SPFsmartGATE / TRIAD-CMD /SPF-FEATURES.md
JosephStoneCellAI's picture
Add files using upload-large-folder tool
001b2a5 verified
# SPFsmartGATE v3.0.0 — Complete Feature List
### 18,500+ lines of Rust | 42 modules | 7 LMDB databases | Multi-agent mesh network | FLINT on-device AI
*Copyright 2026 Joseph Stone — All Rights Reserved*
*Updated: 2026-04-05 (full audit against LIVE src/)*
---
## I. THE GATE — Primary Enforcement Engine
*(gate.rs — 493 lines)*
The core pipeline every single tool call passes through.
### 1. 5-Stage Gate Pipeline
Every tool call goes through this exact sequence:
1. **Calculate** complexity (C value, tier, allocation)
2. **CRITICAL escalation** — if tier = CRITICAL, force max protocol
3. **Validate** against compiled rules (paths, anchor, whitelist, dangerous commands)
4. **Content inspection** on Write/Edit operations
5. **Return** allow/block decision with full audit trail
### 2. GateDecision Struct
Every gate result includes:
- allowed: bool — final verdict
- tool: String — which tool was called
- complexity: ComplexityResult — full C breakdown
- warnings: Vec<String> — non-blocking issues
- errors: Vec<String> — blocking violations
- message: String — human-readable summary
### 3. Source-Centric Decisions
- Gate receives `Source` enum (Stdio/Http/Mesh/Transformer/Pipeline)
- Decision context includes peer identity, role, trust level
---
## II. COMPLEXITY CALCULATOR — The SPF Formula
*(calculate.rs — 416 lines)*
### 4. 4-Factor Exponential Complexity Formula
C = (basic ^ 1) + (dependencies ^ 7) + (complex ^ 10) + (files × 10)
- basic to the 1st — linear base cost
- deps to the 7th — cascading dependency penalty (2→128, 3→2187, 5→78125)
- complex to the 10th — nuclear escalation (3→59049, 4→1048576)
- files times 10 — linear file count penalty
### 5. Master Allocation Formula
a_optimal(C) = W_eff × (1 - 1/ln(C + e))
- W_eff = 40,000 tokens (effective working memory)
- Maps C → optimal analysis token budget
### 6. 4-Tier Classification
| Tier | C Range | Analyze | Build |
|------|---------|---------|-------|
| SIMPLE | < 500 | 40% | 60% |
| LIGHT | < 2,000 | 60% | 40% |
| MEDIUM | < 10,000| 75% | 25% |
| CRITICAL | > 10,000| 95% | 5% |
### 7. Per-Tool Weight Profiles
| Category | Basic | Deps | Complex | Files |
|----------|-------|------|---------|-------|
| Read | 5 | 1 | 0 | 1 |
| Search | 8 | 2 | 0 | 1 |
| Edit | 10 | 2 | 1 | 1 |
| Write | 20 | 2 | 1 | 1 |
| Bash (dangerous) | 50 | 5 | 2 | 1 |
### 8. Dynamic Complexity Boosters
- Content size escalation (+1 per threshold: 200/1000/5000 bytes)
- Risk keyword detection (delete, drop, force, unsafe, rm, sudo)
- Architectural file detection (config, main, lib, Cargo.toml, yaml, etc.)
- Scope estimation (glob patterns, find, xargs)
---
## III. VALIDATION ENGINE — Compiled Rules
*(validate.rs — 1,481 lines)*
### 9. Compiled Write Allowlist
Only two paths accept writes: LIVE/PROJECTS/PROJECTS/ and LIVE/TMP/TMP/
Resolved dynamically from `spf_root()` — portable, compiled-in, runtime-immutable.
### 10. Symlink-Aware Path Resolution
- `canonicalize()` for existing files, parent canonicalize for new files
- Filenames with `..` rejected outright
- Broken symlinks = blocked
### 11. Build Anchor Protocol
- Hard mode: must read file before edit/overwrite → violation = BLOCKED
- Per-file tracking via `session.files_read`
### 12. Dangerous Command Detection
Blocked patterns: `rm -rf`, `dd if=`, `> /dev/`, `chmod 777`, `curl|sh`, `wget|sh`
### 13. /tmp Access Block — hard block on all platforms (SPF policy)
### 14. File Size Limit — max_write_size default 100,000 bytes
---
## IV. CONTENT INSPECTION — Deep Scan
*(inspect.rs — 253 lines)*
### 15. Credential Detection (20 patterns)
API keys, GitHub tokens, GitLab PATs, Slack tokens, PEM private keys, hardcoded credentials.
