontal padding:\n%q", rendered)
+	}
+	lines := strings.Split(rendered, "\n")
+	if len(lines) < 3 {
+		t.Fatalf("user message background should include vertical padding:\n%q", rendered)
+	}
+	for _, line := range lines {
+		if got := lipgloss.Width(line); got != 80 {
+			t.Fatalf("user message background should span full width, line width = %d:\n%q", got, rendered)
+		}
+	}
+	m.width = 80
+	m.messages = []message{{role: roleUser, content: "hi"}}
+	rendered = m.renderMessages()
+	for _, line := range strings.Split(rendered, "\n") {
+		if got := lipgloss.Width(line); got != 80 {
+			t.Fatalf("rendered user transcript should span full width, line width = %d:\n%q", got, rendered)
+		}
+	}
 }
 
-func TestAssistantMessagesRenderMarkdown(t *testing.T) {
+func TestAssistantMessagesRenderMarkdownWithoutChrome(t *testing.T) {
 	m := newModel(context.Background(), nil, Options{ModelName: "test-model"})
 	next, _ := m.Update(tea.WindowSizeMsg{Width: 100, Height: 30})
 	m = next.(model)
 
 	rendered := m.messageView(roleAssistant, "# Title\n\n- item", 80)
-	for _, want := range []string{"Agentu", "Title", "item"} {
+	for _, want := range []string{"Title", "item"} {
 		if !strings.Contains(rendered, want) {
 			t.Fatalf("assistant markdown missing %q:\n%s", want, rendered)
 		}
 	}
+	for _, unwanted := range []string{"Agentu", "│", "┃", "|"} {
+		if strings.Contains(rendered, unwanted) {
+			t.Fatalf("assistant message should not render chrome %q:\n%s", unwanted, rendered)
+		}
+	}
 	if strings.Contains(rendered, "# Title") {
 		t.Fatalf("assistant markdown heading was not rendered:\n%s", rendered)
 	}
 }
 
+func TestAssistantMessageRendersPaddedBodyWithoutBackground(t *testing.T) {
+	m := newModel(context.Background(), nil, Options{ModelName: "test-model"})
+	next, _ := m.Update(tea.WindowSizeMsg{Width: 100, Height: 30})
+	m = next.(model)
+	if _, ok := m.styles.AssistantMsg.GetBackground().(lipgloss.NoColor); !ok {
+		t.Fatalf("assistant message should not set a background color")
+	}
+
+	rendered := m.messageView(roleAssistant, "hello", 80)
+	if !strings.Contains(rendered, "hello") {
+		t.Fatalf("assistant message missing content:\n%s", rendered)
+	}
+	lines := strings.Split(rendered, "\n")
+	if len(lines) < 3 {
+		t.Fatalf("assistant message should include vertical padding:\n%q", rendered)
+	}
+	if got := lipgloss.Width(rendered); got <= len("hello") {
+		t.Fatalf("assistant message should be wider than text, width = %d:\n%q", got, rendered)
+	}
+	for _, line := range lines {
+		if got := lipgloss.Width(line); got >= 80 {
+			t.Fatalf("assistant message should not fill the row, line width = %d:\n%q", got, rendered)
+		}
+	}
+}
+
 func TestUserMessagesDoNotRenderMarkdown(t *testing.T) {
 	m := newModel(context.Background(), nil, Options{ModelName: "test-model"})
 	next, _ := m.Update(tea.WindowSizeMsg{Width: 100, Height: 30})
@@ -106,7 +227,7 @@ func TestUserMessagesDoNotRenderMarkdown(t *testing.T) {
 	}
 }
 
-func TestMessageRoleLabels(t *testing.T) {
+func TestMessageRoleLabelsDoNotLabelUserOrAssistant(t *testing.T) {
 	m := newModel(context.Background(), nil, Options{ModelName: "test-model"})
 	next, _ := m.Update(tea.WindowSizeMsg{Width: 100, Height: 30})
 	m = next.(model)
@@ -116,9 +237,14 @@ func TestMessageRoleLabels(t *testing.T) {
 	}
 
 	rendered := m.renderMessages()
-	for _, want := range []string{"You", "Agentu"} {
+	for _, want := range []string{"hello", "hi"} {
 		if !strings.Contains(rendered, want) {
-			t.Fatalf("messages missing %q:\n%s", want, rendered)
+			t.Fatalf("messages missing content %q:\n%s", want, rendered)
+		}
+	}
+	for _, unwanted := range []string{"You", "Agentu"} {
+		if strings.Contains(rendered, unwanted) {
+			t.Fatalf("message should not render label %q:\n%s", unwanted, rendered)
 		}
 	}
 }
@@ -153,19 +279,205 @@ func TestInputStartsAtOneLineAndGrows(t *testing.T) {
 	}
 }
 
-func TestSlashCommandSuggestions(t *testing.T) {
+func TestCommandPaletteFuzzyMatches(t *testing.T) {
 	m := newModel(context.Background(), nil, Options{ModelName: "test-model", Yolo: true})
 	next, _ := m.Update(tea.WindowSizeMsg{Width: 100, Height: 30})
 	m = next.(model)
-	m.input.SetValue("/")
+	m.input.SetValue("/mdl")
 	m.afterInputChanged()
 
-	rendered := m.View()
-	for _, want := range []string{"/clear reset context", "/exit quit", "/model model/provider/thinking"} {
-		if !strings.Contains(rendered, want) {
-			t.Fatalf("rendered suggestions missing %q:\n%s", want, rendered)
+	plain := stripANSI(m.View())
+	for _, want := range []string{iconCommand + " Commands", "/model model/provider/thinking"} {
+		if !strings.Contains(plain, want) {
+			t.Fatalf("rendered palette missing %q:\n%s", want, plain)
 		}
 	}
+
+	matches := commandMatches("m")
+	modelIndex := -1
+	renameIndex := -1
+	for i, match := range matches {
+		switch match.command.Name {
+		case "/model":
+			modelIndex = i
+		case "/rename":
+			renameIndex = i
+		}
+	}
+	if modelIndex < 0 || renameIndex < 0 || modelIndex > renameIndex {
+		t.Fatalf("/model should sort before /rename for /m query: %#v", matches)
+	}
+}
+
+func TestInputHistoryNavigation(t *testing.T) {
+	m := newModel(context.Background(), nil, Options{ModelName: "test-model"})
+	m.rememberInput("first")
+	m.rememberInput("second")
+
+	next, _ := m.Update(tea.KeyMsg{Type: tea.KeyCtrlP})
+	m = next.(model)
+	if got := m.input.Value(); got != "second" {
+		t.Fatalf("after ctrl+p input = %q", got)
+	}
+	next, _ = m.Update(tea.KeyMsg{Type: tea.KeyCtrlP})
+	m = next.(model)
+	if got := m.input.Value(); got != "first" {
+		t.Fatalf("after second ctrl+p input = %q", got)
+	}
+	next, _ = m.Update(tea.KeyMsg{Type: tea.KeyCtrlN})
+	m = next.(model)
+	if got := m.input.Value(); got != "second" {
+		t.Fatalf("after ctrl+n input = %q", got)
+	}
+	next, _ = m.Update(tea.KeyMsg{Type: tea.KeyCtrlN})
+	m = next.(model)
+	if got := m.input.Value(); got != "" {
+		t.Fatalf("after second ctrl+n input = %q", got)
+	}
+}
+
+func TestToolMessageRendersStructuredCard(t *testing.T) {
+	m := newModel(context.Background(), nil, Options{ModelName: "test-model"})
+	m.messages = append(m.messages, message{role: roleTool, content: `file_read {"path":"README.md"}`})
+
+	plain := stripANSI(m.renderMessages())
+	for _, want := range []string{iconTool + " file_read", "README.md"} {
+		if !strings.Contains(plain, want) {
+			t.Fatalf("tool card missing %q:\n%s", want, plain)
+		}
+	}
+}
+
+func TestToolMessageRendersOutputPreview(t *testing.T) {
+	m := newModel(context.Background(), nil, Options{ModelName: "test-model"})
+	m.messages = append(m.messages, message{role: roleTool, content: "file_read {\"path\":\"README.md\"}\n[output]\nfirst line\nsecond line"})
+
+	plain := stripANSI(m.renderMessages())
+	for _, want := range []string{iconTool + " file_read", "README.md", "output", "first line", "second line"} {
+		if !strings.Contains(plain, want) {
+			t.Fatalf("tool card missing %q:\n%s", want, plain)
+		}
+	}
+}
+
+func TestToolLogOutputUpdatesExistingToolCard(t *testing.T) {
+	m := newModel(context.Background(), nil, Options{ModelName: "test-model"})
+	m.messages = append(m.messages, message{role: roleTool, content: `file_read {"path":"README.md"}`})
+
+	updated := m.mergeToolLog("file_read {\"path\":\"README.md\"}\n[output]\nhello")
+	if !updated {
+		t.Fatal("expected mergeToolLog to return true")
+	}
+	if len(m.messages) != 1 {
+		t.Fatalf("messages count = %d, want 1", len(m.messages))
+	}
+	plain := stripANSI(m.renderMessages())
+	if !strings.Contains(plain, "hello") {
+		t.Fatalf("rendered output missing 'hello':\n%s", plain)
+	}
+
+	status := toolStatusText("file_read {\"path\":\"README.md\"}\n[output]\nhello world")
+	if strings.Contains(status, "hello") {
+		t.Fatalf("status should not contain output: %q", status)
+	}
+	if !strings.Contains(status, "file_read") {
+		t.Fatalf("status missing tool name: %q", status)
+	}
+}
+
+type tuiCompactProvider struct{}
+
+func (tuiCompactProvider) ChatStream(ctx context.Context, req llm.ChatRequest, emit func(llm.StreamEvent) error) error {
+	return emit(llm.StreamEvent{Content: "summary from tui"})
+}
+
+func TestCompactCommand(t *testing.T) {
+	assistant := agent.New(agent.Options{Provider: tuiCompactProvider{}, Model: "test", SystemPrompt: "system"})
+	assistant.SetMessages([]llm.Message{
+		{Role: llm.RoleSystem, Content: "system"},
+		{Role: llm.RoleUser, Content: "old question 1"},
+		{Role: llm.RoleAssistant, Content: "old answer 1"},
+		{Role: llm.RoleUser, Content: "old question 2"},
+		{Role: llm.RoleAssistant, Content: "old answer 2"},
+		{Role: llm.RoleUser, Content: "recent question 1"},
+		{Role: llm.RoleAssistant, Content: "recent answer 1"},
+		{Role: llm.RoleUser, Content: "recent question 2"},
+		{Role: llm.RoleAssistant, Content: "recent answer 2"},
+	})
+	m := newModel(context.Background(), assistant, Options{ModelName: "test-model"})
+
+	out, err := m.handleLocalCommand("/compact")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if out != "Context compacted." {
+		t.Fatalf("output = %q", out)
+	}
+	if !strings.Contains(messageContents(assistant.Messages()), "[context summary]") {
+		t.Fatalf("messages missing summary: %#v", assistant.Messages())
+	}
+}
+
+func messageContents(messages []llm.Message) string {
+	var b strings.Builder
+	for _, msg := range messages {
+		b.WriteString(msg.Content)
+		b.WriteByte('\n')
+	}
+	return b.String()
+}
+
+func TestWorkflowStatusDiffAndTestCommands(t *testing.T) {
+	if _, err := exec.LookPath("git"); err != nil {
+		t.Skip("git not installed")
+	}
+	dir := t.TempDir()
+	runInDir(t, dir, "git", "init")
+	runInDir(t, dir, "git", "config", "user.email", "agentu@example.test")
+	runInDir(t, dir, "git", "config", "user.name", "agentu")
+	if err := os.WriteFile(filepath.Join(dir, "go.mod"), []byte("module example.com/p1\n\ngo 1.26\n"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(filepath.Join(dir, "main_test.go"), []byte("package p1\n\nimport \"testing\"\n\nfunc TestOK(t *testing.T) {}\n"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	runInDir(t, dir, "git", "add", ".")
+	runInDir(t, dir, "git", "commit", "-m", "init")
+	if err := os.WriteFile(filepath.Join(dir, "main_test.go"), []byte("package p1\n\nimport \"testing\"\n\nfunc TestOK(t *testing.T) { t.Log(\"changed\") }\n"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+
+	m := newModel(context.Background(), nil, Options{ModelName: "test-model", WorkingDir: dir})
+	status, err := m.handleStatusCommand(nil)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !strings.Contains(status, "main_test.go") {
+		t.Fatalf("status missing changed file:\n%s", status)
+	}
+	diff, err := m.handleDiffCommand(nil)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !strings.Contains(diff, "changed") {
+		t.Fatalf("diff missing changed content:\n%s", diff)
+	}
+	out, err := m.handleTestCommand(nil)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !strings.Contains(out, "$ go test ./...") {
+		t.Fatalf("test output missing default command:\n%s", out)
+	}
+}
+
+func runInDir(t *testing.T, dir string, name string, args ...string) {
+	t.Helper()
+	cmd := exec.Command(name, args...)
+	cmd.Dir = dir
+	if output, err := cmd.CombinedOutput(); err != nil {
+		t.Fatalf("%s %v failed: %v\n%s", name, args, err, output)
+	}
 }
 
