fix: show one Jev routing header per completion (0.15.33)

README.md

@@ -80,7 +80,10 @@ approval flow.
 
 ## Agent Model Routing
 
-Ao usar `jev-router`, a CLI mostra o modelo físico escolhido em cada resposta
+Ao usar `jev-router`, a CLI mostra uma única vez o modelo físico escolhido por completion,
+no primeiro bloco visível de texto ou tool. Blocos de thinking ocultos não criam cabeçalhos,
+e uma nova completion pode mostrar outra escolha. A apresentação funciona também no
+histórico retomado e no transcript. O modelo aparece
 como `jev-router → glm-5.3-flash` (exemplo). O alias solicitado permanece no
 protocolo; a escolha vem do metadado da resposta do roteador.
 

package.json

@@ -1,6 +1,6 @@
 {
   "name": "@verboo/code",
-  "version": "0.15.32",
+  "version": "0.15.33",
   "description": "Verboo Code — coding agent for the Verboo platform",
   "type": "module",
   "bin": {

scripts/e2e/cli.e2e.mjs

@@ -22,6 +22,16 @@ async function metadata(cli) {
 const estimated = cli => cli.frames.some(frame => /~[1-9][\d.,k]* tokens/.test(frame.text))
 const completed = cli => cli.frames.some(frame => frame.text.includes('E2E_COMPLETE'))
 
+for (const fullscreen of [false, true]) test(`routing header appears once with thinking and text, fullscreen=${fullscreen}`, { timeout: 60_000 }, async () => {
+  const cli = await startCli({ columns: 120, rows: 40, fullscreen, model: 'jev-router', routerOptions: { routedCompletion: true }, args: ['oi'] })
+  try {
+    await cli.waitFor(() => cli.screen().includes('E2E_ROUTED_COMPLETE'))
+    assert.equal(cli.screen().match(/jev-router → glm-5\.3-flash/g)?.length, 1)
+    assert.ok(!cli.screen().includes('Synthetic hidden reasoning.'))
+    assert.deepEqual(cli.router.unexpected, [])
+  } finally { await cli.stop() }
+})
+
 for (const fullscreen of [false, true]) for (const [columns, rows] of [[40, 12], [80, 24], [120, 40]]) for (const agents of [1, 2, 8, 20]) {
   test(`installed CLI: ${agents} agents, ${columns}x${rows}, fullscreen=${fullscreen}`, { timeout: 60_000 }, async () => {
     const cli = await startCli({ columns, rows, fullscreen, routerOptions: { agents }, args: ['E2E_PARENT: delegate to the fixture agents.'] })

scripts/e2e/fake-router.mjs

@@ -1,19 +1,20 @@
 import { createServer } from 'node:http'
 import { setTimeout as delay } from 'node:timers/promises'
 
