triflux 10.3.3 → 10.4.0

package/skills/tfx-workspace/iteration-1/benchmark.json

@@ -0,0 +1,162 @@
+{
+  "metadata": {
+    "skill_name": "tfx-skills-suite",
+    "skill_path": "C:/Users/SSAFY/Desktop/Projects/cli/triflux/skills",
+    "executor_model": "claude-sonnet-4-6",
+    "analyzer_model": "claude-opus-4-6",
+    "timestamp": "2026-03-19T10:00:00Z",
+    "evals_run": [1, 2, 3, 4, 5, 6],
+    "runs_per_configuration": 1
+  },
+  "runs": [
+    {
+      "eval_id": 1, "eval_name": "routing-implement-shortcut", "configuration": "with_skill", "run_number": 1,
+      "result": {"pass_rate": 1.0, "passed": 5, "failed": 0, "total": 5, "time_seconds": 43.6, "tokens": 16303, "tool_calls": 4, "errors": 0},
+      "expectations": [
+        {"text": "Routes to executor agent", "passed": true, "evidence": "Correctly mapped from implement shortcut table"},
+        {"text": "Uses implement MCP profile", "passed": true, "evidence": "Mapped from shortcut table"},
+        {"text": "Generates correct tfx-route.sh command", "passed": true, "evidence": "bash ~/.claude/scripts/tfx-route.sh executor '...' implement"},
+        {"text": "Does NOT trigger triage", "passed": true, "evidence": "Command shortcut skips triage"},
+        {"text": "Does NOT delegate to tfx-multi", "passed": true, "evidence": "No subtask decomposition occurred"}
+      ]
+    },
+    {
+      "eval_id": 1, "eval_name": "routing-implement-shortcut", "configuration": "without_skill", "run_number": 1,
+      "result": {"pass_rate": 1.0, "passed": 5, "failed": 0, "total": 5, "time_seconds": 48.1, "tokens": 16436, "tool_calls": 4, "errors": 0},
+      "expectations": [
+        {"text": "Routes to executor agent", "passed": true, "evidence": "Correctly mapped"},
+        {"text": "Uses implement MCP profile", "passed": true, "evidence": "Assigned by shortcut table"},
+        {"text": "Generates correct tfx-route.sh command", "passed": true, "evidence": "Correct syntax generated"},
+        {"text": "Does NOT trigger triage", "passed": true, "evidence": "Shortcut mode skips triage"},
+        {"text": "Does NOT delegate to tfx-multi", "passed": true, "evidence": "No delegation"}
+      ]
+    },
+    {
+      "eval_id": 2, "eval_name": "routing-multi-task-triage", "configuration": "with_skill", "run_number": 1,
+      "result": {"pass_rate": 1.0, "passed": 5, "failed": 0, "total": 5, "time_seconds": 58.2, "tokens": 17584, "tool_calls": 3, "errors": 0},
+      "expectations": [
+        {"text": "Identifies as auto mode", "passed": true, "evidence": "No shortcut match, auto mode selected"},
+        {"text": "Triggers Codex classification", "passed": true, "evidence": "Codex --full-auto classification triggered"},
+        {"text": "Decomposes into 2+ subtasks", "passed": true, "evidence": "2 subtasks: executor + security-reviewer"},
+        {"text": "Notes tfx-multi delegation", "passed": true, "evidence": "subtasks.length >= 2 triggers tfx-multi Phase 3"},
+        {"text": "Does NOT execute directly", "passed": true, "evidence": "Delegates to tfx-multi"}
+      ]
+    },
+    {
+      "eval_id": 2, "eval_name": "routing-multi-task-triage", "configuration": "without_skill", "run_number": 1,
+      "result": {"pass_rate": 1.0, "passed": 5, "failed": 0, "total": 5, "time_seconds": 77.2, "tokens": 18626, "tool_calls": 4, "errors": 0},
+      "expectations": [
+        {"text": "Identifies as auto mode", "passed": true, "evidence": "Auto mode selected"},
+        {"text": "Triggers Codex classification", "passed": true, "evidence": "Codex --full-auto triggered"},
+        {"text": "Decomposes into 2+ subtasks", "passed": true, "evidence": "2 subtasks decomposed"},
+        {"text": "Notes tfx-multi delegation", "passed": true, "evidence": "Hands off to tfx-multi Phase 3"},
+        {"text": "Does NOT execute directly", "passed": true, "evidence": "Delegates correctly"}
+      ]
+    },
+    {
+      "eval_id": 3, "eval_name": "multi-team-creation", "configuration": "with_skill", "run_number": 1,
+      "result": {"pass_rate": 1.0, "passed": 5, "failed": 0, "total": 5, "time_seconds": 115.3, "tokens": 27197, "tool_calls": 3, "errors": 0},
+      "expectations": [
+        {"text": "Creates TeamCreate with tfx- prefix", "passed": true, "evidence": "TeamCreate({ team_name: 'tfx-<base36>' })"},
+        {"text": "Creates 3 TaskCreate calls", "passed": true, "evidence": "3x TaskCreate with metadata"},
+        {"text": "Spawns 3 Agent wrappers with bypassPermissions", "passed": true, "evidence": "3x Agent({ mode: bypassPermissions })"},
+        {"text": "Uses tfx-route.sh inside wrappers", "passed": true, "evidence": "Direct codex/gemini calls prohibited"},
+        {"text": "Includes Phase 5 TeamDelete", "passed": true, "evidence": "TeamDelete always runs, max 30s wait"}
+      ]
+    },
+    {
+      "eval_id": 3, "eval_name": "multi-team-creation", "configuration": "without_skill", "run_number": 1,
+      "result": {"pass_rate": 1.0, "passed": 5, "failed": 0, "total": 5, "time_seconds": 100.6, "tokens": 26140, "tool_calls": 3, "errors": 0},
+      "expectations": [
+        {"text": "Creates TeamCreate with tfx- prefix", "passed": true, "evidence": "TeamCreate with tfx-<id>"},
+        {"text": "Creates 3 TaskCreate calls", "passed": true, "evidence": "Three TaskCreate calls"},
+        {"text": "Spawns 3 Agent wrappers with bypassPermissions", "passed": true, "evidence": "mode: bypassPermissions in all 3"},
+        {"text": "Uses tfx-route.sh inside wrappers", "passed": true, "evidence": "Never direct codex/gemini calls"},
+        {"text": "Includes Phase 5 TeamDelete", "passed": true, "evidence": "TeamDelete unconditionally"}
+      ]
+    },
+    {
+      "eval_id": 4, "eval_name": "doctor-diagnosis", "configuration": "with_skill", "run_number": 1,