 func TestModelCommand(t *testing.T) {
@@ -200,6 +512,8 @@ type fakeModelManager struct {
 	provider    string
 	model       string
 	thinking    string
+	sessionName string
+	saved       bool
 }
 
 func (f *fakeModelManager) CurrentProvider() string {
@@ -217,6 +531,17 @@ func (f *fakeModelManager) CurrentThinking() string {
 	return f.thinking
 }
 
+func (f *fakeModelManager) CurrentSessionID() string {
+	return "test-id"
+}
+
+func (f *fakeModelManager) CurrentSessionName() string {
+	if f.sessionName == "" {
+		return "test-session"
+	}
+	return f.sessionName
+}
+
 func (f *fakeModelManager) Info() string {
 	return "provider: " + f.CurrentProvider() + "\nmodel: " + f.model + "\nthinking: " + f.thinking
 }
@@ -236,6 +561,28 @@ func (f *fakeModelManager) SetThinking(level string) (string, error) {
 	return f.Info(), nil
 }
 
+func (f *fakeModelManager) Save() error {
+	f.saved = true
+	return nil
+}
+
+func (f *fakeModelManager) Resume(id string) error {
+	return nil
+}
+
+func (f *fakeModelManager) Rename(name string) (string, error) {
+	f.sessionName = name
+	return "Renamed.", nil
+}
+
+func (f *fakeModelManager) Sessions() (string, error) {
+	return "No sessions.", nil
+}
+
+func (f *fakeModelManager) NewSession() (string, error) {
+	return "New session.", nil
+}
+
 func TestParseThemeMode(t *testing.T) {
 	for input, want := range map[string]ThemeMode{
 		"":      ThemeLight,
@@ -257,6 +604,115 @@ func TestParseThemeMode(t *testing.T) {
 	}
 }
 
+func requireStyleColor(t *testing.T, name string, got lipgloss.TerminalColor, want string) {
+	t.Helper()
+	color, ok := got.(lipgloss.Color)
+	if !ok {
+		t.Fatalf("%s = %T, want lipgloss.Color", name, got)
+	}
+	if string(color) != want {
+		t.Fatalf("%s = %q, want %q", name, string(color), want)
+	}
+}
+
+func requireNoStyleColor(t *testing.T, name string, got lipgloss.TerminalColor) {
+	t.Helper()
+	if _, ok := got.(lipgloss.NoColor); !ok {
+		t.Fatalf("%s should not paint a color, got %T", name, got)
+	}
+}
+
+func TestDarkThemePaintsAppBackground(t *testing.T) {
+	m := newModel(context.Background(), nil, Options{ModelName: "test-model", ThemeMode: ThemeDark})
+	next, _ := m.Update(tea.WindowSizeMsg{Width: 100, Height: 30})
+	m = next.(model)
+	rendered := stripANSI(m.View())
+	if !strings.Contains(rendered, "theme dark") {
+		t.Fatalf("rendered view missing 'theme dark':\n%s", rendered)
+	}
+	if m.theme.Mode != ThemeDark {
+		t.Fatalf("theme.Mode = %q, want %q", m.theme.Mode, ThemeDark)
+	}
+	want := darkTheme()
+	requireStyleColor(t, "App.Background", m.styles.App.GetBackground(), want.Background)
+	requireStyleColor(t, "App.Foreground", m.styles.App.GetForeground(), want.Text)
+}
+
+func TestDarkThemeKeepsComponentBackgroundsLightweight(t *testing.T) {
+	m := newModel(context.Background(), nil, Options{ModelName: "test-model", ThemeMode: ThemeDark})
+	next, _ := m.Update(tea.WindowSizeMsg{Width: 100, Height: 30})
+	m = next.(model)
+	requireNoStyleColor(t, "Viewport.Background", m.styles.Viewport.GetBackground())
+	requireNoStyleColor(t, "InputBox.Background", m.styles.InputBox.GetBackground())
+	requireNoStyleColor(t, "AssistantMsg.Background", m.styles.AssistantMsg.GetBackground())
+	requireNoStyleColor(t, "Activity.Background", m.styles.Activity.GetBackground())
+	m.running = true
+	m.status = "thinking..."
+	m.resize()
+	if backgroundPattern.MatchString(m.activityView()) {
+		t.Fatalf("activity view should not paint a background:\n%q", m.activityView())
+	}
+}
+
+func TestDarkThemeStylesUseDarkPalette(t *testing.T) {
+	m := newModel(context.Background(), nil, Options{ModelName: "test-model", ThemeMode: ThemeDark})
+	want := darkTheme()
+	requireStyleColor(t, "Header.Foreground", m.styles.Header.GetForeground(), want.Title)
+	requireStyleColor(t, "Muted.Foreground", m.styles.Muted.GetForeground(), want.Muted)
+	requireStyleColor(t, "Pill.Background", m.styles.Pill.GetBackground(), want.SurfaceAlt)
+	requireStyleColor(t, "input.FocusedStyle.Base.Foreground", m.input.FocusedStyle.Base.GetForeground(), want.Text)
+	requireStyleColor(t, "input.FocusedStyle.Placeholder.Foreground", m.input.FocusedStyle.Placeholder.GetForeground(), want.Muted)
+}
+
+func TestModelCommandShowsDarkTheme(t *testing.T) {
+	m := newModel(context.Background(), nil, Options{ModelName: "test-model", ThemeMode: ThemeDark})
+	m.input.SetValue("/model")
+	next, _ := m.submit()
+	rendered := next.(model).renderMessages()
+	if !strings.Contains(rendered, "theme: dark") {
+		t.Fatalf("model command output missing 'theme: dark':\n%s", rendered)
+	}
+}
+
+func TestThemeSlashCommandSwitchesAtRuntime(t *testing.T) {
+	m := newModel(context.Background(), nil, Options{ModelName: "test-model"})
+	if m.theme.Mode != ThemeLight {
+		t.Fatalf("initial theme = %q, want light", m.theme.Mode)
+	}
+	requireNoStyleColor(t, "initial App.Background", m.styles.App.GetBackground())
+
+	// Switch to dark
+	m.input.SetValue("/theme dark")
+	next, _ := m.submit()
+	m = next.(model)
+	if m.theme.Mode != ThemeDark {
+		t.Fatalf("after /theme dark: Mode = %q, want dark", m.theme.Mode)
+	}
+	requireStyleColor(t, "dark App.Background", m.styles.App.GetBackground(), darkTheme().Background)
+	last := m.messages[len(m.messages)-1]
+	if !strings.Contains(last.content, "theme: dark") {
+		t.Fatalf("expected confirmation message, got %q", last.content)
+	}
+
+	// Switch back to light
+	m.input.SetValue("/theme light")
+	next, _ = m.submit()
+	m = next.(model)
+	if m.theme.Mode != ThemeLight {
+		t.Fatalf("after /theme light: Mode = %q, want light", m.theme.Mode)
+	}
+	requireNoStyleColor(t, "restored App.Background", m.styles.App.GetBackground())
+
+	// No arg shows usage
+	m.input.SetValue("/theme")
+	next, _ = m.submit()
+	m = next.(model)
+	last = m.messages[len(m.messages)-1]
+	if !strings.Contains(last.content, "usage: /theme") {
+		t.Fatalf("expected usage hint, got %q", last.content)
+	}
+}
+
 func TestCleanToolLog(t *testing.T) {
 	got := cleanToolLog("\n[tool] shell_run {\"command\":\"pwd\"}\n")
 	want := `shell_run {"command":"pwd"}`

