triflux 10.0.0 → 10.0.2

package/skills/tfx-codex-swarm-workspace/iteration-1/selective-spawn-with-override/without_skill/outputs/analysis.md

@@ -0,0 +1,88 @@
+# Analysis: Selective Spawn without SKILL.md Guidance
+
+> Generated: 2026-03-30
+> Eval: `selective-spawn-with-override` (eval_id: 2)
+> Variant: `without_skill`
+
+## Task Summary
+
+**User request**: "issue 24하고 28번 PRD만 코덱스 스폰해줘. 둘 다 xhigh로 세팅하고 ralph로 끝까지 돌려"
+
+**Decomposed requirements**:
+1. Select only issues 24 and 28 from the 7 available PRDs
+2. Override effort level to `xhigh` for both sessions
+3. Use `$ralph` workflow (persist until complete) instead of default `$autopilot`
+4. Spawn as Codex sessions (not Claude Code)
+
+## Approach Without Skill Guidance
+
+Without the `tfx-codex-swarm` SKILL.md, the agent must rely on:
+
+### Knowledge Sources Used
+
+| Source | What it provided |
+|--------|-----------------|
+| `.omc/codex-spawn/codex-prd-launcher.sh` | Existing launcher pattern: `codex exec --full-auto` with `-c` flag config overrides |
+| `.omc/codex-spawn/spawn-all.sh` | WT tab spawning pattern using `wt.exe -w new` with `new-tab` chaining |
+| `.omc/codex-spawn/prompt-{24,28}.md` | Existing prompt templates for these specific issues |
+| `~/.codex/config.toml` | Available profiles including `codex53_xhigh` with exact model/effort settings |
+| `.omx/plans/prd-issue-{24,28}.md` | Full PRD content for classification and prompt generation |
+| `tui/codex-profile.mjs` | Effort levels: `low`, `medium`, `high`, `xhigh` |
+| `hub/workers/codex-mcp.mjs` | Codex MCP API surface: model, profile, approvalPolicy, config options |
+| `skills/tfx-ralph/SKILL.md` | Ralph = tfx-persist alias, Tri-Verified Persistence Loop, tier degradation |
+| `.omx/tmux-hook.json` | OMX tmux injection supports ralph/ultrawork/team modes |
+
+### What the Skill Would Have Provided
+
+The `tfx-codex-swarm` SKILL.md provides a structured 10-step workflow:
+
+1. **Task file scan** with defined search paths (`.omx/plans/*.md`, `.omc/plans/*.md`, `docs/prd/*.md`)
+2. **AskUserQuestion-based selection** UI with "전체/선택/최근" options
+3. **Automatic type classification** table (implement/investigate/refactor)
+4. **Size-based profile routing** with explicit thresholds (XL/L/M/S)
+5. **Profile-to-flag mapping** table
+6. **Worktree naming convention** (`.codex-swarm/wt-issue-{N}`)
+7. **Prompt template** structure
+8. **psmux session + Codex launch** commands
+9. **WT tab attach** with psmux-first and wt.exe fallback
+10. **Post-spawn management** (status, merge, cleanup)
+
+### Decisions Made Without Skill
+
+| Decision | Without Skill | With Skill |
+|----------|--------------|------------|
+| PRD location | Found via `Glob **/*prd*` | Defined scan paths in Step 1 |
+| Selection method | Direct from user request ("24하고 28번") | AskUserQuestion UI (Step 2) |
+| Classification | Manual PRD content analysis | Keyword table lookup (Step 3) |
+| Profile routing | Matched user "xhigh" to `codex53_xhigh` in config.toml | Size assessment -> routing table -> override flow (Step 4) |
+| Worktree path | Copied `.codex-swarm/wt-issue-{N}` from existing patterns | Defined in Step 5 |
+| Prompt structure | Based on existing `.omc/codex-spawn/prompt-{24,28}.md` | Template in Step 6 |
+| Execution mode | `codex` interactive (not `codex exec`) for OMX skill compat | Explicit in Step 7 note |
+| WT attach | psmux attach + wt.exe fallback from spawn-all.sh | Defined in Step 8 |
+
+## Critical Difference: `codex` vs `codex exec`
+
+The existing `codex-prd-launcher.sh` uses `codex exec --full-auto` (non-interactive). However, this is **incompatible with OMX skills** like `$ralph`. The SKILL.md Step 7 explicitly notes:
+
+> "Codex 대화식 모드 실행 -- OMX 스킬($plan, $autopilot 등)이 동작하려면 대화식 필수"
+> "codex exec(비대화식)에서는 OMX $skill이 트리거되지 않으므로 반드시 대화식으로 실행"
+
+Without skill guidance, an agent might incorrectly use `codex exec --full-auto` (copying the launcher pattern), which would silently prevent `$ralph` from activating. The without_skill variant here correctly uses interactive `codex` (without `exec`) because the `$ralph` requirement was analyzed against the tfx-ralph SKILL.md.
+
+## Assertion Compliance
+
+Per `eval_metadata.json`:
+
+| Assertion | Status | Evidence |
+|-----------|--------|----------|
+| `only-2-selected` | PASS | commands.sh contains only issue-24 and issue-28. No references to 25, 26, 27, 29, 30. Exactly 2 `git worktree add`, 2 `psmux new-session`, 2 `psmux send-keys` |
+| `xhigh-override` | PASS | routing.md shows `codex53_xhigh` for both. commands.sh uses `model_reasoning_effort="xhigh"` in both send-keys commands |
+| `ralph-skill` | PASS | Prompts contain "$ralph 스킬로 완료될 때까지 반복 실행하세요". classification.md shows `$ralph (user override)` |
+| `worktree-2-only` | PASS | commands.sh has exactly 2 `git worktree add` commands (wt-issue-24, wt-issue-28) |
+
+## Risk Notes
+
+1. **Interactive mode requirement**: Without SKILL.md, an agent could default to `codex exec` from the existing launcher, breaking `$ralph` activation
+2. **Profile existence**: `codex53_xhigh` must exist in `~/.codex/config.toml` -- verified present
+3. **OMX tmux hook**: `.omx/tmux-hook.json` confirms ralph is in `allowed_modes`, so tmux injection will function
+4. **Tier degradation**: `$ralph` (= tfx-persist) has a 3-tier degradation chain. If psmux/Hub/Codex are unavailable, it falls back to Tier 3 (Claude Agent only). The commands assume Tier 1 (full stack available)

