重构主控编排与运行时预热链路，统一工作区提示词/专家调度协议并补齐 wiki 记忆注入与写回闭环。

同时收敛启动与运维脚本默认行为（含 wiki worker）、更新 Admin 可观测性与相关测试，降低首轮时延并提高运行稳定性。

Made-with: Cursor

openclaw/qa/scenarios/models/anthropic-opus-api-key-smoke.md

@@ -0,0 +1,90 @@
+# Anthropic Opus API key smoke
+
+```yaml qa-scenario
+id: anthropic-opus-api-key-smoke
+title: Anthropic Opus API key smoke
+surface: model-provider
+coverage:
+  primary:
+    - models.provider-auth
+  secondary:
+    - models.anthropic
+objective: Verify the regular Anthropic Opus lane can complete a quick chat turn using API-key auth.
+successCriteria:
+  - A live-frontier run fails fast unless the selected primary provider is anthropic.
+  - The selected primary model is Anthropic Opus 4.6.
+  - The QA gateway worker has an Anthropic API key available through environment auth.
+  - The agent replies through the regular Anthropic provider.
+docsRefs:
+  - docs/concepts/model-providers.md
+  - docs/help/testing.md
+codeRefs:
+  - extensions/anthropic/register.runtime.ts
+  - extensions/qa-lab/src/gateway-child.ts
+  - extensions/qa-lab/src/suite.ts
+execution:
+  kind: flow
+  summary: Run with `pnpm openclaw qa suite --provider-mode live-frontier --model anthropic/claude-opus-4-6 --alt-model anthropic/claude-opus-4-6 --scenario anthropic-opus-api-key-smoke`.
+  config:
+    requiredProvider: anthropic
+    requiredModel: claude-opus-4-6
+    chatPrompt: "Anthropic Opus API key smoke. Reply exactly: ANTHROPIC-OPUS-API-KEY-OK"
+    chatExpected: ANTHROPIC-OPUS-API-KEY-OK
+```
+
+```yaml qa-flow
+steps:
+  - name: confirms regular Anthropic API-key lane
+    actions:
+      - set: selected
+        value:
+          expr: splitModelRef(env.primaryModel)
+      - assert:
+          expr: "env.providerMode !== 'live-frontier' || selected?.provider === config.requiredProvider"
+          message:
+            expr: "`expected live primary provider ${config.requiredProvider}, got ${env.primaryModel}`"
+      - assert:
+          expr: "env.providerMode !== 'live-frontier' || selected?.model === config.requiredModel"
+          message:
+            expr: "`expected live primary model ${config.requiredModel}, got ${env.primaryModel}`"
+      - assert:
+          expr: "env.providerMode !== 'live-frontier' || Boolean(env.gateway.runtimeEnv.ANTHROPIC_API_KEY?.trim())"
+          message: expected ANTHROPIC_API_KEY to be available for API-key QA mode
+    detailsExpr: "env.providerMode === 'live-frontier' ? `provider=${selected?.provider} model=${selected?.model} auth=env-api-key` : `mock-compatible provider=${selected?.provider}`"
+  - name: talks through regular Anthropic Opus
+    actions:
+      - if:
+          expr: "env.providerMode !== 'live-frontier'"
+          then:
+            - assert: "true"
+          else:
+            - call: reset
+            - set: selected
+              value:
+                expr: splitModelRef(env.primaryModel)
+            - call: runAgentPrompt
+              args:
+                - ref: env
+                - sessionKey: agent:qa:anthropic-opus-api-key
+                  message:
+                    expr: config.chatPrompt
+                  provider:
+                    expr: selected?.provider
+                  model:
+                    expr: selected?.model
+                  timeoutMs:
+                    expr: resolveQaLiveTurnTimeoutMs(env, 60000, env.primaryModel)
+            - call: waitForOutboundMessage
+              saveAs: chatOutbound
+              args:
+                - ref: state
+                - lambda:
+                    params: [candidate]
+                    expr: "candidate.conversation.id === 'qa-operator'"
+                - expr: resolveQaLiveTurnTimeoutMs(env, 30000, env.primaryModel)
+            - assert:
+                expr: "chatOutbound.text.includes(config.chatExpected)"
+                message:
+                  expr: "`chat marker missing: ${chatOutbound.text}`"
+    detailsExpr: "env.providerMode !== 'live-frontier' ? 'mock mode: skipped live Anthropic smoke' : chatOutbound.text"
+```