### 16. Shell Injection Detection
Command substitution, backtick exec, eval/exec statements (skipped for code files).
### 17. Path Traversal & Blocked Reference Detection
Scans content for `../` sequences and references to blocked paths.
### 18. Context-Aware Inspection
Code files (.rs/.py/.js/.ts/.sh/.json/.toml) skip shell injection — only credentials, traversal, blocked refs checked.
---
## V. CONFIGURATION SYSTEM
*(config.rs — 917 lines, config_db.rs — 797 lines)*
### 19. SpfConfig — Master Configuration Struct
All settings: enforce_mode, paths, max_write_size, tiers, formula, weights, dangerous_commands, git_force_patterns.
### 20. CONFIG.DB — LMDB-Backed Configuration
- Namespaced key-value: `namespace:key` → JSON value
- Path rules DB, Dangerous patterns DB with severity (1-10)
- Hot reload — no restart needed
- Import/export CLI commands
### 21. Canonicalized Path Blocking
`is_path_blocked()` canonicalizes before checking — catches traversal at config level.
---
## VI. SESSION TRACKING
*(session.rs — 216 lines, storage.rs — 100 lines)*
### 22. In-Memory Session State
Tracks: action_count, files_read, files_written, last_tool, complexity_history (100 entries), manifest (200 decisions), failures.
### 23. SESSION.DB — LMDB Persistence
Serialized to JSON in LMDB under `current_session` key. ACID-compliant, crash-safe.
---
## VII. LMDB VIRTUAL FILESYSTEM
*(fs.rs — 666 lines)*
### 24. Hybrid Storage Architecture
- ≤ 1MB: inline in LMDB (zero-copy via mmap)
- > 1MB: disk blobs with LMDB metadata pointer
- SHA-256 checksums, 4GB map size
### 25. Full POSIX-Like Operations
exists, stat, read, write, mkdir, mkdir_p, ls, rm, rm_rf, rename (atomic).
### 26. FileMetadata
file_type, size, mode, created/modified, checksum, version (auto-increment), vector_id (reverse RAG link), real_path (disk blob pointer).
### 27. Vector Index — bridges filesystem to brain: `path ↔ vector_id`
---
## VIII. LMDB PARTITION ROUTING — Virtual Mount Points
*(mcp.rs — routing section)*
### 28. Multi-LMDB Mount Architecture
| Virtual Path | Backend | R/W |
|-------------|---------|-----|
| / (root) | SPF_FS.DB | R/W |
| /config/ | CONFIG.DB | Read-only |
| /tmp/ | LIVE/TMP/ | Device-backed |
| /projects/ | LIVE/PROJECTS/ | Device-backed |
| /home/agent/ | AGENT_STATE.DB | Read-only |
### 29. Config Virtual Files
/config/version, /config/mode, /config/tiers, /config/formula, /config/weights, /config/paths, /config/patterns
---
## IX. PROJECTS REGISTRY
*(projects_db.rs — 89 lines)*
### 30. PROJECTS.DB — LMDB Registry
Generic KV store for project metadata. get/set/delete/list_all. 20MB map. Starts empty.
---
## X. TMP DATABASE — Project Trust & Tracking
*(tmp_db.rs — 609 lines)*
### 31. 4 Internal LMDB Databases
- **projects** — canonical_path → Project struct
- **access_log** — timestamped file access records
- **resources** — per-project byte/file/command counters
- **active** — currently active project path
### 32. 5-Level Trust System
| Level | Value | Access |
|-------|-------|--------|
| Untrusted | 0 | All denied |
| Low | 1 | Read/Glob/Grep only |
| Medium | 2 | All except Bash |
| High | 3 | All tools |
| Full | 4 | No restrictions |
### 33. Per-Project Controls
allowed_tools, denied_tools, protected_paths, max_write_size, max_writes_per_session, session_writes counter, requires_activation flag, lifetime resource tracking.