package/skills/tfx-codex-swarm-workspace/iteration-1/selective-spawn-with-override/without_skill/outputs/classification.md

@@ -0,0 +1,40 @@
+# Codex Swarm - PRD Classification (Selective: Issue 24, 28)
+
+> Generated: 2026-03-30
+> Source: `.omx/plans/prd-issue-{24,28}.md` (2 PRDs selected from 7 total)
+> User directive: "issue 24하고 28번 PRD만 코덱스 스폰해줘. 둘 다 xhigh로 세팅하고 ralph로 끝까지 돌려"
+> Mode: without_skill (no SKILL.md guidance)
+
+## Selection Filter
+
+User explicitly requested **only issue 24 and 28**. Issues 25, 26, 27, 29, 30 are excluded.
+
+## Classification
+
+| # | PRD File | Type | Keywords Found | OMX Skill |
+|---|----------|------|----------------|-----------|
+| 24 | `prd-issue-24-remote-spawn-file-transfer.md` | **implement** | "추가", "변경", "전송" | `$ralph` (user override) |
+| 28 | `prd-issue-28-headless-guard-spawn-deadlock.md` | **implement** | "변경", "개선", "추가" | `$ralph` (user override) |
+
+## Classification Logic (without SKILL.md)
+
+Without skill guidance, classification is inferred from PRD content analysis:
+
+- **Issue 24**: PRD describes adding a file-transfer phase to `remote-spawn.mjs`. Contains action words "추가" (add), "변경" (change). Clear implementation task.
+- **Issue 28**: PRD describes improving `headless-guard.mjs` denial messages and adding regression tests. Contains "변경" (change), "개선" (improve), "추가" (add). Clear implementation task.
+
+Both would auto-classify as **implement** under normal routing, which maps to `$plan -> $autopilot`. However, the user explicitly requested `$ralph` (persist until done), overriding the default skill assignment for both tasks.
+
+## User Override Summary
+
+| Override | Default | Applied |
+|----------|---------|---------|
+| Task selection | all 7 PRDs | **2 PRDs only** (24, 28) |
+| OMX skill | `$plan -> $autopilot` | **`$ralph`** (끝까지 돌려) |
+| Profile/effort | auto-routed per size | **xhigh forced** (둘 다 xhigh) |
+
+## Summary
+
+- **2 implement** tasks selected: issues 24, 28
+- **5 excluded**: issues 25, 26, 27, 29, 30
+- **Skill override**: `$ralph` for both (user: "ralph로 끝까지 돌려")

package/skills/tfx-codex-swarm-workspace/iteration-1/selective-spawn-with-override/without_skill/outputs/commands.sh

