You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Admin escalation tools (list_escalations/resolve_escalation) unavailable in some Addie sessions #7411
Observed behavior: In a web chat session, an org owner/admin user asked Addie to look up and resolve specific escalations (by ID) using list_escalations and resolve_escalation. Both tool calls failed with Unknown tool "list_escalations" (and presumably the same for resolve_escalation), even though the user has an admin/owner role on their organization.
Expected behavior: Admin-tier users should have consistent access to list_escalations / resolve_escalation across sessions/surfaces where they are authenticated as an org owner or AAO admin.
Actual behavior: The tools were not present in the request-scoped tool catalog for this session, despite the user apparently holding an admin-equivalent role. This reads as an intermittent or scope-configuration issue in how admin tool sets are attached to a request.
Impact: Admins are unable to close out or triage escalations from certain sessions, requiring them to fall back to another surface with no clear indication of why.
Ask: Investigate why list_escalations/resolve_escalation were not attached to this session's tool catalog for an apparently admin-eligible user, and confirm the intended scoping rules (e.g., AAO staff only vs. org owner) are being applied consistently.
prompt-engineer: Confirmed reliability blocker. Per-replica in-memory webAdminStatusCache + uncached error fallback (catch → return false in isWebUserAAOAdmin) causes ADMIN_TOOLS to disappear from the catalog for genuine AAO admins on any transient DB error or cold-cache Fly.io replica. This is the structural source of the intermittent symptom.
internal-tools-strategist: Current AAO-staff-only gate is correct — widening it to org owners risks cross-tenant escalation data leakage. Two real but separable gaps: (a) no org-scoped read-only view for org owners who want to check their own escalations, and (b) "Unknown tool" gives no actionable signal (vs. a permission-denied explanation).
My take: Two problems with different owners. The reliability bug (per-replica Map for auth state on a multi-replica Fly.io deployment) is a clear engineering defect regardless of who should have access. The org-owner scope question is a product call that shouldn't block the reliability fix.
Resolution (2026-09-21):@bokelley directed: bounded reliability path is #7535 (shorten stale negative-cache denial, retain fail-closed authorization). Org-owner read-only visibility and permission-denied UX are separate product decisions; staff escalation access should not be widened as part of the cache fix.
Triage execution update: routed to P0 Bugs. The bounded reliability path is #7535 (shorten stale negative-cache denial, retain fail-closed authorization). Org-owner read-only visibility and permission-denied UX remain separate product decisions; do not widen staff escalation access as part of the cache fix.
Resolved by #7450 (f928b36ff4f721ae64c73cc5f80de65b3d69a071) plus the web-boundary follow-up #7452 (c9a8c4f91b056aea6871c76c80e9afe307971ffa). Current main performs uncached, exact-credential admin checks on both Slack and web paths, distinguishes confirmed denial from retryable authority lookup failure, and includes linked-identity-direction/revocation coverage. The older TTL-only #7535 is now closed as superseded because its cache targets no longer exist.
Activity
bokelley commentedon Sep 10, 2026
Triage
Classification: Bug
Bucket(s): addie, admin-tool
Status: deferred — blocked on #7535
What the experts said:
webAdminStatusCache+ uncached error fallback (catch → return falseinisWebUserAAOAdmin) causesADMIN_TOOLSto disappear from the catalog for genuine AAO admins on any transient DB error or cold-cache Fly.io replica. This is the structural source of the intermittent symptom.My take: Two problems with different owners. The reliability bug (per-replica Map for auth state on a multi-replica Fly.io deployment) is a clear engineering defect regardless of who should have access. The org-owner scope question is a product call that shouldn't block the reliability fix.
Resolution (2026-09-21): @bokelley directed: bounded reliability path is #7535 (shorten stale negative-cache denial, retain fail-closed authorization). Org-owner read-only visibility and permission-denied UX are separate product decisions; staff escalation access should not be widened as part of the cache fix.
Triaged by Claude Code. Session: https://claude.ai/code/session_01VUpozg1azN6ZPfQC7Lmx7i
Generated by Claude Code
bokelley commentedon Sep 21, 2026
Triage execution update: routed to P0 Bugs. The bounded reliability path is #7535 (shorten stale negative-cache denial, retain fail-closed authorization). Org-owner read-only visibility and permission-denied UX remain separate product decisions; do not widen staff escalation access as part of the cache fix.
bokelley commentedon Sep 21, 2026
Blocked-on: #7535 — resurfaces on merge
Generated by Claude Code
bokelley commentedon Sep 23, 2026
Resolved by #7450 (
f928b36ff4f721ae64c73cc5f80de65b3d69a071) plus the web-boundary follow-up #7452 (c9a8c4f91b056aea6871c76c80e9afe307971ffa). Currentmainperforms uncached, exact-credential admin checks on both Slack and web paths, distinguishes confirmed denial from retryable authority lookup failure, and includes linked-identity-direction/revocation coverage. The older TTL-only #7535 is now closed as superseded because its cache targets no longer exist.