iter57 cluster-071: typed ChatRequestEvent.llm_control + WorkflowChatSource (#953)

[loning]

摘要

iter57 cluster-071(severity:high)— 新 typed ChatRequestEvent.llm_control + WorkflowChatSource,删 Metadata/ToolContext-as-LLM-control 双重语义

• Old:workflow 和 LLM 控制信息(NyxID/model/route)挤在 ToolContext / Metadata bag → 双重语义(违反 "API 字段单一语义")
• New:typed LLMControlContext + typed WorkflowChatSource;删 Metadata fallback + ToolContext-as-LLM-control paths

违反:CLAUDE.md「核心语义强类型」「字段命名 Metadata 决策树」「API 字段单一语义」

Phase 9 共识链路(4-round + reflector)

Round	Verdict
r1	converge:typed contract scope
r2	converge:NyxID/model/route carrier + step-control scope
r3	escalate:stalled(2:1 ToolContext vs llm_control)
reflector r1	retry-fix:CLAUDE 单一语义判 ToolContext reuse 出局
r4	META_JUDGE_DONE:consensus:structural:add narrow ChatRequestEvent.llm_control and typed WorkflowChatSource; remove Metadata/ToolContext carrier paths ✅

Scope

• 新 LLMControlContext typed contract + LLMControlContextMapper
• 新 ITypedConversationReplyGenerator
• 新 WorkflowChatSource typed
• 涉及:NyxID/channel reply generation / Studio authoring / scheduled skill runner / workflow / streaming proxy / response paths
• 测试覆盖 typed carrier behavior

local PASS:architecture + test_stability + build + focused suites

closes #953

🤖 Generated with Claude Code via codex-refactor-loop iter57

⟦AI:AUTO-LOOP⟧

[loning]

📊 codex 进展 fix-pr957-r3 (⏳ 进行中; 已跑 3 min)

+        result.Succeeded.Should().BeTrue();
+        result.Request!.Source.Should().BeEquivalentTo(
+            WorkflowChatSource.DefinitionActor("source-actor-1", "direct"));
+        result.Request.WorkflowName.Should().Be("direct");
+        result.Request.ActorId.Should().Be("source-actor-1");
+    }
+
+    [Fact]
     public void ChatRunRequestNormalizer_ShouldDerivePromptAndInputParts_FromMultimodalInput()
     {
         var input = new ChatInput

tokens used
111,306
Fixed PR #957 round 3.

Changed:
- Legacy non-inline `workflow + agentId` now normalizes to `WorkflowChatSource.DefinitionActor(agentId, workflowName)` instead of dropping `agentId` into catalog workflow path.
- Added API regression test for legacy `Workflow + AgentId`.
- Added resolver regression test proving the source actor binding is actually resolved and used.

Verified:
- `dotnet test test/Aevatar.Workflow.Host.Api.Tests/Aevatar.Workflow.Host.Api.Tests.csproj --nologo --filter "FullyQualifiedName~WorkflowCapabilityEndpointsCoverageTests"`
- `dotnet test test/Aevatar.Workflow.Application.Tests/Aevatar.Workflow.Application.Tests.csproj --nologo --filter "FullyQualifiedName~WorkflowRunActorResolverTests"`
- `bash tools/ci/test_stability_guards.sh`

自动更新每 10 分钟;edit-in-place 不堆评论;codex 完成后此 comment 自动删除(per Auric "完成后删掉就好了 否则太占空间")。
🤖 controller progress reporter

[loning]

📊 codex 进展 review-pr957-omnibus-r3 (⏳ 进行中; 已跑 12 min)

Coverage gap: the regression test does not cover legacy non-inline `Workflow + AgentId`, which is the path now regressed.

