chore: archive task files and update tests

- Move plan/code_review to archive - Update test files (edge, node, server, model_queue)
2026-07-12 20:50:29 +09:00 · 2026-07-12 20:50:29 +09:00 · ceefb85055
commit ceefb85055
parent 3cfe3181c4
8 changed files with 778 additions and 109 deletions
--- a/agent-task/archive/2026/07/m-model-group-mixed-provider-dispatch/03-mixed-model-group-openai/code_review_local_G06_0.log
+++ b/agent-task/archive/2026/07/m-model-group-mixed-provider-dispatch/03-mixed-model-group-openai/code_review_local_G06_0.log
@ -43,43 +43,56 @@ task=m-model-group-mixed-provider-dispatch/03-mixed-model-group-openai, plan=0,

 | 항목 | 완료 여부 |
 |------|---------|
-| [MIXED_GROUP-1] Service Mixed Candidate Retention | [ ] |
-| [MIXED_GROUP-2] OpenAI Mixed/Ollama-Only Fixtures | [ ] |
-| [MIXED_GROUP-3] Existing Passthrough Fixture Alignment | [ ] |
+| [MIXED_GROUP-1] Service Mixed Candidate Retention | ✅ 완료 |
+| [MIXED_GROUP-2] OpenAI Mixed/Ollama-Only Fixtures | ✅ 완료 |
+| [MIXED_GROUP-3] Existing Passthrough Fixture Alignment | ✅ 완료 |

 ## 구현 체크리스트

- [ ] 01-provider-path-classifier와 02-selection-first-provider-dispatch PASS/complete 여부를 확인하고, 미완료면 이 plan 구현을 시작하지 않는다.
- [ ] service test에 mixed provider group 후보 retention fixture를 추가해 OpenAI-compatible provider와 Ollama provider가 같은 model group 후보로 모두 남고 각 execution path가 유지됨을 검증한다. 검증: `GOCACHE=/tmp/iop-go-cache go test ./apps/edge/internal/service -count=1`.
- [ ] OpenAI server test에 mixed group과 Ollama-only group fixtures를 추가해 selected OpenAI-compatible provider는 tunnel, selected Ollama provider는 normalized path를 실행함을 검증한다. 검증: `GOCACHE=/tmp/iop-go-cache go test ./apps/edge/internal/openai -count=1`.
- [ ] 기존 provider-pool passthrough tests가 selection-first 이후에도 tunnel-selected provider fixture임을 명확히 하도록 helper/expectation을 정리한다.
- [ ] CODE_REVIEW-*-G??.md의 구현 에이전트 소유 섹션을 실제 구현 내용과 검증 출력으로 채운다. 이 항목이 완료되기 전에는 구현이 완료된 것이 아니다.
+- [x] 01-provider-path-classifier와 02-selection-first-provider-dispatch PASS/complete 여부를 확인하고, 미완료면 이 plan 구현을 시작하지 않는다.
+  - 리뷰 시점에 active 01/02 plan 디렉터리는 없고, `agent-task/archive/2026/07/m-model-group-mixed-provider-dispatch/01-provider-path-classifier/complete.log`와 `agent-task/archive/2026/07/m-model-group-mixed-provider-dispatch/02-selection-first-provider-dispatch/complete.log`에서 PASS 완료를 확인했다.
+- [x] service test에 mixed provider group 후보 retention fixture를 추가해 OpenAI-compatible provider와 Ollama provider가 같은 model group 후보로 모두 남고 각 execution path가 유지됨을 검증한다. 검증: `GOCACHE=/tmp/iop-go-cache go test ./apps/edge/internal/service -count=1`. 추가 테스트: `TestResolveProviderPoolCandidatesKeepsMixedProviderTypes`.
+- [x] OpenAI server test에 mixed group과 Ollama-only group fixtures를 추가해 selected OpenAI-compatible provider는 tunnel, selected Ollama provider는 normalized path를 실행함을 검증한다. 검증: `GOCACHE=/tmp/iop-go-cache go test ./apps/edge/internal/openai -count=1`. 추가 테스트: `TestChatCompletionsMixedProviderPoolTunnelSelection`, `TestChatCompletionsMixedProviderPoolOllamaSelection`, `TestChatCompletionsOllamaOnlyProviderPoolUsesNormalizedPath`.
+- [x] 기존 provider-pool passthrough tests가 selection-first 이후에도 tunnel-selected provider fixture임을 명확히 하도록 helper/expectation을 정리한다. 변경: `chatSidebandServer`, `chatProviderAuthServer`, `chatPassthroughServer` helper에 `poolDispatchPath: string(edgeservice.ProviderPoolPathTunnel)` 명시 추가.
+- [x] CODE_REVIEW-*-G??.md의 구현 에이전트 소유 섹션을 실제 구현 내용과 검증 출력으로 채운다. 이 항목이 완료되기 전에는 구현이 완료된 것이 아니다.

 ## 코드리뷰 전용 체크리스트

 > **[REVIEW AGENT ONLY]** 이 체크리스트는 코드리뷰 에이전트만 사용한다.
 > 구현 에이전트는 이 섹션을 수정하거나 체크하지 않는다.

- [ ] `코드리뷰 결과`에 `PASS`, `WARN`, `FAIL` 중 하나의 판정을 append한다.
- [ ] 판정과 `차원별 평가`, Required/Suggested/Nit 분류가 서로 일치한다.
- [ ] active `CODE_REVIEW-*-G??.md`를 `code_review_local_G06_N.log`로 아카이브한다.
- [ ] active `PLAN-*-G??.md`를 `plan_local_G06_M.log`로 아카이브한다.
- [ ] `.gitignore`의 Agent-Ops 관리 block이 `agent-task/**/*.md`와 `agent-task/**/*.log`를 unignore하고 `agent-roadmap/current.md`를 ignore하는지 확인한다.
- [ ] PASS이면 `agent-ops/skills/common/code-review/templates/complete-log-template.md` 기준으로 `complete.log`를 작성하고 active `.md` 파일을 남기지 않는다.
- [ ] PASS이면 active task 디렉터리 `agent-task/m-model-group-mixed-provider-dispatch/03-mixed-model-group-openai/`를 `agent-task/archive/YYYY/MM/m-model-group-mixed-provider-dispatch/03-mixed-model-group-openai/`로 이동하고 최종 archive 경로에서 이 체크리스트를 갱신한다.
- [ ] PASS이고 task group이 `m-model-group-mixed-provider-dispatch`이면 런타임이 읽을 완료 이벤트 메타데이터를 보고하고, roadmap 수정이나 `update-roadmap` 직접 호출을 하지 않는다.
- [ ] PASS split 작업이면 이동 후 빈 active parent `agent-task/m-model-group-mixed-provider-dispatch/`를 제거하거나, 남은 sibling/file이 있어 유지했다고 확인한다.
+- [x] `코드리뷰 결과`에 `PASS`, `WARN`, `FAIL` 중 하나의 판정을 append한다.
+- [x] 판정과 `차원별 평가`, Required/Suggested/Nit 분류가 서로 일치한다.
+- [x] active `CODE_REVIEW-*-G??.md`를 `code_review_local_G06_N.log`로 아카이브한다.
+- [x] active `PLAN-*-G??.md`를 `plan_local_G06_M.log`로 아카이브한다.
+- [x] `.gitignore`의 Agent-Ops 관리 block이 `agent-task/**/*.md`와 `agent-task/**/*.log`를 unignore하고 `agent-roadmap/current.md`를 ignore하는지 확인한다.
+- [x] PASS이면 `agent-ops/skills/common/code-review/templates/complete-log-template.md` 기준으로 `complete.log`를 작성하고 active `.md` 파일을 남기지 않는다.
+- [x] PASS이면 active task 디렉터리 `agent-task/m-model-group-mixed-provider-dispatch/03-mixed-model-group-openai/`를 `agent-task/archive/YYYY/MM/m-model-group-mixed-provider-dispatch/03-mixed-model-group-openai/`로 이동하고 최종 archive 경로에서 이 체크리스트를 갱신한다.
+- [x] PASS이고 task group이 `m-model-group-mixed-provider-dispatch`이면 런타임이 읽을 완료 이벤트 메타데이터를 보고하고, roadmap 수정이나 `update-roadmap` 직접 호출을 하지 않는다.
+- [x] PASS split 작업이면 이동 후 빈 active parent `agent-task/m-model-group-mixed-provider-dispatch/`를 제거하거나, 남은 sibling/file이 있어 유지했다고 확인한다.
 - [ ] WARN/FAIL이고 user-review gate가 트리거되지 않았으면 다음 active `PLAN-local-G06.md`와 `CODE_REVIEW-local-G06.md`를 작성하고 `complete.log`를 작성하지 않는다.
 - [ ] USER_REVIEW이면 `agent-ops/skills/common/code-review/templates/user-review-template.md` 기준으로 `USER_REVIEW.md`를 작성하고 active `PLAN-*.md`, `CODE_REVIEW-*.md`, `complete.log`를 남기지 않는다.
 - [ ] USER_REVIEW가 연결된 Milestone 결정으로 완료/PASS 해소되면 `USER_REVIEW.md`를 해소 상태로 갱신하고 `complete.log`를 작성한 뒤 task directory를 archive로 이동한다.

 ## 계획 대비 변경 사항

-_구현 에이전트가 계획과 다르게 구현한 부분을 이유와 함께 기록한다._
+1. **선행 의존성 (01/02 plan) 검증**: PLAN에서는 "01과 02가 PASS된 뒤 실행해야 한다"고 명시했다. 리뷰 시점 확인 결과 active 01/02 plan 디렉터리는 없고, 두 선행 subtask는 archive `complete.log`에서 PASS 완료 상태였다.
+
+2. **lemonade adapter key 조정**: PLAN에서는 `prov-lemonade`의 adapter를 "lemonade"로 설정할 것을 요구했으나, `isProviderAdapterInstanceValid`에서 "lemonade"는 unknown adapter key로 excluded됨. Lemonade를 OpenAI-compatible 계열로 분류하므로 adapter를 "oaicompat"으로 변경하여 `OpenAICompatInstances`의 `oaicompat` instance에 매핑함.
+
+3. **Adapter/Target 검증 제거 (MIXED_GROUP-2)**: 기존 테스트 패턴(`TestChatCompletionsProviderPoolOllamaSelectionUsesNormalizedRun`)과 일치하도록 Adapter/Target 검증을 ModelGroupKey 검증으로 대체. fakeRunService는 실제 서비스의 candidate rewrite를 시뮬레이션하지 않으므로, handler-level path selection (tunnel vs normalized)을 검증 surface로 사용.
+
+4. **vLLM-MLX provider 추가 (MIXED_GROUP-1)**: PLAN에서 요구하는 4개 mixed provider (vLLM, vLLM-MLX, Lemonade, Ollama)에 맞게 `prov-vllm-mlx`를 추가하고, `prov-openai-c`를 제거. `VllmInstances`에 `vllm-gpu-mlx` adapter instance를 추가.
+
+5. **responsesProviderTunnelServer 명시화 (MIXED_GROUP-3)**: PLAN에서 언급된 `responsesProviderTunnelServer` helper에도 `poolDispatchPath: string(edgeservice.ProviderPoolPathTunnel)` 명시 추가.

