memd 1.9.0

Local memory CLI for AI agents with hybrid retrieval
Documentation
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
//! Metrics collection and reporting
//!
//! Collects latency breakdown and index statistics for observability.
//! Includes tiered metrics for cache/hot/warm tier performance tracking.

use std::collections::HashMap;
use std::sync::atomic::{AtomicU64, Ordering};
use std::time::{Duration, Instant};

use parking_lot::RwLock;
use serde::{Deserialize, Serialize};

/// Per-query latency breakdown
#[derive(Debug, Clone, Default, Serialize, Deserialize)]
pub struct QueryMetrics {
    /// Time spent generating query embedding (ms)
    pub embed_ms: u64,
    /// Time spent in dense HNSW search (ms)
    pub dense_search_ms: u64,
    /// Time spent fetching chunk data (ms)
    pub fetch_ms: u64,
    /// Total query time (ms)
    pub total_ms: u64,
}

impl QueryMetrics {
    /// Create from timing measurements
    pub fn from_timings(
        embed: Duration,
        dense_search: Duration,
        fetch: Duration,
        total: Duration,
    ) -> Self {
        Self {
            embed_ms: embed.as_millis() as u64,
            dense_search_ms: dense_search.as_millis() as u64,
            fetch_ms: fetch.as_millis() as u64,
            total_ms: total.as_millis() as u64,
        }
    }
}

/// Aggregated latency statistics
#[derive(Debug, Clone, Default, Serialize, Deserialize)]
pub struct LatencyStats {
    /// Number of queries
    pub count: u64,
    /// Average total latency (ms)
    pub avg_total_ms: f64,
    /// Average embed latency (ms)
    pub avg_embed_ms: f64,
    /// Average dense search latency (ms)
    pub avg_dense_search_ms: f64,
    /// P50 total latency (ms)
    pub p50_total_ms: u64,
    /// P90 total latency (ms)
    pub p90_total_ms: u64,
    /// P99 total latency (ms)
    pub p99_total_ms: u64,
}

/// Index statistics
#[derive(Debug, Clone, Default, Serialize, Deserialize)]
pub struct IndexStats {
    /// Number of chunks indexed
    pub chunks_indexed: u64,
    /// Number of embeddings stored
    pub embeddings_count: u64,
    /// Embedding dimension
    pub embedding_dimension: usize,
    /// Total index memory estimate (bytes)
    pub index_memory_bytes: u64,
}

/// Tiered search metrics (aggregated)
#[derive(Debug, Clone, Default, Serialize, Deserialize)]
pub struct TieredMetrics {
    /// Total cache lookups
    pub cache_lookups: u64,
    /// Cache hits
    pub cache_hits: u64,
    /// Cache misses
    pub cache_misses: u64,
    /// Cache hit rate (0.0-1.0)
    pub cache_hit_rate: f32,
    /// Hot tier searches performed
    pub hot_tier_searches: u64,
    /// Queries satisfied by hot tier alone
    pub hot_tier_hits: u64,
    /// Warm tier searches performed
    pub warm_tier_searches: u64,
    /// Chunks promoted to hot tier
    pub promotions: u64,
    /// Chunks demoted from hot tier
    pub demotions: u64,
    /// Average cache lookup latency (ms)
    pub avg_cache_lookup_ms: f64,
    /// Average hot tier search latency (ms)
    pub avg_hot_tier_ms: f64,
    /// Average warm tier search latency (ms)
    pub avg_warm_tier_ms: f64,
}

impl TieredMetrics {
    /// Calculate cache hit rate from hits and lookups
    pub fn calculate_hit_rate(&mut self) {
        if self.cache_lookups > 0 {
            self.cache_hit_rate = self.cache_hits as f32 / self.cache_lookups as f32;
        } else {
            self.cache_hit_rate = 0.0;
        }
    }
}

/// Per-query tiered metrics
#[derive(Debug, Clone, Default, Serialize, Deserialize)]
pub struct TieredQueryMetrics {
    /// Source tier that provided results ("cache" | "hot" | "warm" | "hybrid")
    pub source_tier: String,
    /// Cache lookup latency (ms)
    pub cache_lookup_ms: u64,
    /// Hot tier search latency (ms)
    pub hot_tier_ms: u64,
    /// Warm tier search latency (ms)
    pub warm_tier_ms: u64,
    /// Whether cache was hit
    pub cache_hit: bool,
    /// Whether hot tier returned sufficient results
    pub hot_tier_hit: bool,
}

