v1.1.0 to v1.1.0

---

name: watch-rollout

description: >

After a merge, watch the GitHub Actions deploy run until the gating job lands, then run a

smoke test scoped to the change (browser for UI, GET for read-only API) against staging.

Goal-terminating loop — stops when the deploy lands and the smoke completes, or when it fails.

model-tier: standard

model: sonnet

effort: medium

version: "1.1.0"

author: "flurdy"

---

# Watch Rollout

Watch a post-merge GitHub Actions deploy to staging, then smoke-test the change once it lands —

the generic, GitHub-Actions-based cousin of `/watch-release` (which is hardwired to letterbox's

kubectl/CircleCI/Flux stack). Built for client repos with extensive CD workflows.

Chains naturally after `/ready-to-merge`: merge → watch the deploy job → confirm the change is live.

## Usage

```

/watch-rollout # deploy of latest origin/main commit, smoke staging

/watch-rollout 6790 # resolve via PR number (uses its merge commit)

/watch-rollout <sha> # specific commit

/watch-rollout --run 28440286944 # specific workflow run id

/watch-rollout --prod # allow a read-only prod smoke (safety-gated, see below)

/watch-rollout --no-smoke # watch the deploy only, skip the smoke test

```

## Procedure

### Phase 0 — Load config (optional)

Read `.claude/rollout.yaml` at the repo root if present. Recognised keys (all optional):

```yaml

workflow: CMS Pages # deploy workflow name to watch

gating_job: Deploy blc-uk # substring of the job that gates the target env

staging_url: https://legacy.staging.bluelightcard.co.uk

prod_url: https://www.bluelightcard.co.uk

smoke: |

Load /en logged-out, capture the first Amplitude "Page Viewed" request,

assert login_state=logged_out is in the event user properties.

```

Anything missing is inferred or asked for below. Never require the file — it's a convenience for

repos you watch repeatedly.

### Phase 1 — Resolve the commit and run

Determine the target commit (first match wins): `--run` id (skip to job listing) → PR number's

merge commit (`gh pr view {n} --json mergeCommit --jq .mergeCommit.oid`) → explicit sha → latest

default-branch commit:

```bash

~/.agents/skills/watch-rollout/scripts/default-head-sha.sh # fetches origin/main, prints its sha

```

Fallback if the script is unavailable — run as two steps, never `&&` (each stays prefix-matchable):

```bash

git fetch origin main -q

git rev-parse origin/main

```

List runs on that commit and pick the deploy workflow:

```bash

gh run list --commit {sha} --limit 20 --json databaseId,name,status,conclusion,workflowName,event

```

Choose the run by config `workflow`, else by name heuristic (`deploy`/`release`/`cd`/`publish`),

else **ask** via `AskUserQuestion` when more than one plausible candidate exists. Don't guess

silently across ambiguous workflows.

### Phase 2 — Identify the gating job

```bash

~/.agents/skills/watch-rollout/scripts/run-jobs.sh {run_id}

```

Emits a JSON object: `{status, conclusion, jobs: [{name, status, conclusion}]}`. Fallback if the

script is unavailable: `gh run view {run_id} --json status,conclusion,jobs` (read the raw JSON).

A deploy run is often a matrix (per brand / per region). Watch the **one job that gates the env you

care about** — the others don't block your target (e.g. `Deploy blc-uk [preview]` gates BLC-UK

staging; au/dds/prod jobs are irrelevant to it). Pick it by config `gating_job`, else infer from

the branch/PR context, else **ask** which job(s) gate the target. If tests gate the deploy job and

the deploy job isn't spawned yet, note that — a test failure means the deploy never starts.

### Phase 3 — Derive the smoke test (derive + confirm)

Skip entirely if `--no-smoke`. Otherwise assemble a candidate smoke from, in order: config `smoke`,

`/ready-to-merge`'s post-merge follow-up note (if this session just merged), the PR description, and

the Jira AC. Classify the change:

- **UI change** → browser smoke (Chrome MCP): a staging URL + what to observe (screenshot, a network

request, a console signal).

- **Read-only API change** → an HTTP `GET` + the expected status/payload assertion.

- **Neither / unclear** → present what you found and ask the user to supply the check.

Present the candidate (target URL + assertion) and let the user confirm or tweak **before** the

watch starts. Don't run a derived smoke unconfirmed.

### Phase 4 — Watch the deploy

Keep the resolved run id, workflow, sha, gating-job name, and confirmed smoke spec in the scheduling

prompt. Polling is unattended; the smoke step (Phase 5) may need the user present for a browser

permission prompt.

#### Pi protocol v1

If `watch_loop` is available, use this path instead of Claude scheduling:

1. Call `watch_loop` with `action: status`. Continue only when its result reports

`protocolVersion: 1`. If another watch is `armed`, `running`, or `paused`, do not replace it;

show the status and point at `/watch-status`, `/watch-stop`, or `/watch-resume`. If the version

differs, explain the mismatch and stop without scheduling anything.

