Merge pull request #994 from is-Xiaoen/feat/model-routing

feat(routing): intelligent model routing based on structural complexity scoring

pkg/agent/instance.go

@@ -37,6 +37,14 @@ type AgentInstance struct {
 	Subagents                 *config.SubagentsConfig
 	SkillsFilter              []string
 	Candidates                []providers.FallbackCandidate
+
+	// Router is non-nil when model routing is configured and the light model
+	// was successfully resolved. It scores each incoming message and decides
+	// whether to route to LightCandidates or stay with Candidates.
+	Router *routing.Router
+	// LightCandidates holds the resolved provider candidates for the light model.
+	// Pre-computed at agent creation to avoid repeated model_list lookups at runtime.
+	LightCandidates []providers.FallbackCandidate
 }
 
 // NewAgentInstance creates an agent instance from config.
@@ -180,6 +188,25 @@ func NewAgentInstance(
 
 	candidates := providers.ResolveCandidatesWithLookup(modelCfg, defaults.Provider, resolveFromModelList)
 
+	// Model routing setup: pre-resolve light model candidates at creation time
+	// to avoid repeated model_list lookups on every incoming message.
+	var router *routing.Router
+	var lightCandidates []providers.FallbackCandidate
+	if rc := defaults.Routing; rc != nil && rc.Enabled && rc.LightModel != "" {
+		lightModelCfg := providers.ModelConfig{Primary: rc.LightModel}
+		resolved := providers.ResolveCandidatesWithLookup(lightModelCfg, defaults.Provider, resolveFromModelList)
+		if len(resolved) > 0 {
+			router = routing.New(routing.RouterConfig{
+				LightModel: rc.LightModel,
+				Threshold:  rc.Threshold,
+			})
+			lightCandidates = resolved
+		} else {
+			log.Printf("routing: light_model %q not found in model_list — routing disabled for agent %q",
+				rc.LightModel, agentID)
+		}
+	}
+
 	return &AgentInstance{
 		ID:                        agentID,
 		Name:                      agentName,
@@ -200,6 +227,8 @@ func NewAgentInstance(
 		Subagents:                 subagents,
 		SkillsFilter:              skillsFilter,
 		Candidates:                candidates,
+		Router:                    router,
+		LightCandidates:           lightCandidates,
 	}
 }
 

pkg/agent/loop.go

@@ -824,6 +824,12 @@ func (al *AgentLoop) runLLMIteration(
 	iteration := 0
 	var finalContent string
 
+	// Determine effective model tier for this conversation turn.
+	// selectCandidates evaluates routing once and the decision is sticky for
+	// all tool-follow-up iterations within the same turn so that a multi-step
+	// tool chain doesn't switch models mid-way through.
+	activeCandidates, activeModel := al.selectCandidates(agent, opts.UserMessage, messages)
+
 	for iteration < agent.MaxIterations {
 		iteration++
 
@@ -842,7 +848,7 @@ func (al *AgentLoop) runLLMIteration(
 			map[string]any{
 				"agent_id":          agent.ID,
 				"iteration":         iteration,
-				"model":             agent.Model,
+				"model":             activeModel,
 				"messages_count":    len(messages),
 				"tools_count":       len(providerToolDefs),
 				"max_tokens":        agent.MaxTokens,
@@ -858,7 +864,7 @@ func (al *AgentLoop) runLLMIteration(
 				"tools_json":    formatToolsForLog(providerToolDefs),
 			})
 
-		// Call LLM with fallback chain if candidates are configured.
+		// Call LLM with fallback chain if multiple candidates are configured.
 		var response *providers.LLMResponse
 		var err error
 
@@ -879,10 +885,10 @@ func (al *AgentLoop) runLLMIteration(
 		}
 
 		callLLM := func() (*providers.LLMResponse, error) {
-			if len(agent.Candidates) > 1 && al.fallback != nil {
+			if len(activeCandidates) > 1 && al.fallback != nil {
 				fbResult, fbErr := al.fallback.Execute(
 					ctx,
-					agent.Candidates,
+					activeCandidates,
 					func(ctx context.Context, provider, model string) (*providers.LLMResponse, error) {
 						return agent.Provider.Chat(ctx, messages, providerToolDefs, model, llmOpts)
 					},
@@ -900,7 +906,7 @@ func (al *AgentLoop) runLLMIteration(
 				}
 				return fbResult.Response, nil
 			}
-			return agent.Provider.Chat(ctx, messages, providerToolDefs, agent.Model, llmOpts)
+			return agent.Provider.Chat(ctx, messages, providerToolDefs, activeModel, llmOpts)
 		}
 
 		// Retry loop for context/token errors
@@ -1169,6 +1175,44 @@ func (al *AgentLoop) runLLMIteration(
 	return finalContent, iteration, nil
 }
 
