cozystack/.github/workflows/issue-triage.yaml

name: Issue triage

# Every open issue should carry a triage/* label marking its lifecycle stage.
# The issue templates apply `triage/needs-triage` from their front matter, but
# that only covers issues created through the web form: issues created via
# the API or `gh issue create` bypass templates and arrive unlabeled. The
# `issues` events below label those on arrival; the daily sweep is the
# backstop for removed labels and for the pre-existing backlog.
on:
  issues:
    types: [opened, reopened]
  schedule:
    - cron: '53 5 * * *'  # daily at 05:53 UTC, after stale (04:37)
  workflow_dispatch:
    inputs:
      dry_run:
        description: 'Log what would be labeled without applying'
        type: boolean
        # On by default, so the one entry point a human can reach writes
        # nothing until somebody unchecks it. The scheduled sweep reads no
        # inputs and is unaffected: it writes.
        default: true

# Top-level token defaults to read-only. A job-level block replaces this one
# rather than adding to it, so the job below holds `issues: write` and
# nothing else.
permissions:
  contents: read

# One queue for every run of this workflow, on purpose. GitHub keeps a
# single pending run per group, so issues filed while a sweep is running
# lose all but the last of their own runs and wait for the next sweep,
# which is what the sweep is there for. Splitting the group per issue would
# save that wait at the price of a real bug: a sweep works from a listing
# taken when it started, so it and a concurrent per-issue run can read the
# same issue differently and leave it holding two contradictory triage
# labels, which nothing later removes.
#
# The queue closes that door against a second run of this workflow, not
# against a maintainer labeling an issue by hand while a sweep is running:
# the sweep classifies from the listing it took when it started, so a label
# applied during its paced writes is invisible to it and the issue ends up
# carrying a second triage label that the already-triaged skip then keeps
# every later sweep from revisiting. Losing that race costs the spurious
# `triage/needs-triage`, which a maintainer deletes and no later sweep
# re-adds, because the label it raced against is still on the issue. That
# was judged cheaper than re-reading every issue immediately before every
# write: that would double the request budget the write cap below is derived
# against. Deleting is the repair for this one case and not a general undo:
# an issue left carrying no triage label at all is relabeled by the next
# sweep, which is what AGENTS.md's labeling bullet warns about.
concurrency:
  group: issue-triage
  cancel-in-progress: false

jobs:
  triage:
    runs-on: ubuntu-latest
    # A run that reaches the write cap paces roughly seven minutes of work, so
    # anything near this ceiling is a call that hung rather than a long sweep.
    # Without it the default is six hours, and the queue above means one hung
    # run holds every issue filed behind it for all six.
    timeout-minutes: 20
    permissions:
      issues: write
    steps:
      - name: Apply triage labels
        uses: actions/github-script@f28e40c7f34bde8b3046d885e986cb6290c5673b # v7
        with:
          script: |
            // Labels applied with the workflow's GITHUB_TOKEN do not trigger
            // other workflows (GitHub's recursion prevention), so this job
            // cannot re-enter itself or wake any `labeled`-subscribed workflow.
            //
            // Adding a label bumps the issue's `updated_at`, which stale.yaml
            // reads as fresh activity: the stale bot un-marks `lifecycle/stale`
            // and restarts its 60-day clock on the next run. For issues labeled
            // on arrival that is moot; for the sweep over the historical
            // backlog it is a deliberate one-time reset. Triage first, then
            // let staleness take its course.
            //
            // `triage/accepted` is also on stale.yaml's exempt-issue-labels,
            // so applying it is more than a clock reset: the issue stops
            // being auto-closable at all. For every label in
            // ACCEPTED_SIGNAL_LABELS that changes nothing, because each one
            // is on the same exempt list already. An assignee is the single
            // signal that is not a label and not already exempt, so an issue
            // whose only accepted-signal is an assignee gains something it
            // did not have. That is new, and it is the
            // intended reading rather than a side effect: somebody took
            // ownership, which makes the issue a known long-tail task and
            // not an abandoned one — the same argument stale.yaml's own
            // comment makes for exempting `triage/accepted`. Part of the
            // backlog this reaches is marked stale today, some of it inside
            // the 14-day close window, and the sweep spares it.
            const NEEDS_TRIAGE = 'triage/needs-triage';
            const ACCEPTED = 'triage/accepted';

