Extensions and Skill loading

Choose by purpose, then use the relevant API guide. Skills, external tools and pets have distinct inputs and lifecycles.

Open

View on GitHub

Choose an extension

Mechanism	Purpose	Entry	API guide
Skill	Procedures, checklists, conventions	SKILL.md and references	This page
MCP client/server	External tools or phone-tool exposure	HTTP(S), JSON-RPC	MCP
API 2 pet	Floating appearance, animation, feedback	ZIP and ES module	Pet runtime
Built-in tool	Platform/application capability	Kotlin schema/Handler	Tools

Each follows its execution policy. Skills add no system permission; pet scripts do not own host approvals. Local MCP separately skips per-call chat approvals, as documented in its guide.

Skill package

SKILL.md
references/checklist.md
scripts/optional-script.sh

Describe triggers, prerequisites, steps and acceptance criteria in SKILL.md. Keep long references in relative files. Packaged scripts still require an execution environment and tools.

Import and management APIs

data/settings/AppSettings.kt provides these entries. Callers must handle Result failures:

API	Contract
importSkillFile(uri) / importSkillZip(uri)	Import Android URI
importSkillMarkdown(sourceName, skillText)	Install nonempty Markdown
importSkillRepository(repositoryUrl)	Resolve/download supported repository/raw URLs
importSkillZipBytes(sourceName, bytes)	Install existing ZIP data
installedSkills()	List installed metadata
setSkillEnabled(id, enabled) / updateSkillMeta(...)	Enable/disable or edit metadata
deleteSkill(id)	Delete installed content
listSkillFiles(id) / readSkillFile(id, path)	Read by ID and relative path

Loading lifecycle

Import

Resolve source → stage directory → validate entry, archive paths and size → metadata → final directory → enable. Remove temporary files on failure. Do not register success before download/validation.

Request

activeSkillsPrompt(forcedSkillIds) provides available names/descriptions. The Agent reads relevant entries/references through list_skill_files and read_skill_file. Turn-selected IDs are retained by OpenAiAgent.chat and cleared in finally, without leaking to other requests.

Execution

Actual tools execute procedures within their capability/path limits. Adapt unavailable desktop tools to Android/Termux. Successful import does not prove the procedure can run on-device.

Import example

// Existing host settings/text; perform downloads/reads on appropriate dispatchers.
val skill = settings.importSkillMarkdown("project-review.md", markdown).getOrThrow()
val files = settings.listSkillFiles(skill.id).getOrThrow()
val entry = settings.readSkillFile(skill.id, "SKILL.md").getOrThrow()

See the project review Skill for complete Markdown, checklist and ZIP. Verify metadata, enabled state, references, read-only execution and unmet dependencies, then check availability after updates/deletion.