File size: 41,358 Bytes
1269259 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 | // SPF Smart Gateway - FLINT Transformer MCP Tool Handlers
// Copyright 2026 Joseph Stone - All Rights Reserved
//
// FLINT — Focused Learning Intelligence for Network Threats
// Named: Stone + Flint = Fire
//
// BLOCK K — Separate module for transformer tool logic.
// mcp.rs gets minimal forwarding match arms (5 tools).
// gate.rs gets 5 tools added to allowlist.
//
// All handlers follow the mcp.rs pattern:
// fn handle_*(args, state) -> Value
// Returns json!({"type": "text", "text": "..."}) matching MCP protocol.
//
// Depends on: Block E (transformer.rs, checkpoint.rs), Block I (TransformerConfig)
use serde_json::{json, Value};
use std::sync::{Arc, RwLock};
use crate::config::TransformerConfig;
use crate::agent_state::AgentStateDb;
use crate::paths::spf_root;
// ============================================================================
// FLINT IDENTITY
// ============================================================================
/// FLINT's name — used in all tool responses
pub const FLINT_NAME: &str = "FLINT";
pub const FLINT_VERSION: &str = "0.1.0";
// ============================================================================
// TRANSFORMER STATE — held in ServerState behind RwLock
// ============================================================================
/// Runtime state for the transformer, wrapped in RwLock in ServerState.
/// Multiple readers (inference) can proceed simultaneously.
/// Single writer (training) acquires exclusive lock during weight update.
pub struct TransformerState {
/// The transformer model (Block E)
pub model: crate::transformer::SPFTransformer,
/// Configuration
pub config: TransformerConfig,
/// Current training step
pub training_step: u64,
/// Last checkpoint path
pub last_checkpoint: String,
/// Whether model is currently training
pub is_training: bool,
/// Cumulative loss (for metrics)
pub total_loss: f64,
/// Number of training batches completed
pub batches_completed: u64,
/// Gate alignment score (0.0 - 1.0)
pub gate_alignment: f64,
/// Model role: "writer" or "researcher"
pub role: String,
/// Reference to GateTrainingCollector for draining training signals.
/// Shared with ServerState.listeners via Arc — same buffer, same signals.
pub collector: Option<std::sync::Arc<crate::gate_training::GateTrainingCollector>>,
/// AdamW optimizer
pub optimizer: Option<crate::train::AdamW>,
/// Block PP: Whether FLINT auto-responds to chat (default: off)
pub chat_enabled: bool,
}
impl TransformerState {
/// Create a new TransformerState from config
pub fn from_config(config: &TransformerConfig, role: &str) -> Self {
let model_config = crate::transformer::TransformerModelConfig {
vocab_size: config.vocab_size,
d_model: config.d_model,
n_heads: config.n_heads,
n_layers: config.n_layers,
d_ff: config.d_ff,
max_seq_len: config.max_seq_len,
ln_eps: 1e-5,
};
let model = crate::transformer::SPFTransformer::new(model_config, 42);
// FL-9: LearningController removed — training driven by handle_train() + LMDB.
// Only AdamW optimizer needed (used directly in handle_train).
let optimizer = if config.online_learning {
let param_sizes: Vec<usize> = model.weights().iter().map(|w| w.numel()).collect();
Some(crate::train::AdamW::new(
crate::train::AdamWConfig { lr: config.learning_rate as f32, ..Default::default() },
¶m_sizes,
))
} else {
None
};
Self {
model,
config: config.clone(),
training_step: 0,
last_checkpoint: String::new(),
is_training: false,
total_loss: 0.0,
batches_completed: 0,
gate_alignment: 0.0,
role: role.to_string(),
collector: None,
optimizer,
chat_enabled: false, // Block PP: off by default
}
}
}
// ============================================================================
// TOOL DEFINITIONS — called by mcp.rs tool_definitions()
// ============================================================================
/// Returns the 5 transformer tool definitions for MCP registration.
