oh-my-customcode 0.33.0 → 0.34.0

package/templates/.claude/skills/evaluator-optimizer/SKILL.md

@@ -0,0 +1,256 @@
+---
+name: evaluator-optimizer
+description: Parameterized evaluator-optimizer loop for quality-critical output with configurable rubrics
+scope: core
+user-invocable: false
+---
+
+# Evaluator-Optimizer Skill
+
+## Purpose
+
+General-purpose iterative refinement loop. A generator agent produces output, an evaluator agent scores it against a configurable rubric, and the loop continues until the quality gate is met or max iterations are reached.
+
+This skill generalizes the worker-reviewer-pipeline pattern beyond code review to any domain requiring quality-critical output: documentation, architecture decisions, test plans, configurations, and more.
+
+## Configuration Schema
+
+```yaml
+evaluator-optimizer:
+  generator:
+    agent: {subagent_type}       # Agent that produces output
+    model: sonnet                 # Default model
+  evaluator:
+    agent: {subagent_type}       # Agent that reviews output
+    model: opus                   # Evaluator benefits from stronger reasoning
+  rubric:
+    - criterion: {name}
+      weight: {0.0-1.0}
+      description: {what to evaluate}
+  quality_gate:
+    type: all_pass | majority_pass | score_threshold
+    threshold: 0.8                # For score_threshold type
+  max_iterations: 3               # Default, hard cap: 5
+```
+
+### Parameter Details
+
+| Parameter | Required | Default | Description |
+|-----------|----------|---------|-------------|
+| `generator.agent` | Yes | — | Subagent type that produces output |
+| `generator.model` | No | `sonnet` | Model for generation |
+| `evaluator.agent` | Yes | — | Subagent type that evaluates output |
+| `evaluator.model` | No | `opus` | Model for evaluation (stronger reasoning preferred) |
+| `rubric` | Yes | — | List of evaluation criteria with weights |
+| `quality_gate.type` | No | `score_threshold` | Gate strategy |
+| `quality_gate.threshold` | No | `0.8` | Score threshold (for `score_threshold` type) |
+| `max_iterations` | No | `3` | Max refinement loops (hard cap: 5) |
+
+## Quality Gate Types
+
+| Type | Behavior |
+|------|----------|
+| `all_pass` | Every rubric criterion must pass |
+| `majority_pass` | >50% of weighted criteria pass |
+| `score_threshold` | Weighted average score >= threshold |
+
+### Gate Evaluation Logic
+
+- **all_pass**: Each criterion scored individually. All must receive `pass: true`.
+- **majority_pass**: Sum weights of passing criteria. If > 0.5 of total weight, gate passes.
+- **score_threshold**: Compute weighted average: `sum(score_i * weight_i) / sum(weight_i)`. Compare against threshold.
+
+## Workflow
+
+```
+1. Generator produces output
+   → Orchestrator spawns generator agent with task prompt
+   → Generator returns output artifact
+
+2. Evaluator scores against rubric
+   → Orchestrator spawns evaluator agent with:
+     - The output artifact
+     - The rubric criteria
+     - Instructions to produce verdict JSON
+   → Evaluator returns structured verdict
+
+3. Quality gate check:
+   - PASS → return output + final verdict
+   - FAIL → extract feedback, append to generator prompt → iteration N+1
+
+4. Max iterations reached → return best output + warning
+   → "Best" = output from iteration with highest weighted score
+```
+
+### Iteration Flow Diagram
+
+```
+┌─────────────────────────────────────────────────┐
+│                  Orchestrator                    │
+│                                                  │
+│  ┌──────────┐    ┌──────────┐    ┌──────────┐  │
+│  │ Generate  │───→│ Evaluate │───→│  Gate    │  │
+│  │ (iter N)  │    │          │    │  Check   │  │
+│  └──────────┘    └──────────┘    └────┬─────┘  │
+│       ↑                               │         │
+│       │         ┌──────────┐    FAIL  │  PASS   │
+│       └─────────│ Feedback │←────────┘    │     │
+│                 └──────────┘              ↓     │
+│                                     Return      │
+└─────────────────────────────────────────────────┘
+```
+
+## Stopping Criteria Display
+
+```
+[Evaluator-Optimizer]
+├── Generator: {agent}:{model}
+├── Evaluator: {agent}:{model}
+├── Max iterations: {max_iterations} (hard cap: 5)
+├── Quality gate: {type} (threshold: {threshold})
+└── Rubric: {N} criteria
+```
+
+Display this at the start of the loop to provide transparency into the refinement configuration.
+
+## Verdict Format
+
+The evaluator MUST return a structured verdict in this format:
+
+```json
+{
+  "status": "pass | fail",
+  "iteration": 2,
+  "score": 0.85,
+  "rubric_results": [
+    {"criterion": "clarity", "pass": true, "score": 0.9, "feedback": "Clear structure and logical flow"},
+    {"criterion": "accuracy", "pass": true, "score": 0.8, "feedback": "All facts verified, one minor imprecision in section 3"}
