Goal Orchestrator

How the orchestrator runs (Path A)

Three sections render in the canvas:

Chat-style workflow — ONE Claude agent runs the loop. Tool calls + decisions stream into the action log live. <<DONE>> triggers blind Codex review. <<ASK_USER>> pauses for human input. <<NEW_SUBGOAL>> decomposes the goal.
Goal tree — projects contain goals; goals can decompose into sub-goals.
Scheduler — concurrency cap per goal, autospawn-on-start, depends_on sequencing, manual override.

Hover any bubble for the underlying behavior.

Models

• The working agent is claude -p --model claude-opus-4-7 --effort xhigh (Opus 4.7 at max reasoning).
• The blind reviewer is codex exec at the user's default config (gpt-5.5 at reasoning=xhigh).

What changed from the original

• No more propose / score / plan / score / pick ceremony.
• One Claude agent per workflow, streaming straight into the action log.
• Codex is the blind external reviewer (not a duplicate proposer).
• No 1-10 confidence gates — pauses happen only on explicit <<ASK_USER>>.
• Sub-goals can come from the agent itself, not just the user.
• Autospawn fires at iteration 1 (no separate execute phase to gate on).

Loading goals…

—

Send a nudge…