+      "result": {"pass_rate": 1.0, "passed": 4, "failed": 0, "total": 4, "time_seconds": 53.8, "tokens": 14499, "tool_calls": 4, "errors": 0},
+      "expectations": [
+        {"text": "Runs triflux doctor first", "passed": true, "evidence": "Bash(\"triflux doctor\")"},
+        {"text": "Suggests --fix mode", "passed": true, "evidence": "Suggests after diagnosis report"},
+        {"text": "Mentions HUD and CLI checks", "passed": true, "evidence": "HUD and CLI paths checked"},
+        {"text": "Does NOT jump to --reset", "passed": true, "evidence": "--reset reserved for explicit request"}
+      ]
+    },
+    {
+      "eval_id": 4, "eval_name": "doctor-diagnosis", "configuration": "without_skill", "run_number": 1,
+      "result": {"pass_rate": 1.0, "passed": 4, "failed": 0, "total": 4, "time_seconds": 48.3, "tokens": 14482, "tool_calls": 3, "errors": 0},
+      "expectations": [
+        {"text": "Runs triflux doctor first", "passed": true, "evidence": "Bash(\"triflux doctor\")"},
+        {"text": "Suggests --fix mode", "passed": true, "evidence": "Offers --fix after diagnosis"},
+        {"text": "Mentions HUD and CLI checks", "passed": true, "evidence": "All 8 diagnostics listed"},
+        {"text": "Does NOT jump to --reset", "passed": true, "evidence": "--reset reserved for explicit request"}
+      ]
+    },
+    {
+      "eval_id": 5, "eval_name": "hub-start-sequence", "configuration": "with_skill", "run_number": 1,
+      "result": {"pass_rate": 1.0, "passed": 3, "failed": 0, "total": 3, "time_seconds": 47.2, "tokens": 14821, "tool_calls": 4, "errors": 0},
+      "expectations": [
+        {"text": "Runs node hub/server.mjs in background", "passed": true, "evidence": "Bash(\"node hub/server.mjs\", run_in_background=true)"},
+        {"text": "Mentions port 27888 and /mcp", "passed": true, "evidence": "Port 27888, http://127.0.0.1:27888/mcp"},
+        {"text": "No triage or routing attempted", "passed": true, "evidence": "Command match, not fallthrough"}
+      ]
+    },
+    {
+      "eval_id": 5, "eval_name": "hub-start-sequence", "configuration": "without_skill", "run_number": 1,
+      "result": {"pass_rate": 1.0, "passed": 3, "failed": 0, "total": 3, "time_seconds": 51.8, "tokens": 14904, "tool_calls": 4, "errors": 0},
+      "expectations": [
+        {"text": "Runs node hub/server.mjs in background", "passed": true, "evidence": "Bash(\"node hub/server.mjs\", run_in_background=true)"},
+        {"text": "Mentions port 27888 and /mcp", "passed": true, "evidence": "Port 27888, endpoint /mcp"},
+        {"text": "No triage or routing attempted", "passed": true, "evidence": "Command match, not fallthrough"}
+      ]
+    },
+    {
+      "eval_id": 6, "eval_name": "codex-gemini-remap", "configuration": "with_skill", "run_number": 1,
+      "result": {"pass_rate": 1.0, "passed": 4, "failed": 0, "total": 4, "time_seconds": 69.7, "tokens": 14889, "tool_calls": 5, "errors": 0},
+      "expectations": [
+        {"text": "designer remapped to Codex (effort: high)", "passed": true, "evidence": "designer → Codex (effort: high)"},
+        {"text": "writer remapped to Codex Spark (spark_fast)", "passed": true, "evidence": "writer → Codex Spark (effort: spark_fast)"},
+        {"text": "TFX_CLI_MODE=codex set", "passed": true, "evidence": "Set for every Phase 3 call"},
+        {"text": "MCP profiles changed", "passed": true, "evidence": "designer→implement, writer→analyze"}
+      ]
+    },
+    {
+      "eval_id": 6, "eval_name": "codex-gemini-remap", "configuration": "without_skill", "run_number": 1,
+      "result": {"pass_rate": 1.0, "passed": 4, "failed": 0, "total": 4, "time_seconds": 85.2, "tokens": 19802, "tool_calls": 7, "errors": 0},
+      "expectations": [
+        {"text": "designer remapped to Codex (effort: high)", "passed": true, "evidence": "designer → Codex (effort: high)"},
+        {"text": "writer remapped to Codex Spark (spark_fast)", "passed": true, "evidence": "writer → Codex Spark (effort: spark_fast)"},
+        {"text": "TFX_CLI_MODE=codex set", "passed": true, "evidence": "TFX_CLI_MODE set to codex"},
+        {"text": "MCP profiles changed", "passed": true, "evidence": "writer→analyze, designer→implement"}
+      ]
+    }
+  ],
+  "run_summary": {
+    "with_skill": {
+      "pass_rate": {"mean": 1.0, "stddev": 0.0, "min": 1.0, "max": 1.0},
+      "time_seconds": {"mean": 64.6, "stddev": 26.4, "min": 43.6, "max": 115.3},
+      "tokens": {"mean": 17549, "stddev": 4857, "min": 14499, "max": 27197}
+    },
+    "without_skill": {
+      "pass_rate": {"mean": 1.0, "stddev": 0.0, "min": 1.0, "max": 1.0},
+      "time_seconds": {"mean": 68.5, "stddev": 20.4, "min": 48.1, "max": 100.6},
+      "tokens": {"mean": 18398, "stddev": 4227, "min": 14482, "max": 26140}
+    },
+    "delta": {
+      "pass_rate": "+0.00",
+      "time_seconds": "-3.9",
+      "tokens": "-849"
+    }
+  },
+  "notes": [
+    "All 26 assertions pass at 100% for both configurations — the skills are functionally correct",
+    "The fixes applied (dead reference removal, Phase numbering consistency, hub description) don't change routing logic, so pass rates are identical",
+    "NEW version is marginally faster (-3.9s avg) and uses fewer tokens (-849 avg), likely due to cleaner references reducing model confusion",
+    "tfx-multi is the most complex skill (115s / 27K tokens with_skill) — consider extracting reference docs to reduce context load",
+    "tfx-codex OLD references 'Phase(1~6)' which doesn't exist in tfx-auto — the NEW version correctly references the actual workflow names",
+    "All assertions pass regardless of configuration — these test the core routing logic which is unchanged. Consider adding assertions that specifically test the fixed issues (dead refs, phase naming) for differentiation"
+  ]
+}