openclaw/qa/scenarios/models/anthropic-opus-setup-token-smoke.md

@@ -0,0 +1,95 @@
+# Anthropic Opus setup-token smoke
+
+```yaml qa-scenario
+id: anthropic-opus-setup-token-smoke
+title: Anthropic Opus setup-token smoke
+surface: model-provider
+coverage:
+  primary:
+    - models.provider-auth
+  secondary:
+    - models.anthropic
+objective: Verify the regular Anthropic Opus lane can complete a quick chat turn using setup-token auth.
+successCriteria:
+  - A live-frontier run fails fast unless the selected primary provider is anthropic.
+  - The selected primary model is Anthropic Opus 4.6.
+  - The QA gateway worker stages a token auth profile in the isolated agent store.
+  - The agent replies through the regular Anthropic provider.
+docsRefs:
+  - docs/concepts/model-providers.md
+  - docs/help/testing.md
+codeRefs:
+  - extensions/anthropic/register.runtime.ts
+  - extensions/qa-lab/src/gateway-child.ts
+  - extensions/qa-lab/src/suite.ts
+execution:
+  kind: flow
+  summary: Run with `OPENCLAW_LIVE_SETUP_TOKEN_VALUE=<setup-token> pnpm openclaw qa suite --provider-mode live-frontier --model anthropic/claude-opus-4-6 --alt-model anthropic/claude-opus-4-6 --scenario anthropic-opus-setup-token-smoke`.
+  config:
+    requiredProvider: anthropic
+    requiredModel: claude-opus-4-6
+    profileId: "anthropic:qa-setup-token"
+    chatPrompt: "Anthropic Opus setup-token smoke. Reply exactly: ANTHROPIC-OPUS-SETUP-TOKEN-OK"
+    chatExpected: ANTHROPIC-OPUS-SETUP-TOKEN-OK
+```
+
+```yaml qa-flow
+steps:
+  - name: confirms regular Anthropic setup-token lane
+    actions:
+      - set: selected
+        value:
+          expr: splitModelRef(env.primaryModel)
+      - assert:
+          expr: "env.providerMode !== 'live-frontier' || selected?.provider === config.requiredProvider"
+          message:
+            expr: "`expected live primary provider ${config.requiredProvider}, got ${env.primaryModel}`"
+      - assert:
+          expr: "env.providerMode !== 'live-frontier' || selected?.model === config.requiredModel"
+          message:
+            expr: "`expected live primary model ${config.requiredModel}, got ${env.primaryModel}`"
+      - assert:
+          expr: "env.providerMode !== 'live-frontier' || env.gateway.cfg.auth?.profiles?.[config.profileId]?.mode === 'token'"
+          message:
+            expr: "`expected token profile ${config.profileId} in QA config`"
+      - assert:
+          expr: "env.providerMode !== 'live-frontier' || !env.gateway.runtimeEnv.OPENCLAW_LIVE_SETUP_TOKEN_VALUE"
+          message: setup-token value should not be passed to the gateway child env
+    detailsExpr: "env.providerMode === 'live-frontier' ? `provider=${selected?.provider} model=${selected?.model} auth=setup-token profile=${config.profileId}` : `mock-compatible provider=${selected?.provider}`"
+  - name: talks through regular Anthropic Opus
+    actions:
+      - if:
+          expr: "env.providerMode !== 'live-frontier'"
+          then:
+            - assert: "true"
+          else:
+            - call: reset
+            - set: selected
+              value:
+                expr: splitModelRef(env.primaryModel)
+            - call: runAgentPrompt
+              args:
+                - ref: env
+                - sessionKey: agent:qa:anthropic-opus-setup-token
+                  message:
+                    expr: config.chatPrompt
+                  provider:
+                    expr: selected?.provider
+                  model:
+                    expr: selected?.model
+                  timeoutMs:
+                    expr: resolveQaLiveTurnTimeoutMs(env, 60000, env.primaryModel)
+            - call: waitForOutboundMessage
+              saveAs: chatOutbound
+              args:
+                - ref: state
+                - lambda:
+                    params: [candidate]
+                    expr: "candidate.conversation.id === 'qa-operator'"
+                - expr: resolveQaLiveTurnTimeoutMs(env, 30000, env.primaryModel)
+            - assert:
+                expr: "chatOutbound.text.includes(config.chatExpected)"
+                message:
+                  expr: "`chat marker missing: ${chatOutbound.text}`"
+    detailsExpr: "env.providerMode !== 'live-frontier' ? 'mock mode: skipped live Anthropic smoke' : chatOutbound.text"
+```