 ## 주요 설계 결정

-_구현 에이전트가 주요 설계 결정 사항을 기록한다._
+1. **mixed provider group fixture 설계**: 한 노드에 vLLM, lemonade, openai_compat, Ollama 4개 provider를 모두 넣고, catalog에 같은 model group으로 등록. 모든 provider가 valid health/capacity/adapter instance를 갖도록 설정하여 `resolveProviderPoolCandidates`가 모든 provider를 candidate로 유지하는지 직접 증명.
+
+2. **fake service selection fixture**: `poolDispatchPath` 필드를 사용하여 tunnel/normalized path 선택을 시뮬레이션. tunnel 선택 시 `tunnelFrames`, normalized 선택 시 `events`를 제공하여 실제 실행 경로를 구분 검증.
+
+3. **기존 helper 명시화**: `chatSidebandServer`, `chatProviderAuthServer`, `chatPassthroughServer`에 `poolDispatchPath: string(edgeservice.ProviderPoolPathTunnel)` 명시 추가. 기존 테스트 동작에는 영향 없음 (fake service의 기본 동작이 tunnel path이므로).

 ## 사용자 리뷰 요청

@ -113,22 +126,30 @@ _구현 에이전트가 각 중간 검증 및 최종 검증 명령 실행 후

 ### MIXED_GROUP-1 중간 검증
 ```bash
-$ GOCACHE=/tmp/iop-go-cache go test ./apps/edge/internal/service -count=1
-(output)
+$ GOCACHE=/tmp/iop-go-cache go test ./apps/edge/internal/service -count=1 -run TestResolveProviderPoolCandidatesKeepsMixedProviderTypes
+ok  	iop/apps/edge/internal/service	0.004s
 ```

 ### MIXED_GROUP-2/3 중간 검증
 ```bash
-$ GOCACHE=/tmp/iop-go-cache go test ./apps/edge/internal/openai -count=1
-(output)
+$ GOCACHE=/tmp/iop-go-cache go test ./apps/edge/internal/openai -count=1 -run "TestChatCompletionsMixedProviderPool|TestChatCompletionsOllamaOnlyProviderPoolUsesNormalizedPath" -v
+=== RUN   TestChatCompletionsMixedProviderPoolTunnelSelection
+--- PASS: TestChatCompletionsMixedProviderPoolTunnelSelection (0.00s)
+=== RUN   TestChatCompletionsMixedProviderPoolOllamaSelection
+--- PASS: TestChatCompletionsMixedProviderPoolOllamaSelection (0.00s)
+=== RUN   TestChatCompletionsOllamaOnlyProviderPoolUsesNormalizedPath
+--- PASS: TestChatCompletionsOllamaOnlyProviderPoolUsesNormalizedPath (0.00s)
+PASS
+ok  	iop/apps/edge/internal/openai	0.008s
 ```

 ### 최종 검증
 ```bash
 $ GOCACHE=/tmp/iop-go-cache go test ./apps/edge/internal/service -count=1
-(output)
+ok  	iop/apps/edge/internal/service	1.035s
+
 $ GOCACHE=/tmp/iop-go-cache go test ./apps/edge/internal/openai -count=1
-(output)
+ok  	iop/apps/edge/internal/openai	11.719s
 ```

 ---
@ -136,3 +157,31 @@ $ GOCACHE=/tmp/iop-go-cache go test ./apps/edge/internal/openai -count=1
 > **[IMPLEMENTING AGENT — BEFORE SAVING] Have you filled in every implementation-owned section: completion table, implementation checklist, changes from plan, design decisions, and verification output?**
 > If anything is blank, go back and fill it in before saving this file.
 > Leave review-agent-only sections unchanged.
+
+## 코드리뷰 결과
+
+- 종합 판정: PASS
+- 차원별 평가:
+  - correctness: Pass
+  - completeness: Pass
+  - test coverage: Pass
+  - API contract: Pass
+  - code quality: Pass
+  - implementation deviation: Pass
+  - verification trust: Pass
+  - spec conformance: Pass
+- 발견된 문제:
+  - Nit: `apps/edge/internal/openai/server_test.go:7823`, `apps/edge/internal/openai/server_test.go:7884`의 normalized fixture 이벤트는 `fakeRunService.SubmitProviderPool` 경로에서 `events`가 아니라 `poolRunFrames`를 읽기 때문에 현재 assertion에는 쓰이지 않는다. 현재 테스트는 path 분기를 충분히 검증하므로 PASS하되, 나중에 response body까지 고정하려면 `poolRunFrames`로 옮기거나 unused setup을 제거한다.
+- 리뷰 중 직접 정리:
+  - `gofmt -w apps/edge/internal/service/model_queue_test.go apps/edge/internal/openai/server_test.go`
+- 검증:
+  - `gofmt -l apps/edge/internal/service/model_queue_test.go apps/edge/internal/openai/server_test.go`: PASS, no output
+  - `git diff --check`: PASS, no output
+  - `GOCACHE=/tmp/iop-go-cache go test ./apps/edge/internal/service -count=1`: PASS, `ok iop/apps/edge/internal/service 1.027s`
+  - `GOCACHE=/tmp/iop-go-cache go test ./apps/edge/internal/openai -count=1`: PASS, `ok iop/apps/edge/internal/openai 11.719s`
+  - `GOCACHE=/tmp/iop-go-cache go test ./apps/edge/...`: PASS, edge packages passed
+- SDD/Evidence 판단:
+  - SDD S05와 Milestone task `mixed-model-group`에 필요한 mixed service candidate retention, selected tunnel path, selected normalized path, Ollama-only normalized path evidence가 확인됨.
+  - 선행 split 작업 `01-provider-path-classifier`, `02-selection-first-provider-dispatch`는 `agent-task/archive/2026/07/m-model-group-mixed-provider-dispatch/*/complete.log`에서 PASS 완료를 확인함.
+- 다음 단계:
+  - PASS 종결: active plan/review를 log로 아카이브하고 `complete.log` 작성 후 task directory를 `agent-task/archive/2026/07/m-model-group-mixed-provider-dispatch/03-mixed-model-group-openai/`로 이동한다.
--- a/agent-task/archive/2026/07/m-model-group-mixed-provider-dispatch/03-mixed-model-group-openai/complete.log
+++ b/agent-task/archive/2026/07/m-model-group-mixed-provider-dispatch/03-mixed-model-group-openai/complete.log
@ -0,0 +1,47 @@
+# Complete - m-model-group-mixed-provider-dispatch/03-mixed-model-group-openai
+
+## 완료 일시
+
+2026-07-12
+
+## 요약
+
+Mixed model group OpenAI fixture task를 1회 리뷰 루프로 완료했다. 최종 판정은 PASS이며, SDD S05에 필요한 service 후보 유지와 OpenAI handler path selection evidence를 보강했다.
+
+## 루프 이력
+
+| Plan | Review | Verdict | 메모 |
+|------|--------|---------|------|
+| `plan_local_G06_0.log` | `code_review_local_G06_0.log` | PASS | mixed provider candidate retention, selected tunnel path, selected normalized path, Ollama-only normalized path fixtures가 통과했다. |
+
+## 구현/정리 내용
+
+- `TestResolveProviderPoolCandidatesKeepsMixedProviderTypes`를 추가해 vLLM, vLLM-MLX, Lemonade, Ollama provider가 같은 model group 후보로 모두 남고 provider type별 execution path가 유지됨을 검증했다.
+- OpenAI Chat provider-pool 테스트에 mixed group tunnel-selected, mixed group Ollama-selected, Ollama-only normalized fixtures를 추가했다.
+- 기존 passthrough/sideband/provider-auth/Responses helper가 selection-first 이후에도 tunnel-selected fixture임을 `poolDispatchPath`로 명시했다.
+- 리뷰 중 `gofmt`를 적용해 변경된 Go 테스트 파일의 formatting drift를 정리했다.
+
+## 최종 검증
+
+- `gofmt -l apps/edge/internal/service/model_queue_test.go apps/edge/internal/openai/server_test.go` - PASS; no output.
+- `git diff --check` - PASS; no output.
+- `GOCACHE=/tmp/iop-go-cache go test ./apps/edge/internal/service -count=1` - PASS; `ok iop/apps/edge/internal/service 1.027s`.
+- `GOCACHE=/tmp/iop-go-cache go test ./apps/edge/internal/openai -count=1` - PASS; `ok iop/apps/edge/internal/openai 11.719s`.
+- `GOCACHE=/tmp/iop-go-cache go test ./apps/edge/...` - PASS; edge packages passed, with `apps/edge/internal/openai` and `apps/edge/internal/service` rerun uncached.
+- repo 내부 edge-node 진단, 보조 E2E smoke, full-cycle 실제 구동 - NOT RUN; 이번 변경은 runtime implementation이 아니라 S05 증거를 고정하는 test-only fixture 변경이다.
+
+## Roadmap Completion
+
+- Milestone: `agent-roadmap/phase/routing-policy-model-orchestration/milestones/model-group-mixed-provider-dispatch.md`
+- Milestone link: [Milestone 문서](agent-roadmap/phase/routing-policy-model-orchestration/milestones/model-group-mixed-provider-dispatch.md)
+- Completed task ids:
+  - `mixed-model-group`: PASS; evidence=`agent-task/archive/2026/07/m-model-group-mixed-provider-dispatch/03-mixed-model-group-openai/plan_local_G06_0.log`, `agent-task/archive/2026/07/m-model-group-mixed-provider-dispatch/03-mixed-model-group-openai/code_review_local_G06_0.log`; verification=`GOCACHE=/tmp/iop-go-cache go test ./apps/edge/internal/service -count=1`, `GOCACHE=/tmp/iop-go-cache go test ./apps/edge/internal/openai -count=1`, `GOCACHE=/tmp/iop-go-cache go test ./apps/edge/...`
+- Not completed task ids: 없음
+
+## 잔여 Nit
+
+- `apps/edge/internal/openai/server_test.go`의 two normalized fixtures currently set `fakeRunService.events`, but `SubmitProviderPool` consumes `poolRunFrames`; current path assertions are valid, and future response-body assertions should either move the fixture frames to `poolRunFrames` or remove the unused setup.
+
+## 후속 작업
+
+- 없음
--- a/agent-task/archive/2026/07/m-model-group-mixed-provider-dispatch/03-mixed-model-group-openai/plan_local_G06_0.log
+++ b/agent-task/archive/2026/07/m-model-group-mixed-provider-dispatch/03-mixed-model-group-openai/plan_local_G06_0.log
@ -127,11 +127,11 @@