+// selectCandidates returns the model candidates and resolved model name to use
+// for a conversation turn. When model routing is configured and the incoming
+// message scores below the complexity threshold, it returns the light model
+// candidates instead of the primary ones.
+//
+// The returned (candidates, model) pair is used for all LLM calls within one
+// turn — tool follow-up iterations use the same tier as the initial call so
+// that a multi-step tool chain doesn't switch models mid-way.
+func (al *AgentLoop) selectCandidates(
+	agent *AgentInstance,
+	userMsg string,
+	history []providers.Message,
+) (candidates []providers.FallbackCandidate, model string) {
+	if agent.Router == nil || len(agent.LightCandidates) == 0 {
+		return agent.Candidates, agent.Model
+	}
+
+	_, usedLight, score := agent.Router.SelectModel(userMsg, history, agent.Model)
+	if !usedLight {
+		logger.DebugCF("agent", "Model routing: primary model selected",
+			map[string]any{
+				"agent_id":  agent.ID,
+				"score":     score,
+				"threshold": agent.Router.Threshold(),
+			})
+		return agent.Candidates, agent.Model
+	}
+
+	logger.InfoCF("agent", "Model routing: light model selected",
+		map[string]any{
+			"agent_id":    agent.ID,
+			"light_model": agent.Router.LightModel(),
+			"score":       score,
+			"threshold":   agent.Router.Threshold(),
+		})
+	return agent.LightCandidates, agent.Router.LightModel()
+}
+
 // maybeSummarize triggers summarization if the session history exceeds thresholds.
 func (al *AgentLoop) maybeSummarize(agent *AgentInstance, sessionKey, channel, chatID string) {
 	newHistory := agent.Sessions.GetHistory(sessionKey)

pkg/config/config.go

@@ -167,22 +167,35 @@ type SessionConfig struct {
 	IdentityLinks map[string][]string `json:"identity_links,omitempty"`
 }
 
+// RoutingConfig controls the intelligent model routing feature.
+// When enabled, each incoming message is scored against structural features
+// (message length, code blocks, tool call history, conversation depth, attachments).
+// Messages scoring below Threshold are sent to LightModel; all others use the
+// agent's primary model. This reduces cost and latency for simple tasks without
+// requiring any keyword matching — all scoring is language-agnostic.
+type RoutingConfig struct {
+	Enabled    bool    `json:"enabled"`
+	LightModel string  `json:"light_model"` // model_name from model_list to use for simple tasks
+	Threshold  float64 `json:"threshold"`   // complexity score in [0,1]; score >= threshold → primary model
+}
+
 type AgentDefaults struct {
-	Workspace                 string   `json:"workspace"                       env:"PICOCLAW_AGENTS_DEFAULTS_WORKSPACE"`
-	RestrictToWorkspace       bool     `json:"restrict_to_workspace"           env:"PICOCLAW_AGENTS_DEFAULTS_RESTRICT_TO_WORKSPACE"`
-	AllowReadOutsideWorkspace bool     `json:"allow_read_outside_workspace"    env:"PICOCLAW_AGENTS_DEFAULTS_ALLOW_READ_OUTSIDE_WORKSPACE"`
-	Provider                  string   `json:"provider"                        env:"PICOCLAW_AGENTS_DEFAULTS_PROVIDER"`
-	ModelName                 string   `json:"model_name,omitempty"            env:"PICOCLAW_AGENTS_DEFAULTS_MODEL_NAME"`
-	Model                     string   `json:"model"                           env:"PICOCLAW_AGENTS_DEFAULTS_MODEL"` // Deprecated: use model_name instead
-	ModelFallbacks            []string `json:"model_fallbacks,omitempty"`
-	ImageModel                string   `json:"image_model,omitempty"           env:"PICOCLAW_AGENTS_DEFAULTS_IMAGE_MODEL"`
-	ImageModelFallbacks       []string `json:"image_model_fallbacks,omitempty"`
-	MaxTokens                 int      `json:"max_tokens"                      env:"PICOCLAW_AGENTS_DEFAULTS_MAX_TOKENS"`
-	Temperature               *float64 `json:"temperature,omitempty"           env:"PICOCLAW_AGENTS_DEFAULTS_TEMPERATURE"`
-	MaxToolIterations         int      `json:"max_tool_iterations"             env:"PICOCLAW_AGENTS_DEFAULTS_MAX_TOOL_ITERATIONS"`
-	SummarizeMessageThreshold int      `json:"summarize_message_threshold"     env:"PICOCLAW_AGENTS_DEFAULTS_SUMMARIZE_MESSAGE_THRESHOLD"`
-	SummarizeTokenPercent     int      `json:"summarize_token_percent"         env:"PICOCLAW_AGENTS_DEFAULTS_SUMMARIZE_TOKEN_PERCENT"`
-	MaxMediaSize              int      `json:"max_media_size,omitempty"        env:"PICOCLAW_AGENTS_DEFAULTS_MAX_MEDIA_SIZE"`
+	Workspace                 string         `json:"workspace"                       env:"PICOCLAW_AGENTS_DEFAULTS_WORKSPACE"`
+	RestrictToWorkspace       bool           `json:"restrict_to_workspace"           env:"PICOCLAW_AGENTS_DEFAULTS_RESTRICT_TO_WORKSPACE"`
+	AllowReadOutsideWorkspace bool           `json:"allow_read_outside_workspace"    env:"PICOCLAW_AGENTS_DEFAULTS_ALLOW_READ_OUTSIDE_WORKSPACE"`
+	Provider                  string         `json:"provider"                        env:"PICOCLAW_AGENTS_DEFAULTS_PROVIDER"`
+	ModelName                 string         `json:"model_name,omitempty"            env:"PICOCLAW_AGENTS_DEFAULTS_MODEL_NAME"`
+	Model                     string         `json:"model"                           env:"PICOCLAW_AGENTS_DEFAULTS_MODEL"` // Deprecated: use model_name instead
+	ModelFallbacks            []string       `json:"model_fallbacks,omitempty"`
+	ImageModel                string         `json:"image_model,omitempty"           env:"PICOCLAW_AGENTS_DEFAULTS_IMAGE_MODEL"`
+	ImageModelFallbacks       []string       `json:"image_model_fallbacks,omitempty"`
+	MaxTokens                 int            `json:"max_tokens"                      env:"PICOCLAW_AGENTS_DEFAULTS_MAX_TOKENS"`
+	Temperature               *float64       `json:"temperature,omitempty"           env:"PICOCLAW_AGENTS_DEFAULTS_TEMPERATURE"`
+	MaxToolIterations         int            `json:"max_tool_iterations"             env:"PICOCLAW_AGENTS_DEFAULTS_MAX_TOOL_ITERATIONS"`
+	SummarizeMessageThreshold int            `json:"summarize_message_threshold"     env:"PICOCLAW_AGENTS_DEFAULTS_SUMMARIZE_MESSAGE_THRESHOLD"`
+	SummarizeTokenPercent     int            `json:"summarize_token_percent"         env:"PICOCLAW_AGENTS_DEFAULTS_SUMMARIZE_TOKEN_PERCENT"`
+	MaxMediaSize              int            `json:"max_media_size,omitempty"        env:"PICOCLAW_AGENTS_DEFAULTS_MAX_MEDIA_SIZE"`
+	Routing                   *RoutingConfig `json:"routing,omitempty"`
 }
 