openclaw/qa/scenarios/models/claude-cli-provider-capabilities-subscription.md

@@ -0,0 +1,258 @@
+# Claude CLI provider capabilities subscription
+
+```yaml qa-scenario
+id: claude-cli-provider-capabilities-subscription
+title: Claude CLI provider capabilities subscription
+surface: model-provider
+coverage:
+  primary:
+    - models.provider-capabilities
+  secondary:
+    - models.claude-cli
+objective: Verify the Claude CLI model-provider lane can use native Claude subscription auth to talk, read an attached image, use bundled MCP tools, and apply workspace skills.
+successCriteria:
+  - A live-frontier run fails fast unless the selected primary provider is claude-cli.
+  - The Claude CLI backend does not preserve ANTHROPIC_API_KEY for this run, forcing native Claude subscription auth.
+  - The agent replies through the Claude CLI provider in a direct chat turn.
+  - The agent describes an attached image through the Claude CLI image path.
+  - The agent can reach memory via the bundled MCP/tool bridge.
+  - The agent sees and follows a workspace skill.
+docsRefs:
+  - docs/gateway/cli-backends.md
+  - docs/tools/skills.md
+  - docs/cli/mcp.md
+  - docs/tools/index.md
+codeRefs:
+  - extensions/anthropic/cli-backend.ts
+  - src/agents/cli-backends.ts
+  - src/mcp/plugin-tools-serve.ts
+  - extensions/qa-lab/src/suite.ts
+execution:
+  kind: flow
+  summary: Run with `pnpm openclaw qa suite --provider-mode live-frontier --cli-auth-mode subscription --model claude-cli/claude-sonnet-4-6 --alt-model claude-cli/claude-sonnet-4-6 --scenario claude-cli-provider-capabilities-subscription`.
+  config:
+    authMode: subscription
+    requiredProvider: claude-cli
+    chatPrompt: "Claude CLI provider marker check. Reply exactly: CLAUDE-CLI-CHAT-OK"
+    chatExpected: CLAUDE-CLI-CHAT-OK
+    imagePrompt: "Image understanding check: describe the top and bottom colors in the attached image in one short sentence."
+    imageColorGroups:
+      - [red, scarlet, crimson]
+      - [blue, azure, teal, cyan, aqua]
+    memoryFact: "Hidden Claude CLI MCP fact: the provider bridge codename is ORBIT-9."
+    memoryQuery: "provider bridge codename ORBIT-9"
+    memoryExpected: ORBIT-9
+    memoryPrompt: "Memory tools check: use the available memory search MCP/tool bridge to find the hidden provider bridge codename stored only in memory. Reply with the codename."
+    memoryPromptSnippet: "Memory tools check"
+    skillName: qa-claude-cli-skill
+    skillExpected: VISIBLE-SKILL-OK
+    skillBody: |-
+      ---
+      name: qa-claude-cli-skill
+      description: Claude CLI QA skill marker
+      ---
+      When the user asks for the Claude CLI skill marker exactly, or explicitly asks you to use qa-claude-cli-skill, reply with exactly: VISIBLE-SKILL-OK
+    skillPrompt: "Use qa-claude-cli-skill now. Reply exactly with the visible skill marker and nothing else."
+```
+
+```yaml qa-flow
+steps:
+  - name: confirms the selected live provider and Claude CLI auth mode
+    actions:
+      - set: selected
+        value:
+          expr: splitModelRef(env.primaryModel)
+      - set: preserveEnv
+        value:
+          expr: "String(env.gateway.runtimeEnv.OPENCLAW_LIVE_CLI_BACKEND_PRESERVE_ENV ?? '')"
+      - assert:
+          expr: "env.providerMode !== 'live-frontier' || selected?.provider === config.requiredProvider"
+          message:
+            expr: "`expected live primary provider ${config.requiredProvider}, got ${env.primaryModel}`"
+      - assert:
+          expr: "env.providerMode !== 'live-frontier' || env.gateway.runtimeEnv.OPENCLAW_LIVE_CLI_BACKEND_AUTH_MODE === config.authMode"
+          message:
+            expr: "`expected Claude CLI auth mode ${config.authMode}, got ${env.gateway.runtimeEnv.OPENCLAW_LIVE_CLI_BACKEND_AUTH_MODE ?? 'unset'}`"
+      - assert:
+          expr: "env.providerMode !== 'live-frontier' || !preserveEnv.includes('ANTHROPIC_API_KEY')"
+          message:
+            expr: "`expected ANTHROPIC_API_KEY not to be preserved for Claude CLI subscription QA mode, got ${preserveEnv}`"
+    detailsExpr: "env.providerMode === 'live-frontier' ? `provider=${selected?.provider} model=${selected?.model} auth=${env.gateway.runtimeEnv.OPENCLAW_LIVE_CLI_BACKEND_AUTH_MODE} preserve=${preserveEnv}` : `mock-compatible provider=${selected?.provider}`"
+  - name: talks through the selected provider
+    actions:
+      - call: reset
+      - set: selected
+        value:
+          expr: splitModelRef(env.primaryModel)
+      - call: runAgentPrompt
+        args:
+          - ref: env
+          - sessionKey: agent:qa:claude-cli-chat
+            message:
+              expr: config.chatPrompt
+            provider:
+              expr: selected?.provider
+            model:
+              expr: selected?.model
+            timeoutMs:
+              expr: resolveQaLiveTurnTimeoutMs(env, 45000, env.primaryModel)
+      - call: waitForOutboundMessage
+        saveAs: chatOutbound
+        args:
+          - ref: state
+          - lambda:
+              params: [candidate]
+              expr: "candidate.conversation.id === 'qa-operator'"
+          - expr: resolveQaLiveTurnTimeoutMs(env, 20000, env.primaryModel)
+      - assert:
+          expr: "chatOutbound.text.includes(config.chatExpected)"
+          message:
+            expr: "`chat marker missing: ${chatOutbound.text}`"
+    detailsExpr: chatOutbound.text
+  - name: describes an attached image through the selected provider
+    actions:
+      - call: reset
+      - set: selected
+        value:
+          expr: splitModelRef(env.primaryModel)
+      - call: runAgentPrompt
+        args:
+          - ref: env
+          - sessionKey: agent:qa:claude-cli-image
+            message:
+              expr: config.imagePrompt
+            provider:
+              expr: selected?.provider
+            model:
+              expr: selected?.model
+            attachments:
+              - mimeType: image/png
+                fileName: claude-cli-red-top-blue-bottom.png
+                content:
+                  expr: imageUnderstandingValidPngBase64
+            timeoutMs:
+              expr: resolveQaLiveTurnTimeoutMs(env, 60000, env.primaryModel)
+      - call: waitForOutboundMessage
+        saveAs: imageOutbound
+        args:
+          - ref: state
+          - lambda:
+              params: [candidate]
+              expr: "candidate.conversation.id === 'qa-operator'"
+          - expr: resolveQaLiveTurnTimeoutMs(env, 30000, env.primaryModel)
+      - assert:
+          expr: "config.imageColorGroups.every((group) => group.some((color) => normalizeLowercaseStringOrEmpty(imageOutbound.text).includes(color)))"
+          message:
+            expr: "`missing expected image colors: ${imageOutbound.text}`"
+      - assert:
+          expr: "!env.mock || (((await fetchJson(`${env.mock.baseUrl}/debug/requests`)).find((request) => String(request.prompt ?? '').includes('Image understanding check'))?.imageInputCount ?? 0) >= 1)"
+          message: expected image input to reach mock provider
+    detailsExpr: imageOutbound.text
+  - name: reaches memory through the MCP/tool bridge
+    actions:
+      - call: reset
+      - call: fs.writeFile
+        args:
+          - expr: "path.join(env.gateway.workspaceDir, 'MEMORY.md')"
+          - expr: "`${config.memoryFact}\\n`"
+          - utf8
+      - call: forceMemoryIndex
+        args:
+          - env:
+              ref: env
+            query:
+              expr: config.memoryQuery
+            expectedNeedle:
+              expr: config.memoryExpected
+      - call: createSession
+        saveAs: mcpSessionKey
+        args:
+          - ref: env
+          - Claude CLI MCP bridge
+      - call: readEffectiveTools
+        saveAs: mcpTools
+        args:
+          - ref: env
+          - ref: mcpSessionKey
+      - assert:
+          expr: "mcpTools.has('memory_search')"
+          message: memory_search missing from effective tools before MCP bridge check
+      - set: selected
+        value:
+          expr: splitModelRef(env.primaryModel)
+      - call: runAgentPrompt
+        args:
+          - ref: env
+          - sessionKey:
+              ref: mcpSessionKey
+            message:
+              expr: config.memoryPrompt
+            provider:
+              expr: selected?.provider
+            model:
+              expr: selected?.model
+            timeoutMs:
+              expr: resolveQaLiveTurnTimeoutMs(env, 90000, env.primaryModel)
+      - call: waitForOutboundMessage
+        saveAs: mcpOutbound
+        args:
+          - ref: state
+          - lambda:
+              params: [candidate]
+              expr: "candidate.conversation.id === 'qa-operator'"
+          - expr: resolveQaLiveTurnTimeoutMs(env, 45000, env.primaryModel)
+      - assert:
+          expr: "mcpOutbound.text.includes(config.memoryExpected)"
+          message:
+            expr: "`MCP memory result missing ${config.memoryExpected}: ${mcpOutbound.text}`"
+      - assert:
+          expr: "!env.mock || (await fetchJson(`${env.mock.baseUrl}/debug/requests`)).filter((request) => String(request.allInputText ?? '').includes(config.memoryPromptSnippet)).some((request) => request.plannedToolName === 'memory_search')"
+          message: expected mock model to plan memory_search for MCP bridge prompt
+    detailsExpr: mcpOutbound.text
+  - name: applies a workspace skill through the selected provider
+    actions:
+      - call: reset
+      - call: writeWorkspaceSkill
+        args:
+          - env:
+              ref: env
+            name:
+              expr: config.skillName
+            body:
+              expr: config.skillBody
+      - call: waitForCondition
+        args:
+          - lambda:
+              async: true
+              expr: "((await readSkillStatus(env)).find((skill) => skill.name === config.skillName)?.eligible ? true : undefined)"
+          - 15000
+          - 200
+      - set: selected
+        value:
+          expr: splitModelRef(env.primaryModel)
+      - call: runAgentPrompt
+        args:
+          - ref: env
+          - sessionKey: agent:qa:claude-cli-skill
+            message:
+              expr: config.skillPrompt
+            provider:
+              expr: selected?.provider
+            model:
+              expr: selected?.model
+            timeoutMs:
+              expr: resolveQaLiveTurnTimeoutMs(env, 60000, env.primaryModel)
+      - call: waitForOutboundMessage
+        saveAs: skillOutbound
+        args:
+          - ref: state
+          - lambda:
+              params: [candidate]
+              expr: "candidate.conversation.id === 'qa-operator'"
+          - expr: resolveQaLiveTurnTimeoutMs(env, 30000, env.primaryModel)
+      - assert:
+          expr: "skillOutbound.text.includes(config.skillExpected)"
+          message:
+            expr: "`skill marker missing: ${skillOutbound.text}`"
+    detailsExpr: skillOutbound.text
+```

