Crate cap_rs_orchestrator

Expand description

cap-rs-orchestrator — headless engine that runs N collaborating CLI agents in one process, driven by a declarative fleet.yaml.

See docs/cap-orchestrator-engine-design.md.

Modules§

audit: Immutable, ordered record of every cross-session route the engine performs. Human-auditable per CAP’s “orchestrator-mediated, human-auditable” rule.
config: Declarative fleet.yaml schema + validation.
event: Types crossing the engine↔consumer boundary. This boundary is an in-process mpsc channel today and the seam for the future remote (WebSocket) layer.
executor: Deterministic state machine. Owns the registry + audit log; interprets the DSL to drive fan-out, joins, and routing. Runs in its own task; the consumer reads OrchestratorEvents from the returned channel.
factory: Constructs a concrete Driver for a session. Behind a trait so tests use scripted stubs and the engine uses real CLI agents.
real_factory: Builds real cap-rs drivers. Each first-class agent name maps to its highest-fidelity structured path:
registry: Owns all live sessions: maps id → inbox sender + task handle.
routing: Pluggable routing strategies for the orchestrator.
session: One tokio task per session, owning a Box<dyn Driver>. Communicates only over channels — no shared mutable state, no Mutex<Driver>.
worktree: Per-session workspace allocation. Default is one git worktree per session, branched off the fleet’s base_branch.

chat: Convenience façade: run an interactive chat fleet against real CLI agents in repo.
run: Convenience façade: run a fleet against real CLI agents in repo with the default static YAML routing strategy.
run_with_strategy: Convenience façade: run a fleet with a custom routing strategy.