fix: preserve exact answers and native model region settings

Clarify that completion receipts are stripped transport metadata, preserving user-visible exact output. Include validated dynamic Vertex model region keys in shell collection, cache identity and native Claude environment without changing packaged Electron fuses.

Validation: natural discussion/receipt 56 passed; final readiness/main 91 passed; real Codex and OpenClaw exact-answer runs completed; real Electron shell extraction passed. Full release checks remain pending.

desktop/src/agents/local-agent-readiness.cjs

@@ -30,6 +30,7 @@ const MAX_CREDENTIAL_FILE_BYTES = 2 * 1024 * 1024
 const MAX_SHELL_ENV_BYTES = 256 * 1024
 const SHELL_ENV_CACHE_TTL_MS = 30000
 const SHELL_ENV_MARKER = '__MELDWORK_NATIVE_ENV_V1__'
+const VERTEX_MODEL_REGION_KEY = /^VERTEX_REGION_[A-Z0-9_]{1,120}$/
 const CREDENTIAL_ENV_KEYS = Object.freeze({
   codex: ['OPENAI_API_KEY'],
   hermes: ['OPENAI_API_KEY', 'OPENROUTER_API_KEY', 'ANTHROPIC_API_KEY'],
@@ -160,7 +161,9 @@ function textContainsCredential(filename) {
 
 function nativeCredentialEnvironment(kind, env = process.env, options = {}) {
   const result = {}
-  for (const key of RUNTIME_ENV_KEYS[kind] || []) {
+  const keys = [...(RUNTIME_ENV_KEYS[kind] || []), ...(kind === 'claude'
+    ? Object.keys(env).filter(key => VERTEX_MODEL_REGION_KEY.test(key)) : [])]
+  for (const key of keys) {
     if (options.providerConfigured && !NATIVE_CONFIG_ROOT_KEYS.has(key)) continue
     if (typeof env[key] === 'string') result[key] = env[key]
   }
@@ -178,7 +181,7 @@ function nativeCredentialKeyEnvironment(kind, env = process.env) {
 function allowedShellEnvironment(env = process.env, platform = process.platform) {
   const source = { ...env, ...networkEnvironment(env, platform) }
   const result = {}
-  for (const key of SHELL_ENV_KEYS) {
+  for (const key of [...SHELL_ENV_KEYS, ...Object.keys(source).filter(key => VERTEX_MODEL_REGION_KEY.test(key))]) {
     const value = source[key]
     if (typeof value !== 'string' || value.length > MAX_SHELL_ENV_BYTES) continue
     result[key] = value
@@ -186,11 +189,27 @@ function allowedShellEnvironment(env = process.env, platform = process.platform)
   return result
 }
 
-function nativeShellCommand() {
+function nativeShellCommand(shellName) {
   const keys = SHELL_ENV_KEYS.join(' ')
+  // Enumerate names only; do not dump unrelated values or depend on Electron RunAsNode.
+  const regionNames = shellName === 'zsh'
+    ? 'builtin print -rl -- ${(k)parameters[(I)VERTEX_REGION_*]}'
+    : [
+        'if [ -n "${BASH_VERSION-}" ]; then',
+        '  builtin compgen -e VERTEX_REGION_',
+        'elif [ -x /bin/bash ]; then',
+        "  BASH_ENV= ENV= /bin/bash --noprofile --norc -c 'builtin compgen -e VERTEX_REGION_'",
+        'fi',
+      ].join('\n')
   return [
     `printf '${SHELL_ENV_MARKER}\\0'`,
-    `for __meldwork_key in ${keys}; do`,
+    `for __meldwork_key in ${keys} $(${regionNames}); do`,
+    '  case "$__meldwork_key" in ""|*[!A-Za-z0-9_]*) continue ;; esac',
+    '  case "$__meldwork_key" in VERTEX_REGION_*)',
+    '    __meldwork_model=${__meldwork_key#VERTEX_REGION_}',
+    '    case "$__meldwork_model" in ""|*[!A-Z0-9_]*) continue ;; esac',
+    '    [ ${#__meldwork_model} -le 120 ] || continue ;;',
+    '  esac',
     '  eval "__meldwork_value=\\${$__meldwork_key-}"',
     '  eval "__meldwork_set=\\${$__meldwork_key+x}"',
     '  if [ -n "$__meldwork_set" ]; then',
@@ -212,7 +231,8 @@ function parseNativeShellEnvironment(output) {
     if (separator <= 0) continue
     const key = entry.slice(0, separator)
     const value = entry.slice(separator + 1)
-    if (!allowed.has(key) || value.length > MAX_SHELL_ENV_BYTES) continue
+    if ((!allowed.has(key) && !VERTEX_MODEL_REGION_KEY.test(key))
+        || value.length > MAX_SHELL_ENV_BYTES) continue
     result[key] = value
   }
   return result
@@ -237,7 +257,7 @@ async function queryNativeShellEnvironment(options = {}) {
   try {
     const result = await execFileFn(
       shell,
-      [shellName === 'sh' ? '-lc' : '-lic', nativeShellCommand()],
+      [shellName === 'sh' ? '-lc' : '-lic', nativeShellCommand(shellName)],
       {
         timeout: 5000,
         maxBuffer: MAX_SHELL_ENV_BYTES,