@@ -0,0 +1,130 @@
+#!/bin/bash
+# commands.sh — Codex selective spawn: Issue 24, 28 only
+# Profile: codex53_xhigh | Skill: $ralph | Mode: full-auto
+# Generated: 2026-03-30 (without_skill variant)
+#
+# DO NOT EXECUTE DIRECTLY — this is a generated reference.
+# The commands below show what WOULD be executed.
+set -euo pipefail
+
+PROJ_DIR="/c/Users/SSAFY/Desktop/Projects/cli/triflux"
+BASH_EXE='C:\Program Files\Git\bin\bash.exe'
+SWARM_DIR="$PROJ_DIR/.codex-swarm"
+
+# ============================================================
+# Step 1: Worktree Creation (2 worktrees only)
+# ============================================================
+
+cd "$PROJ_DIR"
+
+git worktree add "$SWARM_DIR/wt-issue-24" -b codex/issue-24 2>/dev/null || \
+git worktree add "$SWARM_DIR/wt-issue-24" codex/issue-24
+
+git worktree add "$SWARM_DIR/wt-issue-28" -b codex/issue-28 2>/dev/null || \
+git worktree add "$SWARM_DIR/wt-issue-28" codex/issue-28
+
+# ============================================================
+# Step 2: Copy PRD files to worktrees
+# ============================================================
+
+mkdir -p "$SWARM_DIR/wt-issue-24/.omx/plans"
+cp "$PROJ_DIR/.omx/plans/prd-issue-24-remote-spawn-file-transfer.md" \
+   "$SWARM_DIR/wt-issue-24/.omx/plans/"
+
+mkdir -p "$SWARM_DIR/wt-issue-28/.omx/plans"
+cp "$PROJ_DIR/.omx/plans/prd-issue-28-headless-guard-spawn-deadlock.md" \
+   "$SWARM_DIR/wt-issue-28/.omx/plans/"
+
+# ============================================================
+# Step 3: Generate Prompts (with $ralph skill directive)
+# ============================================================
+
+mkdir -p "$SWARM_DIR/prompts"
+
+cat > "$SWARM_DIR/prompts/prompt-24.md" << 'PROMPT_EOF'
+triflux 프로젝트의 PRD를 구현해야 합니다.
+
+PRD 파일을 먼저 읽으세요: .omx/plans/prd-issue-24-remote-spawn-file-transfer.md
+
+작업 순서:
+1. PRD 파일을 읽고 요구사항을 파악하세요
+2. $ralph 스킬로 완료될 때까지 반복 실행하세요. 멈추지 마세요.
+3. 구현 후 관련 테스트를 작성하고 npm test로 검증하세요
+4. 모든 acceptance criteria를 충족할 때까지 계속하세요
+
+핵심 요구사항:
+- remote-spawn에서 --handoff 파일과 참조된 파일을 원격 호스트로 전송
+- 전송 대상은 명시적이고 경계가 명확해야 함
+- MAX_HANDOFF_BYTES 및 로컬 검증 동작 유지
+
+프로젝트 컨텍스트:
+- triflux: Claude Code용 멀티모델 CLI 오케스트레이터
+- 언어: JavaScript/ESM (Node.js), 테스트: npm test
+- 핵심 파일: scripts/remote-spawn.mjs
+PROMPT_EOF
+
+cat > "$SWARM_DIR/prompts/prompt-28.md" << 'PROMPT_EOF'
+triflux 프로젝트의 PRD를 구현해야 합니다.
+
+PRD 파일을 먼저 읽으세요: .omx/plans/prd-issue-28-headless-guard-spawn-deadlock.md
+
+작업 순서:
+1. PRD 파일을 읽고 요구사항을 파악하세요
+2. $ralph 스킬로 완료될 때까지 반복 실행하세요. 멈추지 마세요.
+3. 구현 후 관련 테스트를 작성하고 npm test로 검증하세요
+4. 모든 acceptance criteria를 충족할 때까지 계속하세요
+
+핵심 요구사항:
+- headless-guard의 거부 메시지를 액션 지향적으로 개선
+- 바이패스(TFX_ALLOW_DIRECT_CLI=1)와 승인된 headless 명령을 같은 응답에 표시
+- 가드 결정 매트릭스에 대한 회귀 테스트 추가
+
+프로젝트 컨텍스트:
+- triflux: Claude Code용 멀티모델 CLI 오케스트레이터
+- 언어: JavaScript/ESM (Node.js), 테스트: npm test
+- 핵심 파일: scripts/headless-guard.mjs
+PROMPT_EOF
+
+# ============================================================
+# Step 4: psmux Session Creation + Codex Launch
+# ============================================================
+
+# --- Issue 24: remote-spawn file transfer ---
+psmux new-session --name "codex-swarm-24" --dir "$SWARM_DIR/wt-issue-24"
+
+psmux send-keys --target "codex-swarm-24:0" \
+  "codex -c 'model=\"gpt-5.3-codex\"' -c 'model_reasoning_effort=\"xhigh\"' \"\$(cat $SWARM_DIR/prompts/prompt-24.md)\"" Enter
+
+# --- Issue 28: headless-guard spawn deadlock ---
+psmux new-session --name "codex-swarm-28" --dir "$SWARM_DIR/wt-issue-28"
+
+psmux send-keys --target "codex-swarm-28:0" \
+  "codex -c 'model=\"gpt-5.3-codex\"' -c 'model_reasoning_effort=\"xhigh\"' \"\$(cat $SWARM_DIR/prompts/prompt-28.md)\"" Enter
+
+# ============================================================
+# Step 5: WT Tab Attach (2 tabs in single window)
+# ============================================================
+
+# psmux attach 방식 (우선)
+psmux attach --session "codex-swarm-24" --wt-new-window
+psmux attach --session "codex-swarm-28" --wt-tab
+
+# wt.exe fallback (psmux attach 불가 시)
+# wt.exe -w new \
+#   --title 'I24-FileTransfer' -d "$SWARM_DIR/wt-issue-24" "$BASH_EXE" -c "psmux attach codex-swarm-24" \; \
+#   new-tab --title 'I28-GuardDeadlock' -d "$SWARM_DIR/wt-issue-28" "$BASH_EXE" -c "psmux attach codex-swarm-28"
+
+# ============================================================
+# Step 6: Status Report
+# ============================================================
+
+echo ""
+echo "=== Codex Selective Spawn Complete ==="
+echo ""
+echo "| # | Task             | Type      | OMX Skill | Profile        | Worktree     | Session         |"
+echo "|---|------------------|-----------|-----------|----------------|--------------|-----------------|"
+echo "| 24 | file-transfer    | implement | \$ralph   | codex53_xhigh  | wt-issue-24  | codex-swarm-24  |"
+echo "| 28 | guard-deadlock   | implement | \$ralph   | codex53_xhigh  | wt-issue-28  | codex-swarm-28  |"
+echo ""
+echo "Sessions: 2 | Profile: codex53_xhigh | Skill: \$ralph | Mode: full-auto"
+echo "Excluded: issues 25, 26, 27, 29, 30"