/// Called from mcp.rs tool_definitions() to avoid bloating that file.
pub fn tool_definitions() -> Vec<Value> {
vec![
json!({
"name": "spf_transformer_status",
"description": "Get FLINT transformer status (loaded, params, checkpoint, role)",
"inputSchema": {
"type": "object",
"properties": {},
"required": []
}
}),
json!({
"name": "spf_transformer_infer",
"description": "Run FLINT inference: prompt → response. Returns generated tokens.",
"inputSchema": {
"type": "object",
"properties": {
"prompt": {"type": "string", "description": "Input text prompt"},
"max_tokens": {"type": "integer", "description": "Max tokens to generate (default: 64)"},
"temperature": {"type": "number", "description": "Sampling temperature (default: from config)"}
},
"required": ["prompt"]
}
}),
json!({
"name": "spf_transformer_chat",
"description": "Chat with FLINT (conversation tracking, multi-turn)",
"inputSchema": {
"type": "object",
"properties": {
"message": {"type": "string", "description": "User message"},
"conversation_id": {"type": "string", "description": "Conversation ID (optional, auto-generated if omitted)"}
},
"required": ["message"]
}
}),
json!({
"name": "spf_transformer_train",
"description": "Trigger FLINT manual training batch from accumulated gate signals",
"inputSchema": {
"type": "object",
"properties": {
"batch_size": {"type": "integer", "description": "Override batch size (optional)"}
},
"required": []
}
}),
json!({
"name": "spf_transformer_metrics",
"description": "Get FLINT learning metrics (loss, accuracy, gate alignment, training step)",
"inputSchema": {
"type": "object",
"properties": {},
"required": []
}
}),
// FL-8: Evil/Good training input tools
json!({
"name": "spf_flint_train_evil",
"description": "Mark a tool call as evil/harmful. Creates negative training signal for FLINT.",
"inputSchema": {
"type": "object",
"properties": {
"tool": {"type": "string", "description": "Tool name that was evil"},
"reason": {"type": "string", "description": "Why this call was evil (optional)"}
},
"required": ["tool"]
}
}),
json!({
"name": "spf_flint_train_good",
"description": "Mark a tool call as good/safe. Creates positive training signal for FLINT.",
"inputSchema": {
"type": "object",
"properties": {
"tool": {"type": "string", "description": "Tool name that was good"},
"reason": {"type": "string", "description": "Why this call was good (optional)"}
},
"required": ["tool"]
}
}),
]
}
// ============================================================================
// TOOL HANDLERS — called from mcp.rs match arms
// ============================================================================
/// Handle spf_transformer_status
/// Returns: model loaded status, param count, checkpoint info, role, config summary
pub fn handle_status(
transformer: &Option<Arc<RwLock<TransformerState>>>,
config: &TransformerConfig,
) -> Value {
match transformer {
None => {
json!({"type": "text", "text": format!(
"FLINT: NOT LOADED\nVersion: {}\nEnabled: {}\nConfig: d_model={}, n_heads={}, n_layers={}, vocab={}\nEstimated params: {}\nEstimated memory: {}MB",
FLINT_VERSION,
config.enabled,
config.d_model, config.n_heads, config.n_layers, config.vocab_size,
config.estimated_params(),
config.estimated_memory_bytes() / 1_000_000
)})
}
Some(state_lock) => {
let state = state_lock.read().unwrap();
let avg_loss = if state.batches_completed > 0 {
state.total_loss / state.batches_completed as f64
} else {
0.0
};
json!({"type": "text", "text": format!(
"FLINT: LOADED v{} ({})\n\
Role: {}\n\
Training step: {}\n\
Batches completed: {}\n\
Avg loss: {:.6}\n\
Gate alignment: {:.2}%\n\
Currently training: {}\n\
Last checkpoint: {}\n\
Config: d_model={}, n_heads={}, n_layers={}, vocab={}\n\
Online learning: {}\n\
EWC lambda: {}",
FLINT_VERSION,
if state.is_training { "training" } else { "idle" },
state.role,
state.training_step,
state.batches_completed,
avg_loss,
state.gate_alignment * 100.0,
state.is_training,
if state.last_checkpoint.is_empty() { "none" } else { &state.last_checkpoint },
config.d_model, config.n_heads, config.n_layers, config.vocab_size,
config.online_learning,
config.ewc_lambda,
)})
}
}
}
/// Handle spf_transformer_infer
/// Runs forward pass on prompt, generates tokens autoregressively.
