| # WIN2 chain-aware version of the coordinator's prev-ok gate (campaign launch/prev-ok.sh), for adoption by the coordinator. | |
| # Usage: WINDOW_COORDINATOR=Deepseek/coordinator chain_prev_ok.sh PREV_TMUX PREV_FILE|- (coordinator-only, S96 FIX 1) | |
| # Same as prev-ok: wait for PREV_TMUX, no HOLD, no stale containers. The previous window must have ended EITHER restored | |
| # with watch PASS (normal) OR chained (WINDOW_CHAIN=1, summary 'chain' == 'CHAINED', marker present). Production must be | |
| # serving K1 unless the chain marker says it is intentionally down (python -m offload.launch.windows.chain gate). | |
| # Exit 0: admitted (prints "chain gate: chained|production"). | |
| # Exit 2: refused with NO chain marker (production is not down because of a chain; as prev-ok.sh). | |
| # Exit 3: "CHAIN STOPPED: PRODUCTION DOWN" (S92 FIX 1): any refusal while the chain marker is present. Production was left | |
| # down by the previous chained window, so the gate runs the chain abort before exiting (chain_abort.sh: strict W8 | |
| # start|check|close + post-close watch when no HOLD exists, marker cleared only on RESTORED + watch PASS; alert and | |
| # touch nothing when a HOLD exists). The next window must not start; read the abort's summary.json. | |
| # Test-only overrides: DSV41_CHAIN_CAMPAIGN, DSV41_CHAIN_PY, DSV41_CHAIN_ABORT, DSV41_CHAIN_MARKER (chain.py). | |
| set -uo pipefail # no -e: every refusal goes through refuse() | |
| # S96 FIX 1: the chain abort below is coordinator-only; refuse at once (before any wait or probe) rather than at a later | |
| # stop, when the abort would be refused and production would stay down. RUN templates export it BEFORE the gate. | |
| [ "${WINDOW_COORDINATOR:-}" = Deepseek/coordinator ] || { echo "STOP: chain_prev_ok.sh is coordinator-only; export WINDOW_COORDINATOR=Deepseek/coordinator before the gate (its chain abort needs it)"; exit 2; } | |
| PREV_TMUX=$1 PREV_FILE=$2 | |
| C=${DSV41_CHAIN_CAMPAIGN:-/opt/dsv41/campaigns/DSV41-OFFLOAD} | |
| PY=${DSV41_CHAIN_PY:-/home/operator/miniconda3/bin/python3} | |
| HERE=$(cd "$(dirname "$0")" && pwd) | |
| cd "$HERE/../../.." || exit 2 | |
| chain() { PYTHONPATH=$PWD "$PY" -m offload.launch.windows.chain "$@"; } | |
| refuse() { | |
| # Only an explicit ABSENT counts as "no marker"; an unreadable answer is treated as present (restore, never ignore). | |
| if [ "$(chain present 2>/dev/null)" = ABSENT ]; then echo "$1"; exit 2; fi | |
| echo "CHAIN STOPPED: PRODUCTION DOWN ($1)" | |
| "${DSV41_CHAIN_ABORT:-$HERE/chain_abort.sh}" | |
| echo "chain abort exit $?; the chain is stopped (exit 3)" | |
| exit 3 | |
| } | |
| while tmux has-session -t "$PREV_TMUX" 2>/dev/null; do sleep 30; done | |
| if [ "$PREV_FILE" != - ]; then | |
| test -s "$PREV_FILE" || refuse "previous window never started" | |
| "$PY" - "$(cat "$PREV_FILE")/summary.json" <<'PY' || refuse "previous window ended neither restored+watch PASS nor chained" | |
| import json, sys | |
| s = json.load(open(sys.argv[1])) | |
| chained = s.get('chain') == 'CHAINED' | |
| restored = str(s.get('restore', {}).get('status', '')).startswith('RESTORED') and str(s.get('post_close_watch', {}).get('status', '')) == 'PASS' | |
| assert chained or restored, (s.get('chain'), s.get('restore'), s.get('post_close_watch')) | |
| PY | |
| fi | |
| MODE=$(chain gate) || refuse "chain gate refused" | |
| echo "chain gate: $MODE" | |
| "$C/launch/no-stale-containers.sh" || refuse "stale serve/experiment container check refused" | |