package/skills/tfx-codex-swarm-workspace/iteration-1/selective-spawn-with-override/without_skill/outputs/routing.md

@@ -0,0 +1,61 @@
+# Codex Swarm - Profile Routing (Selective: Issue 24, 28 with xhigh override)
+
+> Generated: 2026-03-30
+> Mode: without_skill (no SKILL.md guidance)
+
+## Size Assessment
+
+| # | PRD File | Lines | Affected Files | High-cost Keywords | Dependencies | Size |
+|---|----------|-------|----------------|-------------------|--------------|------|
+| 24 | prd-issue-24 | 73 | 4 (`remote-spawn.mjs`, handoff, scp, staging) | 0 | none | **L** |
+| 28 | prd-issue-28 | 69 | 4 (`headless-guard.mjs`, `hub/team/psmux.mjs`, codex, gemini) | 0 | none | **L** |
+
+Both PRDs fall in the **L (standard)** range.
+
+## Default Routing (what auto-routing would produce)
+
+Under normal routing for L-size implement tasks:
+
+| # | Task | Size | Type | Auto Profile | Auto Flags |
+|---|------|------|------|-------------|------------|
+| 24 | file-transfer | L | implement | `codex53_high` | `-c 'model="gpt-5.3-codex"' -c 'model_reasoning_effort="high"'` |
+| 28 | guard-deadlock | L | implement | `codex53_high` | `-c 'model="gpt-5.3-codex"' -c 'model_reasoning_effort="high"'` |
+
+## Applied Routing (after user xhigh override)
+
+User directive: "둘 다 xhigh로 세팅하고" overrides the auto-routed `codex53_high` to `codex53_xhigh`.
+
+| # | Task | Size | Type | Profile | Codex Flags |
+|---|------|------|------|---------|-------------|
+| 24 | file-transfer | L | implement | **`codex53_xhigh`** | `-c 'model="gpt-5.3-codex"' -c 'model_reasoning_effort="xhigh"'` |
+| 28 | guard-deadlock | L | implement | **`codex53_xhigh`** | `-c 'model="gpt-5.3-codex"' -c 'model_reasoning_effort="xhigh"'` |
+
+### Profile Resolution
+
+`codex53_xhigh` maps to the existing Codex config profile `[profiles.codex53_xhigh]`:
+```toml
+[profiles.codex53_xhigh]
+model = "gpt-5.3-codex"
+model_reasoning_effort = "xhigh"
+```
+
+This profile exists in `~/.codex/config.toml`, so either `-c` inline flags or `--profile codex53_xhigh` can be used.
+
+## Execution Defaults
+
+| Setting | Value | Note |
+|---------|-------|------|
+| Mode | `--full-auto` | Sandbox auto-approve (YOLO) |
+| Worktree | enabled | Per-session independent worktree |
+| psmux | enabled | Session management via send-keys |
+| WT | 2 tabs in single window | Only 2 sessions (not 7) |
+| OMX skill | `$ralph` | Persist loop until task fully complete |
+
+## Cost / Duration Estimate
+
+| Factor | Value |
+|--------|-------|
+| Model | gpt-5.3-codex (SWE-Bench 72%, Terminal-Bench 77%) |
+| Effort | xhigh (maximum reasoning, slowest) |
+| Sessions | 2 concurrent |
+| Expected duration | 15-45 min per session (xhigh adds ~2-3x vs high) |

