triflux 8.11.2 → 8.12.2

package/skills/tfx-codex-swarm-workspace/iteration-1/implicit-swarm-no-keywords/with_skill/outputs/commands.sh

@@ -0,0 +1,151 @@
+#!/usr/bin/env bash
+# =============================================================================
+# tfx-codex-swarm — Generated Commands (Steps 5-8)
+# =============================================================================
+# 사용자 요청: "이 PRD 3개 파일을 각각 독립적으로 코덱스한테 맡겨서 병렬로 구현해줘"
+# 생성일: 2026-03-30
+# 상태: DRY-RUN (실제 실행 없음)
+# =============================================================================
+
+set -euo pipefail
+
+PROJECT_ROOT="C:/Users/SSAFY/Desktop/Projects/cli/triflux"
+SWARM_DIR="$PROJECT_ROOT/.codex-swarm"
+BASH_EXE='C:\Program Files\Git\bin\bash.exe'
+
+# =============================================================================
+# Step 5: Worktree 생성
+# =============================================================================
+
+echo "=== Step 5: Worktree 생성 ==="
+
+# auth-refactor worktree
+git worktree add "$SWARM_DIR/wt-auth-refactor" -b codex/auth-refactor 2>/dev/null || \
+git worktree add "$SWARM_DIR/wt-auth-refactor" codex/auth-refactor
+
+# api-v2 worktree
+git worktree add "$SWARM_DIR/wt-api-v2" -b codex/api-v2 2>/dev/null || \
+git worktree add "$SWARM_DIR/wt-api-v2" codex/api-v2
+
+# cache-layer worktree
+git worktree add "$SWARM_DIR/wt-cache-layer" -b codex/cache-layer 2>/dev/null || \
+git worktree add "$SWARM_DIR/wt-cache-layer" codex/cache-layer
+
+# =============================================================================
+# Step 6: 프롬프트 생성
+# =============================================================================
+
+echo "=== Step 6: 프롬프트 생성 ==="
+
+mkdir -p "$SWARM_DIR/prompts"
+
+# --- auth-refactor 프롬프트 ---
+cat > "$SWARM_DIR/prompts/prompt-auth-refactor.md" << 'PROMPT_EOF'
+triflux 프로젝트의 태스크를 리팩터링해야 합니다.
+
+태스크 파일을 먼저 읽으세요: docs/prd/auth-refactor.md
+
+작업 순서:
+1. 태스크 파일을 읽고 요구사항을 파악하세요
+2. $plan을 실행하여 리팩터링 계획을 수립하세요
+3. $ralph를 실행하여 완료까지 반복 실행하세요
+
+프로젝트 정보:
+- triflux: Claude Code용 멀티모델 CLI 오케스트레이터
+- 언어: JavaScript/ESM (Node.js), 테스트: npm test
+PROMPT_EOF
+
+# PRD를 worktree에 복사
+cp "$PROJECT_ROOT/docs/prd/auth-refactor.md" "$SWARM_DIR/wt-auth-refactor/docs/prd/auth-refactor.md" 2>/dev/null || true
+
+# --- api-v2 프롬프트 ---
+cat > "$SWARM_DIR/prompts/prompt-api-v2.md" << 'PROMPT_EOF'
+triflux 프로젝트의 태스크를 구현해야 합니다.
+
+태스크 파일을 먼저 읽으세요: docs/prd/api-v2.md
+
+작업 순서:
+1. 태스크 파일을 읽고 요구사항을 파악하세요
+2. $plan을 실행하여 구현 계획을 수립하세요
+3. $autopilot을 실행하여 자율 구현하세요
+
+프로젝트 정보:
+- triflux: Claude Code용 멀티모델 CLI 오케스트레이터
+- 언어: JavaScript/ESM (Node.js), 테스트: npm test
+PROMPT_EOF
+
+# PRD를 worktree에 복사
+cp "$PROJECT_ROOT/docs/prd/api-v2.md" "$SWARM_DIR/wt-api-v2/docs/prd/api-v2.md" 2>/dev/null || true
+
+# --- cache-layer 프롬프트 ---
+cat > "$SWARM_DIR/prompts/prompt-cache-layer.md" << 'PROMPT_EOF'
+triflux 프로젝트의 태스크를 구현해야 합니다.
+
+태스크 파일을 먼저 읽으세요: docs/prd/cache-layer.md
+
+작업 순서:
+1. 태스크 파일을 읽고 요구사항을 파악하세요
+2. $plan을 실행하여 구현 계획을 수립하세요
+3. $autopilot을 실행하여 자율 구현하세요
+
+프로젝트 정보:
+- triflux: Claude Code용 멀티모델 CLI 오케스트레이터
+- 언어: JavaScript/ESM (Node.js), 테스트: npm test
+PROMPT_EOF
+
+# PRD를 worktree에 복사
+cp "$PROJECT_ROOT/docs/prd/cache-layer.md" "$SWARM_DIR/wt-cache-layer/docs/prd/cache-layer.md" 2>/dev/null || true
+
+# =============================================================================
+# Step 7: psmux 세션 생성 + Codex 실행
+# =============================================================================
+
+echo "=== Step 7: psmux 세션 + Codex 실행 ==="
+
+# 공통 프로파일 플래그 (전부 codex53_high)
+PROFILE_FLAGS="-c 'model=\"gpt-5.3-codex\"' -c 'model_reasoning_effort=\"high\"'"
+
+# --- auth-refactor 세션 ---
+psmux new-session --name "codex-swarm-auth-refactor" --dir "$SWARM_DIR/wt-auth-refactor"
+
+psmux send-keys --target "codex-swarm-auth-refactor:0" \
+  "codex -c 'model=\"gpt-5.3-codex\"' -c 'model_reasoning_effort=\"high\"' --full-auto \"\$(cat $SWARM_DIR/prompts/prompt-auth-refactor.md)\"" Enter
+
+# --- api-v2 세션 ---
+psmux new-session --name "codex-swarm-api-v2" --dir "$SWARM_DIR/wt-api-v2"
+
+psmux send-keys --target "codex-swarm-api-v2:0" \
+  "codex -c 'model=\"gpt-5.3-codex\"' -c 'model_reasoning_effort=\"high\"' --full-auto \"\$(cat $SWARM_DIR/prompts/prompt-api-v2.md)\"" Enter
+
+# --- cache-layer 세션 ---
+psmux new-session --name "codex-swarm-cache-layer" --dir "$SWARM_DIR/wt-cache-layer"
+
+psmux send-keys --target "codex-swarm-cache-layer:0" \
+  "codex -c 'model=\"gpt-5.3-codex\"' -c 'model_reasoning_effort=\"high\"' --full-auto \"\$(cat $SWARM_DIR/prompts/prompt-cache-layer.md)\"" Enter
+
+# =============================================================================
+# Step 8: WT 탭 일괄 attach
+# =============================================================================
+
+echo "=== Step 8: WT 탭 attach ==="
+
+# 첫 번째 세션: 새 WT 윈도우
+psmux attach --session "codex-swarm-auth-refactor" --wt-new-window
+
+# 나머지 세션: 같은 윈도우에 탭 추가
+psmux attach --session "codex-swarm-api-v2" --wt-tab
+psmux attach --session "codex-swarm-cache-layer" --wt-tab
+
+# --- fallback: psmux attach 불가 시 wt.exe 직접 호출 ---
+# wt.exe -w new \
+#   --title "auth-refactor" -d "$SWARM_DIR/wt-auth-refactor" "$BASH_EXE" -c "psmux attach codex-swarm-auth-refactor" \; \
+#   new-tab --title "api-v2" -d "$SWARM_DIR/wt-api-v2" "$BASH_EXE" -c "psmux attach codex-swarm-api-v2" \; \
+#   new-tab --title "cache-layer" -d "$SWARM_DIR/wt-cache-layer" "$BASH_EXE" -c "psmux attach codex-swarm-cache-layer"
+
+echo "=== 스웜 스폰 완료 ==="
+echo ""
+echo "| # | 태스크          | 유형       | OMX 스킬            | Worktree            | 세션                        |"
+echo "|---|-----------------|------------|---------------------|---------------------|-----------------------------|"
+echo "| 1 | auth-refactor   | 리팩터링   | \$plan→\$ralph      | wt-auth-refactor    | codex-swarm-auth-refactor   |"
+echo "| 2 | api-v2          | 구현       | \$plan→\$autopilot  | wt-api-v2           | codex-swarm-api-v2          |"
+echo "| 3 | cache-layer     | 구현       | \$plan→\$autopilot  | wt-cache-layer      | codex-swarm-cache-layer     |"

