Plan 0053: Model Capability Requirements in Skill Specs

Problem

Skills declare what they do, what they read and write, and what they depend on. They do not declare what agent capability they require.

This is a gap: interpret-message requires a model capable of multi-step reasoning across ambiguous natural language input. A 3B quantized model cannot reliably execute it. The skill spec has no way to encode this requirement, so an agent routing interpret-message to qwen2.5:3b would produce degraded output with no warning.

As local models are used for more operations (plan 0054), routing by capability becomes critical. Without a capability field in skill specs, we cannot prevent a 0.5B classification model from attempting a reasoning task designed for a 20B+ model.

Goal

Extend the skill spec system to allow skills to declare minimum model capability requirements. Agents and scripts must check these requirements before routing work to a model.

Approach

1. Define a capability taxonomy

Capability is not just parameter count. A well-chosen 7B reasoning model can outperform a poorly-chosen 13B model on specific tasks. The taxonomy should be task-based, not parameter-based:

Capability level	Description	Example models
trivial	Keyword extraction, format detection, simple classification	qwen2.5-0.5b, qwen3:1.7b
classification	Multi-label classification, tag generation, type inference	qwen2.5:3b, phi4-mini
extraction	Structured output, entity extraction, frontmatter generation	qwen2.5:7b, mistral-7b
generation	Coherent prose, summarization, definition writing	qwen2.5:7b, gemma3:12b
reasoning	Multi-step inference, contradiction detection, planning	deepseek-r1-7b, claude-sonnet-4
long-context	Processing documents >8k tokens	phi-3-mini-128k
cloud-required	Tasks requiring cloud model quality (final review, editorial)	claude-opus-4

2. Add min-capability field to skill frontmatter

Extend the skill specification to include an optional min-capability field:

min-capability: reasoning

Skills without this field are assumed to require trivial capability — they can run on any model.

3. Define the skill specification field

Update content/technology/specifications/agential-semioverse-repository/ specifications/skill-specification.md to document the min-capability field, its valid values, and the enforcement semantics.

4. Update the skill registry

The skill registry (.claude/skills/registry.md) should include a min-capability column so agents can read capability requirements without opening individual skill files.

5. Update routing in interpret-message and MCP tools

The interpret-message skill (and the MCP delegate_task tool) should check min-capability against the available model before dispatching.

For MCP delegate_task: if the caller supplies a model argument, validate it against the skill’s min-capability. If the model is insufficient, return an error with the minimum requirement rather than running with a degraded model.

6. Update local_llm.suggest_model()

Add a min_capability parameter to suggest_model() that filters out models below the required capability level:

suggest_model(task="classification", min_capability="extraction")

7. Annotate existing skills

After the spec is written, do a pass over all skills in the registry and add min-capability where the default (trivial) is wrong:

• interpret-message → reasoning
• make-specification → reasoning
• enrich-triage (inference path) → classification
• mine-triage-relevance → classification
• write-advisory-report → generation

Files to create/modify

File	Change
specifications/skill-specification.md	Add min-capability field
.claude/skills/registry.md	Add min-capability column
skills/interpret-message/SKILL.md	Set min-capability: reasoning
scripts/local_llm.py	Add min_capability param to suggest_model()
scripts/mcp-server.py	Add capability check in delegate_task

Not in scope

• Automatic model selection based on capability (that is suggest_model’s job — this plan only adds the declaration layer)
• Fine-grained benchmarks of specific models against specific tasks
• Hardware-aware capability declaration (NPU vs CPU)

Log

• 2026-03-08: Plan created. Captures the need identified when emsenn noted that interpret-message requires 20B+ reasoning but the skill spec provides no way to encode this constraint.