/// Per-(tool, reason) rejection counter.
///
/// Phase 4.4: every time a tool call surfaces an `OperationError` (invalid
/// params, method not found, internal error, etc.) we bump the
/// matching counter. Operators can then see at a glance which tool
/// is producing which kind of rejection — historically these errors
/// flew straight back to the client with no aggregate visibility.
#[derive(Debug, Clone, Default, Serialize, Deserialize)]
pub struct RejectionStats {
    /// Total rejections across all tools/reasons.
    pub total: u64,
    /// Per-tool rejection counts, keyed by tool name (e.g. `"memory.search"`).
    pub by_tool: HashMap<String, u64>,
    /// Per-reason rejection counts, keyed by operation-error variant name
    /// (e.g. `"invalid_params"`, `"method_not_found"`).
    pub by_reason: HashMap<String, u64>,
}

/// Per-call operation payload token estimate.
///
/// These counters estimate tokens added by operation request params and
/// handler response payloads that `memd` can observe. They do not include the
/// agent's hidden reasoning, system prompt, repo/tool transcripts outside
/// `memd`, or provider-side cache accounting.
#[derive(Debug, Clone, Default, Serialize, Deserialize)]
pub struct ToolTokenUsage {
    /// Timestamp (Unix ms)
    pub timestamp: i64,
    /// Operation name, e.g. `"memory.search"`.
    pub tool: String,
    /// Serialized operation request params size.
    pub request_bytes: u64,
    /// Serialized operation response or error size.
    pub response_bytes: u64,
    /// Estimated request tokens.
    pub estimated_request_tokens: u64,
    /// Estimated response tokens.
    pub estimated_response_tokens: u64,
    /// Estimated request + response tokens.
    pub estimated_total_tokens: u64,
    /// Whether the tool call returned `Ok`.
    pub success: bool,
}

impl ToolTokenUsage {
    /// Build usage from serialized payload sizes.
    pub fn from_payload_sizes(
        tool: impl Into<String>,
        request_bytes: usize,
        response_bytes: usize,
        success: bool,
    ) -> Self {
        let request_bytes = request_bytes as u64;
        let response_bytes = response_bytes as u64;
        let estimated_request_tokens = estimate_tokens_from_bytes(request_bytes);
        let estimated_response_tokens = estimate_tokens_from_bytes(response_bytes);

        Self {
            timestamp: now_unix_ms(),
            tool: tool.into(),
            request_bytes,
            response_bytes,
            estimated_request_tokens,
            estimated_response_tokens,
            estimated_total_tokens: estimated_request_tokens + estimated_response_tokens,
            success,
        }
    }
}

/// Aggregated token estimate for one operation.
#[derive(Debug, Clone, Default, Serialize, Deserialize)]
pub struct ToolTokenAggregate {
    /// Tool calls observed.
    pub calls: u64,
    /// Calls that returned an error.
    pub errors: u64,
    /// Total serialized request bytes.
    pub request_bytes: u64,
    /// Total serialized response/error bytes.
    pub response_bytes: u64,
    /// Total estimated request tokens.
    pub estimated_request_tokens: u64,
    /// Total estimated response tokens.
    pub estimated_response_tokens: u64,
    /// Total estimated request + response tokens.
    pub estimated_total_tokens: u64,
}

impl ToolTokenAggregate {
    fn record(&mut self, usage: &ToolTokenUsage) {
        self.calls += 1;
        if !usage.success {
            self.errors += 1;
        }
        self.request_bytes += usage.request_bytes;
        self.response_bytes += usage.response_bytes;
        self.estimated_request_tokens += usage.estimated_request_tokens;
        self.estimated_response_tokens += usage.estimated_response_tokens;
        self.estimated_total_tokens += usage.estimated_total_tokens;
    }
}

/// Aggregated operation payload token estimates.
#[derive(Debug, Clone, Default, Serialize, Deserialize)]
pub struct TokenUsageStats {
    /// Estimator used for all token fields.
    pub estimator: String,
    /// Aggregate over all operation calls.
    pub total: ToolTokenAggregate,
    /// Aggregate per operation.
    pub by_tool: HashMap<String, ToolTokenAggregate>,
    /// Recent operation payload estimates.
    pub recent_tool_calls: Vec<ToolTokenUsage>,
}