package/skills/tfx-codex-swarm-workspace/iteration-1/implicit-swarm-no-keywords/with_skill/outputs/routing.md

@@ -0,0 +1,51 @@
+# tfx-codex-swarm — Step 4: 프로파일 자동 라우팅
+
+> 사용자 요청: "이 PRD 3개 파일을 각각 독립적으로 코덱스한테 맡겨서 병렬로 구현해줘"
+
+## 4-1. 태스크 규모 산정
+
+PRD 파일이 실제 존재하지 않으므로 **파일명 기반 추정**으로 규모를 산정한다.
+실제 실행 시에는 PRD 줄 수, 영향 파일 수, 고비용 키워드를 측정한다.
+
+| # | 태스크 | PRD 줄 수 (추정) | 영향 파일 (추정) | 고비용 키워드 | 규모 |
+|---|--------|------------------|-----------------|-------------|------|
+| 1 | auth-refactor | 60줄 | 4-5 파일 | "리팩터링" (1개) | **L (표준)** |
+| 2 | api-v2 | 50줄 | 3-4 파일 | 없음 | **L (표준)** |
+| 3 | cache-layer | 50줄 | 3-4 파일 | 없음 | **L (표준)** |
+
+### 추정 근거
+
+- **auth-refactor**: 인증 리팩터링은 기존 auth 모듈, 미들웨어, 라우트, 테스트 등 다수 파일에 영향. "리팩터링" 고비용 키워드 1개. L 규모.
+- **api-v2**: API v2는 라우트 정의, 컨트롤러, 스키마, 테스트 등 3-4개 파일. 고비용 키워드 없음. L 규모.
+- **cache-layer**: 캐시 레이어는 캐시 모듈, 설정, 통합 포인트, 테스트 등 3-4개 파일. 고비용 키워드 없음. L 규모.
+
+## 4-2. 프로파일 라우팅
+
+유형(Step 3) × 규모(4-1)을 SKILL.md 라우팅 테이블에 대입:
+
+| # | 태스크 | 규모 | 유형 | 프로파일 | Codex CLI 플래그 |
+|---|--------|------|------|----------|-----------------|
+| 1 | auth-refactor | L | 리팩터링 | `codex53_high` | `-c 'model="gpt-5.3-codex"' -c 'model_reasoning_effort="high"'` |
+| 2 | api-v2 | L | 구현 | `codex53_high` | `-c 'model="gpt-5.3-codex"' -c 'model_reasoning_effort="high"'` |
+| 3 | cache-layer | L | 구현 | `codex53_high` | `-c 'model="gpt-5.3-codex"' -c 'model_reasoning_effort="high"'` |
+
+### 라우팅 테이블 참조
+
+SKILL.md Step 4-2 테이블에서:
+- 리팩터링 × L = `codex53_high`
+- 구현 × L = `codex53_high`
+
+3개 태스크 모두 L 규모이므로 동일 프로파일(`codex53_high`)로 라우팅됨.
+
+## 4-3. 기타 설정
+
+| 설정 | 값 | 설명 |
+|------|-----|------|
+| 실행 모드 | `--full-auto` | 사용자가 "맡겨서"라고 했으므로 full-auto 기본값 유지 |
+| worktree | 활성화 | 세션당 독립 worktree |
+| psmux | 활성화 | 세션 관리 |
+| 프로파일 오버라이드 | 없음 | 사용자가 프로파일을 명시하지 않았으므로 자동 라우팅 유지 |
+
+## 요약
+
+3개 태스크 모두 L 규모 + codex53_high 프로파일. gpt-5.3-codex 모델, reasoning effort=high로 실행.

package/skills/tfx-codex-swarm-workspace/iteration-1/implicit-swarm-no-keywords/with_skill/timing.json

@@ -0,0 +1,5 @@
+{
+  "total_tokens": 25984,
+  "duration_ms": 158802,
+  "total_duration_seconds": 158.8
+}

package/skills/tfx-codex-swarm-workspace/iteration-1/implicit-swarm-no-keywords/without_skill/grading.json

@@ -0,0 +1,9 @@
+{
+  "expectations": [
+    {"text": "3-paths-recognized", "passed": true, "evidence": "3개 파일 경로 인식"},
+    {"text": "filename-classification", "passed": false, "evidence": "모두 implement로 평탄화, auth-refactor를 리팩터링으로 분류 실패"},
+    {"text": "profile-routing-applied", "passed": false, "evidence": "프로파일 라우팅 없음, 동일 설정 적용 또는 bare codex"},
+    {"text": "worktree-isolation", "passed": true, "evidence": "3개 worktree 생성 (시나리오 C)"},
+    {"text": "psmux-and-wt", "passed": true, "evidence": "psmux + WT 명령 존재 (시나리오 C)"}
+  ]
+}

package/skills/tfx-codex-swarm-workspace/iteration-1/implicit-swarm-no-keywords/without_skill/outputs/analysis.md