-export async function createFakeRouter({ agents = 2, omitUsage = false, zeroUsage = false, partialUsage = false, stall = false, background = false, childFailure = false, childToolLoop = false } = {}) {
+export async function createFakeRouter({ agents = 2, omitUsage = false, zeroUsage = false, partialUsage = false, stall = false, background = false, childFailure = false, childToolLoop = false, routedCompletion = false } = {}) {
   const requests = []
   const unexpected = []
   let sequence = 0
   const activeAgents = new Set()
   const activeRequests = new Set()
+  const model = routedCompletion ? 'jev-router' : 'fixture-model'
   async function handleRequest(req, res) {
     const path = new URL(req.url, 'http://fixture').pathname
     const json = value => { res.setHeader('Content-Type', 'application/json'); res.end(JSON.stringify(value)) }
     if (path === '/') return json({ ok: true })
     if (path === '/api/claude_code/settings') return json({ settings: {} })
     if (path === '/api/plugins/marketplace.json') return json({ name: 'fixture', owner: { name: 'Fixture' }, plugins: [] })
-    if (path.endsWith('/models')) return json({ data: [{ id: 'fixture-model', object: 'model' }], agent_model_roles: { explore: 'fixture-model', balanced: 'fixture-model', powerful: 'fixture-model' } })
+    if (path.endsWith('/models')) return json({ data: [{ id: model, object: 'model' }], agent_model_roles: { explore: model, balanced: model, powerful: model } })
     if (path === '/api/me') return json({ data: { id: '11111111-1111-4111-8111-111111111111', email: 'fixture@example.test', name: 'Fixture', confirmed: true } })
     if (path === '/api/me/subscriptions') return json({ data: [{ id: '11111111-1111-4111-8111-111111111111', groupId: '22222222-2222-4222-8222-222222222222', status: 'active', currentPeriodEnd: '2099-01-01T00:00:00Z' }] })
     if (path === '/api/me/terms/status') return json({ data: { configured: false, mustAccept: false, pendingReacceptance: false } })
@@ -31,10 +32,14 @@ export async function createFakeRouter({ agents = 2, omitUsage = false, zeroUsag
     }
     if (isChild) { activeAgents.add(id); res.once('close', () => activeAgents.delete(id)) }
     const hasResult = body.messages.some(message => message.role === 'tool')
-    res.writeHead(200, { 'Content-Type': 'text/event-stream', 'Cache-Control': 'no-cache' })
-    const emit = (delta, finish_reason = null, usage) => res.write(`data: ${JSON.stringify({ id, object: 'chat.completion.chunk', model: 'fixture-model', choices: [{ index: 0, delta, finish_reason }], ...(usage && { usage }) })}\n\n`)
+    res.writeHead(200, { 'Content-Type': 'text/event-stream', 'Cache-Control': 'no-cache', ...(routedCompletion && { 'X-Verboo-Selected-Model': 'glm-5.3-flash' }) })
+    const emit = (delta, finish_reason = null, usage) => res.write(`data: ${JSON.stringify({ id, object: 'chat.completion.chunk', model, choices: [{ index: 0, delta, finish_reason }], ...(usage && { usage }) })}\n\n`)
     emit({ role: 'assistant' })
-    if (isChild && childToolLoop) {
+    if (routedCompletion && body.tools?.length) {
+      emit({ reasoning_content: 'Synthetic hidden reasoning.' })
+      emit({ content: 'Oi! E2E_ROUTED_COMPLETE' })
+      emit({}, 'stop', { prompt_tokens: 10, completion_tokens: 4 })
+    } else if (isChild && childToolLoop) {
       emit({ tool_calls: [{ index: 0, id: `read-${sequence}`, type: 'function', function: { name: 'Read', arguments: JSON.stringify({ file_path: 'README.md' }) } }] })
       emit({}, 'tool_calls', { prompt_tokens: 120, completion_tokens: 24 })
     } else if (!isChild && !hasResult && body.tools?.length) {

scripts/e2e/terminal.mjs

@@ -9,7 +9,7 @@ import { spawn, spawnSync } from 'node:child_process'
 import { once } from 'node:events'
 import { createFakeRouter } from './fake-router.mjs'
 
-export async function startCli({ columns = 80, rows = 24, fullscreen = false, usePty = true, args = [], routerOptions } = {}) {
+export async function startCli({ columns = 80, rows = 24, fullscreen = false, usePty = true, model = 'fixture-model', args = [], routerOptions } = {}) {
   const consumer = (await readFile(resolve('.artifacts/package/consumer-path.txt'), 'utf8')).trim()
   const root = resolve('.artifacts/pty')
   await mkdir(root, { recursive: true })
@@ -49,7 +49,7 @@ export async function startCli({ columns = 80, rows = 24, fullscreen = false, us
   terminal.unicode.activeVersion = '11'
   let child
   let stderr = ''
-  const cliArgs = ['--import', pathToFileURL(resolve('scripts/e2e/transport-preload.mjs')).href, join(consumer, 'node_modules/@verboo/code/bin/verboo'), '--model', 'fixture-model', '--dangerously-skip-permissions', '--setting-sources', 'user', '--strict-mcp-config', '--mcp-config', '{"mcpServers":{}}', '--disable-slash-commands', '--debug-file', join(dir, 'debug.log'), ...args]
+  const cliArgs = ['--import', pathToFileURL(resolve('scripts/e2e/transport-preload.mjs')).href, join(consumer, 'node_modules/@verboo/code/bin/verboo'), '--model', model, '--dangerously-skip-permissions', '--setting-sources', 'user', '--strict-mcp-config', '--mcp-config', '{"mcpServers":{}}', '--disable-slash-commands', '--debug-file', join(dir, 'debug.log'), ...args]
   try {
     if (usePty) {
       child = pty.spawn(process.execPath, cliArgs, { cwd: project, env, cols: columns, rows, name: 'xterm-256color' })

src/components/MessageModel.test.tsx

@@ -2,6 +2,7 @@ import { expect, test } from 'bun:test'
 import React from 'react'
 import { renderToString } from '../utils/staticRender.js'
 import { MessageModel } from './MessageModel.js'
+import { routingModelHeaderOwners } from '../utils/routingModelHeaders.js'
 
