NAME

PLAN — generate and execute step-by-step task plans with bounded context

SYNOPSIS

/plan TASK /plan next /plan status /plan show /plan clear

DESCRIPTION

/plan breaks a complex task into a numbered GFM checklist, saves it to agents/plan.md, and executes each step using a bounded context — only the system prompt and the current plan state are sent to the model per step. This keeps per-step turn times constant regardless of conversation length and allows large multi-step tasks without filling the context window.

The plan persists to agents/plan.md across /clear and Harvey restarts. Use /plan status at any time to review progress.

SUBCOMMANDS

/plan TASK Ask the model to generate a step-by-step GFM checklist for TASK and save it to agents/plan.md. Each step becomes a checkbox item: - [ ] Step description An existing plan is overwritten; use /plan clear first if you want a clean start on a different task.

/plan next Execute the next uncompleted step. Harvey sends only the system prompt and the current plan state — not the full conversation history — keeping context usage bounded. When a step’s tool calls are blocked or fail, the step is NOT auto-marked complete; fix the underlying issue and run /plan next again.

/plan status Print the plan checklist with completion markers, showing which steps are done and which remain.

/plan show Print the raw agents/plan.md file.

/plan clear Delete agents/plan.md. Does not affect conversation history.

BOUNDED CONTEXT MODEL

Each /plan next call sends only: 1. The system prompt (HARVEY.md). 2. The current agents/plan.md content.

The full conversation history is NOT included. This means: - Per-step token usage is constant regardless of plan length. - The model has no memory of earlier steps beyond the plan file. - For steps that need context from prior steps, inject that context explicitly via /context add before running /plan next, or note it directly in agents/plan.md.

WORKFLOW EXAMPLE

  # Break a large task into a plan
  /plan Refactor the auth package to use short-lived JWT tokens

  # Review the generated steps
  /plan status

  # Execute one step at a time
  /plan next
  /plan next

  # Check progress
  /plan status

  # If a step gets stuck, investigate and retry
  /plan next

FILES

agents/plan.md — persisted plan checklist; human-editable