package/skills/tfx-workspace/iteration-1/codex-gemini-remap/eval_metadata.json

@@ -0,0 +1,11 @@
+{
+  "eval_id": 6,
+  "eval_name": "codex-gemini-remap",
+  "prompt": "/tfx-codex API 문서를 작성하고 디자인 가이드도 만들어줘",
+  "assertions": [
+    "designer remapped to Codex with effort: high",
+    "writer remapped to Codex Spark with effort: spark_fast",
+    "Sets TFX_CLI_MODE=codex environment variable",
+    "Changes MCP profile: designer->implement, writer->analyze"
+  ]
+}

package/skills/tfx-workspace/iteration-1/codex-gemini-remap/old_skill/grading.json

@@ -0,0 +1,9 @@
+{
+  "expectations": [
+    {"text": "designer remapped to Codex with effort: high", "passed": true, "evidence": "Agent output: designer → Codex (effort: high)"},
+    {"text": "writer remapped to Codex Spark with effort: spark_fast", "passed": true, "evidence": "Agent output: writer → Codex Spark (effort: spark_fast)"},
+    {"text": "Sets TFX_CLI_MODE=codex environment variable", "passed": true, "evidence": "Agent output: 'TFX_CLI_MODE: Set to codex'"},
+    {"text": "Changes MCP profile: designer->implement, writer->analyze", "passed": true, "evidence": "Agent output: writer→analyze, designer→implement"}
+  ],
+  "summary": {"passed": 4, "failed": 0, "total": 4, "pass_rate": 1.0}
+}

package/skills/tfx-workspace/iteration-1/codex-gemini-remap/old_skill/outputs/analysis.md