pub fn handle_infer(
transformer: &Option<Arc<RwLock<TransformerState>>>,
args: &Value,
config: &TransformerConfig,
tokenizer_path: &str,
) -> Value {
let state_lock = match transformer {
Some(s) => s,
None => return json!({"type": "text", "text": "ERROR: FLINT not loaded. Enable in transformer.json."}),
};
let prompt = match args.get("prompt").and_then(|v| v.as_str()) {
Some(p) => p,
None => return json!({"type": "text", "text": "ERROR: 'prompt' parameter required"}),
};
let max_tokens = args.get("max_tokens")
.and_then(|v| v.as_u64())
.unwrap_or(64) as usize;
let temperature = args.get("temperature")
.and_then(|v| v.as_f64())
.unwrap_or(config.temperature);
// Load tokenizer
let tokenizer = match crate::tokenizer::Tokenizer::load(tokenizer_path) {
Ok(t) => t,
Err(e) => return json!({"type": "text", "text": format!("ERROR: Failed to load tokenizer: {}", e)}),
};
// Tokenize input
let input_ids = tokenizer.encode(prompt);
if input_ids.is_empty() {
return json!({"type": "text", "text": "ERROR: Empty input after tokenization"});
}
// Read lock for inference (concurrent with other readers)
let state = state_lock.read().unwrap();
// Generate
let output_ids = match state.model.generate(&input_ids, max_tokens, temperature as f32, 42) {
Ok(ids) => ids,
Err(e) => return json!({"type": "text", "text": format!("ERROR: Generation failed: {}", e)}),
};
// Decode
let output_text = tokenizer.decode(&output_ids);
json!({"type": "text", "text": format!(
"[FLINT] Input: {} ({} tokens)\nOutput: {} ({} tokens)\nTemperature: {:.2}",
prompt, input_ids.len(),
output_text, output_ids.len(),
temperature
)})
}
/// Handle spf_transformer_chat
/// Multi-turn conversation with context window management.
pub fn handle_chat(
transformer: &Option<Arc<RwLock<TransformerState>>>,
args: &Value,
config: &TransformerConfig,
tokenizer_path: &str,
) -> Value {
let state_lock = match transformer {
Some(s) => s,
None => return json!({"type": "text", "text": "ERROR: FLINT not loaded. Enable in transformer.json."}),
};
// Block PP: Handle chat toggle (enabled: true/false)
if let Some(enabled) = args.get("enabled").and_then(|v| v.as_bool()) {
let mut state = state_lock.write().unwrap();
state.chat_enabled = enabled;
return json!({"type": "text", "text": format!(
"FLINT chat {}", if enabled { "ON — will respond to messages" } else { "OFF — silent mode" }
)});
}
// Block PP: Check if chat is enabled before processing
{
let state = state_lock.read().unwrap();
if !state.chat_enabled {
return json!({"type": "text", "text":
"FLINT chat is OFF. Use spf_transformer_chat with {\"enabled\": true} to activate."
});
}
}
let message = match args.get("message").and_then(|v| v.as_str()) {
Some(m) => m,
None => return json!({"type": "text", "text": "ERROR: 'message' parameter required"}),
};
let conversation_id = args.get("conversation_id")
.and_then(|v| v.as_str())
.unwrap_or("default");
// Load tokenizer
let tokenizer = match crate::tokenizer::Tokenizer::load(tokenizer_path) {
Ok(t) => t,
Err(e) => return json!({"type": "text", "text": format!("ERROR: Failed to load tokenizer: {}", e)}),
};
// SB-6: Prepend brain context before generation
// Pulls from episodic memory (past Q+A) and knowledge (moral framework, etc.)
let episodic = crate::brain_local::brain_context(message, "flint_episodic", 1000);
let knowledge = crate::brain_local::brain_context(message, "flint_knowledge", 500);
let has_context = !episodic.trim().is_empty() || !knowledge.trim().is_empty();
// Format as chat turn — context prepended when available
let chat_prompt = if has_context {
format!("{}\n{}\n<user> {} <assistant>", knowledge, episodic, message)
} else {
format!("<user> {} <assistant>", message)
};
let input_ids = tokenizer.encode(&chat_prompt);
// Read lock for inference
let state = state_lock.read().unwrap();
// Generate response (capped at reasonable chat length)
let max_response = 128.min(config.max_seq_len.saturating_sub(input_ids.len()));
let output_ids = match state.model.generate(&input_ids, max_response, config.temperature as f32, 42) {
Ok(ids) => ids,
Err(e) => return json!({"type": "text", "text": format!("ERROR: Chat generation failed: {}", e)}),
};
let response_text = tokenizer.decode(&output_ids);
// SB-6: Index Q+A pair into flint_episodic for future context recall
let qa_entry = format!("Q: {}\nA: {}", message, response_text);
let _ = crate::brain_local::brain_store(&qa_entry, conversation_id, "flint_episodic");
json!({"type": "text", "text": format!(
"conversation: {}\nuser: {}\nFLINT: {}",
conversation_id, message, response_text
)})
}
/// Handle spf_transformer_train
/// Triggers a manual training batch from accumulated gate signals.
pub fn handle_train(
transformer: &Option<Arc<RwLock<TransformerState>>>,
args: &Value,
config: &TransformerConfig,
) -> Value {
let state_lock = match transformer {
Some(s) => s,
None => return json!({"type": "text", "text": "ERROR: FLINT not loaded. Enable in transformer.json."}),
};
let batch_size = args.get("batch_size")
.and_then(|v| v.as_u64())
.unwrap_or(config.batch_size as u64) as usize;
// Write lock for training (exclusive access)
let mut state = state_lock.write().unwrap();
if state.is_training {
return json!({"type": "text", "text": "BUSY: FLINT training already in progress. Wait for completion."});
}
state.is_training = true;
// Step 1: Read training signals from agent_state LMDB (FL-2).
