feat: wire the llm-router into the harness and retire the legacy routing surfaces

console/web/src/components/chat/ChatView.tsx

@@ -18,18 +18,20 @@ import { useConversationsCtxOptional } from '@/lib/conversations-context'
 import { formatStopReason } from '@/lib/format-stop-reason'
 import { newMessageId } from '@/lib/session-id'
 import { cn } from '@/lib/utils'
-import type {
-  AssistantMessage,
-  Conversation,
-  FunctionCallMessage,
-  Message,
-  MessagePatch,
-  Mode,
-  ModelId,
-  ModelOption,
-  SystemMessage,
-  ThoughtMessage,
-  UserMessage,
+import {
+  type AssistantMessage,
+  type Conversation,
+  DEFAULT_THINKING_LEVEL,
+  type FunctionCallMessage,
+  type Message,
+  type MessagePatch,
+  type Mode,
+  type ModelId,
+  type ModelOption,
+  type SystemMessage,
+  type ThinkingLevel,
+  type ThoughtMessage,
+  type UserMessage,
 } from '@/types/chat'
 import { Composer, type ComposerSubmitPayload } from './Composer'
 import { ContextUsage } from './ContextUsage'
@@ -99,6 +101,9 @@ export function ChatView({
   onCompactConversation,
 }: ChatViewProps) {
   const [isStreaming, setIsStreaming] = useState(false)
+  const [thinkingLevel, setThinkingLevel] = useState<ThinkingLevel>(
+    DEFAULT_THINKING_LEVEL,
+  )
   const abortRef = useRef<AbortController | null>(null)
   const [copied, setCopied] = useState(false)
   const { functionEntries } = useFunctionsCatalog(backend.id)
@@ -303,7 +308,7 @@ export function ChatView({
           payload.text || '(attachments only)',
           conversation.mode,
           model,
-          { signal: controller.signal, sessionId, messageId },
+          { signal: controller.signal, sessionId, messageId, thinkingLevel },
         )) {
           switch (event.kind) {
             case 'thought-start': {
@@ -530,6 +535,7 @@ export function ChatView({
       conversation.id,
       conversation.mode,
       conversation.model,
+      thinkingLevel,
       sessionId,
       contextWindow,
       backend,
@@ -680,6 +686,8 @@ export function ChatView({
             functionEntries={functionEntries}
             permissionMode={approvalSettings.settings.mode}
             permissionModeLoading={!approvalSettings.loaded}
+            thinkingLevel={thinkingLevel}
+            onThinkingLevelChange={setThinkingLevel}
             onModeChange={(next) => onUpdateMode(conversation.id, next)}
             onModelChange={(next) => onUpdateModel(conversation.id, next)}
             onPermissionModeChange={(next) =>

console/web/src/components/chat/Composer.stories.tsx

@@ -69,6 +69,8 @@ function ComposerHarness({
       modelOptions={STORY_MODEL_OPTIONS}
       functionEntries={STATIC_FUNCTIONS}
       permissionMode="manual"
+      thinkingLevel="off"
+      onThinkingLevelChange={fn()}
       onModeChange={setMode}
       onModelChange={setModel}
       onPermissionModeChange={fn()}

console/web/src/components/chat/Composer.tsx

@@ -4,7 +4,15 @@ import { PermissionModePicker } from '@/components/permissions/PermissionModePic
 import { Button } from '@/components/ui/Button'
 import type { PermissionMode } from '@/lib/backend/approval-settings'
 import type { FunctionEntry } from '@/lib/functions'
-import type { Attachment, Mode, ModelId, ModelOption } from '@/types/chat'
+import { Select } from '@/components/ui/Select'
+import {
+  type Attachment,
+  type Mode,
+  type ModelId,
+  type ModelOption,
+  THINKING_LEVELS,
+  type ThinkingLevel,
+} from '@/types/chat'
 import { AttachmentButton } from './AttachmentButton'
 import { AttachmentChip } from './AttachmentChip'
 import { LexicalShell } from './LexicalShell'
@@ -28,8 +36,10 @@ interface ComposerProps {
    */
   permissionMode: PermissionMode
   permissionModeLoading?: boolean
+  thinkingLevel: ThinkingLevel
   onModeChange: (next: Mode) => void
   onModelChange: (next: ModelId) => void
+  onThinkingLevelChange: (next: ThinkingLevel) => void
   onPermissionModeChange: (next: PermissionMode) => void
   onSubmit: (payload: ComposerSubmitPayload) => void
   onStop?: () => void
@@ -52,8 +62,10 @@ export function Composer({
   catalogLoading,
   permissionMode,
   permissionModeLoading,
+  thinkingLevel,
   onModeChange,
   onModelChange,
+  onThinkingLevelChange,
   onPermissionModeChange,
   onSubmit,
   onStop,
@@ -133,6 +145,16 @@ export function Composer({
           disabled={inputDisabled || !!permissionModeLoading}
         />
         <div className="flex-1 min-w-0" />
+        <Select<ThinkingLevel>
+          value={thinkingLevel}
+          options={THINKING_LEVELS.map((l) => ({
+            value: l,
+            label: l === 'off' ? 'thinking off' : `thinking ${l}`,
+          }))}
+          onChange={onThinkingLevelChange}
+          disabled={inputDisabled}
+          aria-label="thinking level"
+        />
         <ModelPicker
           value={model}
           options={modelOptions}