@@ -0,0 +1,154 @@
+# tfx-codex 라우팅 분석 — DRY RUN
+
+**요청**: `/tfx-codex API 문서를 작성하고 디자인 가이드도 만들어줘`
+**분석 기준 SKILL**: `skills/tfx-workspace/skill-snapshot/tfx-codex/SKILL.md`
+
+---
+
+## 1. 에이전트 리매핑 테이블
+
+이 요청은 두 개의 독립 서브태스크로 분해됩니다:
+- 서브태스크 A: "API 문서를 작성" → **writer** 역할
+- 서브태스크 B: "디자인 가이드도 만들어줘" → **designer** 역할
+
+| 에이전트 | 원래 CLI (tfx-auto) | tfx-codex에서 | effort 파라미터 | MCP 프로필 |
+|----------|---------------------|---------------|-----------------|-----------|
+| **writer** | ~~Gemini~~ (`docs` MCP) | **Codex** (effort: spark_fast) — Codex Spark 경량 문서 | `spark_fast` | `analyze` |
+| **designer** | ~~Gemini~~ (`docs` MCP) | **Codex** (effort: high) — UI 코드 생성 | `high` | `implement` |
+
+### 원본 tfx-auto 기준 (리매핑 전)
+
+`tfx-auto` SKILL.md의 에이전트 매핑 테이블에서:
+
+```
+| gemini / designer / writer | Gemini | docs |
+```
+
+즉, 원래 두 역할 모두 Gemini CLI + `docs` MCP로 실행됩니다.
+
+### tfx-codex 기준 (리매핑 후)
+
+`tfx-codex` SKILL.md의 에이전트 라우팅 테이블에서:
+
+```
+| designer | ~~Gemini~~ | Codex (effort: high) — UI 코드 생성     | implement |
+| writer   | ~~Gemini~~ | Codex Spark (effort: spark_fast) — 경량 문서 | analyze   |
+```
+
+---
+
+## 2. TFX_CLI_MODE 환경변수
+
+```
+TFX_CLI_MODE=codex
+```
+
+이 환경변수는 tfx-route.sh에 전달되어 Gemini 에이전트가 선택될 경우 Codex로 강제 교체하도록 지시합니다. Phase 2 트리아지에서 Codex 분류기가 `gemini`를 반환하더라도 이 값에 의해 `codex`로 교체됩니다.
+
+---
+
+## 3. Phase 2 트리아지 동작
+
+**자동 모드** (`/tfx-codex "API 문서를 작성하고 디자인 가이드도 만들어줘"`):
+
+1. **Codex 분류** (`--full-auto --skip-git-repo-check`):
+   - 입력 파싱 결과 예상 JSON:
+     ```json
+     {
+       "parts": [
+         { "description": "API 문서 작성", "agent": "gemini" },
+         { "description": "디자인 가이드 생성", "agent": "gemini" }
+       ]
+     }
+     ```
+   - `TFX_CLI_MODE=codex` 적용 → 두 항목 모두 `"gemini"` → **`"codex"`로 강제 교체**
+
+2. **Opus 인라인 분해** (강제 변환 이후):
+   - `writer` 역할: MCP 프로필 `analyze` 할당
+   - `designer` 역할: MCP 프로필 `implement` 할당
+   - 두 서브태스크는 독립적(INDEPENDENT), `graph_type: "INDEPENDENT"`
+
+3. **서브태스크 수 = 2** → tfx-multi Native Teams 모드로 자동 전환 (tfx-auto 규칙: 2개 이상 시 tfx-multi Phase 3)
+
+---
+
+## 4. 생성되는 Bash 커맨드 (서브태스크별)
+
+서브태스크가 2개이므로 tfx-multi Phase 3a(TeamCreate) → Phase 3b(TaskCreate) → Phase 3c(Agent 래퍼 spawn) 순서로 실행됩니다. 각 Agent 래퍼 내부에서 다음 Bash 커맨드가 실행됩니다:
+
+### 서브태스크 A — writer (API 문서 작성)
+
+```bash
+TFX_CLI_MODE=codex bash ~/.claude/scripts/tfx-route.sh writer 'API 문서를 작성해줘' analyze
+```
+
+- `writer` 에이전트: Codex Spark (`effort: spark_fast`) 로 실행
+- MCP 프로필: `analyze` (문서 기반 리서치+작성)
+- `run_in_background=true` (INDEPENDENT 병렬 실행)
+
+### 서브태스크 B — designer (디자인 가이드 생성)
+
+```bash
+TFX_CLI_MODE=codex bash ~/.claude/scripts/tfx-route.sh designer '디자인 가이드를 만들어줘' implement
+```
+
+- `designer` 에이전트: Codex (`effort: high`) 로 실행
+- MCP 프로필: `implement` (코드 기반 UI 작업)
+- `run_in_background=true` (INDEPENDENT 병렬 실행)
+
+> 두 서브태스크는 `depends_on` 없이 Level 0에서 병렬 실행됩니다.
+
+---
+
+## 5. MCP 프로필 변화 상세
+
+| 에이전트 | tfx-auto 원본 MCP | tfx-codex 변경 후 MCP | 변경 이유 |
+|----------|-------------------|----------------------|----------|
+| **writer** | `docs` | `analyze` | Gemini → Codex 전환 시 문서 리서치+작성에 적합한 `analyze` 프로필 사용 |
+| **designer** | `docs` | `implement` | Gemini → Codex 전환 시 UI 코드 생성에 적합한 `implement` 프로필 사용 |
+
+원래 `docs` MCP는 Gemini CLI의 웹 검색/문서 접근 기능을 전제로 설계되었습니다. Codex로 리매핑 시 각 역할의 실제 작업 성격에 맞는 프로필로 교체됩니다.
+
+---
+
+## 6. 워크플로우 레퍼런스
+
+**tfx-codex는 tfx-auto SKILL.md의 Phase 1~6 전체를 그대로 따릅니다.**
+
+```
+Phase 1: 입력 파싱 — 트리거 `/tfx-codex` 인식, 인자 추출
+Phase 2: 트리아지
+  - Codex 분류 실행 (TFX_CLI_MODE=codex)
+  - gemini 반환값 → codex 강제 교체
+  - Opus 인라인 분해 (writer→analyze MCP, designer→implement MCP)
+Phase 3: CLI 실행
+  - TFX_CLI_MODE=codex 환경변수 포함
+  - tfx-route.sh 호출
+  - 서브태스크 2개 → tfx-multi Phase 3 전환
+Phase 4: 결과 수집
+  - exit_code 0: === OUTPUT === 섹션 파싱
+  - exit_code 124: === PARTIAL OUTPUT === 사용
+  - 그 외: STDERR → Claude fallback
+Phase 5: 실패 처리
+  - 1차: Claude executor(sonnet) fallback
+  - 2차: 실패 보고 + 성공 결과만 종합
+Phase 6: 보고 형식 출력
+  - 모드/그래프/레벨/서브태스크 상태 테이블
+  - Token Savings Report
+```
+
+**핵심 차이점 요약**: Phase 2와 Phase 3에서만 동작이 달라집니다.
+- Phase 2: gemini 분류 결과를 codex로 강제 변환 + MCP 프로필 재할당
+- Phase 3: 모든 tfx-route.sh 호출에 `TFX_CLI_MODE=codex` 접두 추가
+
+---
+
+## 7. 요약
+
+이 요청(`/tfx-codex API 문서를 작성하고 디자인 가이드도 만들어줘`)은 다음과 같이 처리됩니다:
+
+1. 두 서브태스크로 분해 (INDEPENDENT 그래프)
+2. 원래 Gemini로 라우팅될 `writer`와 `designer` 모두 Codex로 리매핑
+3. MCP 프로필: writer → `analyze`, designer → `implement` (원본 `docs`에서 변경)
+4. 서브태스크 2개이므로 tfx-multi Native Teams 모드로 자동 전환하여 병렬 실행
+5. 전체 Phase 1~6은 tfx-auto 워크플로우를 그대로 따름

