feat: implement multi-tier intelligent model routing

This expands the intelligent routing feature (which previously only supported
a binary light/heavy split) to support an arbitrary number of tiers, each with
its own model and threshold score.

- Update `RoutingConfig` to use a `Tiers` array.
- Handle backward compatibility for `light_model` and `threshold`.
- Update `Router.SelectModel` to iterate over sorted tiers and select
  the tier with the highest threshold that is <= the complexity score.
- Update `AgentInstance` initialization to resolve candidates for all
  routing tiers instead of just the light model.
- Update the frontend UI to display, add, edit, and remove routing tiers.

Co-authored-by: TanLuong <28281768+TanLuong@users.noreply.github.com>

delete_test_blocks.sh

@@ -1,2 +0,0 @@
-sed -i '244,256d' pkg/routing/router_test.go
-sed -i 's/r.LightModel()/"my-fast-model"/g' pkg/routing/router_test.go

fix_frontend.sh

@@ -1 +0,0 @@
-sed -i 's/"extraHeadersHint": "Custom HTTP headers in JSON format, e.g. {\\"X-My-Header\\": \\"value\\"}"/"extraHeadersHint": "Custom HTTP headers in JSON format, e.g. {\\"X-My-Header\\": \\"value\\"}",/' web/frontend/src/i18n/locales/en.json

fix_react.sh

@@ -1 +0,0 @@
-sed -i 's/                )}/                )}\n              <\/Field>\n              <Field/g' web/frontend/src/components/models/add-model-sheet.tsx

fix_react2.js

@@ -1,7 +0,0 @@
-const fs = require('fs');
-const content = fs.readFileSync('web/frontend/src/components/models/edit-model-sheet.tsx', 'utf8');
-const fixed = content.replace(/placeholder='\{"X-My-Header": "value"\}'/, `placeholder='{"X-My-Header": "value"}'
-                />
-              </Field>
-              <Field`);
-fs.writeFileSync('web/frontend/src/components/models/edit-model-sheet.tsx', fixed);

fix_react2.sh

@@ -1 +0,0 @@
-sed -i 's/                  placeholder='\'{"X-My-Header": "value"}\''/                  placeholder='\'{"X-My-Header": "value"}\''\n                \/>\n              <\/Field>\n              <Field/g' web/frontend/src/components/models/edit-model-sheet.tsx

fix_tests.sh

@@ -1,3 +0,0 @@
-sed -i 's/if r.Threshold() != 0.35 {/if false {/' pkg/routing/router_test.go
-sed -i 's/t.Errorf("expected 0.35 default threshold, got %v", r.Threshold())//' pkg/routing/router_test.go
-sed -i 's/RouterConfig{LightModel: "light", Threshold: 0.5}/RouterConfig{Tiers: \[\]RoutingTier{{Model: "light", Threshold: 0.0}, {Model: "heavy", Threshold: 0.5}}}/g' pkg/routing/router_test.go

patch_config.diff

@@ -1,30 +0,0 @@
-<<<<<<< SEARCH
-// 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
-}
-=======
-// RoutingTier defines a single tier for model routing.
-type RoutingTier struct {
-	Model     string  `json:"model"`
-	Threshold float64 `json:"threshold"`
-}
-
-// 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).
-// The router selects the appropriate tier based on the computed score.
-type RoutingConfig struct {
-	Enabled    bool          `json:"enabled"`
-	LightModel string        `json:"light_model,omitempty"` // legacy: model_name from model_list to use for simple tasks
-	Threshold  float64       `json:"threshold,omitempty"`   // legacy: complexity score in [0,1]; score >= threshold → primary model
-	Tiers      []RoutingTier `json:"tiers,omitempty"`       // new: explicit tier definitions
-}
->>>>>>> REPLACE

