SAMURAI 4.6

Releases 4.6.0 through 4.6.1, July 6, 2026.

4.6.1 — July 6, 2026

Bug Fixes

• Deploy: auto-generate + persist DEVICE_ENCRYPTION_KEY; drop LOG_LEVEL from env

4.6.0 — July 6, 2026

Features

• Ai Eval: memory-recall harness — recall@K / MRR retrieval-quality gate
• AI: bi-temporal validity + supersession-on-contradiction
• AI: configurable council cadence + async ‘Correlate now’ button
• AI: entity anchoring + council/reflect review hardening
• AI: full self-management - council follows the autonomy switch
• AI: no em dashes in AI text, URL-synced filter bar, animated collapse, export, polish
• AI: record provider/model that authored a memory
• AI: review council auto-approves/declines proposals to cut the review queue
• AI: self-audit / re-validation pass for stored claims
• AI: self-managing confidence - usage bump, staleness decay, lowering-only weaken
• AI: async eval-sweep runs so the full golden set can be scored
• AI: conversational assistant v0 — ai.chat permission, bounded chat loop, floating widget
• AI: correlation nudge in agent prompts - measured fix, not more rounds
• AI: durable AI-analysis pool + per-provider rate limits
• AI: evaluation harness for verdict accuracy + model comparison
• AI: feed-aware empty tool results + empty-result metric (#938 code-side)
• AI: gated get_endpoint_identity tool for auth/endpoint changes
• AI: get_auth_sessions interface/mac filter args - drill instead of paginate
• AI: get_syslog_events tool - RADIUS/AAA window queries over syslog ingest
• AI: per-provider analysis hourly limit in Settings UI
• AI: Phase-1 analysis tools — fleet/temporal/health context
• AI: token accounting + memory-from-chat
• AI: verdict-quality instrument from confirm/override feedback
• Changes: AI-analysis state badge in the card header
• Changes: analyze in footer, collapsible AI panel, unified device/global footer
• Changes: device Changes tabs reach design parity with the global page
• Changes: show AI analyzed-at time in the analysis panel
• Integrations: syslog notification channel + source-aware log forwarding
• Memories: server-side pagination + search + sort
• Memories: show created / updated / last-used dates
• Observability: Prometheus metrics for the agentic AI loop
• Settings: syslog forwarding source selector + Syslog (SIEM) integration card UI
• UI: promote Integrations to a standalone /integrations sidebar page
• UI: standardized dialog sizes + Memories dialog/table polish

Bug Fixes

• AI: repair edit dialog (400 on save) + round-trip category/tags/device
• AI: type-gated autonomy, only neutral facts self-activate
• Ai Tools: credential-free device read + recursive secret strip
• Ai Tools: drop normalize-backfilled credential keys from public device read
• AI: agent-loop polish batch
• AI: budget-aware finalization + persist partial trail on failure
• AI: cap executed tool calls per native turn at 5
• AI: identity claims must name their evidence source
• AI: native-loop forced-final double-user 400 + add get_device_inventory
• AI: ParseVerdict scans balanced JSON candidates, not the outermost span
• AI: tool_result must precede text in Bedrock user turns (#956 root cause)
• AI: wire-level transcript sanitizer for recurring Bedrock 400
• Audit: register ai_memory_council + ai_eval_run audit categories
• Database: make read concern an explicit, documented decision
• Endpoints: drop the older generation during the materialization window
• Cisco FMC: make free-text access-policy search subnet- and port-range-aware
• Cisco FMC: make port filter/search “any”-aware
• Frontend: pin eslint-plugin-react-hooks to 7.0.1 (unblock CI lint)
• Integrations: escape free-form event text for telegram & slack
• Memories: keep tag/category badges single-line in the Tags column
• Memories: show Confidence column in 0.000 format
• Sync: fortigate hash-strip now mirrors resolved/resolved_value + #447 superset conformance test
• Sync: key the running-config data_hash off the encryption key