console/web/src/components/chat/MessageList.tsx

@@ -201,7 +201,7 @@ function resolveEmptyState(
     onInstallHarness: harnessStatus.install,
     onRetryInstall: harnessStatus.retry,
     onConfigureProvider: () => {
-      window.location.hash = '#/configuration/workers/harness'
+      window.location.hash = '#/configuration/workers/llm-router'
     },
   }
 

console/web/src/components/chat/ModelPicker.tsx

@@ -9,10 +9,10 @@ import {
   type ModelOption,
 } from '@/types/chat'
 
-// Deep link to the harness configuration entry in the workers/config editor,
+// Deep link to the llm-router configuration entry in the workers/config editor,
 // where api keys + per-provider settings are now edited (the bespoke
 // per-provider dialog was retired in favour of the schema-driven form).
-const HARNESS_CONFIG_HASH = '#/configuration/workers/harness'
+const HARNESS_CONFIG_HASH = '#/configuration/workers/llm-router'
 
 interface ModelPickerProps {
   value: ModelId | null
@@ -47,7 +47,7 @@ export function ModelPicker({
   // Optional: present in the app, absent in isolated Storybook renders.
   const ctx = useConversationsCtxOptional()
 
-  // Providers present as harness workers (from harness::provider::list).
+  // Providers present as workers (from router::provider::list).
   // Absent in Storybook or before the list resolves, in which case no empty
   // provider groups or gears appear until the dynamic list arrives.
   const presentIds = ctx?.presentProviders.map((p) => p.id) ?? []

console/web/src/components/chat/ThoughtMessage.tsx

@@ -15,8 +15,13 @@ function thoughtLabel(durationMs: number): string {
 
 export function ThoughtMessage({ message, defaultOpen }: ThoughtMessageProps) {
   const streaming = !!message.streaming
+  // Auto-open while the thought streams so reasoning is visible in real
+  // time; the flip back when streaming ends collapses it to its summary.
   return (
-    <details className="iii-details group/thought" open={defaultOpen}>
+    <details
+      className="iii-details group/thought"
+      open={defaultOpen || streaming}
+    >
       <summary
         className={cn(
           'inline-flex items-center gap-2 font-mono text-[12px] text-ink-faint hover:text-ink transition-colors lowercase select-none',

console/web/src/hooks/use-model-picker-source.ts

@@ -18,7 +18,7 @@ import type { ModelOption } from '@/types/chat'
  * shows present-but-unconfigured providers as gear groups). When the engine is
  * unreachable (catalog fetch throws) the list stays empty.
  *
- * `presentProviders` comes from `harness::provider::list` so the picker can
+ * `presentProviders` comes from `router::provider::list` so the picker can
  * surface providers that exist as workers but aren't configured yet.
  * `refresh()` re-reads both on demand (after a provider `refresh_models`).
  *

console/web/src/lib/backend/real.ts

@@ -111,6 +111,9 @@ async function* realStream(
             provider,
             model: modelId,
             mode,
+            ...(opts?.thinkingLevel && opts.thinkingLevel !== 'off'
+              ? { thinking_level: opts.thinkingLevel }
+              : {}),
             messages: [
               {
                 role: 'user',