Focused tests run on PR head `c83751e9`:
- `dotnet test test/Aevatar.Workflow.Host.Api.Tests/Aevatar.Workflow.Host.Api.Tests.csproj --no-restore --filter "FullyQualifiedName~WorkflowCapabilityEndpointsCoverageTests|FullyQualifiedName~ChatEndpointsInternalTests" --nologo`
- `dotnet test test/Aevatar.Workflow.Application.Tests/Aevatar.Workflow.Application.Tests.csproj --no-restore --filter "FullyQualifiedName~WorkflowRunActorResolverTests|FullyQualifiedName~WorkflowRunFallbackCoverageTests" --nologo`

REVIEW_DONE:957:omnibus:reject
⟦AI:AUTO-LOOP⟧
tokens used
111,324
reject

[P1] Legacy non-inline `workflow + agentId` compatibility is broken. In [ChatRunRequestNormalizer.cs](/tmp/aevatar-pr957-review/src/workflow/Aevatar.Workflow.Infrastructure/CapabilityApi/ChatRunRequestNormalizer.cs:131), legacy input with both `Workflow` and `AgentId` now returns `WorkflowChatSource.CatalogWorkflow(...)` before considering `AgentId`. The final `WorkflowChatRunRequest.ActorId` becomes `null`, and [WorkflowRunActorResolver.cs](/tmp/aevatar-pr957-review/src/workflow/Aevatar.Workflow.Application/Runs/WorkflowRunActorResolver.cs:69) bypasses source-actor resolution and creates a fresh registry-backed run. Before this PR, the same request preserved `ActorId` and resolved through the specified actor binding. This violates “legacy compatibility 保留” and can still silently create-new for a caller that explicitly supplied an actor.

Verified fixed parts:
- typed inline-yaml `source.ActorId` is now preserved by the normalizer.
- resolver reads `request.Source.ActorId`.
- regression tests cover typed inline-yaml and legacy inline-yaml behavior.

Coverage gap: the regression test does not cover legacy non-inline `Workflow + AgentId`, which is the path now regressed.

Focused tests run on PR head `c83751e9`:
- `dotnet test test/Aevatar.Workflow.Host.Api.Tests/Aevatar.Workflow.Host.Api.Tests.csproj --no-restore --filter "FullyQualifiedName~WorkflowCapabilityEndpointsCoverageTests|FullyQualifiedName~ChatEndpointsInternalTests" --nologo`
- `dotnet test test/Aevatar.Workflow.Application.Tests/Aevatar.Workflow.Application.Tests.csproj --no-restore --filter "FullyQualifiedName~WorkflowRunActorResolverTests|FullyQualifiedName~WorkflowRunFallbackCoverageTests" --nologo`

自动更新每 10 分钟;edit-in-place 不堆评论;codex 完成后此 comment 自动删除(per Auric "完成后删掉就好了 否则太占空间")。
🤖 controller progress reporter

[loning]

Verified r3 fix at head 0283beb.

Verdict: approve. GitHub refused a formal approve review because this checkout is authenticated as the PR author, so recording this as a PR comment.

No blocking findings. The legacy Workflow + AgentId non-inline path now normalizes to DefinitionActor(agentId, workflowName), and resolver coverage validates that the source actor binding is read and used for the run. The typed inline-yaml + source.actorId path is also covered at both normalizer and resolver levels.

Validation run from detached PR worktree:

• dotnet test test/Aevatar.Workflow.Host.Api.Tests/Aevatar.Workflow.Host.Api.Tests.csproj --nologo --filter "FullyQualifiedName~WorkflowCapabilityEndpointsCoverageTests"
• dotnet test test/Aevatar.Workflow.Application.Tests/Aevatar.Workflow.Application.Tests.csproj --nologo --filter "FullyQualifiedName~WorkflowRunActorResolverTests"
• bash tools/ci/test_stability_guards.sh

⟦AI:AUTO-LOOP⟧

[loning]

📊 当前状态 — 🔧 fixing(test r4 — 不需要人介入)

