Stuck-PR digest

Once a day, Notifycat reminds each channel about the open PRs nobody has touched since the previous day. It's the safety net under the quiet: a PR that got buried yesterday resurfaces this morning, every morning, until someone deals with it.

What it posts

Two items per channel with stuck PRs: a parent message carrying the count and pinging the channel's configured mentions, and a single threaded reply listing the PRs. The list lives in the thread, so the channel feed pays one line per day — and channels with nothing stuck get nothing at all.

Mentions sit on the parent because Slack thread replies don't notify the channel.

It's on by default

With no digest: section in config.yaml, the digest runs at 9am UTC on weekdays. This is deliberate — the digest is the "nothing slips through" half of the product. To opt out:

digest:
  enabled: false

Schedule and timezone

digest:
  enabled: true
  schedule: "0 9 * * *"      # standard 5-field cron
  timezone: "Europe/Kyiv"    # IANA zone; default UTC

Both the firing time and the "stuck since before today" cutoff are evaluated in timezone. An invalid cron expression or an unrecognized zone fails server startup — same fail-fast contract as the rest of the config. Setting the container's TZ variable does nothing; use this key.

Weekends and holidays

Weekends are always skipped. Saturday and Sunday never get a digest, whatever the schedule says — there is no key to turn this off. A digest nobody reads is worse than no digest: the same PRs are still stuck on Monday and get announced then. If your cron spec deliberately targets a weekend (0 9 * * 6), it will never fire.

Public holidays need a country, because holidays vary by country and there is no sane default:

digest:
  timezone: "Europe/Berlin"
  country: DE               # ISO 3166-1 alpha-2; case-insensitive

Both checks are evaluated in digest.timezone, so "Saturday" means Saturday where the team is, not in UTC.

With no country set, holidays are not skipped — only weekends. The server says so once at boot:

WARN digest holidays not configured detail="digest.country is unset; weekends are skipped but public holidays are not"

Supported countries

Code	Country	Code	Country	Code	Country
AT	Austria	FR	France	NO	Norway
BE	Belgium	GB	United Kingdom	PL	Poland
CH	Switzerland	IE	Ireland	PT	Portugal
DE	Germany	IT	Italy	SE	Sweden
DK	Denmark	LU	Luxembourg	UA	Ukraine
ES	Spain	NL	Netherlands	US	United States
FI	Finland

An unrecognized code does not fail startup. The server warns once, lists the supported set, and runs weekends-only — the same state as no country at all:

WARN digest country not recognized country=Germany detail="digest.country is not a supported country code; weekends are skipped but public holidays are not" supported="AT, BE, CH, ..."

That is deliberate: digest.schedule and digest.timezone decide whether the server runs at all and so fail fast, but the country only enriches one feature. A typo in it must not take the deployment down.

Each table carries that country's national public holidays. Two deliberate additions and several known limits:

• December 24 and 31 are treated as holidays in every country. In most of them they are not legal public holidays, but they are near-universal shutdown days for engineering teams. (In Poland, Wigilia became statutory in 2025, so there it is a real holiday.)
• Regional holidays are not covered. DE is the nine federal days — a Bavarian team still gets a digest on Fronleichnam. Likewise GB is England and Wales, ES and FR are the national sets, and CH is approximate by nature, since only August 1 is federal there and the rest is cantonal.
• IE models St Brigid's Day as the first Monday in February; the statutory rule shifts it to February 1 when that day is a Friday, so 2030 and 2036 are wrong.
• UA carries fixed-date holidays only. Orthodox Pascha needs Julian-calendar arithmetic and both Pascha and Trinity Sunday fall on a Sunday anyway; martial law currently suspends Ukrainian public holidays entirely, which no calendar can model.

Where a holiday falls on a weekend, the country's real rule applies: US shifts to the nearest weekday (Saturday back to Friday, Sunday forward to Monday), GB and IE add a substitute day — so Christmas 2027, a Saturday, is observed on Monday the 27th and Boxing Day on Tuesday the 28th — and continental Europe simply loses the day.

Skipped runs in the log

A skipped day makes no Slack call at all and logs one line:

INFO skipped digest schedule="0 9 * * *" reason=weekend date=2026-07-04 weekday=Saturday
INFO skipped digest schedule="0 9 * * *" reason=holiday date=2026-12-25 country=DE holiday="1. Weihnachtstag"

reason is weekend or holiday, the same shape as the ignored webhook event reason contract.

What counts as stuck

A PR is stuck when its last activity predates the start of the current day (in the configured zone). Activity is anything Notifycat sees on the PR: the open announcement, a review — approve, comment, request changes — or a PR/line comment. Two deliberate exclusions:

• Suppressed bot reviews don't count. With reviews.ignore_ai_reviews: true, an AI-only review pass leaves the PR stuck — it still needs a human.
• Merged, closed, and drafted PRs drop out. Merge/close marks the row; converting to draft removes it entirely.

A monorepo PR that fanned out to several channels appears in each of those channels' digests until it's done.

Per-repository overrides

A repository tier (or org "*" tier) can set its own digest.enabled and digest.schedule — timezone and country are global only, since the server runs a single clock and serves a single team calendar. Note that two repositories posting to the same channel on different schedules produce two digests a day in that channel; each tier's schedule runs independently.

First run after an upgrade

If you enabled the digest on a deployment that predates it, the first run can list PRs that merged long ago — old rows have no open/closed marker, so the digest assumes open. Fix it once with the reconciler, which asks the git host for each PR's real state and marks the closed ones:

docker compose run --rm notifycat /usr/local/bin/notifycat-reconcile -dry-run   # preview
docker compose run --rm notifycat /usr/local/bin/notifycat-reconcile           # apply

It's idempotent, needs a read token (GITHUB_TOKEN / BITBUCKET_TOKEN) plus the same database the server uses, and leaves any PR it can't read untouched rather than wrongly hiding it. It prints a summary like reconcile (applied): checked=37 closed=34 still_open=3 errors=0; a non-zero error count exits non-zero — usually a token-scope issue, fix and re-run. After this one run the close handler keeps rows marked by itself. See also Upgrading → 0.16.0.