feat: Lark bot 不支持自然语言问答（无通用对话 turn 路由层）

[YueZh127]

id: cluster-triage-lark-channel-turn-routing
severity: medium
requires_design: true

来源

本 issue 由 maintainer 手动开，triage codex 深度调研补 evidence。

核心问题

当前代码已经不是“Lark 自然语言完全不走 LLM”：普通文本会进入 LlmReplyRequested，unknown slash 也有 Ornn skill discovery fallback。真正需要进入 refactor loop 的问题是：Lark/Nyx relay 的 turn routing 仍由 ChannelConversationTurnRunner 的 ordered branches、NyxRelayAgentBuilderFlow 白名单 fall-through、ConversationReplyGenerator 工具注入共同隐式表达，缺少一个 typed route decision / policy contract 来统一表达“本地 slash、agent-builder、workflow resume、Ornn skill shortcut、普通 LLM fallback”的边界。

这会让后续新增自然语言、tool、approval、skill discovery 策略继续堆进 runner 中间层，容易破坏分层、命令骨架内聚和 command/query 边界。Fix boundary 必须限制在 aevatar 本仓库内，不能要求 NyxID/chrono-* 新增能力或改 scope。

Evidence

1. Turn runner 同时承担多种业务路由与最终 LLM fallback

agents/Aevatar.GAgents.NyxidChat/ChannelConversationTurnRunner.cs:111

public async Task<ConversationTurnResult> RunInboundAsync(
    ChatActivity activity,
    ConversationTurnRuntimeContext runtimeContext,
    CancellationToken ct)
{
    var registration = await ResolveRegistrationAsync(activity, ct);
    ...
    if (await TryHandleWorkflowResumeAsync(inbound, ct) is { } workflowResumeResult)
        return workflowResumeResult;

    if (await TryHandleSlashCommandAsync(activity, inbound, registration, runtimeContext, ct) is { } slashResult)
        return slashResult;
    ...
    if (await TryHandleAgentBuilderAsync(activity, inboundEvent, registration, runtimeContext, typingReactionTask, ct) is { } agentBuilderResult)
        return agentBuilderResult;
    ...
    return ConversationTurnResult.LlmReplyRequested(
        await BuildLlmReplyRequestAsync(...));
}

违反点：routing policy 由方法内顺序和早返回表达；新增策略时只能继续扩张 runner，而不是组合 typed route decision。

2. Agent-builder route 是独立白名单，靠 false fall-through 表达“交给后续 LLM/skill”

agents/Aevatar.GAgents.Authoring.Lark/NyxRelayAgentBuilderFlow.cs:19

public static bool TryResolve(ChannelInboundEvent evt, out AgentBuilderFlowDecision? decision)
{
    ...
    var command = tokens[0];
    if (!IsKnownCommand(command))
        return false;

    if (!IsPrivateChat(evt.ChatType))
    {
        decision = AgentBuilderFlowDecision.DirectReply(BuildPrivateChatRestrictionReply(command));
        return true;
    }

    return TryResolveKnownCommand(command, tokens, out decision);
}

违反点：false 同时可能表示“不是 slash”“unknown slash”“应交给 Ornn skill/LLM”，没有强类型决策结果区分，调用方只能通过外部顺序隐式解释。

3. Runner 内部直接执行 AgentBuilderTool 并选择 renderer，route decision 与执行耦合

agents/Aevatar.GAgents.NyxidChat/ChannelConversationTurnRunner.cs:970

var relayDecisionMatched = NyxRelayAgentBuilderFlow.TryResolve(inboundEvent, out decision);
...
if (decision.RequiresToolExecution)
{
    using (AgentToolContextScope.Push(...))
    {
        var tool = ActivatorUtilities.CreateInstance<AgentBuilderTool>(_toolServiceProvider);
        var toolResult = await tool.ExecuteAsync(decision.ToolArgumentsJson!, ct);
        replyContent = relayDecisionMatched
            ? NyxRelayAgentBuilderFlow.FormatToolResult(decision, toolResult)
            : AgentBuilderCardFlow.FormatToolResult(decision, toolResult);
    }
}

违反点：中间层 runner 同时解析 route、构造 tool context、执行 tool、选择渲染器；它不再只是把 typed command/event 投入 actor 主链或把 query 交给 readmodel。

4. Canon 文档确认当前事实是多分支 ordered flow，而不是“没有 LLM 路径”

docs/canon/daily-command-pipeline.md:31

`/daily` slash shortcut 不走本地 agent-builder 创建逻辑；`ChannelConversationTurnRunner` 把它改写为一次 LLM turn，要求模型先调用 `use_skill(skill="chrono-ai-daily", args="<slash args>")`。

docs/canon/daily-command-pipeline.md:165

`/daily` 与其他未知 slash（如 `/goal`）是 Ornn skill shortcut：本路由放行给 LLM reply path，不走 `agent_builder`

违反点：文档事实说明普通/unknown 路径已进入 LLM/skill fallback；所以本 issue 的 actionable 方向应是收敛 route/policy 设计，而不是重复添加一条已存在的 LLM 路径。

违反条款

• 严格分层：Domain / Application / Infrastructure / Host；API 仅做宿主与组合，不承载业务编排。

• 读写分离：Command -> Event，Query -> ReadModel；异步完成通过事件通知，不在会话内拼装流程。

• 依赖反转：上层依赖抽象，禁止跨层反向依赖和对具体实现的直接耦合。

• 命令骨架内聚：标准生命周期 Normalize -> Resolve Target -> Build Context -> Build Envelope -> Dispatch -> Receipt -> Observe；业务模块只负责目标解析与载荷/结果映射。

• AI 对话主链必须流式化：实时会话入口必须使用 ChatStreamAsync；ChatAsync 仅可用于明确的非交互式离线场景。

• 本仓库的功能实现禁止依赖外部仓库的新增 / 修改。NyxID、chrono-storage、chrono-ornn 等都是独立产品，aevatar 的需求不是它们的功能路线图。

新原则

Channel inbound 应先归一化为 typed turn route decision，再进入对应 command/event/LLM execution path；false fall-through 不应承载业务语义。

Agent-builder、workflow resume、slash command、Ornn skill shortcut、普通 LLM reply 应共享同一 decision contract，decision 只描述目标语义，不直接执行 tool 或拼装最终流程。

自然语言问答接入必须继续走 ChatStreamAsync 主链；不得新增 ChatAsync 或 webhook 同步 RPC 旁路。

Fix boundary

scope_paths:

• agents/Aevatar.GAgents.NyxidChat/ChannelConversationTurnRunner.cs
• agents/Aevatar.GAgents.NyxidChat/ConversationReplyGenerator.cs
• agents/Aevatar.GAgents.NyxidChat/NyxIdChatEndpoints.Relay.cs
• agents/Aevatar.GAgents.NyxidChat/ServiceCollectionExtensions.cs
• agents/Aevatar.GAgents.Authoring.Lark/NyxRelayAgentBuilderFlow.cs
• agents/Aevatar.GAgents.Authoring.Lark/AgentBuilderCardFlow.cs
• agents/Aevatar.GAgents.Authoring.Lark/AgentBuilderTool.cs
• agents/Aevatar.GAgents.Channel.Runtime/Conversation/IConversationTurnRunner.cs
• agents/Aevatar.GAgents.Channel.Runtime/Conversation/ConversationGAgent.cs
• src/Aevatar.ChatRouting.Core/ChatRouteResolver.cs
• src/Aevatar.ChatRouting.Core/*
• src/Aevatar.AI.ToolProviders.NyxId/Tools/NyxIdProxyTool.cs
• src/Aevatar.AI.ToolProviders.NyxId/Tools/NyxIdCodeExecuteTool.cs
• src/Aevatar.AI.ToolProviders.Skills/UseSkillTool.cs
• src/Aevatar.AI.ToolProviders.Ornn/*
• test/Aevatar.GAgents.ChannelRuntime.Tests/ChannelConversationTurnRunnerTests.cs
• test/Aevatar.GAgents.ChannelRuntime.Tests/NyxRelayAgentBuilderFlowTests.cs
• test/Aevatar.ChatRouting.Core.Tests/*
• docs/canon/daily-command-pipeline.md