internal/tui/markdown.go

@@ -2,6 +2,8 @@ package tui
 
 import (
 	"strings"
+	"unicode"
+	"unicode/utf8"
 
 	"github.com/charmbracelet/glamour"
 	"github.com/charmbracelet/glamour/ansi"
@@ -24,7 +26,7 @@ func renderMarkdown(content string, width int, th theme) string {
 	if err != nil {
 		return content
 	}
-	return strings.TrimRight(rendered, "\n")
+	return trimRenderedMarkdown(rendered)
 }
 
 func markdownStyle(th theme) ansi.StyleConfig {
@@ -32,6 +34,7 @@ func markdownStyle(th theme) ansi.StyleConfig {
 	if th.Mode == ThemeDark {
 		cfg = glamstyles.DarkStyleConfig
 	}
+	clearMarkdownBackgrounds(&cfg)
 
 	zero := uint(0)
 	cfg.Document.Margin = &zero
@@ -51,7 +54,7 @@ func markdownStyle(th theme) ansi.StyleConfig {
 	cfg.LinkText.Color = stringPtr(th.User)
 
 	cfg.Code.Color = stringPtr(th.Tool)
-	cfg.Code.BackgroundColor = stringPtr(th.SurfaceAlt)
+	cfg.Code.BackgroundColor = nil
 	cfg.CodeBlock.Color = stringPtr(th.Text)
 	cfg.CodeBlock.BackgroundColor = nil
 	cfg.CodeBlock.Margin = &zero
@@ -59,6 +62,93 @@ func markdownStyle(th theme) ansi.StyleConfig {
 	return cfg
 }
 
+func clearMarkdownBackgrounds(cfg *ansi.StyleConfig) {
+	cfg.Document.BackgroundColor = nil
+	cfg.BlockQuote.BackgroundColor = nil
+	cfg.Paragraph.BackgroundColor = nil
+	cfg.Heading.BackgroundColor = nil
+	cfg.H1.BackgroundColor = nil
+	cfg.H2.BackgroundColor = nil
+	cfg.H3.BackgroundColor = nil
+	cfg.H4.BackgroundColor = nil
+	cfg.H5.BackgroundColor = nil
+	cfg.H6.BackgroundColor = nil
+	cfg.Text.BackgroundColor = nil
+	cfg.Strikethrough.BackgroundColor = nil
+	cfg.Emph.BackgroundColor = nil
+	cfg.Strong.BackgroundColor = nil
+	cfg.HorizontalRule.BackgroundColor = nil
+	cfg.Item.BackgroundColor = nil
+	cfg.Enumeration.BackgroundColor = nil
+	cfg.Task.BackgroundColor = nil
+	cfg.Link.BackgroundColor = nil
+	cfg.LinkText.BackgroundColor = nil
+	cfg.Image.BackgroundColor = nil
+	cfg.ImageText.BackgroundColor = nil
+	cfg.Code.BackgroundColor = nil
+	cfg.CodeBlock.BackgroundColor = nil
+	cfg.Table.BackgroundColor = nil
+	cfg.DefinitionList.BackgroundColor = nil
+	cfg.DefinitionTerm.BackgroundColor = nil
+	cfg.DefinitionDescription.BackgroundColor = nil
+	cfg.HTMLBlock.BackgroundColor = nil
+	cfg.HTMLSpan.BackgroundColor = nil
+}
+
+func trimRenderedMarkdown(value string) string {
+	lines := strings.Split(strings.TrimRight(value, "\n"), "\n")
+	for i, line := range lines {
+		lines[i] = trimStyledTrailingSpaces(line)
+	}
+	return strings.Join(lines, "\n")
+}
+
+func trimStyledTrailingSpaces(line string) string {
+	cut := 0
+	seenNonSpace := false
+	seenWhitespaceAfterLastNonSpace := false
+	for i := 0; i < len(line); {
+		if end, ok := ansiSequenceEnd(line, i); ok {
+			if seenNonSpace && !seenWhitespaceAfterLastNonSpace {
+				cut = end
+			}
+			i = end
+			continue
+		}
+
+		r, size := utf8.DecodeRuneInString(line[i:])
+		if r == utf8.RuneError && size == 0 {
+			break
+		}
+		end := i + size
+		if unicode.IsSpace(r) {
+			if seenNonSpace {
+				seenWhitespaceAfterLastNonSpace = true
+			}
+		} else {
+			seenNonSpace = true
+			seenWhitespaceAfterLastNonSpace = false
+			cut = end
+		}
+		i = end
+	}
+	if cut == 0 {
+		return ""
+	}
+	return line[:cut]
+}
+
+func ansiSequenceEnd(value string, start int) (int, bool) {
+	if start+2 >= len(value) || value[start] != '\x1b' || value[start+1] != '[' {
+		return 0, false
+	}
+	for i := start + 2; i < len(value); i++ {
+		if value[i] >= 0x40 && value[i] <= 0x7e {
+			return i + 1, true
+		}
+	}
+	return 0, false
+}
 func stringPtr(value string) *string {
 	return &value
 }

internal/tui/markdown_test.go

@@ -37,6 +37,26 @@ func TestRenderMarkdownSupportsThemesAndNarrowWidth(t *testing.T) {
 	}
 }
 
+func TestRenderMarkdownDoesNotEmitBackgroundColors(t *testing.T) {
+	content := "# Title\n\n`inline`\n\n```go\nfmt.Println(\"x\")\n```"
+	for _, mode := range []ThemeMode{ThemeLight, ThemeDark} {
+		rendered := renderMarkdown(content, 80, themeForMode(mode))
+		if backgroundPattern.MatchString(rendered) {
+			t.Fatalf("markdown output for %s should not emit background colors:\n%q", mode, rendered)
+		}
+	}
+}
+
+func TestRenderMarkdownTrimsGlamourTrailingSpaces(t *testing.T) {
+	rendered := renderMarkdown("hello", 80, themeForMode(ThemeLight))
+	plain := stripANSI(rendered)
+	if plain != "hello" {
+		t.Fatalf("markdown should not pad a plain assistant line:\n%q", plain)
+	}
+}
+
+var backgroundPattern = regexp.MustCompile(`\x1b\[[0-9;:]*48[;:][0-9;:]*m`)
+
 var ansiPattern = regexp.MustCompile(`\x1b\[[0-9;:]*[A-Za-z]`)
 
 func stripANSI(value string) string {

internal/tui/theme.go

@@ -15,6 +15,17 @@ const (
 	ThemeDark  ThemeMode = "dark"
 )
 
+const (
+	iconReady   = "✓"
+	iconWorking = "◆"
+	iconTool    = "▣"
+	iconError   = "✕"
+	iconCommand = "⌘"
+	iconSession = "◌"
+	iconContext = "◔"
+	iconModel   = "◉"
+)
+
 func ParseThemeMode(value string) (ThemeMode, error) {
 	switch ThemeMode(strings.ToLower(strings.TrimSpace(value))) {
 	case "", ThemeLight:
@@ -46,17 +57,26 @@ type theme struct {
 
 type styles struct {
 	App           lipgloss.Style
-	Header         lipgloss.Style
-	Title          lipgloss.Style
 	Muted         lipgloss.Style
 	Activity      lipgloss.Style
+	Header        lipgloss.Style
 	ModelMeta     lipgloss.Style
 	Footer        lipgloss.Style
 	Viewport      lipgloss.Style
 	InputBox      lipgloss.Style
 	CommandHint   lipgloss.Style
+	StatusReady   lipgloss.Style
+	StatusWarn    lipgloss.Style
+	StatusError   lipgloss.Style
+	StatusInfo    lipgloss.Style
+	Pill          lipgloss.Style
+	PillAccent    lipgloss.Style
+	Palette       lipgloss.Style
+	PaletteMatch  lipgloss.Style
+	ToolName      lipgloss.Style
+	ToolKey       lipgloss.Style
+	ToolValue     lipgloss.Style
 	UserLabel     lipgloss.Style
-	AssistantLabel lipgloss.Style
 	ToolLabel     lipgloss.Style
 	ErrorLabel    lipgloss.Style
 	SystemLabel   lipgloss.Style
@@ -110,67 +130,103 @@ func darkTheme() theme {
 	}
 }
 
