Rally

Rally is a turn-based coordinator for multi-agent coding sessions.

Its core job is simple: each agent polls for work and gets one clear instruction at a time.

Why It Exists

Rally helps agents follow pre-existing workflows without improvising protocol.

• keeps each agent on a strict loop (next -> do work -> done)
• reduces prompt ambiguity by issuing concrete next actions
• preserves workflow state and transitions centrally

Extension Model

Today, workflow extension is build-time:

• workflows are registered in Rust via WorkflowRegistry
• built-ins are wired in src/main.rs
• workflow behavior/state lives in workflow crates under crates/

Runtime plugin workflows are planned but not shipped yet. See todos/runtime-workflows-final.md.

Architecture Boundaries

• Rally core owns session envelope, persistence, locking, and dispatch.
• Workflow crates own workflow state schemas, transitions, and workflow-specific instruction text.
• Wrapper prompts in prompts/ are for /rally command resolution and loop bootstrap only.

If prompt text is specific to a workflow's behavior, keep it in the workflow crate, not in wrapper templates.

Skill Wrappers

Install wrappers:

rally skill install --target all

Canonical wrapper entrypoint:

rally skill run ...

Alias:

rally skill exec ...

Wrapper helpers:

• rally skill agent-howto ...
• rally skill context
• rally skill doctor --target ...
• rally skill uninstall --target ...

For install/run details, see docs/skill-install-run.md.