Out of scope:

• 不要求 NyxID 增加 im:message:readonly 或任何新 endpoint/scope。
• 不要求 chrono-ornn/chrono-storage 修改 schema 或协议。
• 不把 Lark service 权限配置问题作为本 cluster 的完成条件；若现有 surface 不足，solver 应在 aevatar 内降级或不做相关能力。

Decision questions

1. typed turn route decision 应放在 Aevatar.GAgents.Channel.Runtime 还是 Aevatar.GAgents.NyxidChat？如果放 Runtime，必须避免引入 AI/NyxID/Studio 反向依赖。
2. NyxRelayAgentBuilderFlow.TryResolve 是否应返回 Matched/NotMatched/PassToLlm/Reject/DirectReply/DispatchCommand 这样的强类型结果，替代 bool fall-through？
3. unknown slash 的 Ornn skill discovery 是 route policy 的一种 action，还是 LLM prompt builder 的内部策略？
4. AgentBuilderTool 执行应继续由 turn runner 调用，还是拆成 command handler / port，使 runner 只发 typed command 并返回 accepted/direct reply？
5. 是否复用 src/Aevatar.ChatRouting.Core/ChatRouteResolver.cs 的 policy/readmodel 语义，还是只抽取 channel-turn-local decision contract，避免扩大 scope？

original_authors

• agents/Aevatar.GAgents.NyxidChat/ChannelConversationTurnRunner.cs: eanzhao, maintainer-redacted
• agents/Aevatar.GAgents.Authoring.Lark/NyxRelayAgentBuilderFlow.cs: eanzhao, maintainer-redacted
• agents/Aevatar.GAgents.NyxidChat/ConversationReplyGenerator.cs: eanzhao, maintainer-redacted
• agents/Aevatar.GAgents.NyxidChat/NyxIdChatEndpoints.Relay.cs: eanzhao, maintainer-redacted, louis.li

cc

原始作者见上；本自动 triage 遵守 shared hard rules，不写 private maintainer name 或 @-mention。

⟦AI:AUTO-LOOP⟧

[loning]

This was generated by AI during triage.

🤖 Triage codex — accept: cluster-triage-lark-channel-turn-routing

TL;DR:

• 已 accept，但不是按原文“完全没有 LLM 路径”接入；当前代码已有普通消息 LLM fallback。
• actionable cluster 是：Lark/Nyx relay 的 turn routing 仍散落在 runner ordered branches、agent-builder 白名单 fall-through、LLM/Ornn skill fallback 之间。
• 调研找到 4 条主要 evidence，命中分层、命令骨架内聚、AI 流式主链、外部仓库无改动权条款。
• Fix boundary 限定在 aevatar 本仓库内；不要求 NyxID 增加 scope 或 endpoint。

Investigation 摘要

• Title tokens: Lark, bot, 自然语言, 问答, 通用对话, turn, 路由层。
• 强 signal 文件:
  • agents/Aevatar.GAgents.NyxidChat/ChannelConversationTurnRunner.cs
  • agents/Aevatar.GAgents.Authoring.Lark/NyxRelayAgentBuilderFlow.cs
  • agents/Aevatar.GAgents.NyxidChat/ConversationReplyGenerator.cs
  • agents/Aevatar.GAgents.NyxidChat/NyxIdChatEndpoints.Relay.cs
  • docs/canon/daily-command-pipeline.md
• 当前事实:
  • ChannelConversationTurnRunner.RunInboundAsync 对普通消息返回 LlmReplyRequested。
  • NyxRelayAgentBuilderFlow.TryResolve 只匹配 agent-management 白名单；unknown slash 通过 fall-through 交给后续 LLM/Ornn skill path。
  • docs/canon/daily-command-pipeline.md 已记录 /daily 和 unknown slash 的 LLM/skill fallback。
• 设计风险:
  • false fall-through 没有表达“不是 slash / unknown slash / pass to LLM / pass to skill discovery”的业务差异。
  • runner 同时做 route、tool context、tool execution、renderer selection，后续自然语言/tool/approval 策略会继续堆叠。
• 查重:
  • #1102 workflow、llm、nyx耦合 范围更大；feat: Lark bot 不支持自然语言问答（无通用对话 turn 路由层） #489 聚焦 Lark channel turn routing，不按 duplicate reject。

