@theihtisham/agent-shadow-brain 6.2.0 → 7.0.0

package/CHANGELOG.md

@@ -1,372 +1,411 @@
-# Changelog
-
-## [6.0.3] - 2026-06-03 - Live Brain Release
-
-This release makes Shadow Brain more realistic and more useful for complex agent work. It focuses on source-backed long memory, live memory health, recall evals, agent briefings, and release-ready docs.
-
-### Added
-
-- `src/brain/live-brain.ts`: Live Brain report engine for memory health, recall evals, hygiene actions, top memories, and agent briefings.
-- `src/brain/memory-seeder.ts`: verified memory seeding workflow with draft, user approval, agent approval, dual approval, reject, and seed-to-brain states.
-- `shadow-brain live-brain status`: shows memory quality, source coverage, recall health, and recommended actions.
-- `shadow-brain live-brain brief`: writes a markdown briefing for fresh sessions, handoffs, and context recovery.
-- `shadow-brain live-brain eval`: checks whether useful memory can be recalled for agent workflows.
-- `shadow-brain live-brain hygiene`: lists stale, duplicate, low-quality, and source-light memory issues.
-- `shadow-brain live-brain train-pack`: exports source-backed JSONL examples for evals, distillation, or local fine-tuning pipelines.
-- `/api/v6/live-brain` and the v6 dashboard Live Brain tab: dashboard/API route and UI for the same report and markdown briefing.
-- `/api/v6/memory-intelligence`: live memory data for source coverage, freshness, categories, agents, tags, recent writes, and approval queue counts.
-- `/api/v6/seed-*`: dashboard/API workflow for creating research drafts, approving/rejecting them, and committing approved drafts to global memory.
-- Dashboard Memory Seeder tab: select research agent, approval agent, approval mode, project path, sources, category, tags, and importance before writing durable memory.
-- Dashboard launch controls: `shadow-brain dash --host 127.0.0.1 --no-analysis` for fast UI/API startup on large repos.
-- MCP tools: `live_brain_status`, `live_brain_brief`, and `live_brain_train_pack`.
-- Focused tests for Live Brain scoring, briefing, recall eval, JSONL export, and memory seeding approval gates.
-
-### Changed
-
-- README rewritten to be realistic about what the project does and does not do.
-- npm package version updated to `6.0.3`.
-- npm package description replaced with a practical local-first memory description.
-- Runtime-visible versions updated in CLI, auto-update, portability export, orchestrator status, and MCP server metadata.
-- Resume-pack source discovery now includes common repo docs such as `README.md`, `CHANGELOG.md`, `package.json`, and roadmap docs in addition to project-specific agent/progress files.
-- CLI config merging now honors explicit `watchMode: false` from one-shot commands.
-- Token Economy dashboard now separates the current runtime agent from historical model spend records, avoiding misleading labels such as an old Claude model appearing as the active Codex runtime.
-
-### Notes
-
-- "Training" in this release means data export for user-run evals, distillation, or local fine-tuning. Shadow Brain does not modify remote model weights.
-- No new runtime dependencies were added.
-- Verification: TypeScript build passed; the full Vitest suite passed; `npm pack --dry-run` passed.
-
----
-
-All notable changes to `@theihtisham/agent-shadow-brain` are documented here.
-
-This project follows [Semantic Versioning](https://semver.org/) and [Keep a Changelog](https://keepachangelog.com/) conventions.
-
-> For detailed release notes with migration guides, see the [docs/versions/](docs/versions/) directory.
-
----
-
-## [6.0.2] — 2026-05-12 — Singularity Edition (Current)
-
-> **5 world-first viral features + real ANN vector index + Constitution layer + Confidence Gate + multimodal vision + LoRA distillation pipeline + SWE-Bench-Lite eval harness. Semver-strict this is v7.0.0; numbered 6.0.2 by author preference. See `docs/launch/RELEASE_NOTES_v6.0.2.md` for the full story.**
-
-### Added — 5 viral world-first features
-
-1. **Brain Replay** (`src/brain/brain-replay.ts`) — scrubbable timeline of every brain event since day zero. Exports tweet-ready SVG with density histogram + milestone markers.
-2. **Brain DNA** (`src/brain/brain-dna.ts`) — Spotify-Wrapped-style 1080×1080 fingerprint card per codebase. 8 personality archetypes (Architect, Debugger, Speedster, Scholar, Cartographer, Diplomat, Sentinel, Wanderer).
-3. **Hive Voice** (`src/brain/hive-voice.ts`) — multi-model debate-vote on any question. Outputs SVG consensus bar chart with explicit dissent + confidence distribution.
-4. **Brain Time Capsule** (`src/brain/brain-time-capsule.ts`) — freeze + label + resurrect brain snapshots. Ed25519-signed (stub) `.tcz` bundles.
-5. **Brain Diff** (`src/brain/brain-diff.ts`) — semantic diff between two brains. Outputs Venn-diagram SVG + English narrative.
-
-### Added — infrastructure
-
-6. **Real ANN vector index v2** (`src/brain/embeddings-v2.ts`) — hand-rolled HNSW-style graph (M=8, efSearch=32). Sub-100ms recall over 10k vectors. Sharded persistence (5MB cap). Auto-pulls `nomic-embed-text` from Ollama on first run. Zero new deps.
-7. **Constitution layer** (`src/brain/constitution.ts`) — `<project>/.shadow-brain/constitution.md` parser + injector + validator. Hot-reload via fs.watch. Logs violations to JSONL.
-8. **Confidence Gate** (`src/brain/confidence-gate.ts`) — active output filter (`block` / `downgrade` / `flag` modes). Persists calibration history; reports Brier score + ECE.
-9. **Multimodal vision** (`src/brain/multimodal-vision.ts`) — Ollama llava/llava-llama3/bakllava integration. Image ingestion → description → embedding → indexed search.
-10. **Markdown brain export/import** (`src/brain/markdown-export.ts`) — git-trackable `.shadow-brain.md` round-trip. Diff before write.
-11. **OpenTelemetry-lite tracing** (`src/brain/otel-lite.ts`) — zero-dep OTel-compatible JSONL tracer. Daily file rotation. `tracer.withSpan(name, fn)` helper.
-
-### Added — research-grade scaffolds
-
-12. **Project-LoRA distillation pipeline** (`tools/lora-pipeline/`) — `generate_training_data.py`, `train_lora.py`, `eval_lora.py`, `serve_lora.py`. Distill the brain into a Qwen2.5-Coder-1.5B LoRA adapter. Run when ready.
-13. **SWE-Bench-Lite micro-eval harness** (`tools/swe-bench-lite/`) — 10 hand-curated public-repo bugs. `runner.py` + `score.py` + pre-filled `report-template.md`. Realistic expected lift: +20-40% resolution rate with brain on.
-
-### Added — DX
-
-14. **`shadow-brain demo` command** — 30-second hello-world: auto-ingests sample repo, runs queries, opens dashboard.
-15. **Forgetting curve transparency + pinning** — every memory shows decay reason; `pin: true` flag prevents auto-decay.
-16. **Dashboard tour video** (`docs/launch/shadow-brain-dashboard-tour.mp4`) — 5-minute narrated walkthrough of every panel.