+  ],
+  "improvement_summary": "Section 3 terminology tightened. Examples added to section 2."
+}
+```
+
+### Verdict Fields
+
+| Field | Type | Description |
+|-------|------|-------------|
+| `status` | `pass` or `fail` | Overall quality gate result |
+| `iteration` | number | Current iteration number (1-indexed) |
+| `score` | number (0.0-1.0) | Weighted average score across all criteria |
+| `rubric_results` | array | Per-criterion evaluation details |
+| `improvement_summary` | string | Summary of changes from previous iteration (empty on iteration 1) |
+
+## Domain Examples
+
+| Domain | Generator | Evaluator | Rubric Focus |
+|--------|-----------|-----------|--------------|
+| Code review | `lang-*-expert` | opus reviewer | Correctness, style, security |
+| Documentation | `arch-documenter` | opus reviewer | Completeness, clarity, accuracy |
+| Architecture | Plan agent | opus reviewer | No SPOFs, no circular deps |
+| Test plans | `qa-planner` | `qa-engineer` | Coverage, edge cases, feasibility |
+| Agent creation | `mgr-creator` | opus reviewer | Frontmatter validity, R006 compliance |
+| Security audit | `sec-codeql-expert` | opus reviewer | Vulnerability coverage, false positive rate |
+
+### Example: Documentation Review
+
+```yaml
+evaluator-optimizer:
+  generator:
+    agent: arch-documenter
+    model: sonnet
+  evaluator:
+    agent: general-purpose
+    model: opus
+  rubric:
+    - criterion: completeness
+      weight: 0.3
+      description: All sections present, no gaps in coverage
+    - criterion: clarity
+      weight: 0.3
+      description: Clear language, no ambiguity, proper examples
+    - criterion: accuracy
+      weight: 0.25
+      description: All technical details correct and verifiable
+    - criterion: consistency
+      weight: 0.15
+      description: Consistent terminology, formatting, and style
+  quality_gate:
+    type: score_threshold
+    threshold: 0.8
+  max_iterations: 3
+```
+
+### Example: Code Implementation
+
+```yaml
+evaluator-optimizer:
+  generator:
+    agent: lang-typescript-expert
+    model: sonnet
+  evaluator:
+    agent: general-purpose
+    model: opus
+  rubric:
+    - criterion: correctness
+      weight: 0.35
+      description: Code compiles, logic is correct, edge cases handled
+    - criterion: style
+      weight: 0.2
+      description: Follows project conventions, clean and readable
+    - criterion: security
+      weight: 0.25
+      description: No injection risks, proper input validation
+    - criterion: performance
+      weight: 0.2
+      description: No unnecessary allocations, efficient algorithms
+  quality_gate:
+    type: all_pass
+  max_iterations: 3
+```
+
+## Integration
+
+| Rule | Integration |
+|------|-------------|
+| R009 | Generator and evaluator run sequentially (dependent — evaluator needs generator output) |
+| R010 | Orchestrator configures and invokes the loop; generator and evaluator agents execute via Agent tool |
+| R007 | Each iteration displays agent identification for both generator and evaluator |
+| R008 | Tool calls within generator/evaluator follow tool identification rules |
+| R013 | Ecomode: return verdict summary only, skip per-criterion details |
+| R015 | Display configuration block at loop start for intent transparency |
+
+## Ecomode Behavior
+
+When ecomode is active (R013), compress output:
+
+**Normal mode:**
+```
+[Evaluator-Optimizer] Iteration 2/3
+├── Generator: lang-typescript-expert:sonnet → produced 45-line module
+├── Evaluator: general-purpose:opus → scored 0.85
+├── Rubric: correctness ✓(0.9), style ✓(0.8), security ✓(0.85), performance ✓(0.8)
+└── Gate: score_threshold(0.8) → PASS
+```
+
+**Ecomode:**
+```
+[EO] iter 2/3 → 0.85 → PASS
+```
+
+## Error Handling
+
+| Scenario | Action |
+|----------|--------|
+| Generator fails to produce output | Retry once with simplified prompt; if still fails, abort with error |
+| Evaluator returns malformed verdict | Retry once; if still malformed, treat as fail with score 0.0 |
+| Max iterations reached without passing | Return best-scored output with warning: "Quality gate not met after {N} iterations" |
+| Rubric has zero total weight | Reject configuration, report error before starting loop |
+| Hard cap exceeded in config | Clamp `max_iterations` to 5, emit warning |
+
+## Constraints
+
+- This skill does NOT use `context: fork` — it operates within the caller's context
+- Generator and evaluator MUST be different agent invocations (no self-review)
+- The evaluator prompt MUST include the full rubric to ensure consistent scoring
+- Iteration state (best score, best output) is tracked by the orchestrator
+- The hard cap of 5 iterations prevents runaway refinement loops