openclaw/qa/scenarios/models/claude-cli-provider-capabilities.md

@@ -0,0 +1,258 @@
+# Claude CLI provider capabilities API key
+
+```yaml qa-scenario
+id: claude-cli-provider-capabilities
+title: Claude CLI provider capabilities API key
+surface: model-provider
+coverage:
+  primary:
+    - models.provider-capabilities
+  secondary:
+    - models.claude-cli
+objective: Verify the Claude CLI model-provider lane can use the Anthropic API key path to talk, read an attached image, use bundled MCP tools, and apply workspace skills.
+successCriteria:
+  - A live-frontier run fails fast unless the selected primary provider is claude-cli.
+  - The Claude CLI backend preserves ANTHROPIC_API_KEY for this run instead of using native subscription auth.
+  - The agent replies through the Claude CLI provider in a direct chat turn.
+  - The agent describes an attached image through the Claude CLI image path.
+  - The agent can reach memory via the bundled MCP/tool bridge.
+  - The agent sees and follows a workspace skill.
+docsRefs:
+  - docs/gateway/cli-backends.md
+  - docs/tools/skills.md
+  - docs/cli/mcp.md
+  - docs/tools/index.md
+codeRefs:
+  - extensions/anthropic/cli-backend.ts
+  - src/agents/cli-backends.ts
+  - src/mcp/plugin-tools-serve.ts
+  - extensions/qa-lab/src/suite.ts
+execution:
+  kind: flow
+  summary: Run with `pnpm openclaw qa suite --provider-mode live-frontier --cli-auth-mode api-key --model claude-cli/claude-sonnet-4-6 --alt-model claude-cli/claude-sonnet-4-6 --scenario claude-cli-provider-capabilities`.
+  config:
+    authMode: api-key
+    requiredProvider: claude-cli
+    chatPrompt: "Claude CLI provider marker check. Reply exactly: CLAUDE-CLI-CHAT-OK"
+    chatExpected: CLAUDE-CLI-CHAT-OK
+    imagePrompt: "Image understanding check: describe the top and bottom colors in the attached image in one short sentence."
+    imageColorGroups:
+      - [red, scarlet, crimson]
+      - [blue, azure, teal, cyan, aqua]
+    memoryFact: "Hidden Claude CLI MCP fact: the provider bridge codename is ORBIT-9."
+    memoryQuery: "provider bridge codename ORBIT-9"
+    memoryExpected: ORBIT-9
+    memoryPrompt: "Memory tools check: use the available memory search MCP/tool bridge to find the hidden provider bridge codename stored only in memory. Reply with the codename."
+    memoryPromptSnippet: "Memory tools check"
+    skillName: qa-claude-cli-skill
+    skillExpected: VISIBLE-SKILL-OK
+    skillBody: |-
+      ---
+      name: qa-claude-cli-skill
+      description: Claude CLI QA skill marker
+      ---
+      When the user asks for the Claude CLI skill marker exactly, or explicitly asks you to use qa-claude-cli-skill, reply with exactly: VISIBLE-SKILL-OK
+    skillPrompt: "Use qa-claude-cli-skill now. Reply exactly with the visible skill marker and nothing else."
+```
+
+```yaml qa-flow
+steps:
+  - name: confirms the selected live provider and Claude CLI auth mode
+    actions:
+      - set: selected
+        value:
+          expr: splitModelRef(env.primaryModel)
+      - set: preserveEnv
+        value:
+          expr: "String(env.gateway.runtimeEnv.OPENCLAW_LIVE_CLI_BACKEND_PRESERVE_ENV ?? '')"
+      - assert:
+          expr: "env.providerMode !== 'live-frontier' || selected?.provider === config.requiredProvider"
+          message:
+            expr: "`expected live primary provider ${config.requiredProvider}, got ${env.primaryModel}`"
+      - assert:
+          expr: "env.providerMode !== 'live-frontier' || env.gateway.runtimeEnv.OPENCLAW_LIVE_CLI_BACKEND_AUTH_MODE === config.authMode"
+          message:
+            expr: "`expected Claude CLI auth mode ${config.authMode}, got ${env.gateway.runtimeEnv.OPENCLAW_LIVE_CLI_BACKEND_AUTH_MODE ?? 'unset'}`"
+      - assert:
+          expr: "env.providerMode !== 'live-frontier' || preserveEnv.includes('ANTHROPIC_API_KEY')"
+          message:
+            expr: "`expected ANTHROPIC_API_KEY to be preserved for Claude CLI API-key QA mode, got ${preserveEnv}`"
+    detailsExpr: "env.providerMode === 'live-frontier' ? `provider=${selected?.provider} model=${selected?.model} auth=${env.gateway.runtimeEnv.OPENCLAW_LIVE_CLI_BACKEND_AUTH_MODE} preserve=${preserveEnv}` : `mock-compatible provider=${selected?.provider}`"
+  - name: talks through the selected provider
+    actions:
+      - call: reset
+      - set: selected
+        value:
+          expr: splitModelRef(env.primaryModel)
+      - call: runAgentPrompt
+        args:
+          - ref: env
+          - sessionKey: agent:qa:claude-cli-chat
+            message:
+              expr: config.chatPrompt
+            provider:
+              expr: selected?.provider
+            model:
+              expr: selected?.model
+            timeoutMs:
+              expr: resolveQaLiveTurnTimeoutMs(env, 45000, env.primaryModel)
+      - call: waitForOutboundMessage
+        saveAs: chatOutbound
+        args:
+          - ref: state
+          - lambda:
+              params: [candidate]
+              expr: "candidate.conversation.id === 'qa-operator'"
+          - expr: resolveQaLiveTurnTimeoutMs(env, 20000, env.primaryModel)
+      - assert:
+          expr: "chatOutbound.text.includes(config.chatExpected)"
+          message:
+            expr: "`chat marker missing: ${chatOutbound.text}`"
+    detailsExpr: chatOutbound.text
+  - name: describes an attached image through the selected provider
+    actions:
+      - call: reset
+      - set: selected