维度	值
阶段	Phase 8 fix r4(test-only,迁旧 metadata 测试到 typed LLMControl)
上轮触发	test-add r1 (commit 757feb7c) 只补了 LLMControlContext/Mapper/WorkflowChatSource 覆盖,未迁 NyxIdChatEndpointsCoverageTests / NyxIdLLMProviderRoutingTests / StreamingProxyNyxParticipantCoordinatorTests / ConnectedServicesContextMiddlewareTests / ScopeServiceEndpointsTests(共 ~30 旧 metadata-based 测试)
CI 现状	coverage-quality FAIL(dotnet test 阶段挂),fast-gates / projection-provider-e2e / host-composition-smoke / slow-test-guards 全 pass
派出	fix codex r4(test-only,timeout 5400s)
是否需要人介入	❌ 否(纯测试迁移,production 已切 typed,无规则解读分歧)

下一步自动会做:fix codex 把每个失败测试的 Metadata-based setup/assertion 迁到 typed LLMControlContext API → controller commit/push → CI 重检 → 全绿即合并。

何时需要人介入:

• 同 r4 fix 后仍同类 fail(说明 test 不仅是 carrier 问题,而是 production 真 bug),累计 r6 触发 reflector
• reflector 也 escalate-human 才升 maintainer

🤖 controller status banner

⟦AI:AUTO-LOOP⟧

[loning]

🤖 Phase 8 fix r4 — 28 测试已迁完到 typed LLMControl

维度	值
Round	r4(test-only,production 不动)
Applied	28(NyxIdLLMProviderRouting + NyxIdChatEndpointsCoverage + StreamingProxyNyxParticipantCoordinator + ConnectedServicesContextMiddleware + ScopeServiceEndpoints)
Rejected as false-positive	0
Blocked	1(codex 误判 gh 不可用,controller 接手 post 本评论)
dotnet test Aevatar.AI.Tests	653 passed
dotnet test Aevatar.GAgentService.Integration.Tests	297 passed
dotnet build Aevatar.AI.Core	passed
修复 commit	1b7170fc

变更摘要:

• NyxId LLM provider routing 测试改 LLMControlContext 承载 access_token / route_preference / model_override
• NyxId chat endpoint / interaction 测试断言 ChatRequestEvent.LlmControl(非 Metadata["nyxid.*"])
• StreamingProxy participant coordinator 测试断言 LlmControl token/route
• ConnectedServices middleware 测试 typed ToolContext.ConnectedServices
• ScopeService helper 测试 scoped LLM config → BuildScopedLlmControlAsync,headers 只保留真 header/passthrough

下一步自动:CI 重检(coverage-quality)→ 全绿即 Phase 8 unanimous + auto-merge → close issue #953。

⟦AI:AUTO-LOOP⟧

[codecov]

Codecov Report

❌ Patch coverage is 84.19811% with 67 lines in your changes missing coverage. Please review.
✅ Project coverage is 83.03%. Comparing base (87e9f2d) to head (1b7170f).
⚠️ Report is 2 commits behind head on auto-refact-dev.

Files with missing lines	Patch %	Lines
...ervice.Hosting/Endpoints/ScopeWorkflowEndpoints.cs	50.81%	25 Missing and 5 partials ⚠️
...Service.Hosting/Endpoints/ScopeServiceEndpoints.cs	75.92%	9 Missing and 4 partials ⚠️
...low/Aevatar.Workflow.Core/Modules/LLMCallModule.cs	55.55%	8 Missing ⚠️
...tructure/CapabilityApi/ChatRunRequestNormalizer.cs	93.68%	0 Missing and 6 partials ⚠️
...atar.AI.Core/LLMProviders/OwnerLlmConfigApplier.cs	80.00%	0 Missing and 3 partials ⚠️
...ication/Runs/WorkflowChatRequestEnvelopeFactory.cs	0.00%	1 Missing and 1 partial ⚠️
....AI.Abstractions/LLMProviders/LLMControlContext.cs	97.43%	0 Missing and 1 partial ⚠️
...s/ToolProviders/AgentToolExecutionContextMapper.cs	50.00%	0 Missing and 1 partial ⚠️
src/Aevatar.AI.Core/AIGAgentBase.cs	85.71%	0 Missing and 1 partial ⚠️
src/Aevatar.AI.Core/Chat/ChatRuntime.cs	95.00%	1 Missing ⚠️
... and 1 more

@@                 Coverage Diff                 @@
##           auto-refact-dev     #957      +/-   ##
===================================================
- Coverage            83.04%   83.03%   -0.02%     
===================================================
  Files                  987      990       +3     
  Lines                63558    63862     +304     
  Branches              8214     8248      +34     
===================================================
+ Hits                 52780    53025     +245     
- Misses                7246     7282      +36     
- Partials              3532     3555      +23     

Flag	Coverage Δ
ci	83.03% <84.19%> (-0.02%)	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.