package/templates/.claude/skills/fix-refs/SKILL.md

@@ -1,5 +1,5 @@
 ---
-name: fix-refs
+name: omcustom:fix-refs
 description: Fix broken agent references and symlinks
 scope: harness
 argument-hint: "[agent-name] [--all] [--dry-run]"

package/templates/.claude/skills/help/SKILL.md

@@ -1,7 +1,7 @@
 ---
-name: help
+name: omcustom:help
 description: Show help information for commands and system
-scope: core
+scope: harness
 argument-hint: "[command] [--agents] [--rules]"
 ---
 

package/templates/.claude/skills/lists/SKILL.md

@@ -1,7 +1,7 @@
 ---
-name: lists
+name: omcustom:lists
 description: Show all available commands
-scope: core
+scope: harness
 argument-hint: "[--category <category>] [--verbose]"
 ---
 

package/templates/.claude/skills/monitoring-setup/SKILL.md

@@ -1,5 +1,5 @@
 ---
-name: monitoring-setup
+name: omcustom:monitoring-setup
 description: Enable/disable OpenTelemetry console monitoring for Claude Code usage tracking
 scope: package
 argument-hint: "[enable|disable|status]"

package/templates/.claude/skills/npm-audit/SKILL.md

@@ -1,5 +1,5 @@
 ---
-name: npm-audit
+name: omcustom:npm-audit
 description: Audit npm dependencies for security and updates
 scope: package
 argument-hint: "[--fix] [--production]"

package/templates/.claude/skills/npm-publish/SKILL.md

@@ -1,5 +1,5 @@
 ---
-name: npm-publish
+name: omcustom:npm-publish
 description: Publish package to npm registry with pre-checks
 scope: package
 argument-hint: "[--tag <tag>] [--dry-run]"

package/templates/.claude/skills/npm-version/SKILL.md

@@ -1,5 +1,5 @@
 ---
-name: npm-version
+name: omcustom:npm-version
 description: Manage semantic versions for npm packages
 scope: package
 argument-hint: "<major|minor|patch> [--no-tag] [--no-commit]"

