Skills

​

Supported format

• SKILL.md (YAML frontmatter + markdown body)

<scope>/skills/<skill-name>/SKILL.md

​

SKILL.md required fields

• name
• description

• name must be 1-64 chars, lowercase unicode letters, digits, and hyphen with no leading or trailing hyphen and no consecutive hyphens
• name must match the parent skill directory name after normalization
• description must be 1-1024 chars
• compatibility, if provided, must be 1-500 chars
• unknown frontmatter fields are rejected

• license
• compatibility
• metadata as a string-to-string map
• allowed-tools as a space-delimited tool list such as Bash(git:*) Read

​

Discovery scopes and precedence

1. .reliant.local/skills/**
2. .reliant/skills/**
3. ~/.reliant/skills/**
4. embedded builtins

​

Supporting files

• examples/*.md
• templates/*.txt
• reference.md

• the definition file itself is excluded
• legal and meta files are excluded by default
• OpenAI-specific metadata paths such as agents/** are excluded
• image assets under assets/** are excluded
• symlinks are ignored
• content is bounded and truncated for prompt safety
• invalid or unreadable files are skipped with diagnostics

​

Configuring supporting-file and retrieval limits

skills:
  supportingFiles:
    maxFiles: 8
    maxBytes: 4096
  retrieval:
    maxFiles: 8
    maxChunks: 12
    chunkBytes: 1200
    chunkOverlap: 200
    maxPromptBytes: 12000

• skills.supportingFiles.maxFiles: hard cap on supporting files read from disk for the active skill
• skills.supportingFiles.maxBytes: hard cap per supporting file read from disk
• skills.retrieval.maxFiles: max supporting files considered by retrieval and ranking
• skills.retrieval.maxChunks: max ranked chunks retained across supporting files
• skills.retrieval.chunkBytes: chunk size used for segmenting supporting-file content
• skills.retrieval.chunkOverlap: overlap between consecutive chunks
• skills.retrieval.maxPromptBytes: final prompt budget for supporting-file content

​

Invocation

skills:
  activationMode: auto   # default: auto | explicit

• auto (default): Reliant chooses the most relevant skill from message context.
• explicit: Reliant activates a skill only when the user invokes one directly.

skills:
  integrationMode: filesystem   # default: filesystem | tool

• filesystem (default): include filesystem location hints in <available_skills> and load supporting files for active-skill prompt context.
• tool: suppress filesystem location hints in <available_skills> and avoid injecting supporting-file payloads into <active_skill>.

​

Discovery boundary behavior

• External provider folders like .claude, .codex, and .agents are not scanned during runtime discovery.
• To use skills from those folders, import or install them into a Reliant-managed root first.

• /skill-name for an exact match
• /skill skill-name for unique prefix resolution

​

Progressive disclosure

• always injects compact <available_skills> metadata as canonical XML
• discovery is metadata-first for startup and selection efficiency
• injects full skill instructions only for the active skill
• retrieves supporting-file content with bounded chunking, relevance ranking, and prompt-budget caps
• adds prompt-level <skills_notice> entries for warnings
• emits chat info or warning notices for activation and supporting-file truncation or limit behavior

​

Compatibility matrix

• ✅ = fully supported in Reliant
• ⚠️ = partially supported or adapted

​

Explicitly not supported

• Claude or Codex plugin ecosystems and plugin execution contracts
• external-tool-specific runtime side effects that bypass Reliant tool governance
• external-tool installer UX semantics that conflict with Reliant installation and policy model

​

Migration guide

​

From external skill folders

1. Keep shared or team skills in .reliant/skills/
2. Keep personal variants in .reliant.local/skills/
3. Use Settings → Skills import actions or install_skill to copy from external folders into Reliant roots

​

Skill installer

Feature gate: the entire skills feature, including runtime activation, settings APIs, and install_skill, is disabled unless features.skills_enabled=true.

• local skill directory path that contains SKILL.md
• git URL with optional source_subpath
• GitHub tree and blob URLs

• project → .reliant/skills/...
• project_local → .reliant.local/skills/...
• global → ~/.reliant/skills/...

• skip (default): keep existing files and skip conflicts
• overwrite: replace existing files
• rename: install into a suffixed directory name when destination exists

Note: rename is not supported for SKILL.md skills because Agent Skills require the skill name to match the parent directory name.

{
  "source": "/path/to/skill-dir",
  "scope": "project",
  "conflict_policy": "skip",
  "dry_run": true
}

{
  "source": "/path/to/skill-dir",
  "scope": "project_local",
  "conflict_policy": "overwrite"
}

{
  "source": "https://github.com/acme/skills.git",
  "source_subpath": "skills/incident-response",
  "ref": "main",
  "scope": "project",
  "conflict_policy": "rename",
  "dry_run": true
}

{
  "source": "https://github.com/anthropics/skills/tree/main/skills/frontend-design",
  "scope": "project",
  "conflict_policy": "rename",
  "dry_run": true
}

• strictly validates installer scope and conflict_policy values
• rejects symlinks in source and destination write targets
• rejects path traversal and escaping paths, including source_subpath
• uses staging and rollback semantics to avoid partial installs on failure
• performs post-install discovery validation
• skips excluded non-runtime files so unusable files do not consume disk or prompt budget

​

Frontend skills management

• view discovered skills in .reliant.local/skills, .reliant/skills, and ~/.reliant/skills
• see scope, format, and definition location
• install from a backend-defined recommended skills section
• quick-import skills from common tool folders such as .claude/skills, .codex/skills, .agents/skills, and .cursor/skills
• run installer directly from the UI in dry-run or install mode
• copy ready-to-run installer prompts for chat execution
• copy skill definitions on demand from summary rows
• copy a starter prompt for the built-in skill creator

• Direct execution from Settings → Skills
• Assistant-assisted execution via generated prompts and chat

​

Codex interoperability notes

1. ingest what is portable
2. ignore or map what is tool-specific where safe
3. prefer Reliant-native controls for policy, limits, governance, and workflow attachment

​

Related topics

• MCP Servers
• Custom Workflows
• Presets