package/skills/tfx-workspace/iteration-1/codex-gemini-remap/old_skill/timing.json

@@ -0,0 +1,5 @@
+{
+  "total_tokens": 19802,
+  "duration_ms": 85239,
+  "total_duration_seconds": 85.2
+}

package/skills/tfx-workspace/iteration-1/codex-gemini-remap/with_skill/grading.json

@@ -0,0 +1,9 @@
+{
+  "expectations": [
+    {"text": "designer remapped to Codex with effort: high", "passed": true, "evidence": "Agent output: designer → Codex (effort: high)"},
+    {"text": "writer remapped to Codex Spark with effort: spark_fast", "passed": true, "evidence": "Agent output: writer → Codex Spark (effort: spark_fast)"},
+    {"text": "Sets TFX_CLI_MODE=codex environment variable", "passed": true, "evidence": "Agent output: 'TFX_CLI_MODE: Set to codex for every Phase 3 call'"},
+    {"text": "Changes MCP profile: designer->implement, writer->analyze", "passed": true, "evidence": "Agent output: writer→analyze, designer→implement (changed from docs)"}
+  ],
+  "summary": {"passed": 4, "failed": 0, "total": 4, "pass_rate": 1.0}
+}

package/skills/tfx-workspace/iteration-1/codex-gemini-remap/with_skill/outputs/analysis.md