Artifacts:

• .refactor-loop/runs/triage-issue-489-investigation.md
• .refactor-loop/runs/triage-issue-489.md

⟦AI:AUTO-LOOP⟧

[loning]

🤖 delete-solver：不新增 turn router，删除 Lark 专用 agent-builder 分支

TL;DR

• 这是什么：issue feat: Lark bot 不支持自然语言问答（无通用对话 turn 路由层） #489 的删除优先 solver 结论，现有 ChatRoutePolicyGAgent + ChatRouteResolver 已经是 typed ingress route 主干。
• 结论是什么：不建议再加“自然语言 turn router”；建议删除 NyxRelayAgentBuilderFlow 这条 Lark 专用重复分支，把调用方并到 AgentBuilderCardFlow。
• controller 下一步：让实现 solver 走删除/合并 slice，保留 ChatRouting 主干，不碰 NyxID / chrono-* 外部仓库。

---

详细说明

反向链路显示，ChatRouteResolver 不是死代码：它已经被 relay conversation admission、NyxId direct chat、Responses/Messages、voice policy-aware endpoint 使用；ChatRoutePolicyGAgent 也拥有 actor 状态并经 readmodel 查询。删除整个 ChatRouting 会破坏多入口策略能力，所以不能按“无 router”再造一套新抽象。

真正可删的是 agents/Aevatar.GAgents.Authoring.Lark/NyxRelayAgentBuilderFlow.cs：它只被 ChannelConversationTurnRunner 和自己的测试引用，功能与 AgentBuilderCardFlow 的私聊文本命令、card action、格式化路径高度重叠。删除后 runner 不再需要 relayDecisionMatched 分叉；unknown slash 和普通自然语言继续落到现有 LLM / Ornn skill fallback，agent-builder 管理命令走同一个 flow。

---

📎 完整 codex 原始输出(存档备查)

---

solver: delete
issue: 489
cluster: lark-bot-no-natural-language-turn-router
verdict: propose

Classification

(c) Replaceable with existing

Recommended action

不要新增“Lark 自然语言 turn router”，也不要删除已经投入使用的 ChatRoutePolicyGAgent / ChatRouteResolver 主干。删除对象应收窄到 Lark 专用重复分支 NyxRelayAgentBuilderFlow：它用 bool fall-through 表达 unknown slash / pass-to-LLM，与 AgentBuilderCardFlow 已有私聊文本命令和 card action resolver 重复。调用方 ChannelConversationTurnRunner.TryHandleAgentBuilderAsync 统一调用 AgentBuilderCardFlow.TryResolveAsync 和 AgentBuilderCardFlow.FormatToolResult；普通文本、/daily、unknown slash 继续返回 null 并进入现有 LlmReplyRequested / Ornn skill fallback。保留 ChatRouting typed policy/readmodel/actor，因为它已经服务 relay、direct chat、Responses/Messages、voice 四类入口。

Concrete plan (if propose)

• Files to delete: agents/Aevatar.GAgents.Authoring.Lark/NyxRelayAgentBuilderFlow.cs
• Caller migrations: agents/Aevatar.GAgents.NyxidChat/ChannelConversationTurnRunner.cs TryHandleAgentBuilderAsync 删除 relayDecisionMatched 分支，只调用 AgentBuilderCardFlow.TryResolveAsync(...); tool result 统一走 AgentBuilderCardFlow.FormatToolResult(...)
• Caller migrations: agents/Aevatar.GAgents.Authoring.Lark/AgentBuilderActionIds.cs 删除指向 NyxRelayAgentBuilderFlow 的注释引用
• Caller migrations: agents/Aevatar.GAgents.Authoring.Lark/AgentBuilderJson.cs 删除指向 NyxRelayAgentBuilderFlow 的注释引用
• Tests to delete: test/Aevatar.GAgents.ChannelRuntime.Tests/NyxRelayAgentBuilderFlowTests.cs
• Tests to keep/update: test/Aevatar.GAgents.ChannelRuntime.Tests/AgentBuilderCardFlowTests.cs 覆盖 /agents、/agent-status、/delete-agent <id> confirm、unknown slash fall-through、非 slash fall-through、group restriction 若仍要求群聊限制
• LOC delta: about -555 direct deletion before small caller/test adjustments (wc -l: 328 production + 227 test)
• First slice: deletion/collapse slice: remove NyxRelayAgentBuilderFlow, migrate runner to AgentBuilderCardFlow, move any still-load-bearing command assertions into AgentBuilderCardFlowTests, run channel runtime tests and bash tools/ci/test_stability_guards.sh
• Later design slice: none for deletion solver; if structural solver wants more, it should refine existing ChatRoutePolicy action semantics, not add a second router