// FL-1 persists signals as tlog:{timestamp} keys in route_signals().
// This eliminates the drain race: route_signals() drains the collector
// for brain storage, then persists to LMDB. handle_train() reads from
// LMDB instead of the (already-drained) collector.
let db_path = spf_root().join("LIVE/LMDB5/LMDB5.DB");
let mut signals: Vec<crate::gate_training::TrainingSignal> = Vec::new();
let mut consumed_keys: Vec<String> = Vec::new();
if let Ok(db) = AgentStateDb::open(&db_path) {
if let Ok(keys) = db.list_state_keys() {
let mut tlog_keys: Vec<String> = keys.into_iter()
.filter(|k| k.starts_with("tlog:"))
.collect();
tlog_keys.sort();
for key in &tlog_keys {
if let Ok(Some(json)) = db.get_state(key) {
if let Ok(signal) = serde_json::from_str::<crate::gate_training::TrainingSignal>(&json) {
signals.push(signal);
consumed_keys.push(key.clone());
}
}
}
// Delete consumed tlog entries after successful read
for key in &consumed_keys {
let _ = db.delete_state(key);
}
}
}
// Fallback: if LMDB yielded nothing, try collector drain (backward compat)
if signals.is_empty() {
signals = match &state.collector {
Some(collector) => collector.drain_signals(),
None => Vec::new(),
};
}
if signals.is_empty() {
state.is_training = false;
return json!({"type": "text", "text": "No pending training signals. Gate decisions accumulate signals automatically."});
}
let signal_count = signals.len();
// Step 2: Convert signals to training examples via Block M (learning.rs)
let mut examples: Vec<crate::train::TrainingExample> = signals.iter().map(|signal| {
crate::train::TrainingExample {
input_tokens: crate::learning::signal_to_tokens(signal),
target: crate::train::TrainingTarget::GateDecision(signal.label()),
weight: signal.weight(),
}
}).collect();
// Limit to requested batch_size
examples.truncate(batch_size);
// Step 3: Record previous loss for comparison
let previous_avg_loss = if state.batches_completed > 0 {
(state.total_loss / state.batches_completed as f64) as f32
} else {
f32::MAX
};
// Step 4: Clone weights for training (agentic observer — microsecond read)
let cloned_weights: Vec<crate::tensor::Tensor> = state.model.weights().iter()
.map(|w| crate::tensor::Tensor { data: w.data.clone(), shape: w.shape.clone() })
.collect();
// Release write lock intent — from here, train on clone only
// (In full agentic mode, the lock is released here. In MCP handler mode,
// we still hold the write lock but training is on the clone, not the model.)