### 34. Operation Validation Pipeline
Project lookup → Activation check → Trust level → Protected paths → Write size → Session rate limit
---
## XI. AGENT STATE — Persistent Memory
*(agent_state.rs — 684 lines)*
### 35. AGENT_STATE.DB — 4 Internal Databases
- **memory** — id → MemoryEntry
- **sessions** — session_id → SessionContext
- **state** — generic KV
- **tags** — tag:name → list of memory IDs
### 36. 6-Type Memory System
| Type | Expiry |
|------|--------|
| Preference | Never |
| Fact | Never |
| Instruction | Never |
| Context | Configurable |
| Working | Session-bound |
| Pinned | Never |
### 37. Session Context Continuity
Parent session linking, active project, files modified, total complexity, summary.
### 38. Tag-Based Memory Index — retrieve memories by category tag
---
## XII. DISPATCH — Unified Routing Protocol
*(dispatch.rs — 506 lines)*
### 39. Single Entry Point for All Tool Calls
- `dispatch::call()` — converges stdio, HTTP, mesh, transformer, pipeline sources
- Gate called at entry before execution (`gate::process()`)
- Execution inside `handle_tool_call` — pure routing, no gate logic
### 40. Source Classification
- **Stdio** — blocking listener lock (local user)
- **Http/Mesh/Transformer/Pipeline** — timed try_lock with SERVER_BUSY fallback
### 41. Listener Pattern
Layers register as `DispatchListener`. Dispatch fires `on_response()` with tool name, source, status, and duration. Dispatch never imports listeners — loose coupling.
### 42. FLINT Interception
- Pre-execution brain query: `process_request()` queries brain for relevant context
- Post-execution store/enrich: `process_result()` stores results, triggers auto-train
---
## XIII. FLINT MEMORY ROUTER
*(flint_memory.rs — 1,065 lines)*
### 43. Background Memory Thread
Always running, no user session required:
- GateTrainingCollector → FLINT scores → route → brain_store()
- Watches knowledge/ drop folder → auto-index
### 44. Tiered Memory Promotion (MB-FT)
24hr → 7day → pinned. FLINT scores Working memories by relevance × access_count.
Top 20% promoted. >50% active in window → promote all + touches.
### 45. Build Anchor Context (R3-04 to R3-07)
- Brain-assisted source context for Write/Edit
- Read tracking (R3-06): first read = full passthrough, subsequent = compressed
- Agent Read Interface: blocks brain WRITE from Stdio
### 46. Auto-Train Engine
Every hour: expire → Working→Fact → Fact→Pinned → auto-train if 16+ tlog signals.
---
## XIV. TENSOR ENGINE — Pure Rust Math
*(tensor.rs — 772 lines)*
### 47. N-Dimensional Tensor (f32)
Row-major (C-order) storage with shape metadata. Supports arbitrary dimensions.
### 48. Operations Suite
Add, subtract, multiply, divide, matmul, reshape, permute, softmax, relu, layer_norm, embedding.
### 49. ARM NEON SIMD Acceleration
With scalar fallback for non-ARM platforms.
### 50. Backward Pass Support
Gradient computation for all operations — enables training.
---
## XV. TOKENIZER — BPE Vocabulary
*(tokenizer.rs — 381 lines)*
### 51. Byte-Pair Encoding Tokenizer
Pure Rust, no external dependencies. Trains on SPF corpus (brain data, source code, rules).
### 52. Special Tokens
[PAD], [BOS], [EOS], [UNK], [TOOL], [GATE], [USER], [SPF], [ALLOWED], [BLOCKED]
### 53. Vocabulary Size: 8,192 tokens (SPF Writer default)
---
## XVI. TRANSFORMER MODEL
*(attention.rs — 517 lines, ffn.rs — 277 lines, encoder.rs — 404 lines, decoder.rs — 508 lines)*
### 54. Encoder-Decoder Architecture
- d_model=256, n_heads=8, n_layers=6, d_ff=1024
- ~5M parameters total
- Configurable: Writer (tool execution) vs Researcher (chat/analysis)
### 55. Multi-Head Self-Attention
Scaled dot-product attention with KV cache for generation. Supports masking for causal mode.