console/web/src/lib/backend/types.ts

@@ -83,6 +83,12 @@ export type StreamEvent =
 
 export interface ChatStreamOptions {
   signal?: AbortSignal
+  /**
+   * Reasoning/thinking level for the turn. Sent to `run::start` as
+   * `thinking_level`; omitted when 'off' or absent. The provider degrades
+   * with a warning when the model can't honor it.
+   */
+  thinkingLevel?: import('@/types/chat').ThinkingLevel
   /** mean delay between assistant tokens, in ms */
   meanDelayMs?: number
   /**

console/web/src/lib/harness-config-events.ts

@@ -1,5 +1,5 @@
-/** Fired after `configuration::set` succeeds for the harness entry. */
-const HARNESS_CONFIG_ID = 'harness'
+/** Fired after `configuration::set` succeeds for the llm-router entry (provider credentials + settings). */
+const LLM_ROUTER_CONFIG_ID = 'llm-router'
 
 type Listener = () => void
 
@@ -13,7 +13,7 @@ export function onHarnessConfigSaved(listener: Listener): () => void {
 }
 
 export function notifyHarnessConfigSaved(configId: string): void {
-  if (configId !== HARNESS_CONFIG_ID) return
+  if (configId !== LLM_ROUTER_CONFIG_ID) return
   for (const listener of listeners) {
     listener()
   }

console/web/src/lib/models-catalog.ts

@@ -2,7 +2,7 @@ import { makeCatalogModelKey } from '@/lib/catalog-model-key'
 import { getIiiClient } from '@/lib/iii-client'
 import type { ModelOption } from '@/types/chat'
 
-/** Wire shape returned by `models::list` over the iii bus. */
+/** Wire shape returned by `router::models::list` over the iii bus. */
 export interface CatalogModelRow {
   id: string
   provider: string
@@ -12,7 +12,7 @@ export interface CatalogModelRow {
 
 export async function fetchModelsCatalog(): Promise<CatalogModelRow[]> {
   const client = await getIiiClient()
-  const res = await client.call<{ models?: unknown }>('models::list', {})
+  const res = await client.call<{ models?: unknown }>('router::models::list', {})
   const rows = res?.models
   if (!Array.isArray(rows)) return []
   const out: CatalogModelRow[] = []
@@ -50,7 +50,7 @@ export function catalogRowsToModelOptions(
  * Ask each provider to re-pull its upstream model list into the catalog via
  * `provider::<id>::refresh_models`. Best-effort and parallel — a provider
  * that's offline or has no credential simply registers nothing. Callers
- * re-read `models::list` afterwards to pick up the refreshed catalog.
+ * re-read `router::models::list` afterwards to pick up the refreshed catalog.
  */
 export async function refreshProviderModels(
   providers: readonly string[],
@@ -96,7 +96,7 @@ export async function subscribeModelChanges(
   }
 }
 
-/** A provider declared to the harness, from `harness::provider::list`. */
+/** A provider declared to the harness, from `router::provider::list`. */
 export interface ProviderListEntry {
   id: string
   display_name: string
@@ -111,7 +111,7 @@ export interface ProviderListEntry {
 export async function fetchProviderList(): Promise<ProviderListEntry[]> {
   const client = await getIiiClient()
   const res = await client.call<{ providers?: unknown }>(
-    'harness::provider::list',
+    'router::provider::list',
     {},
   )
   const rows = res?.providers

console/web/src/lib/providers.ts

@@ -1,7 +1,7 @@
 /**
  * Client-side validation helpers + error normalization shared by provider
  * surfaces. Credentials and per-provider settings are now edited through the
- * schema-driven `configuration` form (harness entry), so the former
+ * schema-driven `configuration` form (llm-router entry), so the former
  * `auth::*` / `provider_config::*` bus wrappers were removed — the
  * `configuration::*` calls in `WorkersTab/api.ts` cover that path.
  */

console/web/src/pages/Configuration/tabs/ConsoleSettingsTab.tsx

@@ -19,9 +19,9 @@ import {
   saveDefaultAllowlist,
 } from '@/lib/storage'
 
-// Provider credentials + settings now live in the harness `configuration`
+// Provider credentials + settings now live in the llm-router `configuration`
 // entry, edited via the schema-driven form on the workers tab.
-const HARNESS_CONFIG_HASH = '#/configuration/workers/harness'
+const HARNESS_CONFIG_HASH = '#/configuration/workers/llm-router'
 
 interface ConsoleSettingsTabProps {
   theme: Theme

console/web/src/types/chat.ts

@@ -19,6 +19,20 @@ export const MODES: { id: Mode; label: string }[] = [
 
 export const DEFAULT_MODE: Mode = 'agent'
 