/// Complete metrics snapshot
#[derive(Debug, Clone, Default, Serialize, Deserialize)]
pub struct MetricsSnapshot {
    /// Timestamp (Unix ms)
    pub timestamp: i64,
    /// Per-tenant index stats
    pub index: HashMap<String, IndexStats>,
    /// Latency statistics
    pub latency: LatencyStats,
    /// Recent query metrics (last N queries)
    pub recent_queries: Vec<QueryMetrics>,
    /// Tiered search metrics
    pub tiered: TieredMetrics,
    /// Phase 4.4: per-tool / per-reason rejection counts.
    #[serde(default)]
    pub rejections: RejectionStats,
    /// Estimated operation payload token usage by operation.
    #[serde(default)]
    pub token_usage: TokenUsageStats,
}

/// Metrics collector
pub struct MetricsCollector {
    /// Recent query latencies for percentile calculation
    latencies: RwLock<Vec<QueryMetrics>>,
    /// Maximum history size
    max_history: usize,
    /// Total query count
    query_count: AtomicU64,
    /// Cumulative totals for averages
    total_embed_ms: AtomicU64,
    total_search_ms: AtomicU64,
    total_fetch_ms: AtomicU64,
    total_total_ms: AtomicU64,
    /// Tiered metrics counters
    tiered_lookups: AtomicU64,
    tiered_cache_hits: AtomicU64,
    tiered_hot_hits: AtomicU64,
    tiered_promotions: AtomicU64,
    tiered_demotions: AtomicU64,
    /// Recent tiered query metrics for latency averaging
    tiered_latencies: RwLock<Vec<TieredQueryMetrics>>,
    /// Phase 4.4: per-tool and per-reason rejection counts.
    rejections_total: AtomicU64,
    rejections_by_tool: RwLock<HashMap<String, u64>>,
    rejections_by_reason: RwLock<HashMap<String, u64>>,
    /// Operation payload token estimates.
    tool_token_calls: AtomicU64,
    tool_token_errors: AtomicU64,
    tool_token_request_bytes: AtomicU64,
    tool_token_response_bytes: AtomicU64,
    tool_token_request_tokens: AtomicU64,
    tool_token_response_tokens: AtomicU64,
    tool_token_by_tool: RwLock<HashMap<String, ToolTokenAggregate>>,
    tool_token_recent: RwLock<Vec<ToolTokenUsage>>,
}

impl Default for MetricsCollector {
    fn default() -> Self {
        Self::new(1000)
    }
}

impl Clone for MetricsCollector {
    fn clone(&self) -> Self {
        Self {
            latencies: RwLock::new(self.latencies.read().clone()),
            max_history: self.max_history,
            query_count: AtomicU64::new(self.query_count.load(Ordering::Relaxed)),
            total_embed_ms: AtomicU64::new(self.total_embed_ms.load(Ordering::Relaxed)),
            total_search_ms: AtomicU64::new(self.total_search_ms.load(Ordering::Relaxed)),
            total_fetch_ms: AtomicU64::new(self.total_fetch_ms.load(Ordering::Relaxed)),
            total_total_ms: AtomicU64::new(self.total_total_ms.load(Ordering::Relaxed)),
            tiered_lookups: AtomicU64::new(self.tiered_lookups.load(Ordering::Relaxed)),
            tiered_cache_hits: AtomicU64::new(self.tiered_cache_hits.load(Ordering::Relaxed)),
            tiered_hot_hits: AtomicU64::new(self.tiered_hot_hits.load(Ordering::Relaxed)),
            tiered_promotions: AtomicU64::new(self.tiered_promotions.load(Ordering::Relaxed)),
            tiered_demotions: AtomicU64::new(self.tiered_demotions.load(Ordering::Relaxed)),
            tiered_latencies: RwLock::new(self.tiered_latencies.read().clone()),
            rejections_total: AtomicU64::new(self.rejections_total.load(Ordering::Relaxed)),
            rejections_by_tool: RwLock::new(self.rejections_by_tool.read().clone()),
            rejections_by_reason: RwLock::new(self.rejections_by_reason.read().clone()),
            tool_token_calls: AtomicU64::new(self.tool_token_calls.load(Ordering::Relaxed)),
            tool_token_errors: AtomicU64::new(self.tool_token_errors.load(Ordering::Relaxed)),
            tool_token_request_bytes: AtomicU64::new(
                self.tool_token_request_bytes.load(Ordering::Relaxed),
            ),
            tool_token_response_bytes: AtomicU64::new(
                self.tool_token_response_bytes.load(Ordering::Relaxed),
            ),
            tool_token_request_tokens: AtomicU64::new(
                self.tool_token_request_tokens.load(Ordering::Relaxed),
            ),
            tool_token_response_tokens: AtomicU64::new(
                self.tool_token_response_tokens.load(Ordering::Relaxed),
            ),
            tool_token_by_tool: RwLock::new(self.tool_token_by_tool.read().clone()),
            tool_token_recent: RwLock::new(self.tool_token_recent.read().clone()),
        }
    }
}