+        value:
+          expr: splitModelRef(env.primaryModel)
+      - call: runAgentPrompt
+        args:
+          - ref: env
+          - sessionKey: agent:qa:claude-cli-image
+            message:
+              expr: config.imagePrompt
+            provider:
+              expr: selected?.provider
+            model:
+              expr: selected?.model
+            attachments:
+              - mimeType: image/png
+                fileName: claude-cli-red-top-blue-bottom.png
+                content:
+                  expr: imageUnderstandingValidPngBase64
+            timeoutMs:
+              expr: resolveQaLiveTurnTimeoutMs(env, 60000, env.primaryModel)
+      - call: waitForOutboundMessage
+        saveAs: imageOutbound
+        args:
+          - ref: state
+          - lambda:
+              params: [candidate]
+              expr: "candidate.conversation.id === 'qa-operator'"
+          - expr: resolveQaLiveTurnTimeoutMs(env, 30000, env.primaryModel)
+      - assert:
+          expr: "config.imageColorGroups.every((group) => group.some((color) => normalizeLowercaseStringOrEmpty(imageOutbound.text).includes(color)))"
+          message:
+            expr: "`missing expected image colors: ${imageOutbound.text}`"
+      - assert:
+          expr: "!env.mock || (((await fetchJson(`${env.mock.baseUrl}/debug/requests`)).find((request) => String(request.prompt ?? '').includes('Image understanding check'))?.imageInputCount ?? 0) >= 1)"
+          message: expected image input to reach mock provider
+    detailsExpr: imageOutbound.text
+  - name: reaches memory through the MCP/tool bridge
+    actions:
+      - call: reset
+      - call: fs.writeFile
+        args:
+          - expr: "path.join(env.gateway.workspaceDir, 'MEMORY.md')"
+          - expr: "`${config.memoryFact}\\n`"
+          - utf8
+      - call: forceMemoryIndex
+        args:
+          - env:
+              ref: env
+            query:
+              expr: config.memoryQuery
+            expectedNeedle:
+              expr: config.memoryExpected
+      - call: createSession
+        saveAs: mcpSessionKey
+        args:
+          - ref: env
+          - Claude CLI MCP bridge
+      - call: readEffectiveTools
+        saveAs: mcpTools
+        args:
+          - ref: env
+          - ref: mcpSessionKey
+      - assert:
+          expr: "mcpTools.has('memory_search')"
+          message: memory_search missing from effective tools before MCP bridge check
+      - set: selected
+        value:
+          expr: splitModelRef(env.primaryModel)
+      - call: runAgentPrompt
+        args:
+          - ref: env
+          - sessionKey:
+              ref: mcpSessionKey
+            message:
+              expr: config.memoryPrompt
+            provider:
+              expr: selected?.provider
+            model:
+              expr: selected?.model
+            timeoutMs:
+              expr: resolveQaLiveTurnTimeoutMs(env, 90000, env.primaryModel)
+      - call: waitForOutboundMessage
+        saveAs: mcpOutbound
+        args:
+          - ref: state
+          - lambda:
+              params: [candidate]
+              expr: "candidate.conversation.id === 'qa-operator'"
+          - expr: resolveQaLiveTurnTimeoutMs(env, 45000, env.primaryModel)
+      - assert:
+          expr: "mcpOutbound.text.includes(config.memoryExpected)"
+          message:
+            expr: "`MCP memory result missing ${config.memoryExpected}: ${mcpOutbound.text}`"
+      - assert:
+          expr: "!env.mock || (await fetchJson(`${env.mock.baseUrl}/debug/requests`)).filter((request) => String(request.allInputText ?? '').includes(config.memoryPromptSnippet)).some((request) => request.plannedToolName === 'memory_search')"
+          message: expected mock model to plan memory_search for MCP bridge prompt
+    detailsExpr: mcpOutbound.text
+  - name: applies a workspace skill through the selected provider
+    actions:
+      - call: reset
+      - call: writeWorkspaceSkill
+        args:
+          - env:
+              ref: env
+            name:
+              expr: config.skillName
+            body:
+              expr: config.skillBody
+      - call: waitForCondition
+        args:
+          - lambda:
+              async: true
+              expr: "((await readSkillStatus(env)).find((skill) => skill.name === config.skillName)?.eligible ? true : undefined)"
+          - 15000
+          - 200
+      - set: selected
+        value:
+          expr: splitModelRef(env.primaryModel)
+      - call: runAgentPrompt
+        args:
+          - ref: env
+          - sessionKey: agent:qa:claude-cli-skill
+            message:
+              expr: config.skillPrompt
+            provider:
+              expr: selected?.provider
+            model:
+              expr: selected?.model
+            timeoutMs:
+              expr: resolveQaLiveTurnTimeoutMs(env, 60000, env.primaryModel)
+      - call: waitForOutboundMessage
+        saveAs: skillOutbound
+        args:
+          - ref: state
+          - lambda:
+              params: [candidate]
+              expr: "candidate.conversation.id === 'qa-operator'"
+          - expr: resolveQaLiveTurnTimeoutMs(env, 30000, env.primaryModel)
+      - assert:
+          expr: "skillOutbound.text.includes(config.skillExpected)"
+          message:
+            expr: "`skill marker missing: ${skillOutbound.text}`"
+    detailsExpr: skillOutbound.text
+```

