v4.0.0 - Open Source Release

🧬 Concierge Graph v4.0.0 — Phase 4: Extreme Resilience & Secure by Default

This production version consolidates Phase 4 (Extreme Resilience) of the Concierge Graph, introducing advanced concurrent data integrity algorithms, token economy via deep Abstract Syntax Tree (AST) analysis, and local network infrastructure protection.

We leave behind the conceptual v3.8.3 version to deliver a stable, end-to-end audited engine, fully covered by a rigorous suite of green E2E integration tests (without mocks).

---

🛡️ The 4 Survival Shields (SDDs 09 to 12)

1. SDD-09: Shield Against Indirect Cyclic Loops (GraphRAG)

• What changed: Update of the recursive traversal algorithm (WITH RECURSIVE in SQLite WAL).
• Solution: Introduction of a visited path accumulator (path_visited) using rigid delimiters (e.g., |ID|) and low-latency search in C via SQLite's native instr function.
• Impact: Permanently resolves infinite loops in complex graphs (e.g., $A \rightarrow B \rightarrow C \rightarrow A$) and eliminates substring collisions for files with similar names (such as auth.js vs oauth.js).

2. SDD-10: Opportunistic Write Queue with Auto-Batching and Atomic Fallback

• What changed: Physical optimization of the concurrent write pipeline in SQLite.
• Solution: Implementation of Adaptive Opportunistic Batching (groups up to 50 concurrent writes into a single transaction without artificial timers) combined with Single-Item Fallback (if the batch fails due to any key or integrity constraint, the system rolls back the transaction and attempts to write healthy items individually).
• Impact: Up to 10x increase in disk write throughput under intense concurrency from multiple agents.

3. SDD-11: Semantic Drift Guard (Semantic Deviation Detection)

• What changed: Smart token consumption optimization in the DeltaManager.
• Solution: Double hashing mechanism. In addition to SSH (public signature), we implemented the LBH (Logical Body Hash) by analyzing the syntactic body of functions through Python's AST trees. The system utilizes a custom DocstringStripper to ignore extra spaces, cosmetic formatting, and changes in comments or docstrings.
• Impact: Purely cosmetic changes do not consume tokens. Internal logic changes of just a single character inside a function body are surgically detected and force context re-summarization in the Graph.

4. SDD-12: Smart Checkpoint Pruning (Auto-Pruning of Checkpoints)

• What changed: Physical control over the growth/bloating of the relational state database.
• Solution: Asynchronous cleanup algorithm based on Smart LRU per session, integrated into the BackgroundJanitor class.
• Impact: Limits saved checkpoints strictly to a healthy threshold of $N$ recent records, but unconditionally protects and never deletes the first initialization checkpoint (init), which is essential for agents to perform a full rollback or Hard Reset in case of failure.

---

🔒 Infrastructure: Secure by Default

• Localhost Binding: The Qdrant vector database and satellite services now strictly listen on address 127.0.0.1 by default in our Docker structure, shielding network ports from malicious access on public Wi-Fi networks.
• Tailscale Ready: Users operating across multiple devices can securely expose network traffic by changing the CONCIERGE_BIND_ADDRESS environment variable to 0.0.0.0 in the .env file, routing traffic through Tailscale's encrypted virtual interface.

---

🧪 Test Coverage & Code Quality

• Integration Test Suite: 100% success (23 automated tests without mocks covering all survival slices end-to-end).
• Compatibility: Full backward compatibility with legacy MCP search and ingestion tools.