-17. **MCP server hardening** — better discovery, full feature surface, docs at `/mcp/info`.
-
-### Changed
-
-- `embeddings.ts` kept untouched; `embeddings-v2.ts` composes it (no breaking change for callers that imported v1)
-- `package.json` version → 6.0.2; description updated to reflect new features
-- README hero updated with v6.0.2 feature row + new install command
-
-### Docs
-
-- `docs/launch/RELEASE_NOTES_v6.0.2.md` — full release notes
-- `docs/ROADMAP-v7.md` — v7+ strategic plan
-- `docs/VIRAL-PLAYBOOK.md` — launch playbook
-
-### Notes
-
-- **No new npm runtime dependencies.** Zero-dep philosophy preserved.
-- TypeScript strict mode passes. All new modules compile cleanly.
-- Backwards compatible with v6.0 brains; no migration required.
-
----
-
-## [6.0.0] — 2026-04-24 — Hive Mind Edition
-
-> **Hive Mind for AI agents — the first sub-agent-aware brain. 22 genuinely new features. 100% local-first, free, open source.**
-
-### The Hive Mind
-
-Before v6.0, every AI agent and sub-agent started from zero. Claude's `Task` tool, Cursor's Composer, CrewAI/LangGraph/AutoGen sub-agents — none shared memory with the parent. Shadow Brain v6 is the first tool that bridges the brain across frameworks AND into sub-agents.
-
-### Added — 22 genuinely novel modules
-
-1. **Sub-Agent Brain Bridge (SABB)** (`src/brain/subagent-bridge.ts`) — detects sub-agent spawns, computes focused context slivers, quarantines sub-agent memories until graduated (confidence ≥ 0.7).
-2. **Causal Memory Chains** (`src/brain/causal-chains.ts`) — tracks cause→effect relationships, renders as Graphviz DAG, answers "why did X happen?"
-3. **Agent Collision Detective** (`src/brain/collision-detective.ts`) — real-time overlapping edit intent detection with advisory locks.
-4. **Dream Engine** (`src/brain/dream-engine.ts`) — background idle reflection via local LLM (Ollama): revisit, counterfactual, consolidation, contradiction, pattern-discovery.
-5. **Agent Reputation Ledger** (`src/brain/reputation-ledger.ts`) — Ed25519-signed decision receipts with portable accuracy scores and shields.io badges.
-6. **Swarm Debate Protocol** (`src/brain/swarm-debate.ts`) — pro/con/arbiter multi-agent debate for critical decisions.
-7. **Pre-Mortem Assistant** (`src/brain/pre-mortem.ts`) — surfaces past failures from project memory before significant tasks.
-8. **Branch Brains** (`src/brain/branch-brain.ts`) — git-branch-aware memory context switching.
-9. **Attention Heatmap** (`src/brain/attention-heatmap.ts`) — weighted attribution of which memories influenced each decision.
-10. **Token Economy Engine** (`src/brain/token-economy.ts`) — cross-agent cost tracking, monthly projections, savings suggestions.
-11. **Forgetting Curve + Sleep Consolidation** (`src/brain/forgetting-curve.ts`) — Ebbinghaus-inspired natural decay + tier promotion.
-12. **Formal Verification Bridge** (`src/brain/formal-verification-bridge.ts`) — natural-language rules → ESLint / Semgrep / LSP diagnostics.
-13. **Confidence Calibration Monitor** (`src/brain/calibration-monitor.ts`) — Brier-score based agent trust weighting.
-14. **Cost-Aware Sub-Agent Spawner** (`src/brain/cost-aware-spawner.ts`) — routes sub-agent spawns to cheapest model that satisfies task.
-15. **Air-Gap Mode** (`src/brain/air-gap.ts`) — block outbound network; localhost-only.
-16. **E2E Encryption** (`src/brain/brain-encryption.ts`) — ChaCha20-Poly1305 at rest with scrypt key derivation.
-17. **Hallucination Quarantine** (`src/brain/hallucination-quarantine.ts`) — isolate suspect memories; auto-delete after 7 days.
-18. **Voice Mode** (`src/brain/voice-mode.ts`) — transcript processing with brain integration.
-19. **Brain Garden** (`src/brain/brain-garden.ts`) — living constellation visualization data source.
-20. **PR Auto-Review** (`src/brain/pr-auto-review.ts`) — GitHub PR comment generator citing project memories.
-21. **Team Brain Sync** (`src/brain/team-brain-sync.ts`) — P2P peer state layer for WebRTC team brain.
-22. **Brain Exchange** (`src/brain/brain-exchange.ts`) — export + import curated shareable brain slices with redaction.
-
-### Added — supporting infrastructure
-
-- **Local-First LLM Adapter** (`src/brain/local-llm.ts`) — unified Ollama-first adapter used by Dream Engine, Swarm Debate, Pre-Mortem. Graceful fallback when unreachable.
-- **Hive Accelerator** (`src/brain/hive-accelerator.ts`) — wires SSSP (arXiv 2504.17033) for causal traversal + TurboQuant (ICLR 2026) for similarity scoring.
-- **Complete Web Control Dashboard** (`src/dashboard/v6-dashboard-html.ts` + server routes) — 23-tab SPA at `http://localhost:7341/hive`. Controls every feature: enable/disable, configure, monitor. Agent connect/disconnect. Unified activity log. API+MCP configuration panels. Brain Garden D3 canvas.
-- **48 new CLI commands** grouped under `hive`, `subagent`, `causal`, `collision`, `dream`, `reputation`, `debate`, `premortem`, `branch-brain`, `attention`, `tokens`, `forget`, `formal`, `calibrate`, `airgap`, `encrypt`, `quarantine`, `voice`, `garden`, `pr-review`, `team-sync`, `exchange`.
-- **36 new MCP tool handlers** for every v6 module — any MCP-compatible AI tool can now call them.
-- **v6 integration tests** (`tests/brain/hivemind-v6.test.ts`) — 24 smoke tests covering every module.
-
-### Changed
-
-- Version: `5.2.0` → `6.0.0`
-- Total brain modules: **26 → 48** (22 new v6 modules + Hive Accelerator + Local LLM)
-- CLI commands: **70+ → 90+**
-- MCP tools: **~27 → 60+** (existing + 36 v6 handlers)
-- README hero: v6.0 Hive Mind positioning with 22 novel features prominently listed
-- Web dashboard: `/` (legacy) + `/hive` or `/v6` (complete Hive Mind control)
-- `package.json` description + 25 new v6 keywords for maximum discoverability
-
-### Removed
-
-- Repo hygiene cleanup — removed stray files (`C:Userstheihtmp_b64_3.txt`, `eng.traineddata`, `write_pattern.py`, `.playwright-mcp/`, `seo-gists/`, empty `.codex`). Updated `.gitignore` to prevent recurrence.
-
-### Security
-
-- ChaCha20-Poly1305 AEAD encryption for brain-at-rest
-- Ed25519 cryptographic signatures on agent reputation receipts
-- Air-gap mode with RFC1918 + metadata-endpoint blocking
-- Hallucination quarantine prevents unverified claims from polluting global brain
-
-### Philosophy
-
-**Local-first default.** All AI features (Dream Engine, Swarm Debate, Pre-Mortem, Voice) route to Ollama by default. Zero API keys needed. Zero network required. Zero cost to run. Shadow Brain works offline, in air-gapped environments, and in compliance-constrained enterprises. Remote providers (Anthropic, OpenAI, GLM) remain an optional upgrade — never a requirement.
