Tool Family: runtime-skills

Registration source: tools/mcp_tools/runtime_skills.py:8 register_runtime_skill_tools(mcp, *, ctx) — auto-discovered but conditionally registers only when _runtime_skills_enabled(config) (tools/mcp_tools/registry.py:271-273): skills.enabled: true and skills.allow_model_lookup: true. Otherwise no tools are added.
Gate: all @audit_tool (local read-only methodology catalog; never changes scope/permission/approval).

Tools Exported (4) — conditional

Tool	Params	Result Shape	Notes
`list_runtime_skills`	`include_maybe: bool=False`, `limit: int=50` (capped 1..200)	`RUNTIME_SKILLS:\n- name maybe?	tags: ...
`search_runtime_skills`	`query: str`, `tags: str=""` (comma/space CSV), `include_maybe=False`, `limit=10` (capped 1..25)	`RUNTIME_SKILL_SEARCH:` + matches `name	tags
`load_runtime_skill`	`name: str`, `reason: str=""`	`RUNTIME_SKILL_LOAD: loaded\nNAME: ...\nPATH: ...\nSAFETY: Advisory only; scope, ... still apply.\n\n<rendered markdown>` or `not found: ...` / `blocked: ... is under maybe/ and maybe_enabled is false.`	`registry.get(name)` exact; `skills.maybe_enabled` gates `maybe`; `max_chars_per_skill = skills.max_chars_per_skill` (default 2500 via `_positive_int`) and `render_skill_context([skill], reasons, max_chars_per_skill, max_total_chars)`.
`list_skill_references`	`name: str`	`RUNTIME_SKILL_REFERENCES: name\n- references/.md\nNIST CSF: ...\nMITRE ATT&CK: ...` or `... has no references/.md files.` / `not found: ...` / `listing disabled by skills.allow_reference_listing.`	Requires `skills.allow_reference_listing: true` (default true) else blocked. Surfaces `metadata.references` paths only — not contents — plus `nist_csf`/`mitre_attack`.

tools/skill_registry_cache.get_registry, tools/skill_registry.load_skill_registry, render_skill_context
tools/mcp_tools/registry._skills_config (overlay CONFIG_SCHEMA.skills), _runtime_skills_enabled, _positive_int, _truncate_text
tools/config_manager.CONFIG_SCHEMA

skills.enabled: bool (default true), skills.allow_model_lookup: bool (default true) — both required for enable
skills.maybe_enabled: bool (default false) — gate maybe/ skills
skills.allow_reference_listing: bool (default true) — gates list_skill_references
skills.max_chars_per_skill: int (default 2500) — render cap
skills.roots: list[str] (default ["skills"]) — catalog discovery

All @audit_tool with _redact_args (no secrets in these params); reason not wholesale-redacted but still content-masked if it contained an inline cred shape.
Advisory only disclaimer included in load_runtime_skill output.

limit clamped via _positive_int + max(1, min(...)).
maybe gate checked both in list and load.
name exact match required; search returns empty string when no matches (no matches. sentinel).

tests/test_mcp_runtime_skills.py:67,90,121,140,150,171,183,198 — list/search/load + audit, disabling via skills.enabled: false / allow_model_lookup: false, list_skill_references paths / unknown / disabled, references not read into context, metadata parsing (nist_csf/mitre)
tests/test_mcp_tool_registration.py — expects list_runtime_skills, search_runtime_skills, load_runtime_skill when enabled

source: repo docs (build sync)Edit this page on GitHub →