LIA

Highlights

Three independent hardening fronts shipped in one release.

🛡️ ADR-084 — Indexable vs Semantic Criteria (universal planning principle)

Repeated diagnostics of "mes deux prochains rdv médicaux" surfaced a class of failures invisible to the existing validator: the planner was passing semantic qualifiers (medical, urgent, important) as the query of literal-text-search tools (Calendar / Gmail / Notion-style stores) which then returned 0 hits or false positives.

4-layer defense:

• Universal prompt section in smart_planner_prompt.txt (370 tokens, prompt-cached after first call)
• Structured semantic_filter_terms hint emitted by the query analyzer
• Universal PlanValidator._check_semantic_leak gated by PLANNER_SEMANTIC_LEAK_MODE (off / observe / autocorrect, default observe)
• ToolManifest.text_search_mode declarative opt-out for native semantic-search backends

Ships in observe mode this release: zero plan mutation, only telemetry. Phase 2 promotion to autocorrect gated on operational metrics. Tests: 19 cases including French apostrophe regression and required-param protection.

🔧 Gemini 3.x response.content shape fix

Switching query_analyzer / planner to gemini-3.1-pro-preview exposed smart_planner_panic_failed with a JSON parse error: langchain_google_genai wraps the response as list[dict] for Gemini 3.x specifically while every other provider keeps str.

33 sites across 26 files migrated to the LangChain Core 1.2+ BaseMessage.text property which handles both shapes:

• Planner ×2, response node, compaction ×3, memory extraction + reference resolution ×4
• HITL classification + question generation streams ×6
• Voice TTS, broadcast notifications
• Journals extraction + consolidation ×3, interests ×3, psyche ×2, heartbeat
• MCP service, llm_cache key serialization

End-to-end Gemini 3.1 Pro Preview pipeline now runs cleanly. Test mocks for compaction, draft_modifier, item_filter and psyche updated to mirror .content on .text.

⏱️ Vagues 1-5 timeout centralization

Audit inventoried ~80 backend timeouts and surfaced:

• 28 hardcoded values used in production paths but never operator-tunable
• 2 inline magic numbers in parallel_executor
• 1 orphan field (task_orchestrator_execution_timeout_seconds) declared but never read
• 2 dead constants, 1 duplicated Field (currency API)
• 1 cascade inversion bug (Brave enrichment 3 s wrapper firing before its own 5 s HTTP call → raised to 8 s)

All 28 hardcoded values exposed as Pydantic Fields with operator-validated ranges and matching .env.example / .env.prod.example entries. Two new config modules: LocksSettings (OAuth refresh lock) and SchedulerSettings (scheduled-actions executor). The orphan global timeout was wired in parallel_executor.execute_plan_parallel as a soft cap on multi-wave plans + new Prometheus counter lia_parallel_execution_global_timeout_total{plan_outcome}.

Single source of doc truth: docs/technical/TIMEOUT_REGISTRY.md — 12 sections + Annex A (LLM out-of-scope), Annex B (frontend timing.ts mirror), Annex C (gaps tracked for future Vague 6).

Pydantic ranges fully aligned with scripts/validate_config.py end-to-end; 4 legacy gt=0-only Fields tightened (http_timeout_oauth/token/external_api, apple_connection_timeout).

What changed for users

See the in-app FAQ changelog (v1.20.6 entry) — 4 user/admin items in 6 languages:

1. Web search (Brave) response enrichment more reliable
2. Gemini 3.x LLM fully compatible end-to-end
3. 28 timeouts now operator-tunable via .env + new TIMEOUT_REGISTRY.md
4. Preventive defense layer on the planner (in observe mode)

Documentation

• ADR-084 created
• TIMEOUT_REGISTRY.md created
• PLANNER.md, SMART_SERVICES.md, OBSERVABILITY_AGENTS.md updated
• README banner bumped (v1.20.5 demoted)
• GETTING_STARTED.md compatibility line updated
• Blog article execution-plans enriched with an "Indexable vs Semantic criteria distinction" section in 6 languages

Commit

91eb2ad — see CHANGELOG.md for the full diff narrative.