 test('shows the physical Jev choice during a normal CLI conversation', async () => {
   const message = {
@@ -18,3 +19,45 @@ test('shows the physical Jev choice during a normal CLI conversation', async ()
   )
   expect(rendered).toContain('jev-router → glm-5.3-flash')
 })
+
+function assistant(uuid: string, id: string, content: Array<{ type: string; text?: string; thinking?: string }>, requestId = 'request-one') {
+  return { type: 'assistant', uuid, requestId, message: { id, model: 'ultra/jev-router', metadata: { selected_model: 'glm-5.3' }, content } }
+}
+
+test.each([false, true])('renders one routing header per completion across normalized blocks (transcript=%s)', async isTranscriptMode => {
+  const messages = [
+    assistant('thinking', 'completion-one', [{ type: 'thinking', thinking: 'internal reasoning' }]),
+    assistant('redacted', 'completion-one', [{ type: 'redacted_thinking' }]),
+    assistant('blank', 'completion-one', [{ type: 'text', text: '   ' }]),
+    assistant('first-text', 'completion-one', [{ type: 'text', text: 'Oi!' }]),
+    assistant('second-text', 'completion-one', [{ type: 'text', text: 'Em que posso ajudar?' }]),
+    assistant('tool', 'completion-one', [{ type: 'tool_use' }]),
+    assistant('next-request', 'completion-two', [{ type: 'text', text: 'Outra resposta' }], 'request-two'),
+  ]
+  const owners = routingModelHeaderOwners(messages)
+  expect([...owners]).toEqual(['first-text', 'next-request'])
+  const rendered = await renderToString(<>{messages.map(message =>
+    <MessageModel key={message.uuid} message={message} isTranscriptMode={isTranscriptMode} showRoutingModel={owners.has(message.uuid)} />,
+  )}</>, 100)
+  expect(rendered.match(/ultra\/jev-router → glm-5\.3/g)?.length).toBe(2)
+})
+
+test('resumed and grouped tool blocks preserve completion identity', () => {
+  const first = assistant('tool-first', 'completion-one', [{ type: 'tool_use' }])
+  const later = assistant('tool-later', 'completion-one', [{ type: 'tool_use' }])
+  const messages = [
+    { type: 'grouped_tool_use', uuid: 'group', displayMessage: first, messages: [first, later] },
+    assistant('text-after-tools', 'completion-one', [{ type: 'text', text: 'Done' }]),
+    assistant('next', 'completion-two', [{ type: 'text', text: 'Next' }]),
+  ]
+  expect([...routingModelHeaderOwners(messages)]).toEqual(['group', 'next'])
+  expect([...routingModelHeaderOwners(structuredClone(messages))]).toEqual(['group', 'next'])
+})
+
+test('equal provider message IDs in independent requests are not deduplicated', () => {
+  const messages = [
+    assistant('one', 'provider-id', [{ type: 'text', text: 'One' }], 'request-one'),
+    assistant('two', 'provider-id', [{ type: 'text', text: 'Two' }], 'request-two'),
+  ]
+  expect([...routingModelHeaderOwners(messages)]).toEqual(['one', 'two'])
+})

src/components/MessageModel.tsx

@@ -2,20 +2,20 @@ import React from 'react'
 import { stringWidth } from '../ink/stringWidth.js'
 import { Box, Text } from '../ink.js'
 import type { NormalizedMessage } from '../types/message.js'
+import { hasVisibleModelContent } from '../utils/routingModelHeaders.js'
 type Props = {
   message: NormalizedMessage;
   isTranscriptMode: boolean;
+  showRoutingModel?: boolean;
 };
-export function MessageModel({ message, isTranscriptMode }: Props) {
-  if (
-    message.type !== 'assistant' ||
-    message.message.content.length === 0
-  ) {
+export function MessageModel({ message, isTranscriptMode, showRoutingModel = true }: Props) {
+  if (!hasVisibleModelContent(message)) {
     return null
   }
 
   const requested = message.message.model
   const selected = message.message.metadata?.selected_model
+  if (typeof selected === 'string' && !showRoutingModel) return null
   const display =
     typeof selected === 'string' && selected.length > 0
       ? `${requested} → ${selected}`