### 56. Position-wise Feed-Forward Network
2-layer MLP: linear → ReLU → linear. Expansion ratio 4×.
### 57. Encoder Stack
N layers, each: self-attention → add&norm → FFN → add&norm.
### 58. Decoder Stack
N layers, each: masked self-attention → add&norm → cross-attention → add&norm → FFN → add&norm.
Cached decoder layers for O(1) autoregressive generation.
---
## XVII. FULL TRANSFORMER MODEL
*(transformer.rs — 552 lines)*
### 59. Complete Model Assembly
Encoder + decoder tied together with embedding projection and token output head.
### 60. Forward Cache
Cached activations from forward pass for backward pass (train-on-copy pattern).
### 61. Two Configurations
- **Writer**: tool selection, gate prediction, task execution
- **Researcher**: conversational analysis, question answering
Both share identical architecture — only training data differs.
---
## XVIII. TRAINING ENGINE
*(train.rs — 1,064 lines)*
### 62. Backward Passes for All Layers
Cross-entropy loss → decoder backward → encoder backward → embedding backward.
### 63. Attention Backpropagation
5-step gradient flow: dV → dP → dS (softmax Jacobian) → dQ → dK.
### 64. AdamW Optimizer
Decoupled weight decay (Loshchilov & Hutter 2017). Per-parameter adaptive moments.
### 65. LayerNorm Backward
Gradient computation through normalization layer.
### 66. Train-on-Copy Pattern
Atomic weight merge — inference never blocked by training.
---
## XIX. ONLINE LEARNING ENGINE
*(learning.rs — 645 lines)*
### 67. Elastic Weight Consolidation (EWC)
Prevents catastrophic forgetting by penalizing changes to important weights. Fisher Information Matrix persisted to LMDB. Lambda = 0.4.
### 68. Experience Replay Buffer
10,000 slots (50/50 new:samples mix). False-positive examples locked — never evicted.
### 69. Gate-as-Teacher
Every gate approve/deny is a supervised training label. FLINT learns to predict gate decisions.
---
## XX. GATE TRAINING BRIDGE
*(gate_training.rs — 1,302 lines)*
### 70. Training Signal Collection
Every gate decision captured as TrainingSignal with context, severity, and result.
### 71. Confusion Matrix
Tracks TP/TN/FP/FN rates. Persisted to LMDB. Used for alignment scoring.
### 72. False Positive Handling
- FP detection + reporting
- Severity-weighted signals (4× FP, 6× repeated FP)
- FP-locked buffer slots — never evicted
- Result-based FP candidate flagging
### 73. Evil Detection Layer (4-Tier Moral Framework)
Based on MORAL_FRAMEWORK.txt:
- Tier 1: Deception (self-serving lies)
- Tier 2: Exploitation (power vs powerless)
- Tier 3: Destruction (deliberate harm)
- Tier 4: Systemic Evil (corruption of trust)
Evil signals receive 4-8× training weight amplification.
### 74. Sequence Context
Last N tool calls tracked for pattern detection — identifies multi-step attack chains.
---
## XXI. FLINT TRANSFORMER MCP TOOLS
*(transformer_tools.rs — 1,052 lines)*
### 75. Transformer State Management
Singleton `TransformerState` — model loading, checkpoint save/load, inference, training.
### 76. Checkpoint System (CP-1/CP-2)
- Save after training batch
- Load on startup — confirmed persistent (step ≠ 0 on boot)
- LIVE/MODELS/writer_v1.spfc
### 77. Training Tools
- `spf_flint_train_evil` — mark tool call as evil/harmful
- `spf_flint_train_good` — mark tool call as good/safe
- Auto-train: 16+ tlog signals OR 1hr interval
### 78. FLINT Status & Metrics
Query current step, loss, alignment, learning rate, EWC lambda, replay buffer occupancy.
---
## XXII. FRAMING PROTOCOL — Persistent Streams
*(framing.rs — 498 lines)*
### 79. Length-Prefixed Message Framing
Replaces one-shot read-to-end pattern. Enables persistent bidirectional mesh streams.