openclaw/qa/scenarios/models/codex-harness-no-meta-leak.md

@@ -0,0 +1,147 @@
+# Codex harness no meta leak
+
+```yaml qa-scenario
+id: codex-harness-no-meta-leak
+title: Codex harness no meta leak
+surface: dm
+coverage:
+  primary:
+    - models.codex-cli
+  secondary:
+    - runtime.no-meta-leak
+objective: Verify the Codex app-server harness keeps coordination/meta chatter out of the visible reply.
+successCriteria:
+  - The scenario forces the Codex embedded harness and disables PI fallback.
+  - The final visible reply includes the requested confirmation token.
+  - The visible reply does not include internal coordination or progress chatter.
+docsRefs:
+  - docs/plugins/sdk-agent-harness.md
+  - docs/concepts/qa-e2e-automation.md
+  - docs/help/testing.md
+codeRefs:
+  - extensions/codex/src/app-server/event-projector.ts
+  - extensions/qa-lab/src/reply-failure.ts
+  - extensions/qa-lab/src/suite.ts
+execution:
+  kind: flow
+  summary: Run with `pnpm openclaw qa suite --provider-mode live-frontier --model codex/gpt-5.4 --alt-model codex/gpt-5.4 --scenario codex-harness-no-meta-leak`.
+  config:
+    requiredProvider: codex
+    requiredModel: gpt-5.4
+    harnessRuntime: codex
+    harnessFallback: none
+    expectedReply: QA_LEAK_OK
+    prompt: |-
+      Think through your answer privately, but do not expose any internal planning, thread-context checks, or progress narration.
+
+      Reply with exactly one line:
+      QA_LEAK_OK
+    forbiddenReplySubstrings:
+      - checking thread context
+      - thread context thin
+      - post a tight progress reply here
+      - posting a coordination nudge
+      - posted a short coordination reply
+      - not inventing status
+```
+
+```yaml qa-flow
+steps:
+  - name: confirms GPT-5.4 Codex harness target
+    actions:
+      - set: selected
+        value:
+          expr: splitModelRef(env.primaryModel)
+      - assert:
+          expr: "env.providerMode !== 'live-frontier' || selected?.provider === config.requiredProvider"
+          message:
+            expr: "`expected live primary provider ${config.requiredProvider}, got ${env.primaryModel}`"
+      - assert:
+          expr: "env.providerMode !== 'live-frontier' || selected?.model === config.requiredModel"
+          message:
+            expr: "`expected live primary model ${config.requiredModel}, got ${env.primaryModel}`"
+      - if:
+          expr: "env.providerMode !== 'live-frontier'"
+          then:
+            - assert: "true"
+          else:
+            - call: patchConfig
+              saveAs: patchResult
+              args:
+                - env:
+                    ref: env
+                  patch:
+                    agents:
+                      defaults:
+                        embeddedHarness:
+                          runtime:
+                            expr: config.harnessRuntime
+                          fallback:
+                            expr: config.harnessFallback
+            - call: waitForGatewayHealthy
+              args:
+                - ref: env
+                - 60000
+            - call: waitForQaChannelReady
+              args:
+                - ref: env
+                - 60000
+            - call: readConfigSnapshot
+              saveAs: snapshot
+              args:
+                - ref: env
+            - assert:
+                expr: "snapshot.config.agents?.defaults?.embeddedHarness?.runtime === config.harnessRuntime"
+                message:
+                  expr: "`expected embeddedHarness.runtime=${config.harnessRuntime}, got ${JSON.stringify(snapshot.config.agents?.defaults?.embeddedHarness)}`"
+            - assert:
+                expr: "snapshot.config.agents?.defaults?.embeddedHarness?.fallback === config.harnessFallback"
+                message:
+                  expr: "`expected embeddedHarness.fallback=${config.harnessFallback}, got ${JSON.stringify(snapshot.config.agents?.defaults?.embeddedHarness)}`"
+    detailsExpr: "env.providerMode === 'live-frontier' ? `provider=${selected?.provider} model=${selected?.model} runtime=${snapshot.config.agents?.defaults?.embeddedHarness?.runtime} fallback=${snapshot.config.agents?.defaults?.embeddedHarness?.fallback}` : `mock mode: parsed ${scenario.id}`"
+  - name: keeps codex coordination chatter out of the visible reply
+    actions:
+      - if:
+          expr: "env.providerMode !== 'live-frontier'"
+          then:
+            - assert: "true"
+          else:
+            - call: reset
+            - call: runAgentPrompt
+              args:
+                - ref: env
+                - sessionKey: agent:qa:codex-meta-leak
+                  message:
+                    expr: config.prompt
+                  provider:
+                    expr: selected?.provider
+                  model:
+                    expr: selected?.model
+                  timeoutMs:
+                    expr: resolveQaLiveTurnTimeoutMs(env, 180000, env.primaryModel)
+            - call: waitForOutboundMessage
+              saveAs: outbound
+              args:
+                - ref: state
+                - lambda:
+                    params: [candidate]
+                    expr: "candidate.conversation.id === 'qa-operator' && candidate.text.includes(config.expectedReply)"
+                - expr: resolveQaLiveTurnTimeoutMs(env, 60000, env.primaryModel)
+            - set: outboundLower
+              value:
+                expr: normalizeLowercaseStringOrEmpty(outbound.text)
+            - assert:
+                expr: "outbound.text.trim() === config.expectedReply"
+                message:
+                  expr: "`expected exact visible reply ${config.expectedReply}, got ${outbound.text}`"
+            - forEach:
+                items:
+                  expr: "config.forbiddenReplySubstrings ?? []"
+                item: forbidden
+                actions:
+                  - assert:
+                      expr: "!outboundLower.includes(normalizeLowercaseStringOrEmpty(forbidden))"
+                      message:
+                        expr: "`visible reply leaked internal meta text (${forbidden}): ${outbound.text}`"
+    detailsExpr: "env.providerMode !== 'live-frontier' ? 'mock mode: skipped live codex leak check' : outbound.text"
+```