-
----
-
-## [5.2.0] — 2026-04-22 — Subconscious Singularity Edition
-
-> **One brain. Every agent. Every project. Every session. Forever.**
-
-### The Singularity
-
-Before v5.2.0, every project had its own brain and every AI agent (Claude, Cursor, Cline...) had to be told what each project knew. Now there is **one brain** — a singleton global JSON store at `~/.shadow-brain/global.json` that every agent reads from and writes to.
-
-What Cursor learns in project A is instantly known to Claude Code in project B.
-
-### Added
-
-- **Singleton Global Brain** (`src/brain/global-brain.ts`) — one `~/.shadow-brain/global.json` is the source of truth for every project + agent on the machine. Atomic write-and-rename plus an in-process write queue handles concurrent multi-agent writes without native dependencies. Auto-prune at 1GB. Project-scoped views via `GlobalBrain.projectIdFor(rootDir)`.
-- **Subconscious Engine** (`src/brain/subconscious.ts`) — proactive context injection on every session start. The agent doesn't have to ask — the brain surfaces recent decisions, active tasks, similar past work, project state, cross-agent insights, and warnings. Hard-capped at 2K tokens, relevance-ranked, agent-configurable.
-- **Universal Bootstrap** (`src/brain/session-hooks.ts`) — `asb attach-all` detects every installed AI agent on your machine and installs SessionStart hooks per agent's native mechanism (Claude Code settings.json, Cursor `.cursor/rules`, Cline VS Code extension, Windsurf `.windsurfrules`, Codex config.json, Aider `.aider.conf.yml`, Copilot `.github/copilot-instructions.md`, etc.). Every agent calls Shadow Brain on first prompt of every session.
-- **L0 In-Memory Hot Tier** (`src/brain/l0-cache.ts`) — sub-millisecond recall via byte-budgeted Map LRU. Sits in front of the global JSON brain. Per-namespace caches with `getCache(name)`. Default 64MB budget, configurable.
-- **Agent Safety Firewall** (`src/brain/agent-firewall.ts`) — blocks risky agent actions such as secret file access, destructive commands, curl-to-shell installs, and prompt-injection payloads before tools execute.
-- **Agent Handoff Packets** (`src/brain/agent-handoff.ts`) — `asb handoff cursor codex` generates a cross-agent continuation packet with task state, changed files, recent memories, and safety warnings.
-- **Proof Timeline + Share Report** — `asb timeline` and `asb proof report` expose the viral proof stream: what agents learned, when, and what the firewall blocked.
-- **Cursor Adapter** (`src/adapters/cursor.ts`) — first-class adapter, no more falling back to ClineAdapter. Reads conversations from `~/.cursor/conversations`, injects into `.cursor/rules/shadow-brain-insights.md`, escalates critical insights to `.cursorrules`.
-- **Windsurf Adapter** (`src/adapters/windsurf.ts`) — first-class adapter. Reads from `~/.windsurf/conversations`, writes to `.windsurfrules` with `# === Shadow Brain Insights ===` marker for safe re-injection.
-- **GitHub Copilot Adapter** (`src/adapters/copilot.ts`) — workspace-level integration via `.github/copilot-instructions.md` (Copilot's official context file). Replaces a marker block on each update so injection is idempotent.
-- **4 New CLI Commands**:
-  - `asb attach-all` — universal bootstrap for every detected agent
-  - `asb subconscious inject|status|configure` — manage proactive context injection
-  - `asb global stats|recall|cache|sync` — inspect the singleton global brain
-  - `asb handoff`, `asb timeline`, `asb firewall`, `asb detach-all`, `asb audit-hooks`, `asb proof report` — trust, safety, and launch-proof commands
-- **4 New MCP Tools** — `subconscious_inject`, `global_recall`, `global_stats`, `attach_status`. Now any MCP client can talk to the global brain directly.
-
-### Changed
-
-- **AgentTool union** now includes `'copilot'` — full type-safe support for 10 agents (was 9)
-- **adapters/index.ts** — Cursor and Windsurf no longer alias ClineAdapter; each gets its own implementation
-- **Description + 80 npm keywords** rewritten for v5.2.0 positioning and maximum discoverability
-- **Total brain modules: 26** (4 new + 22 from v5.1.1)
-
-### Performance
-
-- Hot recalls now <1ms via L0 cache (was 10-30ms via SQLite)
-- Cross-project queries replace per-project DB lookup overhead with a single indexed query
-
-### Fixed
-
-- Cursor and Windsurf adapters previously reused ClineAdapter — they now have correct path detection and conversation parsing for their actual file layouts
-
-### Migration Notes
-
-v5.2.0 is **fully backward-compatible** with v5.1.1. The global brain is created on first init and coexists with project-local brains. No breaking changes to existing CLI commands or MCP tools.
-
-To opt into the new singleton model: run `asb attach-all` once and every future session of every agent will auto-bootstrap.
-
----
-
-## [5.1.1] — 2026-04-18 — Hyper-Cognitive Intelligence Edition
-
-### Added
-- **Built-in LSP Server** — pure Node.js Language Server Protocol implementation with stdio + TCP transport modes, real-time diagnostics, hover info, and code actions
-- **Custom LLM Fine-Tuning Engine** — automatic training data generation from code changes, JSONL export, multi-model training pipeline
-- **Smart Cache** — 3-tier (hot/warm/cold) LRU cache with predictive prefetch via co-access graph, tag-based invalidation, and adaptive tier promotion
-- **Intent Engine** — multi-strategy NLP command understanding with keyword, fuzzy, and context-aware matching strategies
-- **Code DNA Fingerprinting** — 9-gene structural analysis covering structure, complexity, style, dependency, and evolution gene categories
-- **Temporal Intelligence** — velocity metrics, file heatmaps, 5-factor bug prediction, anomaly detection, and peak hour/day analysis
-- 6 new CLI commands: `lsp`, `fine-tune`, `cache`, `intent`, `dna`, `temporal`
-- 27 README badges for maximum discoverability
-
-### Changed
-- Total brain modules: **22** (6 new + 16 from v5.0.1)
-- All v5.0.1, v5.0.0, and v4.0.0 features included and enhanced
-- Expanded npm keywords to 50+
-
-### Fixed
-- 18 TypeScript compilation errors in CLI — aligned property references with actual module interfaces
-
----
-
-## [5.0.1] — 2026-04-15 — Zero-Config Intelligence Edition
-
-### Added
-- **Zero-Config Auto-Setup** — auto-detect project type, AI tools, languages, frameworks; setup is explicit and reversible in v5.2.0
-- **MCP Server** — 19-tool Model Context Protocol server for Claude Code, Cursor, and all AI tools
-- **Rich Web Dashboard** — real-time WebSocket dashboard with 8 panels (health, memory tiers, insights, fixes, AI tools, modules, stats, controls)
-- **Natural Language Queries** — `shadow-brain ask "..."` for plain-English brain interaction
-- **Plugin System** — hookable analysis pipeline with npm plugin discovery
-- **Brain Export/Import** — full brain state portability for team sync and backup
-- 5 new CLI commands: `off`, `ask`, `export`, `import`, `plugin`
-- **Cursor MCP Compatibility** — root path POST + `/v1/chat` endpoint support
-- **Postinstall Auto-Setup** — runs automatically after `npm install`
-
-### Changed
-- 50+ npm keywords for discoverability
-- All v5.0.0 and v4.0.0 features included and enhanced
-
----
-
-## [5.0.0] — 2026-04-12 — Infinite Intelligence Edition
-
-### Added
-- **Hierarchical Memory Compression** — 4-tier pyramid (raw -> summary -> pattern -> principle) with drill-down/up navigation
-- **Context-Triggered Associative Recall** — automatic memory activation based on file paths, keywords, and categories
-- **Multi-Agent Consensus Protocol** — trust-weighted voting, confidence intervals, conflict resolution across multiple AI agents
-- **Collective Cross-Project Learning** — verified rule sharing, viral propagation, accuracy tracking across projects
-- 5 new CLI commands: `memory`, `recall`, `consensus`, `collective`, `v5`
-
-### Changed
-- Memory architecture fully rewritten for infinite retention
-- Cross-agent communication protocol standardized
-
----
-
-## [4.0.0] — 2026-04-08 — Hyper-Intelligence Edition
-
-### Added
-- **TurboQuant Infinite Memory** — 6x compression, zero forgetting (based on Google Research, ICLR 2026)
-- **SSSP BMSSP Routing** — O(m log^(2/3) n) deterministic routing (based on arXiv 2504.17033, Duan et al.)
