Smoke Test
Run the smoke checks declared per target in .claude/deploy-targets.json and
report whether the deployed thing actually answers.
disable-model-invocation: true is deliberate. These checks execute commands
the repository declares and send requests to production, so this skill runs
when a person asks for it, not when a model infers it might be useful.
Workflow
Resolve the helper script — the same two-step resolution deploy-status
uses, because it is the same helper and a skill that found only one of the
two locations would work for exactly one install shape:
if [ -f "$HOME/.claude/skills/deploy-status/scripts/deploy-ops.mjs" ]; then
DEPLOY_OPS="$HOME/.claude/skills/deploy-status/scripts/deploy-ops.mjs"
elif [ -f "skills/deploy-status/scripts/deploy-ops.mjs" ]; then
DEPLOY_OPS="skills/deploy-status/scripts/deploy-ops.mjs"
else
echo "deploy-status helper not found; re-run dotbabel bootstrap or dotbabel init" >&2
exit 2
fi
Run the smoke command from the consuming project root, forwarding only
the flags this skill documents:
node "$DEPLOY_OPS" smoke $ARGUMENTS
$ARGUMENTS is raw caller text spliced into a command that reaches
production, so pass only --dry-run and --json — drop anything else
rather than forwarding it. A value carrying ;, |, &, a backtick,
$(, or a newline is shell syntax, not a flag: refuse it and say so.
Preserve the helper's exit code:
0 — every declared check passed, or none is declared, or it was a dry
run. None of those is a failure.
1 — at least one check failed, or the run hit its 300-second budget.
2 — target discovery failed.
Read verdict, not just the exit code. The payload distinguishes four
states, and two of them exit 0 without having verified anything:
verdict |
Meaning |
pass |
Every declared check ran and passed |
fail |
At least one check failed, or the budget was hit |
not_configured |
No target declares smoke — nothing was checked |
not_run |
--dry-run; checks were listed, none executed |
Report not_configured and not_run as exactly that. Reporting either as a
pass claims a verification that did not happen.
On failure
Report the failing checks and recommend /rollback-prod. Then stop.
Never invoke /rollback-prod, and never run the rollback helper directly.
Rolling back is a production change, and both the rule floor and KD-13 require
an explicit human instruction for one. /rollback-prod additionally demands a
typed confirmation of its own, which exists precisely so that a chain of
automated judgments cannot end in an unattended production change.
Recommend it, name the failing checks, and let the person decide.
What the checks are
Declared per target in .claude/deploy-targets.json; see
skills/deploy-status/SKILL.md for the full schema and the security guards.
In short: http checks are GET only, retry 3 times with 2s/4s/8s backoff, and
abort after 10 seconds; command checks run their argv exactly once, because
a command may not be idempotent.
Secrets are read only from named environment variables, and header values never
appear in the output. If a check needs a token, export it before invoking:
SMOKE_BEARER=… /smoke-test
Rules
- Never roll back. Recommend
/rollback-prod and stop.
- Never deploy or promote. This skill only observes.
- Never invent a smoke check. If a target declares none, report
not_configured rather than guessing at a health endpoint.
- Never report a dry run as a pass.
--dry-run lists what would run and
verifies nothing; its verdict is not_run.
- Never print a header value. The helper redacts them; do not echo the
environment variable it read them from either.
