Skill package maintenance (DCC adapters + bundled core skills)

This guide is the single maintenance contract for SKILL packages in dcc-mcp-core and downstream adapters (e.g. dcc-mcp-maya). In-repo reference implementations live under:

• python/dcc_mcp_core/skills/dcc-diagnostics/ — infrastructure skill: rich frontmatter description, search-hint, layer, tool purposes spelled out in SKILL.md body.
• python/dcc_mcp_core/skills/media/ — infrastructure skill: typed vx-managed FFmpeg/FFprobe wrappers for DCC render/playblast artifacts without exposing arbitrary shell or vx execution.
• python/dcc_mcp_core/skills/workflow/ — orchestration skill: example JSON chains and explicit “when not to use” boundaries.

Use those two trees when authoring or reviewing any new skill.

Ownership before implementation

Before adding or changing bundled adapter skills, read docs/POLICY_SKILL_OWNERSHIP.md and the relevant adapter's SKILL_OWNERSHIP.yml if it exists.

• Common file operations (open, save, import, export, read_file, write_file, path probes) must have one primary owning skill package for each adapter.
• Do not copy a file-operation tool into a second skill just to improve discoverability; add aliases, search hints, recipes, or next-tools pointing to the primary owner instead.
• If a duplicate is unavoidable, record the rationale and owner in SKILL_OWNERSHIP.yml in the same PR.

Frontmatter (SKILL.md)

• Keep description as the primary agent-facing contract (MCP get_skill_info / search summaries do not ship the Markdown body).
• Under metadata.dcc-mcp: always set tools, layer, dcc, version, search-hint, tags as required by your adapter policy.
• Optional but recommended for long-form notes:
  • recipes: — anchor-based snippets (recipes__* tools when the host registers them).
  • skill-reference-docs: — glob list relative to the skill root so skill_refs__list / skill_refs__read can serve arbitrary Markdown/text under references/ (or other dirs) without hard-coding one filename.
• Legacy introspection: single path is still honoured by skill-reference-docs resolution; prefer skill-reference-docs for new packages.

tools.yaml

• Every tool must declare execution, affinity, and realistic timeout_hint_secs when execution: async.
• Import / export / save / paths: descriptions should state absolute vs workspace-relative paths, required plugins, and common failure follow-ups (e.g. parent directory must exist, use file_exists, save scene before export). Short descriptions make gateway describe_tool useless — aim for enough text that an agent can succeed without guessing.

Python (or other) scripts

• Validate inputs early; return skill_error / ToolResult-style envelopes with possible_solutions instead of letting Maya open modal dialogs.
• For writes: ensure parent directory exists or return a structured error.
• After export: verify output file exists and non-zero size when feasible.

Linting (dcc-mcp-maya)

Run from the Maya adapter repo:

python tools/lint_skills.py

Rules include IO description length hints and references/ metadata coverage. Extend tools/lint_skills.py when you add new cross-cutting conventions.

Gateway-facing agents

• Prefer gateway MCP search → describe followed by REST /v1/call (or per-host load_skill then typed tools). Long prose belongs in recipes / skill-reference-docs, not only in SKILL.md body below the frontmatter.

Adapter startup registration

Adapters that expose optional metadata-driven tools should use register_metadata_driven_tools(...) instead of copying scan/import/register wrappers. The helper scans with scan_and_load_lenient(...) when skills are not supplied, registers the default recipes and skill-reference-docs extensions, and returns a report with per-extension registered, failed, or skipped status.

from dcc_mcp_core import register_metadata_driven_tools

report = register_metadata_driven_tools(
    server,
    dcc_name="maya",
    extra_paths=[studio_skill_root],
)
logger.info("metadata tools: %s", report.to_dict())

Pass explicit skills=loaded_skills when the adapter has already scanned skill roots for server startup; that keeps discovery single-pass while preserving the same optional-extension error policy.