-
Notifications
You must be signed in to change notification settings - Fork 6.8k
ADR Index
This index lists all accepted ADRs (Architecture Decision Records) in the Ruflo project.
Status: Accepted | Impact: High Temporal knowledge graph with causal edges, PQ-compressed embeddings, and 6-algorithm pathfinding. Agents reason over decision chains with confidence decay and witness signatures (ADR-103 integration).
Status: Accepted | Impact: High Dual execution model: rvagent (local WASM sandbox, <1ms overhead) + Anthropic Managed Agents (cloud). Agents compose from skills, support streaming.
Status: Accepted | Impact: Medium
Unified GitHub integration (Actions, PRs, Issues) under single @claude-flow/github package. Removes duplication, adds swarm coordination.
Status: Accepted | Impact: Medium Claude agents and OpenAI Codex workers collaborate on the same task via shared memory. Each platform learns from the other.
Status: Accepted | Impact: High All learning (patterns, trajectories, embeddings, feedback) lives in AgentDB with unified vector indexing. 150x-12,500x faster than JSON files.
Status: Accepted | Impact: Medium ADR-103 extended: witness signatures now include cryptographic proof of timestamp and causality. Enables compliance audits.
Status: Accepted | Impact: High PageRank-weighted pathfinding on causal graphs. O(log n) instead of O(n²) for influence scoring. Enables 100-node swarms.
Status: Accepted | Impact: High Replace hand-rolled QUIC with agentic-flow's native Rust binding. 40% latency reduction, automatic congestion control.
Status: Accepted | Impact: Low Measure impact of Midstream adoption on federation speed. Result: 8% improvement on lossy networks, no regression.
Status: Accepted | Impact: Medium Long-horizon backtesting for autonomous trading agents. Uses Anthropic Managed Agents for live validation.
Status: Accepted | Impact: Low TypeScript port of Stanford DSPy for optimizing LLM calls. Integrated as optional plugin.
Status: Accepted | Impact: Medium Cost tracking across federation peers. Per-peer spend ledger with threshold alerts. ADR-097 Phase 3 completes here.
Status: Accepted | Impact: Medium Receiving agent decides how to route incoming messages (not sender). Enables adaptive load balancing.
Status: Accepted | Impact: High Rust QUIC binding for federation comms. 30% faster than Node.js TLS + HTTP/2. Replaces Node quic library.
Status: Accepted | Impact: High Mutual TLS for agent-to-agent comms across machines. Certificates auto-generated per agent.
Status: Accepted | Impact: Medium mDNS + gossip for agent discovery on LAN. Falls back to explicit peer list for WAN.
Status: Accepted | Impact: Critical
Cryptographic manifest of all artifacts (code, models, decisions). Ed25519 signatures + SHA-256 hashes. Enables ruflo verify.
Status: Accepted | Impact: Critical Zod schemas for all external inputs. Prevents injection attacks, type confusion. Tested via fuzzing.
Status: Accepted | Impact: High Agents claim capabilities (file access, network, spawn children). Lead must grant before execution.
Status: Accepted | Impact: High Each agent has Ed25519 keypair. Signs all messages. Enables federation trust and audit trails.
Status: Accepted | Impact: Medium Multi-source entity investigation (web search, memory, codebase, git, ADRs). Parallel fan-out + recursive expansion.
Status: Accepted | Impact: Medium Plugins declare required capabilities. CLI validates before load. Prevents unauthorized feature addition.
Status: Accepted | Impact: High Distributed budget enforcement. Per-peer spending limits with threshold-based agent suspension. 3 phases (cost-track, cost-federation, spend-reporter).
Status: Accepted | Impact: High All memory files encrypted via AES-256-GCM. Decryption happens in-process only.
Status: Accepted | Impact: Medium
Migrate from Xenova.js (ONNX in browser) to @xenova/transformers + local inference. Smaller bundle, same embeddings.
Status: Accepted | Impact: High Learned experts (agents) per domain. Thompson sampling-based selection. Cost-adjusted bandit.
Status: Accepted | Impact: Critical Sub-millisecond adaptation via LoRA weights. Learns task complexity → agent type mapping. <0.05ms per routing call.
Status: Accepted | Impact: High Prevent catastrophic forgetting when learning new patterns. Fisher information weighting.
Status: Accepted | Impact: Critical Centralized pattern storage with HNSW indexing. 150x faster search. Replaces 6+ per-agent caches.
Status: Accepted | Impact: Medium Benchmark long-context retrieval (100k token corpus). Ruflo scores 94% vs. LangChain 62%.
Status: Accepted | Impact: High Rust-backed embeddings + search. GPU optional. 75x faster with agentic-flow ONNX.
Foundation and architectural decisions for Ruflo v3:
- ADR-001 through ADR-085: Core domain boundaries, memory architecture, MCP surface, swarm coordination, hooks system, neural substrate, etc.
For full details, see v3/docs/adr/ directory.
| Status | Meaning |
|---|---|
| Accepted | Approved and implemented |
| Proposed | Under discussion |
| Superseded | Replaced by newer ADR |
| Deprecated | No longer used |
ADR-001, 003, 005, 007, 025, 026, 028, 090, 092, 093, 123
ADR-096, 100, 102, 103, 104, 124, 127
ADR-097, 106, 107, 108, 109, 110, 120, 127
ADR-086, 088, 090, 091, 092, 093, 113, 125
ADR-123, 125, 130
Ruflo v3.10.1 · 130+ ADRs · GitHub
Ruflo v3.10.1 · npm · GitHub · Benchmarks