// Step 5: Forward + loss on each example using the LIVE model (read-only)
let mut total_loss = 0.0f32;
let mut correct = 0u64;
let mut processed = 0u64;
let mut all_grads: Option<Vec<crate::tensor::Tensor>> = None;
for example in &examples {
let tokens: Vec<u32> = example.input_tokens.iter()
.map(|&t| t as u32)
.collect();
let seq_len = tokens.len().min(config.max_seq_len);
if seq_len == 0 { continue; }
// Forward pass WITH CACHE for backward
match state.model.forward_causal_with_cache(&tokens[..seq_len], 1, seq_len) {
Ok((logits, cache)) => {
// Gate decision score: sigmoid of last-position first logit
let last_offset = (seq_len - 1) * config.vocab_size;
let gate_logit = logits.data.get(last_offset).copied().unwrap_or(0.0);
let prediction = 1.0 / (1.0 + (-gate_logit).exp());
let label = match &example.target {
crate::train::TrainingTarget::GateDecision(l) => l.clamp(0.0, 1.0),
crate::train::TrainingTarget::NextToken(_) => continue,
};
let pred_t = crate::tensor::Tensor::from_data(
vec![prediction], vec![1]
).unwrap();
if let Ok((loss, d_pred)) = crate::train::binary_ce_loss(
&pred_t, &[label], &[example.weight]
) {
total_loss += loss;
let predicted_allow = prediction > 0.5;
let actual_allow = label > 0.5;
if predicted_allow == actual_allow { correct += 1; }
processed += 1;
// Step 6: Composed backward pass
// Build d_logits from d_pred (scatter to full logits shape)
let mut d_logits_data = vec![0.0f32; seq_len * config.vocab_size];
d_logits_data[last_offset] = d_pred.data[0];
if let Ok(d_logits) = crate::tensor::Tensor::from_data(
d_logits_data, vec![1, seq_len, config.vocab_size]
) {
if let Ok(grads) = crate::train::model_backward_causal(
&d_logits, &cache, &state.model,
) {
// Accumulate gradients
match &mut all_grads {
None => all_grads = Some(grads),
Some(acc) => {
for (a, g) in acc.iter_mut().zip(grads.iter()) {
for (av, gv) in a.data.iter_mut().zip(g.data.iter()) {
*av += *gv;
}
}
}
}
}
}
}
}
Err(e) => {
eprintln!("[FLINT-TRAIN] Forward pass error: {}", e);
continue;
}
}
}
// Step 7: Apply optimizer on cloned weights if we have gradients
let avg_loss = if processed > 0 { total_loss / processed as f32 } else { 0.0 };
let alignment = if processed > 0 { correct as f64 / processed as f64 } else { 0.0 };
let mut weights_adopted = false;
// FL-4: Load EWC state for penalty + Fisher update
let ewc_path = spf_root().join("LIVE/MODELS/ewc_state.bin");
let total_params: usize = cloned_weights.iter().map(|t| t.data.len()).sum();
let mut ewc = if ewc_path.exists() {
crate::learning::OnlineEWC::load_from_file(&ewc_path)
.unwrap_or_else(|_| crate::learning::OnlineEWC::new(total_params, config.ewc_lambda as f32))
} else {
crate::learning::OnlineEWC::new(total_params, config.ewc_lambda as f32)
};
if let Some(grads) = all_grads {
// Average gradients over processed examples
let scale = 1.0 / processed.max(1) as f32;
let mut grad_refs: Vec<crate::tensor::Tensor> = grads.iter().map(|g| {
let scaled: Vec<f32> = g.data.iter().map(|&v| v * scale).collect();
crate::tensor::Tensor { data: scaled, shape: g.shape.clone() }
}).collect();
// FL-4: Apply EWC penalty gradients to prevent catastrophic forgetting
if ewc.active {
let flat_weights: Vec<f32> = cloned_weights.iter()
.flat_map(|t| t.data.iter().copied())
.collect();
let (_ewc_loss, ewc_grads) = ewc.penalty(&flat_weights);
let mut offset = 0;
for grad_tensor in grad_refs.iter_mut() {
let n = grad_tensor.data.len();
for i in 0..n {
if offset + i < ewc_grads.len() {
grad_tensor.data[i] += ewc_grads[offset + i];
}
}
offset += n;
}
}
// Apply AdamW step on cloned weights
let mut cloned = cloned_weights;
if let Some(ref mut optimizer) = state.optimizer {
let mut param_refs: Vec<&mut crate::tensor::Tensor> = cloned.iter_mut().collect();
let grad_ref_slice: Vec<&crate::tensor::Tensor> = grad_refs.iter().collect();
let lr = config.learning_rate as f32;
optimizer.step(&mut param_refs, &grad_ref_slice, lr);
// FL-3: Always adopt weights. Gate labels are ground truth —
// loss gating blocked learning of difficult patterns by discarding
// weight updates whenever loss temporarily increased.
let mut model_weights = state.model.weights_mut();
for (mw, cw) in model_weights.iter_mut().zip(cloned.iter()) {
mw.data.copy_from_slice(&cw.data);
}
weights_adopted = true;
eprintln!("[FLINT-TRAIN] Batch loss: {:.6} (prev: {:.6}). Weights adopted.", avg_loss, previous_avg_loss);
// FL-4: Update Fisher matrix with this batch's gradients
let flat_grads: Vec<f32> = grad_refs.iter()
.flat_map(|t| t.data.iter().copied())
.collect();
ewc.update_fisher(&flat_grads);
// Snapshot adopted weights as EWC reference point
let flat_adopted: Vec<f32> = state.model.weights().iter()
.flat_map(|t| t.data.iter().copied())
.collect();
ewc.snapshot_weights(&flat_adopted);
}
}
// FL-4: Persist EWC state after training
if let Err(e) = ewc.save_to_file(&ewc_path) {
eprintln!("[FLINT-TRAIN] EWC save error: {}", e);
} else if ewc.update_count > 0 {
if let Ok(db) = AgentStateDb::open(&db_path) {
let meta = format!(
"{{\"update_count\":{},\"lambda\":{},\"params\":{},\"active\":{}}}",
ewc.update_count, ewc.lambda, total_params, ewc.active
);
let _ = db.set_state("ewc:meta", &meta);
}
}
// Step 9: Update state metrics
state.training_step += 1;
state.batches_completed += 1;
state.total_loss += avg_loss as f64;
state.gate_alignment = alignment;
state.is_training = false;
// CP-1: Persist checkpoint so RC-2 can restore weights + step on restart.