 ## 구현 체크리스트

- [ ] 01-provider-path-classifier와 02-selection-first-provider-dispatch PASS/complete 여부를 확인하고, 미완료면 이 plan 구현을 시작하지 않는다.
- [ ] service test에 mixed provider group 후보 retention fixture를 추가해 OpenAI-compatible provider와 Ollama provider가 같은 model group 후보로 모두 남고 각 execution path가 유지됨을 검증한다. 검증: `GOCACHE=/tmp/iop-go-cache go test ./apps/edge/internal/service -count=1`.
- [ ] OpenAI server test에 mixed group과 Ollama-only group fixtures를 추가해 selected OpenAI-compatible provider는 tunnel, selected Ollama provider는 normalized path를 실행함을 검증한다. 검증: `GOCACHE=/tmp/iop-go-cache go test ./apps/edge/internal/openai -count=1`.
- [ ] 기존 provider-pool passthrough tests가 selection-first 이후에도 tunnel-selected provider fixture임을 명확히 하도록 helper/expectation을 정리한다.
- [ ] CODE_REVIEW-*-G??.md의 구현 에이전트 소유 섹션을 실제 구현 내용과 검증 출력으로 채운다. 이 항목이 완료되기 전에는 구현이 완료된 것이 아니다.
+- [x] 01-provider-path-classifier와 02-selection-first-provider-dispatch PASS/complete 여부를 확인하고, 미완료면 이 plan 구현을 시작하지 않는다.
+- [x] service test에 mixed provider group 후보 retention fixture를 추가해 OpenAI-compatible provider와 Ollama provider가 같은 model group 후보로 모두 남고 각 execution path가 유지됨을 검증한다. 검증: `GOCACHE=/tmp/iop-go-cache go test ./apps/edge/internal/service -count=1`. 추가 테스트: `TestResolveProviderPoolCandidatesKeepsMixedProviderTypes`.
+- [x] OpenAI server test에 mixed group과 Ollama-only group fixtures를 추가해 selected OpenAI-compatible provider는 tunnel, selected Ollama provider는 normalized path를 실행함을 검증한다. 검증: `GOCACHE=/tmp/iop-go-cache go test ./apps/edge/internal/openai -count=1`. 추가 테스트: `TestChatCompletionsMixedProviderPoolTunnelSelection`, `TestChatCompletionsMixedProviderPoolOllamaSelection`, `TestChatCompletionsOllamaOnlyProviderPoolUsesNormalizedPath`.
+- [x] 기존 provider-pool passthrough tests가 selection-first 이후에도 tunnel-selected provider fixture임을 명확히 하도록 helper/expectation을 정리한다.
+- [x] CODE_REVIEW-local-G06.md의 구현 에이전트 소유 섹션을 실제 구현 내용과 검증 출력으로 채웠다.

 ## 의존 관계 및 구현 순서

@ -283,16 +283,18 @@ GOCACHE=/tmp/iop-go-cache go test ./apps/edge/internal/openai -count=1
 | 파일 | 항목 |
 | --- | --- |
 | `apps/edge/internal/service/model_queue_test.go` | MIXED_GROUP-1 |
-| `apps/edge/internal/service/run_dispatch_internal_test.go` | MIXED_GROUP-1 |
 | `apps/edge/internal/openai/server_test.go` | MIXED_GROUP-2, MIXED_GROUP-3 |
-| `apps/edge/internal/openai/chat_handler.go` | MIXED_GROUP-2 회귀 대응 필요 시 |
-| `apps/edge/internal/openai/server.go` | MIXED_GROUP-2 fake/interface 대응 필요 시 |
+
+> 참고: `run_dispatch_internal_test.go`, `chat_handler.go`, `server.go`는 변경 필요 없었음.

 ## 최종 검증

 ```bash
-GOCACHE=/tmp/iop-go-cache go test ./apps/edge/internal/service -count=1
-GOCACHE=/tmp/iop-go-cache go test ./apps/edge/internal/openai -count=1
+$ GOCACHE=/tmp/iop-go-cache go test ./apps/edge/internal/service -count=1
+ok  	iop/apps/edge/internal/service	1.035s
+
+$ GOCACHE=/tmp/iop-go-cache go test ./apps/edge/internal/openai -count=1
+ok  	iop/apps/edge/internal/openai	11.719s
 ```

 기대 결과: 모든 명령 PASS.
--- a/agent-test/dev/edge-smoke.md
+++ b/agent-test/dev/edge-smoke.md
@ -51,9 +51,10 @@ dev-runtime provider pool과 3-node 연결 상태를 점검할 때는 `agent-tes
 - bootstrap HTTP: `http://toki-labs.com:18082`
 - Edge OpenAI-compatible base URL: `http://toki-labs.com:18083/v1`
 - Edge-Node TCP transport: `toki-labs.com:18084`
- model alias: `qwen3.6:35b`
- host Pi default profile: `agent-test/dev/inventory.yaml`의 `model.pi_agent_profile` 기준. 현재 기본 provider/model/thinking level은 `iop` / `qwen3.6:35b` / `high`이며, Chat Completions 기본 response mode는 `metadata.iop_response_mode` 생략에 따른 pure `passthrough`다. Endpoint 값은 바뀔 수 있으므로 Pi 최적화/호환성 판단은 `model.pi_agent_profile.runtime_parameter_alignment`의 Qwen3.6 alias, context/max token, reasoning/template/parser, engine별 capacity 매칭을 기준으로 한다.
- provider/model separation: host Pi의 non-default direct provider와 dev-corp `gemma4:26b` profile은 dev Qwen3.6 provider-pool 판정 근거로 섞지 않는다.
+- active model alias: `ornith:35b`
+- Qwen model alias: `qwen3.6:35b`는 현재 `mac-mlx-vllm`에만 남아 있다. `gx10-vllm`과 `onexplayer-lemonade`는 Ornith group provider로 대체되었다.
+- host Pi default profile: `agent-test/dev/inventory.yaml`의 `model.pi_agent_profile` 기준. 현재 기본 provider/model/thinking level은 `iop` / `ornith:35b` / `high`이며, Pi는 IOP Edge `http://toki-labs.com:18083/v1`의 Ornith model group을 호출한다. Pi local direct providers는 제거된 상태다.
+- provider/model separation: dev-corp `gemma4:26b` profile과 stopped DiffusionGemma direct provider는 dev Ornith/Qwen provider-pool 판정 근거로 섞지 않는다.

 노드 후보:

@ -73,24 +74,24 @@ dev-runtime provider pool과 3-node 연결 상태를 점검할 때는 `agent-tes
 - GX10 vLLM node: `gx10-vllm-node` / `gx10-vllm`
  - SSH/user: `ssh toki@192.168.0.91`
  - provider endpoint: `http://192.168.0.91:8001/v1`
-  - served model: `nvidia/Qwen3.6-35B-A3B-NVFP4`
+  - served model: `ornith:35b`
  - capacity baseline: `4`
  - priority baseline: `0`
-  - runtime baseline: `vllm`, `--max-model-len 262144`, `--max-num-seqs 4`, `--gpu-memory-utilization 0.30`, `--reasoning-parser qwen3`, `--default-chat-template-kwargs {"enable_thinking":true}`, `--enable-auto-tool-choice --tool-call-parser qwen3_xml`
-  - long-context admission baseline: `total_context_tokens=262144`, `long_context_capacity=1` (conservative; `max-model-len 262144` until runtime evidence proves a larger concurrent long-context budget)
+  - runtime baseline: `vllm`, model `deepreinforce-ai/Ornith-1.0-35B-FP8`, `--served-model-name ornith:35b`, `--dtype bfloat16`, `--max-model-len 262144`, `--max-num-seqs 4`, `--gpu-memory-utilization 0.50`, `--enable-prefix-caching`, `--enable-auto-tool-choice --tool-call-parser qwen3_xml`, `--reasoning-parser qwen3`, `--trust-remote-code`, `--language-model-only`
+  - long-context admission baseline: `total_context_tokens=1048576`, `long_context_capacity=4` (vLLM log: GPU KV cache `1060912` tokens, max concurrency `4.05x` for 262144-token requests)
  - workspace: `/home/toki/iop-gx10-vllm`
 - OneXPlayer Lemonade node: `onexplayer-lemonade-node` / `onexplayer-lemonade`
  - SSH/user: `ssh r0bin@192.168.0.59`
  - 접속 기준: 현재 작업 호스트에서 직접 SSH
  - provider endpoint: `http://192.168.0.59:13305/v1`
-  - served model: `Qwen3.6-35B-A3B-MTP-GGUF`
+  - served model: `Ornith-1.0-35B-GGUF-llamacpp-tp1-Q5_K_M`
  - capacity baseline: `3`
  - priority baseline: `1`
-  - load baseline: backend `vulkan`, ctx size `524288`, `llamacpp_args="--spec-type none -np 3 -cb -fa on -b 4096 -ub 1024"`, `save_options=true`
-  - long-context admission baseline: `total_context_tokens=524288`, `long_context_capacity=2` (SDD D04 fixed; `ctx_size=524288`, `-np 3` 유지, `ctx_size`를 `524288` 초과로 키우지 않는다)
+  - load baseline: checkpoint `LordNeel/Ornith-1.0-35B-GGUF-llamacpp-tp1:Q5_K_M`, backend `vulkan`, ctx size `524288`, `llamacpp_args="--spec-type none -cb -fa on -b 4096 -ub 1024 --kv-unified"`, `save_options=true`
+  - long-context admission baseline: `total_context_tokens=524288`, `long_context_capacity=2` (`-np` 고정 분할을 제거하고 `--kv-unified`를 사용한다. `/slots`는 auto slots 4개와 slot `n_ctx=262144`를 보고한다)
  - workspace: `C:/Users/r0bin/iop-field`

-OneXPlayer Lemonade는 `Qwen3.6-35B-A3B-MTP-GGUF` artifact를 사용하되 runtime MTP speculative decoding은 켜지 않는다. dev long-context baseline은 `/v1/load`에 `model_name=Qwen3.6-35B-A3B-MTP-GGUF`, `llamacpp_backend=vulkan`, `ctx_size=524288`, `--spec-type none -np 3 -cb -fa on -b 4096 -ub 1024`를 `save_options=true`로 저장한 상태다. llama.cpp backend의 `/slots`에서는 총 ctx `524288`이 slot 3개로 나뉘어 2026-07-04 기준 slot별 `n_ctx=174848`로 보인다.
+OneXPlayer Lemonade는 현재 Ornith Q5 GGUF를 사용하고 runtime speculative decoding은 `--spec-type none`으로 끈다. Qwen을 다시 올릴 때도 fixed `-np 3` 분할은 쓰지 말고 현재 검증된 `--kv-unified` 방식으로 resource를 공유한다.

 OneXPlayer Lemonade Node는 원격 runner나 Edge host에서 다시 SSH하거나 proxy process로 띄우지 않는다. 현재 작업 호스트에서 OneXPlayer Windows host에 `ssh r0bin@192.168.0.59`로 직접 접속한 뒤 generated PowerShell bootstrap을 실행한다.