Reverse-evidence (why this is safe to delete)

• No public API breaks from deleting NyxRelayAgentBuilderFlow: rg shows references only in ChannelConversationTurnRunner, comments, and NyxRelayAgentBuilderFlowTests; no Host endpoint, proto, SDK, or external repo reference.
• No external repo dependency on the code: rg in ../NyxID found no Aevatar.ChatRouting, ChatRoute, NyxRelayAgentBuilderFlow, ChannelConversationTurnRunner, or LlmReplyRequested references; chrono-* untouched.
• Tests covering the path are themselves not load-bearing: NyxRelayAgentBuilderFlowTests.cs asserts the duplicate flow’s rendering and fall-through behavior; equivalent current behavior can be asserted against AgentBuilderCardFlowTests and runner fallback tests.
• Existing natural-language route already exists: ChannelConversationTurnRunner falls through to ConversationTurnResult.LlmReplyRequested; docs/canon/daily-command-pipeline.md records /daily and unknown slash LLM/skill fallback.
• Existing typed ingress router must stay: ChatRouteResolver is called by ConversationGAgent.ResolveInboundTargetRefAsync, NyxIdChatConversationCreateCommandTargetResolver, ResponsesChatRouteDecisionPort, and voice policy endpoints; deleting it would break active route policy behavior.
• Git history confirms intent: ChatRouteResolver was added for issue Ingress v1 · Phase 2 — ChatRouteResolver lib + QueryPort + env fallback #693 as ingress v1 policy resolver; NyxRelayAgentBuilderFlow started as Nyx-backed Lark relay cutover code and later was narrowed while AgentBuilderCardFlow became the broader shared command/card resolver.

Risks

• If NyxRelayAgentBuilderFlow.FormatToolResult has Lark-specific card rendering that AgentBuilderCardFlow.FormatToolResult lacks, deletion could regress Lark UI. The visible tests show most of that rendering already uses shared AgentBuilderCardContent or equivalent card paths; migrate only missing assertions before deleting.
• If group-chat known-command restriction is intentionally different between relay text and card flow, AgentBuilderCardFlow may need a small no-new-abstraction condition before deleting the duplicate file.
• If operators depend on exact English text from NyxRelayAgentBuilderFlow replies, text snapshots may change. No public API surface found, but user-visible copy can still regress.

Escalation triggers (if any)

• none

Reasoning trace (internal — for meta-judge)

• Why I claim this can be deleted: the issue’s desired “natural language router” is already covered by fall-through to LLM and by the broader ChatRouting policy stack; the remaining Lark-specific flow is a duplicate specialized parser/formatter with one production caller.
• What I checked to verify safety: issue feat: Lark bot 不支持自然语言问答（无通用对话 turn 路由层） #489 body/comments, audit iter 290, CLAUDE deletion-first clause, rg call sites for ChatRouteResolver, NyxRelayAgentBuilderFlow, AgentBuilderCardFlow, IChatRoutePolicyQueryPort, ChatRouteDecision, LlmReplyRequested; git logs for ChatRouteResolver, ChannelConversationTurnRunner, NyxRelayAgentBuilderFlow; sibling ../NyxID grep.
• What I cannot decide alone: whether to preserve exact Lark card copy and group-command restriction text; implementation solver should pin those with migrated tests if product cares.