 const DefaultMaxMediaSize = 20 * 1024 * 1024 // 20 MB

pkg/routing/classifier.go

@@ -0,0 +1,80 @@
+package routing
+
+// Classifier evaluates a feature set and returns a complexity score in [0, 1].
+// A higher score indicates a more complex task that benefits from a heavy model.
+// The score is compared against the configured threshold: score >= threshold selects
+// the primary (heavy) model; score < threshold selects the light model.
+//
+// Classifier is an interface so that future implementations (ML-based, embedding-based,
+// or any other approach) can be swapped in without changing routing infrastructure.
+type Classifier interface {
+	Score(f Features) float64
+}
+
+// RuleClassifier is the v1 implementation.
+// It uses a weighted sum of structural signals with no external dependencies,
+// no API calls, and sub-microsecond latency. The raw sum is capped at 1.0 so
+// that the returned score always falls within the [0, 1] contract.
+//
+// Individual weights (multiple signals can fire simultaneously):
+//
+//	token > 200 (≈600 chars): 0.35  — very long prompts are almost always complex
+//	token 50-200:             0.15  — medium length; may or may not be complex
+//	code block present:       0.40  — coding tasks need the heavy model
+//	tool calls > 3 (recent):  0.25  — dense tool usage signals an agentic workflow
+//	tool calls 1-3 (recent):  0.10  — some tool activity
+//	conversation depth > 10:  0.10  — long sessions carry implicit complexity
+//	attachments present:      1.00  — hard gate; multi-modal always needs heavy model
+//
+// Default threshold is 0.35, so:
+//   - Pure greetings / trivial Q&A:                 0.00 → light  ✓
+//   - Medium prose message (50–200 tokens):          0.15 → light  ✓
+//   - Message with code block:                       0.40 → heavy  ✓
+//   - Long message (>200 tokens):                    0.35 → heavy  ✓
+//   - Active tool session + medium message:          0.25 → light  (acceptable)
+//   - Any message with an image/audio attachment:    1.00 → heavy  ✓
+type RuleClassifier struct{}
+
+// Score computes the complexity score for the given feature set.
+// The returned value is in [0, 1]. Attachments short-circuit to 1.0.
+func (c *RuleClassifier) Score(f Features) float64 {
+	// Hard gate: multi-modal inputs always require the heavy model.
+	if f.HasAttachments {
+		return 1.0
+	}
+
+	var score float64
+
+	// Token estimate — primary verbosity signal
+	switch {
+	case f.TokenEstimate > 200:
+		score += 0.35
+	case f.TokenEstimate > 50:
+		score += 0.15
+	}
+
+	// Fenced code blocks — strongest indicator of a coding/technical task
+	if f.CodeBlockCount > 0 {
+		score += 0.40
+	}
+
+	// Recent tool call density — indicates an ongoing agentic workflow
+	switch {
+	case f.RecentToolCalls > 3:
+		score += 0.25
+	case f.RecentToolCalls > 0:
+		score += 0.10
+	}
+
+	// Conversation depth — accumulated context implies compound task
+	if f.ConversationDepth > 10 {
+		score += 0.10
+	}
+
+	// Cap at 1.0 to honor the [0, 1] contract even when multiple signals fire
+	// simultaneously (e.g., long message + code block + tool chain = 1.10 raw).
+	if score > 1.0 {
+		score = 1.0
+	}
+	return score
+}