desktop/src/workspace/local-workspace-auto-runner.cjs

@@ -2704,7 +2704,7 @@ class LocalWorkspaceAutoRunner {
           ].join('\n')
         : '',
       phase === 'discussion'
-        ? 'After your natural answer append exactly one receipt: [[MELDWORK_COLLABORATION:{"summary":"concise result","taskDecision":{"status":"continue","reason":"what you accomplished and what remains","deliverables":[],"nextKinds":[]}}]]. List actual deliverables when available. Only the current delivery owner judges the whole task; other participants report their contribution and may request peers with nextKinds. Merge any owner judgment or handoffTo into this same receipt. Append nothing after it.'
+        ? 'After your natural answer append exactly one receipt: [[MELDWORK_COLLABORATION:{"summary":"concise result","taskDecision":{"status":"continue","reason":"what you accomplished and what remains","deliverables":[],"nextKinds":[]}}]]. This receipt is transport metadata: Meldwork removes it before displaying your answer. User requirements such as exact text, one line, JSON only, or no explanation apply to the visible answer before the receipt; preserve that requested answer exactly and still append the receipt separately. Do not explain the receipt in the visible answer. List actual deliverables when available. Only the current delivery owner judges the whole task; other participants report their contribution and may request peers with nextKinds. Merge any owner judgment or handoffTo into this same receipt. Append nothing after it.'
         : '',
     ].filter(Boolean).join('\n\n')
   }

desktop/test/agents/local-agent-readiness.test.cjs

@@ -869,6 +869,59 @@ test('real shell extraction preserves explicit empty cloud and credential overri
   assert.equal(fallback.env.ANTHROPIC_API_KEY, '')
 })
 
+test('model-specific Vertex regions participate in cache identity and stay scoped to the native Agent', async () => {
+  const key = 'VERTEX_REGION_CLAUDE_FUTURE_MODEL'
+  const base = { platform: 'win32', home: 'C:\\vertex-region-test', env: { PATH: 'C:\\bin', [key]: 'us-east5' } }
+  const before = await resolveNativeShellEnvironment(base)
+  const after = await resolveNativeShellEnvironment({ ...base, env: { ...base.env, [key]: '' } })
+  assert.equal(before.env[key], 'us-east5')
+  assert.equal(after.env[key], '')
+  assert.deepEqual(nativeCredentialEnvironment('claude', after.env), { [key]: '' })
+  assert.deepEqual(nativeCredentialEnvironment('claude', before.env, { providerConfigured: true }), {})
+  assert.deepEqual(nativeCredentialEnvironment('gemini', before.env), {})
+  assert.deepEqual(nativeCredentialEnvironment('codex', before.env), {})
+  const removed = await resolveNativeShellEnvironment({ ...base, env: { PATH: base.env.PATH } })
+  assert.equal(removed.env[key], undefined)
+  assert.deepEqual(nativeCredentialEnvironment('claude', {
+    VERTEX_REGION_: 'empty-model', VERTEX_REGION_lowercase: 'invalid',
+    'VERTEX_REGION_MODEL;bad': 'invalid', NODE_OPTIONS: '--require untrusted',
+  }), {})
+})
+
+for (const shell of ['/bin/sh', '/bin/bash', '/bin/zsh']) {
+  test(`native shell ${shell} extracts exported dynamic regions without unrelated values or value evaluation`, {
+    skip: process.platform === 'win32' || !fs.existsSync(shell),
+  }, async t => {
+    const home = fs.mkdtempSync(path.join(os.tmpdir(), 'meldwork-region-'))
+    t.after(() => fs.rmSync(home, { recursive: true, force: true }))
+    const execFile = require('node:util').promisify(require('node:child_process').execFile)
+    let output = ''
+    const resolved = await resolveNativeShellEnvironment({
+      cache: false, home, shell, env: { PATH: '/usr/bin:/bin' },
+      execFileFn: async (command, args, options) => {
+        const result = await execFile(command, ['-c', [
+          "export VERTEX_REGION_CLAUDE_FUTURE_MODEL=us-east5",
+          "export VERTEX_REGION_CLAUDE_EMPTY=''",
+          "export VERTEX_REGION_LITERAL='$(touch should-not-exist)'",
+          "export MELDWORK_UNRELATED='synthetic-unrelated-secret'",
+          "export NODE_OPTIONS='--require /nonexistent/untrusted-module.cjs'",
+          args[1],
+        ].join('\n')], { ...options, cwd: home })
+        output = result.stdout
+        return result
+      },
+    })
+    assert.equal(resolved.source, 'native-shell')
+    assert.equal(resolved.env.VERTEX_REGION_CLAUDE_FUTURE_MODEL, 'us-east5')
+    assert.equal(resolved.env.VERTEX_REGION_CLAUDE_EMPTY, '')
+    assert.equal(resolved.env.VERTEX_REGION_LITERAL, '$(touch should-not-exist)')
+    assert.equal(resolved.env.MELDWORK_UNRELATED, undefined)
+    assert.equal(resolved.env.NODE_OPTIONS, undefined)
+    assert.equal(output.includes('synthetic-unrelated-secret'), false)
+    assert.equal(fs.existsSync(path.join(home, 'should-not-exist')), false)
+  })
+}
+
 test('native cloud configuration is forwarded only to compatible Agents', () => {
   const cloud = {
     CLAUDE_CODE_USE_VERTEX: '1', AWS_PROFILE: 'native', AWS_ACCESS_KEY_ID: 'access-key',

