claude-sdk-rs: Command for Claude Code

Orchestration Commander — a stateless drain that re-derives and reports the remainder

One drain: a single Claude turn, woken by scripts/commander_drain.sh via bastion ask, that routes this repo's message queue, re-derives the fleet's surfaces the only safe way (running the derivation, never guessing which dirty files are derived), and reports whatever is left over that a human or a lane needs to see. It never implements a block and never edits another lane's chain.

Variables

None. A drain takes no arguments — everything it needs is already on disk: the queue, the lease and registry records, and the repo's own planning/state.json. See "Stateless per drain" below for why that is deliberate, not a limitation.

The six steps, in order

1. Validate and survey the whole queue tree, then drain this lane's inbox

drain_queue() and complete_message() are Python functions inside scripts/check_messages.py — a drain is a Claude turn with shell access, not a Python process, so naming a function is not an instruction a drain can execute. Everything below is a shell command, run as written.

a. Validate every message record and layout invariant across every lane, not just this one's.

python3 scripts/check_messages.py --quiet

discover_queues() walks <lock_dir>/queue/<repo>/<lane>/ for every repo and lane under the resolved lock dir — this single invocation already covers the whole tree, which is exactly what thirteen consecutive drains never ran. Exit 0 means every record and receipt-backed transition in every lane's inbox//processing//done/ is well-formed; a nonzero exit means at least one is broken, and its FAIL <path> lines name which. Resolve <lock_dir> the same way the script does — --lock-dir, else FLEET_LOCK_DIR, else a brain.toml found by walking up from cwd — never assume a repo-relative path (see scripts/check_lane_agents.py's identical precedence).

b. Age-check every inbox file, and report an absent queue directory and an empty one as different findings. check_messages.py validates message shape; it does not report how long a file has waited, and it does not distinguish "this lane has never written a queue dir" from "this lane's queue dir exists and is empty." Both distinctions matter — a count of undrained messages reads as backlog, "10 hours old" reads as a failure; and "no queue dir for my lane" is why the first thirteen drains read "nothing to drain" instead of "never checked."

for q in "$LOCK_DIR"/queue/*/*/; do
  q="${q%/}"
  inbox="$q/inbox"
  if [ ! -d "$inbox" ]; then
    echo "ABSENT  $inbox"
  elif [ -z "$(ls -A "$inbox" 2>/dev/null)" ]; then
    echo "EMPTY   $inbox"
  else
    for f in "$inbox"/*.json; do
      [ -e "$f" ] || continue
      now=$(date +%s)
      mtime=$(stat -f %m "$f" 2>/dev/null || stat -c %Y "$f")
      echo "MSG     $f  age=$(( (now - mtime) / 60 ))m"
    done
  fi
done

Run this over every <lock_dir>/queue/<repo>/<lane>/ the tree contains — not only the drain's own lane. Every MSG line surfaced for a lane other than this one is reported, not acted on (step 4's board is where it lands) — routing another lane's message is that lane's call, never this drain's (see "Out of scope" precedent in the block that shipped this step, BT.ticket.commander-must-validate-the-whole-queue-tree).