pkg/routing/features.go

@@ -0,0 +1,127 @@
+package routing
+
+import (
+	"strings"
+	"unicode/utf8"
+
+	"github.com/sipeed/picoclaw/pkg/providers"
+)
+
+// lookbackWindow is the number of recent history entries scanned for tool calls.
+// Six entries covers roughly one full tool-use round-trip (user → assistant+tool_call → tool_result → assistant).
+const lookbackWindow = 6
+
+// Features holds the structural signals extracted from a message and its session context.
+// Every dimension is language-agnostic by construction — no keyword or pattern matching
+// against natural-language content. This ensures consistent routing for all locales.
+type Features struct {
+	// TokenEstimate is a proxy for token count.
+	// CJK runes count as 1 token each; non-CJK runes as 0.25 tokens each.
+	// This avoids API calls while giving accurate estimates for all scripts.
+	TokenEstimate int
+
+	// CodeBlockCount is the number of fenced code blocks (``` pairs) in the message.
+	// Coding tasks almost always require the heavy model.
+	CodeBlockCount int
+
+	// RecentToolCalls is the count of tool_call messages in the last lookbackWindow
+	// history entries. A high density indicates an active agentic workflow.
+	RecentToolCalls int
+
+	// ConversationDepth is the total number of messages in the session history.
+	// Deep sessions tend to carry implicit complexity built up over many turns.
+	ConversationDepth int
+
+	// HasAttachments is true when the message appears to contain media (images,
+	// audio, video). Multi-modal inputs require vision-capable heavy models.
+	HasAttachments bool
+}
+
+// ExtractFeatures computes the structural feature vector for a message.
+// It is a pure function with no side effects and zero allocations beyond
+// the returned struct.
+func ExtractFeatures(msg string, history []providers.Message) Features {
+	return Features{
+		TokenEstimate:     estimateTokens(msg),
+		CodeBlockCount:    countCodeBlocks(msg),
+		RecentToolCalls:   countRecentToolCalls(history),
+		ConversationDepth: len(history),
+		HasAttachments:    hasAttachments(msg),
+	}
+}
+
+// estimateTokens returns a token count proxy that handles both CJK and Latin text.
+// CJK runes (U+2E80–U+9FFF, U+F900–U+FAFF, U+AC00–U+D7AF) map to roughly one
+// token each, while non-CJK runes average ~0.25 tokens/rune (≈4 chars per token
+// for English). Splitting the count this way avoids the 3x underestimation that a
+// flat rune_count/3 would produce for Chinese, Japanese, and Korean text.
+func estimateTokens(msg string) int {
+	total := utf8.RuneCountInString(msg)
+	if total == 0 {
+		return 0
+	}
+	cjk := 0
+	for _, r := range msg {
+		if r >= 0x2E80 && r <= 0x9FFF || r >= 0xF900 && r <= 0xFAFF || r >= 0xAC00 && r <= 0xD7AF {
+			cjk++
+		}
+	}
+	return cjk + (total-cjk)/4
+}
+
+// countCodeBlocks counts the number of complete fenced code blocks.
+// Each ``` delimiter increments a counter; pairs of delimiters form one block.
+// An unclosed opening fence (odd count) is treated as zero complete blocks
+// since it may just be an inline code span or a typo.
+func countCodeBlocks(msg string) int {
+	n := strings.Count(msg, "```")
+	return n / 2
+}
+
+// countRecentToolCalls counts messages with tool calls in the last lookbackWindow
+// entries of history. It examines the ToolCalls field rather than parsing
+// the content string, so it is robust to any message format.
+func countRecentToolCalls(history []providers.Message) int {
+	start := len(history) - lookbackWindow
+	if start < 0 {
+		start = 0
+	}
+
+	count := 0
+	for _, msg := range history[start:] {
+		if len(msg.ToolCalls) > 0 {
+			count += len(msg.ToolCalls)
+		}
+	}
+	return count
+}
+
+// hasAttachments returns true when the message content contains embedded media.
+// It checks for base64 data URIs (data:image/, data:audio/, data:video/) and
+// common image/audio URL extensions. This is intentionally conservative —
+// false negatives (missing an attachment) just mean the routing falls back to
+// the primary model anyway.
+func hasAttachments(msg string) bool {
+	lower := strings.ToLower(msg)
+
+	// Base64 data URIs embedded directly in the message
+	if strings.Contains(lower, "data:image/") ||
+		strings.Contains(lower, "data:audio/") ||
+		strings.Contains(lower, "data:video/") {
+		return true
+	}
+
+	// Common image/audio extensions in URLs or file references
+	mediaExts := []string{
+		".jpg", ".jpeg", ".png", ".gif", ".webp", ".bmp",
+		".mp3", ".wav", ".ogg", ".m4a", ".flac",
+		".mp4", ".avi", ".mov", ".webm",
+	}
+	for _, ext := range mediaExts {
+		if strings.Contains(lower, ext) {
+			return true
+		}
+	}
+
+	return false
+}

