Skip to content

v1.21.3 — Reliability: full-codebase audit (wave 1) + hardening

Choose a tag to compare

View all tags
@jgouviergmail jgouviergmail released this 03 Jul 13:58
v1.21.3
ae86152

Reliability release: wave 1 of a full-codebase audit — 14 confirmed, surgical, single-function fixes (each with a regression test written red-first) plus a hardening pass. No new user feature, no DB schema change, no migration.

Fixed — Connected services

  • Apple Calendar events now honor the requested timezone (a "10:00 Europe/Paris" event no longer lands at 10:00 UTC; covers create + update).
  • Microsoft Calendar search tolerates apostrophes (l'anniversaire no longer breaks the OData filter → 400).
  • Google Tasks "list task lists" works again (a missing () raised AttributeError on every call).
  • Route maps render correctly (a polyline-encoder operator-precedence bug corrupted multi-segment geometry).

Fixed — Assistant behavior

  • Ambiguous Gmail labels now ask which one instead of a generic error (KeyError).
  • Stopping a ReAct exploration (closing the conversation) actually cancels it (CancelledError was swallowed as a tool error).
  • Connector "not activated" messages are localized in all 6 languages (no more inline French in LLM-facing scaffolding).

Security & configuration

  • /users/search/by-email is superuser-only (account-enumeration surface closed).
  • Planner budget/steps .env overrides are honored (one path always served a hardcoded 50$/50).
  • Two previously-hardcoded values become .env-configurable: metrics sync interval, MCP ReAct step timeout.

Reliability & dead code

  • Registry LRU eviction no longer crashes on mixed timestamp forms in long conversations.
  • QueryIntelligence field silently dropped on every HITL resume restored (guarded by a full round-trip test).
  • LLM capability check unified (tool_calling → tools); web client no longer leaks an abort timer per request.
  • Dead never-wired per-node message-windowing subsystem removed (ADR-094).

Test infrastructure

  • The long-silently-rotted tests/agents suite (83 stale failures, never gated) is rehabilitated to green (1069 tests) and wired into CI.
  • A value-blind-test scanner (scripts/analyze_value_blind_tests.py) closes the class of gap the polyline bug exposed — zero fully value-blind files remain.

Quality: Ruff / Black / MyPy strict clean; backend 8572 unit + 1069 agents green, frontend 120 green; i18n parity OK; Docker healthy. Full details in CHANGELOG.md.

Assets 2
Loading