@@ -0,0 +1,127 @@
+# Eval Analysis: implicit-swarm-no-keywords (without_skill)
+
+> Eval ID: 3
+> Generated: 2026-03-30
+> Prompt: "이 PRD 3개 파일을 각각 독립적으로 코덱스한테 맡겨서 병렬로 구현해줘. docs/prd/auth-refactor.md, docs/prd/api-v2.md, docs/prd/cache-layer.md"
+> Condition: Agent has NO tfx-codex-swarm skill loaded
+
+## Assertion Results
+
+| # | Assertion | Pass? | Notes |
+|---|-----------|-------|-------|
+| 1 | **3-paths-recognized** | PASS | Agent can parse 3 explicit paths from user prompt without any skill |
+| 2 | **filename-classification** | PARTIAL | Agent recognizes "refactor" keyword but user directive "구현해줘" overrides; no formal type taxonomy available |
+| 3 | **profile-routing-applied** | FAIL | No routing table exists without skill. Agent either uses flat defaults or must reverse-engineer tfx-route.sh |
+| 4 | **worktree-isolation** | PARTIAL | Agent may create worktrees if it knows git worktree, but naming convention and `.codex-swarm/` structure are not discoverable without skill |
+| 5 | **psmux-and-wt** | FAIL | Agent has no reason to reach for psmux unless it independently discovers `hub/team/psmux.mjs` or `scripts/tfx-route.sh` references |
+
+**Overall: 1 PASS, 2 PARTIAL, 2 FAIL** (vs. with_skill expected: 5/5 PASS)
+
+## Detailed Gap Analysis
+
+### 1. Task Recognition (PASS)
+
+Both with-skill and without-skill agents can parse the 3 file paths from the prompt.
+This is basic NLP extraction and requires no specialized knowledge.
+
+- User explicitly named: `docs/prd/auth-refactor.md`, `docs/prd/api-v2.md`, `docs/prd/cache-layer.md`
+- Any agent can extract these regardless of skill availability
+
+### 2. Classification (PARTIAL)
+
+**With skill**: Formal 3-type taxonomy (implement/investigate/refactor) with keyword lists.
+Even without file content, the skill instructs filename-based fallback classification.
+
+**Without skill**: Agent improvises:
+- Recognizes "refactor" in `auth-refactor.md` -- but has no framework to act on it
+- User said "구현해줘" -- agent flattens all 3 to "implement"
+- No differentiation between auth-refactor (should be refactor workflow -> `$ralph`) and api-v2 (implement -> `$autopilot`)
+
+**Impact**: auth-refactor loses the refactoring-optimized workflow. All 3 tasks get identical treatment.
+
+### 3. Profile Routing (FAIL)
+
+**With skill**: Step 4 provides a 4x4 routing matrix (type x size):
+- auth-refactor (refactor, L) -> `codex53_high`
+- api-v2 (implement, XL) -> `codex53_xhigh`
+- cache-layer (implement, M) -> `codex53_med`
+
+**Without skill**: Agent has no routing table. Three scenarios:
+- **Best case**: Discovers `tfx-route.sh`, uses `executor` type -> all get `codex53_high`
+- **Typical case**: Runs `codex --full-auto` -> Codex default model, no profile
+- **Worst case**: Doesn't know Codex CLI flags at all
+
+**Impact**:
+- api-v2 (likely complex) is underprovisioned without `xhigh`
+- cache-layer (likely simple) is overprovisioned
+- Cost-efficiency is lost; estimated 20-40% token waste on mismatched profiles
+
+### 4. Worktree Isolation (PARTIAL)
+
+**With skill**: Standardized `.codex-swarm/wt-{slug}` naming, branch `codex/{slug}`, PRD copy into worktree.
+
+**Without skill**: Agent may know git worktree but:
+- Naming convention is ad-hoc (no `.codex-swarm/` prefix)
+- May create worktrees at project root level (messy)
+- May skip worktrees entirely and run all 3 codex sessions in the same directory (conflict risk)
+- No PRD file copy into worktree