2. Make this prompt self-contained with the resolved values before passing it to `action: start`:

```text

Load and follow the skill named `watch-rollout` now. This is one continuation tick, not new watcher setup. Watch GitHub Actions run {run_id} ({workflow} on {sha}); the gating job is "{gating_job}" and the confirmed smoke is {smoke spec with URL, or "disabled (--no-smoke)"}. Run ~/.agents/skills/watch-rollout/scripts/run-jobs.sh {run_id}. Render the current gating-job status as visible text. If it is in_progress, waiting, queued, or not yet started and no upstream job has failed, call the matching `watch_loop` action: complete with outcome: continue. If it or an upstream prerequisite failed, report the failure and complete with outcome: stop. On gating-job success, either report deploy success when smoke is disabled, or run the confirmed read-only smoke and report pass/fail with captured evidence; then complete with outcome: stop. Do not complete before rendering the status or terminal evidence. Never re-trigger, cancel, re-run, or approve a workflow; never deploy.

```

3. State that the watcher starts after about one minute, polls every four minutes, and is capped at

30 ticks (just under two hours). Then make the terminating start call:

```yaml

action: start

protocolVersion: 1

label: Rollout

mode: fixed

initialDelaySeconds: 60

intervalSeconds: 240

missedCompletionPolicy: retry

maxTicks: 30

tickPrompt: <the prompt above>

```

A pending tick uses `action: complete` with `outcome: continue`. Deploy failure or completed smoke

uses `action: complete` with `outcome: stop` and a concise reason. Protocol v1 also exposes

model-facing `action: stop` for a terminal abort with the matching watcher and generation tokens.

The runtime owns the cadence and finite budget; do not set `allowIndefinite`.

#### Claude Code fallback

If `watch_loop` is unavailable, retain the existing `/loop` path. Hand it the same resolved values

in a self-contained dynamic-loop prompt:

```

/loop Watch GitHub Actions run {run_id} ({workflow} on {sha}).

Run: ~/.agents/skills/watch-rollout/scripts/run-jobs.sh {run_id}.

While the gating job "{gating_job}" is in_progress, waiting, queued, or not yet started,

reschedule ~240s.

On gating-job success → if smoke is disabled, report deploy success and stop; otherwise run the

smoke test: {smoke spec, with URL}. Report pass/fail with captured evidence.

On gating-job failure → report which job failed and stop. Stop the loop once the smoke is done or the deploy failed.

```

If neither `watch_loop` nor `/loop` is available, explain that recurring watches are unsupported

and stop. ~240s keeps each wake inside the prompt-cache window and catches completion promptly.

Both paths are **goal-terminating** — they end when the smoke completes or the deploy fails, not at

end of day.

### Phase 5 — Smoke test (the loop's terminal tick)

When the gating job is green:

- **Browser (UI):** drive a browser to the staging URL and capture per the spec — screenshot, a

specific network request, or a console signal. Pick the driver to fit the check:

- **Chrome MCP** (`mcp__claude-in-chrome__*`) — when you want it in the user's visible Chrome

session (shared cookies/auth, watching it happen). `/browser-screenshot` covers the screenshot

case. Captures network via `read_network_requests`, console via `read_console_messages`.

- **Playwright MCP** (`mcp__playwright__*`) — when a clean, headless, scriptable check suits better

(no visible session needed, deterministic network capture via `browser_network_requests`).

**Read-only navigation only** with either driver — no form submits, clicks, or actions that mutate

state; never trigger a JS dialog.

- **API:** `curl`/`gh api` a `GET` and assert the status code / payload field.

Report **pass/fail with the captured evidence** (the screenshot path, the network payload, the

status code), then stop the loop.

### Phase 6 — Offer to save config

If config was inferred (not loaded from `.claude/rollout.yaml`) and the run went cleanly, offer to

write the resolved `workflow`, `gating_job`, URL, and `smoke` to `.claude/rollout.yaml` so the next

run is one command. Only on explicit yes.

## Safety rules

- **Staging by default.** Smoke prod only with `--prod` **and** only for strictly read-only checks

(GET / navigation). **Never** issue writes (POST/PUT/PATCH/DELETE), submit forms, or perform

auth-mutating actions against prod. If a derived prod smoke isn't clearly read-only, refuse it and

ask.

- **Watch only.** Never re-trigger, cancel, re-run, or approve a workflow; never deploy. This skill

observes a rollout, it doesn't drive one.

- **Don't thrash.** If a browser permission is denied or a fetch fails twice, stop and report — don't

retry in a loop.

## Failure modes

- **No deploy workflow on the commit** → report it and stop (the merge may not trigger a deploy, or

the workflow runs on a tag/schedule instead).

- **Gating job never spawns** (gated behind failing tests/lint) → detect the upstream failure, report

which job failed, stop.

- **Ambiguous workflow or gating job** → ask; never guess across candidates.

- **Browser blocked** (bot detection, auth wall, permission denied) → stop and report what was

reached, rather than retrying.

## Notes

- Goal-terminating, so unlike `/watch-prs` and `/watch-release` it does not take a stop-hour.

- This skill is for GitHub-Actions-deployed repos. For CircleCI + FluxCD repos use

`/watch-flux-rollout` (same shape, kubectl/CircleCI watch); for letterbox's multirepo release

flow use `/watch-release` / `/release-status`.