openclaw/qa/scenarios/models/model-switch-follow-up.md

@@ -0,0 +1,79 @@
+# Model switch follow-up
+
+```yaml qa-scenario
+id: model-switch-follow-up
+title: Model switch follow-up
+surface: models
+coverage:
+  primary:
+    - models.switching
+  secondary:
+    - runtime.session-continuity
+objective: Verify the agent can switch to a different configured model and continue coherently.
+successCriteria:
+  - Agent reflects the model switch request.
+  - Follow-up answer remains coherent with prior context.
+  - Final report notes whether the switch actually happened.
+docsRefs:
+  - docs/help/testing.md
+  - docs/web/dashboard.md
+codeRefs:
+  - extensions/qa-lab/src/report.ts
+execution:
+  kind: flow
+  summary: Verify the agent can switch to a different configured model and continue coherently.
+  config:
+    initialPrompt: "Say hello from the default configured model."
+    followupPrompt: "Continue the exchange after switching models and note the handoff."
+```
+
+```yaml qa-flow
+steps:
+  - name: runs on the default configured model
+    actions:
+      - call: reset
+      - call: runAgentPrompt
+        args:
+          - ref: env
+          - sessionKey: agent:qa:model-switch
+            message:
+              expr: config.initialPrompt
+            timeoutMs:
+              expr: liveTurnTimeoutMs(env, 30000)
+      - call: waitForOutboundMessage
+        saveAs: outbound
+        args:
+          - ref: state
+          - lambda:
+              params: [candidate]
+              expr: "candidate.conversation.id === 'qa-operator'"
+    detailsExpr: "env.mock ? String((await fetchJson(`${env.mock.baseUrl}/debug/last-request`))?.body?.model ?? '') : outbound.text"
+  - name: switches to the alternate model and continues
+    actions:
+      - set: alternate
+        value:
+          expr: splitModelRef(env.alternateModel)
+      - call: runAgentPrompt
+        args:
+          - ref: env
+          - sessionKey: agent:qa:model-switch
+            message:
+              expr: config.followupPrompt
+            provider:
+              expr: alternate?.provider
+            model:
+              expr: alternate?.model
+            timeoutMs:
+              expr: resolveQaLiveTurnTimeoutMs(env, 30000, env.alternateModel)
+      - call: waitForCondition
+        saveAs: outbound
+        args:
+          - lambda:
+              expr: "state.getSnapshot().messages.filter((candidate) => candidate.direction === 'outbound' && candidate.conversation.id === 'qa-operator' && (() => { const lower = normalizeLowercaseStringOrEmpty(candidate.text); return lower.includes('switch') || lower.includes('handoff'); })()).at(-1)"
+          - expr: resolveQaLiveTurnTimeoutMs(env, 20000, env.alternateModel)
+      - assert:
+          expr: "!env.mock || ((await fetchJson(`${env.mock.baseUrl}/debug/last-request`))?.body?.model === 'gpt-5.4-alt')"
+          message:
+            expr: "`expected gpt-5.4-alt, got ${String((await fetchJson(`${env.mock.baseUrl}/debug/last-request`))?.body?.model ?? '')}`"
+    detailsExpr: outbound.text
+```

openclaw/qa/scenarios/models/model-switch-tool-continuity.md

@@ -0,0 +1,95 @@
+# Model switch with tool continuity
+
+```yaml qa-scenario
+id: model-switch-tool-continuity
+title: Model switch with tool continuity
+surface: models
+coverage:
+  primary:
+    - models.switching
+  secondary:
+    - runtime.tool-continuity
+objective: Verify switching models preserves session context and tool use instead of dropping into plain-text only behavior.
+successCriteria:
+  - Alternate model is actually requested.
+  - A tool call still happens after the model switch.
+  - Final answer acknowledges the handoff and uses the tool-derived evidence.
+docsRefs:
+  - docs/help/testing.md
+  - docs/concepts/model-failover.md
+codeRefs:
+  - extensions/qa-lab/src/suite.ts
+  - extensions/qa-lab/src/mock-openai-server.ts
+execution:
+  kind: flow
+  summary: Verify switching models preserves session context and tool use instead of dropping into plain-text only behavior.
+  config:
+    initialPrompt: "Read repo/qa/scenarios/index.md and summarize the QA scenario pack mission in one clause before any model switch."
+    followupPrompt: "The harness has already requested the alternate model for this turn. Do not call session_status or change models yourself. Tool continuity check: use the read tool to reread repo/qa/scenarios/index.md, then mention the model handoff and QA mission in one short sentence."
+    promptSnippet: "Tool continuity check"
+```
+
+```yaml qa-flow
+steps:
+  - name: keeps using tools after switching models
+    actions:
+      - call: waitForGatewayHealthy
+        args:
+          - ref: env
+          - 60000
+      - call: reset
+      - call: runAgentPrompt
+        args:
+          - ref: env
+          - sessionKey: agent:qa:model-switch-tools
+            message:
+              expr: config.initialPrompt
+            timeoutMs:
+              expr: liveTurnTimeoutMs(env, 30000)
+      - set: alternate
+        value:
+          expr: splitModelRef(env.alternateModel)
+      - set: beforeSwitchCursor
+        value:
+          expr: state.getSnapshot().messages.length
+      - call: runAgentPrompt
+        args:
+          - ref: env
+          - sessionKey: agent:qa:model-switch-tools
+            message:
+              expr: config.followupPrompt
+            provider:
+              expr: alternate?.provider
+            model:
+              expr: alternate?.model
+            timeoutMs:
+              expr: resolveQaLiveTurnTimeoutMs(env, 30000, env.alternateModel)
+      - call: waitForCondition
+        saveAs: outbound
+        args:
+          - lambda:
+              expr: "state.getSnapshot().messages.slice(beforeSwitchCursor).filter((candidate) => candidate.direction === 'outbound' && candidate.conversation.id === 'qa-operator' && hasModelSwitchContinuityEvidence(candidate.text)).at(-1)"
+          - expr: resolveQaLiveTurnTimeoutMs(env, 20000, env.alternateModel)
+      - assert:
+          expr: hasModelSwitchContinuityEvidence(outbound.text)
+          message:
+            expr: "`switch reply missed kickoff continuity: ${outbound.text}`"
+      - if:
+          expr: "Boolean(env.mock)"
+          then:
+            - set: switchDebugRequests
+              value:
+                expr: "await fetchJson(`${env.mock.baseUrl}/debug/requests`)"
+            - set: switchRequest
+              value:
+                expr: "switchDebugRequests.find((request) => String(request.allInputText ?? '').includes(config.promptSnippet))"
+            - assert:
+                expr: "switchRequest?.plannedToolName === 'read'"
+                message:
+                  expr: "`expected read after switch, got ${String(switchRequest?.plannedToolName ?? '')}`"
+            - assert:
+                expr: "String(switchRequest?.model ?? '') === String(alternate?.model ?? '')"
+                message:
+                  expr: "`expected alternate model, got ${String(switchRequest?.model ?? '')}`"
+    detailsExpr: outbound.text
+```