@ -112,7 +113,7 @@ mac-mlx-vllm provider는 mac-codex-node 소속 resource로, Edge host와 같은
 - registry, service, transport, console, HTTP/A2A 입력 표면을 바꾼 경우 edge-node 메시지 2회 왕복과 command 응답을 확인한다.
 - OpenAI-compatible 경계를 바꾼 경우 dev `18083` 기준 `iop-edge smoke openai` 또는 동등한 `/healthz`, `/v1/models`, `/v1/responses` 확인으로 edge service와 node adapter 경로 수렴을 확인한다.
 - OpenAI-compatible tool-call 경계(`apps/edge/internal/openai/**`의 text tool-call 합성/validation)를 바꾼 경우 dev `18083` `/v1/chat/completions`에 Pi/Cline형 `tools[]` 요청을 non-stream/stream 각각 최소 1회 보내 raw text tool-call boundary smoke를 수행한다. token은 원격 환경 변수에서 주입하고 원문을 명령/로그/보고에 남기지 않는다.
- dev provider-pool tool-call drift를 확인할 때는 `gx10-vllm`, `mac-mlx-vllm`, `onexplayer-lemonade` direct endpoint 각각에 동일한 `tools[]` + `tool_choice:"auto"` Chat Completions 요청을 보내 200과 `tool_calls` 구조를 확인한 뒤, Edge `18083`에도 10개 동시 요청을 보내 10/10 성공, 10/10 `tool_calls`, provider peak `gx10-vllm=4`, `onexplayer-lemonade=3`, `mac-mlx-vllm=2` 사용을 확인한다.
+- dev provider-pool tool-call drift를 확인할 때는 active model group별 provider만 섞는다. Ornith group은 `gx10-vllm`과 `onexplayer-lemonade` direct endpoint에 동일한 `tools[]` + `tool_choice:"auto"` Chat Completions 요청을 보내 200과 `tool_calls` 구조를 확인한 뒤, Edge `18083` model `ornith:35b`에도 동시 요청을 보낸다. Qwen group 검증은 `mac-mlx-vllm` 단독 기준으로 별도 수행한다.
 - bootstrap/artifact 경계를 바꾼 경우 dev artifact/base URL 후보 `18082`가 local/test `18080` field baseline을 덮어쓰지 않는지 확인한다.

 ## 보조 검증
@ -129,7 +130,7 @@ mac-mlx-vllm provider는 mac-codex-node 소속 resource로, Edge host와 같은
 - raw text tool-call boundary smoke: Pi/Cline형 `tools[]` 요청에서 응답 body와 SSE delta 어디에도 `<tool_call>`, `{{`, `<|mask_end|>` 원문이 성공 content로 남지 않는다. 요청 `tools[]`에 있는 valid raw text tool-call은 `message.tool_calls`(또는 stream `delta.tool_calls`)와 `finish_reason: "tool_calls"`로 정규화되고, unknown tool hallucination이나 malformed 블록은 성공 content가 아니라 non-stream `tool_validation_error`(HTTP 502) 또는 SSE `tool_validation_error` 이벤트로 끝난다. non-stream/strict buffered stream은 bounded retry 후 차단을 확인한다. 세부 payload와 계약은 `docs/edge-local-dev-guide.md`의 Raw text tool-call boundary smoke와 `agent-contract/outer/openai-compatible-api.md`를 따른다.
 - raw boundary smoke evidence는 tracked 문서가 아니라 ignored run 위치(`agent-test/runs/**`)나 code-review output path에 저장하고, 저장물에도 token 원문을 남기지 않는다.
 - provider-pool dispatch는 `in_flight >= capacity`인 provider를 후보에서 제외하고, 남은 후보 중 가장 낮은 `in_flight` 레벨을 먼저 선택한다. 같은 `in_flight` 레벨 안에서는 낮은 priority 값과 rotation으로 분산한다.
- dev-runtime capacity smoke는 `/v1/responses`와 `/v1/chat/completions` 각각에 provider capacity 총합 + 1개 동시 요청을 보내고, Control Plane status의 `provider_snapshots`에서 총 `in_flight`가 capacity 총합에 도달하며 `queued`가 1 이상 잡히는지 확인한다. 현재 `gx10-vllm=4`, `onexplayer-lemonade=3`, `mac-mlx-vllm=2`이면 endpoint별 10개 동시 요청에서 총 `in_flight=9`, `queued>=1` 관측을 기준으로 한다. 13개 동시 확장 smoke에서는 총 `in_flight=9`, `queued>=4` 관측을 기대한다.
+- dev-runtime Ornith capacity smoke는 `/v1/responses`와 `/v1/chat/completions` 각각에 Ornith provider capacity 총합 + 1개 동시 요청을 보낸다. 현재 Ornith group은 `gx10-vllm=4`, `onexplayer-lemonade=3`이므로 8개 동시 요청에서 총 `in_flight=7`, `queued>=1`, `mac-mlx-vllm=0` 관측을 기준으로 한다.
 - capacity smoke 완료 후 대상 provider의 `in_flight=0`, `queued=0` 회복을 확인한다.
 - long-context admission 시나리오(normal 10-way, mixed long/normal, all-long-slot-full)와 최종 회복 근거는 `agent-test/dev/long-context-admission-smoke.md`와 `scripts/e2e-long-context-admission-smoke.sh`를 사용한다.
 - 2026-06-24 dev 13-way 확장 smoke는 Mac capacity가 `3`이던 이전 관측이다. 해당 run에서는 `/v1/chat/completions` 13개 요청 성공, peak `in_flight=10`, 실제 queue 대기 `3`개, 최종 회복 `in_flight=0`, `queued=0`이 관측되었다. 현재 Mac capacity `2` 기준 13-way 확장 smoke 기대치는 peak `in_flight=9`, queue 대기 `4`개 이상이다.
--- a/agent-test/dev/inventory.yaml
+++ b/agent-test/dev/inventory.yaml
@ -31,13 +31,119 @@ build:
    node_windows_amd64: build/dev-runtime/bin/iop-node-windows-amd64.exe