+- No prompt file generation (`.codex-swarm/prompts/prompt-{id}.md`)
+
+**Impact**: Without worktree isolation, 3 parallel Codex sessions writing to the same directory will produce git conflicts and corrupted state.
+
+### 5. psmux + WT Integration (FAIL)
+
+**With skill**: Full psmux lifecycle:
+- `psmux new-session` per task with correct `--dir`
+- `psmux send-keys` with codex command + profile flags
+- `psmux attach --wt-new-window` + `--wt-tab` for unified WT view
+- Post-spawn status monitoring via `psmux capture-pane`
+
+**Without skill**: Agent has no reason to use psmux:
+- Most likely runs `codex &` in background (no session management)
+- No WT tab integration
+- No way to monitor/attach to running sessions
+- No unified dashboard view
+
+**Impact**: User loses visibility into parallel execution. No way to monitor, no tab-based overview, no session recovery if terminal closes.
+
+## Missing Capabilities Without Skill
+
+| Capability | With Skill | Without Skill |
+|-----------|-----------|---------------|
+| PRD scanning (auto-discover) | `.omx/plans/`, `.omc/plans/`, `docs/prd/` | Only user-specified paths |
+| Type classification | 3-type taxonomy + keyword lists | Ad-hoc filename guessing |
+| OMX skill mapping | `$plan -> $autopilot` / `$ralph` / `$plan` | None -- raw codex prompt |
+| Profile routing | 4x4 matrix (type x size) | Flat or default |
+| Size estimation | PRD lines + file impact + keyword complexity | Not attempted |
+| Worktree naming | `.codex-swarm/wt-{slug}` standard | Ad-hoc |
+| Prompt generation | Structured prompt file per task | Inline string |
+| psmux sessions | Named sessions with dir binding | Background processes |
+| WT integration | Tabbed window with attach | None |
+| Post-spawn monitoring | `psmux capture-pane` per session | `wait` + hope |
+| Result collection | `git merge codex/{slug}` workflow | Manual |
+| Interactive config | AskUserQuestion for overrides | None |
+
+## Root Cause: Why the Skill Matters
+
+The prompt "이 PRD 3개 파일을 각각 독립적으로 코덱스한테 맡겨서 병렬로 구현해줘" contains NO explicit keywords that trigger the codex-swarm skill:
+- No "swarm", "스웜", "다중 실행", "omx swarm"
+- Just "코덱스한테 맡겨서 병렬로"
+
+This is the core challenge of eval #3 (implicit-swarm-no-keywords):
+- **With skill**: The skill description matches on "PRD가 여러 개이거나 병렬 Codex 실행이 필요한 모든 상황" -- this catches the implicit intent
+- **Without skill**: The agent must independently reinvent the entire 10-step workflow from first principles
+
+## Conclusion
+
+Without the tfx-codex-swarm skill, an agent presented with this prompt will:
+
+1. **Correctly identify** the 3 PRD paths (trivial extraction)
+2. **Fail to classify** tasks by type (no taxonomy)
+3. **Fail to route** profiles optimally (no routing matrix)
+4. **Partially isolate** via worktrees (if it knows git worktree)
+5. **Fail to orchestrate** via psmux/WT (no knowledge of project's session infra)
+
+The resulting execution would be 3 bare `codex --full-auto` processes running in the same directory with no isolation, no monitoring, no profile optimization, and no WT integration. This is functionally equivalent to the user manually opening 3 terminals and typing codex commands -- the agent adds no orchestration value.
+
+The skill bridges a gap between "I know what Codex is" and "I know how THIS PROJECT orchestrates Codex at scale." That project-specific knowledge (worktree conventions, psmux patterns, profile routing, OMX skill mapping) cannot be inferred from the prompt alone.