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

@@ -0,0 +1,5 @@
+{
+  "total_tokens": 65928,
+  "duration_ms": 259460,
+  "total_duration_seconds": 259.5
+}

package/skills/tfx-deep-interview/SKILL.md

@@ -1,9 +1,8 @@
 ---
 name: tfx-deep-interview
 description: >
-  구현 전 소크라테스식 질의로 요구사항 깊이 탐색 (Gemini 위임, 3-CLI consensus 미사용).
+  구현 전 소크라테스식 질의로 요구사항 깊이 탐색.
   5단계 인터뷰를 통해 요구사항을 명확히 하고 최적 구현 경로를 도출합니다.
-  NOTE: deep- 접두사이지만 Gemini 단독 분석 스킬. 3-CLI Deep 규약의 예외.
 triggers:
   - deep-interview
   - 딥인터뷰

package/skills/tfx-plan/SKILL.md.tmpl

@@ -9,10 +9,9 @@ triggers:
 argument-hint: "<구현할 기능 설명>"
 ---
 
-# tfx-plan — Light Implementation Plan
+# {{SKILL_NAME}} — Light Implementation Plan
 
-> **ARGUMENTS 처리**: 이 스킬이 `ARGUMENTS: <값>`과 함께 호출되면, 해당 값을 사용자 입력으로 취급하여
-> 워크플로우의 첫 단계 입력으로 사용한다. ARGUMENTS가 비어있거나 없으면 기존 절차대로 사용자에게 입력을 요청한다.
+{{> base}}
 
 
 > **Deep 버전**: tfx-deep-plan. "제대로/꼼꼼히" 수정자로 자동 에스컬레이션.

package/skills/tfx-psmux-rules/SKILL.md

@@ -294,14 +294,23 @@ wt.exe -w 0 sp -V -p triflux ...
 
 ---
 
-## RULE 6: spark53 프로파일은 Pro 전용
+## RULE 6: WT 탭/창은 wt-manager 경유 필수
+
+- `wt.exe new-tab ...` 직접 호출 금지
+- `wt.exe split-pane ...` 직접 호출 금지
+- `Start-Process wt.exe ...` PowerShell 호출 금지
+- 반드시 wt-manager.mjs의 createTab() / applyLayout() 사용
+
+---
+
+## RULE 7: spark53 프로파일은 Pro 전용
 
 `spark53_med`, `spark53_low` 등 spark 모델 프로파일은 **Codex Pro 구독 전용**이다.
 비-Pro 환경에서는 `codex53_low`로 폴백한다.
 
 ---
 
-## RULE 7: WT 레이아웃 선택 필수
+## RULE 8: WT 레이아웃 선택 필수
 
 WT에 패인을 배치하기 전에 **반드시** 사용자에게 레이아웃을 확인한다.
 새 탭(`nt`)은 금지 — split + dashboard가 기본.

package/skills/tfx-qa/SKILL.md.tmpl

@@ -9,10 +9,9 @@ triggers:
 argument-hint: "[테스트 명령 또는 파일 경로]"
 ---
 
-# tfx-qa — Light Test-Fix Cycle
+# {{SKILL_NAME}} — Light Test-Fix Cycle
 
-> **ARGUMENTS 처리**: 이 스킬이 `ARGUMENTS: <값>`과 함께 호출되면, 해당 값을 사용자 입력으로 취급하여
-> 워크플로우의 첫 단계 입력으로 사용한다. ARGUMENTS가 비어있거나 없으면 기존 절차대로 사용자에게 입력을 요청한다.
+{{> base}}
 
 
 > **Deep 버전**: tfx-deep-qa. "제대로/꼼꼼히" 수정자로 자동 에스컬레이션.

