菜单
Use when inspecting or repairing desired-state reconciliation, latest-only state, pending or failed retries, superseded deployments, target locks, and pull-agent status.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
| name | deployment-reconciler |
| description | Use when inspecting or repairing desired-state reconciliation, latest-only state, pending or failed retries, superseded deployments, target locks, and pull-agent status. |
mcl deploy-reconcile --state-dir "$STATE_DIR" --event-log "$EVENTS_JSONL" --dry-run
mcl deploy-reconcile --state-dir "$STATE_DIR" --target "$TARGET" \
--target-host "$TARGET=$SSH_HOST" --ssh-user deploy \
--identity-file "$MCL_DEPLOY_SSH_IDENTITY" --ssh-option BatchMode=yes
mcl deploy-agent --target "$TARGET" --manifest-dir "$MANIFEST_DIR" \
--trusted-manifest-public-key "$MCL_DEPLOY_MANIFEST_PUBLIC_KEY" --dry-run
systemctl status mcl-deployment-reconciler.service mcl-deployment-reconciler.timer
systemctl status mcl-deploy-agent.service mcl-deploy-agent.timer
journalctl -u mcl-deployment-reconciler.service -u mcl-deploy-agent.service -b --no-pager -n 200
targets/<target>.json to identify the latest desired deployment.desired/, current/, failed/, superseded/, converged/, and
agent-status/ entries for the same deployment id.accepted and pending as retryable states unless a newer sequence
has superseded them.superseded as terminal; do not apply it after a higher sequence was
accepted.failed as retryable only when the event or agent status says
retryable: true and the retry budget is not exhausted.converged or succeeded as terminal unless the target health check
later proves false.flock -n lock ownership before starting another reconciler.State is latest-only per target. A lower sequence must not replace a higher
sequence. A same-sequence deployment with a different deployment id is
ambiguous for the pull agent and must be non-retryable. Retry state is bounded
by maxAttempts; exhausted retries become non-retryable. Already converged
deployments should short-circuit rather than restore or switch again.
current, failed, superseded, converged, and
agent-status.activate-requested, target apply phases, and final
status.Stop before deleting state files, rewriting sequences, forcing a superseded
deployment, clearing a non-retryable state, bypassing flock, increasing retry
budgets, or changing target-host mappings for production targets.
The reconciler does not invent rollback. It follows the signed manifest
rollbackPolicy and target-side apply result. If rollback must be manual, use
the break-glass skill and preserve the reconciler state files for audit.
Use when central deployment automation is unavailable and an approved operator must recover a target through direct SSH, signed manifests, forced-command deploy keys, and explicit rollback evidence.
Use when running deployment migration rehearsals in NixOS VM tests or Incus/LXC topology simulations, including check-env, check-runtime, dry-run, runtime launch, failure injections, and production evidence gates.
Use when inspecting Attic or Cachix deployment cache health, proving closure substitute coverage, repairing cache misses, or handling cache token and signing-key boundaries.
Use when performing normal deployment operations, confirming deployment health, running direct SSH deployment wrappers, or deciding whether to fall back to rollback.
Use when investigating a failed, stuck, superseded, or ambiguous deployment across GitHub Actions, deployment event artifacts, mcl status summaries, target journals, cache health, and post-switch health checks.