Mellea Fy Classify

Melleafy Step 0: Five-Axis Classification

Version: 4.0.0 | Prereq: None | Produces: classification.json

Schema: Output intermediate/classification.json MUST conform to schemas/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).

Processing order

Run axes in this order — each informs the next:

1. Axis 4 — Source runtime (drives which dialect doc applies to Axes 2 and 3)
2. Axis 5 — Interaction modality (explicit-first, then inferential)
3. Axis 3 — Tool involvement (informed by Axis 4's dialect mapping table)
4. Axis 1 — Reasoning archetype (informed by Axes 2, 3, 4)
5. Axis 2 — Pipeline shape (informed by Axis 1)

Axis 1: Reasoning Archetype

What kind of thinking does the skill do?

Detection signals:

• Count EXTRACT + CLASSIFY + VALIDATE_DOMAIN tags → high count suggests A
• Count GENERATE + SCHEMA tags → high count suggests B
• "hypothesis", "root cause", "investigate", "reproduce" → C
• Single-service wrapper with minimal decision logic → D1
• "spawn", "subagent", "dispatch", "CI/CD", "merge", "deploy" → D2
• Mostly guidelines, checklists, or "when X, do Y" rules → E

Axis 2: Pipeline Shape

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.

Axis 3: Tool Involvement

Axis 4: Source Runtime

The source runtime determines which dialect doc applies. Detection is signal-based with weighted scoring.

Signal weighting

Each detection signal has a weight:

• Strong (2.0): definitive marker — unambiguous runtime indicator
• Medium (1.0): characteristic pattern — common in this runtime, rare in others
• Weak (0.5): circumstantial — present in this runtime but also in others

Supported runtimes

R1 Hybrid detection

1. Compute weighted signal score for each runtime.
2. Find the two highest-scoring runtimes.
3. If their scores differ by ≤ 1.0 → classify as hybrid.
4. Tiebreaker precedence (when within threshold): 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).

R2 — Generated SKILL.md suppression (re-run safety)

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:

1. Exclude it from signal computation — do not let its YAML frontmatter contribute to the agent_skills_std score.
2. Add a warning to 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.

Axis 5: Interaction Modality

Eight modalities. Detection is explicit-first: look for declared modality signals in the source first; infer only if none are found.

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 — contradictory

Multiple modalities are valid (e.g., scheduled + event_triggered). Record primary in modality, others in secondary_modalities.

Cross-axis validation

After computing all five axes, run these consistency checks:

Output: 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",
  "modality_confidence": 0.75,
  "secondary_modalities": [],
  "warnings": ["Axis 5 inferred — no explicit modality signals found"],
  "halt": null
}

If any halt condition fires, set "halt": "<reason>" and stop — do not proceed to Step 1a.

Mapping report: Classification section

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.