desktop/test/workspace/local-workspace-v4-natural-discussion.test.cjs

@@ -138,6 +138,33 @@ test('Natural routing retains a request beyond the stored message excerpt', asyn
   assert.deepEqual(turns, ['codex', 'hermes', 'codex'])
 })
 
+for (const discussionStyle of ['agent-led', 'sequential']) {
+  for (const answer of ['EXACT_ANSWER_OK', '{"ok":true}']) {
+    test(`Natural ${discussionStyle} preserves exact visible ${answer.startsWith('{') ? 'JSON' : 'text'} separately from completion`, async t => {
+      const { directory, options } = fixture()
+      t.after(() => fs.rmSync(directory, { recursive: true, force: true }))
+      const ledger = new RunLedger({ storagePath: path.join(directory, 'run-ledger.json') })
+      const workspace = new LocalWorkspace({
+        ...options, runLedger: ledger, naturalAgentResponses: true,
+        runAgent: async (_agent, prompt) => ({ outcome: 'completed',
+          text: naturalPhase(prompt) === 'proposal' ? answer : decisionReply(answer) }),
+      })
+      await workspace.refreshAgents()
+      const group = workspace.createGroup({ name: 'Exact answer', agentKinds: ['codex'], workdir: directory, allowWrite: false })
+      const controller = await runDiscussion(workspace, group, {
+        text: `Reply with exactly ${answer} and no other visible text.`, discussionStyle, maxRounds: 4,
+      })
+      assert.equal(ledger.get(controller.runId).status, 'completed')
+      const replies = workspace.snapshot().messages.filter(message => message.role === 'agent')
+      assert.ok(replies.length > 0)
+      assert.ok(replies.every(message => message.content === answer))
+      const restarted = new LocalWorkspace({ ...options, runLedger: new RunLedger({ storagePath: ledger.storagePath }) })
+      assert.deepEqual(restarted.snapshot().messages.filter(message => message.role === 'agent').map(message => message.content), replies.map(message => message.content))
+      assert.equal(ledger.get(controller.runId).agentRuns.at(-1).context.taskDecision.status, 'completed')
+    })
+  }
+}
+
 for (const status of ['completed', 'blocked']) {
   test(`Natural single-Agent task persists the owner's ${status} judgment`, async (t) => {
     const { directory, options } = fixture()

