clp-plan

Context is the most important resource to manage in Claude Code. A plan that ignores context budget will fail midway through when compaction fires and critical state is lost. CLP-aware planning accounts for this by designing work in context-sized chunks with explicit checkpoint boundaries. Given the user's task description (provided as $ARGUMENTS or in the preceding message), create a CLP-aware execution plan:

<planning_steps> Check current context utilization with /clp:status. Calculate available working tokens for this task.

Break the task into phases that each fit within a single context window. Each phase should be independently completable — if compaction fires between phases, the handoff manifest should be sufficient to continue. Which phases involve heavy file reading or exploration? These should be delegated to subagents to keep the main context clean. Mark these as "subagent: true" in the plan. For each phase, estimate token cost: - File reads: ~100 tokens per KB of source code - Tool outputs: ~500-2000 tokens per bash command - Conversation: ~200 tokens per turn (user + assistant) - Skill loading: check estimated_tokens in skill-registry.json Insert /clp:checkpoint calls at natural breakpoints: - After each phase completes - Before any phase that might trigger compaction - Before switching from planning to implementation

<plan_format>

CLP Execution Plan: [task name]

Available budget: [X]k tokens remaining Estimated total cost: [Y]k tokens Phases: [N] | Checkpoints: [M] | Subagent tasks: [P]

Phase 1: [name] (~[X]k tokens)

• [step 1]
• [step 2]
• Skills needed: [skill IDs]
• Subagent: [yes/no — if yes, what it investigates]
• ⏸ Checkpoint after this phase

Phase 2: [name] (~[X]k tokens)

...

Execution notes

• [Any risks, dependencies, or ordering constraints]
• [Suggested Plan Mode vs Normal Mode for each phase] </plan_format>

- Every plan must include at least one checkpoint. - If estimated total cost exceeds 60% of available budget, split into multiple sessions. - Exploration phases (reading code to understand it) should always use subagents. - Implementation phases should be in the main context. - Suggest Plan Mode (Shift+Tab) for planning/analysis phases to halve token cost. - If $ARGUMENTS is empty, ask the user what they want to accomplish.