+/** Reasoning effort sent to run::start as `thinking_level`; 'off' is omitted. */
+export type ThinkingLevel = 'off' | 'minimal' | 'low' | 'medium' | 'high' | 'xhigh'
+
+export const THINKING_LEVELS: ThinkingLevel[] = [
+  'off',
+  'minimal',
+  'low',
+  'medium',
+  'high',
+  'xhigh',
+]
+
+export const DEFAULT_THINKING_LEVEL: ThinkingLevel = 'off'
+
 export type Role = 'user' | 'assistant' | 'thought' | 'function-call'
 
 export interface Attachment {

harness/package.json

@@ -2,7 +2,7 @@
   "name": "harness",
   "version": "0.5.7",
   "private": true,
-  "description": "Node port of the iii harness stack: harness, approval-gate, turn-orchestrator, llm-budget, providers, and side-cars. Conversations persist in the external session-manager worker.",
+  "description": "Node port of the iii harness stack: harness, approval-gate, turn-orchestrator, llm-budget, and side-cars. Conversations persist in the external session-manager worker; LLM providers are standalone llm-router plugin workers.",
   "license": "Apache-2.0",
   "type": "module",
   "engines": {
@@ -26,12 +26,6 @@
     "dev:turn-orchestrator": "tsx src/turn-orchestrator/main.ts",
     "dev:llm-budget": "tsx src/llm-budget/main.ts",
     "dev:hook-fanout": "tsx src/hook-fanout/main.ts",
-    "dev:models-catalog": "tsx src/models-catalog/main.ts",
-    "dev:provider-anthropic": "tsx src/provider-anthropic/main.ts",
-    "dev:provider-openai": "tsx src/provider-openai/main.ts",
-    "dev:provider-kimi": "tsx src/provider-kimi/main.ts",
-    "dev:provider-lmstudio": "tsx src/provider-lmstudio/main.ts",
-    "dev:provider-llamacpp": "tsx src/provider-llamacpp/main.ts",
     "dev:context-compaction": "tsx src/context-compaction/main.ts",
     "dev:web": "tsx src/web/main.ts"
   },
@@ -42,12 +36,6 @@
     "iii-turn-orchestrator": "./dist/turn-orchestrator/main.js",
     "iii-llm-budget": "./dist/llm-budget/main.js",
     "iii-hook-fanout": "./dist/hook-fanout/main.js",
-    "iii-models-catalog": "./dist/models-catalog/main.js",
-    "iii-provider-anthropic": "./dist/provider-anthropic/main.js",
-    "iii-provider-openai": "./dist/provider-openai/main.js",
-    "iii-provider-kimi": "./dist/provider-kimi/main.js",
-    "iii-provider-lmstudio": "./dist/provider-lmstudio/main.js",
-    "iii-provider-llamacpp": "./dist/provider-llamacpp/main.js",
     "iii-context-compaction": "./dist/context-compaction/main.js",
     "iii-web": "./dist/web/main.js"
   },

harness/src/context-compaction/model-resolver.ts

@@ -30,32 +30,37 @@ export async function fetchModelLimit(
   modelID: string,
 ): Promise<ResolvedModel> {
   try {
+    // router::models::get: payload key is `id`, result is wrapped as
+    // `{ model }`, null on a catalog miss (the cold-window signal).
     const entry = await iii.trigger<
       unknown,
       {
-        id?: string;
-        provider?: string;
-        context_window?: number;
-        max_output_tokens?: number;
+        model?: {
+          id?: string;
+          provider?: string;
+          context_window?: number;
+          max_output_tokens?: number;
+        } | null;
       } | null
     >({
-      function_id: 'models::get',
-      payload: { provider: providerID, model_id: modelID },
+      function_id: 'router::models::get',
+      payload: { provider: providerID, id: modelID },
       timeoutMs: 5_000,
     });
 
-    if (!entry) {
+    const model = entry?.model ?? null;
+    if (!model) {
       logger.debug('model-resolver: model not found in catalog', { providerID, modelID });
       return null;
     }
 
     return {
       providerID,
       modelID,
-      modelLimit: limitFromModel(entry),
+      modelLimit: limitFromModel(model),
     };
   } catch (err) {
-    logger.debug('model-resolver: models::get failed', {
+    logger.debug('model-resolver: router::models::get failed', {
       providerID,
       modelID,
       err: String(err),