docs/releases/Meldwork-V1.0.5-progress.md

@@ -33,7 +33,7 @@
 ## 待完成
 
 1. 群聊写权限：自然讨论的实际写入与未知结果恢复已完成本轮检查点，见末节；仍需纳入最终全量测试与打包验收。
-2. 群聊收尾：自然讨论已使用 AI 完成判断，`needs-human` 已接入可恢复输入 Gate，详见末节。严格限定只输出一行时，真实 Codex 曾漏回执而成为 partial；完成契约的可用性和更多真实多 Agent 场景仍需验证。人类采用与 AI 完成继续分开。
+2. 群聊收尾：自然讨论已使用 AI 完成判断，`needs-human` 已接入可恢复输入 Gate。严格一行输出曾导致真实 Codex 漏回执；澄清回执与可见答案的边界后，Codex/OpenClaw 原要求复测均通过，见末节。仍需最终综合验收；人类采用与 AI 完成继续分开。
 3. 群聊路由及上下文：已限制历史预算、恢复长回复末尾、按原生会话确认状态增量发送，并使用明确 nextKinds 派发替代正文 @ 正则；详见相应检查点，最终版本仍需综合验收。
 4. 启动检测：继续检查环境恢复、探测超时、Keychain 状态与执行事实一致性；首次扫描等待仍长。
 5. 通用性：继续落实 review 中与当前范围有关的品牌/Skill/媒体词路由问题，业务核心不依赖指定 Agent；Pi 空能力探测已收紧。
@@ -244,3 +244,25 @@ nextKinds 使用最多 32 个唯一标识，与现有 V4 成员上限一致；
 - 最终 `MELDWORK_TEST_CLAUDE_EXECUTABLE=/Users/rydersun/.local/opt/npm-global/bin/claude npm --prefix desktop test` 全量 1554/1554，失败和跳过均为 0，退出 0，约 321 秒；日志 `/tmp/meldwork-105-native-media-desktop.log`。测试启动后没有修改产品代码。`git diff --check` 通过，main/workspace 中已无媒体关键词解析与共享生成接线。
 
 本检查点未更新包版本、重建前端或打包，也未替换日常应用。CLI 环境覆盖、完成回执缺失的恢复体验及最终发布验收仍待完成。
+
+## 精确可见答案与完成回执
+
+2026-09-09：自然讨论的通用提示明确回执是传输元数据，展示前会剥离；用户的一行、精确文本、JSON-only 等格式约束适用于回执前的可见答案。没有降低完成判断要求、强制判成功或增加固定复核 Agent，也没有重放已有写入。
+
+- 自然讨论与回执测试 56/56，新增 agent-led/sequential 的精确文本和 JSON 持久化验证：可见正文不混入回执，重新加载保持原答案，账本仍保存 AI 判断。
+- 将真实验证任务恢复为此前失败的原始“answer exactly NATIVE_MEDIA_ROUTING_105_OK”要求，不向用户任务添加完成协议提示。Codex 配置 `/tmp/meldwork-105-native-media-jmhpam`、OpenClaw 配置 `/tmp/meldwork-105-native-media-L8HULT` 均在两次调用后 completed；可见答案精确，共享媒体调用 0，无生成附件，脚本退出 0，Electron 正常关闭。
+- 已查看 `/tmp/meldwork-105-exact-answer-completed.png` 和 `/tmp/meldwork-105-exact-answer-openclaw.png`。该结果证明上述真实样本通过，不承诺任何模型都不会漏回执；缺失判断仍不被系统伪造为成功。
+
+## 动态模型区域配置
+
+2026-09-09：原生环境采集、缓存身份与 Claude 协议适配允许格式受限的 VERTEX_REGION_<MODEL>，不再依赖固定模型名称枚举。值保持原样，显式空值覆盖旧值；显式 Meldwork Provider 不继承这些原生区域，其他 Agent 也不接收。未知凭据证据仍为未知，不因区域配置存在就判断登录成功。
+
+动态名称由 shell 内建能力枚举，再经格式和长度校验读取值；不导出无关环境，不新增 Node 依赖。zsh 使用参数名枚举，bash 使用 compgen，其他受支持 sh 在可用时借助系统 /bin/bash 枚举继承的导出名称。缺少这些 shell 能力时仍保留固定白名单和进程环境回退，不宣称所有 POSIX 环境均支持新增动态名称。
+
+验证：
+
+- 最终 readiness/Main 安全 91/91，覆盖缓存变更/移除/清空、Agent/Provider 隔离、非法名称，以及真实 sh/bash/zsh 采集。包含带命令替换形式的字面值和不相关敏感值，验证没有命令执行或额外值输出。
+- 中间实现曾使用当前 Electron 执行路径采集动态键；开发态通过，但检查 after-pack 发现正式包禁用 RunAsNode，因此在提交前替换为纯 shell。未放开打包安全 fuse。中间 Agent 全量 464/464 对应该被替换实现，不能作为最终采集实现的全量证明。
+- 最终 `/tmp/meldwork-105-region-electron.cjs` 使用真实 Electron 与隔离合成配置 `/tmp/meldwork-105-region-wJecRY` 验证动态键、空值、无 NODE_OPTIONS 透传及 Provider 隔离，采集约 15ms，退出 0；未调用外部模型。这是该次采集耗时，不是启动提速结论，也不是 AWS/Azure/Vertex 账户认证验收。
+
+上述两项尚未纳入新的全量桌面/前端构建与打包，包版本仍为 0.1.4，整体版本目标继续开发。