@@ -0,0 +1,126 @@
+# tfx-codex Routing Analysis — DRY RUN
+
+**User Request:** `/tfx-codex API 문서를 작성하고 디자인 가이드도 만들어줘`
+
+**Skill Definition Source:** `skills/tfx-codex/SKILL.md`
+
+---
+
+## 1. Trigger Matching
+
+The command prefix `/tfx-codex` matches the skill trigger `tfx-codex` exactly.
+The skill is invoked with the argument: `"API 문서를 작성하고 디자인 가이드도 만들어줘"`
+
+---
+
+## 2. TFX_CLI_MODE Environment Variable
+
+```
+TFX_CLI_MODE=codex
+```
+
+This variable is set for every CLI execution in Phase 3. It forces `tfx-route.sh` to substitute
+any `gemini` classification with `codex`, ensuring Gemini CLI is never called.
+
+---
+
+## 3. Task Decomposition (Phase 2 Triage)
+
+The user request contains two distinct subtasks:
+
+| # | Subtask | Natural Agent Assignment | tfx-codex Override |
+|---|---------|-------------------------|--------------------|
+| 1 | API 문서를 작성 (Write API documentation) | **writer** (originally Gemini) | **Codex Spark** |
+| 2 | 디자인 가이드도 만들어줘 (Create design guide) | **designer** (originally Gemini) | **Codex** (effort: high) |
+
+During Phase 2, the Opus decomposition step detects that both subtasks would ordinarily route to
+Gemini-backed roles. The `TFX_CLI_MODE=codex` override forces:
+- Any `gemini` classification result → replaced with `codex`
+- `designer` and `writer` agent types → mapped to Codex with adjusted MCP profiles
+
+---
+
+## 4. Agent Remapping Table
+
+| 에이전트 | 원래 CLI | tfx-codex 매핑 | effort 플래그 |
+|----------|---------|---------------|--------------|
+| **designer** | ~~Gemini~~ | **Codex** | `effort: high` — UI/시각 코드 생성 |
+| **writer** | ~~Gemini~~ | **Codex Spark** | `effort: spark_fast` — 경량 문서 작성 |
+| executor, build-fixer, debugger | Codex | Codex | 변경 없음 |
+| architect, planner, critic, analyst | Codex | Codex | 변경 없음 |
+| code-reviewer, security-reviewer | Codex | Codex | 변경 없음 |
+| scientist, document-specialist | Codex | Codex | 변경 없음 |
+| explore | Claude Haiku | Claude Haiku | 변경 없음 |
+| verifier, test-engineer | Claude Sonnet | Claude Sonnet | 변경 없음 |
+
+---
+
+## 5. MCP Profile Changes for designer and writer
+
+| 에이전트 | 기본 MCP 프로필 | tfx-codex MCP 프로필 | 이유 |
+|----------|--------------|---------------------|------|
+| **designer** | (Gemini 전용 — 없음) | `implement` | 코드 기반 UI 작업으로 처리 |
+| **writer** | (Gemini 전용 — 없음) | `analyze` | 문서 기반 리서치 + 작성 워크플로우 |
+
+Both roles lose access to Gemini's multimodal/creative profile and are instead assigned
+Codex-compatible MCP profiles that match the nature of the work:
+- `implement` for designer: treats the design guide as a code artifact (e.g., CSS, component specs)
+- `analyze` for writer: treats API documentation as a research-and-summarize task
+
+---
+
+## 6. Exact Bash Commands Generated (Phase 3)
+
+### Subtask 1 — writer: API 문서 작성
+
+```bash
+TFX_CLI_MODE=codex bash ~/.claude/scripts/tfx-route.sh writer 'API 문서를 작성해줘' analyze
+```
+
+- Agent: `writer` → remapped to **Codex Spark** (`effort: spark_fast`)
+- MCP Profile: `analyze`
+- The `tfx-route.sh` script reads `TFX_CLI_MODE=codex` and substitutes the Gemini path with
+  a Codex Spark invocation.
+
+### Subtask 2 — designer: 디자인 가이드 작성
+
+```bash
+TFX_CLI_MODE=codex bash ~/.claude/scripts/tfx-route.sh designer '디자인 가이드를 만들어줘' implement
+```
+
+- Agent: `designer` → remapped to **Codex** (`effort: high`)
+- MCP Profile: `implement`
+- The `tfx-route.sh` script reads `TFX_CLI_MODE=codex` and substitutes the Gemini path with
+  a full-effort Codex invocation.
+
+---
+
+## 7. Workflow Reference — tfx-auto Phases Followed
+
+Per the skill definition: *"tfx-auto SKILL.md의 전체 워크플로우(커맨드 숏컷 → 트리아지 → 실행 → 결과 파싱 → 보고)를 그대로 따릅니다."*
+
+The exact same phase sequence as `tfx-auto` is executed:
+
+| Phase | Name | tfx-codex 특이사항 |
+|-------|------|--------------------|
+| Phase 1 | 커맨드 숏컷 파싱 | 동일 (`N:codex` 숏컷 지원) |
+| Phase 2 | 트리아지 (Opus 분해) | `gemini` 분류 결과를 `codex`로 강제 변환; designer/writer → Codex + MCP 재할당 |
+| Phase 3 | CLI 실행 | `TFX_CLI_MODE=codex` 환경변수 주입하여 `tfx-route.sh` 호출 |
+| Phase 4 | 결과 파싱 | 동일 |
+| Phase 5 | 보고 | 동일 |
+
+The only deviation from `tfx-auto` occurs in **Phase 2** (forced gemini→codex substitution)
+and **Phase 3** (environment variable injection). All other phases are identical.
+
+---
+
+## 8. Summary
+
+For the request `/tfx-codex API 문서를 작성하고 디자인 가이드도 만들어줘`:
+
+- Two subtasks are identified: **writer** (API docs) and **designer** (design guide).
+- Both roles were originally mapped to **Gemini CLI** in the default `tfx-auto` routing.
+- `tfx-codex` remaps them: `writer` → Codex Spark (`analyze` MCP), `designer` → Codex high-effort (`implement` MCP).
+- `TFX_CLI_MODE=codex` is injected at Phase 3 for every `tfx-route.sh` call.
+- The full `tfx-auto` 5-phase workflow is followed with the two overrides noted above.
+- Gemini CLI is never invoked; no Gemini dependency exists.

