Skills
English | 中文
Skills are file-backed capability packs discovered at startup and exposed to the model as:
- lightweight metadata in the system prompt (name + description)
- on-demand content via the
readtool againstskill://... - optional interactive
/skill:<name>commands
This document covers current runtime behavior in src/extensibility/skills.ts, src/discovery/builtin.ts, src/internal-urls/skill-protocol.ts, and src/discovery/agents-md.ts.
What a skill is in this codebase
A discovered skill is represented as:
namedescriptionfilePath(theSKILL.mdpath)baseDir(skill directory)- source metadata (
provider,level, path)
The runtime only requires name and path for validity. In practice, matching quality depends on description being meaningful.
Required layout and SKILL.md expectations
Directory layout
For provider-based discovery (native/Claude/Codex/Agents/plugin providers), skills are discovered as one level under skills/:
<skills-root>/<skill-name>/SKILL.md
Nested patterns like <skills-root>/group/<skill>/SKILL.md are not discovered by provider loaders.
For skills.customDirectories, scanning uses the same non-recursive layout (*/SKILL.md).
Provider-discovered layout (non-recursive under skills/):
<root>/skills/
├─ postgres/
│ └─ SKILL.md ✅ discovered
├─ pdf/
│ └─ SKILL.md ✅ discovered
└─ team/
└─ internal/
└─ SKILL.md ❌ not discovered by provider loaders
Custom-directory scanning is also non-recursive, so nested paths are ignored unless you point `customDirectories` at that nested parent.
SKILL.md frontmatter
Supported frontmatter fields on the skill type:
name?: stringdescription?: stringglobs?: string[]alwaysApply?: booleanhide?: booleandisableModelInvocation?: boolean(Agent Skills equivalent ofhide; normalized from kebab-casedisable-model-invocation)- additional keys are preserved as unknown metadata
Current runtime behavior:
namedefaults to the skill directory namedescriptionis required for:- native
.musepiprovider skill discovery (requireDescription: true) omp-pluginsextension-package skills and thegithubprovider (.github/skills/), which also passrequireDescription: trueskills.customDirectoriesscans viascanSkillsFromDirinsrc/discovery/helpers.ts(non-recursive)
- native
- the claude/codex/agents/opencode/claude-plugins providers can load skills without description
Discovery pipeline
loadSkills() in src/extensibility/skills.ts does three passes:
- Capability providers via
loadCapability("skills")(the managed/auto-learn provider’s skills are skipped here and handled in pass 3) - Custom directories via
scanSkillsFromDir(..., { requireDescription: true })(one-level directory enumeration) - Managed (auto-learn) skills (
musepi-managedprovider) resolved dead-last with first-wins, so any same-named authored skill from any provider or custom directory takes precedence
If skills.enabled is false, discovery returns no skills.
Built-in skill providers and precedence
Provider ordering is priority-first (higher wins), then registration order for ties.
Current registered skill providers:
native(priority 100) —.musepiuser/project skills viasrc/discovery/builtin.tsomp-plugins(priority 90) —skills/bundled next to extension packages loaded throughextensions:,--extension/-e, or installed plugins under~/.musepi/plugins/node_modulesclaude(priority 80)- priority 70 group (in registration order):
claude-pluginsagentscodex
opencode(priority 55)github(priority 30) —.github/skills/<name>/SKILL.md(GitHub Agent Skills layout, project-only)musepi-managed(priority 5) — auto-learn skills under~/.musepi/agent/managed-skills, registered insrc/discovery/builtin.tsand discovered unconditionally (only writing/nudging is gated byautolearn.enabled); always defers to a same-named authored skill
Dedup key is skill name. First item with a given name wins.
Source toggles and filtering
loadSkills() applies these controls:
- source toggles:
enableCodexUser,enableClaudeUser,enableClaudeProject,enablePiUser,enablePiProject,enableAgentsUser,enableAgentsProject disabledExtensionsentries withskill:<name>ignoredSkills(exclude; glob patterns)includeSkills(include allowlist; glob patterns; empty means include all)
Filter order is:
- not disabled by
disabledExtensions - source enabled
- not ignored
- included (if include list present)
The agents provider (.agent[s]/skills) is the canonical OMP-native location and has its own enableAgentsUser/enableAgentsProject toggles — disabling Claude/Codex/Pi does not turn it off. For providers without a dedicated toggle (claude-plugins, opencode, gemini, github, …), enablement falls back to: enabled if any named source toggle is enabled.