            // Label-shaped accepted-signals, admitted on two conditions.
            // Necessary: the label is on stale.yaml's exempt-issue-labels,
            // because `triage/accepted` is itself exempt and awarding it for
            // a label that is not would hand the issue a permanent reprieve
            // as a side effect — which is why `priority/backlog`, the tier
            // stale.yaml leaves reapable, is not a signal. Not sufficient:
            // a `lifecycle/*` label states what the stale bot may do and
            // nothing about whether anyone reviewed the issue, so exempt
            // `lifecycle/frozen` is still not a signal here. Freezing an
            // untriaged issue to keep the bot off it is exactly the case
            // where that inference would be false.
            const ACCEPTED_SIGNAL_LABELS = [
              'priority/critical-urgent',
              'priority/important-soon',
              'priority/important-longterm',
              'epic',
            ];

            // The sweep is the only place this workflow writes in bulk, and
            // GitHub's secondary rate limits allow 80 content-generating
            // requests a minute and 500 an hour. This client has no
            // throttling plugin and no retries, and a throttled write
            // answers 403, which is exempt from retry anyway: an unpaced
            // sweep does not slow down when it is throttled, it fails the
            // job and leaves the backlog half labeled. Pace the writes
            // under the per-minute ceiling, and stop short of the hourly
            // one instead of walking into it. Whatever a run leaves behind
            // stays unlabeled, so the next one picks it up.
            //
            // Those are the secondary limits. GITHUB_TOKEN also carries a
            // primary one of 1000 requests an hour for the whole repository,
            // shared with every other workflow, and a run that reaches the
            // cap spends a large part of it. That is affordable at 05:53 and
            // only while the backlog drains, but it is the ceiling to lower
            // the cap against if this ever moves to a busier slot.
            const WRITE_INTERVAL_MS = 1000;
            const MAX_WRITES_PER_RUN = 400;
            const sleep = (ms) =>
              new Promise((resolve) => setTimeout(resolve, ms));

            // Returns the label to add, or null when the item needs nothing.
            // The signals stand for a maintainer decision already recorded on
            // the issue, a weaker claim than the one `triage/accepted` makes
            // in .github/labels.yml: `priority/important-longterm` says
            // outright the work may not be staffed. Of the two labels this
            // workflow writes it is still the closer one — `needs-triage` on
            // an issue somebody prioritised or took asks for a triage that
            // already happened — and a maintainer who disagrees replaces it.
            const classify = (issue) => {
              if (issue.pull_request) return null; // listForRepo returns PRs too
              const labels = (issue.labels || []).map((l) =>
                typeof l === 'string' ? l : l.name
              );
              if (labels.some((n) => n.startsWith('triage/'))) return null;
              const looked = // somebody has already looked at this one
                (issue.assignees || []).length > 0 ||
                labels.some((n) => ACCEPTED_SIGNAL_LABELS.includes(n));
              return looked ? ACCEPTED : NEEDS_TRIAGE;
            };

            const apply = async (issue, label, dryRun) => {
              if (dryRun) {
                core.info(`[dry-run] #${issue.number}: would add ${label}`);
                return;
              }
              await github.rest.issues.addLabels({
                owner: context.repo.owner,
                repo: context.repo.repo,
                issue_number: issue.number,
                labels: [label],
              });
              core.info(`#${issue.number}: added ${label}`);
            };

            if (context.eventName === 'issues') {
              const issue = context.payload.issue;
              const label = classify(issue);
              if (label) {
                await apply(issue, label, false);
              } else {
                core.info(`#${issue.number}: already triaged, nothing to do`);
              }
              return;
            }

            // schedule / workflow_dispatch: sweep every open issue.
            // workflow_dispatch inputs arrive as strings in the webhook
            // payload even when declared `type: boolean`.
            //
            // Keyed on the event rather than on the input alone, so both
            // unknown cases fail safe: `schedule` is the one event that
            // writes without being asked, so any trigger added here later
            // starts out dry rather than writing on its first run, and a
            // dispatch sending anything but the literal `false` gets a dry
            // run. The declared default above is what a human sees in the
            // form; this does not depend on it reaching the payload.
            const dryRun =
              context.eventName !== 'schedule' &&
              String(context.payload.inputs?.dry_run ?? 'true') !== 'false';

            const issues = await github.paginate(github.rest.issues.listForRepo, {
              owner: context.repo.owner,
              repo: context.repo.repo,
              state: 'open',
              per_page: 100,
            });

