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

internal/agent/agent.go

@@ -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]...)