impl MetricsCollector {
    /// Create a new metrics collector
    pub fn new(max_history: usize) -> Self {
        Self {
            latencies: RwLock::new(Vec::with_capacity(max_history)),
            max_history,
            query_count: AtomicU64::new(0),
            total_embed_ms: AtomicU64::new(0),
            total_search_ms: AtomicU64::new(0),
            total_fetch_ms: AtomicU64::new(0),
            total_total_ms: AtomicU64::new(0),
            tiered_lookups: AtomicU64::new(0),
            tiered_cache_hits: AtomicU64::new(0),
            tiered_hot_hits: AtomicU64::new(0),
            tiered_promotions: AtomicU64::new(0),
            tiered_demotions: AtomicU64::new(0),
            tiered_latencies: RwLock::new(Vec::with_capacity(max_history)),
            rejections_total: AtomicU64::new(0),
            rejections_by_tool: RwLock::new(HashMap::new()),
            rejections_by_reason: RwLock::new(HashMap::new()),
            tool_token_calls: AtomicU64::new(0),
            tool_token_errors: AtomicU64::new(0),
            tool_token_request_bytes: AtomicU64::new(0),
            tool_token_response_bytes: AtomicU64::new(0),
            tool_token_request_tokens: AtomicU64::new(0),
            tool_token_response_tokens: AtomicU64::new(0),
            tool_token_by_tool: RwLock::new(HashMap::new()),
            tool_token_recent: RwLock::new(Vec::with_capacity(max_history)),
        }
    }

    /// Phase 4.4: record a rejected tool call.
    ///
    /// `tool` is the operation name (e.g. `"memory.search"`). `reason`
    /// is a short kebab-case label like `"invalid-params"` or
    /// `"method-not-found"` — typically derived from the `McpError`
    /// variant. Cheap path: three atomic/lock ops, no allocation
    /// beyond the HashMap entry when a new key is seen.
    pub fn record_rejection(&self, tool: &str, reason: &str) {
        self.rejections_total.fetch_add(1, Ordering::Relaxed);
        *self
            .rejections_by_tool
            .write()
            .entry(tool.to_string())
            .or_insert(0) += 1;
        *self
            .rejections_by_reason
            .write()
            .entry(reason.to_string())
            .or_insert(0) += 1;
    }

    /// Phase 4.4: read the current rejection aggregate.
    pub fn get_rejection_stats(&self) -> RejectionStats {
        RejectionStats {
            total: self.rejections_total.load(Ordering::Relaxed),
            by_tool: self.rejections_by_tool.read().clone(),
            by_reason: self.rejections_by_reason.read().clone(),
        }
    }

    /// Record estimated operation request/response tokens for one call.
    pub fn record_tool_token_usage(&self, usage: ToolTokenUsage) {
        self.tool_token_calls.fetch_add(1, Ordering::Relaxed);
        if !usage.success {
            self.tool_token_errors.fetch_add(1, Ordering::Relaxed);
        }
        self.tool_token_request_bytes
            .fetch_add(usage.request_bytes, Ordering::Relaxed);
        self.tool_token_response_bytes
            .fetch_add(usage.response_bytes, Ordering::Relaxed);
        self.tool_token_request_tokens
            .fetch_add(usage.estimated_request_tokens, Ordering::Relaxed);
        self.tool_token_response_tokens
            .fetch_add(usage.estimated_response_tokens, Ordering::Relaxed);

        self.tool_token_by_tool
            .write()
            .entry(usage.tool.clone())
            .or_default()
            .record(&usage);

        let mut recent = self.tool_token_recent.write();
        if recent.len() >= self.max_history {
            recent.remove(0);
        }
        recent.push(usage);
    }