-- **Self-Evolving Genetic Rules** — tournament selection, Gaussian mutation, Bayesian meta-learning
-- **Cross-Agent Intelligence Protocol (CAIP)** — 7 agents share insights in real-time via WebSocket
-- **Adversarial Hallucination Defense** — cross-reference verification, evidence scoring, confidence thresholds
-- **Swarm Intelligence** — Ant Colony Optimization pheromone-based file prioritization
-- **Knowledge Graph + PageRank** — code entity impact radar with d=0.85 damping factor
-- **Predictive Engine** — bug risk scoring, technical debt forecasting, anomaly detection
-- 9 new CLI commands: `turbo`, `route`, `caip`, `evolve`, `predict`, `graph`, `swarm`, `defense`, `v4`
-
-### Changed
-- All v3.0.0 features included and enhanced
-- Neural mesh upgraded with TurboMemory integration
-- Pattern memory upgraded with compressed vector storage
-- Learning engine upgraded with genetic optimization
-
----
-
-## [3.0.0] — 2026-04-01 — Hyper-Intelligence Edition
-
-### Added
-- **Cognitive Load Analysis** — estimates mental complexity of code sections
-- **Security Audit Engine** — vulnerability scanning, dependency risk assessment
-- **Influence Map** — tracks which code changes ripple through the codebase
-- Enhanced neural mesh with Shannon entropy and cosine similarity
-- Bayesian confidence scoring for all insights
-
-### Changed
-- Analysis pipeline fully restructured for hyper-intelligence features
-- CLI expanded with security and cognitive commands
-
----
-
-## [2.1.0] — 2026-03-25 — Quantum Neural Mesh
-
-### Added
-- **Cross-Session Shared Intelligence** — persist and share insights across sessions
-- **Shannon Entropy Analysis** — information-theoretic code complexity measurement
-- **Cosine Similarity Matching** — semantic code comparison
-- **Bayesian Confidence** — probabilistic confidence scoring for recommendations
-
-### Changed
-- Neural mesh architecture upgraded for cross-session persistence
-- Insight quality improved with entropy-based scoring
-
----
-
-## [2.0.0] — 2026-03-18 — Super-Intelligence Edition
-
-### Added
-- **Semantic Analyzer** — deep code understanding beyond syntax
-- **Dependency Graph** — full project dependency visualization and analysis
-- **Code Similarity Engine** — detect duplicated and near-duplicated code
-- **MCP Server** — initial Model Context Protocol server implementation
-- **Team Mode** — collaborative AI-assisted development
-
-### Changed
-- Complete architecture rewrite from v1.x
-- CLI expanded from 15 to 35+ commands
-
----
-
-## [1.x] — 2026-02 to 2026-03 — Foundation
-
-### v1.2.0
-- Added vulnerability scanning
-- Added CI/CD integration hooks
-
-### v1.1.0
-- Added smart fix suggestions
-- Added health scoring with weighted metrics
-
-### v1.0.0
-- Initial release
-- 7 agent adapters (Claude Code, Cursor, Cline, Windsurf, Copilot, Kilo Code, Amp)
-- Basic health scoring and analysis pipeline
-- Vulnerability scanning
-- CLI with 15 commands
-
----
-
-[5.2.0]: https://github.com/theihtisham/agent-shadow-brain/releases/tag/v5.2.0
-[5.1.1]: https://github.com/theihtisham/agent-shadow-brain/releases/tag/v5.1.1
-[5.0.1]: https://github.com/theihtisham/agent-shadow-brain/releases/tag/v5.0.1
-[5.0.0]: https://github.com/theihtisham/agent-shadow-brain/releases/tag/v5.0.0
-[4.0.0]: https://github.com/theihtisham/agent-shadow-brain/releases/tag/v4.0.0
-[3.0.0]: https://github.com/theihtisham/agent-shadow-brain/releases/tag/v3.0.0
-[2.1.0]: https://github.com/theihtisham/agent-shadow-brain/releases/tag/v2.1.0
-[2.0.0]: https://github.com/theihtisham/agent-shadow-brain/releases/tag/v2.0.0
+# Changelog
+
+
+## [7.0.0] - 2026-11-11 - Durable Memory and Explained Recall
+
+### Breaking changes
+
+- Require Node 22.12+ in 22.x, 24.x, or 26+; update dependency majors including simple-git, Zod, Anthropic SDK, and Vitest.
+- Remove automatic install-time agent configuration. Setup is explicit.
+- Reject unreadable, invalid, and unsupported stores without silently resetting or renaming them.
+- Recall scoring uses real keyword coverage, creation-age recency, and matching-content agent reports. Existing minimum-confidence thresholds may return different results.
+
+### Added and fixed
+
+- Serialize cross-process commits; merge latest data with pending operations, preserve external deletions, and use unique temporary files.
+- Acknowledge queued writes after persistence, retain failed changes for retry, and expose structured storage diagnostics.
+- Add `storage-health` and `recall-explain` commands, `global_storage_health` MCP tool, and public GlobalBrain library exports/types.
+- Make scored recall finite and deterministic across the full candidate pool; explain matches and supporting identities.
+- Load CLI configuration only when required, redact diagnostic key output, and escape dashboard activity text.
+- Update dependencies, isolate tests from personal state, correct a date-dependent training-export test, add artifact consumer smoke checks, and expand CI to Windows/Linux on Node 22/24/26.
+- Make legacy promotional automation manual-only.
+
+## [6.3.0] - 2026-06-06 - Recall Confidence + Issue-Bot Spam Fix
+
+This release adds per-memory recall confidence scoring and fixes a GitHub Actions
+issue-bot that was opening duplicate milestone/spike issues.
+
+### Added
+
+- `GlobalBrain.recallScored()`: recall enriched with a blended `RecallConfidence` per entry. Signals: keyword relevance, recency (30-day half-life decay), author-assigned importance, cross-agent corroboration, and prior usage. Results are ranked by confidence and support a `minConfidence` floor.
+- `RecallConfidence` and `ScoredGlobalEntry` types in `src/types.ts`.
+- MCP tool `global_recall_scored` exposing scored recall to compatible clients.
+- Dashboard recall results now show a color-coded confidence badge (high/medium/low) with a per-signal breakdown on hover; the `/api/v6/ww2br-recall` endpoint now returns scored results.
+- Focused tests in `tests/brain/global-recall-confidence.test.ts` for scoring, importance ranking, cross-agent corroboration, and `minConfidence` filtering.
+
+### Fixed
+
+- Viral Amplifier bot opened dozens of duplicate `Milestone — 25 stars` and `Momentum spike` issues. Root causes: (1) the milestone job ran on every cron schedule (including the every-6h trending cron) instead of only the daily run, and (2) both scripts trusted local state files (`.github/.milestone-state.json`, `.github/.trending-state.json`) that the workflow never committed, so each fresh checkout reset state and re-fired.
+- `milestone-notifier.mjs` and `trending-detector.mjs` are now idempotent: they verify against GitHub (open/closed issues for milestones, open issues within the 24h window for spikes) before posting, and treat lookup failures as "skip" to stay spam-safe.
+- `viral-amplifier.yml`: milestone job restricted to the daily 14:00 UTC cron; `welcome-new-issue` now ignores bot-authored issues so the bot no longer triages its own issues.
+
+## [6.0.3] - 2026-06-03 - Live Brain Release
+
+This release makes Shadow Brain more realistic and more useful for complex agent work. It focuses on source-backed long memory, live memory health, recall evals, agent briefings, and release-ready docs.
+
+### Added
+
+- `src/brain/live-brain.ts`: Live Brain report engine for memory health, recall evals, hygiene actions, top memories, and agent briefings.
+- `src/brain/memory-seeder.ts`: verified memory seeding workflow with draft, user approval, agent approval, dual approval, reject, and seed-to-brain states.
+- `shadow-brain live-brain status`: shows memory quality, source coverage, recall health, and recommended actions.
+- `shadow-brain live-brain brief`: writes a markdown briefing for fresh sessions, handoffs, and context recovery.
+- `shadow-brain live-brain eval`: checks whether useful memory can be recalled for agent workflows.
+- `shadow-brain live-brain hygiene`: lists stale, duplicate, low-quality, and source-light memory issues.
+- `shadow-brain live-brain train-pack`: exports source-backed JSONL examples for evals, distillation, or local fine-tuning pipelines.
+- `/api/v6/live-brain` and the v6 dashboard Live Brain tab: dashboard/API route and UI for the same report and markdown briefing.
+- `/api/v6/memory-intelligence`: live memory data for source coverage, freshness, categories, agents, tags, recent writes, and approval queue counts.
+- `/api/v6/seed-*`: dashboard/API workflow for creating research drafts, approving/rejecting them, and committing approved drafts to global memory.
+- Dashboard Memory Seeder tab: select research agent, approval agent, approval mode, project path, sources, category, tags, and importance before writing durable memory.
+- Dashboard launch controls: `shadow-brain dash --host 127.0.0.1 --no-analysis` for fast UI/API startup on large repos.
+- MCP tools: `live_brain_status`, `live_brain_brief`, and `live_brain_train_pack`.
+- Focused tests for Live Brain scoring, briefing, recall eval, JSONL export, and memory seeding approval gates.
+
+### Changed
+
+- README rewritten to be realistic about what the project does and does not do.
+- npm package version updated to `6.0.3`.
+- npm package description replaced with a practical local-first memory description.
+- Runtime-visible versions updated in CLI, auto-update, portability export, orchestrator status, and MCP server metadata.
+- Resume-pack source discovery now includes common repo docs such as `README.md`, `CHANGELOG.md`, `package.json`, and roadmap docs in addition to project-specific agent/progress files.
+- CLI config merging now honors explicit `watchMode: false` from one-shot commands.
+- Token Economy dashboard now separates the current runtime agent from historical model spend records, avoiding misleading labels such as an old Claude model appearing as the active Codex runtime.
+
+### Notes
+
+- "Training" in this release means data export for user-run evals, distillation, or local fine-tuning. Shadow Brain does not modify remote model weights.
+- No new runtime dependencies were added.
+- Verification: TypeScript build passed; the full Vitest suite passed; `npm pack --dry-run` passed.
+
+---
+
+All notable changes to `@theihtisham/agent-shadow-brain` are documented here.
+
+This project follows [Semantic Versioning](https://semver.org/) and [Keep a Changelog](https://keepachangelog.com/) conventions.
+
+> For detailed release notes with migration guides, see the [docs/versions/](docs/versions/) directory.
+
+---
+
+## [6.0.2] — 2026-05-12 — Singularity Edition (Current)
+
+> **5 world-first viral features + real ANN vector index + Constitution layer + Confidence Gate + multimodal vision + LoRA distillation pipeline + SWE-Bench-Lite eval harness. Semver-strict this is v7.0.0; numbered 6.0.2 by author preference. See `docs/launch/RELEASE_NOTES_v6.0.2.md` for the full story.**
+
+### Added — 5 viral world-first features
+
+1. **Brain Replay** (`src/brain/brain-replay.ts`) — scrubbable timeline of every brain event since day zero. Exports tweet-ready SVG with density histogram + milestone markers.
+2. **Brain DNA** (`src/brain/brain-dna.ts`) — Spotify-Wrapped-style 1080×1080 fingerprint card per codebase. 8 personality archetypes (Architect, Debugger, Speedster, Scholar, Cartographer, Diplomat, Sentinel, Wanderer).
+3. **Hive Voice** (`src/brain/hive-voice.ts`) — multi-model debate-vote on any question. Outputs SVG consensus bar chart with explicit dissent + confidence distribution.
+4. **Brain Time Capsule** (`src/brain/brain-time-capsule.ts`) — freeze + label + resurrect brain snapshots. Ed25519-signed (stub) `.tcz` bundles.
+5. **Brain Diff** (`src/brain/brain-diff.ts`) — semantic diff between two brains. Outputs Venn-diagram SVG + English narrative.
+
+### Added — infrastructure
+
+6. **Real ANN vector index v2** (`src/brain/embeddings-v2.ts`) — hand-rolled HNSW-style graph (M=8, efSearch=32). Sub-100ms recall over 10k vectors. Sharded persistence (5MB cap). Auto-pulls `nomic-embed-text` from Ollama on first run. Zero new deps.
+7. **Constitution layer** (`src/brain/constitution.ts`) — `<project>/.shadow-brain/constitution.md` parser + injector + validator. Hot-reload via fs.watch. Logs violations to JSONL.
+8. **Confidence Gate** (`src/brain/confidence-gate.ts`) — active output filter (`block` / `downgrade` / `flag` modes). Persists calibration history; reports Brier score + ECE.