Files with missing lines	Coverage Δ
...stractions/LLMProviders/LLMControlContextMapper.cs	100.00% <100.00%> (ø)
...Aevatar.AI.Abstractions/LLMProviders/LLMRequest.cs	77.61% <100.00%> (+0.33%)	⬆️
src/Aevatar.AI.Core/RoleGAgent.cs	82.38% <100.00%> (+0.01%)	⬆️
src/Aevatar.AI.Core/Tools/ToolCallLoop.cs	90.45% <100.00%> (+0.07%)	⬆️
...ice.Application/Responses/MessagesCommandFacade.cs	79.67% <100.00%> (+0.40%)	⬆️
...ce.Application/Responses/ResponsesCommandFacade.cs	77.10% <100.00%> (+0.20%)	⬆️
...ication.Abstractions/Runs/WorkflowChatRunModels.cs	100.00% <100.00%> (ø)
...w.Application/Runs/WorkflowDirectFallbackPolicy.cs	74.54% <100.00%> (+0.47%)	⬆️
...kflow.Application/Runs/WorkflowRunActorResolver.cs	93.01% <100.00%> (+0.42%)	⬆️
...orkflow/Aevatar.Workflow.Core/WorkflowRunGAgent.cs	83.07% <100.00%> (ø)
... and 12 more

... and 6 files with indirect coverage changes

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.

[loning]

🤖 Phase 8 Re-review — tests-r6-quickverify (头 SHA 1b7170f)

Verdict: approve-r6-quick

I quick-verified the r4-only range 757feb7c..1b7170fc against PR head 1b7170fc1bd3c42c2aefd22e10007bed385380b5.

Checks performed:

• Scope is test-only: all 5 changed files are under test/**; no production file changed in this range.
• Carrier migration is real: old keyed Metadata[LLMRequestMetadataKeys.*] / nyxid.* assertions and fixtures were replaced with typed LlmControl / LLMControlContext / AgentToolConnectedServicesContext assertions.
• Assertions remain concrete: token, route, model override, max tool rounds, user memory, connected services, and custom metadata preservation are still checked with exact expected values or strict NotContainKey checks for removed metadata keys.
• No test disabling found: no [Skip], Skip = ..., Explicit, or disable-style marker in the r4 diff.
• CI state on GitHub is green for the relevant gates: coverage-quality, codecov/patch, fast-gates, host-composition-smoke, projection-provider-e2e, and slow-test-guards are successful on this head.

No blocking issue found. This r4 push is a narrow test-only typed-carrier migration and is OK to merge.

REVIEW_DONE:957:tests-r6-quickverify:approve-r6-quick
⟦AI:AUTO-LOOP⟧

[loning]

✅ Phase 8 r4 + r6 quick-verify approve — auto-merging

Review	Verdict
r5 (3/3) on 757feb7	architect approve · tests approve · quality approve
r6 quick-verify on 1b7170f	tests approve(test-only carrier 迁移,无 production 改动,无 [Skip],断言未放宽)

r4 fix 是 28 测试断言迁移(NyxIdLLMProviderRouting / NyxIdChatEndpointsCoverage / StreamingProxyNyxParticipantCoordinator / ConnectedServicesContextMiddleware / ScopeServiceEndpoints)到 typed LLMControlContext carrier;production code 0 改动。CI 现状全绿(coverage-quality / codecov/patch / fast-gates / host-composition-smoke / projection-provider-e2e / slow-test-guards)。

下一步:gh pr merge 957 --squash --delete-branch,close #953。

🤖 controller consensus + auto-merge

⟦AI:AUTO-LOOP⟧