package/templates/.claude/skills/research/SKILL.md

@@ -20,6 +20,17 @@ Orchestrates 10 parallel research teams for comprehensive deep analysis of any t
 /research Rust async runtime comparison
 ```
 
+## When NOT to Use
+
+| Scenario | Better Alternative |
+|----------|--------------------|
+| Simple factual question | Direct answer or single WebSearch |
+| Single-file code review | `/dev-review` with specific file |
+| Known solution implementation | `/structured-dev-cycle` |
+| Topic with < 3 comparison dimensions | Single Explore agent |
+
+**Pre-execution check**: If the query can be answered with < 3 sources, skip 10-team research.
+
 ## Architecture — 4 Phases
 
 ### Phase 1: Parallel Research (10 teams, batched per R009)
@@ -258,6 +269,8 @@ Before execution:
 └── Phase 4: Report + GitHub issue
 
 Estimated: {time} | Teams: 10 | Models: sonnet → opus → codex
+Stopping: max 30 verification rounds, convergence at 0 contradictions
+Cost: ~$8-15 (10 teams × sonnet + opus verification)
 Execute? [Y/n]
 ```
 

package/templates/.claude/skills/sauron-watch/SKILL.md

@@ -1,5 +1,5 @@
 ---
-name: sauron-watch
+name: omcustom:sauron-watch
 description: Full R017 verification (5+3 rounds) before commit
 scope: harness
 disable-model-invocation: true

package/templates/.claude/skills/status/SKILL.md

@@ -1,7 +1,7 @@
 ---
-name: status
+name: omcustom:status
 description: Show system status and health checks
-scope: core
+scope: harness
 argument-hint: "[--verbose] [--health]"
 ---
 

package/templates/.claude/skills/task-decomposition/SKILL.md

@@ -20,6 +20,19 @@ Decomposition is **recommended** when any of these thresholds are met:
 | Domains involved | > 2 domains | Requires multiple specialists |
 | Agent types needed | > 2 types | Cross-specialty coordination |
 
+### Step 0: Pattern Selection
+
+Before decomposing, select the appropriate workflow pattern:
+
+| Pattern | When to Use | Primitive |
+|---------|-------------|-----------|
+| Sequential | Steps must execute in order, each depends on previous | dag-orchestration (linear) |
+| Parallel | Independent subtasks with no shared state | Agent tool (R009) or Agent Teams (R018) |
+| Evaluator-Optimizer | Quality-critical output needing iterative refinement | worker-reviewer-pipeline |
+| Orchestrator | Complex multi-step with dynamic routing | Routing skills (secretary/dev-lead/de-lead/qa-lead) |
+
+**Decision**: If task has independent subtasks → Parallel. If quality-critical → add EO review cycle. If multi-step with dependencies → Sequential/Orchestrator.
+
 ## Decomposition Process
 
 ```

package/templates/.claude/skills/update-docs/SKILL.md

@@ -1,5 +1,5 @@
 ---
-name: update-docs
+name: omcustom:update-docs
 description: Sync documentation with project structure
 scope: harness
 argument-hint: "[--check] [--target <path>]"

package/templates/.claude/skills/update-external/SKILL.md

@@ -1,5 +1,5 @@
 ---
-name: update-external
+name: omcustom:update-external
 description: Update agents from external sources (GitHub, docs, etc.)
 scope: harness
 argument-hint: "[agent-name] [--check] [--force]"

package/templates/.claude/skills/worker-reviewer-pipeline/SKILL.md

@@ -98,6 +98,16 @@ When Agent Teams is NOT available, falls back to sequential Agent tool calls:
 Agent(worker) → result → Agent(reviewer) → verdict → Agent(worker) → ...
 ```
 
