ULTRAPLAN

Remote planning system that offloads complex architectural planning to a Cloud Container Runtime (CCR) running Opus 4.6. While kairos handles ambient continuous operations (fast thinking), ULTRAPLAN handles episodic deep deliberation (slow thinking). The community drew the parallel to Kahneman's System 1 / System 2.

Architecture

ULTRAPLAN runs in a persistent cloud compute environment: - Read-only during planning — the remote session cannot push changes - No push notifications — the local terminal polls every 3 seconds - Browser UI — real-time observation and approve/reject workflow - Archival after execution — session archives after teleport to prevent parallel divergence

Three Variants

Variant Duration Complexity Token Budget
one_shot 10-15 min Simple planning ~100-200K tokens
two_pass Medium Moderate complexity ~200-400K tokens
three_subagents_with_critique 30 min max High complexity ~400-800K tokens ($4-8 at Opus rates)

The three-subagent variant spawns 3 research agents + 1 critique agent. Rarely used due to cost.

Sentinel Teleportation

When planning completes: 1. The local terminal polls and detects the sentinel value __ULTRAPLAN_TELEPORT_LOCAL__ 2. The "teleport approved" signal triggers 3. The remote session archives 4. Results transfer back to the local terminal

This ensures a single execution point — preventing divergence between remote and local state.

Two-Phase Best Practice

The recommended workflow: 1. Local plan (2-5 min, cheap) — quick initial planning in the terminal 2. Promote to ULTRAPLAN — hand off to remote CCR for deep deliberation 3. Browser review — watch the planning session in real-time 4. Teleport back — approve results back to local terminal

This consistently produces better plans than either approach alone.

Coordinator Mode

COORDINATOR_MODE (in coordinator/) implements one-worker-per-logical-CPU-core distribution. The coordinator encodes orchestration logic in natural language prompts rather than procedural DAGs — described as the most radical architectural choice in the multi-agent system.

The four-phase coordinator workflow: 1. Research — workers investigate in parallel 2. Synthesis — coordinator alone reads findings and crafts specs 3. Implementation — workers make targeted changes per spec 4. Verification — workers run tests and verify

Key Claims

Sources