[superlog-app]

Summary

Agent PR delivery stalled completely after the worker restart at 22:05 UTC (incident knotty-manatee). Three independent metrics confirmed zero deliveries: superlog.prs.created, superlog.agent_pr_delivery.guard, and pr_observability.review_passes all flatlined, while agent investigations kept running at above-baseline rate. The worker emitted no api.github.com calls across 115 minutes of runtime.

Root Cause

In loadAgentRunContext, the two-step issue-loading query is unbounded:

// step 1 — fetches ALL issue IDs for the incident, no limit
const issueIds = await db
  .select({ issueId: schema.incidentIssues.issueId })
  .from(schema.incidentIssues)
  .where(eq(schema.incidentIssues.incidentId, incident.id));

// step 2 — passes every ID as a bind parameter
const issueRows = issueIds.length > 0
  ? await db.query.issues.findMany({
      where: inArray(schema.issues.id, issueIds.map(r => r.issueId)),
      ...
    })
  : [];

When an incident accumulates many linked issues, the inArray call grows proportionally. Production logs showed 13,854+ bind parameters (the log line was itself truncated at 30 kB, so the actual count may have been higher). Postgres rejects queries that exceed its 65,535 bind parameter limit with a hard "Failed query" error. The failure is caught in queue.ts and logged as "agent-run advance failed; the next sweep re-enqueues it" — but the next sweep calls the same unbounded loadAgentRunContext again, creating a permanent retry loop that prevents any advance from succeeding and blocks all PR delivery.

Remediation

Add a JOIN on issues to the first query so rows can be sorted by lastSeen DESC before applying a LIMIT 200. The second inArray call is now bounded at ≤ 200 parameters, well under the Postgres limit. The 200 most-recently-seen issues are always included, matching existing downstream caps (MAX_TRACE_ENRICHED_ISSUES = 50 in prompt-context.ts).

const ISSUE_CONTEXT_LIMIT = 200;
const issueIds = await db
  .select({ issueId: schema.incidentIssues.issueId })
  .from(schema.incidentIssues)
  .innerJoin(schema.issues, eq(schema.issues.id, schema.incidentIssues.issueId))
  .where(eq(schema.incidentIssues.incidentId, incident.id))
  .orderBy(desc(schema.issues.lastSeen))
  .limit(ISSUE_CONTEXT_LIMIT);

Typecheck: ✅ tsc --noEmit passes
Tests: ✅ test:agent-run-queue, test:agent-run-start pass

Related incident: c7813ad1-000c-4ee0-b76d-d1b7b3707d03

---

Was this PR helpful? Leave feedback — goes straight to the Superlog team.