+9. **Multimodal vision** (`src/brain/multimodal-vision.ts`) — Ollama llava/llava-llama3/bakllava integration. Image ingestion → description → embedding → indexed search.
+10. **Markdown brain export/import** (`src/brain/markdown-export.ts`) — git-trackable `.shadow-brain.md` round-trip. Diff before write.
+11. **OpenTelemetry-lite tracing** (`src/brain/otel-lite.ts`) — zero-dep OTel-compatible JSONL tracer. Daily file rotation. `tracer.withSpan(name, fn)` helper.
+
+### Added — research-grade scaffolds
+
+12. **Project-LoRA distillation pipeline** (`tools/lora-pipeline/`) — `generate_training_data.py`, `train_lora.py`, `eval_lora.py`, `serve_lora.py`. Distill the brain into a Qwen2.5-Coder-1.5B LoRA adapter. Run when ready.
+13. **SWE-Bench-Lite micro-eval harness** (`tools/swe-bench-lite/`) — 10 hand-curated public-repo bugs. `runner.py` + `score.py` + pre-filled `report-template.md`. Realistic expected lift: +20-40% resolution rate with brain on.
+
+### Added — DX
+
+14. **`shadow-brain demo` command** — 30-second hello-world: auto-ingests sample repo, runs queries, opens dashboard.
+15. **Forgetting curve transparency + pinning** — every memory shows decay reason; `pin: true` flag prevents auto-decay.
+16. **Dashboard tour video** (`docs/launch/shadow-brain-dashboard-tour.mp4`) — 5-minute narrated walkthrough of every panel.
+17. **MCP server hardening** — better discovery, full feature surface, docs at `/mcp/info`.
+
+### Changed
+
+- `embeddings.ts` kept untouched; `embeddings-v2.ts` composes it (no breaking change for callers that imported v1)
+- `package.json` version → 6.0.2; description updated to reflect new features
+- README hero updated with v6.0.2 feature row + new install command
+
+### Docs
+
+- `docs/launch/RELEASE_NOTES_v6.0.2.md` — full release notes
+- `docs/ROADMAP-v7.md` — v7+ strategic plan
+- `docs/VIRAL-PLAYBOOK.md` — launch playbook
+
+### Notes
+
+- **No new npm runtime dependencies.** Zero-dep philosophy preserved.
+- TypeScript strict mode passes. All new modules compile cleanly.
+- Backwards compatible with v6.0 brains; no migration required.
+
+---
+
+## [6.0.0] — 2026-04-24 — Hive Mind Edition
+
+> **Hive Mind for AI agents — the first sub-agent-aware brain. 22 genuinely new features. 100% local-first, free, open source.**
+
+### The Hive Mind
+
+Before v6.0, every AI agent and sub-agent started from zero. Claude's `Task` tool, Cursor's Composer, CrewAI/LangGraph/AutoGen sub-agents — none shared memory with the parent. Shadow Brain v6 is the first tool that bridges the brain across frameworks AND into sub-agents.
+
+### Added — 22 genuinely novel modules
+
+1. **Sub-Agent Brain Bridge (SABB)** (`src/brain/subagent-bridge.ts`) — detects sub-agent spawns, computes focused context slivers, quarantines sub-agent memories until graduated (confidence ≥ 0.7).
+2. **Causal Memory Chains** (`src/brain/causal-chains.ts`) — tracks cause→effect relationships, renders as Graphviz DAG, answers "why did X happen?"
+3. **Agent Collision Detective** (`src/brain/collision-detective.ts`) — real-time overlapping edit intent detection with advisory locks.
+4. **Dream Engine** (`src/brain/dream-engine.ts`) — background idle reflection via local LLM (Ollama): revisit, counterfactual, consolidation, contradiction, pattern-discovery.
+5. **Agent Reputation Ledger** (`src/brain/reputation-ledger.ts`) — Ed25519-signed decision receipts with portable accuracy scores and shields.io badges.
+6. **Swarm Debate Protocol** (`src/brain/swarm-debate.ts`) — pro/con/arbiter multi-agent debate for critical decisions.
+7. **Pre-Mortem Assistant** (`src/brain/pre-mortem.ts`) — surfaces past failures from project memory before significant tasks.
+8. **Branch Brains** (`src/brain/branch-brain.ts`) — git-branch-aware memory context switching.
+9. **Attention Heatmap** (`src/brain/attention-heatmap.ts`) — weighted attribution of which memories influenced each decision.
+10. **Token Economy Engine** (`src/brain/token-economy.ts`) — cross-agent cost tracking, monthly projections, savings suggestions.
+11. **Forgetting Curve + Sleep Consolidation** (`src/brain/forgetting-curve.ts`) — Ebbinghaus-inspired natural decay + tier promotion.
+12. **Formal Verification Bridge** (`src/brain/formal-verification-bridge.ts`) — natural-language rules → ESLint / Semgrep / LSP diagnostics.
+13. **Confidence Calibration Monitor** (`src/brain/calibration-monitor.ts`) — Brier-score based agent trust weighting.
+14. **Cost-Aware Sub-Agent Spawner** (`src/brain/cost-aware-spawner.ts`) — routes sub-agent spawns to cheapest model that satisfies task.
+15. **Air-Gap Mode** (`src/brain/air-gap.ts`) — block outbound network; localhost-only.
+16. **E2E Encryption** (`src/brain/brain-encryption.ts`) — ChaCha20-Poly1305 at rest with scrypt key derivation.
+17. **Hallucination Quarantine** (`src/brain/hallucination-quarantine.ts`) — isolate suspect memories; auto-delete after 7 days.
+18. **Voice Mode** (`src/brain/voice-mode.ts`) — transcript processing with brain integration.
+19. **Brain Garden** (`src/brain/brain-garden.ts`) — living constellation visualization data source.
+20. **PR Auto-Review** (`src/brain/pr-auto-review.ts`) — GitHub PR comment generator citing project memories.
+21. **Team Brain Sync** (`src/brain/team-brain-sync.ts`) — P2P peer state layer for WebRTC team brain.
+22. **Brain Exchange** (`src/brain/brain-exchange.ts`) — export + import curated shareable brain slices with redaction.
+
+### Added — supporting infrastructure
+
+- **Local-First LLM Adapter** (`src/brain/local-llm.ts`) — unified Ollama-first adapter used by Dream Engine, Swarm Debate, Pre-Mortem. Graceful fallback when unreachable.
+- **Hive Accelerator** (`src/brain/hive-accelerator.ts`) — wires SSSP (arXiv 2504.17033) for causal traversal + TurboQuant (ICLR 2026) for similarity scoring.
+- **Complete Web Control Dashboard** (`src/dashboard/v6-dashboard-html.ts` + server routes) — 23-tab SPA at `http://localhost:7341/hive`. Controls every feature: enable/disable, configure, monitor. Agent connect/disconnect. Unified activity log. API+MCP configuration panels. Brain Garden D3 canvas.
+- **48 new CLI commands** grouped under `hive`, `subagent`, `causal`, `collision`, `dream`, `reputation`, `debate`, `premortem`, `branch-brain`, `attention`, `tokens`, `forget`, `formal`, `calibrate`, `airgap`, `encrypt`, `quarantine`, `voice`, `garden`, `pr-review`, `team-sync`, `exchange`.