package/skills/tfx-remote-spawn/SKILL.md

@@ -1,22 +1,19 @@
 ---
 name: tfx-remote-spawn
-deprecated: true
-superseded_by: tfx-swarm
-description: "DEPRECATED — tfx-swarm으로 통합됨. 이 스킬 대신 /tfx-swarm을 사용하세요. 원격 실행, 세션 spawn, 호스트 관리 모두 tfx-swarm에서 처리."
+description: >
+  원격/로컬 머신에 Claude 세션을 psmux 기반으로 spawn하고 관리합니다.
+  자동 핸드오프, 추가 프롬프트 전송, 세션 재부착, 원격 환경 자동 감지를 지원합니다.
+  이 스킬은 다음 상황에서 반드시 사용하세요:
+  원격 실행, 세션 spawn, 다른 머신에서 작업, 원격 Claude, 세션 전달, 핸드오프 전달,
+  원격 세션에 프롬프트 보내기, 세션 목록, 세션 재부착.
+  로컬 호스트 별칭이 references/hosts.json에 등록되어 있으면 호스트명 언급만으로도 트리거됩니다.
 triggers:
   - tfx-remote-spawn
   - remote-spawn
 argument-hint: "[호스트] [프롬프트] 또는 자연어"
 ---
 
-# tfx-remote-spawn — DEPRECATED
-
-> **DEPRECATED:** 이 스킬은 `tfx-swarm`으로 통합되었습니다. `/tfx-swarm`을 사용하세요.
-> 원격 실행, 세션 spawn, 호스트 관리 모두 tfx-swarm에서 처리합니다.
-
----
-
-# (레거시) 원격/로컬 Claude 세션 관리
+# tfx-remote-spawn — 원격/로컬 Claude 세션 관리
 
 > **ARGUMENTS 처리**: 이 스킬이 `ARGUMENTS: <값>`과 함께 호출되면, 해당 값을 사용자 입력으로 취급하여
 > 워크플로우의 첫 단계 입력으로 사용한다. ARGUMENTS가 비어있거나 없으면 기존 절차대로 사용자에게 입력을 요청한다.

package/skills/tfx-research/SKILL.md.tmpl

@@ -11,10 +11,9 @@ triggers:
 argument-hint: "<검색 주제>"
 ---
 
-# tfx-research — Light Web Research
+# {{SKILL_NAME}} — Light Web Research
 
-> **ARGUMENTS 처리**: 이 스킬이 `ARGUMENTS: <값>`과 함께 호출되면, 해당 값을 사용자 입력으로 취급하여
-> 워크플로우의 첫 단계 입력으로 사용한다. ARGUMENTS가 비어있거나 없으면 기존 절차대로 사용자에게 입력을 요청한다.
+{{> base}}
 
 
 > **Deep 버전**: tfx-deep-research. "제대로/꼼꼼히" 수정자로 자동 에스컬레이션.

package/skills/tfx-review/SKILL.md.tmpl

@@ -9,10 +9,9 @@ triggers:
 argument-hint: "[파일 경로 또는 변경 설명]"
 ---
 
-# tfx-review — Light Code Review
+# {{SKILL_NAME}} — Light Code Review
 
-> **ARGUMENTS 처리**: 이 스킬이 `ARGUMENTS: <값>`과 함께 호출되면, 해당 값을 사용자 입력으로 취급하여
-> 워크플로우의 첫 단계 입력으로 사용한다. ARGUMENTS가 비어있거나 없으면 기존 절차대로 사용자에게 입력을 요청한다.
+{{> base}}
 
 
 > **Deep 버전**: tfx-deep-review. "제대로/꼼꼼히" 수정자로 자동 에스컬레이션.