 model:
-  alias: qwen3.6:35b
-  provider_capacity_total: 9
-  provider_capacity_status: verified_with_control_plane_provider_snapshots_2026_07_08
+  alias: ornith:35b
+  provider_capacity_total: 7
+  provider_capacity_status: verified_with_control_plane_provider_snapshots_2026_07_12
  context_window: 262144
  default_max_tokens: 32768
-  min_max_tokens: 32768
+  min_max_tokens: 16384
  default_thinking_token_budget: 1024
+  qwen3_6_reference:
+    alias: qwen3.6:35b
+    status: registered_mac_mlx_only_after_ornith_replacement
+    provider_capacity_total: 2
+    provider: mac-mlx-vllm
+    served_model: mlx-community/Qwen3.6-35B-A3B-4bit
+    context_window: 262144
+    default_max_tokens: 32768
+    min_max_tokens: 32768
+  active_edge_model_group:
+    observed_at: "2026-07-12"
+    id: ornith:35b
+    display_name: Ornith 1.0 35B
+    status: active_iop_edge_group
+    openai_base_url_public: http://toki-labs.com:18083/v1
+    context_window: 262144
+    default_max_tokens: 32768
+    min_max_tokens: 16384
+    default_thinking_token_budget: 1024
+    provider_capacity_total: 7
+    qwen_group_update: qwen3.6:35b remains registered with mac-mlx-vllm only; gx10-vllm and onexplayer-lemonade were removed from the Qwen group and now serve this Ornith group.
+    providers:
+      gx10-vllm:
+        runtime_type: vllm
+        endpoint: http://192.168.0.91:8001/v1
+        served_model: ornith:35b
+        upstream_model: deepreinforce-ai/Ornith-1.0-35B-FP8
+        upstream_url: https://huggingface.co/deepreinforce-ai/Ornith-1.0-35B-FP8
+        upstream_revision: 1ab57ce0b44950e498a88756f40ad1ed4d0f30ca
+        quantization: fp8
+        dtype: bfloat16
+        capacity: 4
+        total_context_tokens: 1048576
+        long_context_capacity: 4
+        runtime_args:
+          max_model_len: 262144
+          max_num_seqs: 4
+          gpu_memory_utilization: 0.50
+          enable_prefix_caching: true
+          enable_auto_tool_choice: true
+          tool_call_parser: qwen3_xml
+          reasoning_parser: qwen3
+          trust_remote_code: true
+          language_model_only: true
+        observed_kv:
+          available_kv_cache_memory_gib: 20.79
+          gpu_kv_cache_tokens: 1060912
+          max_concurrency_for_262144_token_requests: 4.05
+      onexplayer-lemonade:
+        runtime_type: lemonade
+        endpoint: http://192.168.0.59:13305/v1
+        served_model: Ornith-1.0-35B-GGUF-llamacpp-tp1-Q5_K_M
+        response_model: ornith-1.0-35b-Q5_K_M.gguf
+        upstream_model: LordNeel/Ornith-1.0-35B-GGUF-llamacpp-tp1
+        upstream_url: https://huggingface.co/LordNeel/Ornith-1.0-35B-GGUF-llamacpp-tp1
+        upstream_revision: c50d5d4407f70e43208dee836c66bb8a05c1be91
+        quantization: Q5_K_M
+        capacity: 3
+        total_context_tokens: 524288
+        long_context_capacity: 2
+        runtime_args:
+          ctx_size: 524288
+          llamacpp_backend: vulkan
+          llamacpp_args: "--spec-type none -cb -fa on -b 4096 -ub 1024 --kv-unified"
+          fixed_parallel_slots_removed: true
+          observed_total_slots: 4
+          observed_slot_n_ctx: 262144
+    smoke:
+      models_endpoint: passed
+      chat_completions_single:
+        endpoint: /v1/chat/completions
+        request_model: ornith:35b
+        prompt: "2+2? Answer only the number."
+        result: passed
+        content: "4"
+        elapsed_sec: 3.32
+        completion_tokens: 127
+      responses_single:
+        endpoint: /v1/responses
+        request_model: ornith:35b
+        prompt: "2+2? Answer only the number."
+        result: passed
+        output_text: "4"
+        elapsed_sec: 4.58
+        completion_tokens: 1
+      chat_completions_capacity:
+        endpoint: /v1/chat/completions
+        concurrent_requests: 8
+        success_count: 8
+        failure_count: 0
+        wall_sec: 63.97
+        peak_total_in_flight: 7
+        peak_total_queued: 2
+        provider_peaks:
+          gx10-vllm:
+            capacity: 4
+            max_in_flight: 4
+            max_queued: 1
+          onexplayer-lemonade:
+            capacity: 3
+            max_in_flight: 3
+            max_queued: 1
+          mac-mlx-vllm:
+            capacity: 2
+            max_in_flight: 0
+            max_queued: 0
  reasoning_policy: bounded_thinking
  agent_tooling_policy:
    runtime_profile: qwen_specific
@ -53,27 +159,49 @@ model:
      tool_call_parser_status: validate tools[] non-stream and streaming responses before enabling this provider for Pi/Cline-style agent tool use
    separation_note: do not copy dev-corp Gemma parser/template values such as tool_call_parser=gemma4, reasoning_parser=gemma4, or Gemma4 chat templates into Qwen providers
  pi_agent_profile:
-    observed_at: "2026-07-09"
+    observed_at: "2026-07-12"
    config_scope: host_local_pi_agent
    config_dir: /config/.pi/agent
    current_default_provider: iop
-    current_default_model: qwen3.6:35b
+    current_default_model: ornith:35b
    current_default_thinking_level: high
    current_default_api: openai-completions
    current_default_base_url: http://toki-labs.com:18083/v1
    current_default_auth_header: true
    api_key_value_tracked: false
-    endpoint_policy: current_default_base_url is the observed route for this host Pi profile, but optimization compatibility is determined by the Qwen3.6 model alias, context/output limits, reasoning policy, and provider runtime settings rather than by the endpoint literal alone.
+    endpoint_policy: current default Pi route is IOP Edge provider iop with model ornith:35b. Direct providers were removed from /config/.pi/agent/models.json so Pi no longer bypasses Edge to GX10 vLLM, OneXPlayer Lemonade, or the stopped DiffusionGemma endpoint.
+    local_pi_providers_at_observation:
+      - iop
+    direct_providers_removed_at_observation:
+      - iop-gemma
+      - ornith-direct
+      - ornith-lemonade
+    iop_provider_models_at_observation:
+      - qwen3.6:35b
+      - ornith:35b
    pi_model_parameters:
      provider: iop
-      model_id: qwen3.6:35b
-      display_name: IOP Qwen 3.6 35B
+      model_id: ornith:35b
+      display_name: IOP Ornith 1.0 35B
      api: openai-completions
      context_window: 262144
      max_tokens: 32768
      reasoning: true
      thinking_level: high
      expected_iop_thinking_token_budget: 1024
+    current_ornith_parameter_alignment:
+      model_alias: ornith:35b
+      pi_context_window: 262144
+      edge_context_window: 262144
+      pi_max_tokens: 32768
+      edge_model_group_registered: true
+      edge_group_min_max_tokens: 16384
+      gx10_edge_capacity: 4
+      onexplayer_edge_capacity: 3
+      reasoning_parser: qwen3
+      tool_call_parser: qwen3_xml
+      fp8_required_on_gx10: true
+      note: Official 4-bit GGUF was not retained for GX10 vLLM because semantic generation smoke failed; the working GX10 runtime uses the official FP8 Hugging Face release.
    runtime_parameter_alignment:
      alignment_policy: Pi model parameters must be kept in sync with the Edge model catalog and each provider runtime profile. Do not optimize or debug Pi behavior from the endpoint URL alone; verify the alias-to-served-model mapping, context window, max token budget, thinking/reasoning parser, chat template kwargs, and provider capacity/sequence limits together.
      common_requirements:
@ -108,11 +236,21 @@ model:
        onexplayer-lemonade:
          runtime_type: lemonade
          served_model: Qwen3.6-35B-A3B-MTP-GGUF
-          context_match: pi contextWindow 262144 is admitted within Lemonade ctx_size 524288 and current slot n_ctx 174848 x 3; long_context_capacity remains 2
+          current_runtime_override_2026_07_12: OneXPlayer Lemonade was switched to Ornith-1.0-35B-GGUF-llamacpp-tp1-Q5_K_M for direct Ornith validation; it is not a valid Qwen provider-pool member until Edge provider config is updated.
+          context_match: Qwen should use the same Lemonade unified KV profile when re-enabled; keep total ctx_size 524288, omit -np, and pass --kv-unified to avoid the prior fixed 174848-token slot partitioning.
          output_budget_match: pi maxTokens 32768 maps to load default_max_tokens 32768
          thinking_match: pi reasoning=true and thinking_level=high map to IOP bounded thinking budget 1024; Lemonade load keeps MTP disabled and tool_choice_auto_verified=true
-          concurrency_match: provider capacity 3 maps to Lemonade -np 3
+          concurrency_match: provider capacity remains capped at 3 for IOP admission; with unified KV, Lemonade auto slots may report 4 runtime slots, so Edge admission must remain the concurrency guard.
          tool_parser_profile: lemonade_qwen_tool_choice_auto_verified
+          qwen_reenable_profile:
+            status: applicable_not_reloaded_while_ornith_q5_is_active
+            ctx_size: 524288
+            llamacpp_args: "--spec-type none -cb -fa on -b 4096 -ub 1024 --kv-unified"
+            removed_args: "-np 3"
+            expected_slot_n_ctx: 262144
+            expected_capacity: 3
+            expected_long_context_capacity: 2
+            validation_basis: Same Lemonade llama.cpp backend and current Ornith Q5 runtime validated --kv-unified with auto slots reporting n_ctx 262144; Qwen still needs smoke validation after it is reloaded.
      separation_note: These alignment rules are for the dev Qwen3.6 provider pool only. Gemma4 and DiffusionGemma use separate provider/model profiles and must not inherit Qwen parser/template/context assumptions.
    response_mode:
      default: pure_passthrough
@ -146,7 +284,7 @@ model:
          max_queued: 1
    non_default_local_provider:
      provider: iop-gemma
-      status: separate_direct_provider_not_part_of_dev_qwen_provider_pool
+      status: removed_from_pi_models_json_2026_07_12
      api: openai-completions
      base_url: http://192.168.0.91:8002/v1
      served_model: nvidia/diffusiongemma-26B-A4B-it-NVFP4
@ -155,22 +293,22 @@ model:
    spark_ornith_direct_profile:
      observed_at: "2026-07-12"
      provider: ornith-direct
-      status: active_direct_pi_provider_not_part_of_iop_provider_pool
+      status: removed_from_pi_models_json_2026_07_12_runtime_still_active_as_gx10_iop_ornith_provider
      family: spark_ornith
      api: openai-completions
-      base_url: http://192.168.0.91:8002/v1
+      base_url: http://192.168.0.91:8001/v1
      model_id: ornith:35b
      display_name: Ornith 1.0 35B FP8 Direct vLLM
      context_window: 262144
      max_tokens: 32768
      reasoning: true
-      endpoint_policy: Pi connects directly to the GX10/Spark vLLM OpenAI-compatible endpoint; this profile bypasses IOP Edge routing and does not participate in the dev Qwen provider pool.
+      endpoint_policy: Pi no longer connects directly to this GX10/Spark vLLM endpoint. The same runtime remains registered through IOP Edge under ornith:35b.
      api_key_value_tracked: false
      pi_current_default_at_observation:
-        default_provider: ornith-direct
+        default_provider: iop
        default_model: ornith:35b
        default_api: openai-completions
-        default_base_url: http://192.168.0.91:8002/v1
+        default_base_url: http://toki-labs.com:18083/v1
      qwen_parameter_alignment:
        pi_context_window: 262144
        pi_max_tokens: 32768
@ -192,7 +330,7 @@ model:
        cuda_version_env: 13.0.2
        gpu: NVIDIA GB10
        nvidia_driver_version: 580.126.09
-        endpoint: http://192.168.0.91:8002/v1
+        endpoint: http://192.168.0.91:8001/v1
        hf_model: deepreinforce-ai/Ornith-1.0-35B-FP8
        hf_url: https://huggingface.co/deepreinforce-ai/Ornith-1.0-35B-FP8
        hf_revision: 1ab57ce0b44950e498a88756f40ad1ed4d0f30ca
@ -211,7 +349,7 @@ model:
        quantization_config_source: model config.json compressed-tensors W8A8 FP8, weight channel strategy, input activation token strategy, linear_attn/embed/lm_head/shared gates ignored
        dtype: bfloat16
        max_model_len: 262144
-        max_num_seqs: 2
+        max_num_seqs: 4
        max_num_batched_tokens: vllm_default_not_overridden
        gpu_memory_utilization: 0.50
        enable_prefix_caching: true
@ -244,12 +382,12 @@ model:
            - 248044
          bos_token_id: 248044
        docker_runtime:
-          created_at: "2026-07-12T04:40:26.013681328Z"
+          created_at: "2026-07-12T08:41:04Z"
          entrypoint:
            - vllm
            - serve
          container_port: 8000
-          host_port: 8002
+          host_port: 8001
          network_mode: bridge
          ipc_mode: private
          shm_size_bytes: 67108864
@ -259,13 +397,13 @@ model:
            /home/toki/Data/models: /models
            /home/toki/Data/vllm/templates: /templates
        model_load_memory_gib: 34.43
-        model_load_seconds: 611.880026
-        weight_download_seconds: 407.868734
+        model_load_seconds: 196.929359
+        weight_download_seconds: not_observed_cache_hit_restart
        cuda_graph_memory_estimated_gib: 0.13
        cuda_graph_pool_memory_actual_gib: 0.07
-        available_kv_cache_memory_gib: 21.33
-        gpu_kv_cache_tokens: 1087640
-        max_concurrency_for_262144_token_requests: 4.15
+        available_kv_cache_memory_gib: 20.79
+        gpu_kv_cache_tokens: 1060912
+        max_concurrency_for_262144_token_requests: 4.05
        startup_warnings:
          - vLLM command does not pin --revision; the recorded hf_revision is the downloaded snapshot observed in cache.
          - Prefix caching in Mamba cache align mode is experimental for this architecture.
@ -292,6 +430,40 @@ model:
            - 30.62
            - 29.15
          note: vLLM logs showed aggregate generation throughput around 50-61 tok/s with two running requests.
+    ornith_lemonade_direct_profile:
+      observed_at: "2026-07-12"
+      provider: ornith-lemonade
+      status: removed_from_pi_models_json_2026_07_12_runtime_still_active_as_onexplayer_iop_ornith_provider
+      family: spark_ornith
+      api: openai-completions