SOLVER_DONE:delete:propose:delete NyxRelayAgentBuilderFlow and merge runner onto AgentBuilderCardFlow; keep existing ChatRoutePolicy/Resolver主干:first-slice=remove duplicate Lark agent-builder branch, migrate ChannelConversationTurnRunner to AgentBuilderCardFlow, move load-bearing tests

⟦AI:AUTO-LOOP⟧

[loning]

🤖 structural solver：建议先显式化 turn route，再拆 typed executor

TL;DR

• 这是什么：对 issue 489 的 structural 方案，定位为 Lark/Nyx relay turn routing 的隐式 fall-through 和 runner 内执行耦合。
• 结论：verdict=propose；先做一个不新增 port 的窄切片，再做 ChannelTurnRouteDecision + typed AgentBuilder executor。
• 下一步：controller 可让 implementer 先落 first slice；若有人要扩 ChatRouteAction proto 或加新 actor，再升级人工确认。

---

详细说明

当前 ConversationGAgent 已在入站 admission 前调用 ChatRouteResolver，所以问题不是“完全没有 typed route policy”。真正的问题在下一层：ChannelConversationTurnRunner 仍用 ordered branch + bool false 表达本地 slash、agent-builder、unknown slash、普通 LLM fallback 的差异，并且在 runner 内直接创建 AgentBuilderTool。IActorDispatchPort 是 actor 之间发命令的标准通道；本方案不新增 envelope、不新增 projection phase，也不要求 NyxID/Ornn 改能力。

建议拆两步：第一步只把 NyxRelayAgentBuilderFlow.TryResolve 的 false 改成可测试的本地 outcome（例如 non-slash / unknown-slash-forward-to-LLM / known-command），不改运行拓扑。第二步新增 transient ChannelTurnRouteDecision 值对象，并把 AgentBuilder 的 JSON tool adapter 拆到边界，runner 走 typed AgentBuilderActionRequest，让 AgentBuilderTool 和 Lark direct route 共用同一个 executor。

方案	改动	代价	取舍
first slice	显式化 AgentBuilder fall-through	2 文件，约 +80/-25 LOC	消除 false 多义性，但 runner 仍执行 tool
structural slice	ChannelTurnRouteDecision + typed executor	6-8 文件，约 +260/-120 LOC	结构完整，测试面更清楚
不建议	扩 ChatRouteAction oneof 承载本地 slash	proto + policy 语义变更	需要人工确认，容易把 ingress policy 和 local turn 混在一起

---

📎 完整 codex 原始输出(存档备查)

---

solver: structural
issue: 489
cluster: lark-bot-no-natural-language-turn-router
verdict: propose

CLAUDE clause violated (quoted verbatim)

• 命令骨架内聚：标准生命周期 Normalize -> Resolve Target -> Build Context -> Build Envelope -> Dispatch -> Receipt -> Observe；业务模块只负责目标解析与载荷/结果映射。

• 依赖反转：上层依赖抽象，禁止跨层反向依赖和对具体实现的直接耦合。

• 读写分离：Command -> Event，Query -> ReadModel；异步完成通过事件通知，不在会话内拼装流程。

• AI 对话主链必须流式化：实时会话入口必须使用 ChatStreamAsync；ChatAsync 仅可用于明确的非交互式离线场景。

Recommended framing

当前仓库已经有 ChatRouteResolver/ChatRouteDecision 处理入站 policy target，ConversationGAgent 也已经在 actor inbox 内把 resolved ChatRouteAction 填到 deferred LLM request 的 TargetRef。所以 issue 489 不应再造第二套全局 router，也不应扩 NyxID/Ornn。结构上应补的是 channel turn 内部的 transient route decision：先把“本地 slash、agent-builder、card action、Ornn skill shortcut、普通 LLM fallback、ignore”归一成 ChannelTurnRouteDecision，再由 runner 只做 decision execution。AgentBuilder 的执行不再通过 ActivatorUtilities.CreateInstance<AgentBuilderTool> + JSON 参数绕回 tool adapter，而是抽出 typed AgentBuilderActionExecutor，让 LLM tool adapter 与 Lark direct route 两个调用方共用同一业务执行面。