+func appStyle(t theme) lipgloss.Style {
+	style := lipgloss.NewStyle().Foreground(lipgloss.Color(t.Text))
+	if t.Mode == ThemeDark {
+		style = style.Background(lipgloss.Color(t.Background))
+	}
+	return style
+}
+
 func (t theme) styles() styles {
 	return styles{
-		App: lipgloss.NewStyle().
-			Foreground(lipgloss.Color(t.Text)).
-			Background(lipgloss.Color(t.Background)),
-
-		Header: lipgloss.NewStyle().
-			Foreground(lipgloss.Color(t.Text)).
-			Background(lipgloss.Color(t.SurfaceAlt)).
-			Padding(0, 1),
-
-		Title: lipgloss.NewStyle().
-			Bold(true).
-			Foreground(lipgloss.Color(t.Title)),
+		App: appStyle(t),
 
 		Muted: lipgloss.NewStyle().
 			Foreground(lipgloss.Color(t.Muted)),
 
 		Activity: lipgloss.NewStyle().
-			Bold(true).
-			Foreground(lipgloss.Color(t.Background)).
-			Background(lipgloss.Color(t.Title)).
+			Foreground(lipgloss.Color(t.Muted)).
 			Padding(0, 1),
 
+		Header: lipgloss.NewStyle().
+			Bold(true).
+			Foreground(lipgloss.Color(t.Title)),
+
 		ModelMeta: lipgloss.NewStyle().
-			Foreground(lipgloss.Color(t.Text)).
-			Background(lipgloss.Color(t.SurfaceAlt)).
+			Foreground(lipgloss.Color(t.Muted)).
 			Padding(0, 1),
 
 		Footer: lipgloss.NewStyle().
 			Foreground(lipgloss.Color(t.Muted)).
-			Background(lipgloss.Color(t.Background)).
 			Padding(0, 1),
 
 		Viewport: lipgloss.NewStyle().
-			Foreground(lipgloss.Color(t.Text)).
-			Background(lipgloss.Color(t.Background)),
+			Foreground(lipgloss.Color(t.Text)),
 
 		InputBox: lipgloss.NewStyle().
 			Foreground(lipgloss.Color(t.Text)).
-			Background(lipgloss.Color(t.Surface)).
-			Border(lipgloss.NormalBorder()).
+			Border(lipgloss.ThickBorder()).
 			BorderForeground(lipgloss.Color(t.Border)).
-			Padding(0, 1),
+			Padding(0, 2),
 
 		CommandHint: lipgloss.NewStyle().
 			Foreground(lipgloss.Color(t.Muted)).
+			Padding(0, 1),
+
+		StatusReady: lipgloss.NewStyle().
+			Foreground(lipgloss.Color(t.Assistant)),
+
+		StatusWarn: lipgloss.NewStyle().
+			Foreground(lipgloss.Color(t.Tool)),
+
+		StatusError: lipgloss.NewStyle().
+			Foreground(lipgloss.Color(t.Error)),
+
+		StatusInfo: lipgloss.NewStyle().
+			Foreground(lipgloss.Color(t.User)),
+
+		Pill: lipgloss.NewStyle().
+			Foreground(lipgloss.Color(t.Text)).
 			Background(lipgloss.Color(t.SurfaceAlt)).
 			Padding(0, 1),
 
+		PillAccent: lipgloss.NewStyle().
+			Bold(true).
+			Foreground(lipgloss.Color(t.Title)).
+			Background(lipgloss.Color(t.SurfaceAlt)).
+			Padding(0, 1),
+
+		Palette: lipgloss.NewStyle().
+			Foreground(lipgloss.Color(t.Text)).
+			Border(lipgloss.ThickBorder()).
+			BorderForeground(lipgloss.Color(t.Border)).
+			Padding(0, 2),
+
+		PaletteMatch: lipgloss.NewStyle().
+			Bold(true).
+			Foreground(lipgloss.Color(t.User)),
+
+		ToolName: lipgloss.NewStyle().
+			Bold(true).
+			Foreground(lipgloss.Color(t.Tool)),
+
+		ToolKey: lipgloss.NewStyle().
+			Foreground(lipgloss.Color(t.Muted)),
+
+		ToolValue: lipgloss.NewStyle().
+			Foreground(lipgloss.Color(t.Text)),
+
 		UserLabel:   labelStyle(t.User),
-		AssistantLabel: labelStyle(t.Assistant),
 		ToolLabel:   labelStyle(t.Tool),
 		ErrorLabel:  labelStyle(t.Error),
 		SystemLabel: labelStyle(t.System),
 
-		UserMessage:   messageStyle(t.Text, t.User),
-		AssistantMsg:  messageStyle(t.Text, t.Assistant),
-		ToolMessage:   messageStyle(t.Text, t.Tool),
-		ErrorMessage:  messageStyle(t.Error, t.Error),
-		SystemMessage: lipgloss.NewStyle().Foreground(lipgloss.Color(t.System)).PaddingLeft(1),
+		UserMessage:   bubbleMessageStyle(t.Text, t.SurfaceAlt),
+		AssistantMsg:  assistantMessageStyle(t.Text),
+		ToolMessage:   messageStyle(t.Text),
+		ErrorMessage:  messageStyle(t.Error),
+		SystemMessage: lipgloss.NewStyle().Foreground(lipgloss.Color(t.System)).Padding(0, 1),
 	}
 }
 
@@ -180,21 +236,30 @@ func labelStyle(color string) lipgloss.Style {
 		Foreground(lipgloss.Color(color))
 }
 
-func messageStyle(textColor, accentColor string) lipgloss.Style {
+func messageStyle(textColor string) lipgloss.Style {
 	return lipgloss.NewStyle().
 		Foreground(lipgloss.Color(textColor)).
-		Border(lipgloss.ThickBorder(), false, false, false, true).
-		BorderForeground(lipgloss.Color(accentColor)).
-		PaddingLeft(1)
+		Padding(0, 1)
+}
+
+func assistantMessageStyle(textColor string) lipgloss.Style {
+	return lipgloss.NewStyle().
+		Foreground(lipgloss.Color(textColor)).
+		Padding(1, 2)
+}
+
+func bubbleMessageStyle(textColor, backgroundColor string) lipgloss.Style {
+	return lipgloss.NewStyle().
+		Foreground(lipgloss.Color(textColor)).
+		Background(lipgloss.Color(backgroundColor)).
+		Padding(1, 2)
 }
 
 func applyTextareaTheme(input *textarea.Model, t theme) {
 	focused := textarea.Style{
 		Base: lipgloss.NewStyle().
-			Foreground(lipgloss.Color(t.Text)).
-			Background(lipgloss.Color(t.Surface)),
-		CursorLine: lipgloss.NewStyle().
-			Background(lipgloss.Color(t.Surface)),
+			Foreground(lipgloss.Color(t.Text)),
+		CursorLine: lipgloss.NewStyle(),
 		Placeholder: lipgloss.NewStyle().
 			Foreground(lipgloss.Color(t.Muted)),
 		Prompt: lipgloss.NewStyle().
@@ -205,7 +270,7 @@ func applyTextareaTheme(input *textarea.Model, t theme) {
 			Foreground(lipgloss.Color(t.Surface)),
 	}
 	blurred := focused
-	blurred.CursorLine = lipgloss.NewStyle().Background(lipgloss.Color(t.Surface))
+	blurred.CursorLine = lipgloss.NewStyle()
 
 	input.FocusedStyle = focused
 	input.BlurredStyle = blurred

internal/tui/workflow.go

@@ -0,0 +1,143 @@
+package tui
+
+import (
+	"bytes"
+	"context"
+	"errors"
+	"fmt"
+	"os"
+	"os/exec"
+	"path/filepath"
+	"strings"
+	"time"
+
+	"agentu/internal/tools"
+)
+
+const workflowCommandTimeout = 2 * time.Minute
+
+func (m model) handleStatusCommand(args []string) (string, error) {
+	if len(args) != 0 {
+		return "", fmt.Errorf("usage: /status")
+	}
+	out, err := m.runGitCommand("status", "--short")
+	if err != nil {
+		return out, err
+	}
+	if strings.TrimSpace(out) == "" {
+		return "Working tree clean.", nil
+	}
+	return "Changed files:\n" + out, nil
+}
+
+func (m model) handleDiffCommand(args []string) (string, error) {
+	if len(args) > 0 && args[0] == "--stat" {
+		out, err := m.runGitCommand(append([]string{"diff", "--stat", "--"}, args[1:]...)...)
+		if err != nil {
+			return out, err
+		}
+		if strings.TrimSpace(out) == "" {
+			return "No unstaged diff.", nil
+		}
+		return out, nil
+	}
+	out, err := m.runGitCommand(append([]string{"diff", "--"}, args...)...)
+	if err != nil {
+		return out, err
+	}
+	if strings.TrimSpace(out) == "" {
+		return "No unstaged diff.", nil
+	}
+	return out, nil
+}
+
+func (m model) handleTestCommand(args []string) (string, error) {
+	command := strings.TrimSpace(strings.Join(args, " "))
+	if command == "" {
+		var err error
+		command, err = defaultTestCommand(m.workflowDir())
+		if err != nil {
+			return "", err
+		}
+	}
+	out, err := m.runShellCommand(command)
+	if err != nil {
+		if strings.TrimSpace(out) != "" {
+			return out, err
+		}
+		return "", err
+	}
+	if strings.TrimSpace(out) == "" {
+		return fmt.Sprintf("$ %s\n(no output)", command), nil
+	}
+	return out, nil
+}
+
+func (m model) runGitCommand(args ...string) (string, error) {
+	ctx, cancel := context.WithTimeout(m.ctx, workflowCommandTimeout)
+	defer cancel()
+
+	cmd := exec.CommandContext(ctx, "git", append([]string{"-C", m.workflowDir()}, args...)...)
+	output, err := cmd.CombinedOutput()
+	text := tools.FormatResult(string(output))
+	if errors.Is(ctx.Err(), context.DeadlineExceeded) {
+		return text, fmt.Errorf("git command timed out: %w", ctx.Err())
+	}
+	if err != nil {
+		if strings.TrimSpace(text) != "" {
+			return text, fmt.Errorf("git command failed: %w", err)
+		}
+		return "", fmt.Errorf("git command failed: %w", err)
+	}
+	return text, nil
+}
+
+func (m model) runShellCommand(command string) (string, error) {
+	ctx, cancel := context.WithTimeout(m.ctx, workflowCommandTimeout)
+	defer cancel()
+
+	shell := os.Getenv("SHELL")
+	if shell == "" {
+		shell = "/bin/sh"
+	}
+	cmd := exec.CommandContext(ctx, shell, "-lc", command)
+	cmd.Dir = m.workflowDir()
+
+	var combined bytes.Buffer
+	cmd.Stdout = &combined
+	cmd.Stderr = &combined
+	err := cmd.Run()
+	text := strings.TrimRight(tools.FormatResult(combined.String()), "\n")
+	if text != "" {
+		text = fmt.Sprintf("$ %s\n%s", command, text)
+	}
+	if errors.Is(ctx.Err(), context.DeadlineExceeded) {
+		return text, fmt.Errorf("command timed out: %w", ctx.Err())
+	}
+	if err != nil {
+		return text, fmt.Errorf("command failed: %w", err)
+	}
+	return text, nil
+}
+
+func (m model) workflowDir() string {
+	if strings.TrimSpace(m.workingDir) == "" {
+		return "."
+	}
+	return m.workingDir
+}
+
+func defaultTestCommand(dir string) (string, error) {
+	if fileExists(filepath.Join(dir, "go.mod")) {
+		return "go test ./...", nil
+	}
+	if fileExists(filepath.Join(dir, "package.json")) {
+		return "npm test", nil
+	}
+	return "", fmt.Errorf("no default test command found; use /test <command>")
+}
+
+func fileExists(path string) bool {
+	info, err := os.Stat(path)
+	return err == nil && !info.IsDir()
+}