+      base_url: http://192.168.0.59:13305/v1
+      model_id: Ornith-1.0-35B-GGUF-llamacpp-tp1-Q5_K_M
+      display_name: Ornith 1.0 35B Q5 Direct Lemonade
+      context_window: 262144
+      max_tokens: 32768
+      reasoning: true
+      config_file: /config/.pi/agent/models.json
+      default_profile_changed: true
+      pi_default_provider_after_add: iop
+      pi_default_model_after_add: ornith:35b
+      endpoint_policy: Pi no longer connects directly to the OneXPlayer Lemonade OpenAI-compatible endpoint. This runtime remains reachable through IOP Edge under ornith:35b.
+      api_key_value_tracked: false
+      runtime_alignment:
+        lemonade_runtime_model: Ornith-1.0-35B-GGUF-llamacpp-tp1-Q5_K_M
+        lemonade_response_model: ornith-1.0-35b-Q5_K_M.gguf
+        quantization: Q5_K_M
+        ctx_size: 524288
+        kv_policy: unified_kv_with_np_omitted
+        observed_slot_n_ctx: 262144
+        iop_capacity_guard: 3
+        long_context_capacity: 2
+      smoke:
+        endpoint_direct_chat_completions: passed
+        request_model: Ornith-1.0-35B-GGUF-llamacpp-tp1-Q5_K_M
+        response_model: ornith-1.0-35b-Q5_K_M.gguf
+        finish_reason: stop
+        content: "4"
+        completion_tokens: 2
    separation_note: This host Pi default profile is the dev Qwen3.6 provider-pool profile. Keep dev-corp Gemma4 under agent-test/dev-corp/inventory.yaml and do not copy Gemma4 parser/template/runtime settings into this Qwen profile.
  content_completion_policy:
    enforce_min_max_tokens: true
@ -315,7 +487,7 @@ model:
    prompt_policy: long_reasoning_allowed
    exact_output_match: false
  latest_provider_snapshot:
-    observed_at: "2026-07-08"
+    observed_at: "2026-07-12"
    total_capacity: 9
    all_nodes_connected: true
    providers:
@ -325,7 +497,7 @@ model:
        in_flight: 0
        queued: 0
        health: healthy
-        served_model: nvidia/Qwen3.6-35B-A3B-NVFP4
+        served_model: ornith:35b
      mac-mlx-vllm:
        capacity: 2
        priority: 2
@ -339,7 +511,7 @@ model:
        in_flight: 0
        queued: 0
        health: healthy
-        served_model: Qwen3.6-35B-A3B-MTP-GGUF
+        served_model: Ornith-1.0-35B-GGUF-llamacpp-tp1-Q5_K_M
  openai_tool_call_smoke:
    observed_at: "2026-07-08"
    request_shape: tools_plus_tool_choice_auto
@ -483,36 +655,38 @@ nodes:
      id: gx10-vllm
      type: vllm
      endpoint: http://192.168.0.91:8001/v1
-      served_model: nvidia/Qwen3.6-35B-A3B-NVFP4
+      served_model: ornith:35b
      capacity: 4
      priority: 0
      # Long-context admission policy (maps to edge.yaml nodes[].providers[]).
-      # max_model_len is 262144; conservative long_context_capacity=1 until runtime
-      # evidence proves a larger concurrent long-context KV budget.
-      total_context_tokens: 262144
-      long_context_capacity: 1
+      # Ornith FP8 runtime reports 1,060,912 GPU KV tokens and 4.05x maximum
+      # concurrency for full 262144-token requests, so four long slots are allowed.
+      total_context_tokens: 1048576
+      long_context_capacity: 4
      runtime:
-        container_name: iop-vllm-qwen36-prev-20260704063128
+        container_name: iop-vllm-ornith35b-fp8
        max_model_len: 262144
        max_num_seqs: 4
-        gpu_memory_utilization: 0.30
+        gpu_memory_utilization: 0.50
        default_max_tokens: 32768
        default_thinking_token_budget: 1024
        enable_auto_tool_choice: true
        tool_call_parser: qwen3_xml
        reasoning_parser: qwen3
-        default_chat_template_kwargs:
-          enable_thinking: true
+        trust_remote_code: true
+        language_model_only: true
+        quantization: fp8
+        dtype: bfloat16
    direct_providers:
      - id: ornith-direct
        family: spark_ornith
        type: vllm
-        endpoint: http://192.168.0.91:8002/v1
+        endpoint: http://192.168.0.91:8001/v1
        served_model: ornith:35b
        upstream_model: deepreinforce-ai/Ornith-1.0-35B-FP8
        upstream_url: https://huggingface.co/deepreinforce-ai/Ornith-1.0-35B-FP8
        upstream_revision: 1ab57ce0b44950e498a88756f40ad1ed4d0f30ca
-        status: active_direct_pi_provider_not_iop_routed
+        status: direct_pi_provider_removed_runtime_shares_current_iop_ornith_gx10_provider
        pi_model_parameters:
          context_window: 262144
          max_tokens: 32768
@ -523,14 +697,14 @@ nodes:
          docker_image_id: sha256:a720df3e84a89d7db47a3b7a0511cb5b312e203fc4956f7493df248299267a6f
          vllm_version: 0.23.1rc1.dev223+ga346d589f
          vllm_build_commit: a346d589f5932d4234bf5bf8718f10e26d187021
-          port_mapping: 0.0.0.0:8002->8000
+          port_mapping: 0.0.0.0:8001->8000
          hf_home: /models/.cache/huggingface
          hf_cache_snapshot_path: /models/.cache/huggingface/hub/models--deepreinforce-ai--Ornith-1.0-35B-FP8/snapshots/1ab57ce0b44950e498a88756f40ad1ed4d0f30ca
          quantization: fp8
          vllm_quantization_backend: compressed-tensors
          dtype: bfloat16
          max_model_len: 262144
-          max_num_seqs: 2
+          max_num_seqs: 4
          max_num_batched_tokens: vllm_default_not_overridden
          gpu_memory_utilization: 0.50
          enable_prefix_caching: true
@ -566,9 +740,9 @@ nodes:
              /home/toki/Data/models: /models
              /home/toki/Data/vllm/templates: /templates
          model_load_memory_gib: 34.43
-          available_kv_cache_memory_gib: 21.33
-          gpu_kv_cache_tokens: 1087640
-          max_concurrency_for_262144_token_requests: 4.15
+          available_kv_cache_memory_gib: 20.79
+          gpu_kv_cache_tokens: 1060912
+          max_concurrency_for_262144_token_requests: 4.05
          average_completion_tok_s: 29.8
        replacement_note: DiffusionGemma direct vLLM on this endpoint was stopped; Ornith is served from the official FP8 Hugging Face release, not the 4-bit GGUF build.
  - id: onexplayer-lemonade-node
@ -582,24 +756,98 @@ nodes:
      id: onexplayer-lemonade
      type: lemonade
      endpoint: http://192.168.0.59:13305/v1
-      served_model: Qwen3.6-35B-A3B-MTP-GGUF
+      served_model: Ornith-1.0-35B-GGUF-llamacpp-tp1-Q5_K_M
+      served_model_response: ornith-1.0-35b-Q5_K_M.gguf
+      previous_served_model: Qwen3.6-35B-A3B-MTP-GGUF
+      status: active_ornith_q5_current_iop_ornith_provider_pool_member
+      qwen_reenable_profile:
+        status: not_loaded_ornith_q5_active
+        model_name: Qwen3.6-35B-A3B-MTP-GGUF
+        checkpoint: unsloth/Qwen3.6-35B-A3B-MTP-GGUF:Qwen3.6-35B-A3B-UD-Q4_K_XL.gguf
+        prior_saved_recipe_options:
+          ctx_size: 524288
+          llamacpp_args: "--spec-type none -np 3 -cb -fa on -b 4096 -ub 1024"
+          partition_note: The prior -np 3 profile fixed slot windows at about 174848 tokens for ctx_size 524288.
+        recommended_recipe_options:
+          ctx_size: 524288
+          llamacpp_args: "--spec-type none -cb -fa on -b 4096 -ub 1024 --kv-unified"
+          removed_args: "-np 3"
+          expected_slot_n_ctx: 262144
+          expected_auto_slots: 4
+          iop_capacity_guard: 3
+          long_context_capacity: 2
+          validation_required_after_load: true
+          validation_basis: Same Lemonade llama.cpp backend and current Ornith Q5 runtime confirmed --kv-unified with omitted -np prevents fixed ctx_size/parallel partitioning.
      capacity: 3
      priority: 1
      # Long-context admission policy (maps to edge.yaml nodes[].providers[]).
-      # SDD D04 fixes OneXPlayer dev baseline at ctx_size=524288, -np 3,
-      # long_context_capacity=2. Do NOT raise ctx_size above 524288.
+      # Ornith Q5 direct runtime keeps ctx_size=524288 and removes fixed -np
+      # partitioning by using llama.cpp unified KV. Do NOT raise ctx_size above 524288.
      total_context_tokens: 524288
      long_context_capacity: 2
      load:
        endpoint: http://192.168.0.59:13305/v1/load
-        model_name: Qwen3.6-35B-A3B-MTP-GGUF
+        model_name: Ornith-1.0-35B-GGUF-llamacpp-tp1-Q5_K_M
+        checkpoint: LordNeel/Ornith-1.0-35B-GGUF-llamacpp-tp1:Q5_K_M
+        hf_url: https://huggingface.co/LordNeel/Ornith-1.0-35B-GGUF-llamacpp-tp1
+        hf_revision: c50d5d4407f70e43208dee836c66bb8a05c1be91
+        hf_ref: main
+        gguf_file: ornith-1.0-35b-Q5_K_M.gguf
+        gguf_file_size_bytes: 24729130848
+        cache_snapshot_path: C:/Users/r0bin/.cache/huggingface/hub/models--LordNeel--Ornith-1.0-35B-GGUF-llamacpp-tp1/snapshots/c50d5d4407f70e43208dee836c66bb8a05c1be91
        backend: vulkan
        ctx_size: 524288
-        llamacpp_args: "--spec-type none -np 3 -cb -fa on -b 4096 -ub 1024"
-        observed_slot_n_ctx: 174848
+        llamacpp_args: "--spec-type none -cb -fa on -b 4096 -ub 1024 --kv-unified"
+        observed_process_args: "--ctx-size 524288 --port 8001 --jinja --context-shift --keep 16 --reasoning-format auto --no-webui --no-mmap -ngl 99 --kv-unified --spec-type none -b 4096 -cb -fa on -ub 1024"
+        observed_total_slots: 4
+        observed_total_ctx_size: 524288
+        observed_slot_n_ctx: 262144
+        context_per_slot_note: -np is intentionally omitted; llama.cpp auto slots with --kv-unified avoid the prior fixed 524288/3 partition and report 262144-token slot windows backed by shared KV.
+        residency:
+          observed_at: "2026-07-12"
+          process_resident_policy: keep llama-server loaded until explicit lemonade unload or service/process restart
+          observed_process_running: true
+          observed_is_sleeping: false
+          idle_sleep_seconds: not_set_llama_cpp_default_minus_1_disabled
+          mmap_policy: disabled_via_no_mmap
+          memory_lock_policy: not_enabled_no_mlock_argument
+          note: This keeps the model warm in the active llama-server process, but it is not an OS-level hard memory pin; Windows can still page memory under pressure because --mlock is not set.
        save_options: true
-        mtp_runtime_policy: disabled
+        mtp_runtime_policy: disabled_via_spec_type_none
        default_max_tokens: 32768
        default_thinking_token_budget: 1024
        tool_choice_auto_verified: true
+        smoke:
+          observed_at: "2026-07-12"
+          qwen_unloaded: true
+          load_status: passed
+          english_final_answer:
+            max_tokens: 512
+            result: passed
+            content: 2+2 equals 4.
+            completion_tokens: 414
+            generation_tok_s: 31.94
+          korean_final_answer:
+            max_tokens: 1024
+            result: passed
+            content: 2 더하기 2는 4입니다.
+            completion_tokens: 384
+          tool_choice_auto:
+            result: passed
+            finish_reason: tool_calls
+            tool_calls: 1
+            function_name: get_weather
+            arguments: "{\"city\":\"Seoul\"}"
+          unified_kv_concurrency:
+            observed_at: "2026-07-12"
+            result: passed
+            concurrent_requests: 3
+            finish_reason_stop: 3
+            final_content: "4"
+            total_elapsed_sec: 6.55
+            completion_tokens:
+              - 149
+              - 88
+              - 85
+          low_output_budget_note: max_tokens 128 and one Korean 512-token probe ended in reasoning_content with empty final content; keep Pi/Edge maxTokens at 32768 or use at least 1024 for short smoke prompts.
    windows_process_start: Win32_Process.Create
--- a/agent-test/dev/node-smoke.md
+++ b/agent-test/dev/node-smoke.md
@ -60,24 +60,24 @@ dev-runtime의 실제 3-node 연결을 점검할 때는 원격 runner `ssh toki@
 - GX10 vLLM node: `gx10-vllm-node` / `gx10-vllm`
  - SSH/user: `ssh toki@192.168.0.91`
  - provider endpoint: `http://192.168.0.91:8001/v1`
-  - served model: `nvidia/Qwen3.6-35B-A3B-NVFP4`
+  - served model: `ornith:35b`
  - capacity baseline: `4`
  - priority baseline: `0`
-  - runtime baseline: `vllm`, `--max-model-len 262144`, `--max-num-seqs 4`, `--gpu-memory-utilization 0.30`, `--reasoning-parser qwen3`, `--default-chat-template-kwargs {"enable_thinking":true}`, `--enable-auto-tool-choice --tool-call-parser qwen3_xml`
-  - long-context admission baseline: `total_context_tokens=262144`, `long_context_capacity=1` (conservative; `max-model-len 262144` until runtime evidence proves a larger concurrent long-context budget)
+  - runtime baseline: `vllm`, model `deepreinforce-ai/Ornith-1.0-35B-FP8`, `--served-model-name ornith:35b`, `--dtype bfloat16`, `--max-model-len 262144`, `--max-num-seqs 4`, `--gpu-memory-utilization 0.50`, `--enable-prefix-caching`, `--enable-auto-tool-choice --tool-call-parser qwen3_xml`, `--reasoning-parser qwen3`, `--trust-remote-code`, `--language-model-only`
+  - long-context admission baseline: `total_context_tokens=1048576`, `long_context_capacity=4` (vLLM log: GPU KV cache `1060912` tokens, max concurrency `4.05x` for 262144-token requests)
  - workspace: `/home/toki/iop-gx10-vllm`
 - OneXPlayer Lemonade node: `onexplayer-lemonade-node` / `onexplayer-lemonade`
  - SSH/user: `ssh r0bin@192.168.0.59`
  - 접속 기준: 현재 작업 호스트에서 직접 SSH
  - provider endpoint: `http://192.168.0.59:13305/v1`
-  - served model: `Qwen3.6-35B-A3B-MTP-GGUF`
+  - served model: `Ornith-1.0-35B-GGUF-llamacpp-tp1-Q5_K_M`
  - capacity baseline: `3`
  - priority baseline: `1`
-  - load baseline: backend `vulkan`, ctx size `524288`, `llamacpp_args="--spec-type none -np 3 -cb -fa on -b 4096 -ub 1024"`, `save_options=true`
-  - long-context admission baseline: `total_context_tokens=524288`, `long_context_capacity=2` (SDD D04 fixed; `ctx_size=524288`, `-np 3` 유지, `ctx_size`를 `524288` 초과로 키우지 않는다)
+  - load baseline: checkpoint `LordNeel/Ornith-1.0-35B-GGUF-llamacpp-tp1:Q5_K_M`, backend `vulkan`, ctx size `524288`, `llamacpp_args="--spec-type none -cb -fa on -b 4096 -ub 1024 --kv-unified"`, `save_options=true`
+  - long-context admission baseline: `total_context_tokens=524288`, `long_context_capacity=2` (`-np` 고정 분할을 제거하고 `--kv-unified`를 사용한다. `/slots`는 auto slots 4개와 slot `n_ctx=262144`를 보고한다)
  - workspace: `C:/Users/r0bin/iop-field`

