deploy

name: deploy description: > Roll out the latest local code to the Clack GCE VM. Runs scripts/gce-update-image.sh in the background, surfaces each phase (build → push → prune → pull → drain → swap → ready) via a Monitor, reports the downtime, and finally checks whether local tool mappings have diverged from the VM — if so, prompts to push them too. Trigger when the user says "deploy", "deploy again", "deploy now", "ship it", "redeploy", or any near variant.

Deploy to GCE

Orchestrates the standard image-update deploy for the Clack VM. Replaces the manual sequence of "kick off bash, arm monitor, ack each phase, extract downtime."

Step 1 — kick off the deploy in the background

Bash(
  command: "bash scripts/gce-update-image.sh",
  description: "Deploy",
  run_in_background: true
)

Note the returned task_id (e.g. bey7dhw8e) AND the output file path (/private/tmp/.../tasks/<task_id>.output). You need both.

Step 2 — arm a Monitor with the standard phase filter

Monitor(
  description: "deploy progress",
  timeout_ms: 900000,    # 15 min — safely above the script's 5-min readiness wait
  persistent: false,
  command: "tail -f <OUTPUT_FILE> | grep -E --line-buffered \"✓|✗|ERROR|error:|failed|denied|no space|Pre-pulling|Draining|Bot idle|Drain timeout|Drain check skipped|Stopping old|Waiting for|Bot is ready|downtime|Step [0-9]+/[0-9]+ : FROM|Successfully built|Successfully tagged|^DONE|New image pulled|Total reclaimed\""
)

That filter catches every phase marker plus the failure modes the script itself surfaces.

Step 3 — acknowledge each phase event tersely

The user is watching the live stream; don't restate what they already see. One sentence per event, matching the marker:

Step 4 — when the bash task completes, extract the downtime

The bash task fires its own completion notification. On that notification:

Bash(command: "grep -E 'downtime|Bot is ready' <OUTPUT_FILE> | tail -1")

Report it as **Downtime: 28s.** (the actual seconds).

Step 5 — check tool mappings divergence

After downtime is reported, compare every local data/default_configuration/tool_mapping/*.json against the VM:

DIVERGED=()
for f in data/default_configuration/tool_mapping/*.json; do
    name=$(basename "$f")
    local_md5=$(md5 -q "$f")
    remote_md5=$(gcloud compute ssh clack --zone=northamerica-northeast1-a --quiet \
        --command="sudo md5sum /mnt/disks/clack-data/data/default_configuration/tool_mapping/$name | cut -d' ' -f1" 2>/dev/null)
    [ "$local_md5" != "$remote_md5" ] && DIVERGED+=("$name")
done
[ ${#DIVERGED[@]} -gt 0 ] && printf 'DIVERGED: %s\n' "${DIVERGED[@]}" || echo "IN SYNC"

• If IN SYNC → say nothing further beyond the downtime line.

• If any files diverged → say:

  Local tool mappings differ from the VM: <file1>, <file2>, ... Push them with bash scripts/gce-push-config.sh?

  Wait for the user's confirmation. If they say yes, run:

  Bash(command: "bash scripts/gce-push-config.sh --force 2>&1 | grep -vE 'LIBARCHIVE\\.xattr|known_hosts' | grep -E '✓|Streaming|✗'")
  

  Deploy context implies overwrite intent, so --force is appropriate here (the safety check is for accidental clobbers, not authorized ones).

Step 6 — handle the stale monitor event

After the bash task completes, the Monitor often emits one final notification a few minutes later: [Monitor timed out — re-arm if needed.]. That's expected. Acknowledge with Stale monitor. Idle. and stop.

Failure modes (from gce-update-image.sh)

The script exits non-zero on:

• Container crash during swap → script prints docker logs --tail 80 clack command
• 5-min timeout waiting for "Clack is ready" → script prints docker logs -f clack command
• no space left on device → boot disk full; the script's docker image prune -f before pull is meant to prevent this. If it recurs, check /mnt/stateful_partition usage on the VM.
• denied: Unauthenticated request on pull → the configure-docker step failed; usually a one-off and resolved by re-running.

In every case the script's stderr includes a copy-pasteable diagnostic command. Forward it to the user verbatim.

Drain phase (before swap)

Phase 1.5 waits for the running bot to go idle before the hard docker stop, so an in-flight Claude run isn't killed mid-answer. It probes the running container's /status endpoint (busy = active query runs OR executing Changes-Workflow runs) and proceeds the instant the bot is idle.

• A long drain wait is expected, not a hang — the bot is busy. The gate is bounded (DRAIN_MAX_WAIT, default 300s) and then swaps anyway, printing what was still active. Don't kill the deploy just because draining takes a while.
• Drain check skipped means the running image predates the /status endpoint (e.g. the very first deploy of this feature). Harmless — it proceeds.

Gotchas

• macOS ._* xattr files show up in tar-pipe diffs but are not real content differences. The gce-push-config.sh safety check still flags them — that's a known false positive. Use --force to bypass.
• The skill is image-only — it does NOT push config.json, mcp.json, or default_configuration/. Those live on the persistent disk and need gce-push-config.sh. The tool-mapping check at Step 5 catches the most common case where this matters.
• Don't poll for completion. The Bash background task and the Monitor both notify automatically.