소스 정보
- 저장소
- generative-computing/mellea-skills-compiler
- 최근 소스 활동
- 2026년 8월 21일 11:27
- 감지된 SKILL.md 언어
- 영어
- 스타
- 46
- 포크
- 8
설치 방법
기본적으로 소스를 먼저 확인하는 Prompt가 선택됩니다. 직접 명령으로 전환하거나 로컬 사본을 다운로드할 수도 있습니다.
소스 파일 검토
설치 여부를 결정하기 전에 SKILL.md와 SkillsMP에 표시된 보조 파일을 읽어 보세요.
메뉴
기본적으로 소스를 먼저 확인하는 Prompt가 선택됩니다. 직접 명령으로 전환하거나 로컬 사본을 다운로드할 수도 있습니다.
설치 여부를 결정하기 전에 SKILL.md와 SkillsMP에 표시된 보조 파일을 읽어 보세요.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
직접 명령은 검토 Prompt를 거치지 않습니다. 실행하기 전에 소스를 확인하세요.
npx skills add https://github.com/generative-computing/mellea-skills-compiler --skill mellea-fy-classify명령은 한 줄로 유지됩니다. 복사하기 전에 가로로 스크롤해 전체 내용을 확인하세요.
로컬 사본을 원하시나요? SkillsMP에서 현재 제공할 수 있는 파일을 다운로드하세요.
SOC 직업 분류 기준
SKILL.md 표시 중
| name | mellea-fy-classify |
| description | # Melleafy Step 0: Five-Axis Classification |
| metadata | {"user-invocable":true,"disable-model-invocation":true} |
Version: 4.0.0 | Prereq: None | Produces: classification.json
Schema: Output
intermediate/classification.jsonMUST conform toschemas/classification.schema.json.
Classify the source spec along five axes before any other step runs. The classification drives dialect selection (Step 1a), category defaults (Step 2.5), entry-point shape (Step 3), and modality validation (Step 7).
Run axes in this order — each informs the next:
What kind of thinking does the skill do?
| Type | What it does | Decomposition strategy |
|---|---|---|
| A: Analysis | Sequential phases → structured findings | Full slot decomposition, two-step enrichment |
| B: Generation | Intent + constraints → generated artifact | Decompose intent capture and draft generation |
| C: Diagnosis | Hypothesis-driven investigation with gating | Decompose reasoning; live system interaction as stubs or grounding_context |
| D1: Integration | Thin wrapper around a single service/API | Decompose only the intent classification layer |
| D2: Orchestration | Multi-step coordination of tools/agents | Decompose only the decision logic |
| E: Knowledge | Passive rules, conventions, best practices | Depends on Axis 2 (pipeline shape) |
Detection signals:
EXTRACT + CLASSIFY + VALIDATE_DOMAIN tags → high count suggests AGENERATE + SCHEMA tags → high count suggests B| Shape | When | Pipeline structure |
|---|---|---|
| Sequential | 3+ distinct phases where each output feeds the next | Multi-phase m.instruct(format=PhaseSchema) |
| One-shot | Flat spec or all knowledge applies simultaneously | Single m.instruct(format=OutputSchema) |
For Types A, B, C: almost always Sequential. For Types D1, D2: usually One-shot for the decision logic component. For Type E: assess internal structure — named stages or sequential dependencies → Sequential; flat rules → One-shot.
| Pattern | Description | Generated files |
|---|---|---|
| P0: No tools | Pure reasoning | No tools.py, no dependencies.yaml |
| P4: Tools provide input | Tools run BEFORE pipeline; output feeds reasoning as parameters | dependencies.yaml, optionally loader.py |
| P2: Pipeline calls tools (deterministic) | LLM classifies intent; Python constructs tool calls from template | tools.py with allowlist, constrained_slots.py |
| P3: Pipeline calls tools (LLM-directed) | LLM decides which tool to call and with what arguments | tools.py with m.react() |
The source runtime determines which dialect doc applies. Detection is signal-based with weighted scoring.
Each detection signal has a weight:
| Runtime | Strong signals | Medium signals | Weak signals |
|---|---|---|---|
agent_skills_std | --- frontmatter + name:, description: YAML fields | .md extension, model: frontmatter key | Single-file, no Python |
claude_code | CLAUDE.md, .claude/commands/ directory | bash_command:, allowed_tools: | Markdown-primary |
bob | .bob/settings/settings.json, .bob/skills/ directory | bash_command:, allowed_tools: | Markdown-primary |
openclaw | SOUL.md + AGENTS.md in same directory | .md files with ## Identity, ## Rules sections | Multi-file workspace |
letta | .af file extension, JSON with agent_type key | "human_input_pause", "memory" keys | Single JSON file |
crewai | crew.py or Crew( in Python, @CrewBase | @agent, @task, @crew decorators | YAML agents.yaml+tasks.yaml |
langgraph | StateGraph(, add_node(, add_edge( | from langgraph imports | Python with graph construction |
autogen | from autogen import or from autogen_agentchat | OAI_CONFIG_LIST, , |
hybrid.agent_skills_std > claude_code > openclaw > crewai > langgraph > autogen > letta > openai_agents_sdk > smolagents.If a spec provides --source-runtime=<runtime> on the command line, skip detection and use the override (record in classification.json:source_runtime_override).