### 80. 8 Stream Types
| Type | Hex | Purpose |
|------|-----|---------|
| ToolRpc | 0x01 | Tool calls (JSON-RPC) |
| ChatText | 0x02 | Text chat |
| VoiceAudio | 0x03 | Opus voice frames |
| PipelineTask | 0x04 | Orchestrator → worker |
| PipelineResult | 0x05 | Worker → orchestrator |
| BrainSync | 0x06 | Memory synchronization |
| WeightSync | 0x07 | Model weight sharing |
| Control | 0x08 | Session control |
### 81. Frame Format
`[1-byte type][4-byte length BE][payload bytes]`. Max frame: 10MB.
Legacy detection: first byte '{' = old JSON-RPC, 0x01-0x08 = framed.
---
## XXIII. MESH NETWORK TRANSPORT
*(mesh.rs — 1,171 lines)*
### 82. P2P QUIC Mesh via iroh
Ed25519 identity doubles as iroh EndpointId. Inbound: peer connects → JSON-RPC → dispatch::call(Source::Mesh). Outbound: tool call → QUIC stream → peer's ALPN.
### 83. Persistent Bidirectional Streams
`call_peer_stream()` for framed persistent connections. `stream_router()` dispatches by StreamType.
### 84. Discovery Modes
- **auto** — mDNS (LAN) + Pkarr DHT (internet)
- **local** — mDNS only
- **manual** — explicit trust groups only
### 85. Trust Model — Default-deny
Only peers in `groups/*.keys` are accepted. No anonymous connections.
### 86. Identity Integration
Ed25519 signing key → iroh SecretKey via direct Curve25519 byte mapping.
---
## XXIV. IDENTITY — Cryptographic Keys
*(identity.rs — 435 lines)*
### 87. Ed25519 Key Pair Management
Unique identity generated on first boot. Files:
- `identity.key` — private key (hex, 64 chars)
- `identity.pub` — public key (hex, 64 chars)
- `identity.seal` — filesystem-bound clone detection
### 88. Clone Detection
If filesystem seal breaks → archive old identity, generate new, preserve settings.
### 89. Trusted Peer Management
- `groups/*.keys` — one public key per line (simple trust list)
- `groups/*.json` — peer info with addresses, name, role
### 90. Boot Integrity Check (SEC-3)
Verifies .mcp.json routing, scans for rogue agent configs.
---
## XXV. PIPELINE — Batch Task Execution
*(pipeline.rs — 1,107 lines)*
### 91. Pipeline Protocol
Batch submit, chain execution, backpressure handling.
### 92. PipelineTask Type
- task_id, tool, arguments, mode (Batch/Chain)
- Optional chain_next — output of one task feeds next
- chain_pipe_field — which output field to pipe to next input
### 93. PipelineResult Type
- task_id, success (bool), output, error, duration_ms
- Links back to original task for audit trail
### 94. Chain Execution
Tasks linked via `chain_next`. Sequential — result of current feeds input of next.
`build_chain()` from task list → linked structure.
### 95. Worker Dispatch
Tasks routed to mesh workers via StreamType 0x04/0x05.
Results processed via `dispatch::call(Source::Pipeline)`.
### 96. 14 Unit Tests
Chain building, batch processing, result recording, backpressure simulation.
---
## XXVI. NETWORK POOL
*(network.rs — 526 lines)*
### 97. Full Pool Management
PoolState, WorkerStatus, ProofOfWork types. Roles: NetAdmin, Worker, Thinker.
### 98. NetAdmin Orchestrator
Distributes tasks to idle workers over QUIC mesh. Collects results. Logs proof of work via `session.record_manifest_detailed()`.
### 99. Worker Lifecycle
Join pool → accept tasks → execute → return results → proof of work receipt.
---
## XXVII. WORKER MODE
*(worker.rs — 494 lines)*
### 100. Headless Inference Node
Runs transformer without user session:
1. Load config + checkpoint
2. Start mesh (accept pipeline streams)
3. Start HTTP (monitoring)
4. Loop: receive task → execute → return
5. Capture training signals
### 101. CLI: `spf-smart-gate worker [--role writer|researcher]`
---
## XXVIII. ORCHESTRATOR STATE
*(orchestrator.rs — 225 lines)*
### 102. Role Management
AgentRole enum: NetAdmin, Worker, Thinker, Orchestrator. Dynamic role switching.
