[REPLACE: SKILL_NAME] Skill
监控 [REPLACE: VERCEL_PROJECT] 的 Vercel 部署——在过去 [REPLACE: LOOKBACK_HOURS] 小时内对 [REPLACE: ALERT_ON] 发出提醒
安装方式:把技能目录放入 ~/.claude/skills/(Claude Code)或在 claude.ai 设置中启用;也可复制右侧安装命令一键添加。
技能指令原文(SKILL.md)
${var} — Optional. Override the Vercel project slug. If empty, watches [REPLACE: VERCEL_PROJECT].
Today is ${today}. Watch Vercel deployments for [REPLACE: VERCEL_PROJECT] and alert on [REPLACE: ALERT_ON] within the last [REPLACE: LOOKBACK_HOURS] hours.
Required secrets
VERCEL_TOKEN— personal access token from https://vercel.com/account/tokens. Read scope is enough.- (optional)
VERCEL_TEAM_ID— if the project lives under a team, set this so the API queries the right scope.
Read the "Keys set for this run" line in your run notes, never a shell test (those are refused and cost a turn). If VERCEL_TOKEN is not on it, log DEPLOY_WATCH_NO_TOKEN and exit cleanly - never abort the workflow.
Steps
- Resolve scope: the project is
${var}if set, else[REPLACE: VERCEL_PROJECT]; write it into each URL literally (thebelow). IfVERCEL_TEAM_IDis on the "Keys set for this run" line, print it with its own call,node -e 'console.log(process.env.VERCEL_TEAM_ID)'(a team ID is not a credential, so./secretcurldoes not fill a{VERCEL_TEAM_ID}placeholder), and add&teamId=with the printed value to every Vercel URL below.
- Fetch recent deploys — Vercel API v6 lists deployments for a project:
Get the window start in milliseconds with its own call, date -u -d '[REPLACE: LOOKBACK_HOURS] hours ago' +%s000, and write the printed value in as :
./secretcurl -s -o /tmp/vercel-deploys.json -w 'vercel http=%{http_code}\n' -H "Authorization: Bearer {VERCEL_TOKEN}" "https://api.vercel.com/v6/deployments?projectId=<PROJECT>&since=<SINCE_MS>&limit=20"
On a non-2xx code, log DEPLOY_WATCH_FETCH_FAIL http= and stop.
Make every Vercel call in-run with ./secretcurl (write the key as {VERCEL_TOKEN} — a bare $VERCEL_TOKEN on the line is refused by the Bash permission layer). Read-only status checks and any irreversible action (e.g. triggering a deploy) both run in-run — the irreversible one as the skill's final, fail-closed action. Never defer a read.
- Parse and classify — for each deploy, capture:
uid, state (READY / ERROR / CANCELED / BUILDING / QUEUED), url, target (production / preview), creator, createdAt, meta.githubCommitMessage.
- Apply the alert filter —
[REPLACE: ALERT_ON] is one of:
production-failures → alert when target=production AND state in {ERROR, CANCELED}.
any-failures → alert on any state in {ERROR, CANCELED}.
slow-builds → alert when build time > 10× the last-week median for this project.
all → alert on every state transition (noisy — only useful while debugging the skill).
- Compare against last-success baseline — if alerting on a failure, also fetch the most recent successful production deploy and include in the notification: "last green: [commit] · [N hours] ago".
./secretcurl -s -o /tmp/vercel-last-green.json -w 'vercel http=%{http_code}\n' -H "Authorization: Bearer {VERCEL_TOKEN}" "https://api.vercel.com/v6/deployments?projectId=<PROJECT>&target=production&state=READY&limit=1"
- Dedup — track alerted deploy UIDs in
memory/topics/[REPLACE: SKILL_NAME]-alerted.json. Never re-alert for the same UID.
- Notify on every new alert via
./notify:
*Deploy alert — [REPLACE: VERCEL_PROJECT]*
${state}: ${commit_message}
${target} build by ${creator} · ${ago}
Last green: ${last_green_commit} · ${last_green_ago}
Inspect: https://vercel.com/${owner}/${PROJECT}/${uid}
- Write a roll-up to
output/articles/[REPLACE: SKILL_NAME]-${today}.md: total deploys, success/fail counts per target, average build time, list of failed UIDs with commit messages.
- Log to
memory/logs/${today}.md:
## [REPLACE: SKILL_NAME]
- **Deploys (${LOOKBACK_HOURS}h)**: total=N, ready=X, error=Y, canceled=Z, building=W
- **Alerts fired**: N (deduped from M raw matches)
- **Status**: DEPLOY_OK | DEPLOY_QUIET (no deploys) | DEPLOY_ALERT | DEPLOY_DEGRADED
Network note
The Vercel API requires Authorization: Bearer {VERCEL_TOKEN}. A bare $SECRET on a command line is refused by the Bash permission layer, so make every Vercel call in-run with ./secretcurl (write the key as the {VERCEL_TOKEN} placeholder — it keeps the secret off the line). Both the read-only status checks and any irreversible action (like triggering a deploy) run in-run; the irreversible one goes last, as the skill's final fail-closed action. Never defer a read.
Shell shape. Each Bash call is one allowed command (pipes are fine). The run refuses a call that sets a variable (PROJECT=..., SINCE_MS=$(( ... ))), uses if / for, tests a key ([ -n "$VERCEL_TEAM_ID" ]), joins commands with ;, && or ||, redirects into a file (> .vercel-deploys.json), or puts a shell variable on a ./secretcurl line ($URL), and each refusal costs a model turn. So run date as its own call, write the project, team ID and window start into each URL literally, save replies with -o /tmp/, and write alerted.json, the article and the log with Write/Edit.
Constraints
- Dedup is non-negotiable. Re-running the same alert for the same deploy will train operators to mute the channel — once alerted, never again unless the deploy changes state.
- Production beats preview for alerting. A failed preview deploy is interesting but not urgent. Default to
production-failures until the operator opts into more.
- Compare against baseline. A failed build means more when paired with "last green was 3 hours ago" than alone.