compact: retain newest ~20% of context; refactor startup into internal/app

- Replace fixed 6-message compact retention with ratio-based logic:
  retain ~20% of compactable estimated tokens, with a 6-message floor
  and user-turn boundary alignment
- Add compactRetainedTokenBudget, compactRecentStart, compactTurnBoundary
  helpers in internal/agent/agent.go
- Add TestAgentCompactKeepsTwentyPercentRecentContext
- Move startup/session/repl/bootstrap logic from root into internal/app
  so main.go is a thin binary entry point (15 lines)
- Update /compact docs in README.md
This commit is contained in:
loveuer
2026-06-24 23:59:21 -07:00
parent 9abd5c3b5f
commit dabf5bfecc
9 changed files with 408 additions and 224 deletions
+1 -1
View File
@@ -109,7 +109,7 @@ go run . [flags]
Slash commands:
- `/clear` resets conversation history.
- `/compact` summarizes older conversation history and keeps recent context.
- `/compact` summarizes older conversation history and keeps the newest ~20% of context, with a small recent-message floor.
- `/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.
+59 -7
View File
@@ -17,7 +17,8 @@ import (
const (
defaultMaxToolRounds = 50
doomLoopThreshold = 3
compactRecentMessages = 6
compactMinRecentMessages = 6
compactRecentRetentionPercent = 20
summaryMessageLimit = 8000
contextSummaryPrefix = "[context summary]"
toolLogArgumentLimit = 400
@@ -181,6 +182,61 @@ func (a *Agent) SetLastUsage(usage *llm.Usage) {
copy := *usage
a.lastUsage = &copy
}
// compactRetainedTokenBudget returns the number of estimated tokens to retain
// based on compactRecentRetentionPercent, using integer ceiling division.
func compactRetainedTokenBudget(totalTokens int) int {
if totalTokens <= 0 {
return 1
}
return (totalTokens*compactRecentRetentionPercent + 99) / 100
}
// compactTurnBoundary walks backward from start until it finds a user message
// or reaches prefixEnd, ensuring the retained suffix starts at a user turn.
func compactTurnBoundary(messages []llm.Message, prefixEnd int, start int) int {
for start > prefixEnd && messages[start].Role != llm.RoleUser {
start--
}
return start
}
// compactRecentStart returns the index where the retained recent suffix begins.
// It balances the 20% token-budget ratio with the compactMinRecentMessages floor
// and aligns to a user-turn boundary.
func compactRecentStart(messages []llm.Message, prefixEnd int) int {
compactableCount := len(messages) - prefixEnd
if compactableCount <= compactMinRecentMessages {
return len(messages)
}
compactable := messages[prefixEnd:]
totalTokens := llm.EstimateTokens(compactable, nil)
retainTokens := compactRetainedTokenBudget(totalTokens)
// Walk backward from the end to find the earliest suffix fitting the budget.
ratioStart := len(messages) - 1
for i := len(messages) - 1; i >= prefixEnd; i-- {
if llm.EstimateTokens(messages[i:], nil) <= retainTokens {
ratioStart = i
} else {
break
}
}
// Enforce the minimum recent-message floor.
minStart := len(messages) - compactMinRecentMessages
if minStart < prefixEnd {
minStart = prefixEnd
}
// Choose the earlier start so the floor is never weakened.
start := minStart
if ratioStart < start {
start = ratioStart
}
return compactTurnBoundary(messages, prefixEnd, start)
}
// Compact summarizes older conversation messages while preserving the system
// prompt and recent messages intact.
@@ -196,12 +252,8 @@ func (a *Agent) Compact(ctx context.Context) error {
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 {
recentStart := compactRecentStart(a.messages, prefixEnd)
if recentStart <= prefixEnd || recentStart >= len(a.messages) {
return nil
}
middle := append([]llm.Message(nil), a.messages[prefixEnd:recentStart]...)
