groom-stale

groom-stale

Three-pass triage that uses only the Sentry MCP — no git operations, no PRs. Safe for unattended runs.

Inputs

• SENTRY_ORG_SLUG — required
• SENTRY_DEFAULT_PROJECT_SLUG — optional
• STALE_AGE_DAYS — default 30 (from userConfig)

If $ARGUMENTS contains --dry-run, log every decision but do not call update_issue. Default behavior is to apply changes.

Workflow

All passes use search_issues with the query field (literal Sentry syntax). To pull deeper details on a single issue, use get_sentry_resource with the issue short_id as resourceId.

Pass 1 — Close obviously stale

search_issues with query: "is:unresolved lastSeen:-${STALE_AGE_DAYS}d sort:lastSeen", limit: 50. For each result where:

• Total events in the last STALE_AGE_DAYS days is zero (verify via get_sentry_resource event count), AND
• The issue is at least 60 days old (firstSeen < now − 60d)

call update_issue with status: "ignored". Track these in a digest section labelled Closed as stale.

Pass 2 — Re-open regressions

search_issues with query: "is:resolved lastSeen:-7d sort:lastSeen", limit: 50. These are resolved issues seeing new events in the last week — almost always regressions. For each one:

• Confirm at least 5 new events since the resolve timestamp (visible in get_sentry_resource).
• Call update_issue with status: "unresolved".

Track these as Re-opened regressions. Do not also assign — leave that for humans or for Pass 3.

Pass 3 — Assign hot unassigned issues

search_issues with query: "is:unresolved is:unassigned has:stack sort:freq", limit: 20. For each:

• Pull get_sentry_resource to read the top-of-stack file path.
• Run git log -n 5 --format='%an|%ae' -- <file> in the current repo to find the most recent authors. If the file doesn't exist in the current repo, skip — this skill is conservative and won't guess owners across repos.
• Pick the most recent author. Look them up in Sentry via find_teams / member lookup; if there's a matching member, call update_issue with assignedTo: "user:<member-id>".
• If no match, fall back to the team that owns the project (from find_projects) via assignedTo: "team:<team-slug>".

Track these as Assigned. Cap this pass at 20 assignments per run to avoid notification storms.

Final — Post digest

Print a markdown digest to stdout:

# Sentry grooming digest — <ISO date>
**Org**: <org>  **Project**: <project or "all">

## Closed as stale (N)
- <SHORT-ID> <title> — last seen <relative time>

## Re-opened regressions (N)
- <SHORT-ID> <title> — <N> new events since resolve

## Assigned (N)
- <SHORT-ID> → @<handle> (suspected from <file>)

## Skipped (N)
- <SHORT-ID> — <reason>

If a Slack connector is enabled on the user's claude.ai account, also post the digest to the channel the user has configured (the routine prompt asks for the channel; for interactive use just print).

Hard rules

• Never delete issues. update_issue to ignored is the strongest action.
• Cap each pass at 50 issues to prevent runaway calls.
• Never assign to a user the skill cannot verify is a Sentry org member.
• --dry-run produces the same digest but skips every write call.