pkg/routing/router.go

@@ -0,0 +1,82 @@
+package routing
+
+import (
+	"github.com/sipeed/picoclaw/pkg/providers"
+)
+
+// defaultThreshold is used when the config threshold is zero or negative.
+// At 0.35 a message needs at least one strong signal (code block, long text,
+// or an attachment) before the heavy model is chosen.
+const defaultThreshold = 0.35
+
+// RouterConfig holds the validated model routing settings.
+// It mirrors config.RoutingConfig but lives in pkg/routing to keep the
+// dependency graph simple: pkg/agent resolves config → routing, not the reverse.
+type RouterConfig struct {
+	// LightModel is the model_name (from model_list) used for simple tasks.
+	LightModel string
+
+	// Threshold is the complexity score cutoff in [0, 1].
+	// score >= Threshold → primary (heavy) model.
+	// score <  Threshold → light model.
+	Threshold float64
+}
+
+// Router selects the appropriate model tier for each incoming message.
+// It is safe for concurrent use from multiple goroutines.
+type Router struct {
+	cfg        RouterConfig
+	classifier Classifier
+}
+
+// New creates a Router with the given config and the default RuleClassifier.
+// If cfg.Threshold is zero or negative, defaultThreshold (0.35) is used.
+func New(cfg RouterConfig) *Router {
+	if cfg.Threshold <= 0 {
+		cfg.Threshold = defaultThreshold
+	}
+	return &Router{
+		cfg:        cfg,
+		classifier: &RuleClassifier{},
+	}
+}
+
+// newWithClassifier creates a Router with a custom Classifier.
+// Intended for unit tests that need to inject a deterministic scorer.
+func newWithClassifier(cfg RouterConfig, c Classifier) *Router {
+	if cfg.Threshold <= 0 {
+		cfg.Threshold = defaultThreshold
+	}
+	return &Router{cfg: cfg, classifier: c}
+}
+
+// SelectModel returns the model to use for this conversation turn along with
+// the computed complexity score (for logging and debugging).
+//
+//   - If score < cfg.Threshold: returns (cfg.LightModel, true, score)
+//   - Otherwise:               returns (primaryModel, false, score)
+//
+// The caller is responsible for resolving the returned model name into
+// provider candidates (see AgentInstance.LightCandidates).
+func (r *Router) SelectModel(
+	msg string,
+	history []providers.Message,
+	primaryModel string,
+) (model string, usedLight bool, score float64) {
+	features := ExtractFeatures(msg, history)
+	score = r.classifier.Score(features)
+	if score < r.cfg.Threshold {
+		return r.cfg.LightModel, true, score
+	}
+	return primaryModel, false, score
+}
+
+// LightModel returns the configured light model name.
+func (r *Router) LightModel() string {
+	return r.cfg.LightModel
+}
+
+// Threshold returns the complexity threshold in use.
+func (r *Router) Threshold() float64 {
+	return r.cfg.Threshold
+}