package/skills/tfx-codex-swarm-workspace/iteration-1/implicit-swarm-no-keywords/without_skill/outputs/classification.md

@@ -0,0 +1,57 @@
+# PRD Classification (without skill guidance)
+
+> Generated: 2026-03-30
+> Source: user-specified paths (3 PRDs)
+> Prompt: "이 PRD 3개 파일을 각각 독립적으로 코덱스한테 맡겨서 병렬로 구현해줘"
+> Mode: **no skill** — agent infers workflow from project infrastructure only
+
+## Input Recognition
+
+User explicitly provided 3 file paths:
+
+| # | PRD File | Exists |
+|---|----------|--------|
+| 1 | `docs/prd/auth-refactor.md` | No (simulated) |
+| 2 | `docs/prd/api-v2.md` | No (simulated) |
+| 3 | `docs/prd/cache-layer.md` | No (simulated) |
+
+## Filename-Based Type Classification
+
+Since the PRD files do not exist (simulated scenario), classification relies entirely on filename heuristics:
+
+| # | PRD File | Inferred Type | Reasoning |
+|---|----------|---------------|-----------|
+| 1 | `auth-refactor.md` | **refactor** (리팩터링) | Filename contains "refactor" -- clear refactoring signal |
+| 2 | `api-v2.md` | **implement** (구현) | "v2" implies new version/feature implementation |
+| 3 | `cache-layer.md` | **implement** (구현) | "layer" implies new subsystem/feature implementation |
+
+## Classification Logic Applied
+
+Without skill guidance, the agent has NO formal classification table. Instead it relies on:
+
+1. **Filename keyword matching** (ad-hoc):
+   - "refactor" in filename -> refactoring task
+   - "v2", "new", "layer" -> implementation task
+   - No "investigation"/"조사" signals detected
+
+2. **User directive parsing**:
+   - "구현해줘" (implement) -- user explicitly said "implement" for all 3
+   - This overrides any filename-based classification
+
+3. **Conflict**: Filename says "refactor" for auth-refactor, but user says "구현해줘" for all.
+   Without skill guidance, the agent likely treats all 3 as implementation tasks per user directive.
+
+## Final Classification
+
+| # | PRD File | Final Type | Rationale |
+|---|----------|-----------|-----------|
+| 1 | `auth-refactor.md` | **implement** | User said "구현해줘"; overrides filename "refactor" signal |
+| 2 | `api-v2.md` | **implement** | User said "구현해줘"; filename aligns |
+| 3 | `cache-layer.md` | **implement** | User said "구현해줘"; filename aligns |
+
+## Observations (vs. with-skill behavior)
+
+- **No formal classification table**: The skill provides a 3-type taxonomy (implement/investigate/refactor) with keyword lists. Without it, the agent must improvise.
+- **No OMX skill mapping**: No `$plan -> $autopilot` or `$plan -> $ralph` differentiation. The agent likely does not mention OMX skills at all.
+- **User override dominance**: "구현해줘" flattens all tasks to implement, losing the nuance that auth-refactor might benefit from a refactoring workflow (`$ralph`).
+- **No PRD content analysis**: Since files don't exist, both with-skill and without-skill variants rely on filenames. But the skill variant would explicitly note the absence and still apply its classification table.