package/skills/tfx-workspace/async-tests/run-tests.sh

@@ -0,0 +1,203 @@
+#!/usr/bin/env bash
+# tfx-route.sh v2.5 async job system — 통합 테스트
+set -uo pipefail
+
+ROUTE="scripts/tfx-route.sh"
+PASS=0
+FAIL=0
+TOTAL=0
+
+assert_eq() {
+  local name="$1" expected="$2" actual="$3"
+  TOTAL=$((TOTAL + 1))
+  if [[ "$actual" == *"$expected"* ]]; then
+    echo "  ✓ $name"
+    PASS=$((PASS + 1))
+  else
+    echo "  ✗ $name — expected: '$expected', got: '$actual'"
+    FAIL=$((FAIL + 1))
+  fi
+}
+
+assert_neq() {
+  local name="$1" unexpected="$2" actual="$3"
+  TOTAL=$((TOTAL + 1))
+  if [[ "$actual" != *"$unexpected"* ]]; then
+    echo "  ✓ $name"
+    PASS=$((PASS + 1))
+  else
+    echo "  ✗ $name — should NOT contain: '$unexpected', got: '$actual'"
+    FAIL=$((FAIL + 1))
+  fi
+}
+
+assert_exit() {
+  local name="$1" expected="$2" actual="$3"
+  TOTAL=$((TOTAL + 1))
+  if [[ "$actual" -eq "$expected" ]]; then
+    echo "  ✓ $name"
+    PASS=$((PASS + 1))
+  else
+    echo "  ✗ $name — expected exit=$expected, got exit=$actual"
+    FAIL=$((FAIL + 1))
+  fi
+}
+
+echo "═══ tfx-route.sh v2.5 Async Job System Tests ═══"
+echo ""
+
+# ── Test 1: --async 기본 동작 ──
+echo "Test 1: --async 기본 시작 + job_id 반환"
+JOB_ID=$(bash "$ROUTE" --async executor "echo hello" none 30 2>/dev/null)
+EC=$?
+assert_exit "exit code 0" 0 "$EC"
+TOTAL=$((TOTAL + 1))
+if [[ -n "$JOB_ID" ]]; then echo "  ✓ job_id not empty ($JOB_ID)"; PASS=$((PASS + 1)); else echo "  ✗ job_id is empty"; FAIL=$((FAIL + 1)); fi
+assert_neq "job_id not error" "error" "$JOB_ID"
+echo ""
+
+# ── Test 2: --job-status running → done 전이 ──
+echo "Test 2: --job-status 상태 전이 (running → done)"
+LONG_JOB=$(bash "$ROUTE" --async executor "sleep 3 && echo done" none 60 2>/dev/null)
+STATUS_EARLY=$(bash "$ROUTE" --job-status "$LONG_JOB" 2>/dev/null)
+assert_eq "initial status: running" "running" "$STATUS_EARLY"
+
+# Codex 시작 ~10초 + sleep 3초 + 후처리 → 최대 25초 대기
+for i in $(seq 1 5); do
+  sleep 5
+  STATUS_LATE=$(bash "$ROUTE" --job-status "$LONG_JOB" 2>/dev/null)
+  [[ "$STATUS_LATE" == "done" ]] && break
+done
+assert_eq "final status: done" "done" "$STATUS_LATE"
+echo ""
+
+# ── Test 3: --job-status 존재하지 않는 job ──
+echo "Test 3: --job-status 존재하지 않는 job"
+RESULT=$(bash "$ROUTE" --job-status "nonexistent-12345" 2>/dev/null)
+EC=$?
+assert_eq "returns error" "error" "$RESULT"
+assert_exit "exit code 1" 1 "$EC"
+echo ""
+
+# ── Test 4: --job-result 완료된 job ──
+echo "Test 4: --job-result 완료된 job 결과 읽기"
+# Test 1의 JOB_ID 재사용 — Codex 완료 대기
+for i in $(seq 1 6); do
+  S=$(bash "$ROUTE" --job-status "$JOB_ID" 2>/dev/null)
+  [[ "$S" == "done" ]] && break
+  sleep 5
+done
+RESULT=$(bash "$ROUTE" --job-result "$JOB_ID" 2>/dev/null)
+EC=$?
+assert_exit "exit code 0" 0 "$EC"
+TOTAL=$((TOTAL + 1))
+if [[ -n "$RESULT" ]]; then echo "  ✓ result not empty (${#RESULT} bytes)"; PASS=$((PASS + 1)); else echo "  ✗ result is empty"; FAIL=$((FAIL + 1)); fi
+assert_neq "result not error" "error:" "$RESULT"
+echo ""
+
+# ── Test 5: --job-result 아직 실행 중인 job ──
+echo "Test 5: --job-result 실행 중인 job → 에러"
+RUNNING_JOB=$(bash "$ROUTE" --async executor "sleep 30" none 60 2>/dev/null)
+RESULT=$(bash "$ROUTE" --job-result "$RUNNING_JOB" 2>/dev/null)
+EC=$?
+assert_eq "returns error" "error: job still running" "$RESULT"
+assert_exit "exit code 1" 1 "$EC"
+# cleanup
+JOB_DIR="${TMPDIR:-/tmp}/tfx-jobs/$RUNNING_JOB"
+[[ -f "$JOB_DIR/pid" ]] && kill "$(cat "$JOB_DIR/pid")" 2>/dev/null
+echo ""
+
+# ── Test 6: --job-wait 완료 감지 ──