// Only save when weights were actually adopted — skip no-op batches.
if weights_adopted {
let ckpt_dir = spf_root().join("LIVE/MODELS");
let _ = std::fs::create_dir_all(&ckpt_dir);
let ckpt_path = ckpt_dir.join(&config.writer_checkpoint);
let weight_refs = state.model.weights();
match crate::checkpoint::serialize_weights(&weight_refs, "flint_writer", state.training_step) {
Ok(bytes) => match std::fs::write(&ckpt_path, &bytes) {
Ok(()) => {
state.last_checkpoint = ckpt_path.to_string_lossy().to_string();
eprintln!("[FLINT-TRAIN] Checkpoint saved: step={}", state.training_step);
}
Err(e) => eprintln!("[FLINT-TRAIN] Checkpoint write failed: {}", e),
},
Err(e) => eprintln!("[FLINT-TRAIN] Checkpoint serialize failed: {}", e),
}
}
json!({"type": "text", "text": format!(
"FLINT training batch completed\n\
Signals drained: {}\n\
Examples processed: {}/{}\n\
Average loss: {:.6}\n\
Gate alignment: {:.1}%\n\
Weights adopted: {}\n\
Training step: {}\n\
Total batches: {}",
signal_count,
processed,
examples.len(),
avg_loss,
alignment * 100.0,
weights_adopted,
state.training_step,
state.batches_completed,
)})
}
/// Handle spf_transformer_metrics
/// Returns current learning metrics.
pub fn handle_metrics(
transformer: &Option<Arc<RwLock<TransformerState>>>,
config: &TransformerConfig,
) -> Value {
match transformer {
None => {
json!({"type": "text", "text": "FLINT: NOT LOADED\nNo metrics available."})
}
Some(state_lock) => {
let state = state_lock.read().unwrap();
let avg_loss = if state.batches_completed > 0 {
state.total_loss / state.batches_completed as f64
} else {
f64::NAN
};
json!({"type": "text", "text": format!(
"=== FLINT Metrics ===\n\
Version: {}\n\
Role: {}\n\
Training step: {}\n\
Batches completed: {}\n\
Average loss: {:.6}\n\
Gate alignment: {:.2}%\n\
Learning rate: {:.2e}\n\
EWC lambda: {}\n\
Online learning: {}\n\
Replay buffer: {} slots\n\
Last checkpoint: {}",
FLINT_VERSION,
state.role,
state.training_step,
state.batches_completed,
avg_loss,
state.gate_alignment * 100.0,
config.learning_rate,
config.ewc_lambda,
config.online_learning,
config.replay_buffer_size,
if state.last_checkpoint.is_empty() { "none" } else { &state.last_checkpoint },
)})
}
}
}
// ============================================================================
// FL-8: Evil/Good Training Input Tools
// ============================================================================
/// Handle spf_flint_train_evil — user labels a tool call as evil/harmful.
/// Creates a negative training signal (false_positive=true, evil_score=1.0)
/// and stores it in LMDB for the next training batch.
pub fn handle_train_evil(args: &Value) -> Value {
let tool = args.get("tool").and_then(|v| v.as_str()).unwrap_or("");
let reason = args.get("reason").and_then(|v| v.as_str()).unwrap_or("user labeled evil");
if tool.is_empty() {
return json!({"type": "text", "text": "ERROR: 'tool' parameter required. Specify which tool call was evil."});
}
let timestamp = chrono::Utc::now().to_rfc3339();
let signal = crate::gate_training::TrainingSignal {
tool: tool.to_string(),
source: "evil_label".to_string(),
allowed: true,
status: "evil".to_string(),
duration_ms: 0,
timestamp: timestamp.clone(),
user_override: false,
false_positive: true,
recent_call_count: 0,
preceding_tools: Vec::new(),
evil_score: 1.0,
};
// Store in LMDB as tlog entry for handle_train() consumption
let db_path = spf_root().join("LIVE/LMDB5/LMDB5.DB");
if let Ok(db) = AgentStateDb::open(&db_path) {
if let Ok(json) = serde_json::to_string(&signal) {
let tlog_key = format!("tlog:{}", timestamp);
let _ = db.set_state(&tlog_key, &json);
}
}
json!({"type": "text", "text": format!(
"FLINT evil label recorded for '{}': {}. Will be included in next training batch.",
tool, reason
)})
}
/// Handle spf_flint_train_good — user labels a tool call as good/safe.