src/components/MessageRow.routing.test.tsx

@@ -0,0 +1,27 @@
+import { expect, test } from 'bun:test'
+import React from 'react'
+import { buildMessageLookups } from '../utils/messages.js'
+import { routingModelHeaderOwners } from '../utils/routingModelHeaders.js'
+import { renderToString } from '../utils/staticRender.js'
+import { MessageRow } from './MessageRow.js'
+
+test('a hidden thinking row leaves no routing header or metadata margin', async () => {
+  const messages = ['thinking', 'text'].map(type => ({
+    type: 'assistant', uuid: `block-${type}`, requestId: 'one-request',
+    message: { id: 'one-completion', model: 'ultra/jev-router', metadata: { selected_model: 'glm-5.3' }, content: type === 'thinking' ? [{ type, thinking: 'internal', signature: '' }] : [{ type, text: 'Oi!' }] },
+  }))
+  const owners = routingModelHeaderOwners(messages)
+  const props = {
+    isUserContinuation: false, hasContentAfter: false, tools: [], commands: [], verbose: false,
+    inProgressToolUseIDs: new Set<string>(), streamingToolUseIDs: new Set<string>(),
+    screen: 'prompt', canAnimate: false, lastThinkingBlockId: null, latestBashOutputUUID: null,
+    columns: 100, isLoading: false, lookups: buildMessageLookups(messages, messages),
+  }
+  const hidden = await renderToString(<MessageRow {...props} message={messages[0]} showRoutingModel={false} />, 100)
+  expect(hidden.trim()).toBe('')
+  const rendered = await renderToString(<>{messages.map(message =>
+    <MessageRow key={message.uuid} {...props} message={message} showRoutingModel={owners.has(message.uuid)} />,
+  )}</>, 100)
+  expect(rendered.match(/ultra\/jev-router → glm-5\.3/g)?.length).toBe(1)
+  expect(rendered).toContain('Oi!')
+})

src/components/MessageRow.tsx

@@ -12,6 +12,7 @@ import { MessageModel } from './MessageModel.js';
 import { shouldRenderStatically } from './Messages.js';
 import { MessageTimestamp } from './MessageTimestamp.js';
 import { OffscreenFreeze } from './OffscreenFreeze.js';
+import { hasVisibleModelContent } from '../utils/routingModelHeaders.js';
 export type Props = {
   message: RenderableMessage;
   /** Whether the previous message in renderableMessages is also a user message. */
@@ -22,6 +23,7 @@ export type Props = {
    * if the collapsed group spinner should stay active. Pass `false` otherwise.
    */
   hasContentAfter: boolean;
+  showRoutingModel?: boolean;
   tools: Tools;
   commands: Command[];
   verbose: boolean;
@@ -96,6 +98,7 @@ function MessageRowImpl(t0) {
     message: msg,
     isUserContinuation,
     hasContentAfter,
+    showRoutingModel = true,
     tools,
     commands,
     verbose,
@@ -216,16 +219,10 @@ function MessageRowImpl(t0) {
       }
     }
   }
-  let t5;
-  if ($[34] !== displayMsg || $[35] !== isTranscriptMode) {
-    t5 = displayMsg.type === "assistant" && (isTranscriptMode && displayMsg.message.content.some(_temp) && (displayMsg.timestamp || displayMsg.message.model) || displayMsg.message.content.length > 0 && typeof displayMsg.message.metadata?.selected_model === "string");
-    $[34] = displayMsg;
-    $[35] = isTranscriptMode;
-    $[36] = t5;
-  } else {
-    t5 = $[36];
-  }
-  const hasMetadata = t5;
+  const hasMetadata = hasVisibleModelContent(displayMsg) && (
+    isTranscriptMode && displayMsg.message.content.some(_temp) && (displayMsg.timestamp || displayMsg.message.model) ||
+    showRoutingModel && typeof displayMsg.message.metadata?.selected_model === 'string'
+  );
   const t6 = !hasMetadata;
   const t7 = hasMetadata ? undefined : columns;
   let t8;
@@ -265,8 +262,9 @@ function MessageRowImpl(t0) {
     return t9;
   }
   let t9;
-  if ($[57] !== displayMsg || $[58] !== isTranscriptMode) {
-    t9 = <Box flexDirection="row" justifyContent="flex-end" gap={1} marginTop={1}><MessageTimestamp message={displayMsg} isTranscriptMode={isTranscriptMode} /><MessageModel message={displayMsg} isTranscriptMode={isTranscriptMode} /></Box>;
+  if ($[57] !== displayMsg || $[58] !== isTranscriptMode || $[34] !== showRoutingModel) {
+    t9 = <Box flexDirection="row" justifyContent="flex-end" gap={1} marginTop={1}><MessageTimestamp message={displayMsg} isTranscriptMode={isTranscriptMode} /><MessageModel message={displayMsg} isTranscriptMode={isTranscriptMode} showRoutingModel={showRoutingModel} /></Box>;
+    $[34] = showRoutingModel;
     $[57] = displayMsg;
     $[58] = isTranscriptMode;
     $[59] = t9;
@@ -348,6 +346,7 @@ export function areMessageRowPropsEqual(prev: Props, next: Props): boolean {
 