pkg/routing/router_test.go

@@ -0,0 +1,414 @@
+package routing
+
+import (
+	"strings"
+	"testing"
+
+	"github.com/sipeed/picoclaw/pkg/providers"
+)
+
+// ── ExtractFeatures ──────────────────────────────────────────────────────────
+
+func TestExtractFeatures_EmptyMessage(t *testing.T) {
+	f := ExtractFeatures("", nil)
+	if f.TokenEstimate != 0 {
+		t.Errorf("TokenEstimate: got %d, want 0", f.TokenEstimate)
+	}
+	if f.CodeBlockCount != 0 {
+		t.Errorf("CodeBlockCount: got %d, want 0", f.CodeBlockCount)
+	}
+	if f.RecentToolCalls != 0 {
+		t.Errorf("RecentToolCalls: got %d, want 0", f.RecentToolCalls)
+	}
+	if f.ConversationDepth != 0 {
+		t.Errorf("ConversationDepth: got %d, want 0", f.ConversationDepth)
+	}
+	if f.HasAttachments {
+		t.Error("HasAttachments: got true, want false")
+	}
+}
+
+func TestExtractFeatures_TokenEstimate(t *testing.T) {
+	// 30 ASCII runes: 0 CJK + 30/4 = 7 tokens
+	msg := strings.Repeat("a", 30)
+	f := ExtractFeatures(msg, nil)
+	if f.TokenEstimate != 7 {
+		t.Errorf("TokenEstimate: got %d, want 7", f.TokenEstimate)
+	}
+}
+
+func TestExtractFeatures_TokenEstimate_CJK(t *testing.T) {
+	// 9 CJK runes → 9 tokens (each CJK rune ≈ 1 token).
+	// Using a rune slice literal avoids CJK string literals in source.
+	msg := string([]rune{
+		0x4F60, 0x597D, 0x4E16, 0x754C,
+		0x4F60, 0x597D, 0x4E16, 0x754C,
+		0x4F60,
+	})
+	f := ExtractFeatures(msg, nil)
+	if f.TokenEstimate != 9 {
+		t.Errorf("CJK TokenEstimate: got %d, want 9", f.TokenEstimate)
+	}
+}
+
+func TestExtractFeatures_TokenEstimate_Mixed(t *testing.T) {
+	// Mixed: 4 CJK runes + 8 ASCII runes → 4 + 8/4 = 6 tokens.
+	msg := string([]rune{0x4F60, 0x597D, 0x4E16, 0x754C}) + "hello ok"
+	f := ExtractFeatures(msg, nil)
+	if f.TokenEstimate != 6 {
+		t.Errorf("Mixed TokenEstimate: got %d, want 6", f.TokenEstimate)
+	}
+}
+
+func TestExtractFeatures_CodeBlocks(t *testing.T) {
+	cases := []struct {
+		msg  string
+		want int
+	}{
+		{"no code here", 0},
+		{"```go\nfmt.Println()\n```", 1},
+		{"```python\npass\n```\n```js\nconsole.log()\n```", 2},
+		{"```unclosed", 0}, // odd number of fences = 0 complete blocks
+	}
+	for _, tc := range cases {
+		f := ExtractFeatures(tc.msg, nil)
+		if f.CodeBlockCount != tc.want {
+			t.Errorf("msg=%q: CodeBlockCount got %d, want %d", tc.msg, f.CodeBlockCount, tc.want)
+		}
+	}
+}
+
+func TestExtractFeatures_RecentToolCalls(t *testing.T) {
+	// History longer than lookbackWindow — only last lookbackWindow entries count.
+	history := make([]providers.Message, 10)
+	// Put 2 tool calls at positions 8 and 9 (within the last 6)
+	history[8] = providers.Message{Role: "assistant", ToolCalls: []providers.ToolCall{{Name: "exec"}}}
+	history[9] = providers.Message{
+		Role:      "assistant",
+		ToolCalls: []providers.ToolCall{{Name: "read_file"}, {Name: "write_file"}},
+	}
+	// Position 3 is outside the lookback window and must NOT be counted
+	history[3] = providers.Message{Role: "assistant", ToolCalls: []providers.ToolCall{{Name: "old_tool"}}}
+
+	f := ExtractFeatures("test", history)
+	// 1 (position 8) + 2 (position 9) = 3
+	if f.RecentToolCalls != 3 {
+		t.Errorf("RecentToolCalls: got %d, want 3", f.RecentToolCalls)
+	}
+}
+
+func TestExtractFeatures_ConversationDepth(t *testing.T) {
+	history := make([]providers.Message, 7)
+	f := ExtractFeatures("msg", history)
+	if f.ConversationDepth != 7 {
+		t.Errorf("ConversationDepth: got %d, want 7", f.ConversationDepth)
+	}
+}
+
+func TestExtractFeatures_HasAttachments_DataURI(t *testing.T) {
+	cases := []struct {
+		msg  string
+		want bool
+	}{
+		{"plain text", false},
+		{"here is an image: data:image/png;base64,abc123", true},
+		{"audio: data:audio/mp3;base64,xyz", true},
+		{"video: data:video/mp4;base64,xyz", true},
+	}
+	for _, tc := range cases {
+		f := ExtractFeatures(tc.msg, nil)
+		if f.HasAttachments != tc.want {
+			t.Errorf("msg=%q: HasAttachments got %v, want %v", tc.msg, f.HasAttachments, tc.want)
+		}
+	}
+}
+
+func TestExtractFeatures_HasAttachments_Extension(t *testing.T) {
+	cases := []struct {
+		msg  string
+		want bool
+	}{
+		{"check out photo.jpg", true},
+		{"see screenshot.png", true},
+		{"listen to audio.mp3", true},
+		{"watch clip.mp4", true},
+		{"just a .go file", false},
+		{"document.pdf", false}, // pdf is not in the media list
+	}
+	for _, tc := range cases {
+		f := ExtractFeatures(tc.msg, nil)
+		if f.HasAttachments != tc.want {
+			t.Errorf("msg=%q: HasAttachments got %v, want %v", tc.msg, f.HasAttachments, tc.want)
+		}
+	}
+}
+
+// ── RuleClassifier ───────────────────────────────────────────────────────────
+
+func TestRuleClassifier_ZeroFeatures(t *testing.T) {