+45
View File
@@ -145,6 +145,9 @@ func TestAgentCompact(t *testing.T) {
if messages[len(messages)-1].Content != "recent-answer-2" {
t.Fatalf("last recent message = %#v", messages[len(messages)-1])
}
if messages[2].Content != "recent-user-0" {
t.Fatalf("first retained recent message = %#v", messages[2])
}
if usage := a.LastUsage(); usage == nil || usage.TotalTokens != 12 {
t.Fatalf("usage = %#v", usage)
}
@@ -217,6 +220,48 @@ func longConversation() []llm.Message {
return messages
}
func manyTurnConversation(turns int) []llm.Message {
messages := []llm.Message{{Role: llm.RoleSystem, Content: "system"}}
for i := 0; i < turns; i++ {
messages = append(messages,
llm.Message{Role: llm.RoleUser, Content: fmt.Sprintf("user-%02d", i)},
llm.Message{Role: llm.RoleAssistant, Content: fmt.Sprintf("assistant-%02d", i)},
)
}
return messages
}
func TestAgentCompactKeepsTwentyPercentRecentContext(t *testing.T) {
provider := &compactProvider{}
a := New(Options{Provider: provider, Model: "test", SystemPrompt: "system"})
a.SetMessages(manyTurnConversation(25))
if err := a.Compact(context.Background()); err != nil {
t.Fatal(err)
}
messages := a.Messages()
// 1 system + 1 summary + 10 retained recent messages (5 user/assistant turns)
if len(messages) != 12 {
t.Fatalf("messages = %d, want 12; contents: %v", len(messages), joinMessageContents(messages))
}
if provider.summaryCalls != 1 {
t.Fatalf("summaryCalls = %d, want 1", provider.summaryCalls)
}
if messages[2].Content != "user-20" {
t.Fatalf("first retained message = %#v, want user-20", messages[2])
}
if messages[len(messages)-1].Content != "assistant-24" {
t.Fatalf("last retained message = %#v, want assistant-24", messages[len(messages)-1])
}
if !strings.Contains(provider.lastSummary, "user-19") {
t.Fatalf("summary should contain user-19 (part of summarized 80%%): %q", provider.lastSummary)
}
if strings.Contains(provider.lastSummary, "user-20") {
t.Fatalf("summary should not contain user-20 (part of retained 20%%): %q", provider.lastSummary)
}
}
func joinMessageContents(messages []llm.Message) string {
var b strings.Builder
for _, msg := range messages {
+164
View File
@@ -0,0 +1,164 @@
package app
import (
"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"
)
type cliOptions struct {
configPath string
yolo bool
plain bool
themeName string
resumeID string
}
func Run() error {
opts := parseCLIOptions()
themeMode, err := tui.ParseThemeMode(opts.themeName)
if err != nil {
return err
}
cfg, err := loadConfig(opts.configPath)
if err != nil {
return err
}
providerName, providerConfig, provider := defaultProvider(cfg, http.DefaultClient)
assistant := newAssistant(cfg, providerName, providerConfig, provider, opts.yolo, http.DefaultClient)
modelManager, err := newSessionManager(cfg, assistant, http.DefaultClient)
if err != nil {
return err
}
if err := initializeSession(modelManager, opts.resumeID); err != nil {
return err
}
ctx, stop := signal.NotifyContext(context.Background(), os.Interrupt)
defer stop()
return runInterface(ctx, assistant, modelManager, providerConfig, cfg.Agent.WorkingDir, opts, themeMode)
}
func parseCLIOptions() cliOptions {
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()
return cliOptions{
configPath: *configPath,
yolo: *yolo,
plain: *plain,
themeName: *themeName,
resumeID: *resumeID,
}
}
func loadConfig(configPath string) (*config.Config, error) {
cfg, err := config.Load(configPath)
if err == nil {
return cfg, nil
}
if !errors.Is(err, os.ErrNotExist) {
return nil, err
}
if bootstrapErr := bootstrapConfig(configPath); bootstrapErr != nil {
return nil, bootstrapErr
}
return nil, fmt.Errorf("created default config at %s; edit it to set your API key and provider, then restart agentu", configPath)
