Use co-agent
Start with readiness
/co-agent:setup
/co-agent review the current diff
/co-agent decide between a queue and an event bus
/co-agent adr record the selected deployment model
Setup discovers each peer's plugin/CLI access path, performs a real probe, and records READY or the failure reason. A CLI binary existing on PATH is not proof that authentication or a model call works. Ordinary review, decide, and ADR work can continue solo with notice. Consensus and harness stop for setup when no peer is READY.
Choose an implementation workflow
/co-agent:consensus docs/spec.md
/co-agent:harness docs/spec.md
Consensus takes the document through planning, a plan gate, host implementation,
final review, and reporting. Harness keeps design, tests, verification, and commits
with the host. An eligible peer implements in isolated worktrees when READY;
when none is available, the workflow must explicitly select an authorized host
implementation plan with --allow-host-implementation. An enabled, fresh READY
raw-CLI reviewer remains mandatory. Fix invalid configuration or readiness before
continuing. Kiro can review but cannot be the external harness writer because it
lacks the required write sandbox.
Keep configuration, readiness and flag queries tied to the orchestration project where setup ran. Only the writer changes cwd to its task worktree. Follow the harness contract for the exact planner calls, exit statuses and capture/review gates.
Tune and inspect
/co-agent:configure
/co-agent:configure set autosync on
/co-agent:configure set pr_autofix max_iterations 5
/co-agent:sync-context
Inspect effective settings before changing models or reasoning effort. Committed defaults merge with user and repository-local overrides; the current host determines the applicable paths and peer roster. Use the configuration output and canonical defaults rather than copying model IDs from an example.
Read the result
The host sends comparable prompts, checks peer findings against source, separates confirmed problems from unsupported claims, and records peer failures. Large inputs can exceed a peer's configured context limit; missing coverage must be disclosed. Review verdicts are advisory and do not override repository CI or merge requirements.
Use commands for settings and workflow entry points, and PR autofix for the feedback loop.