| name | horizon |
| description | Run a durable Horizon workflow for a multi-feature goal with bounded autonomous retries and an audit trail. |
| argument-hint | [long-horizon goal] |
| user-invocable | true |
Horizon
Switch the current core session with parallax_horizon. For $ARGUMENTS, resume only an explicitly matching Horizon session or initialize a unique ID with horizon_init_session.
Research first and persist concise findings. Write a schema-valid milestone and feature plan with measurable acceptance criteria. Use full Parallax protocol for implementation, refactors, architecture, security, and complex fixes; use none only for bounded research or trivial changes.
Execute one feature at a time. Keep plan and orchestration state synchronized and log consequential autonomous decisions. Evaluate each result with horizon_evaluate_subagent; it independently runs detected project checks and derives the verification dimension rather than trusting caller prose. Passing checks and an aggregate score of at least 75 are required; otherwise run a corrective pass. Never exceed the configured retry cap. Pause mid-run only for an external resource or authorization that cannot be researched or safely defaulted.
When dispatching parallax-claudecode:horizon-worker or parallax-claudecode:horizon-auditor, the first prompt line must be exactly HORIZON_DISPATCH {"sessionId":"<horizon-session>","featureId":"<feature>"} followed by the brief. A human brief alone is rejected because it has no durable session/feature identity.
Complete with a final debug audit, broad verification, archived traces where useful, consistent completion stats, and an honest report of completed, failed, and deferred work.