/// Creates a positive training signal (allowed=true, evil_score=0.0)
/// and stores it in LMDB for the next training batch.
pub fn handle_train_good(args: &Value) -> Value {
let tool = args.get("tool").and_then(|v| v.as_str()).unwrap_or("");
let reason = args.get("reason").and_then(|v| v.as_str()).unwrap_or("user labeled good");
if tool.is_empty() {
return json!({"type": "text", "text": "ERROR: 'tool' parameter required. Specify which tool call was good."});
}
let timestamp = chrono::Utc::now().to_rfc3339();
let signal = crate::gate_training::TrainingSignal {
tool: tool.to_string(),
source: "good_label".to_string(),
allowed: true,
status: "ok".to_string(),
duration_ms: 0,
timestamp: timestamp.clone(),
user_override: false,
false_positive: false,
recent_call_count: 0,
preceding_tools: Vec::new(),
evil_score: 0.0,
};
let db_path = spf_root().join("LIVE/LMDB5/LMDB5.DB");
if let Ok(db) = AgentStateDb::open(&db_path) {
if let Ok(json) = serde_json::to_string(&signal) {
let tlog_key = format!("tlog:{}", timestamp);
let _ = db.set_state(&tlog_key, &json);
}
}
json!({"type": "text", "text": format!(
"FLINT good label recorded for '{}': {}. Will be included in next training batch.",
tool, reason
)})
}
// ============================================================================
// GATE ALLOWLIST ADDITIONS (for gate.rs — document only)
// ============================================================================
//
// Add to gate.rs allowlist at line 197 (after "spf_mesh_call"):
//
// "spf_mesh_status" | "spf_mesh_peers" | "spf_mesh_call" |
// // Transformer tools — Block K
// "spf_transformer_status" | "spf_transformer_infer" |
// "spf_transformer_chat" | "spf_transformer_train" |
// "spf_transformer_metrics"
// => validate::ValidationResult::ok(),
//
// ============================================================================
// MCP.RS ADDITIONS (document only — minimal forwarding)
// ============================================================================
//
// In tool_definitions():
// // ====== TRANSFORMER TOOLS (Block K) ======
// // Tool defs are in transformer_tools::tool_definitions()
// tools.extend(crate::transformer_tools::tool_definitions());
//
// In handle_tool_call() match:
// "spf_transformer_status" => {
// let gate_params = ToolParams { ..Default::default() };
// let decision = gate::process("spf_transformer_status", &gate_params, config, session);
// if !decision.allowed { /* blocked response */ }
// crate::transformer_tools::handle_status(&state.transformer, &state.transformer_config)
// }
// "spf_transformer_infer" => { /* gate check + */ handle_infer(&state.transformer, args, ...) }
// "spf_transformer_chat" => { /* gate check + */ handle_chat(&state.transformer, args, ...) }
// "spf_transformer_train" => { /* gate check + */ handle_train(&state.transformer, args, ...) }
// "spf_transformer_metrics" => { /* gate check + */ handle_metrics(&state.transformer, ...) }
//
// ============================================================================
// TESTS
// ============================================================================
#[cfg(test)]
mod tests {
use super::*;
fn make_test_config() -> TransformerConfig {
TransformerConfig {
enabled: true,
d_model: 64,
n_heads: 4,
n_layers: 2,
vocab_size: 256,
max_seq_len: 64,
d_ff: 256,
learning_rate: 1e-4,
batch_size: 4,
online_learning: true,
ewc_lambda: 0.4,
replay_buffer_size: 100,
temperature: 0.7,
writer_checkpoint: "test_writer.spfc".to_string(),
researcher_checkpoint: "test_researcher.spfc".to_string(),
}
}
#[test]
fn test_flint_identity() {
assert_eq!(FLINT_NAME, "FLINT");
assert_eq!(FLINT_VERSION, "0.1.0");
}
#[test]
fn test_tool_definitions_count() {
let defs = tool_definitions();
assert_eq!(defs.len(), 7, "Should have 5 transformer + 2 FL-8 tools");
}
#[test]
fn test_tool_definitions_names() {
let defs = tool_definitions();
let names: Vec<&str> = defs.iter()
.map(|d| d["name"].as_str().unwrap())
.collect();
assert!(names.contains(&"spf_transformer_status"));