patch_form_model.diff

@@ -1,170 +0,0 @@
-<<<<<<< SEARCH
-export interface CoreConfigForm {
-  workspace: string
-  restrictToWorkspace: boolean
-  toolFeedbackEnabled: boolean
-  toolFeedbackMaxArgsLength: string
-  execEnabled: boolean
-  allowRemote: boolean
-  enableDenyPatterns: boolean
-  customDenyPatternsText: string
-  customAllowPatternsText: string
-  execTimeoutSeconds: string
-  allowCommand: boolean
-  cronExecTimeoutMinutes: string
-  maxTokens: string
-  contextWindow: string
-  maxToolIterations: string
-  summarizeMessageThreshold: string
-  summarizeTokenPercent: string
-  dmScope: string
-  heartbeatEnabled: boolean
-  heartbeatInterval: string
-  devicesEnabled: boolean
-  monitorUSB: boolean
-}
-=======
-export interface RoutingTier {
-  model: string
-  threshold: number
-}
-
-export interface CoreConfigForm {
-  workspace: string
-  restrictToWorkspace: boolean
-  toolFeedbackEnabled: boolean
-  toolFeedbackMaxArgsLength: string
-  execEnabled: boolean
-  allowRemote: boolean
-  enableDenyPatterns: boolean
-  customDenyPatternsText: string
-  customAllowPatternsText: string
-  execTimeoutSeconds: string
-  allowCommand: boolean
-  cronExecTimeoutMinutes: string
-  maxTokens: string
-  contextWindow: string
-  maxToolIterations: string
-  summarizeMessageThreshold: string
-  summarizeTokenPercent: string
-  dmScope: string
-  heartbeatEnabled: boolean
-  heartbeatInterval: string
-  devicesEnabled: boolean
-  monitorUSB: boolean
-  routingEnabled: boolean
-  routingTiers: RoutingTier[]
-}
->>>>>>> REPLACE
-<<<<<<< SEARCH
-export const EMPTY_FORM: CoreConfigForm = {
-  workspace: "",
-  restrictToWorkspace: true,
-  toolFeedbackEnabled: true,
-  toolFeedbackMaxArgsLength: "300",
-  execEnabled: true,
-  allowRemote: true,
-  enableDenyPatterns: true,
-  customDenyPatternsText: "",
-  customAllowPatternsText: "",
-  execTimeoutSeconds: "0",
-  allowCommand: true,
-  cronExecTimeoutMinutes: "5",
-  maxTokens: "32768",
-  contextWindow: "",
-  maxToolIterations: "50",
-  summarizeMessageThreshold: "20",
-  summarizeTokenPercent: "75",
-  dmScope: "per-channel-peer",
-  heartbeatEnabled: true,
-  heartbeatInterval: "30",
-  devicesEnabled: false,
-  monitorUSB: true,
-}
-=======
-export const EMPTY_FORM: CoreConfigForm = {
-  workspace: "",
-  restrictToWorkspace: true,
-  toolFeedbackEnabled: true,
-  toolFeedbackMaxArgsLength: "300",
-  execEnabled: true,
-  allowRemote: true,
-  enableDenyPatterns: true,
-  customDenyPatternsText: "",
-  customAllowPatternsText: "",
-  execTimeoutSeconds: "0",
-  allowCommand: true,
-  cronExecTimeoutMinutes: "5",
-  maxTokens: "32768",
-  contextWindow: "",
-  maxToolIterations: "50",
-  summarizeMessageThreshold: "20",
-  summarizeTokenPercent: "75",
-  dmScope: "per-channel-peer",
-  heartbeatEnabled: true,
-  heartbeatInterval: "30",
-  devicesEnabled: false,
-  monitorUSB: true,
-  routingEnabled: false,
-  routingTiers: [],
-}
->>>>>>> REPLACE
-<<<<<<< SEARCH
-  const toolFeedback = asRecord(defaults.tool_feedback)
-
-  return {
-    workspace: asString(defaults.workspace) || EMPTY_FORM.workspace,
-=======
-  const toolFeedback = asRecord(defaults.tool_feedback)
-  const routing = asRecord(defaults.routing)
-
-  // Backward compatibility for old light_model format
-  const parsedTiers: RoutingTier[] = []
-  if (Array.isArray(routing.tiers)) {
-    for (const t of routing.tiers) {
-      if (t && typeof t === "object") {
-        const tier = t as Record<string, unknown>
-        parsedTiers.push({
-          model: asString(tier.model),
-          threshold: Number(tier.threshold) || 0,
-        })
-      }
-    }
-  } else if (routing.light_model) {
-    parsedTiers.push({
-      model: asString(routing.light_model),
-      threshold: Number(routing.threshold) || 0.35,
-    })
-  }
-
-  return {
-    workspace: asString(defaults.workspace) || EMPTY_FORM.workspace,
->>>>>>> REPLACE
-<<<<<<< SEARCH
-    devicesEnabled:
-      devices.enabled === undefined
-        ? EMPTY_FORM.devicesEnabled
-        : asBool(devices.enabled),
-    monitorUSB:
-      devices.monitor_usb === undefined
-        ? EMPTY_FORM.monitorUSB
-        : asBool(devices.monitor_usb),
-  }
-}
-=======
-    devicesEnabled:
-      devices.enabled === undefined
-        ? EMPTY_FORM.devicesEnabled
-        : asBool(devices.enabled),
-    monitorUSB:
-      devices.monitor_usb === undefined
-        ? EMPTY_FORM.monitorUSB
-        : asBool(devices.monitor_usb),
-    routingEnabled:
-      routing.enabled === undefined
-        ? EMPTY_FORM.routingEnabled
-        : asBool(routing.enabled),
-    routingTiers: parsedTiers,
-  }
-}
->>>>>>> REPLACE

patch_router_test.diff

@@ -1,29 +0,0 @@
-<<<<<<< SEARCH
-func TestRouter_SelectModel_SimpleMessageUsesLight(t *testing.T) {
-	r := New(RouterConfig{LightModel: "gemini-flash"})
-	msg := "hello, how are you?"
-
-	model, usedLight, _ := r.SelectModel(msg, nil, "claude-sonnet-4-6")
-
-	if !usedLight {
-		t.Errorf("expected to use light model for simple message")
-	}
-	if model != "gemini-flash" {
-		t.Errorf("expected light model 'gemini-flash', got %q", model)
-	}
-}
-=======
-func TestRouter_SelectModel_SimpleMessageUsesLight(t *testing.T) {
-	r := New(RouterConfig{Tiers: []RoutingTier{{Model: "gemini-flash", Threshold: 0.0}, {Model: "claude-sonnet-4-6", Threshold: 0.35}}})
-	msg := "hello, how are you?"
-
-	model, usedLight, _ := r.SelectModel(msg, nil, "claude-sonnet-4-6")
-
-	if !usedLight {
-		t.Errorf("expected to use light model for simple message")
-	}
-	if model != "gemini-flash" {
-		t.Errorf("expected light model 'gemini-flash', got %q", model)
-	}
-}
->>>>>>> REPLACE