Adds lattice-aware sibling dedup between scorer and floor filter in both `search` and `context` commands, with D2 routing rules, DB column updates, and event emission.
Hardened GWells controller pause, resume, stop, pinning, and live override behavior to avoid duplicate loops and stale runtime caches.
Replaced dry-run-only planning with an explicit --apply flag for actual deletion. Default behaviour (no --apply) is unchanged — plan preview only.
Added deterministic top-level ring spacing so large explicit-spine layouts avoid origin crowding while preserving small-graph behavior.
PM-aware lockfile snapshots and diff capture for all four package managers, plus pnpm transitive dependency analysis. yarn and bun skip transitive (incompatible list output formats, documented).
Added deterministic fallback seeding for no-spine, orphan, and unreached graph nodes so non-filesystem layouts avoid origin clumping.
Extends `cerebra memory status` to show truth tower contents, and ships the Phase 5 end-to-end integration test suite covering all 6 behavioral classes.
Refreshed IMPLEMENTATION_STATUS.md, INSTALL.md, and INTEGRATION_BOUNDARIES.md to reflect the completed state of all 13 Tier 2/3 plans — no code changes.
Multi-layer lifecycle script analysis wired non-fatally into the sandbox flow — JS static analysis, Python AST visitor, dependency confusion detection, transitive npm tree scan, and publish anomaly signals.
Routed GWells well assignment through structural analysis fallback while preserving existing radial-backbone and parallel-spines behavior.
Wires T1 auto-promotion into `cerebra context` so every retrieval run feeds selected memory into the Truth Tower, with --no-promote and abstention guard rails.
Added a behavior-neutral structural resolver so GWells can analyze non-filesystem graph shapes before routing layout behavior through topology-aware roles.
TruthTower class with T1 auto-promotion from retrieval results, T2 manual promotion citing a T1 anchor, staleness cascade on T1 eviction, chronological render, and full inspector event suite.
When ≥2 categories score ≥0.65, classify_record_lattice() writes sibling memory_records for each additional category, all linked by a shared lattice_lineage_id; existing classify_record() untouched.
WorkingMemory class with capacity eviction, tower-cited penalty, and pin guards; `cerebra memory` CLI (status / promote / evict).
Migration009 (sessions/wm/tower tables), slot constants, PHASE_5_EVENT_TYPES, new_session(), vault lockfile, and `cerebra session` CLI commands.
Semver bump to 0.19.0; v112 arc reconciled in NOW.md + ROADMAP with full commit table; 10 new KNOWN_SHARP_EDGES lessons appended.
MCP server (5 tools, PreToolUse hook), prompt injection scanner (8 categories + base64 rescan), Linux namespace sandbox (overlayfs + strace), package intel adapter (typosquatting, known-bad registry), and real-time fs watch daemon with event routing.
Diagnoses and closes the v112.4 opacity regression report (binding intact, no code fix needed) and resolves the pre-existing i18n.spec.ts:187 stale assertion.
Policy management UI delivered in three ADR-008 phases: JSON schema contracts and TypeScript domain types; strict type coverage with browser preview mocks and pagination; then the full management card suite with rule display and simulation controls.
Six Tier 2 safety subsystems shipped: HMAC audit chain, integrity checks, incident response with lockdown and playbooks, git pre-commit hooks, secret scanning, and policy management with simulation. CLI wired; ADRs 001–008 and plans 01–13 added.
Audited and documented the full data visibility story for the native dashboard: CLI-first setup path, data path catalog, report evidence surfaces, audit events display, and a consolidated v0.2 smoke checklist. Added a local alpha release checklist.
Consolidated the dark-mode CSS architecture to eliminate per-card overrides and fixed contrast failures. Added the full decision check flow: Rust adapter, TypeScript card shell, invoke wiring, and result display.
Replaced cleanup dry-run wrappers with validated target selectors, extracted audit event ID validation into a dedicated Rust helper, added full unit test coverage for adapter validators, and wired a Tauri desktop CI job.
Wires the agent chat frontend — AgentChatTile with ephemeral conversation state, ChatInput with Enter-to-send, Agents settings category (endpoint/model/API key/test connection), and schema v95 migration.
Adds the Rust inference module — InferenceBackend trait, RemoteClient (OpenAI-compat HTTP), and chat/test_inference_connection Tauri commands. No frontend wiring yet; that's Commit 2.