claude-sdk-rs: Command for Claude Code

Handoff — Hand off an in-flight session cleanly to a fresh agent.

Use this when the current session has grown large enough that continuing in a new context is better than pressing on. It writes planning/handoff.md so the next agent can orient instantly after /prime, logs the work, and commits.

Variables

$ARGUMENTS — optional free-text note to include in the handoff (e.g. "focus on the parser next, the renderer is blocked on the tokenizer fix"). If omitted, derive context from git history and status.md.

Execution Model

Run entirely inline. Spawn no subagents, and do NOT invoke /log-work, /commit, or /backlog-ticket as skills — their steps are inlined below.

This is deliberate and is the whole point of the command. /handoff exists to serialize the current session's context to disk, and the main agent is the only thing that holds that context. A subagent cold-starts and has to reconstruct the session narrative from git log — slow, lossy, and exactly the reconstruction the handoff is meant to prevent. Delegating also fires at the end of a long session, when re-serializing instructions into a subagent prompt costs the most.

Budget: ~3 file reads, ~3 file writes, 2 shell calls. If you find yourself spawning an agent or reading more than five files, stop — you are re-deriving something you already know.

Instructions

Step 1 — Gather (3 reads + 1 shell call)

Read:

• planning/status.md — current focus, block statuses
• planning/state.json — tracks[].blocks[] statuses and the existing carryover[]
• planning/handoff.md — only if it exists (you are updating, not blindly replacing)

Run once: git log --oneline -10 && git status --short

Do not re-read log.md, the task spec, or docs/state/state-schema.md — you have the session context those would reconstruct. Read them only if you genuinely lack something.

Step 2 — Flip closed blocks, then drain durable context

2a — Flip any block this session closed to status: "closed" in planning/state.json tracks[].blocks[]. Do this before Step 4's emit-state: that authored field is the input the derivation reads, and emit-state never infers completion from status.md (the sync is one-way by design). Skipping this leaves focus and every generated surface stale until someone reconciles by hand — the engine-rs state-json-block-status-stale incident, 2026-07-03.