Collision and duplicate handling
- Capability dedup already keeps first skill per name (highest-precedence provider)
extensibility/skills.tsadditionally:- de-duplicates identical files by
realpath(symlink-safe) - emits collision warnings when a later skill name conflicts
- keeps the convenience
loadSkillsFromDir({ dir, source })API as a thin adapter overscanSkillsFromDir
- de-duplicates identical files by
- Custom-directory skills are merged after provider skills and follow the same collision behavior
Runtime usage behavior
System prompt exposure
System prompt construction (src/system-prompt.ts) uses discovered skills as follows:
- if
readtool is available:- include discovered skills list in prompt, excluding skills with
hide: true
- include discovered skills list in prompt, excluding skills with
- otherwise:
- omit discovered list
hide: true does not disable the skill. Hidden skills are still loaded and remain reachable through skill://<name> and /skill:<name> when skill commands are enabled.
Task tool subagents receive the session’s discovered/provided skills list via normal session creation; there is no per-task skill pinning override.
Interactive /skill:<name> commands
If skills.enableSkillCommands is true, interactive mode registers one slash command per discovered skill.
/skill:<name> [args] behavior:
- reads the skill file directly from
filePath - strips frontmatter
- injects skill body as a custom message
- delivery mode follows the submission keybinding:
- Enter → invokes the skill on the
steerqueue while streaming (matches free-text Enter, which also steers), or as a normal idle prompt when the agent is not streaming - Ctrl+Enter (
app.message.followUp) → invokes the skill on thefollowUpqueue while streaming, or as a normal idle prompt when the agent is not streaming
- Enter → invokes the skill on the
- appends metadata (
Skill: <path>, optionalUser: <args>)
There is no flag, mode-selector, or frontmatter knob to override this — the keybinding is the choice, identical to how free text is routed during streaming (input-controller.ts:562-568 for Enter, input-controller.ts:961-966 for Ctrl+Enter; both dispatch through #invokeSkillCommand).
skill:// URL behavior
src/internal-urls/skill-protocol.ts supports:
skill://<name>→ resolves to that skill’sSKILL.mdskill://<name>/<relative-path>→ resolves inside that skill directory
skill:// URL resolution
skill://pdf
-> <pdf-base>/SKILL.md
skill://pdf/references/tables.md
-> <pdf-base>/references/tables.md
Guards:
- reject absolute paths
- reject `..` traversal
- reject any resolved path escaping <pdf-base>
Resolution details:
- skill name must match exactly
- relative paths are URL-decoded
- absolute paths are rejected
- path traversal (
..) is rejected - resolved path must remain within
baseDir - missing files return an explicit
File not founderror
Content type:
.md=>text/markdown- everything else =>
text/plain
No fallback search is performed for missing assets.
Skills vs AGENTS.md, commands, tools, hooks
Skills vs AGENTS.md
- Skills: named, optional capability packs selected by task context or explicitly requested
- AGENTS.md/context files: persistent instruction files loaded as context-file capability and merged by level/depth rules
src/discovery/agents-md.ts walks ancestor directories from cwd to discover standalone AGENTS.md files. For repositories nested under the user’s home directory, it continues through enclosing workspace directories up to but not including the home directory. With no repository root under home, the home boundary remains included. Otherwise it stops at the repository root, or at the filesystem root when no repository root is known outside home. Files in hidden owner directories are skipped.
Skills vs slash commands
- Skills: model-readable knowledge/workflow content
- Slash commands: user-invoked command entry points
/skill:<name>is a convenience wrapper that injects skill text; it does not change skill discovery semantics
Skills vs custom tools
- Skills: documentation/workflow content loaded through prompt context and
read - Custom tools: executable tool APIs callable by the model with schemas and runtime side effects
Skills vs hooks
- Skills: passive content
- Hooks: event-driven runtime interceptors that can block/modify behavior during execution
Practical authoring guidance tied to discovery logic
- Put each skill in its own directory:
<skills-root>/<skill-name>/SKILL.md - Always include explicit
nameanddescriptionfrontmatter - Keep referenced assets under the same skill directory and access with
skill://<name>/... - For nested taxonomy (
team/domain/skill), pointskills.customDirectoriesto the nested parent directory; scanning itself remains non-recursive - Avoid duplicate skill names across sources; first match wins by provider precedence