-OneXPlayer Lemonade는 `Qwen3.6-35B-A3B-MTP-GGUF` artifact를 사용하되 runtime MTP speculative decoding은 끈 상태를 dev 기준으로 삼는다. Node 검증 전 `/v1/load`의 `recipe_options`가 `model_name=Qwen3.6-35B-A3B-MTP-GGUF`, Vulkan, `ctx_size=524288`, `--spec-type none -np 3 -cb -fa on -b 4096 -ub 1024`를 포함하는지 확인한다. backend `/slots`에서는 총 ctx `524288`이 slot 3개로 나뉘어 2026-07-04 기준 slot별 `n_ctx=174848`로 보인다.
+OneXPlayer Lemonade는 현재 Ornith Q5 GGUF를 사용하고 runtime speculative decoding은 `--spec-type none`으로 끈다. Qwen을 다시 올릴 때도 fixed `-np 3` 분할은 쓰지 말고 현재 검증된 `--kv-unified` 방식으로 resource를 공유한다.

 GX10은 Linux/ARM64 bootstrap, OneXPlayer는 Windows native PowerShell bootstrap을 기본으로 한다. OneXPlayer는 현재 작업 호스트에서 직접 접속해 세팅하며, 원격 runner나 Edge host에서 `node-onexplayer-lemonade.yaml`로 proxy 실행하지 않는다.