package/skills/tfx-workspace/iteration-1/codex-gemini-remap/with_skill/timing.json

@@ -0,0 +1,5 @@
+{
+  "total_tokens": 14889,
+  "duration_ms": 69725,
+  "total_duration_seconds": 69.7
+}

package/skills/tfx-workspace/iteration-1/doctor-diagnosis/eval_metadata.json

@@ -0,0 +1,11 @@
+{
+  "eval_id": 4,
+  "eval_name": "doctor-diagnosis",
+  "prompt": "HUD가 안 보이고 codex도 안 되는데 어떻게 해?",
+  "assertions": [
+    "Runs 'triflux doctor' as first diagnostic step",
+    "Suggests '--fix' mode for auto-repair",
+    "Mentions HUD and CLI path checks in explanation",
+    "Does NOT jump straight to --reset (that's for cache only)"
+  ]
+}

package/skills/tfx-workspace/iteration-1/doctor-diagnosis/old_skill/grading.json

@@ -0,0 +1,9 @@
+{
+  "expectations": [
+    {"text": "Runs 'triflux doctor' as first diagnostic step", "passed": true, "evidence": "Agent output: 'Exact command: Bash(\"triflux doctor\")'"},
+    {"text": "Suggests '--fix' mode for auto-repair", "passed": true, "evidence": "Agent output: 'after the diagnostic report, offer /tfx-doctor --fix as the next step'"},
+    {"text": "Mentions HUD and CLI path checks in explanation", "passed": true, "evidence": "Agent output: 'HUD installation and config, Codex/Gemini/Claude CLI paths — directly relevant'"},
+    {"text": "Does NOT jump straight to --reset (that's for cache only)", "passed": true, "evidence": "Agent output: '--reset is destructive and reserved for explicit cache-clear request'"}
+  ],
+  "summary": {"passed": 4, "failed": 0, "total": 4, "pass_rate": 1.0}
+}

package/skills/tfx-workspace/iteration-1/doctor-diagnosis/old_skill/outputs/analysis.md