Concrete plan

• New abstractions (if any):
  • ChannelTurnRouteDecision：typed transient value object, not a port; Layer=Application/runtime boundary; Project=agents/Aevatar.GAgents.NyxidChat; File=Routing/ChannelTurnRouteDecision.cs. Cases: WorkflowResume, SlashCommand, AgentBuilderDirectReply, AgentBuilderAction, ForwardToLLM, Ignore.
  • ChannelTurnRoutePlanner：concrete coordinator, not an interface; Layer=Application/runtime boundary; Project=agents/Aevatar.GAgents.NyxidChat; File=Routing/ChannelTurnRoutePlanner.cs. It composes existing slash registry, workflow resume service, NyxRelayAgentBuilderFlow, AgentBuilderCardFlow, and LLM fallback classification.
  • AgentBuilderActionExecutor + AgentBuilderActionRequest：typed shared executor justified by 2 concrete callers: AgentBuilderTool JSON adapter and Lark direct route. Layer=Application boundary for authoring; Project=agents/Aevatar.GAgents.Authoring.Lark; Files=AgentBuilderActionExecutor.cs, AgentBuilderActionRequest.cs.
• Existing primitives:
  • Reuse ChatRouteResolver/ChatRouteDecision only for outer policy target; do not make it own local slash semantics.
  • Reuse actor inbox already present in ConversationGAgent; no new actor hop for routing.
  • Reuse ChatStreamAsync/IAsyncEnumerable in ConversationReplyGenerator; natural language and skill shortcut remain deferred LLM reply path.
  • Reuse IActorDispatchPort/existing catalog and runner command ports inside AgentBuilderActionExecutor; ACK remains accepted-only.
  • Reuse projection/readmodel only for existing ChatRoutePolicy and AgentBuilder query ports; no query-time priming.
• Files:
  • agents/Aevatar.GAgents.Authoring.Lark/NyxRelayAgentBuilderFlow.cs: replace bool fall-through with typed local outcome; keep old wrapper only if tests still need it during first slice.
  • agents/Aevatar.GAgents.Authoring.Lark/AgentBuilderCardFlow.cs: return typed action request instead of JSON-only AgentBuilderFlowDecision for direct route.
  • agents/Aevatar.GAgents.Authoring.Lark/AgentBuilderTool.cs: become JSON external tool adapter -> AgentBuilderActionRequest -> executor.
  • agents/Aevatar.GAgents.Authoring.Lark/AgentBuilderActionExecutor.cs: new typed executor shared by tool and direct Lark route.
  • agents/Aevatar.GAgents.Authoring.Lark/AgentBuilderActionRequest.cs: new typed request/result contract.
  • agents/Aevatar.GAgents.NyxidChat/ChannelConversationTurnRunner.cs: move ordered route classification into planner; execute ChannelTurnRouteDecision; remove direct ActivatorUtilities.CreateInstance<AgentBuilderTool>.
  • agents/Aevatar.GAgents.NyxidChat/Routing/ChannelTurnRouteDecision.cs: new transient decision value object.
  • agents/Aevatar.GAgents.NyxidChat/Routing/ChannelTurnRoutePlanner.cs: new concrete planner.
  • agents/Aevatar.GAgents.NyxidChat/ServiceCollectionExtensions.cs: register executor/planner as needed.
  • docs/canon/daily-command-pipeline.md: small wording update only if implementation changes observable route terminology.
• LOC delta: structural slice ~+260 / -120 net +140. First slice ~+80 / -25 net +55.
• Tests to add:
  • test/Aevatar.GAgents.ChannelRuntime.Tests/NyxRelayAgentBuilderFlowTests.cs: unknown slash, /daily, non-slash, and known commands return distinct outcomes, not the same false/null.
  • test/Aevatar.GAgents.ChannelRuntime.Tests/ChannelConversationTurnRunnerTests.cs: /daily routes to LlmReplyRequested with skill shortcut prompt and does not call AgentBuilder executor.
  • test/Aevatar.GAgents.ChannelRuntime.Tests/ChannelConversationTurnRunnerTests.cs: unknown slash routes to LLM/Ornn skill discovery, not local unknown-command reply.
  • test/Aevatar.GAgents.ChannelRuntime.Tests/ChannelConversationTurnRunnerTests.cs: known /agents routes to typed AgentBuilder action and returns direct reply/rendered card.
  • test/Aevatar.GAgents.ChannelRuntime.Tests/AgentBuilderActionExecutorTests.cs or existing authoring test project: JSON tool adapter and direct Lark route produce identical accepted-only behavior for run_agent/disable_agent/delete_agent.
• proto changes (if any): none recommended. ChatRouteDecision in src/Aevatar.ChatRouting.Abstractions/chat_route_policy.proto is already transient and should not be expanded for local turn internals in this issue.
• Runtime cost:
  • First slice: 0 actor hops, 0 network calls, +1 small enum/record allocation per inbound turn if implemented as record; <1 microsecond classification overhead.
  • Structural slice without new actor: 0 new actor hops, 0 new projection writes, same LLM/tool network calls; +1 planner decision object and +1 typed request object on matched agent-builder path, roughly <1 KB allocation per turn.
  • If executor adds a self-message hop, expected local actor-hop latency is runtime dependent, roughly 1-10 ms local/Orleans scheduling; not required for this issue.
• First slice: yes. No-new-port narrow plan: update NyxRelayAgentBuilderFlow to expose explicit outcome for NotSlash, UnknownSlashForwardToLLM, DailySkillShortcut, KnownCommand, PrivateChatRejected; update tests and runner condition only enough to stop interpreting all false as the same semantic.
• Later design slice: introduce ChannelTurnRouteDecision and typed AgentBuilderActionExecutor; keep ChatRouteResolver as outer policy target resolver and keep natural language on ChatStreamAsync.