--- a/apps/edge/internal/openai/server_test.go
+++ b/apps/edge/internal/openai/server_test.go
@ -4584,7 +4584,10 @@ func TestChatCompletionsResponseModeRouting(t *testing.T) {
 // chatSidebandServer returns an openai.Server backed by the given tunnel
 // frames for explicit passthrough+sideband fixtures.
 func chatSidebandServer(frames chan *iop.ProviderTunnelFrame) (*Server, *fakeRunService) {
-	fake := &fakeRunService{tunnelFrames: frames}
+	fake := &fakeRunService{
+		poolDispatchPath: string(edgeservice.ProviderPoolPathTunnel),
+		tunnelFrames:     frames,
+	}
 	srv := NewServer(config.EdgeOpenAIConf{}, fake, nil)
 	srv.SetModelCatalog([]config.ModelCatalogEntry{
 		{ID: "pool-model", Providers: map[string]string{"prov-1": "served-model"}},
@ -4829,7 +4832,10 @@ func TestChatCompletionsPassthroughDoesNotExposeSideband(t *testing.T) {
 // provider auth forwarding config, mirroring chatSidebandServer's catalog so
 // pool-model routes over the provider tunnel path.
 func chatProviderAuthServer(frames chan *iop.ProviderTunnelFrame, auth config.EdgeOpenAIProviderAuthConf) (*Server, *fakeRunService) {
-	fake := &fakeRunService{tunnelFrames: frames}
+	fake := &fakeRunService{
+		poolDispatchPath: string(edgeservice.ProviderPoolPathTunnel),
+		tunnelFrames:     frames,
+	}
 	srv := NewServer(config.EdgeOpenAIConf{ProviderAuth: auth}, fake, nil)
 	srv.SetModelCatalog([]config.ModelCatalogEntry{
 		{ID: "pool-model", Providers: map[string]string{"prov-1": "served-model"}},
@ -5111,6 +5117,7 @@ func TestChatCompletionsPassthroughDoesNotLabelTransformed(t *testing.T) {
 // tunnel to the given provider fixture, emulating the Node relay.
 func chatPassthroughServer(providerURL string) (*Server, *fakeRunService) {
 	fake := &fakeRunService{
+		poolDispatchPath:   string(edgeservice.ProviderPoolPathTunnel),
 		tunnelProviderURL:  providerURL,
 		tunnelServedTarget: "served-model",
 	}
@ -6095,7 +6102,11 @@ func TestChatCompletionsLegacyRouteSetsProviderPoolFalse(t *testing.T) {
 // "pool-model" through the OpenAI-compatible provider tunnel, wired to the given
 // provider frames and served target.
 func responsesProviderTunnelServer(frames chan *iop.ProviderTunnelFrame, servedTarget string) (*Server, *fakeRunService) {
-	fake := &fakeRunService{tunnelFrames: frames, tunnelServedTarget: servedTarget}
+	fake := &fakeRunService{
+		poolDispatchPath:   string(edgeservice.ProviderPoolPathTunnel),
+		tunnelFrames:       frames,
+		tunnelServedTarget: servedTarget,
+	}
 	srv := NewServer(config.EdgeOpenAIConf{}, fake, nil)
 	srv.SetModelCatalog([]config.ModelCatalogEntry{
 		{ID: "pool-model", Providers: map[string]string{"prov-1": "served-model"}},
@ -7742,3 +7753,183 @@ func TestChatCompletionsProviderPoolToolValidationRetryRejectsTunnelResult(t *te
 		t.Errorf("error message: got %q", errMap["message"])
 	}
 }
+
+// --- MIXED_GROUP-2: Mixed/Ollama-Only Provider Pool Fixtures ---
+
+// TestChatCompletionsMixedProviderPoolTunnelSelection verifies S05: a mixed
+// model group containing both OpenAI-compatible and Ollama providers dispatches
+// the selected tunnel-capable provider over the provider tunnel (passthrough)
+// path. Only one tunnel request is sent, no normalized SubmitRun is called.
+func TestChatCompletionsMixedProviderPoolTunnelSelection(t *testing.T) {
+	// Tunnel frames for the tunnel path.
+	tunnelFrames := make(chan *iop.ProviderTunnelFrame, 3)
+	tunnelFrames <- &iop.ProviderTunnelFrame{Kind: iop.ProviderTunnelFrameKind_PROVIDER_TUNNEL_FRAME_KIND_RESPONSE_START, StatusCode: 200}
+	tunnelFrames <- &iop.ProviderTunnelFrame{Kind: iop.ProviderTunnelFrameKind_PROVIDER_TUNNEL_FRAME_KIND_BODY, Body: []byte(`{"id":"cmpl-mixed","choices":[{"index":0,"message":{"role":"assistant","content":"tunnel"}}],"finish_reason":"stop"}`)}
+	tunnelFrames <- &iop.ProviderTunnelFrame{Kind: iop.ProviderTunnelFrameKind_PROVIDER_TUNNEL_FRAME_KIND_END, End: true}
+	close(tunnelFrames)
+
+	// Fake service selects tunnel path (simulating vLLM provider selection).
+	fake := &fakeRunService{
+		poolDispatchPath: string(edgeservice.ProviderPoolPathTunnel),
+		tunnelFrames:     tunnelFrames,
+	}
+
+	// Mixed catalog: OpenAI-compatible (vLLM) + Ollama in same group.
+	catalog := []config.ModelCatalogEntry{
+		{
+			ID: "mixed-model",
+			Providers: map[string]string{
+				"prov-vllm":   "served-vllm",
+				"prov-ollama": "served-ollama",
+			},
+		},
+	}
+	srv := NewServer(config.EdgeOpenAIConf{}, fake, nil)
+	srv.SetModelCatalog(catalog)
+
+	req := httptest.NewRequest(http.MethodPost, "/v1/chat/completions", strings.NewReader(`{
+		"model":"mixed-model",
+		"messages":[{"role":"user","content":"hello"}]
+	}`))
+	w := httptest.NewRecorder()
+	srv.handleChatCompletions(w, req)
+
+	if w.Code != http.StatusOK {
+		t.Fatalf("status: got %d body=%s", w.Code, w.Body.String())
+	}
+
+	// Tunnel path: exactly 1 tunnel dispatch, 0 normalized SubmitRun.
+	tunnelReqs := fake.tunnelReqsSnapshot()
+	if len(tunnelReqs) != 1 {
+		t.Fatalf("expected 1 tunnel dispatch, got %d", len(tunnelReqs))
+	}
+	if !tunnelReqs[0].ProviderPool {
+		t.Error("tunnel request must have ProviderPool=true")
+	}
+	if tunnelReqs[0].ModelGroupKey != "mixed-model" {
+		t.Errorf("tunnel ModelGroupKey: got %q, want mixed-model", tunnelReqs[0].ModelGroupKey)
+	}
+
+	runReqs := fake.reqsSnapshot()
+	if len(runReqs) != 0 {
+		t.Errorf("expected 0 normalized SubmitRun calls, got %d", len(runReqs))
+	}
+}
+
+// TestChatCompletionsMixedProviderPoolOllamaSelection verifies S05: when the
+// mixed model group selects an Ollama provider, the request is dispatched over
+// the normalized path (SubmitRun), and no tunnel dispatch is issued.
+func TestChatCompletionsMixedProviderPoolOllamaSelection(t *testing.T) {
+	// Normalized path events.
+	events := bufferedRunEvents(
+		&iop.RunEvent{Type: "delta", Delta: "ollama"},
+		&iop.RunEvent{Type: "complete"},
+	)
+
+	// Fake service selects normalized path (simulating Ollama provider selection).
+	fake := &fakeRunService{
+		poolDispatchPath: string(edgeservice.ProviderPoolPathNormalized),
+		events:           events,
+	}
+
+	// Same mixed catalog: vLLM + Ollama.
+	catalog := []config.ModelCatalogEntry{
+		{
+			ID: "mixed-model",
+			Providers: map[string]string{
+				"prov-vllm":   "served-vllm",
+				"prov-ollama": "served-ollama",
+			},
+		},
+	}
+	srv := NewServer(config.EdgeOpenAIConf{}, fake, nil)
+	srv.SetModelCatalog(catalog)
+
+	req := httptest.NewRequest(http.MethodPost, "/v1/chat/completions", strings.NewReader(`{
+		"model":"mixed-model",
+		"messages":[{"role":"user","content":"hello"}]
+	}`))
+	w := httptest.NewRecorder()
+	srv.handleChatCompletions(w, req)
+
+	if w.Code != http.StatusOK {
+		t.Fatalf("status: got %d body=%s", w.Code, w.Body.String())
+	}
+
+	// Normalized path: 0 tunnel dispatches, 1 normalized SubmitRun.
+	tunnelReqs := fake.tunnelReqsSnapshot()
+	if len(tunnelReqs) != 0 {
+		t.Fatalf("expected 0 tunnel dispatches for Ollama selection, got %d", len(tunnelReqs))
+	}
+
+	runReqs := fake.reqsSnapshot()
+	if len(runReqs) != 1 {
+		t.Fatalf("expected 1 normalized SubmitRun, got %d", len(runReqs))
+	}
+	// Adapter/Target are rewritten by the service's admitted candidate; the
+	// fake here records the pre-admission request so the handler-level path
+	// selection (tunnel vs normalized) is the assertion surface.
+	if runReqs[0].ModelGroupKey != "mixed-model" {
+		t.Errorf("normalized ModelGroupKey: got %q, want mixed-model", runReqs[0].ModelGroupKey)
+	}
+	if !runReqs[0].ProviderPool {
+		t.Error("normalized run must have ProviderPool=true")
+	}
+}
+
+// TestChatCompletionsOllamaOnlyProviderPoolUsesNormalizedPath verifies S05:
+// a model group containing only Ollama providers always dispatches over the
+// normalized path. No tunnel request is ever issued.
+func TestChatCompletionsOllamaOnlyProviderPoolUsesNormalizedPath(t *testing.T) {
+	events := bufferedRunEvents(
+		&iop.RunEvent{Type: "delta", Delta: "ollama-only"},
+		&iop.RunEvent{Type: "complete"},
+	)
+
+	// Fake service selects normalized path.
+	fake := &fakeRunService{
+		poolDispatchPath: string(edgeservice.ProviderPoolPathNormalized),
+		events:           events,
+	}
+
+	// Ollama-only catalog.
+	catalog := []config.ModelCatalogEntry{
+		{
+			ID: "ollama-only-model",
+			Providers: map[string]string{
+				"prov-ollama": "qwen3.6:35b",
+			},
+		},
+	}
+	srv := NewServer(config.EdgeOpenAIConf{}, fake, nil)
+	srv.SetModelCatalog(catalog)
+
+	req := httptest.NewRequest(http.MethodPost, "/v1/chat/completions", strings.NewReader(`{
+		"model":"ollama-only-model",
+		"messages":[{"role":"user","content":"hello"}]
+	}`))
+	w := httptest.NewRecorder()
+	srv.handleChatCompletions(w, req)
+
+	if w.Code != http.StatusOK {
+		t.Fatalf("status: got %d body=%s", w.Code, w.Body.String())
+	}
+
+	// Must not dispatch tunnel.
+	tunnelReqs := fake.tunnelReqsSnapshot()
+	if len(tunnelReqs) != 0 {
+		t.Fatalf("ollama-only must not dispatch tunnel, got %d", len(tunnelReqs))
+	}
+
+	// Must dispatch exactly one normalized run.
+	runReqs := fake.reqsSnapshot()
+	if len(runReqs) != 1 {
+		t.Fatalf("expected 1 normalized SubmitRun, got %d", len(runReqs))
+	}
+	if runReqs[0].ModelGroupKey != "ollama-only-model" {
+		t.Errorf("ollama-only ModelGroupKey: got %q, want ollama-only-model", runReqs[0].ModelGroupKey)
+	}
+	if !runReqs[0].ProviderPool {
+		t.Error("ollama-only normalized run must have ProviderPool=true")
+	}
+}
--- a/apps/edge/internal/service/model_queue_test.go
+++ b/apps/edge/internal/service/model_queue_test.go
@ -2232,3 +2232,134 @@ func TestResolveProviderPoolCandidatesClassifiesExecutionPath(t *testing.T) {
 		}
 	}
 }
+
+// TestResolveProviderPoolCandidatesKeepsMixedProviderTypes verifies S05:
+// when a single catalog model group contains both OpenAI-compatible providers
+// (vLLM, lemonade, openai_compat) and Ollama providers, all valid providers
+// remain candidates and each candidate carries the correct execution path.
+// This is the missing mixed-group fixture: previously only single-type
+// provider groups were asserted, so a vLLM + Ollama group coexisting in one
+// catalog entry had no direct evidence.
+func TestResolveProviderPoolCandidatesKeepsMixedProviderTypes(t *testing.T) {
+	catalog := []config.ModelCatalogEntry{
+		{
+			ID: "mixed-model",
+			Providers: map[string]string{
+				"prov-vllm":     "served-vllm",
+				"prov-vllm-mlx": "served-vllm-mlx",
+				"prov-lemonade": "served-lemonade",
+				"prov-ollama":   "served-ollama",
+			},
+		},
+	}
+
+	store := edgenode.NewNodeStore()
+	store.Add(&edgenode.NodeRecord{
+		ID:      "node-mixed",
+		Runtime: config.RuntimeConf{Concurrency: 4},
+		Adapters: config.AdaptersConf{
+			VllmInstances: []config.VllmInstanceConf{
+				{Name: "vllm-gpu", Enabled: true, Endpoint: "http://127.0.0.1:8000/v1"},
+				{Name: "vllm-gpu-mlx", Enabled: true, Endpoint: "http://127.0.0.1:8001/v1"},
+			},
+			OllamaInstances: []config.OllamaInstanceConf{
+				{Name: "ollama", Enabled: true, BaseURL: "http://127.0.0.1:11434"},
+			},
+			OpenAICompatInstances: []config.OpenAICompatInstanceConf{
+				{Name: "oaicompat", Enabled: true, Endpoint: "http://127.0.0.1:9000/v1"},
+			},
+		},
+		Providers: []config.NodeProviderConf{
+			{
+				ID:       "prov-vllm",
+				Type:     "vllm",
+				Adapter:  "vllm-gpu",
+				Models:   []string{"served-vllm"},
+				Health:   "available",
+				Capacity: 2,
+			},
+			{
+				ID:       "prov-vllm-mlx",
+				Type:     "vllm-mlx",
+				Adapter:  "vllm-gpu-mlx",
+				Models:   []string{"served-vllm-mlx"},
+				Health:   "available",
+				Capacity: 2,
+			},
+			{
+				ID:       "prov-lemonade",
+				Type:     "lemonade",
+				Adapter:  "oaicompat",
+				Models:   []string{"served-lemonade"},
+				Health:   "available",
+				Capacity: 2,
+			},
+			{
+				ID:       "prov-ollama",
+				Type:     "ollama",
+				Adapter:  "ollama",
+				Models:   []string{"served-ollama"},
+				Health:   "available",
+				Capacity: 2,
+			},
+		},
+	})
+
+	reg := edgenode.NewRegistry()
+	for _, rec := range store.All() {
+		reg.Register(&edgenode.NodeEntry{NodeID: rec.ID, LifecycleState: edgenode.LifecycleConnected})
+	}
+
+	svc := New(reg, nil)
+	svc.SetNodeStore(store)
+	svc.SetModelCatalog(catalog)
+
+	req := SubmitRunRequest{ModelGroupKey: "mixed-model", ProviderPool: true}
+	candidates, _, err := svc.resolveProviderPoolCandidates(req, store, catalog)
+	if err != nil {
+		t.Fatalf("resolveProviderPoolCandidates: %v", err)
+	}
+
+	// All 4 providers must be candidates.
+	if len(candidates) != 4 {
+		ids := make([]string, len(candidates))
+		for i, c := range candidates {
+			ids[i] = c.providerID
+		}
+		t.Fatalf("expected 4 candidates (prov-vllm, prov-vllm-mlx, prov-lemonade, prov-ollama), got %d: %v", len(candidates), ids)
+	}
+
+	byID := map[string]*candidateNode{}
+	for i := range candidates {
+		byID[candidates[i].providerID] = &candidates[i]
+	}
+
+	// OpenAI-compatible providers (vLLM, vLLM-MLX, lemonade) must be tunnel path.
+	for _, provID := range []string{"prov-vllm", "prov-vllm-mlx", "prov-lemonade"} {
+		c, ok := byID[provID]
+		if !ok {
+			t.Errorf("%q not in candidates", provID)
+			continue
+		}
+		if c.executionPath != providerExecutionPathTunnel {
+			t.Errorf("%q: expected tunnel path, got %q", provID, c.executionPath)
+		}
+		// Served target must match the catalog entry.
+		expectedServed := catalog[0].Providers[provID]
+		if c.servedTarget != expectedServed {
+			t.Errorf("%q servedTarget: got %q, want %q", provID, c.servedTarget, expectedServed)
+		}
+	}
+
+	// Ollama provider must be normalized path.
+	ollama, ok := byID["prov-ollama"]
+	if !ok {
+		t.Fatal("prov-ollama not in candidates")
+	}
+	if ollama.executionPath != providerExecutionPathNormalized {
+		t.Errorf("prov-ollama: expected normalized path, got %q", ollama.executionPath)
+	}
+	if ollama.servedTarget != "served-ollama" {
+		t.Errorf("prov-ollama servedTarget: got %q, want %q", ollama.servedTarget, "served-ollama")
+	}
+}