### 103. Pool State Tracking
Worker registration, idle/busy counts, task assignment tracking.
---
## XXIX. CHAT ENGINE
*(chat.rs — 1,286 lines)*
### 104. Text Chat Protocol
Messages over StreamType::ChatText (0x02) + local LMDB storage.
### 105. Message Format
id, from, to, text, timestamp, conversation_id, MessageType (UserText/AgentResponse/System/ToolResult).
### 106. Conversation Tracking
- Context window (last N messages as transformer input)
- Room-based routing (1:1 and group)
- LMDB persistence with "chat:" key prefix
### 107. FLINT Chat Participation (Block QQ)
- Rate-limited auto-response (1 per 5 seconds per peer)
- Decoder-only/causal mode for conversational responses
### 108. MCP Tools
- `spf_chat_send` — send message to peer/conversation
- `spf_chat_history` — retrieve message history with limit
- `spf_chat_rooms` — list all active conversations
---
## XXX. VOICE PIPELINE
*(voice.rs — 2,071 lines)*
### 109. Architecture
AudioInput → encode → VoiceFrame → mesh (0x03) → decode → AudioOutput.
TTS: text → speech → mesh. STT: mesh → text → ChatEngine.
### 110. TTS: espeak-ng FFI In-Process
Zero external process. FFI link to libespeak-ng.so. Light mode = espeak only. Rich mode = Piper ONNX (voice-tts feature).
### 111. Opus Codec
Static linkage (libopus.a). No runtime C deps beyond espeak-ng.
### 112. Voice Mode Selection
- Agent TTS: Light (espeak-ng FFI) or Rich (Piper ONNX)
- Agent STT: Light (text input) or Rich (Whisper via candle)
- Peer audio quality: Light (12kHz) or Rich (24kHz)
### 113. Call Management
Outgoing/incoming calls, P2P P2P call protocol (ring/accept/reject/end/status).
### 114. Voice Team Channels
Group audio channels for team communication.
### 115. MCP Tools
- `spf_voice_call` — start/accept/reject/end voice calls
- `spf_voice_mode` — start/stop/speak/listen/pipe audio
- `spf_voice_team` — create/join/leave team channels
---
## XXXI. REVERSE PROXY BROWSER
*(browser.rs — 934 lines)*
### 116. Architecture
1. Agent calls `spf_web_navigate(url)`
2. Browser opens `http://127.0.0.1:PORT/proxy?url=TARGET`
3. ProxyEngine fetches page, injects control JS, serves it
4. Injected JS opens WebSocket to /ws/browser
5. Agent sends commands (click/fill/eval) through BrowserSession
6. JS executes, returns results via same path
### 117. ProxyEngine
Fetches external pages, injects control JavaScript, strips CSP headers.
Sets `<base href>` for relative URL resolution — no URL rewriting needed.
### 118. BrowserSession
Manages WS channel, command queue, timeout handling.
Commands: click, fill, eval, navigate, screenshot, design, page.
### 119. SSRF Protection
All target URLs validated through `web::validate_url()` before proxying.
### 120. MCP Tools
- `spf_web_connect` — initialize browser engine
- `spf_web_navigate` — navigate to URL
- `spf_web_page` — structured page overview
- `spf_web_click` — click element by CSS selector
- `spf_web_select` — query elements by CSS selector
- `spf_web_fill` — type text into form field
- `spf_web_eval` — execute JavaScript expression
- `spf_web_screenshot` — capture screenshot
- `spf_web_design` — extract design brief
---
## XXXII. WEB ACCESS
*(web.rs — 469 lines)*
### 121. Dual Search Backend
- Brave Search API (if BRAVE_API_KEY set)
- DuckDuckGo HTML fallback
### 122. Web Fetch — HTML-to-Text Conversion
Fetches URL, strips HTML, returns clean text. 30s timeout.
### 123. File Download
Gated through SPF — destination path validated.
### 124. HTTP API Client
Full REST: GET/POST/PUT/DELETE/PATCH. Custom headers and body.
