claude-code-autoconfig: Command for Claude Code

Continue where the previous session in this terminal left off.

This is the no-ceremony wrapper around /recover-context's auto mode: no arguments, it figures everything out itself, and it is plan-aware (below). Recovery is mechanical re-reading — probe, read, reconcile — never the resumed work itself. So this command ends its turn at the report: state recovered, the one next action named, a go-ahead question. The user's next prompt ("go") starts a fresh turn with the recovered context already in the conversation.

⛔ End the turn on a plain text question — never an AskUserQuestion picker. A picker answer continues the SAME turn, so the work it green-lights runs inside the recovery turn — exactly what the stop exists to prevent. This overrides any local end-of-turn convention that prefers a picker for closed choices. Ask in text, end the message, wait.

No model pin, no model advice (v18). Claude Code ignores command model: frontmatter at runtime — a known bug (anthropics/claude-code#45191, closed "not planned") — so v9–v16's model: sonnet pin never actually switched anything (measured 2026-08-23 on 2.1.220: 85 /continue runs, zero switches at the command boundary) and was removed rather than left looking load-bearing. Do not re-add it. Do not surface model advice either: v17's "⚠ recovery ran on {model} — /model sonnet first" report line was cut 2026-08-27 as UI noise — /clear then /continue must stay a two-step habit with zero model choreography around it. The recovery steps are identical on any model; the probe's currentModel field is informational only — never report it, never act on it.

It is also plan-aware: when the previous session was executing a substep of a phased plan doc (the plan-authoring pattern — a docs/*.md or .claude/plans/*.md with a ## Ledger section) and stayed on that plan through to the end of the session, the plan doc + Ledger is the handoff, not the transcript. If the session moved on to something else afterward, the transcript's tail is the handoff — see Step 3.

Every probe runs in ONE call (Step 1). Recovery cost is round trips × resident context, and this command used to spend eight to fifteen sequential round trips on pure filesystem reads whose results never changed what the next probe did — while the context they recovered was ~700 tokens at p50 (measured 2026-08-07 across 245 real /continue sessions). The probes were separate only because this doc used to write them as separate prose steps. Everything below Step 1 is a DECISION, not a lookup.

Step 0: Announce (the turn's first output, before any tool call)

Say exactly:

Recovering context for the previous session's last active use case.

Say that sentence in full — not the vaguer "Recovering the previous session's context.", and not a trimmed hybrid that drops "context for". Both halves earn their place: "context for" names what is being rebuilt, and "last active use case" names how much of it comes back — the in-flight tail of that session (cutoff ladder + 3k cap) or its plan Ledger, never the whole transcript.