}
func defaultProvider(cfg *config.Config, httpClient *http.Client) (string, config.ProviderConfig, llm.Provider) {
providerName := cfg.DefaultProviderName()
providerConfig := cfg.Providers[providerName]
provider := llm.NewOpenAICompatibleClient(providerConfig.BaseURL, providerConfig.APIKey, httpClient)
return providerName, providerConfig, provider
}
func newAssistant(cfg *config.Config, providerName string, providerConfig config.ProviderConfig, provider llm.Provider, yolo bool, httpClient *http.Client) *agent.Agent {
return agent.New(agent.Options{
Provider: provider,
ProviderName: providerName,
Model: providerConfig.Model,
Thinking: providerConfig.Thinking,
ThinkingParam: providerConfig.ThinkingParam,
ThinkingEnabled: providerConfig.ThinkingConfigured,
SystemPrompt: systemPrompt(cfg, yolo),
ToolRegistry: newToolRegistry(cfg.Agent.WorkingDir, yolo, httpClient),
ToolTimeout: cfg.Agent.ToolTimeout,
MaxContextTokens: maxContextTokens(cfg, providerConfig),
})
}
func systemPrompt(cfg *config.Config, yolo bool) string {
instructions := tools.AgentInstructions(cfg.Agent.WorkingDir, yolo)
return strings.TrimSpace(cfg.Agent.SystemPrompt + "\n\n" + instructions)
}
func maxContextTokens(cfg *config.Config, providerConfig config.ProviderConfig) int {
if cfg.Agent.MaxContextTokens != 0 {
return cfg.Agent.MaxContextTokens
}
return providerConfig.ContextTokens
}
func newToolRegistry(workingDir string, yolo bool, httpClient *http.Client) *tools.Registry {
var registry *tools.Registry
if yolo {
registry = tools.Builtins(workingDir)
} else {
registry = tools.ReadOnlyBuiltins(workingDir)
}
registry.Register(tools.NewBuiltinSearchTool(httpClient))
registry.Register(tools.NewBuiltinFetchTool(httpClient))
return registry
}
func newSessionManager(cfg *config.Config, assistant *agent.Agent, httpClient *http.Client) (*session.Manager, error) {
home, err := os.UserHomeDir()
if err != nil {
return nil, fmt.Errorf("resolve home dir: %w", err)
}
sessionsDir := filepath.Join(home, ".agentu", "sessions")
store, err := session.NewStore(sessionsDir)
if err != nil {
return nil, fmt.Errorf("init session store: %w", err)
}
return session.NewManager(cfg, assistant, httpClient, store)
}
func runInterface(ctx context.Context, assistant *agent.Agent, modelManager *session.Manager, providerConfig config.ProviderConfig, workingDir string, opts cliOptions, themeMode tui.ThemeMode) error {
if opts.plain {
return repl(ctx, assistant, modelManager)
}
err := tui.Run(ctx, assistant, tui.Options{
ModelName: providerConfig.Model,
Yolo: opts.yolo,
ThemeMode: themeMode,
ModelManager: modelManager,
WorkingDir: workingDir,
})
printExitSession(modelManager)
return err
}
+38
View File
@@ -0,0 +1,38 @@
package app
import (
"fmt"
"os"
"path/filepath"
"agentu/pkg/config"
)
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 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
}
+1 -1
View File
@@ -1,4 +1,4 @@
package main
package app
import (
"context"
+60
View File
@@ -0,0 +1,60 @@
package app
import (
"bufio"
"context"
"fmt"
"os"
"strings"
"agentu/internal/agent"
"agentu/internal/session"
)
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()
}
}
+38
View File
@@ -0,0 +1,38 @@
package app
import (
"fmt"
"os"
"agentu/internal/session"
)
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 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)
}
}
+2 -215
View File
@@ -1,228 +1,15 @@
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"
"agentu/internal/app"
)
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 {
if err := app.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)
}
}