main.go

@@ -0,0 +1,228 @@
+package main
+
+import (
+	"bufio"
+	"context"
+	"errors"
+	"flag"
+	"fmt"
+	"net/http"
+	"os"
+	"os/signal"
+	"path/filepath"
+	"strings"
+
+	"agentu/internal/agent"
+	"agentu/internal/session"
+	"agentu/internal/tools"
+	"agentu/internal/tui"
+	"agentu/pkg/config"
+	"agentu/pkg/llm"
+)
+
+const defaultConfigContent = `# agentu configuration
+# Replace the placeholder values below with your actual provider settings.
+# See agentu.example.yaml for the full reference.
+providers:
+  default:
+    base_url: https://api.openai.com
+    api_key: YOUR_API_KEY_HERE
+    models:
+      - id: gpt-4o
+        context: 128000
+      - id: gpt-4o-mini
+        context: 128000
+`
+
+func main() {
+	if err := run(); err != nil {
+		fmt.Fprintln(os.Stderr, "agentu:", err)
+		os.Exit(1)
+	}
+}
+
+func run() error {
+	configPath := flag.String("config", config.DefaultPath, "path to YAML config")
+	yolo := flag.Bool("yolo", false, "also enable automatic file writes and shell execution")
+	plain := flag.Bool("plain", false, "use the simple line-based REPL instead of the TUI")
+	themeName := flag.String("theme", "light", "TUI theme: light or dark")
+	resumeID := flag.String("resume", "", "resume a specific session by ID")
+	flag.Parse()
+
+	themeMode, err := tui.ParseThemeMode(*themeName)
+	if err != nil {
+		return err
+	}
+
+	cfg, err := config.Load(*configPath)
+	if err != nil {
+		if errors.Is(err, os.ErrNotExist) {
+			if bootstrapErr := bootstrapConfig(*configPath); bootstrapErr != nil {
+				return bootstrapErr
+			}
+			return fmt.Errorf("created default config at %s; edit it to set your API key and provider, then restart agentu", *configPath)
+		}
+		return err
+	}
+
+	providerName := cfg.DefaultProviderName()
+	providerConfig := cfg.Providers[providerName]
+	provider := llm.NewOpenAICompatibleClient(providerConfig.BaseURL, providerConfig.APIKey, http.DefaultClient)
+
+	var registry *tools.Registry
+	if *yolo {
+		registry = tools.Builtins(cfg.Agent.WorkingDir)
+	} else {
+		registry = tools.ReadOnlyBuiltins(cfg.Agent.WorkingDir)
+	}
+	registry.Register(tools.NewBuiltinSearchTool(http.DefaultClient))
+	registry.Register(tools.NewBuiltinFetchTool(http.DefaultClient))
+	systemPrompt := strings.TrimSpace(cfg.Agent.SystemPrompt + "\n\n" + tools.AgentInstructions(cfg.Agent.WorkingDir, *yolo))
+	maxContextTokens := cfg.Agent.MaxContextTokens
+	if maxContextTokens == 0 {
+		maxContextTokens = providerConfig.ContextTokens
+	}
+
+	assistant := agent.New(agent.Options{
+		Provider:         provider,
+		ProviderName:     providerName,
+		Model:            providerConfig.Model,
+		Thinking:         providerConfig.Thinking,
+		ThinkingParam:    providerConfig.ThinkingParam,
+		ThinkingEnabled:  providerConfig.ThinkingConfigured,
+		SystemPrompt:     systemPrompt,
+		ToolRegistry:     registry,
+		ToolTimeout:      cfg.Agent.ToolTimeout,
+		MaxContextTokens: maxContextTokens,
+	})
+
+	// Initialize session store
+	home, err := os.UserHomeDir()
+	if err != nil {
+		return fmt.Errorf("resolve home dir: %w", err)
+	}
+	sessionsDir := filepath.Join(home, ".agentu", "sessions")
+	store, err := session.NewStore(sessionsDir)
+	if err != nil {
+		return fmt.Errorf("init session store: %w", err)
+	}
+
+	modelManager, err := session.NewManager(cfg, assistant, http.DefaultClient, store)
+	if err != nil {
+		return err
+	}
+
+	if err := initializeSession(modelManager, *resumeID); err != nil {
+		return err
+	}
+
+	ctx, stop := signal.NotifyContext(context.Background(), os.Interrupt)
+	defer stop()
+
+	if !*plain {
+		err := tui.Run(ctx, assistant, tui.Options{
+			ModelName:    providerConfig.Model,
+			Yolo:         *yolo,
+			ThemeMode:    themeMode,
+			ModelManager: modelManager,
+			WorkingDir:   cfg.Agent.WorkingDir,
+		})
+		printExitSession(modelManager)
+		return err
+	}
+
+	return repl(ctx, assistant, modelManager)
+}
+
+func initializeSession(modelManager *session.Manager, resumeID string) error {
+	if resumeID != "" {
+		if err := modelManager.Resume(resumeID); err != nil {
+			return fmt.Errorf("resume session: %w", err)
+		}
+		return nil
+	}
+	if _, err := modelManager.NewSession(); err != nil {
+		return fmt.Errorf("start new session: %w", err)
+	}
+	return nil
+}
+
+func bootstrapConfig(configPath string) error {
+	resolved, err := config.ResolvePath(configPath)
+	if err != nil {
+		return fmt.Errorf("resolve config path: %w", err)
+	}
+	dir := filepath.Dir(resolved)
+	if err := os.MkdirAll(dir, 0o755); err != nil {
+		return fmt.Errorf("create config directory: %w", err)
+	}
+	if err := os.WriteFile(resolved, []byte(defaultConfigContent), 0o644); err != nil {
+		return fmt.Errorf("write default config: %w", err)
+	}
+	return nil
+}
+
+func repl(ctx context.Context, assistant *agent.Agent, modelManager *session.Manager) error {
+	scanner := bufio.NewScanner(os.Stdin)
+	scanner.Buffer(make([]byte, 0, 64*1024), 1024*1024)
+
+	fmt.Println("agentu")
+	fmt.Println("Type /exit to quit, /clear to reset context, /compact to compress context.")
+	for {
+		fmt.Print("> ")
+		if !scanner.Scan() {
+			if err := scanner.Err(); err != nil {
+				return err
+			}
+			fmt.Println()
+			printExitSession(modelManager)
+			return nil
+		}
+
+		input := strings.TrimSpace(scanner.Text())
+		switch input {
+		case "":
+			continue
+		case "/exit", "/quit":
+			printExitSession(modelManager)
+			return nil
+		case "/clear":
+			assistant.Clear()
+			fmt.Println("context cleared")
+			continue
+		case "/compact":
+			if err := assistant.Compact(ctx); err != nil {
+				fmt.Fprintln(os.Stderr, "compact error:", err)
+			} else {
+				fmt.Println("context compacted")
+				if modelManager != nil {
+					_ = modelManager.Save()
+				}
+			}
+			continue
+		}
+
+		if err := assistant.RunTurn(ctx, input, os.Stdout, os.Stderr); err != nil {
+			fmt.Fprintln(os.Stderr, "\nerror:", err)
+		}
+		_ = modelManager.Save()
+		fmt.Println()
+	}
+}
+
+func printExitSession(m *session.Manager) {
+	if m == nil {
+		return
+	}
+	_ = m.Save()
+	id := m.CurrentSessionID()
+	name := m.CurrentSessionName()
+	if id != "" {
+		if name != "" && name != id {
+			fmt.Fprintf(os.Stderr, "session: %s (%s)\n", id, name)
+		} else {
+			fmt.Fprintf(os.Stderr, "session: %s\n", id)
+		}
+		fmt.Fprintf(os.Stderr, "resume:  agentu --resume %s\n", id)
+	}
+}