1---2id: smoke-test3name: smoke-test4type: skill5version: 1.0.06domain:7 - infra8 - devex9platform:10 - vercel11 - flyio12 - aws13task:14 - diagnostics15 - runtime-ops16maturity: draft17owner: "@kaiohenricunha"18created: 2026-09-18T00:00:00.000Z19updated: 2026-09-18T00:00:00.000Z20description: |21 Run the smoke checks a repository declares for its deploy targets and report whether production actually answers after a release. Use after a deploy, during release verification, and when confirming a rollback landed. Triggers on: "smoke test", "run smoke checks", "is production healthy", "did the deploy work", "verify the release is live".22argument-hint: "[--dry-run] [--json]"23tools: Bash, Read, Grep, Glob24allowed-tools: Read Grep Glob Bash25model: sonnet26effort: medium27disable-model-invocation: true28user-invocable: true29dotbabel:30 compute:31 mode: dynamic32 requirement: routine33 binding: consumer34metadata:35 source: https://skillmd.com/skills/kaiohenricunha/smoke-test36---3738# Smoke Test3940Run the `smoke` checks declared per target in `.claude/deploy-targets.json` and41report whether the deployed thing actually answers.4243`disable-model-invocation: true` is deliberate. These checks execute commands44the repository declares and send requests to production, so this skill runs45when a person asks for it, not when a model infers it might be useful.4647## Workflow48491. Resolve the helper script — the same two-step resolution `deploy-status`50 uses, because it is the same helper and a skill that found only one of the51 two locations would work for exactly one install shape:5253 ```bash54 if [ -f "$HOME/.claude/skills/deploy-status/scripts/deploy-ops.mjs" ]; then55 DEPLOY_OPS="$HOME/.claude/skills/deploy-status/scripts/deploy-ops.mjs"56 elif [ -f "skills/deploy-status/scripts/deploy-ops.mjs" ]; then57 DEPLOY_OPS="skills/deploy-status/scripts/deploy-ops.mjs"58 else59 echo "deploy-status helper not found; re-run dotbabel bootstrap or dotbabel init" >&260 exit 261 fi62 ```63642. Run the smoke command from the consuming project root, forwarding **only**65 the flags this skill documents:6667 ```bash68 node "$DEPLOY_OPS" smoke $ARGUMENTS69 ```7071 `$ARGUMENTS` is raw caller text spliced into a command that reaches72 production, so pass only `--dry-run` and `--json` — drop anything else73 rather than forwarding it. A value carrying `;`, `|`, `&`, a backtick,74 `$(`, or a newline is shell syntax, not a flag: refuse it and say so.75763. Preserve the helper's exit code:77 - `0` — every declared check passed, or none is declared, or it was a dry78 run. None of those is a failure.79 - `1` — at least one check failed, or the run hit its 300-second budget.80 - `2` — target discovery failed.81824. **Read `verdict`, not just the exit code.** The payload distinguishes four83 states, and two of them exit 0 without having verified anything:8485 | `verdict` | Meaning |86 | ---------------- | ------------------------------------------------ |87 | `pass` | Every declared check ran and passed |88 | `fail` | At least one check failed, or the budget was hit |89 | `not_configured` | No target declares `smoke` — nothing was checked |90 | `not_run` | `--dry-run`; checks were listed, none executed |9192 Report `not_configured` and `not_run` as exactly that. Reporting either as a93 pass claims a verification that did not happen.9495## On failure9697Report the failing checks and **recommend `/rollback-prod`**. Then stop.9899**Never invoke `/rollback-prod`, and never run the rollback helper directly.**100Rolling back is a production change, and both the rule floor and KD-13 require101an explicit human instruction for one. `/rollback-prod` additionally demands a102typed confirmation of its own, which exists precisely so that a chain of103automated judgments cannot end in an unattended production change.104105Recommend it, name the failing checks, and let the person decide.106107## What the checks are108109Declared per target in `.claude/deploy-targets.json`; see110`skills/deploy-status/SKILL.md` for the full schema and the security guards.111In short: `http` checks are GET only, retry 3 times with 2s/4s/8s backoff, and112abort after 10 seconds; `command` checks run their `argv` exactly once, because113a command may not be idempotent.114115Secrets are read only from named environment variables, and header values never116appear in the output. If a check needs a token, export it before invoking:117118```bash119SMOKE_BEARER=… /smoke-test120```121122## Rules123124- **Never roll back.** Recommend `/rollback-prod` and stop.125- **Never deploy or promote.** This skill only observes.126- **Never invent a smoke check.** If a target declares none, report127 `not_configured` rather than guessing at a health endpoint.128- **Never report a dry run as a pass.** `--dry-run` lists what would run and129 verifies nothing; its verdict is `not_run`.130- **Never print a header value.** The helper redacts them; do not echo the131 environment variable it read them from either.