claude-sdk-rs: Command for Claude Code

Assess — Fan out recon agents over an existing codebase and produce a verified assessment.

Stage 1 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 topic, plus optional flags. Example: "autonomous multi-repo orchestration --slug orchestration-extensions"

Flag	What it does
--slug <name>	Output directory planning/<name>/. Default: kebab-case of the topic
--areas "a; b; c"	Name the recon areas explicitly instead of deriving them
--depth quick|standard|deep	3 / 5–6 / 8+ scouts. Default standard
--no-verify	Skip the verification pass. Only for a throwaway look — never before /plan
--resume	Re-read an existing assessment.md and run only the stages that are missing

Purpose

Turn "I want to understand this system well enough to plan work on it" into one dated, cited, independently re-checked artifact — planning/<slug>/assessment.md — that a later session can plan from without re-deriving anything.

This command produces evidence, not a plan. It may not propose a sequence, a wave, a block, or an estimate. Those are /seams and /sequence.

Scope. One coherent question about one system (possibly spanning repos). If you cannot state the question in a sentence, you have two assessments.

Instructions

1. If $ARGUMENTS is empty, stop and ask what to assess.

2. Frame the question. Restate the topic as one sentence and 3–8 numbered questions the assessment must answer. Show them to the user and proceed — these become the scout briefs and the section headings, so a vague question here produces a vague dossier.

   • Plan-quality floor. If framing the question requires inventing a load-bearing fact about intent (what "done" means, which repos are in scope, what the operator actually wants), STOP and ask 2–4 targeted questions. Do not guess the objective.