main_test.go

@@ -0,0 +1,101 @@
+package main
+
+import (
+	"context"
+	"path/filepath"
+	"testing"
+
+	"agentu/internal/agent"
+	"agentu/internal/session"
+	"agentu/pkg/config"
+	"agentu/pkg/llm"
+)
+
+type startupProvider struct{}
+
+func (p *startupProvider) ChatStream(ctx context.Context, req llm.ChatRequest, emit func(llm.StreamEvent) error) error {
+	return emit(llm.StreamEvent{Content: "ok"})
+}
+
+func startupTestManager(t *testing.T) (*session.Manager, *agent.Agent) {
+	t.Helper()
+	cfg := &config.Config{
+		Providers: map[string]config.ProviderConfig{
+			"one": {
+				Name:          "one",
+				BaseURL:       "https://one.example.com",
+				APIKey:        "sk-test",
+				Model:         "model-a",
+				Models:        []string{"model-a"},
+				Thinking:      "none",
+				ThinkingParam: "thinking",
+			},
+		},
+	}
+	assistant := agent.New(agent.Options{
+		Provider:     &startupProvider{},
+		ProviderName: "one",
+		Model:        "model-a",
+	})
+	store, err := session.NewStore(filepath.Join(t.TempDir(), "sessions"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	manager, err := session.NewManager(cfg, assistant, nil, store)
+	if err != nil {
+		t.Fatal(err)
+	}
+	return manager, assistant
+}
+
+func TestInitializeSessionStartsNewByDefault(t *testing.T) {
+	manager, assistant := startupTestManager(t)
+	if _, err := manager.NewSession(); err != nil {
+		t.Fatalf("new session: %v", err)
+	}
+	savedID := manager.CurrentSessionID()
+	assistant.SetMessages([]llm.Message{{Role: llm.RoleUser, Content: "old context"}})
+	if err := manager.Save(); err != nil {
+		t.Fatalf("save: %v", err)
+	}
+
+	if err := initializeSession(manager, ""); err != nil {
+		t.Fatalf("initialize: %v", err)
+	}
+	if manager.CurrentSessionID() == savedID {
+		t.Fatalf("default startup resumed latest session %s", savedID)
+	}
+	for _, msg := range assistant.Messages() {
+		if msg.Content == "old context" {
+			t.Fatal("default startup carried old context into new session")
+		}
+	}
+}
+
+func TestInitializeSessionResumesExplicitID(t *testing.T) {
+	manager, assistant := startupTestManager(t)
+	if _, err := manager.NewSession(); err != nil {
+		t.Fatalf("new session: %v", err)
+	}
+	savedID := manager.CurrentSessionID()
+	assistant.SetMessages([]llm.Message{{Role: llm.RoleUser, Content: "old context"}})
+	if err := manager.Save(); err != nil {
+		t.Fatalf("save: %v", err)
+	}
+
+	if err := initializeSession(manager, savedID); err != nil {
+		t.Fatalf("initialize: %v", err)
+	}
+	if manager.CurrentSessionID() != savedID {
+		t.Fatalf("expected resumed session %s, got %s", savedID, manager.CurrentSessionID())
+	}
+	found := false
+	for _, msg := range assistant.Messages() {
+		if msg.Content == "old context" {
+			found = true
+		}
+	}
+	if !found {
+		t.Fatal("explicit resume did not restore saved context")
+	}
+}

pkg/config/config.go

@@ -21,19 +21,28 @@ type Config struct {
 
 type ProviderConfig struct {
 	Name               string
-	BaseURL            string   `yaml:"base_url"`
-	Model              string   `yaml:"model"`
-	Models             []string `yaml:"models"`
-	APIKey             string   `yaml:"api_key"`
-	Thinking           string   `yaml:"thinking"`
-	ThinkingParam      string   `yaml:"thinking_param"`
+	BaseURL            string
+	Model              string
+	Models             []string
+	ModelConfigs       map[string]ModelConfig
+	ContextTokens      int
+	APIKey             string
+	Thinking           string
+	ThinkingParam      string
 	ThinkingConfigured bool
 }
 
+type ModelConfig struct {
+	ID            string `yaml:"id"`
+	Thinking      string `yaml:"thinking"`
+	ContextTokens int    `yaml:"context"`
+}
+
 type AgentConfig struct {
 	SystemPrompt     string        `yaml:"system_prompt"`
 	WorkingDir       string        `yaml:"working_dir"`
 	ToolTimeout      time.Duration `yaml:"tool_timeout"`
+	MaxContextTokens int           `yaml:"max_context_tokens"`
 }
 
 type rawConfig struct {
@@ -42,16 +51,20 @@ type rawConfig struct {
 		SystemPrompt     string `yaml:"system_prompt"`
 		WorkingDir       string `yaml:"working_dir"`
 		ToolTimeout      string `yaml:"tool_timeout"`
+		MaxContextTokens int    `yaml:"max_context_tokens"`
 	} `yaml:"agent"`
 }
 
 type rawProviderConfig struct {
 	BaseURL string           `yaml:"base_url"`
-	Model         string   `yaml:"model"`
-	Models        []string `yaml:"models"`
+	Models  []rawModelConfig `yaml:"models"`
 	APIKey  string           `yaml:"api_key"`
-	Thinking      *string  `yaml:"thinking"`
-	ThinkingParam string   `yaml:"thinking_param"`
+}
+
+type rawModelConfig struct {
+	ID            string `yaml:"id"`
+	Thinking      string `yaml:"thinking"`
+	ContextTokens int    `yaml:"context"`
 }
 
 func Load(path string) (*Config, error) {
@@ -78,6 +91,7 @@ func Load(path string) (*Config, error) {
 		Agent: AgentConfig{
 			SystemPrompt:     raw.Agent.SystemPrompt,
 			WorkingDir:       raw.Agent.WorkingDir,
+			MaxContextTokens: raw.Agent.MaxContextTokens,
 		},
 	}
 
@@ -142,8 +156,8 @@ func (c *Config) Validate() error {
 		if strings.TrimSpace(provider.BaseURL) == "" {
 			missing = append(missing, prefix+".base_url")
 		}
-		if strings.TrimSpace(provider.Model) == "" {
-			missing = append(missing, prefix+".model")
+		if len(provider.Models) == 0 {
+			missing = append(missing, prefix+".models")
 		}
 		if strings.TrimSpace(provider.APIKey) == "" {
 			missing = append(missing, prefix+".api_key")
@@ -151,6 +165,19 @@ func (c *Config) Validate() error {
 		if len(missing) > 0 {
 			return fmt.Errorf("missing required config: %s", strings.Join(missing, ", "))
 		}
+		for _, model := range provider.Models {
+			if strings.TrimSpace(model) == "" {
+				return fmt.Errorf("%s.models.id is required", prefix)
+			}
+		}
+		for _, model := range provider.ModelConfigs {
+			if model.ContextTokens < 0 {
+				return fmt.Errorf("%s.models.context must be greater than or equal to zero", prefix)
+			}
+			if model.Thinking != "" && !IsThinkingLevel(model.Thinking) {
+				return fmt.Errorf("%s.models.thinking must be one of: %s", prefix, strings.Join(ThinkingLevels(), ", "))
+			}
+		}
 		if provider.ThinkingConfigured && !IsThinkingLevel(provider.Thinking) {
 			return fmt.Errorf("%s.thinking must be one of: %s", prefix, strings.Join(ThinkingLevels(), ", "))
 		}
@@ -158,6 +185,9 @@ func (c *Config) Validate() error {
 	if c.Agent.ToolTimeout <= 0 {
 		return errors.New("agent.tool_timeout must be greater than zero")
 	}
+	if c.Agent.MaxContextTokens < 0 {
+		return errors.New("agent.max_context_tokens must be greater than or equal to zero")
+	}
 	return nil
 }
 
@@ -187,27 +217,59 @@ func normalizeProviders(raw rawConfig) map[string]ProviderConfig {
 }
 
 func normalizeProvider(name string, raw rawProviderConfig) ProviderConfig {
+	models, modelConfigs := normalizeModelConfigs(raw.Models)
 	provider := ProviderConfig{
 		Name:         name,
 		BaseURL:      raw.BaseURL,
-		Model:         raw.Model,
-		Models:        append([]string(nil), raw.Models...),
+		Models:       models,
+		ModelConfigs: modelConfigs,
 		APIKey:       raw.APIKey,
-		ThinkingParam: raw.ThinkingParam,
 	}
-	if provider.Model == "" && len(provider.Models) > 0 {
+	if len(provider.Models) > 0 {
 		provider.Model = provider.Models[0]
 	}
-	if raw.Thinking != nil {
-		provider.Thinking = strings.ToLower(strings.TrimSpace(*raw.Thinking))
-		provider.ThinkingConfigured = true
-		if provider.ThinkingParam == "" {
+	applyModelDefaults(&provider)
+	if provider.ThinkingConfigured && provider.ThinkingParam == "" {
 		provider.ThinkingParam = "thinking"
 	}
-	}
 	return provider
 }
 