    /// Get recent operation token estimates.
    pub fn get_recent_tool_token_usage(&self, limit: usize) -> Vec<ToolTokenUsage> {
        let recent = self.tool_token_recent.read();
        recent.iter().rev().take(limit).cloned().collect()
    }

    /// Get aggregated operation payload token estimates.
    pub fn get_token_usage_stats(&self) -> TokenUsageStats {
        let request_tokens = self.tool_token_request_tokens.load(Ordering::Relaxed);
        let response_tokens = self.tool_token_response_tokens.load(Ordering::Relaxed);

        TokenUsageStats {
            estimator: "ceil(serialized_mcp_payload_bytes / 4)".to_string(),
            total: ToolTokenAggregate {
                calls: self.tool_token_calls.load(Ordering::Relaxed),
                errors: self.tool_token_errors.load(Ordering::Relaxed),
                request_bytes: self.tool_token_request_bytes.load(Ordering::Relaxed),
                response_bytes: self.tool_token_response_bytes.load(Ordering::Relaxed),
                estimated_request_tokens: request_tokens,
                estimated_response_tokens: response_tokens,
                estimated_total_tokens: request_tokens + response_tokens,
            },
            by_tool: self.tool_token_by_tool.read().clone(),
            recent_tool_calls: self.get_recent_tool_token_usage(10),
        }
    }

    /// Record a query's latency breakdown
    pub fn record_query(&self, metrics: QueryMetrics) {
        // Update totals
        self.query_count.fetch_add(1, Ordering::Relaxed);
        self.total_embed_ms
            .fetch_add(metrics.embed_ms, Ordering::Relaxed);
        self.total_search_ms
            .fetch_add(metrics.dense_search_ms, Ordering::Relaxed);
        self.total_fetch_ms
            .fetch_add(metrics.fetch_ms, Ordering::Relaxed);
        self.total_total_ms
            .fetch_add(metrics.total_ms, Ordering::Relaxed);

        // Add to history (circular buffer)
        let mut latencies = self.latencies.write();
        if latencies.len() >= self.max_history {
            latencies.remove(0);
        }
        latencies.push(metrics);
    }

    /// Get latency statistics
    pub fn get_latency_stats(&self) -> LatencyStats {
        let count = self.query_count.load(Ordering::Relaxed);
        if count == 0 {
            return LatencyStats::default();
        }

        let latencies = self.latencies.read();

        // Calculate percentiles
        let mut totals: Vec<u64> = latencies.iter().map(|m| m.total_ms).collect();
        totals.sort_unstable();

        let p50 = percentile(&totals, 50);
        let p90 = percentile(&totals, 90);
        let p99 = percentile(&totals, 99);

        LatencyStats {
            count,
            avg_total_ms: self.total_total_ms.load(Ordering::Relaxed) as f64 / count as f64,
            avg_embed_ms: self.total_embed_ms.load(Ordering::Relaxed) as f64 / count as f64,
            avg_dense_search_ms: self.total_search_ms.load(Ordering::Relaxed) as f64 / count as f64,
            p50_total_ms: p50,
            p90_total_ms: p90,
            p99_total_ms: p99,
        }
    }

    /// Get recent query metrics
    pub fn get_recent_queries(&self, limit: usize) -> Vec<QueryMetrics> {
        let latencies = self.latencies.read();
        latencies.iter().rev().take(limit).cloned().collect()
    }

    /// Record a tiered query's metrics
    pub fn record_tiered_query(&self, metrics: TieredQueryMetrics) {
        // Update counters
        self.tiered_lookups.fetch_add(1, Ordering::Relaxed);
        if metrics.cache_hit {
            self.tiered_cache_hits.fetch_add(1, Ordering::Relaxed);
        }
        if metrics.hot_tier_hit {
            self.tiered_hot_hits.fetch_add(1, Ordering::Relaxed);
        }

        // Add to history (circular buffer)
        let mut tiered_latencies = self.tiered_latencies.write();
        if tiered_latencies.len() >= self.max_history {
            tiered_latencies.remove(0);
        }
        tiered_latencies.push(metrics);
    }

    /// Record a chunk promotion to hot tier
    pub fn record_promotion(&self) {
        self.tiered_promotions.fetch_add(1, Ordering::Relaxed);
    }

