ultra

Pursue the task below at high intensity by fanning it out across a fleet of peer engine agents, then synthesizing. The fleet bills to the OpenAI, xAI, and Cursor subscriptions, so the depth costs peer quota, not Claude quota.

Task: $ARGUMENTS

Size gate first (cheap, in the main loop):

• If the task is small, a single question, or a one file change, do NOT convene a fleet. Return it to the normal routing policy: questions stay in the main session, and requested changes normally use the Codex primary lane. Say you skipped the fleet because the task did not warrant it.
• Width is a judgment call, not a floor. A goal that stays narrow because its packages interlock, depend on each other, or are one coherent change needs no decline statement. The real defect is dispatching independent packages one per turn instead of together.
• Otherwise decompose the task into independent facets. Research decomposes by angle or source (different subsystems, different questions, different documents). Implementation decomposes by work package. Facet count follows the goal's natural package count with a floor of three: 6 to 8 facets when the seams allow it; up to about 12 only when the user asked for maximum coverage and the facets are truly independent. A three package goal is fleet shaped even though it sits below six.

Compose one self contained brief per facet:

• Each brief states its facet's goal, the shared context needed to work it alone, the relevant paths, and what a good result must cover. Claude worker facets use the fusion-brief: v1 isolated envelope. Implementation facets carry a verification command. Consult and research facets instead carry explicit coverage or acceptance criteria, with collection review as their verification. A facet brief never depends on another facet's output. Pass every brief directly as the Agent prompt; do not write transport files. Companion adapters own any private staging required by their transports.
• Every Grok facet brief begins with a single routing header line containing exactly one grok-role: <role> field, choosing one of burst, independence, live-web, or large-context from the facet's actual purpose. Do not include a second Grok role anywhere in that brief.
• Every Cursor facet brief begins with a single routing header line containing exactly one cursor-role: <role> field: diversity for a read only consult that needs a model family no other lane carries, or burst for a bounded implementation facet in its own orchestrator owned worktree.
• Research briefs go to grok with --web when live sources help; implementation briefs state write permission explicitly.

Launch the fleet in one message:

• The convening ceremony runs inside this skill, not prepaid by the main loop before invoking it: record the checkpoint or staged patch baseline when the tree is uncommitted, create the orchestrator owned worktrees for eligible independent Codex facets, and generate each brief's sibling declaration list mechanically from the decomposition so every facet's files appear in every other facet's forbidden list. The main loop's job is the decomposition decision and final judgment, not the staging ritual.
• The fleet changes capacity, not lane ownership. Codex remains the primary builder and deep reviewer. Grok supplies breadth under its burst, live-web, large-context, and independence protected roles. Cursor supplies cross family consults under diversity and Composer implementation breadth under burst.
• Invoke all independent Grok, Claude, and Codex facets together as parallel foreground-delivery Agent calls in one tool message. Do not set run_in_background. If the runtime still returns an async launch receipt, retain ownership and collect it before synthesis; Fusion's lifecycle hook blocks a turn that tries to abandon a Claude worker. Additional independent Codex implementation facets may use distinct orchestrator owned worktrees when their files are disjoint and verification needs no heavy setup. For each such facet, use --write --cwd "<absolute worktree path>" -- <brief> as the complete direct prompt to codex:codex-rescue. A Cursor burst facet uses the same direct prompt shape with cursor:cursor-rescue. A natural language Working directory: line does not select the Codex sandbox, and --cwd must never be appended to the Agent call, wrapper Bash command, or --raw-args-token invocation. Grok carries the remaining burst breadth, research digests, large context reads, and isolated scoped fixes. Packages that overlap files or require ordering are consolidated or sequenced, never fanned out as parallel overflow.
• The harness runs at most 20 subagents at once (CLAUDE_CODE_MAX_CONCURRENT_SUBAGENTS). A fleet message never launches more Agent calls than 20 minus the tasks still in flight; a wider decomposition launches the remainder as completion notifications free slots, and a refused launch is settled by the lifecycle as an infrastructure death rather than retried at once.
• Never set background mode on the Codex Agent or add --background to its brief. Complexity, expected duration, and model choice do not authorize detachment inside a fleet.
• Track the facets visibly with the harness's task list when one exists, otherwise enumerate the facets in the dispatch note. Never poll; completions arrive as notifications.
• If Grok is unavailable, keep one Codex facet in each available workspace, use isolated Codex or Cursor burst worktrees only for eligible independent packages, and send remaining generic implementation breadth to the Codex volume tiers; reserve fusion:claude-worker for packages whose load bearing capability is the Claude tool surface or privacy boundary, or that are structurally stranded with a stated claude-fallback: <reason> header. Use fusion:deep-reasoner for read only adversarial analysis. If Codex is unavailable, bounded and safely isolated implementation may use Grok under burst, while packages needing the Claude tool surface or privacy boundary, or that are structurally stranded with a stated claude-fallback: <reason> header, use fusion:claude-worker. Say every substitution in the synthesis.