+## Stopping Criteria Display
+
+Before execution, display:
+```
+[Worker-Reviewer Pipeline]
+├── Max iterations: {max_iterations} (default: 3, hard cap: 5)
+├── Quality gate: {pass_threshold}% approval required
+└── Early stop: All reviewers approve → stop immediately
+```
+
 ## Display Format
 
 ```

package/templates/CLAUDE.md.en

@@ -151,30 +151,31 @@ Violation = immediate correction. No exception for "small changes".
 
 | Command | Description |
 |---------|-------------|
-| `/analysis` | Analyze project and auto-configure customizations |
-| `/create-agent` | Create a new agent |
-| `/update-docs` | Sync documentation with project structure |
-| `/update-external` | Update agents from external sources |
-| `/audit-agents` | Audit agent dependencies |
-| `/fix-refs` | Fix broken references |
+| `/omcustom:analysis` | Analyze project and auto-configure customizations |
+| `/omcustom:create-agent` | Create a new agent |
+| `/omcustom:update-docs` | Sync documentation with project structure |
+| `/omcustom:update-external` | Update agents from external sources |
+| `/omcustom:audit-agents` | Audit agent dependencies |
+| `/omcustom:fix-refs` | Fix broken references |
 | `/dev-review` | Review code for best practices |
 | `/dev-refactor` | Refactor code |
 | `/memory-save` | Save session context to claude-mem |
 | `/memory-recall` | Search and recall memories |
-| `/monitoring-setup` | Enable/disable OTel console monitoring |
-| `/npm-publish` | Publish package to npm registry |
-| `/npm-version` | Manage semantic versions |
-| `/npm-audit` | Audit dependencies |
+| `/omcustom:monitoring-setup` | Enable/disable OTel console monitoring |
+| `/omcustom:npm-publish` | Publish package to npm registry |
+| `/omcustom:npm-version` | Manage semantic versions |
+| `/omcustom:npm-audit` | Audit dependencies |
 | `/codex-exec` | Execute Codex CLI prompts |
 | `/optimize-analyze` | Analyze bundle and performance |
 | `/optimize-bundle` | Optimize bundle size |
 | `/optimize-report` | Generate optimization report |
 | `/research` | 10-team parallel deep analysis and cross-verification |
-| `/sauron-watch` | Full R017 verification |
+| `/deep-plan` | Research-validated planning (research → plan → verify) |
+| `/omcustom:sauron-watch` | Full R017 verification |
 | `/structured-dev-cycle` | 6-stage structured development cycle (Plan → Verify → Implement → Verify → Compound → Done) |
-| `/lists` | Show all available commands |
-| `/status` | Show system status |
-| `/help` | Show help information |
+| `/omcustom:lists` | Show all available commands |
+| `/omcustom:status` | Show system status |
+| `/omcustom:help` | Show help information |
 
 ## Project Structure
 
@@ -183,7 +184,7 @@ project/
 +-- CLAUDE.md                    # Entry point
 +-- .claude/
 |   +-- agents/                  # Subagent definitions (44 files)
-|   +-- skills/                  # Skills (69 directories)
+|   +-- skills/                  # Skills (71 directories)
 |   +-- rules/                   # Global rules (R000-R019)
 |   +-- hooks/                   # Hook scripts (memory, HUD)
 |   +-- contexts/                # Context files (ecomode)
@@ -249,15 +250,15 @@ Task tool + routing skills remain the fallback for simple/cost-sensitive tasks.
 
 ```bash
 # Project analysis
-/analysis
+/omcustom:analysis
 
 # Show all commands
-/lists
+/omcustom:lists
 
 # Agent management
-/create-agent my-agent
-/update-docs
-/audit-agents
+/omcustom:create-agent my-agent
+/omcustom:update-docs
+/omcustom:audit-agents
 
 # Code review
 /dev-review src/main.go
@@ -267,7 +268,7 @@ Task tool + routing skills remain the fallback for simple/cost-sensitive tasks.
 /memory-recall authentication
 