Melleafy Step 6 writes a generated SKILL.md inside <package_name>/ for non-.md source runtimes (CrewAI, LangGraph, Letta, etc.). Because <package_name>/ is a subdirectory, Step 0 does not scan it during signal computation — no suppression logic is needed for the normal case.
Residual guard: if a SKILL.md is found at the skill root itself (e.g., left from an older melleafy run that placed it there, or manually created), read the first 30 lines and look for "auto-generated by Melleafy". If found:
agent_skills_std score.classification.json:warnings: "SKILL.md at skill root appears to be a melleafy-generated file — excluded from source runtime detection."Scope: applies only to .md files at the skill root containing the auto-generation marker. A genuine SKILL.md written by the user (without the marker) participates in signal scoring normally.
Unrecognised specs with no signals → unknown runtime. Melleafy can still process these using the generic inventory pass, but no dialect-specific mapping table applies.
Eight modalities. Detection is explicit-first: look for declared modality signals in the source first; infer only if none are found.
| Modality | Declared by | Description |
|---|---|---|
synchronous_oneshot | Absence of other modality signals | Request-in, response-out, no state |
streaming | "stream", "streaming output", stream=True references | Token-by-token output |
conversational_session | Session state references, "conversation history", m.chat() patterns | Multi-turn with memory |
review_gated | "approve", "human review", "gate", result.interruptions | Human-in-the-loop approval step |
scheduled | Cron patterns, "every N hours", schedule: frontmatter key | Time-triggered execution |
event_triggered | Webhook references, "on push", "on PR", event handler patterns | Event-driven |
heartbeat | "monitor", "watchdog", "poll every", "continuous" | Polling loop |
realtime_media | Audio/video/image stream references | Real-time media processing |
Explicit-first rule: If the source has a modality: frontmatter key or an activeHours: / isolatedSession: directive (OpenClaw), treat those as definitive. Only apply inferential detection when no explicit signals exist.
Composition validation — some combinations are impossible (halt with error):
heartbeat + synchronous_oneshot — contradictoryMultiple modalities are valid (e.g., scheduled + event_triggered). Record primary in modality, others in secondary_modalities.
After computing all five axes, run these consistency checks:
| Condition | Action |
|---|---|
| Axis 3 = P0 (no tools) AND Axis 1 = D1 (integration) | Halt — D1 without tools is incoherent |
| Axis 2 = Sequential AND phase count ≤ 1 | Warn — Sequential classification needs at least 2 phases |
Axis 5 = realtime_media AND Axis 1 ∈ {A, B} | Warn — realtime media with analysis/generation archetypes may need host adapter |
Axis 4 = hybrid AND no --source-runtime override | Warn — hybrid specs need manual review |
classification.json{
"format_version": "1.0",
"archetype": "A",
"archetype_confidence": 0.88,
"shape": "Sequential",
"shape_phase_count": 3,
"tool_involvement": "Tools provide input",
"tool_involvement_variant": "P4",
"source_runtime": "openclaw",
"source_runtime_override": null,
"source_runtime_scores": { "openclaw": 5.5, "agent_skills_std": 1.0 },
"hybrid_threshold_triggered": false,
"modality": "synchronous_oneshot"
If any halt condition fires, set "halt": "<reason>" and stop — do not proceed to Step 1a.
Document the classification as the first section of the mapping report (generated in Step 6):
## Classification
**Axis 1 — Reasoning Archetype**: Type A (Analytical Pipeline)
Evidence: 8 EXTRACT/CLASSIFY/VALIDATE_DOMAIN elements identified across §2–§5.
**Axis 2 — Pipeline Shape**: Sequential (3 phases)
Rationale: Phase 1 (extraction) feeds Phase 2 (assessment) feeds Phase 3 (report generation).
**Axis 3 — Tool Involvement**: P4 (Tools provide input)
Rationale: Spec reads code files and test outputs before analysis; no tool calls inside the pipeline.
**Axis 4 — Source Runtime**: openclaw (score 5.5; next: agent_skills_std 1.0)
Signals: SOUL.md + AGENTS.md directory structure (strong, 2.0); `## Identity` section in SOUL.md (medium, 1.0).
**Axis 5 — Interaction Modality**: synchronous_oneshot (inferred, confidence 0.75)
Rationale: No explicit modality declarations found; no session state, scheduling, or event signals detected.
AssistantAgent(UserProxyAgent(| Python multi-agent |
openai_agents_sdk | Agent(instructions=, Runner.run_sync( | @function_tool, handoffs=[ | from agents import |
smolagents | CodeAgent(, ToolCallingAgent( | from smolagents, additional_authorized_imports | HfApiModel( |