    /// Record a chunk demotion from hot tier
    pub fn record_demotion(&self) {
        self.tiered_demotions.fetch_add(1, Ordering::Relaxed);
    }

    /// Get tiered search statistics
    pub fn get_tiered_stats(&self) -> TieredMetrics {
        let lookups = self.tiered_lookups.load(Ordering::Relaxed);
        let cache_hits = self.tiered_cache_hits.load(Ordering::Relaxed);
        let hot_hits = self.tiered_hot_hits.load(Ordering::Relaxed);

        let tiered_latencies = self.tiered_latencies.read();

        // Calculate average latencies from recent queries
        let (avg_cache_ms, avg_hot_ms, avg_warm_ms) = if tiered_latencies.is_empty() {
            (0.0, 0.0, 0.0)
        } else {
            let count = tiered_latencies.len() as f64;
            let sum_cache: u64 = tiered_latencies.iter().map(|m| m.cache_lookup_ms).sum();
            let sum_hot: u64 = tiered_latencies.iter().map(|m| m.hot_tier_ms).sum();
            let sum_warm: u64 = tiered_latencies.iter().map(|m| m.warm_tier_ms).sum();
            (
                sum_cache as f64 / count,
                sum_hot as f64 / count,
                sum_warm as f64 / count,
            )
        };

        // Count warm tier searches (non-cache-hit queries)
        let warm_tier_searches = lookups.saturating_sub(cache_hits);

        let mut metrics = TieredMetrics {
            cache_lookups: lookups,
            cache_hits,
            cache_misses: lookups.saturating_sub(cache_hits),
            cache_hit_rate: 0.0,                   // Calculated below
            hot_tier_searches: warm_tier_searches, // Hot tier searched on non-cache-hit
            hot_tier_hits: hot_hits,
            warm_tier_searches,
            promotions: self.tiered_promotions.load(Ordering::Relaxed),
            demotions: self.tiered_demotions.load(Ordering::Relaxed),
            avg_cache_lookup_ms: avg_cache_ms,
            avg_hot_tier_ms: avg_hot_ms,
            avg_warm_tier_ms: avg_warm_ms,
        };

        metrics.calculate_hit_rate();
        metrics
    }

    /// Create a metrics snapshot
    pub fn snapshot(&self, index_stats: HashMap<String, IndexStats>) -> MetricsSnapshot {
        let timestamp = std::time::SystemTime::now()
            .duration_since(std::time::UNIX_EPOCH)
            .map(|d| d.as_millis() as i64)
            .unwrap_or(0);

        MetricsSnapshot {
            timestamp,
            index: index_stats,
            latency: self.get_latency_stats(),
            recent_queries: self.get_recent_queries(10),
            tiered: self.get_tiered_stats(),
            rejections: self.get_rejection_stats(),
            token_usage: self.get_token_usage_stats(),
        }
    }
}

fn now_unix_ms() -> i64 {
    std::time::SystemTime::now()
        .duration_since(std::time::UNIX_EPOCH)
        .map(|d| d.as_millis() as i64)
        .unwrap_or(0)
}

fn estimate_tokens_from_bytes(bytes: u64) -> u64 {
    bytes.saturating_add(3) / 4
}

fn percentile(sorted: &[u64], p: usize) -> u64 {
    if sorted.is_empty() {
        return 0;
    }
    // Use an inclusive rank over n-1 so p50/p90 on small samples behave as expected.
    let idx = (p * (sorted.len() - 1) / 100).min(sorted.len() - 1);
    sorted[idx]
}

/// Timer for measuring durations
pub struct Timer {
    start: Instant,
}

impl Timer {
    pub fn start() -> Self {
        Self {
            start: Instant::now(),
        }
    }

    pub fn elapsed(&self) -> Duration {
        self.start.elapsed()
    }

    pub fn elapsed_ms(&self) -> u64 {
        self.elapsed().as_millis() as u64
    }
}

#[cfg(test)]
mod tests {
    use super::*;

    #[test]
    fn test_record_and_stats() {
        let collector = MetricsCollector::new(100);

        // Record some queries
        for i in 1..=10 {
            collector.record_query(QueryMetrics {
                embed_ms: i * 10,
                dense_search_ms: i * 5,
                fetch_ms: i * 2,
                total_ms: i * 17,
            });
        }

        let stats = collector.get_latency_stats();
        assert_eq!(stats.count, 10);
        assert!(stats.avg_total_ms > 0.0);
        assert!(stats.p50_total_ms > 0);
    }