+- **36 new MCP tool handlers** for every v6 module — any MCP-compatible AI tool can now call them.
+- **v6 integration tests** (`tests/brain/hivemind-v6.test.ts`) — 24 smoke tests covering every module.
+
+### Changed
+
+- Version: `5.2.0` → `6.0.0`
+- Total brain modules: **26 → 48** (22 new v6 modules + Hive Accelerator + Local LLM)
+- CLI commands: **70+ → 90+**
+- MCP tools: **~27 → 60+** (existing + 36 v6 handlers)
+- README hero: v6.0 Hive Mind positioning with 22 novel features prominently listed
+- Web dashboard: `/` (legacy) + `/hive` or `/v6` (complete Hive Mind control)
+- `package.json` description + 25 new v6 keywords for maximum discoverability
+
+### Removed
+
+- Repo hygiene cleanup — removed stray files (`C:Userstheihtmp_b64_3.txt`, `eng.traineddata`, `write_pattern.py`, `.playwright-mcp/`, `seo-gists/`, empty `.codex`). Updated `.gitignore` to prevent recurrence.
+
+### Security
+
+- ChaCha20-Poly1305 AEAD encryption for brain-at-rest
+- Ed25519 cryptographic signatures on agent reputation receipts
+- Air-gap mode with RFC1918 + metadata-endpoint blocking
+- Hallucination quarantine prevents unverified claims from polluting global brain
+
+### Philosophy
+
+**Local-first default.** All AI features (Dream Engine, Swarm Debate, Pre-Mortem, Voice) route to Ollama by default. Zero API keys needed. Zero network required. Zero cost to run. Shadow Brain works offline, in air-gapped environments, and in compliance-constrained enterprises. Remote providers (Anthropic, OpenAI, GLM) remain an optional upgrade — never a requirement.
+
+---
+
+## [5.2.0] — 2026-04-22 — Subconscious Singularity Edition
+
+> **One brain. Every agent. Every project. Every session. Forever.**
+
+### The Singularity
+
+Before v5.2.0, every project had its own brain and every AI agent (Claude, Cursor, Cline...) had to be told what each project knew. Now there is **one brain** — a singleton global JSON store at `~/.shadow-brain/global.json` that every agent reads from and writes to.
+
+What Cursor learns in project A is instantly known to Claude Code in project B.
+
+### Added
+
+- **Singleton Global Brain** (`src/brain/global-brain.ts`) — one `~/.shadow-brain/global.json` is the source of truth for every project + agent on the machine. Atomic write-and-rename plus an in-process write queue handles concurrent multi-agent writes without native dependencies. Auto-prune at 1GB. Project-scoped views via `GlobalBrain.projectIdFor(rootDir)`.
+- **Subconscious Engine** (`src/brain/subconscious.ts`) — proactive context injection on every session start. The agent doesn't have to ask — the brain surfaces recent decisions, active tasks, similar past work, project state, cross-agent insights, and warnings. Hard-capped at 2K tokens, relevance-ranked, agent-configurable.
+- **Universal Bootstrap** (`src/brain/session-hooks.ts`) — `asb attach-all` detects every installed AI agent on your machine and installs SessionStart hooks per agent's native mechanism (Claude Code settings.json, Cursor `.cursor/rules`, Cline VS Code extension, Windsurf `.windsurfrules`, Codex config.json, Aider `.aider.conf.yml`, Copilot `.github/copilot-instructions.md`, etc.). Every agent calls Shadow Brain on first prompt of every session.
+- **L0 In-Memory Hot Tier** (`src/brain/l0-cache.ts`) — sub-millisecond recall via byte-budgeted Map LRU. Sits in front of the global JSON brain. Per-namespace caches with `getCache(name)`. Default 64MB budget, configurable.
+- **Agent Safety Firewall** (`src/brain/agent-firewall.ts`) — blocks risky agent actions such as secret file access, destructive commands, curl-to-shell installs, and prompt-injection payloads before tools execute.
+- **Agent Handoff Packets** (`src/brain/agent-handoff.ts`) — `asb handoff cursor codex` generates a cross-agent continuation packet with task state, changed files, recent memories, and safety warnings.
+- **Proof Timeline + Share Report** — `asb timeline` and `asb proof report` expose the viral proof stream: what agents learned, when, and what the firewall blocked.
+- **Cursor Adapter** (`src/adapters/cursor.ts`) — first-class adapter, no more falling back to ClineAdapter. Reads conversations from `~/.cursor/conversations`, injects into `.cursor/rules/shadow-brain-insights.md`, escalates critical insights to `.cursorrules`.
+- **Windsurf Adapter** (`src/adapters/windsurf.ts`) — first-class adapter. Reads from `~/.windsurf/conversations`, writes to `.windsurfrules` with `# === Shadow Brain Insights ===` marker for safe re-injection.
+- **GitHub Copilot Adapter** (`src/adapters/copilot.ts`) — workspace-level integration via `.github/copilot-instructions.md` (Copilot's official context file). Replaces a marker block on each update so injection is idempotent.
+- **4 New CLI Commands**:
+  - `asb attach-all` — universal bootstrap for every detected agent
+  - `asb subconscious inject|status|configure` — manage proactive context injection
+  - `asb global stats|recall|cache|sync` — inspect the singleton global brain
+  - `asb handoff`, `asb timeline`, `asb firewall`, `asb detach-all`, `asb audit-hooks`, `asb proof report` — trust, safety, and launch-proof commands
+- **4 New MCP Tools** — `subconscious_inject`, `global_recall`, `global_stats`, `attach_status`. Now any MCP client can talk to the global brain directly.
+
+### Changed
+
+- **AgentTool union** now includes `'copilot'` — full type-safe support for 10 agents (was 9)
+- **adapters/index.ts** — Cursor and Windsurf no longer alias ClineAdapter; each gets its own implementation
+- **Description + 80 npm keywords** rewritten for v5.2.0 positioning and maximum discoverability
+- **Total brain modules: 26** (4 new + 22 from v5.1.1)
+
+### Performance
+
+- Hot recalls now <1ms via L0 cache (was 10-30ms via SQLite)
+- Cross-project queries replace per-project DB lookup overhead with a single indexed query
+
+### Fixed
+
+- Cursor and Windsurf adapters previously reused ClineAdapter — they now have correct path detection and conversation parsing for their actual file layouts
+
+### Migration Notes
+
+v5.2.0 is **fully backward-compatible** with v5.1.1. The global brain is created on first init and coexists with project-local brains. No breaking changes to existing CLI commands or MCP tools.
+
+To opt into the new singleton model: run `asb attach-all` once and every future session of every agent will auto-bootstrap.