   // Verbose toggle changes thinking block visibility
   if (prev.verbose !== next.verbose) return false;
+  if (prev.showRoutingModel !== next.showRoutingModel) return false;
 
   // collapsed_read_search is never static in prompt mode (matches shouldRenderStatically)
   if (prev.message.type === 'collapsed_read_search' && next.screen !== 'transcript') {

src/components/Messages.tsx

@@ -83,6 +83,7 @@ const SEND_USER_FILE_TOOL_NAME: string | null = feature('KAIROS') ? (require('..
 
 /* eslint-enable @typescript-eslint/no-require-imports */
 import { VirtualMessageList } from './VirtualMessageList.js';
+import { routingModelHeaderOwners } from '../utils/routingModelHeaders.js';
 
 /**
  * In brief-only mode, filter messages to show ONLY Brief tool_use blocks,
@@ -529,6 +530,8 @@ const MessagesImpl = ({
     };
   }, [verbose, normalizedMessages, isTranscriptMode, syntheticStreamingToolUseMessages, shouldTruncate, tools, isBriefOnly]);
 
+  const routingHeaderOwners = useMemo(() => routingModelHeaderOwners(collapsed_0), [collapsed_0]);
+
   // Cheap slice — only runs when scroll range or slice config changes.
   const renderableMessages = useMemo(() => {
     // Safety cap for the non-virtualized render path. Applied here (not at
@@ -622,7 +625,7 @@ const MessagesImpl = ({
     // streaming instead of waiting for the block to finalize.
     const hasContentAfter = msg_8.type === 'collapsed_read_search' && (!!streamingText || hasContentAfterIndex(renderableMessages, index, tools, streamingToolUseIDs));
     const k_0 = messageKey(msg_8);
-    const row = <MessageRow key={k_0} message={msg_8} isUserContinuation={isUserContinuation} hasContentAfter={hasContentAfter} tools={tools} commands={commands} verbose={verbose || isItemExpanded(msg_8) || cursor?.expanded === true && index === selectedIdx} inProgressToolUseIDs={inProgressToolUseIDs} streamingToolUseIDs={streamingToolUseIDs} screen={screen} canAnimate={canAnimate} onOpenRateLimitOptions={onOpenRateLimitOptions} lastThinkingBlockId={lastThinkingBlockId} latestBashOutputUUID={latestBashOutputUUID} columns={columns} isLoading={isLoading} lookups={lookups_0} />;
+    const row = <MessageRow key={k_0} message={msg_8} isUserContinuation={isUserContinuation} hasContentAfter={hasContentAfter} showRoutingModel={routingHeaderOwners.has(msg_8.uuid)} tools={tools} commands={commands} verbose={verbose || isItemExpanded(msg_8) || cursor?.expanded === true && index === selectedIdx} inProgressToolUseIDs={inProgressToolUseIDs} streamingToolUseIDs={streamingToolUseIDs} screen={screen} canAnimate={canAnimate} onOpenRateLimitOptions={onOpenRateLimitOptions} lastThinkingBlockId={lastThinkingBlockId} latestBashOutputUUID={latestBashOutputUUID} columns={columns} isLoading={isLoading} lookups={lookups_0} />;
 
     // Per-row Provider — only 2 rows re-render on selection change.
     // Wrapped BEFORE divider branch so both return paths get it.

src/services/api/openaiShim.test.ts

@@ -322,14 +322,15 @@ test.each([false, true])('jev-router exposes the selected model in message metad
     const headers = { 'X-Verboo-Selected-Model': 'glm-5.3-flash' }
     if (stream) {
       return makeSseResponse(makeStreamChunks([
+        { choices: [{ delta: { reasoning_content: 'internal reasoning' }, finish_reason: null }] },
         { choices: [{ delta: { content: 'ok' }, finish_reason: null }] },
         { choices: [{ delta: {}, finish_reason: 'stop' }] },
       ]), headers)
     }
     return new Response(JSON.stringify({
       id: 'chatcmpl-selected',
       model: 'jev-router',
-      choices: [{ message: { role: 'assistant', content: 'ok' }, finish_reason: 'stop' }],
+      choices: [{ message: { role: 'assistant', reasoning_content: 'internal reasoning', content: 'ok' }, finish_reason: 'stop' }],
     }), { headers: { ...headers, 'Content-Type': 'application/json' } })
   }) as unknown as FetchType