+echo "Test 6: --job-wait 완료 감지"
+WAIT_JOB=$(bash "$ROUTE" --async executor "echo wait-test-ok" none 30 2>/dev/null)
+sleep 15  # codex 실행 대기
+WAIT_RESULT=$(bash "$ROUTE" --job-wait "$WAIT_JOB" 60 2>/dev/null)
+assert_eq "wait returns done" "done" "$WAIT_RESULT"
+echo ""
+
+# ── Test 7: --job-wait still_running (max_wait < 실행시간) ──
+echo "Test 7: --job-wait still_running (짧은 max_wait)"
+SLOW_JOB=$(bash "$ROUTE" --async executor "sleep 60" none 120 2>/dev/null)
+sleep 1
+WAIT_RESULT=$(bash "$ROUTE" --job-wait "$SLOW_JOB" 5 2>/dev/null)
+assert_eq "wait returns still_running" "still_running" "$WAIT_RESULT"
+# cleanup
+JOB_DIR="${TMPDIR:-/tmp}/tfx-jobs/$SLOW_JOB"
+[[ -f "$JOB_DIR/pid" ]] && kill "$(cat "$JOB_DIR/pid")" 2>/dev/null
+echo ""
+
+# ── Test 8: exit code 전파 ──
+echo "Test 8: 실패한 job의 exit code 전파"
+FAIL_JOB=$(bash "$ROUTE" --async executor "exit 42" none 30 2>/dev/null)
+# Codex 완료 대기
+for i in $(seq 1 8); do
+  S=$(bash "$ROUTE" --job-status "$FAIL_JOB" 2>/dev/null)
+  [[ "$S" != *"running"* ]] && break
+  sleep 5
+done
+STATUS=$(bash "$ROUTE" --job-status "$FAIL_JOB" 2>/dev/null)
+# Codex가 exit 42를 감싸서 성공/실패 둘 다 가능 — "running이 아님"만 확인
+TOTAL=$((TOTAL + 1))
+if [[ "$STATUS" == "done" || "$STATUS" == *"failed"* || "$STATUS" == "timeout" ]]; then
+  echo "  ✓ status is terminal: $STATUS"; PASS=$((PASS + 1))
+else
+  echo "  ✗ status not terminal: $STATUS"; FAIL=$((FAIL + 1))
+fi
+# Codex는 exit 42를 감싸서 다른 코드로 반환할 수 있음 — 완료 자체만 확인
+TOTAL=$((TOTAL + 1))
+if [[ "$STATUS" != *"running"* ]]; then echo "  ✓ job completed (not stuck running)"; PASS=$((PASS + 1)); else echo "  ✗ job still running"; FAIL=$((FAIL + 1)); fi
+echo ""
+
+# ── Test 9: job 디렉토리 구조 검증 ──
+echo "Test 9: job 디렉토리 구조"
+STRUCT_JOB=$(bash "$ROUTE" --async executor "echo structure-test" none 30 2>/dev/null)
+JOB_DIR="${TMPDIR:-/tmp}/tfx-jobs/$STRUCT_JOB"
+assert_eq "pid file exists" "true" "$([ -f "$JOB_DIR/pid" ] && echo true || echo false)"
+assert_eq "agent_type file exists" "true" "$([ -f "$JOB_DIR/agent_type" ] && echo true || echo false)"
+assert_eq "start_time file exists" "true" "$([ -f "$JOB_DIR/start_time" ] && echo true || echo false)"
+AGENT=$(cat "$JOB_DIR/agent_type" 2>/dev/null)
+assert_eq "agent_type == executor" "executor" "$AGENT"
+echo ""
+
+# ── Test 10: native.mjs 프롬프트 검증 ──
+echo "Test 10: native.mjs buildSlimWrapperPrompt async 키워드"
+PROMPT_CHECK=$(node -e "
+import('./hub/team/native.mjs').then(m => {
+  const p = m.buildSlimWrapperPrompt('codex', {
+    subtask: 'test task',
+    role: 'scientist',
+    teamName: 'test-team',
+    taskId: 'task-1',
+    agentName: 'codex-worker-1',
+  });
+  const checks = {
+    has_async: p.includes('--async'),
+    has_job_wait: p.includes('--job-wait'),
+    has_job_result: p.includes('--job-result'),
+    has_route_timeout: p.includes('auto 1800'),
+    no_old_bashTimeout: !p.includes('timeout: 1860000'),
+    has_launch_timeout: p.includes('timeout: 15000'),
+    has_wait_timeout: p.includes('timeout: 570000'),
+    has_result_timeout: p.includes('timeout: 30000'),
+  };
+  for (const [k, v] of Object.entries(checks)) {
+    console.log(k + '=' + v);
+  }
+});
+" 2>/dev/null)
+for line in $PROMPT_CHECK; do
+  key="${line%%=*}"
+  val="${line##*=}"
+  assert_eq "$key" "true" "$val"
+done
+echo ""
+
+# ── 결과 요약 ──
+echo "═══════════════════════════════════════════════════"
+echo "  Results: $PASS/$TOTAL passed, $FAIL failed"
+echo "═══════════════════════════════════════════════════"
+
+if [[ "$FAIL" -gt 0 ]]; then
+  exit 1
+fi
+exit 0