---
## XIII. BRAIN LOCAL — In-Process Vector Search
*(brain_local.rs — 344 lines)*
### 125. stoneshell-brain Integration
Embeds MiniLM model (all-MiniLM-L6-v2) for local vector search.
Functions: search, recall, store, index, list, status.
### 126. Retry Logic (RC-1)
3 attempts with 2s delay on brain load failure.
---
## XXXIV. CHECKPOINT SYSTEM
*(checkpoint.rs — 483 lines)*
### 127. Binary Checkpoint Format
`[SPFC magic:4][version:4][num_tensors:4][headers][data]`
### 128. Delta Format for Mesh Weight Sharing
Sends only changed weights between nodes — bandwidth efficient.
### 129. Versioned Checkpoints
Rollback safety — format version tracked.
---
## XXXV. UTF-8 SAFETY UTILITIES
*(utf8_safe.rs — 101 lines)*
### 130. Lossy Text Conversion
Binary-safe UTF-8 handling. Binary files get fingerprint info, not corrupt text.
---
### XXXVI. INTEGRATION TESTS
*(integration_tests.rs — 763 lines)*
### 131. Full Build Verification
Tests for gate pipeline, complexity calculation, dispatch routing, mesh framing, pipeline chains, transformer inference, checkpoint round-trip.
---
## XXXVII. HOOK SYSTEM — 31 Shell Hooks
*(31 files in hooks/)*
### 132. Native Tool Blocking Hooks (9 hooks)
Exit code 1, hard-block: Read, Edit, Write, Bash, Glob, Grep, WebFetch, WebSearch, NotebookEdit.
### 133. MCP Tool Tracking Hooks (20 hooks)
Exit code 0, log to spf.log: all 83 MCP tools grouped by category.
### 134. Lifecycle Hooks
- session-start — resets state, injects SPF formula
- session-end — handoff note, brain checkpoint
- stop-check — saves state, prevents infinite loop
- user-prompt — prompt complexity scoring
- post-action — updates STATUS.txt, triggers brain checkpoint
- post-failure — logs to failures.log
---
## XXXVIII. STATE AND OBSERVABILITY
### 135. STATUS.txt — Memory Triad System 2
Auto-updated after every tool action. Current state, files read/written, Build Anchor ratio.
### 136. spf.log — Full Action Log
Timestamped log with rotation (0.5GB max, 5,000 lines, one backup).
### 137. failures.log — Dedicated Failure Log
Separate file for failed tool calls.
### 138. Brain Checkpoint on Writes
post-action.sh auto-stores to brain on Write/Edit. Async, tagged: spf, checkpoint, tool_name.
### 139. LEARNING PIPELINE — ACTIVE
- PRE: init_brain() + index_knowledge_docs() + index_spf_sources()
- DURING (30s loop): GateTrainingCollector → FLINT scores → route_signals → brain_store()
- AFTER (1hr loop): expire → Working→Fact → Fact→Pinned → auto-train (16+ tlog or 1hr)
---
## XXXIX. BRAIN STATE (verified 2026-04-05)
### 140. Collections
| Collection | Docs | Purpose |
|------------|------|---------|
| spf_source | 465 | Source code chunks |
| flint_training | 2,013 | Training signals |
| flint_results | 448 | Gate decisions |
| default | 392 | General knowledge |
| flint_knowledge | 49 | FLINT facts |
| session_state | 4 | Current session |
| flint_episodic | 2 | Episodic memory |
Total: 3,371 docs | Storage: 31.41 MB | Model: all-MiniLM-L6-v2
---
## XL. PERMISSIONS AND ACCESS CONTROL
### 141. 3-Layer Defense
| Layer | Mechanism | Bypass? |
|-------|-----------|---------|
| 1. Permissions Deny | settings.json denies native tools | Claude enforced |
| 2. Hook Blocking | pre hooks exit 1 on native tools | Cannot bypass |
| 3. Gate Validation | Compiled Rust allowlist + path blocking | Compiled binary |
### 142. Self-Protection (Dynamic from spf_root)
- SPFsmartGATE/src/ — own source code
- Cargo.toml, Cargo.lock — dependencies
- LIVE/SPF_FS/blobs/ — vfs blob storage
- Home/.claude/ — Claude settings
- System dirs: /etc, /system, /usr, PREFIX
### 143. Config Get/Set Removed from MCP
Config changes are CLI-only. AI cannot modify its own config or invoke the gate directly.