assert!(names.contains(&"spf_transformer_infer"));
assert!(names.contains(&"spf_transformer_chat"));
assert!(names.contains(&"spf_transformer_train"));
assert!(names.contains(&"spf_transformer_metrics"));
}
#[test]
fn test_tool_definitions_mention_flint() {
let defs = tool_definitions();
for def in &defs {
let desc = def["description"].as_str().unwrap();
assert!(desc.contains("FLINT"),
"Tool {} description should mention FLINT", def["name"]);
}
}
#[test]
fn test_tool_definitions_have_schemas() {
let defs = tool_definitions();
for def in &defs {
assert!(def.get("inputSchema").is_some(),
"Tool {} missing inputSchema", def["name"]);
assert_eq!(def["inputSchema"]["type"], "object",
"Tool {} schema should be object", def["name"]);
}
}
#[test]
fn test_status_not_loaded() {
let config = make_test_config();
let result = handle_status(&None, &config);
let text = result["text"].as_str().unwrap();
assert!(text.contains("FLINT: NOT LOADED"));
assert!(text.contains("d_model=64"));
}
#[test]
fn test_status_loaded() {
let config = make_test_config();
let state = TransformerState::from_config(&config, "writer");
let locked = Arc::new(RwLock::new(state));
let result = handle_status(&Some(locked), &config);
let text = result["text"].as_str().unwrap();
assert!(text.contains("FLINT: LOADED"));
assert!(text.contains("writer"));
assert!(text.contains("Training step: 0"));
}
#[test]
fn test_infer_not_loaded() {
let config = make_test_config();
let args = json!({"prompt": "hello"});
let result = handle_infer(&None, &args, &config, "/nonexistent");
let text = result["text"].as_str().unwrap();
assert!(text.contains("FLINT not loaded"));
}
#[test]
fn test_infer_missing_prompt() {
let config = make_test_config();
let state = TransformerState::from_config(&config, "writer");
let locked = Arc::new(RwLock::new(state));
let args = json!({});
let result = handle_infer(&Some(locked), &args, &config, "/nonexistent");
let text = result["text"].as_str().unwrap();
assert!(text.contains("ERROR: 'prompt' parameter required"));
}
#[test]
fn test_train_not_loaded() {
let config = make_test_config();
let args = json!({});
let result = handle_train(&None, &args, &config);
let text = result["text"].as_str().unwrap();
assert!(text.contains("FLINT not loaded"));
}
#[test]
fn test_train_no_collector() {
let config = make_test_config();
let state = TransformerState::from_config(&config, "writer");
let locked = Arc::new(RwLock::new(state));
let args = json!({"batch_size": 8});
let result = handle_train(&Some(locked.clone()), &args, &config);
let text = result["text"].as_str().unwrap();
// No collector connected → error message
assert!(text.contains("No training signal collector"));
}
#[test]
fn test_metrics_not_loaded() {
let config = make_test_config();
let result = handle_metrics(&None, &config);
let text = result["text"].as_str().unwrap();
assert!(text.contains("FLINT: NOT LOADED"));
}
#[test]
fn test_metrics_loaded() {
let config = make_test_config();
let mut state = TransformerState::from_config(&config, "researcher");
state.batches_completed = 10;
state.total_loss = 5.0;
state.gate_alignment = 0.85;
let locked = Arc::new(RwLock::new(state));
let result = handle_metrics(&Some(locked), &config);
let text = result["text"].as_str().unwrap();
assert!(text.contains("FLINT Metrics"));
assert!(text.contains("researcher"));
assert!(text.contains("85.00%"));
assert!(text.contains("Batches completed: 10"));
}
#[test]
fn test_transformer_state_from_config() {
let config = make_test_config();
let state = TransformerState::from_config(&config, "writer");
assert_eq!(state.role, "writer");
assert_eq!(state.training_step, 0);
assert!(!state.is_training);
assert!(state.last_checkpoint.is_empty());
}
#[test]
fn test_chat_not_loaded() {
let config = make_test_config();
let args = json!({"message": "hello"});
let result = handle_chat(&None, &args, &config, "/nonexistent");
let text = result["text"].as_str().unwrap();
assert!(text.contains("FLINT not loaded"));
}
}
|