feat(config): add auto-compaction block and CLI wiring

Adds a configurable compaction section to portfolio-bot.yaml with
threshold_ratio, keep_recent_turns and an optional summary prompt.
Wires the new fields through config.Validate() and cmd/chat-bot/main.go
into agent.Runner.WithCompaction() so the runner can opt in to
auto-compaction at startup.

cmd/chat-bot/main.go

@@ -115,7 +115,13 @@ func serveCmd() *cobra.Command {
 			if err != nil {
 				return err
 			}
-			runner := agent.New(client, p, cfg.SystemPrompt, store, cfg.RAG.TopK)
+			runner := agent.New(client, p, cfg.SystemPrompt, store, cfg.RAG.TopK).
+				WithCompaction(agent.CompactionConfig{
+					Enabled:             cfg.Compaction.Enabled,
+					ThresholdRatio:      cfg.Compaction.ThresholdRatio,
+					KeepRecentTurns:     cfg.Compaction.KeepRecentTurns,
+					SummarySystemPrompt: cfg.Compaction.SummarySystemPrompt,
+				})
 
 			h := server.NewHandlers(cfg, runner, store, version)
 			srv := server.New(cfg, h)
@@ -212,7 +218,13 @@ func runAsk(cfg *config.Config, question string, noStream bool) error {
 	if err != nil {
 		return err
 	}
-	runner := agent.New(client, p, cfg.SystemPrompt, store, cfg.RAG.TopK)
+	runner := agent.New(client, p, cfg.SystemPrompt, store, cfg.RAG.TopK).
+		WithCompaction(agent.CompactionConfig{
+			Enabled:             cfg.Compaction.Enabled,
+			ThresholdRatio:      cfg.Compaction.ThresholdRatio,
+			KeepRecentTurns:     cfg.Compaction.KeepRecentTurns,
+			SummarySystemPrompt: cfg.Compaction.SummarySystemPrompt,
+		})
 	history := []agent.Message{{Role: agent.RoleUser, Content: question}}
 
 	if noStream {

configs/portfolio-bot.yaml

@@ -21,7 +21,7 @@ providers:
     type: llamacpp
     model: qwen2.5-3b-instruct
     endpoint: http://localhost:9100/v1
-    context_size: 4096
+    context_size: 2048
     max_tokens: 2048
     default: true
 
@@ -124,3 +124,17 @@ logging:
   level: info           # debug | info | warn | error
   format: json          # json | text
   output: stderr
+
+# Auto-compaction: fold the older part of a long conversation into a single
+# summary message before sending it to the model, so context overflow doesn't
+# kill long threads. Triggered when the previous turn's input tokens exceed
+# `threshold_ratio` of the provider's reported MaxContextWindow.
+#
+# Off by default — most portfolio chats are short. Turn it on for chatty
+# visitors or for the small-context local models (qwen2.5-1.5b / 3b) where
+# 4–6 turns is already most of the window.
+compaction:
+  enabled: true
+  threshold_ratio: 0.75      # compact at 75% of context window
+  keep_recent_turns: 2       # last 2 user turns kept verbatim; older → summary
+  # summary_system_prompt: "" # leave empty for the built-in bilingual default

internal/config/config.go

@@ -59,13 +59,41 @@ type Logging struct {
 	Output string `yaml:"output"`
 }
 
+// Compaction controls automatic context-window management. When the most
+// recent request's input tokens exceed threshold_ratio × the provider's
+// reported MaxContextWindow, the older portion of the conversation is
+// summarized by the LLM and replaced with a single system-role note.
+// Recent turns (keep_recent_turns) are always preserved verbatim.
+//
+// Disabled by default; flip enabled: true in the YAML to opt in. The
+// model's MaxContextWindow comes from llm.LLMClient.Capabilities(), so
+// the same setting works for llamacpp, ollama, openai and anthropic
+// without per-provider configuration.
+type Compaction struct {
+	Enabled     bool    `yaml:"enabled"`
+	// ThresholdRatio is the fraction of the context window that, once
+	// exceeded by the previous turn's input tokens, triggers compaction.
+	// Range 0–1; default 0.75 (75% of the window).
+	ThresholdRatio float64 `yaml:"threshold_ratio"`
+	// KeepRecentTurns is how many of the most recent user turns (each
+	// being a user message plus everything up to the next user message)
+	// to keep verbatim after compaction. Older turns are folded into the
+	// summary. Default 4.
+	KeepRecentTurns int `yaml:"keep_recent_turns"`
+	// SummarySystemPrompt is the instruction sent to the LLM when it
+	// summarizes the older turns. Empty falls back to a built-in
+	// bilingual prompt that auto-matches the conversation language.
+	SummarySystemPrompt string `yaml:"summary_system_prompt"`
+}
+
 type Config struct {
-	Server       Server    `yaml:"server"`
+	Server       Server     `yaml:"server"`
 	Providers    []Provider `yaml:"providers"`
-	RAG          RAG       `yaml:"rag"`
-	Persona      Persona   `yaml:"persona"`
-	SystemPrompt string    `yaml:"system_prompt"`
-	Logging      Logging   `yaml:"logging"`
+	RAG          RAG        `yaml:"rag"`
+	Persona      Persona    `yaml:"persona"`
+	SystemPrompt string     `yaml:"system_prompt"`
+	Compaction   Compaction `yaml:"compaction"`
+	Logging      Logging    `yaml:"logging"`
 }
 
 func Load(path string) (*Config, error) {
@@ -120,6 +148,14 @@ func (c *Config) validate() error {
 	if c.RAG.Tokenize == "" {
 		c.RAG.Tokenize = "unicode61"
 	}
+	if c.Compaction.Enabled {
+		if c.Compaction.ThresholdRatio <= 0 || c.Compaction.ThresholdRatio > 1 {
+			c.Compaction.ThresholdRatio = 0.75
+		}
+		if c.Compaction.KeepRecentTurns <= 0 {
+			c.Compaction.KeepRecentTurns = 4
+		}
+	}
 	return nil
 }