Files
goclaw/internal/agent/intent_classify.go
T
viettranx 49441f7305 refactor: remove dead delegate code, rename lane/channel to team/teammate
- Remove handleDelegateAnnounce() dead code (no sender emits delegate:* messages)
- Remove delegate tool reference from intent_classify.go
- Rename LaneDelegate → LaneTeam with backward-compat env var fallback
- Rename ChannelDelegate → ChannelTeammate across all team tool files
- Comment out lifecycle guards in team_tasks_lifecycle.go (TODO: reviewer workflow)
- Update string literals in cron.go, task_ticker.go
- Gate tool_status placeholder_update to non-streaming runs only
- Skip FinalizeStream on tool.call to prevent mid-run content loss
2026-03-18 11:04:45 +07:00

154 lines
4.8 KiB
Go

package agent
import (
"context"
"strings"
"time"
"github.com/nextlevelbuilder/goclaw/internal/i18n"
"github.com/nextlevelbuilder/goclaw/internal/providers"
)
// IntentType represents the classified intent of a user message.
type IntentType string
const (
IntentStatusQuery IntentType = "status_query"
IntentCancel IntentType = "cancel"
IntentSteer IntentType = "steer"
IntentNewTask IntentType = "new_task"
)
const intentClassifyTimeout = 5 * time.Second
const intentSystemPrompt = `You are an intent classifier. The user has sent a message while the AI assistant is busy processing a previous request.
Classify the user's intent into exactly ONE of these categories:
- status_query: The user is asking about progress, status, or what the assistant is currently doing (e.g., "what are you doing?", "status?", "how far along?", "đang làm gì?", "bao giờ xong?")
- cancel: The user wants to stop or cancel the current task (e.g., "stop", "cancel", "nevermind", "thôi", "dừng lại")
- steer: The user wants to add instructions or redirect the current task (e.g., "also check X", "focus on Y instead", "thêm phần Z nữa")
- new_task: The user is sending a new unrelated request or message
Respond with ONLY the category name, nothing else.`
// intentPatterns provides regex-free fast-path detection for common keywords.
// Avoids LLM call for obvious patterns, saving cost and latency.
var statusKeywords = []string{
"status", "progress", "đang làm gì", "bao giờ xong", "做什么", "进度",
"what are you doing", "how far", "?",
}
var cancelKeywords = []string{
"stop", "cancel", "abort", "thôi", "dừng", "hủy", "取消", "停",
"nevermind", "never mind",
}
// quickClassify attempts keyword-based classification before calling the LLM.
// Returns (intent, true) if a match is found, or ("", false) to fall through to LLM.
func quickClassify(msg string) (IntentType, bool) {
lower := strings.ToLower(strings.TrimSpace(msg))
// Short messages (< 30 chars) are more likely to be simple intents.
if len(lower) > 60 {
return "", false
}
for _, kw := range cancelKeywords {
if strings.Contains(lower, kw) {
return IntentCancel, true
}
}
for _, kw := range statusKeywords {
if strings.Contains(lower, kw) {
return IntentStatusQuery, true
}
}
return "", false
}
// ClassifyIntent determines the intent of a user message sent while the agent is busy.
// Uses keyword fast-path first, then falls back to LLM classification.
// Falls back to IntentNewTask on any error.
func ClassifyIntent(ctx context.Context, provider providers.Provider, model, userMessage string) IntentType {
// Fast-path: keyword matching for obvious patterns (no LLM cost).
if intent, ok := quickClassify(userMessage); ok {
return intent
}
ctx, cancel := context.WithTimeout(ctx, intentClassifyTimeout)
defer cancel()
resp, err := provider.Chat(ctx, providers.ChatRequest{
Messages: []providers.Message{
{Role: "system", Content: intentSystemPrompt},
{Role: "user", Content: userMessage},
},
Model: model,
Options: map[string]any{
providers.OptMaxTokens: 20,
providers.OptTemperature: 0.0,
},
})
if err != nil {
return IntentNewTask
}
result := strings.TrimSpace(strings.ToLower(resp.Content))
switch {
case strings.Contains(result, "status_query"):
return IntentStatusQuery
case strings.Contains(result, "cancel"):
return IntentCancel
case strings.Contains(result, "steer"):
return IntentSteer
default:
return IntentNewTask
}
}
// FormatStatusReply builds a user-friendly status response from the current agent activity.
func FormatStatusReply(status *AgentActivityStatus, locale string) string {
if status == nil {
return i18n.T(locale, i18n.MsgStatusWorking)
}
elapsed := time.Since(status.StartedAt).Round(time.Second)
phase := formatPhase(status.Phase, status.Tool, locale)
return i18n.T(locale, i18n.MsgStatusDetailed, phase, status.Iteration, elapsed)
}
// formatPhase returns a human-readable phase description.
func formatPhase(phase, tool, locale string) string {
switch phase {
case "thinking":
return i18n.T(locale, i18n.MsgStatusPhaseThinking)
case "tool_exec":
if tool != "" {
return i18n.T(locale, i18n.MsgStatusPhaseToolExec, formatToolLabel(tool))
}
return i18n.T(locale, i18n.MsgStatusPhaseTools)
case "compacting":
return i18n.T(locale, i18n.MsgStatusPhaseCompact)
default:
return i18n.T(locale, i18n.MsgStatusPhaseDefault)
}
}
// formatToolLabel returns a user-friendly label for a tool name.
func formatToolLabel(tool string) string {
switch {
case strings.HasPrefix(tool, "web"):
return "web search"
case tool == "exec":
return "code execution"
case tool == "browser":
return "browser"
case tool == "spawn":
return "delegation"
case strings.HasPrefix(tool, "memory"):
return "memory"
case strings.HasPrefix(tool, "file"):
return "file operations"
default:
return tool
}
}