 # Verification
-/sauron-watch
+/omcustom:sauron-watch
 ```
 
 ## External Dependencies

package/templates/CLAUDE.md.ko

@@ -151,30 +151,31 @@ oh-my-customcode로 구동됩니다.
 
 | 커맨드 | 설명 |
 |--------|------|
-| `/analysis` | 프로젝트 분석 및 자동 커스터마이징 |
-| `/create-agent` | 새 에이전트 생성 |
-| `/update-docs` | 프로젝트 구조와 문서 동기화 |
-| `/update-external` | 외부 소스에서 에이전트 업데이트 |
-| `/audit-agents` | 에이전트 의존성 감사 |
-| `/fix-refs` | 깨진 참조 수정 |
+| `/omcustom:analysis` | 프로젝트 분석 및 자동 커스터마이징 |
+| `/omcustom:create-agent` | 새 에이전트 생성 |
+| `/omcustom:update-docs` | 프로젝트 구조와 문서 동기화 |
+| `/omcustom:update-external` | 외부 소스에서 에이전트 업데이트 |
+| `/omcustom:audit-agents` | 에이전트 의존성 감사 |
+| `/omcustom:fix-refs` | 깨진 참조 수정 |
 | `/dev-review` | 코드 베스트 프랙티스 리뷰 |
 | `/dev-refactor` | 코드 리팩토링 |
 | `/memory-save` | 세션 컨텍스트를 claude-mem에 저장 |
 | `/memory-recall` | 메모리 검색 및 리콜 |
-| `/monitoring-setup` | OTel 콘솔 모니터링 활성화/비활성화 |
-| `/npm-publish` | npm 레지스트리에 패키지 배포 |
-| `/npm-version` | 시맨틱 버전 관리 |
-| `/npm-audit` | 의존성 감사 |
+| `/omcustom:monitoring-setup` | OTel 콘솔 모니터링 활성화/비활성화 |
+| `/omcustom:npm-publish` | npm 레지스트리에 패키지 배포 |
+| `/omcustom:npm-version` | 시맨틱 버전 관리 |
+| `/omcustom:npm-audit` | 의존성 감사 |
 | `/codex-exec` | Codex CLI 프롬프트 실행 |
 | `/optimize-analyze` | 번들 및 성능 분석 |
 | `/optimize-bundle` | 번들 크기 최적화 |
 | `/optimize-report` | 최적화 리포트 생성 |
 | `/research` | 10-team 병렬 딥 분석 및 교차 검증 |
-| `/sauron-watch` | 전체 R017 검증 |
+| `/deep-plan` | 연구 검증 기반 계획 수립 (research → plan → verify) |
+| `/omcustom:sauron-watch` | 전체 R017 검증 |
 | `/structured-dev-cycle` | 6단계 구조적 개발 사이클 (Plan → Verify → Implement → Verify → Compound → Done) |
-| `/lists` | 모든 사용 가능한 커맨드 표시 |
-| `/status` | 시스템 상태 표시 |
-| `/help` | 도움말 표시 |
+| `/omcustom:lists` | 모든 사용 가능한 커맨드 표시 |
+| `/omcustom:status` | 시스템 상태 표시 |
+| `/omcustom:help` | 도움말 표시 |
 
 ## 프로젝트 구조
 
@@ -183,7 +184,7 @@ project/
 +-- CLAUDE.md                    # 진입점
 +-- .claude/
 |   +-- agents/                  # 서브에이전트 정의 (44 파일)
-|   +-- skills/                  # 스킬 (69 디렉토리)
+|   +-- skills/                  # 스킬 (71 디렉토리)
 |   +-- rules/                   # 전역 규칙 (R000-R019)
 |   +-- hooks/                   # 훅 스크립트 (메모리, HUD)
 |   +-- contexts/                # 컨텍스트 파일 (ecomode)
@@ -249,15 +250,15 @@ Claude Code의 Agent Teams 기능이 활성화되어 있으면 (`CLAUDE_CODE_EXP
 
 ```bash
 # 프로젝트 분석
-/analysis
+/omcustom:analysis
 
 # 모든 커맨드 표시
-/lists
+/omcustom:lists
 
 # 에이전트 관리
-/create-agent my-agent
-/update-docs
-/audit-agents
+/omcustom:create-agent my-agent
+/omcustom:update-docs
+/omcustom:audit-agents
 
 # 코드 리뷰
 /dev-review src/main.go
@@ -267,7 +268,7 @@ Claude Code의 Agent Teams 기능이 활성화되어 있으면 (`CLAUDE_CODE_EXP
 /memory-recall authentication
 
 # 검증
-/sauron-watch
+/omcustom:sauron-watch
 ```
 
 ## 외부 의존성