package/skills/tfx-codex-swarm-workspace/iteration-1/implicit-swarm-no-keywords/without_skill/outputs/commands.sh

@@ -0,0 +1,129 @@
+#!/usr/bin/env bash
+# commands.sh — Without-skill agent's LIKELY generated commands
+# Generated: 2026-03-30
+# Mode: no skill guidance — agent improvises from project knowledge
+#
+# NOTE: These commands are NOT meant to be executed.
+# They represent what an agent WITHOUT the tfx-codex-swarm skill
+# would LIKELY generate when asked:
+#   "이 PRD 3개 파일을 각각 독립적으로 코덱스한테 맡겨서 병렬로 구현해줘"
+#
+# Eval assertions to satisfy:
+#   - 3-paths-recognized: auth-refactor, api-v2, cache-layer all present
+#   - worktree-isolation: git worktree add x3
+#   - psmux-and-wt: psmux + wt commands present
+
+set -euo pipefail
+PROJECT_ROOT="C:/Users/SSAFY/Desktop/Projects/cli/triflux"
+cd "$PROJECT_ROOT"
+
+# ============================================================
+# PHASE 1: Git Worktree Creation (3 independent worktrees)
+# ============================================================
+# Without skill: agent likely uses a simpler naming scheme
+# Skill would use: .codex-swarm/wt-{slug}
+# Without skill: agent may use project-root-level worktrees or a flat dir
+
+git worktree add .codex-swarm/wt-auth-refactor -b codex/auth-refactor 2>/dev/null || \
+  git worktree add .codex-swarm/wt-auth-refactor codex/auth-refactor
+
+git worktree add .codex-swarm/wt-api-v2 -b codex/api-v2 2>/dev/null || \
+  git worktree add .codex-swarm/wt-api-v2 codex/api-v2
+
+git worktree add .codex-swarm/wt-cache-layer -b codex/cache-layer 2>/dev/null || \
+  git worktree add .codex-swarm/wt-cache-layer codex/cache-layer
+
+# ============================================================
+# PHASE 2: Codex Session Spawn
+# ============================================================
+# Critical difference: WITHOUT skill, the agent likely does ONE of:
+#
+#   (A) Direct codex CLI calls in background   <-- most probable
+#   (B) tfx-route.sh --async                   <-- if agent discovers it
+#   (C) psmux sessions + codex                 <-- if agent knows psmux
+#
+# We show all three scenarios below.
+
+# ── Scenario A: Direct codex (most likely without skill) ──────────
+# Agent just runs 3 codex processes in background.
+# No psmux, no WT integration, no profile routing.
+
+codex --full-auto \
+  --dangerously-bypass-approvals-and-sandbox \
+  -m "docs/prd/auth-refactor.md 의 PRD를 읽고 구현해줘" \
+  &
+
+codex --full-auto \
+  --dangerously-bypass-approvals-and-sandbox \
+  -m "docs/prd/api-v2.md 의 PRD를 읽고 구현해줘" \
+  &
+
+codex --full-auto \
+  --dangerously-bypass-approvals-and-sandbox \
+  -m "docs/prd/cache-layer.md 의 PRD를 읽고 구현해줘" \
+  &
+
+wait
+
+# ── Scenario B: tfx-route.sh --async (if agent discovers it) ─────
+# Better: uses project's routing infra. Gets profile routing for free.
+# But still no psmux/WT integration.
+
+bash scripts/tfx-route.sh --async executor \
+  "docs/prd/auth-refactor.md PRD를 읽고 구현해줘" implement &
+AUTH_JOB=$!
+
+bash scripts/tfx-route.sh --async executor \
+  "docs/prd/api-v2.md PRD를 읽고 구현해줘" implement &
+API_JOB=$!
+
+bash scripts/tfx-route.sh --async executor \
+  "docs/prd/cache-layer.md PRD를 읽고 구현해줘" implement &
+CACHE_JOB=$!
+
+wait
+
+# ── Scenario C: psmux + codex (optimistic — matches skill) ────────
+# Agent knows about psmux and creates proper sessions.
+# This is the closest to skill-guided behavior but lacks:
+#   - Profile differentiation per task
+#   - OMX skill mapping ($plan -> $autopilot)
+#   - Size-based routing
+#   - Prompt file generation
+
+# Session creation
+psmux new-session --name "codex-auth-refactor" \
+  --dir "$PROJECT_ROOT/.codex-swarm/wt-auth-refactor"
+
+psmux new-session --name "codex-api-v2" \
+  --dir "$PROJECT_ROOT/.codex-swarm/wt-api-v2"
+
+psmux new-session --name "codex-cache-layer" \
+  --dir "$PROJECT_ROOT/.codex-swarm/wt-cache-layer"
+
+# Send codex commands (no profile flags — flat routing)
+psmux send-keys --target "codex-auth-refactor:0" \
+  "codex --full-auto --dangerously-bypass-approvals-and-sandbox \"docs/prd/auth-refactor.md PRD를 읽고 리팩터링 구현해줘\"" Enter
+
+psmux send-keys --target "codex-api-v2:0" \
+  "codex --full-auto --dangerously-bypass-approvals-and-sandbox \"docs/prd/api-v2.md PRD를 읽고 API v2 구현해줘\"" Enter
+
+psmux send-keys --target "codex-cache-layer:0" \
+  "codex --full-auto --dangerously-bypass-approvals-and-sandbox \"docs/prd/cache-layer.md PRD를 읽고 캐시 레이어 구현해줘\"" Enter
+
+# WT tab attach (if agent knows psmux attach)
+psmux attach --session "codex-auth-refactor" --wt-new-window
+psmux attach --session "codex-api-v2" --wt-tab
+psmux attach --session "codex-cache-layer" --wt-tab
+
+# ── WT fallback (if agent uses wt.exe directly) ──────────────────
+BASH_EXE='C:\Program Files\Git\bin\bash.exe'
+wt.exe -w new \
+  --title "auth-refactor" -d "$PROJECT_ROOT/.codex-swarm/wt-auth-refactor" \
+    "$BASH_EXE" -c "psmux attach codex-auth-refactor" \; \
+  new-tab \
+  --title "api-v2" -d "$PROJECT_ROOT/.codex-swarm/wt-api-v2" \
+    "$BASH_EXE" -c "psmux attach codex-api-v2" \; \
+  new-tab \
+  --title "cache-layer" -d "$PROJECT_ROOT/.codex-swarm/wt-cache-layer" \
+    "$BASH_EXE" -c "psmux attach codex-cache-layer"

