claude-sdk-rs: Command for Claude Code

Seams — Map where new work attaches to the existing system.

Stage 2 of the pre-plan pipeline: /assess → /seams → /sequence → /plan. Method: docs/how-to-plan-with-agents.md in the brain repo.

Variables

$ARGUMENTS — the assessment slug, plus optional flags. Example: orchestration-extensions

Flag	What it does
--spike <n>	Run up to n spikes instead of the default 1. --spike 0 disables
--from <path>	Read an assessment that is not at planning/<slug>/assessment.md

Purpose

The assessment says what is true. Sequencing says what to build. This stage answers the question in between: where does the new work attach to the old, and what breaks if that attachment is wrong. It is the stage most often skipped and the one whose absence most reliably produces a plan that is coherent on paper and unbuildable in practice.

Output: planning/<slug>/seams.md.

It produces no blocks, no waves, no estimates. Those are /sequence.

Instructions

1. Resolve the slug. Read planning/<slug>/assessment.md, verification.md if present, and CLAUDE.md. Where assessment and verification disagree, verification wins. If no assessment exists, stop and point at /assess.

2. Classify every capability the work depends on into exactly one of three states. This is the core of the command; do not soften the categories.

   State	Definition	The planning consequence
   Built	Exists, is called in production, and has been observed working	Call it. Do not rebuild it
   Half-built	Exists in source but has no production caller, is behind a disabled flag, is untested, or was never actually run	This is where plans die. Each one is either a small wiring block or a large rewrite, and you cannot tell which without looking
   Absent	Does not exist	Honest new work

   For half-built, additionally record: what specifically is missing (a caller, a flag, a config value, a schema agreement), and how you know — with file + symbol. "Shipped and unused" and "shipped and broken" are different blocks; say which.