+func normalizeModelConfigs(rawModels []rawModelConfig) ([]string, map[string]ModelConfig) {
+	models := make([]string, 0, len(rawModels))
+	configs := make(map[string]ModelConfig, len(rawModels))
+	for _, raw := range rawModels {
+		model := ModelConfig{
+			ID:            strings.TrimSpace(raw.ID),
+			Thinking:      strings.ToLower(strings.TrimSpace(raw.Thinking)),
+			ContextTokens: raw.ContextTokens,
+		}
+		models = append(models, model.ID)
+		if model.ID != "" {
+			configs[model.ID] = model
+		}
+	}
+	return models, configs
+}
+
+func applyModelDefaults(provider *ProviderConfig) {
+	model, ok := provider.ModelConfigs[provider.Model]
+	if !ok {
+		return
+	}
+	provider.ContextTokens = model.ContextTokens
+	if model.Thinking != "" {
+		provider.Thinking = model.Thinking
+		provider.ThinkingConfigured = true
+	}
+}
+
+func (p ProviderConfig) WithModel(model string) ProviderConfig {
+	p.Model = strings.TrimSpace(model)
+	applyModelDefaults(&p)
+	return p
+}
+
 func ThinkingLevels() []string {
 	return []string{"none", "middle", "high", "xhigh", "max"}
 }

pkg/config/config_test.go

@@ -16,8 +16,9 @@ func TestLoadExpandsEnvAndDefaults(t *testing.T) {
 providers:
   test:
     base_url: https://example.com
-    model: test-model
     api_key: ${AGENTU_TEST_KEY}
+    models:
+      - id: test-model
 agent:
   working_dir: .
 `), 0o644); err != nil {
@@ -43,6 +44,115 @@ agent:
 	}
 }
 
+func TestLoadMinimalModelObjectConfig(t *testing.T) {
+	dir := t.TempDir()
+	t.Setenv("AGENTU_TEST_KEY", "sk-test")
+	path := filepath.Join(dir, "agentu.yaml")
+	if err := os.WriteFile(path, []byte(`
+providers:
+  test:
+    base_url: https://example.com
+    api_key: ${AGENTU_TEST_KEY}
+    models:
+      - id: model-a
+        thinking: none
+        context: 256000
+`), 0o644); err != nil {
+		t.Fatal(err)
+	}
+
+	cfg, err := Load(path)
+	if err != nil {
+		t.Fatal(err)
+	}
+	provider := cfg.Providers["test"]
+	if provider.Model != "model-a" {
+		t.Fatalf("model = %q", provider.Model)
+	}
+	if len(provider.Models) != 1 || provider.Models[0] != "model-a" {
+		t.Fatalf("models = %#v", provider.Models)
+	}
+	if provider.ContextTokens != 256000 {
+		t.Fatalf("context tokens = %d", provider.ContextTokens)
+	}
+	if !provider.ThinkingConfigured || provider.Thinking != "none" || provider.ThinkingParam != "thinking" {
+		t.Fatalf("thinking config = %#v", provider)
+	}
+	if cfg.Agent.ToolTimeout != 30*time.Second {
+		t.Fatalf("tool timeout = %s", cfg.Agent.ToolTimeout)
+	}
+	if cfg.Agent.SystemPrompt == "" {
+		t.Fatal("system prompt default was not applied")
+	}
+}
+
+func TestLoadRejectsStringModelEntries(t *testing.T) {
+	dir := t.TempDir()
+	path := filepath.Join(dir, "agentu.yaml")
+	if err := os.WriteFile(path, []byte(`
+providers:
+  test:
+    base_url: https://example.com
+    api_key: sk-test
+    models:
+      - test-model
+`), 0o644); err != nil {
+		t.Fatal(err)
+	}
+
+	_, err := Load(path)
+	if err == nil || !strings.Contains(err.Error(), "cannot unmarshal") {
+		t.Fatalf("err = %v", err)
+	}
+}
+
+func TestLoadMaxContextTokens(t *testing.T) {
+	dir := t.TempDir()
+	path := filepath.Join(dir, "agentu.yaml")
+	if err := os.WriteFile(path, []byte(`
+providers:
+  test:
+    base_url: https://example.com
+    api_key: sk-test
+    models:
+      - id: test-model
+agent:
+  max_context_tokens: 120000
+`), 0o644); err != nil {
+		t.Fatal(err)
+	}
+
+	cfg, err := Load(path)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if cfg.Agent.MaxContextTokens != 120000 {
+		t.Fatalf("max context tokens = %d", cfg.Agent.MaxContextTokens)
+	}
+}
+
+func TestLoadRejectsNegativeMaxContextTokens(t *testing.T) {
+	dir := t.TempDir()
+	path := filepath.Join(dir, "agentu.yaml")
+	if err := os.WriteFile(path, []byte(`
+providers:
+  test:
+    base_url: https://example.com
+    api_key: sk-test
+    models:
+      - id: test-model
+agent:
+  max_context_tokens: -1
+`), 0o644); err != nil {
+		t.Fatal(err)
+	}
+
+	_, err := Load(path)
+	if err == nil || !strings.Contains(err.Error(), "agent.max_context_tokens") {
+		t.Fatalf("err = %v", err)
+	}
+}
+
 func TestLoadUsesDefaultConfigPath(t *testing.T) {
 	dir := t.TempDir()
 	t.Setenv("HOME", dir)
@@ -56,8 +166,9 @@ func TestLoadUsesDefaultConfigPath(t *testing.T) {
 providers:
   test:
     base_url: https://example.com
-    model: test-model
     api_key: ${AGENTU_TEST_KEY}
+    models:
+      - id: test-model
 `), 0o644); err != nil {
 		t.Fatal(err)
 	}
@@ -93,8 +204,9 @@ func TestLoadDefaultSystemPromptDiscouragesEagerTools(t *testing.T) {
 providers:
   test:
     base_url: https://example.com
-    model: test-model
     api_key: ${AGENTU_TEST_KEY}
+    models:
+      - id: test-model
 `), 0o644); err != nil {
 		t.Fatal(err)
 	}
@@ -119,17 +231,15 @@ providers:
   loveuer:
     base_url: https://example.com
     api_key: ${AGENTU_TEST_KEY}
-    model: model-a
     models:
-      - model-a
-      - model-b
+      - id: model-a
         thinking: xhigh
-    thinking_param: thinking
+      - id: model-b
   backup:
     base_url: https://backup.example.com
     api_key: ${AGENTU_TEST_KEY}
     models:
-      - model-c
+      - id: model-c
 `), 0o644); err != nil {
 		t.Fatal(err)
 	}
@@ -162,11 +272,13 @@ providers:
   two:
     base_url: https://two.example.com
     api_key: ${AGENTU_TEST_KEY}
-    model: model-b
+    models:
+      - id: model-b
   one:
     base_url: https://one.example.com
     api_key: ${AGENTU_TEST_KEY}
-    model: model-a
+    models:
+      - id: model-a
 `), 0o644); err != nil {
 		t.Fatal(err)
 	}
@@ -188,8 +300,9 @@ func TestLoadRejectsInvalidThinking(t *testing.T) {
 providers:
   test:
     base_url: https://example.com
-    model: test-model
     api_key: ${AGENTU_TEST_KEY}
+    models:
+      - id: test-model
         thinking: huge
 `), 0o644); err != nil {
 		t.Fatal(err)
@@ -224,8 +337,9 @@ func TestLoadRejectsLegacyProviderConfig(t *testing.T) {
 	if err := os.WriteFile(path, []byte(`
 provider:
   base_url: https://example.com
-  model: test-model
   api_key: sk-test
+  models:
+    - id: test-model
 `), 0o644); err != nil {
 		t.Fatal(err)
 	}
@@ -244,8 +358,9 @@ active_provider: loveuer
 providers:
   loveuer:
     base_url: https://example.com
-    model: test-model
     api_key: sk-test
+    models:
+      - id: test-model
 `), 0o644); err != nil {
 		t.Fatal(err)
 	}

pkg/llm/openai.go

@@ -10,8 +10,13 @@ import (
 	"io"
 	"net/http"
 	"strings"
+	"time"
 )
 
+const maxChatAttempts = 3
+
+const retryDelay = 100 * time.Millisecond
+
 type OpenAICompatibleClient struct {
 	baseURL    string
 	apiKey     string
@@ -31,11 +36,21 @@ func NewOpenAICompatibleClient(baseURL, apiKey string, httpClient *http.Client)
 
 func (c *OpenAICompatibleClient) ChatStream(ctx context.Context, req ChatRequest, emit func(StreamEvent) error) error {
 	req.Stream = true
+	extra := make(map[string]any, len(req.Extra)+1)
+	for key, value := range req.Extra {
+		extra[key] = value
+	}
+	if _, exists := extra["stream_options"]; !exists {
+		extra["stream_options"] = map[string]any{"include_usage": true}
+	}
+	req.Extra = extra
 	body, err := json.Marshal(req)
 	if err != nil {
 		return fmt.Errorf("marshal chat request: %w", err)
 	}
 
+	var lastErr error
+	for attempt := 1; attempt <= maxChatAttempts; attempt++ {
 		httpReq, err := http.NewRequestWithContext(ctx, http.MethodPost, c.baseURL+"/v1/chat/completions", bytes.NewReader(body))
 		if err != nil {
 			return fmt.Errorf("create chat request: %w", err)
@@ -48,18 +63,56 @@ func (c *OpenAICompatibleClient) ChatStream(ctx context.Context, req ChatRequest
 
 		resp, err := c.httpClient.Do(httpReq)
 		if err != nil {
-		return fmt.Errorf("send chat request: %w", err)
+			lastErr = fmt.Errorf("send chat request: %w", err)
+			if !shouldRetryRequest(ctx, attempt, 0) {
+				return lastErr
+			}
+			if waitErr := waitBeforeRetry(ctx); waitErr != nil {
+				return waitErr
+			}
+			continue
 		}
-	defer resp.Body.Close()
 
 		if resp.StatusCode < 200 || resp.StatusCode >= 300 {
 			limit := io.LimitReader(resp.Body, 4096)
 			data, _ := io.ReadAll(limit)
-		return fmt.Errorf("chat request failed: status=%d body=%s", resp.StatusCode, strings.TrimSpace(string(data)))
+			_ = resp.Body.Close()
+			lastErr = fmt.Errorf("chat request failed: status=%d body=%s", resp.StatusCode, strings.TrimSpace(string(data)))
+			if !shouldRetryRequest(ctx, attempt, resp.StatusCode) {
+				return lastErr
+			}
+			if waitErr := waitBeforeRetry(ctx); waitErr != nil {
+				return waitErr
+			}
+			continue
 		}
 