+	c := &RuleClassifier{}
+	score := c.Score(Features{})
+	if score != 0.0 {
+		t.Errorf("zero features: got %f, want 0.0", score)
+	}
+}
+
+func TestRuleClassifier_AttachmentsHardGate(t *testing.T) {
+	c := &RuleClassifier{}
+	score := c.Score(Features{HasAttachments: true})
+	if score != 1.0 {
+		t.Errorf("attachments: got %f, want 1.0", score)
+	}
+}
+
+func TestRuleClassifier_CodeBlockAlone(t *testing.T) {
+	c := &RuleClassifier{}
+	// Code block alone = 0.40, above default threshold 0.35
+	score := c.Score(Features{CodeBlockCount: 1})
+	if score < 0.35 {
+		t.Errorf("code block: score %f is below default threshold 0.35", score)
+	}
+}
+
+func TestRuleClassifier_LongMessage(t *testing.T) {
+	c := &RuleClassifier{}
+	// >200 tokens = 0.35, exactly at default threshold → heavy
+	score := c.Score(Features{TokenEstimate: 250})
+	if score < 0.35 {
+		t.Errorf("long message: score %f is below default threshold 0.35", score)
+	}
+}
+
+func TestRuleClassifier_MediumMessage(t *testing.T) {
+	c := &RuleClassifier{}
+	// 50-200 tokens = 0.15, below threshold → light
+	score := c.Score(Features{TokenEstimate: 100})
+	if score >= 0.35 {
+		t.Errorf("medium message: score %f should be below default threshold 0.35", score)
+	}
+}
+
+func TestRuleClassifier_ShortMessage(t *testing.T) {
+	c := &RuleClassifier{}
+	// <50 tokens, no other signals = 0.0 → light
+	score := c.Score(Features{TokenEstimate: 10})
+	if score != 0.0 {
+		t.Errorf("short message: got %f, want 0.0", score)
+	}
+}
+
+func TestRuleClassifier_ToolCallDensity(t *testing.T) {
+	c := &RuleClassifier{}
+
+	scoreNone := c.Score(Features{RecentToolCalls: 0})
+	scoreLow := c.Score(Features{RecentToolCalls: 2})
+	scoreHigh := c.Score(Features{RecentToolCalls: 5})
+
+	if scoreNone != 0.0 {
+		t.Errorf("no tools: got %f, want 0.0", scoreNone)
+	}
+	if scoreLow <= scoreNone {
+		t.Errorf("low tools should score higher than none: %f vs %f", scoreLow, scoreNone)
+	}
+	if scoreHigh <= scoreLow {
+		t.Errorf("high tools should score higher than low: %f vs %f", scoreHigh, scoreLow)
+	}
+}
+
+func TestRuleClassifier_DeepConversation(t *testing.T) {
+	c := &RuleClassifier{}
+	shallow := c.Score(Features{ConversationDepth: 5})
+	deep := c.Score(Features{ConversationDepth: 15})
+	if deep <= shallow {
+		t.Errorf("deep conversation should score higher: %f vs %f", deep, shallow)
+	}
+}
+
+func TestRuleClassifier_ScoreDoesNotExceedOne(t *testing.T) {
+	c := &RuleClassifier{}
+	// Max all signals simultaneously
+	f := Features{
+		TokenEstimate:     500,
+		CodeBlockCount:    3,
+		RecentToolCalls:   10,
+		ConversationDepth: 20,
+	}
+	score := c.Score(f)
+	if score > 1.0 {
+		t.Errorf("score %f exceeds 1.0", score)
+	}
+}
+
+// ── Router ───────────────────────────────────────────────────────────────────
+
+func TestRouter_DefaultThreshold(t *testing.T) {
+	r := New(RouterConfig{LightModel: "gemini-flash"})
+	if r.Threshold() != defaultThreshold {
+		t.Errorf("default threshold: got %f, want %f", r.Threshold(), defaultThreshold)
+	}
+}
+
+func TestRouter_NegativeThresholdFallsBackToDefault(t *testing.T) {
+	r := New(RouterConfig{LightModel: "gemini-flash", Threshold: -0.1})
+	if r.Threshold() != defaultThreshold {
+		t.Errorf("negative threshold: got %f, want %f", r.Threshold(), defaultThreshold)
+	}
+}
+
+func TestRouter_SelectModel_SimpleMessageUsesLight(t *testing.T) {
+	r := New(RouterConfig{LightModel: "gemini-flash", Threshold: 0.35})
+	msg := "hi"
+	model, usedLight, _ := r.SelectModel(msg, nil, "claude-sonnet-4-6")
+	if !usedLight {
+		t.Error("simple message: expected light model to be selected")
+	}
+	if model != "gemini-flash" {
+		t.Errorf("simple message: model got %q, want %q", model, "gemini-flash")
+	}
+}
+
+func TestRouter_SelectModel_CodeBlockUsesPrimary(t *testing.T) {
+	r := New(RouterConfig{LightModel: "gemini-flash", Threshold: 0.35})
+	msg := "```go\nfmt.Println(\"hello\")\n```"
+	model, usedLight, _ := r.SelectModel(msg, nil, "claude-sonnet-4-6")
+	if usedLight {
+		t.Error("code block: expected primary model to be selected")
+	}
+	if model != "claude-sonnet-4-6" {
+		t.Errorf("code block: model got %q, want %q", model, "claude-sonnet-4-6")
+	}
+}
+
+func TestRouter_SelectModel_AttachmentUsesPrimary(t *testing.T) {
+	r := New(RouterConfig{LightModel: "gemini-flash", Threshold: 0.35})
+	msg := "can you analyze this? data:image/png;base64,abc123"
+	model, usedLight, _ := r.SelectModel(msg, nil, "claude-sonnet-4-6")
+	if usedLight {
+		t.Error("attachment: expected primary model to be selected")
+	}
+	if model != "claude-sonnet-4-6" {
+		t.Errorf("attachment: model got %q, want %q", model, "claude-sonnet-4-6")