    #[test]
    fn test_percentile() {
        let data = vec![1, 2, 3, 4, 5, 6, 7, 8, 9, 10];
        assert_eq!(percentile(&data, 50), 5);
        assert_eq!(percentile(&data, 90), 9);
    }

    #[test]
    fn test_timer() {
        let timer = Timer::start();
        std::thread::sleep(Duration::from_millis(10));
        assert!(timer.elapsed_ms() >= 10);
    }

    #[test]
    fn test_query_metrics_from_timings() {
        let metrics = QueryMetrics::from_timings(
            Duration::from_millis(10),
            Duration::from_millis(5),
            Duration::from_millis(2),
            Duration::from_millis(17),
        );

        assert_eq!(metrics.embed_ms, 10);
        assert_eq!(metrics.dense_search_ms, 5);
        assert_eq!(metrics.fetch_ms, 2);
        assert_eq!(metrics.total_ms, 17);
    }

    #[test]
    fn test_snapshot() {
        let collector = MetricsCollector::new(100);

        collector.record_query(QueryMetrics {
            embed_ms: 10,
            dense_search_ms: 5,
            fetch_ms: 2,
            total_ms: 17,
        });

        let mut index_stats = HashMap::new();
        index_stats.insert(
            "test_tenant".to_string(),
            IndexStats {
                chunks_indexed: 100,
                embeddings_count: 100,
                embedding_dimension: 384,
                index_memory_bytes: 1024 * 1024,
            },
        );

        let snapshot = collector.snapshot(index_stats);

        assert!(snapshot.timestamp > 0);
        assert_eq!(snapshot.latency.count, 1);
        assert_eq!(snapshot.recent_queries.len(), 1);
        assert!(snapshot.index.contains_key("test_tenant"));
    }

    #[test]
    fn test_circular_buffer() {
        let collector = MetricsCollector::new(5);

        // Record more than max_history
        for i in 1..=10 {
            collector.record_query(QueryMetrics {
                embed_ms: i,
                dense_search_ms: i,
                fetch_ms: i,
                total_ms: i,
            });
        }

        // Should only keep last 5
        let recent = collector.get_recent_queries(10);
        assert_eq!(recent.len(), 5);

        // Most recent should be 10 (reversed order)
        assert_eq!(recent[0].total_ms, 10);
        assert_eq!(recent[4].total_ms, 6);
    }

    #[test]
    fn test_empty_stats() {
        let collector = MetricsCollector::new(100);
        let stats = collector.get_latency_stats();

        assert_eq!(stats.count, 0);
        assert_eq!(stats.avg_total_ms, 0.0);
        assert_eq!(stats.p50_total_ms, 0);
    }

    #[test]
    fn test_tool_token_usage_recording() {
        let collector = MetricsCollector::new(10);

        collector.record_tool_token_usage(ToolTokenUsage::from_payload_sizes(
            "memory.search",
            15,
            400,
            true,
        ));
        collector.record_tool_token_usage(ToolTokenUsage::from_payload_sizes(
            "memory.search",
            5,
            7,
            false,
        ));

        let stats = collector.get_token_usage_stats();
        assert_eq!(stats.total.calls, 2);
        assert_eq!(stats.total.errors, 1);
        assert_eq!(stats.total.request_bytes, 20);
        assert_eq!(stats.total.response_bytes, 407);
        assert_eq!(stats.total.estimated_request_tokens, 6);
        assert_eq!(stats.total.estimated_response_tokens, 102);
        assert_eq!(stats.total.estimated_total_tokens, 108);

        let search = stats.by_tool.get("memory.search").unwrap();
        assert_eq!(search.calls, 2);
        assert_eq!(search.errors, 1);
        assert_eq!(stats.recent_tool_calls.len(), 2);
        assert!(!stats.recent_tool_calls[0].success);
    }

    #[test]
    fn test_tiered_metrics_recording() {
        let collector = MetricsCollector::new(100);

        // Record a cache hit
        collector.record_tiered_query(TieredQueryMetrics {
            source_tier: "cache".to_string(),
            cache_lookup_ms: 1,
            hot_tier_ms: 0,
            warm_tier_ms: 0,
            cache_hit: true,
            hot_tier_hit: false,
        });