+		defer resp.Body.Close()
 		return readSSE(resp.Body, emit)
 	}
+	return lastErr
+}
+
+func shouldRetryRequest(ctx context.Context, attempt int, status int) bool {
+	if ctx.Err() != nil || attempt >= maxChatAttempts {
+		return false
+	}
+	if status == 0 {
+		return true
+	}
+	return status == http.StatusRequestTimeout || status == http.StatusTooManyRequests || status >= 500
+}
+
+func waitBeforeRetry(ctx context.Context) error {
+	timer := time.NewTimer(retryDelay)
+	defer timer.Stop()
+	select {
+	case <-ctx.Done():
+		return ctx.Err()
+	case <-timer.C:
+		return nil
+	}
+}
 
 func readSSE(r io.Reader, emit func(StreamEvent) error) error {
 	reader := bufio.NewReader(r)
@@ -79,7 +132,7 @@ func readSSE(r io.Reader, emit func(StreamEvent) error) error {
 			if parseErr != nil {
 				return parseErr
 			}
-			if event.Content != "" || len(event.ToolCalls) > 0 || event.FinishReason != "" {
+			if event.Content != "" || len(event.ToolCalls) > 0 || event.FinishReason != "" || event.Usage != nil {
 				if emitErr := emit(event); emitErr != nil {
 					return emitErr
 				}
@@ -100,6 +153,7 @@ type streamPayload struct {
 		} `json:"delta"`
 		FinishReason string `json:"finish_reason"`
 	} `json:"choices"`
+	Usage *Usage `json:"usage,omitempty"`
 	Error *struct {
 		Message string `json:"message"`
 		Type    string `json:"type"`
@@ -127,15 +181,14 @@ func parseStreamPayload(payload string) (StreamEvent, error) {
 		}
 		return StreamEvent{}, fmt.Errorf("provider error: %s", decoded.Error.Message)
 	}
+	event := StreamEvent{Usage: decoded.Usage}
 	if len(decoded.Choices) == 0 {
-		return StreamEvent{}, nil
+		return event, nil
 	}
 
 	choice := decoded.Choices[0]
-	event := StreamEvent{
-		Content:      choice.Delta.Content,
-		FinishReason: choice.FinishReason,
-	}
+	event.Content = choice.Delta.Content
+	event.FinishReason = choice.FinishReason
 	for _, tc := range choice.Delta.ToolCalls {
 		event.ToolCalls = append(event.ToolCalls, ToolCallDelta{
 			Index:     tc.Index,

pkg/llm/openai_test.go

@@ -30,6 +30,10 @@ func TestOpenAICompatibleClientStreamsContentAndToolCalls(t *testing.T) {
 		if raw["thinking"] != "high" {
 			t.Fatalf("thinking = %#v", raw["thinking"])
 		}
+		streamOptions, ok := raw["stream_options"].(map[string]any)
+		if !ok || streamOptions["include_usage"] != true {
+			t.Fatalf("stream_options = %#v", raw["stream_options"])
+		}
 		if !req.Stream {
 			t.Fatal("request did not enable stream")
 		}
@@ -66,6 +70,75 @@ func TestOpenAICompatibleClientStreamsContentAndToolCalls(t *testing.T) {
 	}
 }
 
+func TestChatStreamCapturesUsage(t *testing.T) {
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		body := mustReadBody(t, r)
+		var raw map[string]any
+		if err := json.NewDecoder(strings.NewReader(body)).Decode(&raw); err != nil {
+			t.Fatal(err)
+		}
+		streamOptions, ok := raw["stream_options"].(map[string]any)
+		if !ok || streamOptions["include_usage"] != true {
+			t.Fatalf("stream_options = %#v", raw["stream_options"])
+		}
+
+		w.Header().Set("Content-Type", "text/event-stream")
+		_, _ = w.Write([]byte(`data: {"choices":[{"delta":{"content":"ok"},"finish_reason":""}]}` + "\n\n"))
+		_, _ = w.Write([]byte(`data: {"choices":[],"usage":{"prompt_tokens":12,"completion_tokens":3,"total_tokens":15}}` + "\n\n"))
+		_, _ = w.Write([]byte("data: [DONE]\n\n"))
+	}))
+	defer server.Close()
+
+	client := NewOpenAICompatibleClient(server.URL, "sk-test", server.Client())
+	var usage *Usage
+	err := client.ChatStream(context.Background(), ChatRequest{Model: "test"}, func(event StreamEvent) error {
+		if event.Usage != nil {
+			usage = event.Usage
+		}
+		return nil
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if usage == nil {
+		t.Fatal("usage was not emitted")
+	}
+	if usage.PromptTokens != 12 || usage.CompletionTokens != 3 || usage.TotalTokens != 15 {
+		t.Fatalf("usage = %#v", usage)
+	}
+}
+
+func TestOpenAICompatibleClientRetriesTransientHTTPError(t *testing.T) {
+	attempts := 0
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		attempts++
+		if attempts == 1 {
+			http.Error(w, "temporary", http.StatusBadGateway)
+			return
+		}
+		w.Header().Set("Content-Type", "text/event-stream")
+		_, _ = w.Write([]byte(`data: {"choices":[{"delta":{"content":"ok"},"finish_reason":""}]}` + "\n\n"))
+		_, _ = w.Write([]byte("data: [DONE]\n\n"))
+	}))
+	defer server.Close()
+
+	client := NewOpenAICompatibleClient(server.URL, "sk-test", server.Client())
+	var content strings.Builder
+	err := client.ChatStream(context.Background(), ChatRequest{Model: "test"}, func(event StreamEvent) error {
+		content.WriteString(event.Content)
+		return nil
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if attempts != 2 {
+		t.Fatalf("attempts = %d, want 2", attempts)
+	}
+	if content.String() != "ok" {
+		t.Fatalf("content = %q", content.String())
+	}
+}
+
 func mustReadBody(t *testing.T, r *http.Request) string {
 	t.Helper()
 	data, err := io.ReadAll(r.Body)

pkg/llm/token_estimate.go

@@ -0,0 +1,43 @@
+package llm
+
+import "encoding/json"
+
+const tokenEstimateCharRatio = 4
+
+// EstimateTokens estimates the token cost of messages and tool definitions in a
+// chat completion request. It intentionally uses a simple JSON-size heuristic:
+// roughly four UTF-8 bytes per token plus a small per-message framing overhead.
+func EstimateTokens(messages []Message, tools []Tool) int {
+	if len(messages) == 0 && len(tools) == 0 {
+		return 0
+	}
+
+	payload := struct {
+		Messages []Message `json:"messages,omitempty"`
+		Tools    []Tool    `json:"tools,omitempty"`
+	}{
+		Messages: messages,
+		Tools:    tools,
+	}
+	data, err := json.Marshal(payload)
+	if err != nil {
+		return 0
+	}
+
+	return estimateBytesAsTokens(len(data)) + len(messages)*4
+}
+
+func estimateMessageTokens(msg Message) int {
+	data, err := json.Marshal(msg)
+	if err != nil {
+		return 0
+	}
+	return estimateBytesAsTokens(len(data)) + 4
+}
+
+func estimateBytesAsTokens(n int) int {
+	if n <= 0 {
+		return 0
+	}
+	return (n + tokenEstimateCharRatio - 1) / tokenEstimateCharRatio
+}

pkg/llm/token_estimate_test.go

@@ -0,0 +1,45 @@
+package llm
+
+import "testing"
+
+func TestEstimateTokensEmpty(t *testing.T) {
+	if got := EstimateTokens(nil, nil); got != 0 {
+		t.Fatalf("EstimateTokens(nil, nil) = %d, want 0", got)
+	}
+}
+
+func TestEstimateTokens(t *testing.T) {
+	shortMessages := []Message{{Role: RoleUser, Content: "hello"}}
+	short := EstimateTokens(shortMessages, nil)
+	if short <= 4 {
+		t.Fatalf("short estimate = %d, want message framing plus content", short)
+	}
+	if short > 50 {
+		t.Fatalf("short estimate = %d, looks too high for one short message", short)
+	}
+
+	longMessages := []Message{
+		{Role: RoleUser, Content: "hello"},
+		{Role: RoleAssistant, Content: "This is a longer response with enough content to exceed the short estimate."},
+	}
+	long := EstimateTokens(longMessages, nil)
+	if long <= short {
+		t.Fatalf("long estimate = %d, want > short estimate %d", long, short)
+	}
+
+	withTools := EstimateTokens(shortMessages, []Tool{{
+		Type: "function",
+		Function: ToolFunction{
+			Name:        "file_read",
+			Description: "Read a file",
+			Parameters:  []byte(`{"type":"object"}`),
+		},
+	}})
+	if withTools <= short {
+		t.Fatalf("withTools estimate = %d, want > short estimate %d", withTools, short)
+	}
+
+	if msgTokens := estimateMessageTokens(shortMessages[0]); msgTokens <= 4 {
+		t.Fatalf("estimateMessageTokens = %d, want message framing plus content", msgTokens)
+	}
+}

pkg/llm/types.go

@@ -79,6 +79,13 @@ type StreamEvent struct {
 	Content      string
 	ToolCalls    []ToolCallDelta
 	FinishReason string
+	Usage        *Usage
+}
+
+type Usage struct {
+	PromptTokens     int `json:"prompt_tokens"`
+	CompletionTokens int `json:"completion_tokens"`
+	TotalTokens      int `json:"total_tokens"`
 }
 
 type ToolCallDelta struct {