+	}
+}
+
+func TestRouter_SelectModel_LongMessageUsesPrimary(t *testing.T) {
+	r := New(RouterConfig{LightModel: "gemini-flash", Threshold: 0.35})
+	// >200 token estimate: 210 * 3 = 630 chars
+	msg := strings.Repeat("word ", 210)
+	model, usedLight, _ := r.SelectModel(msg, nil, "claude-sonnet-4-6")
+	if usedLight {
+		t.Error("long message: expected primary model to be selected")
+	}
+	if model != "claude-sonnet-4-6" {
+		t.Errorf("long message: model got %q, want %q", model, "claude-sonnet-4-6")
+	}
+}
+
+func TestRouter_SelectModel_DeepToolChainUsesLight(t *testing.T) {
+	// Tool calls alone (0.25) don't cross the 0.35 threshold — acceptable behavior.
+	// Routing is conservative: only promote to heavy when the signal is unambiguous.
+	r := New(RouterConfig{LightModel: "gemini-flash", Threshold: 0.35})
+	history := []providers.Message{
+		{Role: "assistant", ToolCalls: []providers.ToolCall{{Name: "read_file"}, {Name: "write_file"}}},
+		{Role: "assistant", ToolCalls: []providers.ToolCall{{Name: "exec"}, {Name: "search"}}},
+	}
+	msg := "ok"
+	_, usedLight, _ := r.SelectModel(msg, history, "claude-sonnet-4-6")
+	if !usedLight {
+		t.Error("short message + moderate tool calls: expected light model (score 0.20 < 0.35)")
+	}
+}
+
+func TestRouter_SelectModel_ToolChainPlusMediumUsesHeavy(t *testing.T) {
+	// Tool calls (0.25) + medium message (0.15) = 0.40 >= 0.35 → heavy
+	r := New(RouterConfig{LightModel: "gemini-flash", Threshold: 0.35})
+	history := []providers.Message{
+		{Role: "assistant", ToolCalls: []providers.ToolCall{
+			{Name: "a"}, {Name: "b"}, {Name: "c"}, {Name: "d"},
+		}},
+	}
+	// ~55 tokens * 3 = 165 chars
+	msg := strings.Repeat("word ", 55)
+	_, usedLight, _ := r.SelectModel(msg, history, "claude-sonnet-4-6")
+	if usedLight {
+		t.Error("tool chain + medium message: expected primary model (score >= 0.35)")
+	}
+}
+
+func TestRouter_SelectModel_CustomThreshold(t *testing.T) {
+	// Very low threshold: even a short message triggers heavy model
+	r := New(RouterConfig{LightModel: "gemini-flash", Threshold: 0.05})
+	msg := strings.Repeat("word ", 55) // medium message → 0.15 >= 0.05
+	_, usedLight, _ := r.SelectModel(msg, nil, "claude-sonnet-4-6")
+	if usedLight {
+		t.Error("low threshold: medium message should use primary model")
+	}
+}
+
+func TestRouter_SelectModel_HighThreshold(t *testing.T) {
+	// Very high threshold: even code blocks route to light
+	r := New(RouterConfig{LightModel: "gemini-flash", Threshold: 0.99})
+	msg := "```go\nfmt.Println()\n```"
+	_, usedLight, _ := r.SelectModel(msg, nil, "claude-sonnet-4-6")
+	if !usedLight {
+		t.Error("very high threshold: code block (0.40) should route to light model")
+	}
+}
+
+func TestRouter_LightModel(t *testing.T) {
+	r := New(RouterConfig{LightModel: "my-fast-model", Threshold: 0.35})
+	if r.LightModel() != "my-fast-model" {
+		t.Errorf("LightModel: got %q, want %q", r.LightModel(), "my-fast-model")
+	}
+}
+
+// ── newWithClassifier (internal testing hook) ─────────────────────────────────
+
+type fixedScoreClassifier struct{ score float64 }
+
+func (f *fixedScoreClassifier) Score(_ Features) float64 { return f.score }
+
+func TestRouter_CustomClassifier_LowScore_SelectsLight(t *testing.T) {
+	r := newWithClassifier(
+		RouterConfig{LightModel: "light", Threshold: 0.5},
+		&fixedScoreClassifier{score: 0.2},
+	)
+	_, usedLight, _ := r.SelectModel("anything", nil, "heavy")
+	if !usedLight {
+		t.Error("low score with custom classifier: expected light model")
+	}
+}
+
+func TestRouter_CustomClassifier_HighScore_SelectsPrimary(t *testing.T) {
+	r := newWithClassifier(
+		RouterConfig{LightModel: "light", Threshold: 0.5},
+		&fixedScoreClassifier{score: 0.8},
+	)
+	_, usedLight, _ := r.SelectModel("anything", nil, "heavy")
+	if usedLight {
+		t.Error("high score with custom classifier: expected primary model")
+	}
+}
+
+func TestRouter_CustomClassifier_ExactThreshold_SelectsPrimary(t *testing.T) {
+	// score == threshold → primary (uses >= comparison)
+	r := newWithClassifier(
+		RouterConfig{LightModel: "light", Threshold: 0.5},
+		&fixedScoreClassifier{score: 0.5},
+	)
+	_, usedLight, _ := r.SelectModel("anything", nil, "heavy")
+	if usedLight {
+		t.Error("score == threshold: expected primary model (>= threshold → primary)")
+	}
+}
+
+func TestRouter_SelectModel_ReturnsScore(t *testing.T) {
+	r := newWithClassifier(
+		RouterConfig{LightModel: "light", Threshold: 0.5},
+		&fixedScoreClassifier{score: 0.42},
+	)
+	_, _, score := r.SelectModel("anything", nil, "heavy")
+	if score != 0.42 {
+		t.Errorf("score: got %f, want 0.42", score)
+	}
+}