---
## XLI. CLI COMMANDS
### 144. Binary CLI
- `spf-smart-gate serve` — MCP stdio server
- `spf-smart-gate status` — Gateway status
- `spf-smart-gate session` — Full session state
- `spf-smart-gate reset` — Fresh session
- `spf-smart-gate worker [--role writer|researcher]` — Headless worker mode
- `spf-smart-gate init-config` — Initialize CONFIG.DB
- `spf-smart-gate refresh-paths` — Update CONFIG.DB paths (--dry-run)
- `spf-smart-gate fs-import/fs-export` — LMDB ↔ Device
- `spf-smart-gate config-import/config-export` — JSON ↔ CONFIG.DB
- `spf-smart-gate gate` — One-shot gate check
- `spf-smart-gate calculate` — Complexity calc only
---
## XLII. ARCHITECTURE PROPERTIES
### 145. Multi-Database LMDB Architecture
| # | Database | Location | Purpose |
|---|----------|----------|---------|
| 1 | SPF_FS.DB | LIVE/SPF_FS/ | Virtual filesystem + blobs |
| 2 | CONFIG.DB | LIVE/CONFIG/ | Path rules, tiers, formula |
| 3 | SESSION.DB | LIVE/SESSION/ | Session state persistence |
| 4 | PROJECTS.DB | LIVE/PROJECTS/ | Project registry |
| 5 | TMP.DB | LIVE/TMP/ | Trust, access logs, active project |
| 6 | AGENT_STATE.DB | LIVE/LMDB5/ | Memory, sessions, chat |
| 7 | BRAIN | LIVE/TMP/stoneshell-brain/ | Vector search (chromadb) |
### 146. Zero-Dependency Path Resolution
`spf_root()` walks up from binary location for Cargo.toml → SPF_ROOT env → HOME/SPFsmartGATE. Cached via OnceLock.
### 147. MCP Protocol v2024-11-05
JSON-RPC 2.0 over stdio. Compatible with any MCP client. 83 agent tools (8 user-only tools excluded).
### 148. Cross-Platform Portable
Android ARM64, Linux x86_64/ARM64, macOS ARM/Intel. Zero hardcoded paths. All resolved from binary location.
### 149. Multi-Agent Network
- Ed25519 identity (CL-1: self-test at boot)
- QUIC mesh transport (iroh)
- Qwen tool prefix aliasing (QW-1/QW-2)
- API Session Layer — full duplex agent-to-agent encrypted sessions (PP-2)
---
## XLIII. FLINT AI STATUS
### 150. Training Status (live, 2026-04-05)
- Step: 116 | Batches completed: 6
- Gate alignment: 81.25% | Average loss: 0.648
- EWC lambda: 0.4 | Learning rate: 1e-4
- Online learning: ON | Replay buffer: 10,000 slots
- Checkpoint interval: 1,000 steps
- Checkpoint: LIVE/MODELS/writer_v1.spfc
### 151. Model Architecture
- d_model: 256 | n_heads: 8 | n_layers: 6 | d_ff: 1,024
- Vocab: 8,192 | Params: ~5M
- Writer + Researcher configs
---
## XLIV. SECURITY EXCEPTIONS (Known Gaps)
| Gap | Description | Remediation |
|-----|-------------|-------------|
| GAP-1 | ChatText → handle_mesh_chat() — no gate routing | SEC-CHAT pending |
| GAP-2 | VoiceAudio — intentional, audio frames only | Accepted |
| GAP-3 | /proxy + /proxy/asset — SSRF only, bypasses dispatch | WB-2 pending |
| GAP-4 | /ws/browser — no gate/FLINT | SEC-WS pending |
---
*Generated from source code analysis — verified 2026-04-05*
*8,245 → 18,500+ lines | 88 → 150+ features | 31 → 42 modules | 6 → 7 databases*
*New since February 2026: FLINT Transformer AI, Mesh Network, Chat, Voice, Browser, Pipeline, Network Pool, Channel Hub*