Code Plan

Goal

Produce one self-contained engineering plan that another agent can execute without losing the user's intent, context, boundaries, validation bar, implementation strategy, ordering, or stopping condition.

A plan is not just a Goal contract and not just a task list. It explains why the work matters, what must be true at the end, the approach and order, how to verify it, which risks to watch, and when to pause or stop.

You own the PLAN: which change is correct (the layer that owns the behavior, not the symptom site), whether that change is irreducible, the regression bar, the authorized boundary, and the risk-ordered human reading sequence. code-tasking owns the downstream EXECUTION GRAPH: it consumes this plan plus the real codebase, proves the ripple set empirically, re-sorts the work leaf-first by compile dependency, and makes each irreducible change unsplittable at the toolchain level for a memoryless executor. Drift firewall: if a statement could only be checked by running a command against the real codebase, it is execution-altitude and belongs in code-tasking; if it is a judgment about the chosen approach's shape, it belongs here.

What Every Plan Contains

Every plan is incomplete unless it contains these, in this order. Fill each once; do not restate the same content across sections.