Skills and tiers

The runtime serves a skill library. A skill is a canonical SKILL.md source, rendered per call at one of three tiers. One source directory, rendered per call, replaces the per-harness static mirrors of the upstream design.

This page explains the tier mechanism and the catalog tools. For the full per-skill reference, see the Skill index and The skill workflow.

Read a skill

get_skill reads a skill at a tier. It takes the skill name and an optional tier.

{
  "name": "develop-tdd",
  "tier": "lean"
}

An absent tier uses the TRUENORTH_TIER environment default. The tier argument overrides it for the one call.

The three tiers

• full: the verbatim SKILL.md. Use it when the model has ample context and wants the complete guidance, including rationale and examples.
• reasoning: the directive content with the meta scaffolding stripped. Use it for a native reasoning model that does not need hand-held chain-of-thought.
• lean: the directives compressed toward a tight token target. Whole protected invariant and acceptance-criterion lines are retained even when they exceed that target. Use it for a local model with a small context window.

A tier transform never alters the directive content that encodes an invariant or an acceptance criterion. It strips or compresses only meta and guardrail scaffolding. full is the identity.

The catalog tools

The catalog tools read and search the library:

• index_skills: list every skill and its phase.
• read_skill: parse a SKILL.md into its frontmatter, headings, and sections.
• search_skills: search skill metadata.
• validate_skill: enforce the two-tier structural contract. Every skill requires name, description, and kind, where kind is prose or scripted. Scripted skills also require a nonempty frontmatter verify; naming exceptions use name_exception. The complete SKILL.md is limited to 150 lines. The tool reports command metadata but does not execute it.

The skill graph

The graph tools build and query the entity-relation graph over the skills:

• build_skill_graph: build and persist the graph.
• read_graph, search_nodes, open_nodes: query it.
• get_dependencies: report the forward and reverse dependencies and the handoff chain for one skill.

build_skill_graph classifies skill mentions instead of treating every mention as a dependency:

• depends_on: A requires B before A runs. A after B produces A → B; A before B produces B → A.
• invokes: run, invoke, invokes, and route to name an execution target.
• handoff_to: next, next_skill, and hand off to name the next skill.
• references: any other exact mention or skill link.
• enforces: a skill cites a conventions section.

The build receipt includes unresolved or unclassified mentions, isolated skills, orchestrator-like skills without outgoing control flow, and dependency cycles. Unresolved or unclassified targets fail the build before the graph cache is written.

Git context

get_git_context reports the git status, log, or diff scoped to the tracked areas. It is a read-only report the runtime uses to ground its work.