        // Record a hot tier hit
        collector.record_tiered_query(TieredQueryMetrics {
            source_tier: "hot".to_string(),
            cache_lookup_ms: 1,
            hot_tier_ms: 5,
            warm_tier_ms: 0,
            cache_hit: false,
            hot_tier_hit: true,
        });

        // Record a warm tier query
        collector.record_tiered_query(TieredQueryMetrics {
            source_tier: "warm".to_string(),
            cache_lookup_ms: 1,
            hot_tier_ms: 3,
            warm_tier_ms: 20,
            cache_hit: false,
            hot_tier_hit: false,
        });

        let stats = collector.get_tiered_stats();

        assert_eq!(stats.cache_lookups, 3);
        assert_eq!(stats.cache_hits, 1);
        assert_eq!(stats.hot_tier_hits, 1);
        assert!((stats.cache_hit_rate - 0.333).abs() < 0.01);
    }

    #[test]
    fn test_tiered_promotions_demotions() {
        let collector = MetricsCollector::new(100);

        collector.record_promotion();
        collector.record_promotion();
        collector.record_demotion();

        let stats = collector.get_tiered_stats();

        assert_eq!(stats.promotions, 2);
        assert_eq!(stats.demotions, 1);
    }

    #[test]
    fn test_tiered_latency_averages() {
        let collector = MetricsCollector::new(100);

        for i in 0..5 {
            collector.record_tiered_query(TieredQueryMetrics {
                source_tier: "warm".to_string(),
                cache_lookup_ms: 2,
                hot_tier_ms: 4,
                warm_tier_ms: (i + 1) * 10, // 10, 20, 30, 40, 50
                cache_hit: false,
                hot_tier_hit: false,
            });
        }

        let stats = collector.get_tiered_stats();

        assert!((stats.avg_cache_lookup_ms - 2.0).abs() < 0.01);
        assert!((stats.avg_hot_tier_ms - 4.0).abs() < 0.01);
        assert!((stats.avg_warm_tier_ms - 30.0).abs() < 0.01); // (10+20+30+40+50)/5
    }

    #[test]
    fn test_snapshot_includes_tiered() {
        let collector = MetricsCollector::new(100);

        collector.record_tiered_query(TieredQueryMetrics {
            source_tier: "cache".to_string(),
            cache_lookup_ms: 1,
            hot_tier_ms: 0,
            warm_tier_ms: 0,
            cache_hit: true,
            hot_tier_hit: false,
        });

        collector.record_promotion();

        let snapshot = collector.snapshot(HashMap::new());

        assert_eq!(snapshot.tiered.cache_lookups, 1);
        assert_eq!(snapshot.tiered.cache_hits, 1);
        assert_eq!(snapshot.tiered.promotions, 1);
    }

    #[test]
    fn rejection_stats_record_and_snapshot() {
        let collector = MetricsCollector::new(100);

        collector.record_rejection("memory.search", "invalid-params");
        collector.record_rejection("memory.search", "invalid-params");
        collector.record_rejection("memory.search", "tool-error");
        collector.record_rejection("task.start", "invalid-params");

        let stats = collector.get_rejection_stats();
        assert_eq!(stats.total, 4);
        assert_eq!(stats.by_tool.get("memory.search"), Some(&3));
        assert_eq!(stats.by_tool.get("task.start"), Some(&1));
        assert_eq!(stats.by_reason.get("invalid-params"), Some(&3));
        assert_eq!(stats.by_reason.get("tool-error"), Some(&1));

        // Rejection stats must also flow into the full snapshot.
        let snapshot = collector.snapshot(HashMap::new());
        assert_eq!(snapshot.rejections.total, 4);
        assert_eq!(snapshot.rejections.by_tool.get("memory.search"), Some(&3));
    }

    #[test]
    fn test_tiered_metrics_calculate_hit_rate() {
        let mut metrics = TieredMetrics {
            cache_lookups: 100,
            cache_hits: 75,
            ..Default::default()
        };

        metrics.calculate_hit_rate();
        assert!((metrics.cache_hit_rate - 0.75).abs() < 0.001);

        // Edge case: no lookups
        let mut empty_metrics = TieredMetrics::default();
        empty_metrics.calculate_hit_rate();
        assert_eq!(empty_metrics.cache_hit_rate, 0.0);
    }
}