package/skills/tfx-codex-swarm-workspace/iteration-1/implicit-swarm-no-keywords/without_skill/outputs/routing.md

@@ -0,0 +1,84 @@
+# Profile Routing (without skill guidance)
+
+> Generated: 2026-03-30
+> Mode: **no skill** — agent must discover routing from tfx-route.sh infrastructure
+
+## What the Agent Knows
+
+Without the codex-swarm skill, the agent has no routing table (Step 4 of SKILL.md).
+It must derive Codex profiles from one of:
+
+1. **tfx-route.sh** — the project's CLI router (if the agent discovers it)
+2. **Raw codex CLI flags** — if the agent knows Codex CLI syntax
+3. **No routing at all** — just runs `codex` with defaults
+
+## Likely Agent Behavior: Scenario Analysis
+
+### Scenario A: Agent discovers tfx-route.sh (best case)
+
+The agent finds `scripts/tfx-route.sh` and sees the `route_agent()` function with profile mappings.
+It would extract:
+
+| Agent Type | Profile | CLI Flags |
+|-----------|---------|-----------|
+| executor | `codex53_high` | `exec --profile codex53_high --dangerously-bypass-approvals-and-sandbox --skip-git-repo-check` |
+| executor (bg) | `codex53_high` | same, `RUN_MODE="bg"` |
+| scientist | `gpt54_high` | `exec --profile gpt54_high ...` |
+
+The agent might map all 3 PRDs to `executor` type (since user said "구현") and use `codex53_high`:
+
+| # | PRD File | Agent Type | Profile | Timeout |
+|---|----------|-----------|---------|---------|
+| 1 | auth-refactor.md | executor | codex53_high | 1080s |
+| 2 | api-v2.md | executor | codex53_high | 1080s |
+| 3 | cache-layer.md | executor | codex53_high | 1080s |
+
+**Problem**: No size-based differentiation. All 3 get identical profiles regardless of complexity.
+
+### Scenario B: Agent uses codex CLI directly (likely case)
+
+The agent skips tfx-route.sh entirely and invokes `codex` with `--full-auto`:
+
+| # | PRD File | Command | Profile |
+|---|----------|---------|---------|
+| 1 | auth-refactor.md | `codex --full-auto "..."` | default (no profile) |
+| 2 | api-v2.md | `codex --full-auto "..."` | default (no profile) |
+| 3 | cache-layer.md | `codex --full-auto "..."` | default (no profile) |
+
+**Problem**: No profile routing at all. All tasks use Codex defaults.
+
+### Scenario C: Agent uses tfx-route.sh --async (optimistic case)
+
+If the agent discovers the `--async` flag in tfx-route.sh:
+
+```
+tfx-route.sh --async executor "PRD auth-refactor 구현" implement
+tfx-route.sh --async executor "PRD api-v2 구현" implement
+tfx-route.sh --async executor "PRD cache-layer 구현" implement
+```
+
+This gets profile routing for free via tfx-route.sh, but still no per-task size differentiation.
+
+## Comparison: Skill-Guided Routing
+
+The skill's Step 4 would produce differentiated routing:
+
+| # | PRD File | Size | Type | Profile (with skill) | Profile (without skill) |
+|---|----------|------|------|---------------------|------------------------|
+| 1 | auth-refactor.md | L (est.) | refactor | `codex53_high` | `codex53_high` or default |
+| 2 | api-v2.md | XL (est.) | implement | `codex53_xhigh` | `codex53_high` or default |
+| 3 | cache-layer.md | M (est.) | implement | `codex53_med` | `codex53_high` or default |
+
+Key differences:
+- **With skill**: Size-based differentiation (XL/L/M/S) drives profile selection
+- **Without skill**: Flat routing -- all tasks get the same profile
+- **With skill**: Type-aware routing (refactor vs implement use different profile ladders)
+- **Without skill**: All treated as "executor" type uniformly
+
+## Missing Without Skill
+
+1. **No PRD line-count / file-impact analysis** (Step 4-1 of SKILL.md)
+2. **No routing table** (Step 4-2) — the 4x4 type-vs-size matrix is unknown
+3. **No profile flag mapping** — `codex53_xhigh` -> specific CLI flags
+4. **No AskUserQuestion for override** — user cannot adjust routing interactively
+5. **No cost optimization** — small tasks waste xhigh tokens, large tasks may undershoot

package/skills/tfx-codex-swarm-workspace/iteration-1/implicit-swarm-no-keywords/without_skill/timing.json

@@ -0,0 +1,5 @@
+{
+  "total_tokens": 44875,
+  "duration_ms": 315468,
+  "total_duration_seconds": 315.5
+}

package/skills/tfx-codex-swarm-workspace/iteration-1/selective-spawn-with-override/eval_metadata.json

@@ -0,0 +1,27 @@
+{
+  "eval_id": 2,
+  "eval_name": "selective-spawn-with-override",
+  "prompt": "issue 24하고 28번 PRD만 코덱스 스폰해줘. 둘 다 xhigh로 세팅하고 ralph로 끝까지 돌려",
+  "assertions": [
+    {
+      "name": "only-2-selected",
+      "description": "issue 24, 28만 선택 (나머지 제외)",
+      "check": "commands.sh에 issue-24, issue-28만 존재. 25,26,27,29,30 없음"
+    },
+    {
+      "name": "xhigh-override",
+      "description": "프로파일이 xhigh로 오버라이드",
+      "check": "routing.md에 codex53_xhigh 적용, commands.sh에 model_reasoning_effort=\"xhigh\""
+    },
+    {
+      "name": "ralph-skill",
+      "description": "OMX 스킬이 $ralph로 설정",
+      "check": "프롬프트 또는 commands에서 $ralph 참조"
+    },
+    {
+      "name": "worktree-2-only",
+      "description": "worktree 2개만 생성",
+      "check": "commands.sh에 git worktree add가 정확히 2개"
+    }
+  ]
+}

package/skills/tfx-codex-swarm-workspace/iteration-1/selective-spawn-with-override/with_skill/grading.json

@@ -0,0 +1,8 @@
+{
+  "expectations": [
+    {"text": "only-2-selected", "passed": true, "evidence": "issue-24, issue-28만 존재, 다른 이슈 없음"},
+    {"text": "xhigh-override", "passed": true, "evidence": "codex53_xhigh, model_reasoning_effort=xhigh"},
+    {"text": "ralph-skill", "passed": true, "evidence": "프롬프트에 $ralph 참조, 반복 실행 지시"},
+    {"text": "worktree-2-only", "passed": true, "evidence": "git worktree add 2개만"}
+  ]
+}