+
+---
+
+## [5.1.1] — 2026-04-18 — Hyper-Cognitive Intelligence Edition
+
+### Added
+- **Built-in LSP Server** — pure Node.js Language Server Protocol implementation with stdio + TCP transport modes, real-time diagnostics, hover info, and code actions
+- **Custom LLM Fine-Tuning Engine** — automatic training data generation from code changes, JSONL export, multi-model training pipeline
+- **Smart Cache** — 3-tier (hot/warm/cold) LRU cache with predictive prefetch via co-access graph, tag-based invalidation, and adaptive tier promotion
+- **Intent Engine** — multi-strategy NLP command understanding with keyword, fuzzy, and context-aware matching strategies
+- **Code DNA Fingerprinting** — 9-gene structural analysis covering structure, complexity, style, dependency, and evolution gene categories
+- **Temporal Intelligence** — velocity metrics, file heatmaps, 5-factor bug prediction, anomaly detection, and peak hour/day analysis
+- 6 new CLI commands: `lsp`, `fine-tune`, `cache`, `intent`, `dna`, `temporal`
+- 27 README badges for maximum discoverability
+
+### Changed
+- Total brain modules: **22** (6 new + 16 from v5.0.1)
+- All v5.0.1, v5.0.0, and v4.0.0 features included and enhanced
+- Expanded npm keywords to 50+
+
+### Fixed
+- 18 TypeScript compilation errors in CLI — aligned property references with actual module interfaces
+
+---
+
+## [5.0.1] — 2026-04-15 — Zero-Config Intelligence Edition
+
+### Added
+- **Zero-Config Auto-Setup** — auto-detect project type, AI tools, languages, frameworks; setup is explicit and reversible in v5.2.0
+- **MCP Server** — 19-tool Model Context Protocol server for Claude Code, Cursor, and all AI tools
+- **Rich Web Dashboard** — real-time WebSocket dashboard with 8 panels (health, memory tiers, insights, fixes, AI tools, modules, stats, controls)
+- **Natural Language Queries** — `shadow-brain ask "..."` for plain-English brain interaction
+- **Plugin System** — hookable analysis pipeline with npm plugin discovery
+- **Brain Export/Import** — full brain state portability for team sync and backup
+- 5 new CLI commands: `off`, `ask`, `export`, `import`, `plugin`
+- **Cursor MCP Compatibility** — root path POST + `/v1/chat` endpoint support
+- **Postinstall Auto-Setup** — runs automatically after `npm install`
+
+### Changed
+- 50+ npm keywords for discoverability
+- All v5.0.0 and v4.0.0 features included and enhanced
+
+---
+
+## [5.0.0] — 2026-04-12 — Infinite Intelligence Edition
+
+### Added
+- **Hierarchical Memory Compression** — 4-tier pyramid (raw -> summary -> pattern -> principle) with drill-down/up navigation
+- **Context-Triggered Associative Recall** — automatic memory activation based on file paths, keywords, and categories
+- **Multi-Agent Consensus Protocol** — trust-weighted voting, confidence intervals, conflict resolution across multiple AI agents
+- **Collective Cross-Project Learning** — verified rule sharing, viral propagation, accuracy tracking across projects
+- 5 new CLI commands: `memory`, `recall`, `consensus`, `collective`, `v5`
+
+### Changed
+- Memory architecture fully rewritten for infinite retention
+- Cross-agent communication protocol standardized
+
+---
+
+## [4.0.0] — 2026-04-08 — Hyper-Intelligence Edition
+
+### Added
+- **TurboQuant Infinite Memory** — 6x compression, zero forgetting (based on Google Research, ICLR 2026)
+- **SSSP BMSSP Routing** — O(m log^(2/3) n) deterministic routing (based on arXiv 2504.17033, Duan et al.)
+- **Self-Evolving Genetic Rules** — tournament selection, Gaussian mutation, Bayesian meta-learning
+- **Cross-Agent Intelligence Protocol (CAIP)** — 7 agents share insights in real-time via WebSocket
+- **Adversarial Hallucination Defense** — cross-reference verification, evidence scoring, confidence thresholds
+- **Swarm Intelligence** — Ant Colony Optimization pheromone-based file prioritization
+- **Knowledge Graph + PageRank** — code entity impact radar with d=0.85 damping factor
+- **Predictive Engine** — bug risk scoring, technical debt forecasting, anomaly detection
+- 9 new CLI commands: `turbo`, `route`, `caip`, `evolve`, `predict`, `graph`, `swarm`, `defense`, `v4`
+
+### Changed
+- All v3.0.0 features included and enhanced
+- Neural mesh upgraded with TurboMemory integration
+- Pattern memory upgraded with compressed vector storage
+- Learning engine upgraded with genetic optimization
+
+---
+
+## [3.0.0] — 2026-04-01 — Hyper-Intelligence Edition
+
+### Added
+- **Cognitive Load Analysis** — estimates mental complexity of code sections
+- **Security Audit Engine** — vulnerability scanning, dependency risk assessment
+- **Influence Map** — tracks which code changes ripple through the codebase
+- Enhanced neural mesh with Shannon entropy and cosine similarity
+- Bayesian confidence scoring for all insights
+
+### Changed
+- Analysis pipeline fully restructured for hyper-intelligence features
+- CLI expanded with security and cognitive commands
+
+---
+
+## [2.1.0] — 2026-03-25 — Quantum Neural Mesh
+
+### Added
+- **Cross-Session Shared Intelligence** — persist and share insights across sessions
+- **Shannon Entropy Analysis** — information-theoretic code complexity measurement
+- **Cosine Similarity Matching** — semantic code comparison
+- **Bayesian Confidence** — probabilistic confidence scoring for recommendations
+
+### Changed
+- Neural mesh architecture upgraded for cross-session persistence
+- Insight quality improved with entropy-based scoring
+
+---
+
+## [2.0.0] — 2026-03-18 — Super-Intelligence Edition
+
+### Added
+- **Semantic Analyzer** — deep code understanding beyond syntax
+- **Dependency Graph** — full project dependency visualization and analysis
+- **Code Similarity Engine** — detect duplicated and near-duplicated code
+- **MCP Server** — initial Model Context Protocol server implementation
+- **Team Mode** — collaborative AI-assisted development
+
+### Changed
+- Complete architecture rewrite from v1.x
+- CLI expanded from 15 to 35+ commands
+
+---
+
+## [1.x] — 2026-02 to 2026-03 — Foundation
+
+### v1.2.0
+- Added vulnerability scanning
+- Added CI/CD integration hooks
+
+### v1.1.0
+- Added smart fix suggestions
+- Added health scoring with weighted metrics
+
+### v1.0.0
+- Initial release
+- 7 agent adapters (Claude Code, Cursor, Cline, Windsurf, Copilot, Kilo Code, Amp)
+- Basic health scoring and analysis pipeline
+- Vulnerability scanning
+- CLI with 15 commands
+
+---
+
+[5.2.0]: https://github.com/theihtisham/agent-shadow-brain/releases/tag/v5.2.0
+[5.1.1]: https://github.com/theihtisham/agent-shadow-brain/releases/tag/v5.1.1
+[5.0.1]: https://github.com/theihtisham/agent-shadow-brain/releases/tag/v5.0.1
+[5.0.0]: https://github.com/theihtisham/agent-shadow-brain/releases/tag/v5.0.0
+[4.0.0]: https://github.com/theihtisham/agent-shadow-brain/releases/tag/v4.0.0
+[3.0.0]: https://github.com/theihtisham/agent-shadow-brain/releases/tag/v3.0.0
+[2.1.0]: https://github.com/theihtisham/agent-shadow-brain/releases/tag/v2.1.0
+[2.0.0]: https://github.com/theihtisham/agent-shadow-brain/releases/tag/v2.0.0