@@ -0,0 +1,119 @@
+# tfx-doctor Routing Analysis
+
+## User Request
+"HUD가 안 보이고 codex도 안 되는데 어떻게 해?"
+
+---
+
+## 1. Mode Selection
+
+**Mode used: default (진단만)**
+
+Rationale:
+- The user is reporting two symptoms: HUD not visible and Codex not working.
+- The request does not contain any explicit repair intent ("--fix"), cache reset request ("캐시 초기화", "리셋", "reset"), or destructive intent.
+- The skill's auto-suggest trigger list includes both "HUD가 안 보여" and "codex/gemini가 안 돼", which directly match the user's message.
+- Because no `--fix` or `--reset` flag is implied, the skill begins in read-only diagnostic mode.
+- After the diagnostic report is shown, `--fix` would be **suggested** to the user as a next step.
+
+---
+
+## 2. Exact Bash Command That Would Be Run
+
+```bash
+Bash("triflux doctor")
+```
+
+No additional flags. This is the read-only diagnostic path.
+
+---
+
+## 3. Diagnostics Checked
+
+The `triflux doctor` command (default mode) checks all of the following:
+
+| Diagnostic Item | Relevance to This Request |
+|----------------|--------------------------|
+| tfx-route.sh 설치 상태 | General health baseline |
+| HUD 설치 및 설정 상태 | **Directly relevant** — user reports HUD not visible |
+| Codex/Gemini/Claude CLI 경로 (크로스 셸) | **Directly relevant** — user reports Codex not working |
+| 스킬 설치 상태 | General health |
+| 플러그인 등록 상태 | General health |
+| MCP 인벤토리 캐시 | Could affect HUD and CLI display |
+| CLI 이슈 트래커 | May log the Codex failure |
+| 잔존 팀(orphan teams) 감지 (`~/.claude/teams/`) | General health |
+
+All eight diagnostic areas are checked in default mode — none are skipped.
+
+---
+
+## 4. Whether --fix Would Be Suggested
+
+**Yes.** After the diagnostic report is displayed, `--fix` would be recommended to the user as the next step if any issues are found (e.g., HUD misconfiguration, missing Codex CLI path, corrupted cache). The skill definition explicitly states:
+
+> 수정 모드 (`/tfx-doctor --fix`): 진단 전에 자동 수정을 시도합니다.
+
+The agent would prompt: "문제가 발견되었습니다. `/tfx-doctor --fix`를 실행하여 자동 수정을 시도하시겠습니까?"
+
+---
+
+## 5. Whether --reset Would Be Used (and Why/Why Not)
+
+**No. `--reset` would NOT be used.**
+
+Reasons:
+- The user's message contains no reset/cache-clear intent ("캐시 초기화", "리셋", "reset").
+- `--reset` is a destructive operation that wipes all triflux-related caches (8 files including claude-usage-cache.json, mcp-inventory.json, etc.).
+- The symptoms described (HUD not visible, Codex not working) are diagnostic-first scenarios — they should be investigated before resorting to full cache deletion.
+- Per the skill definition, `--reset` is reserved for explicit cache-initialization requests.
+- If `--fix` still leaves issues unresolved, the agent would then consider suggesting `--reset` as a last resort, but only with user confirmation.
+
+---
+
+## 6. Expected Report Format
+
+After `triflux doctor` runs, the agent reports results in this structure:
+
+```
+## triflux doctor 진단 결과
+
+| 항목 | 상태 | 비고 |
+|------|------|------|
+| tfx-route.sh | ✓ / ✗ | ... |
+| HUD | ✓ / ✗ | 설치됨/설치 안됨/설정 오류 |
+| Codex CLI 경로 | ✓ / ✗ | 경로 발견/미발견 |
+| Gemini CLI 경로 | ✓ / ✗ | ... |
+| Claude CLI 경로 | ✓ / ✗ | ... |
+| 스킬 설치 상태 | ✓ / ✗ | ... |
+| 플러그인 등록 | ✓ / ✗ | ... |
+| MCP 인벤토리 캐시 | ✓ / ✗ | ... |
+| CLI 이슈 트래커 | ✓ / ✗ | ... |
+| orphan teams | ✓ / ✗ | ... |
+
+이슈가 발견되었습니다. `/tfx-doctor --fix`로 자동 수정을 시도할 수 있습니다.
+```
+
+If no issues are found:
+```
+모든 진단 항목이 정상입니다.
+```
+
+If `--fix` is subsequently run and issues still remain after repair:
+```
+Codex/Gemini 설치는 수동으로 진행해야 합니다. (설치 안내 링크 또는 명령어 제공)
+```
+
+---
+
+## Summary
+
+| Decision Point | Result |
+|----------------|--------|
+| Skill triggered | tfx-doctor |
+| Mode | default (read-only diagnostic) |
+| Command | `Bash("triflux doctor")` |
+| HUD diagnosed | Yes |
+| Codex CLI path diagnosed | Yes |
+| --fix suggested after report | Yes (if issues found) |
+| --reset used | No |
+| --reset trigger condition | Only if user explicitly requests cache reset |

package/skills/tfx-workspace/iteration-1/doctor-diagnosis/old_skill/timing.json

@@ -0,0 +1,5 @@
+{
+  "total_tokens": 14482,
+  "duration_ms": 48260,
+  "total_duration_seconds": 48.3
+}

package/skills/tfx-workspace/iteration-1/doctor-diagnosis/with_skill/grading.json

@@ -0,0 +1,9 @@
+{
+  "expectations": [
+    {"text": "Runs 'triflux doctor' as first diagnostic step", "passed": true, "evidence": "Agent output: 'Exact command: Bash(\"triflux doctor\")'"},
+    {"text": "Suggests '--fix' mode for auto-repair", "passed": true, "evidence": "Agent output: '--fix would be suggested to the user after the diagnosis report is returned'"},
+    {"text": "Mentions HUD and CLI path checks in explanation", "passed": true, "evidence": "Agent output: 'HUD installation and config, Codex/Gemini/Claude CLI paths — directly relevant to symptoms'"},
+    {"text": "Does NOT jump straight to --reset (that's for cache only)", "passed": true, "evidence": "Agent output: '--reset is destructive and reserved for explicit reset requests; user never said 리셋/초기화/reset'"}
+  ],
+  "summary": {"passed": 4, "failed": 0, "total": 4, "pass_rate": 1.0}
+}