Risks

• This framing costs more than the minimal patch because it touches both routing and AgentBuilder tool execution, not just the /daily branch.
• The over-engineering risk is adding a generic IChannelTurnRouter port with one production adapter. Keep it bounded by using concrete planner/value objects first; introduce an interface only if a second real adapter appears.
• Moving AgentBuilder execution out of AgentBuilderTool can expose existing JSON-shaped tests as too adapter-focused; keep JSON tests at the tool boundary and add behavior tests against typed executor.
• Do not extend ChatRouteAction for local slash/card actions unless a maintainer accepts that the global chat route policy owns local channel-turn semantics.

Escalation triggers (if any)

• ESCALATE_REASON:core-proto:adding new ChatRouteAction oneof cases for local slash/agent-builder/workflow actions
• ESCALATE_REASON:actor-topology:introducing a new core actor type or making AgentBuilder direct commands a new actor lifecycle
• ESCALATE_REASON:pipeline-phase:adding a new projection or observation phase for route decisions
• ESCALATE_REASON:canon-vocabulary:changing docs/canon/architecture-vocabulary.md router/port/seam definitions
• ESCALATE_REASON:cross-cluster:touching in-flight ChatRun JSON contract or invocation Metadata clusters to make this pass
• ESCALATE_REASON:benchmark:adding a mandatory actor inbox hop on every Lark message without a local latency benchmark

Reasoning trace (internal — for meta-judge)

• Why this structure beats a documented exception: the failure is not an isolated /daily gap; source confirms normal text already reaches LlmReplyRequested and ConversationReplyGenerator already uses ChatStreamAsync. A comment saying “unknown slash falls through” would preserve the same multi-meaning false and keep future approval/tool routes piling into runner branches.
• Why I picked this abstraction over alternatives: ChatRouteResolver already exists but its proto explicitly models ingress policy target and says ChatRouteDecision is transient. Reusing it for local slash/card execution would mix policy target with turn-local command execution. A concrete ChannelTurnRouteDecision value object is narrower than a new port and does not create a one-adapter seam.
• What I cannot decide alone: whether maintainers want local turn route decisions to become part of the public chat route policy proto. That would be a core-proto escalation, not a default implementation detail.
  SOLVER_DONE:structural:propose:typed-turn-route-decision-plus-typed-agent-builder-executor:first-slice=make-agent-builder-fallthrough-explicit-with-local-outcome-enum

⟦AI:AUTO-LOOP⟧