pr-review-sla

<skill name="pr-review-sla"> <overview> Catch a stalled PR the morning it stalls, not the week someone notices. A daily cron spawns a fresh session with read-only access to GitHub through the `gh` CLI; this skill pulls every open PR across `{{target_repos}}` and checks each one against three independent rules — overdue for first review, stale with no activity, and carrying requested changes the author hasn't addressed — then nudges whoever owns the next move in {{alert_channel}}.

Proactive and schedule-driven; recomputes from GitHub's current state every run, with no memory of the prior day's nudges. </overview>

<when-to-load> - The daily cron fires the PR review sweep. - A human asks which PRs are overdue for review, stale, or blocked on unaddressed feedback. </when-to-load> <workflow>

Step 0 — Orient

# List every open PR across the configured repos.
gh pr list --repo {{target_repos}} --state open \
  --json number,title,url,author,createdAt,updatedAt,reviewRequests,reviews,comments,commits

Fresh session, no ledger — recompute everything from this pull. Repeat per repo if {{target_repos}} names more than one.

Step 1 — Pull the review detail per PR

gh pr view --repo {{target_repos}} <PR_NUMBER> \
  --json number,title,url,author,createdAt,updatedAt,reviewRequests,reviews,comments,commits

reviews gives each review's state (APPROVED, CHANGES_REQUESTED, COMMENTED) and timestamp; commits and comments give the last activity; reviewRequests gives who's currently asked to review.

Step 2 — Flag PRs awaiting review past the SLA

For each PR with an outstanding reviewRequests entry and no APPROVED or CHANGES_REQUESTED review from that reviewer since the request:

hours_waiting = now - max(createdAt, last review-request timestamp)

Flag if hours_waiting > {{review_sla_hours}}. Nudge target: the requested reviewer(s).

Step 3 — Flag stale PRs

For every open PR, take the most recent of: last commit, last comment, last review.

days_idle = now - max(last_commit, last_comment, last_review)

Flag if days_idle > {{stale_days}}. Nudge target: the author. A PR already flagged in Step 2 can also be flagged here if it independently qualifies — the two checks are separate signals, not mutually exclusive.

Step 4 — Flag unaddressed requested changes

For each PR whose most recent review is CHANGES_REQUESTED, check whether the author has pushed a commit or posted a comment since that review's timestamp. If not, flag it. Nudge target: the author.

Check	Passes (no nudge)	Fails (nudge)
Overdue review	Reviewer submitted within SLA	No review after {{review_sla_hours}}h
Stale	Any activity within the window	No activity for {{stale_days}}d
Unaddressed changes	Author pushed/replied after the review	Silence since CHANGES_REQUESTED

Step 5 — Write the nudge

For every PR that trips at least one rule, write one line: the PR title and link, which rule(s) it tripped, and how long it's been in that state, e.g. "Awaiting review 3d (SLA 24h) — @reviewer" or "Stale 6d, no activity — @author" or "Changes requested 4d ago, not yet addressed — @author". A PR that trips more than one rule gets one entry listing all of them, not a duplicate nudge per rule.

Step 6 — Post to Slack

Post exactly one summary message to {{alert_channel}} for the whole sweep, grouped by rule (overdue for review / stale / unaddressed changes), each entry tagging the person it's nudging. This is a fresh session — there is no prior list to diff against, so the whole sweep is recomputed and reposted every day.

</workflow> <guardrails> - **Read-only, always.** Every GitHub call is a read (`gh pr list`, `gh pr view`, `gh api` GET). Never merge, close, approve, or dismiss a review, even if `GH_TOKEN` would technically allow it. - **One output.** The Slack post to {{alert_channel}} is the only thing that leaves the sandbox. No PR comments, no direct messages, no GitHub writes. - **No memory between runs.** Each run is a fresh session; recompute every PR's state from GitHub rather than assuming anything from the prior day's nudges. - **Scoped secret.** `GH_TOKEN` is injected as an environment variable at runtime, scoped to read access and to this agent's grant. - **People decide, not the agent.** The nudge flags what's blocking; a human reviews, pings, or closes the PR themselves. </guardrails> </skill>