            const acceptedIssues = [];
            const needsTriageIssues = [];
            let skipped = 0;
            let deferred = 0;
            let failed = 0;
            let refused = 0;
            let writes = 0;
            // Listing index a refusal stopped the run on. The remainder it
            // yields is counted in listing entries, not in issues left
            // needing a label — most of any remainder is pull requests and
            // already-triaged issues the loop would have skipped. It is
            // reported so that a run which stopped early cannot read as one
            // that reached the end of the listing; `deferred` is the write
            // cap's own remainder and does not cover this.
            let stoppedAt = -1;
            for (const [index, issue] of issues.entries()) {
              const label = classify(issue);
              if (!label) {
                skipped += 1;
                continue;
              }
              // The cap counts in dry-run too, so a dry run predicts what a
              // real one would do rather than a longer list nobody gets.
              if (writes >= MAX_WRITES_PER_RUN) {
                deferred += 1;
                continue;
              }
              if (writes > 0 && !dryRun) {
                await sleep(WRITE_INTERVAL_MS);
              }
              writes += 1;
              try {
                await apply(issue, label, dryRun);
              } catch (err) {
                // 403 is what a secondary rate limit and a token without
                // `issues: write` both answer; 429 is the other throttle
                // shape. Those two are the only statuses that stop the run.
                // Every other status carries on, and that is the whole of
                // the else: the 404 or 410 of an issue transferred or
                // deleted mid-sweep, but a 422 or a 5xx just the same,
                // because a failure about one issue says nothing about the
                // next one and the next sweep retries it. A refusal is the
                // opposite kind: it applies to every write left in the run,
                // and repeating a rejected request up to the cap is what
                // gets an integration banned, so stop at the first one. The
                // remainder stays unlabeled and the next run takes it,
                // exactly like the write cap's own deferral.
                if (err.status === 403 || err.status === 429) {
                  refused += 1;
                  stoppedAt = index;
                  core.warning(
                    `#${issue.number}: ${label} refused ` +
                      `(${err.status}); stopping the sweep`
                  );
                  break;
                }
                // One issue that cannot be labeled, because it was moved or
                // deleted mid-sweep, must not cost the run the rest of the
                // backlog and the summary with it. It stays unlabeled, so
                // the next run picks it up like any other. Counted below the
                // refusal branch rather than above it, so that one refusal
                // reports as one problem instead of also inflating the
                // failed count it is already reported by.
                failed += 1;
                core.warning(`#${issue.number}: ${label} failed: ${err.message}`);
                continue;
              }
              if (label === ACCEPTED) {
                acceptedIssues.push(issue.number);
              } else {
                needsTriageIssues.push(issue.number);
              }
            }

            // A sweep over the backlog logs one line per issue, which is a
            // page nobody reads to find out what a run did. Every row is
            // printed whether or not it is zero: a table that hides its empty
            // rows makes "no writes failed" and "the count was never taken"
            // look alike.
            const unexamined =
              stoppedAt < 0 ? 0 : issues.length - 1 - stoppedAt;
            const numbers = (ns) =>
              ns.length ? ns.map((n) => `#${n}`).join(', ') : 'none';
            const report = [
              `## Issue triage${dryRun ? ' — dry run, nothing written' : ''}`,
              '',
              '| outcome | issues |',
              '| --- | --- |',
              `| \`${ACCEPTED}\` | ${acceptedIssues.length} |`,
              `| \`${NEEDS_TRIAGE}\` | ${needsTriageIssues.length} |`,
              `| already triaged, or a pull request | ${skipped} |`,
              `| deferred, this run hit its write cap of ${MAX_WRITES_PER_RUN} | ${deferred} |`,
              `| write failed, left for the next run | ${failed} |`,
              `| write refused, which stopped the sweep | ${refused} |`,
              `| listing entries never reached, the sweep stopped at a refusal | ${unexamined} |`,
              '',
              `\`${ACCEPTED}\`: ${numbers(acceptedIssues)}`,
              '',
              `\`${NEEDS_TRIAGE}\`: ${numbers(needsTriageIssues)}`,
              '',
            ].join('\n');
            core.info(report);
            // `write()` is async: unawaited, the step can end green with an
            // empty summary panel. `addRaw` appends no newline of its own.
            await core.summary.addRaw(report).write();
            // An issue that vanished mid-sweep is benign however many times
            // it happens, and the next run simply does not see it. A refusal
            // is not benign even once: it means the token lost `issues:
            // write`, or the pacing above stopped being enough and the sweep
            // is into the secondary rate limit. Discriminating on the status
            // rather than on a count keeps the alarm live once the backlog is
            // drained and a run writes to a handful of issues at most, which
            // is the regime this workflow spends almost all of its life in.
            if (refused > 0) {
              core.setFailed(
                `A label write was refused after ${writes} attempted, so ` +
                  `the sweep stopped early; check the job's issues: write ` +
                  `permission and the log for rate-limit responses`
              );
            }