Skip to content

[feature request] Add programmatic APIs for slash commands, subagents, background tasks, and plugin reload #156

Open

Description

@shayne-snap

Summary

I’m building oh-my-antigravity

Antigravity CLI currently supports slash commands, plugin skills, hooks, and subagents inside the interactive
host session, but there does not appear to be a stable CLI/API surface for plugin runtimes to invoke those
capabilities programmatically.

This makes it difficult for workflow plugins to build durable orchestration features such as autopilot loops,
review gates, multi-phase execution, or team workflows.

Current limitation

From local CLI discovery, agy exposes plugin lifecycle commands such as install/list/enable/disable/
validate/link, but I could not find stable commands for:

  • running a slash command or plugin skill programmatically
  • invoking a named subagent from an external runtime
  • starting or waiting on a background task via structured JSON
  • reloading plugins without restarting the CLI session

Inside the interactive Antigravity session, these capabilities seem to exist at the host level. For example,
agents can define/invoke subagents and long-running tasks can continue asynchronously. But plugin CLIs cannot
reliably drive those same capabilities.

Related use case

I am building a workflow plugin that tries to model multi-phase autonomous workflows:

plan -> execute -> code review -> fix/replan -> final verification

The plugin can store durable state itself, but it cannot currently invoke the next slash command, spawn a
subagent, or wait on host-level background work through a documented API.

A stable API for these host capabilities would make advanced Antigravity plugins much more reliable.

Requested APIs

It would be very useful to expose something like:

agy skill run omagy:code-review --input '<text-or-json>' --json
agy subagent invoke --agent code-reviewer --prompt '<prompt>' --json
agy task wait <task-id> --json
agy plugin reload

Equivalent API shapes would also work. The important part is that plugins can call these capabilities in a
structured, stable way.

Why this matters

Workflow plugins often need to coordinate multiple host-level capabilities:

  • /plugin:autopilot may need to run planning, implementation, and review phases.
  • /plugin:code-review may need to invoke reviewer subagents and synthesize results.
  • /plugin:team may need to launch multiple agents and track completion.
  • Long-running workflows need durable state plus a reliable way to resume/wait/cancel.
  • During plugin development, restarting the whole CLI session to reload plugin changes slows iteration.

Without programmatic APIs, plugins have to rely on prompt instructions and durable CLI state, while the parent
agent manually follows the orchestration contract. That works, but it is less reliable than a first-class host
API.

Desired behavior

A plugin runtime should be able to:

  1. Start a slash command or skill by name.
  2. Pass structured input.
  3. Receive a structured task/session id.
  4. Poll or wait for completion.
  5. Read structured output/status.
  6. Cancel or recover stalled tasks.
  7. Reload plugin definitions during development without restarting the whole CLI.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Labels

comp:agent capabilitiesBackground Tasks & Subagents, Permissions, Sandboxcomp:customizationsPlugins & Skills, MCP, Status Line, Window Titlefeature requestRequesting for new features

Type

No type

Projects

No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions