dbg-cli 0.3.3

A universal debugger CLI that lets AI agents observe runtime state instead of guessing from source code
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
use std::path::{Path, PathBuf};

use anyhow::{Context, Result, bail};
use rusqlite::{Connection, params};

/// gdbg's schema version.
///
/// Kept separate from `dbg_cli::session_db::SCHEMA_VERSION` for now —
/// gdbg's tables are GPU-specific and don't include the `session_id`
/// column that the unified SessionDb format uses. When gdbg is fully
/// migrated to SessionDb (see plan task 10 "full rewrite deferred"),
/// the two versions will unify.
///
/// Bumping this invalidates every saved `.gpu.db` file: `GpuDb::open`
/// refuses to load anything that doesn't match, pointing the user at
/// the raw `.nsys-rep` + `.csv` artifacts to re-ingest.
pub const GDBG_SCHEMA_VERSION: i64 = 1;

/// A GPU profiling session backed by a SQLite database.
pub struct GpuDb {
    pub conn: Connection,
    pub _path: PathBuf,
    /// Active focus filter (kernel name substring).
    pub focus: Option<String>,
    /// Active ignore filter (kernel name substring).
    pub ignore: Option<String>,
    /// Active region filter (region name substring).
    pub region_filter: Option<String>,
}

impl std::fmt::Debug for GpuDb {
    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        f.debug_struct("GpuDb")
            .field("path", &self._path)
            .field("focus", &self.focus)
            .field("ignore", &self.ignore)
            .field("region_filter", &self.region_filter)
            .finish()
    }
}

impl GpuDb {
    /// Create a new session database at the given path.
    pub fn create(path: &Path) -> Result<Self> {
        if let Some(parent) = path.parent() {
            std::fs::create_dir_all(parent)?;
        }
        let conn = Connection::open(path)
            .with_context(|| format!("cannot create {}", path.display()))?;
        conn.execute_batch("PRAGMA journal_mode=WAL; PRAGMA synchronous=NORMAL;")?;
        init_schema(&conn)?;
        conn.execute(
            &format!("PRAGMA user_version = {GDBG_SCHEMA_VERSION}"),
            [],
        )?;
        Ok(Self {
            conn,
            _path: path.to_path_buf(),
            focus: None,
            ignore: None,
            region_filter: None,
        })
    }

    /// Open an existing session database.
    ///
    /// Refuses to open any DB whose `user_version` doesn't match
    /// `GDBG_SCHEMA_VERSION`. There is no migration path — the raw
    /// `.nsys-rep` + `.csv` files under the session's collection
    /// directory are the durable artifact; re-run `gdbg <target>`
    /// to rebuild the index.
    pub fn open(path: &Path) -> Result<Self> {
        if !path.exists() {
            bail!("session not found: {}", path.display());
        }
        let conn = Connection::open(path)
            .with_context(|| format!("cannot open {}", path.display()))?;
        let found: i64 = conn
            .query_row("PRAGMA user_version", [], |r| r.get(0))
            .unwrap_or(0);
        if found != GDBG_SCHEMA_VERSION {
            bail!(
                "gdbg session DB at {path} has schema_version={found}, \
                 expected {expected}. No migration path — delete it and \
                 re-run `gdbg <target>` to rebuild from the raw captures.",
                path = path.display(),
                expected = GDBG_SCHEMA_VERSION,
            );
        }
        Ok(Self {
            conn,
            _path: path.to_path_buf(),
            focus: None,
            ignore: None,
            region_filter: None,
        })
    }

    /// Get the session storage directory for saved sessions.
    /// Walks up to find `.git` and uses that root; falls back to cwd.
    pub fn session_dir() -> PathBuf {
        find_project_root().join(".dbg").join("gpu")
    }

    /// Save this session by copying the DB to `.dbg/gpu/<name>.gpu.db`.
    pub fn save(&self, name: &str) -> Result<PathBuf> {
        let dir = Self::session_dir();
        std::fs::create_dir_all(&dir)?;
        let dest = dir.join(format!("{name}.gpu.db"));
        // Use SQLite backup API for safe copy of a live DB
        let mut dest_conn = Connection::open(&dest)?;
        let backup = rusqlite::backup::Backup::new(&self.conn, &mut dest_conn)?;
        backup.run_to_completion(100, std::time::Duration::from_millis(10), None)?;
        Ok(dest)
    }

    /// Load a saved session by name or path.
    pub fn load(name_or_path: &str) -> Result<Self> {
        let path = if name_or_path.ends_with(".gpu.db") || name_or_path.contains('/') {
            PathBuf::from(name_or_path)
        } else {
            Self::session_dir().join(format!("{name_or_path}.gpu.db"))
        };
        Self::open(&path)
    }

    /// List all saved sessions.
    pub fn list_saved() -> Result<Vec<SavedSession>> {
        let dir = Self::session_dir();
        if !dir.exists() {
            return Ok(Vec::new());
        }
        let mut sessions = Vec::new();
        for entry in std::fs::read_dir(&dir)? {
            let entry = entry?;
            let path = entry.path();
            if path.extension().is_some_and(|e| e == "db")
                && path.to_str().is_some_and(|s| s.contains(".gpu."))
            {
                if let Ok(db) = Self::open(&path) {
                    let name = path
                        .file_stem()
                        .unwrap_or_default()
                        .to_str()
                        .unwrap_or_default()
                        .strip_suffix(".gpu")
                        .unwrap_or_default()
                        .to_string();
                    sessions.push(SavedSession {
                        name,
                        _target: db.meta("target"),
                        device: db.meta("device"),
                        kernel_count: db.unique_kernel_count(),
                        _launch_count: db.total_launch_count(),
                        layers: db.layer_names(),
                        created: db.meta("created"),
                    });
                }
            }
        }
        sessions.sort_by(|a, b| b.created.cmp(&a.created));
        Ok(sessions)
    }

    // -----------------------------------------------------------------------
    // Meta
    // -----------------------------------------------------------------------

    pub fn set_meta(&self, key: &str, value: &str) -> Result<()> {
        self.conn.execute(
            "INSERT OR REPLACE INTO meta (key, value) VALUES (?1, ?2)",
            params![key, value],
        )?;
        Ok(())
    }

    pub fn meta(&self, key: &str) -> String {
        self.conn
            .query_row("SELECT value FROM meta WHERE key = ?1", params![key], |row| {
                row.get(0)
            })
            .unwrap_or_default()
    }

    // -----------------------------------------------------------------------
    // Layers
    // -----------------------------------------------------------------------

    pub fn add_layer(
        &self,
        source: &str,
        file: &str,
        command: Option<&str>,
        secs: Option<f64>,
        target_hash: Option<&str>,
    ) -> Result<i64> {
        self.conn.execute(
            "INSERT INTO layers (source, file, collected_at, command_used, collection_secs, target_hash)
             VALUES (?1, ?2, datetime('now'), ?3, ?4, ?5)",
            params![source, file, command, secs, target_hash],
        )?;
        Ok(self.conn.last_insert_rowid())
    }

    /// Undo an `add_layer` when a downstream import fails. Without
    /// this, a failed nsys/ncu import leaves an empty layer row that
    /// makes `has_layer("nsys")` return true and the session summary
    /// falsely claim both layers are present.
    pub fn remove_layer(&self, layer_id: i64) -> Result<()> {
        self.conn
            .execute("DELETE FROM layers WHERE id = ?1", params![layer_id])?;
        Ok(())
    }

    /// Execute a query and collect all rows via a mapping function.
    /// Returns an empty Vec on any error (safe for diagnostic/display code).
    pub fn query_vec<T>(
        &self,
        sql: &str,
        params: impl rusqlite::Params,
        f: impl FnMut(&rusqlite::Row<'_>) -> rusqlite::Result<T>,
    ) -> Vec<T> {
        let Ok(mut stmt) = self.conn.prepare(sql) else {
            return Vec::new();
        };
        let Ok(rows) = stmt.query_map(params, f) else {
            return Vec::new();
        };
        rows.filter_map(|r| r.ok()).collect()
    }

    /// Check if target hashes are consistent across all layers.
    /// Returns None if consistent, Some(warning) if mismatched.
    pub fn check_target_consistency(&self) -> Option<String> {
        let rows: Vec<(String, String)> = self.query_vec(
            "SELECT source, target_hash FROM layers WHERE target_hash IS NOT NULL",
            [],
            |row| Ok((row.get(0)?, row.get(1)?)),
        );

        if rows.len() < 2 {
            return None;
        }

        let first_hash = &rows[0].1;
        let mismatched: Vec<&str> = rows
            .iter()
            .filter(|(_, h)| h != first_hash)
            .map(|(s, _)| s.as_str())
            .collect();

        if mismatched.is_empty() {
            None
        } else {
            Some(format!(
                "target file changed between collection phases: {} vs {}",
                rows[0].0,
                mismatched.join(", ")
            ))
        }
    }

    /// Check kernel population consistency across layers.
    /// Returns warnings about kernels that appear in some layers but not others.
    pub fn check_kernel_consistency(&self) -> Vec<String> {
        let mut warnings = Vec::new();

        let has_nsys = self.has_layer("nsys");
        let has_torch = self.has_layer("torch");

        if !has_nsys || !has_torch {
            return warnings;
        }

        let orphans: Vec<String> = self.query_vec(
            "SELECT DISTINCT l.kernel_name
             FROM launches l
             WHERE l.layer_id IN (SELECT id FROM layers WHERE source = 'torch')
               AND l.kernel_name NOT IN (
                 SELECT DISTINCT kernel_name FROM launches
                 WHERE layer_id IN (SELECT id FROM layers WHERE source = 'nsys')
               )",
            [],
            |row| row.get(0),
        );

        if !orphans.is_empty() {
            warnings.push(format!(
                "{} kernels in torch layer but not nsys (different run?): {}",
                orphans.len(),
                orphans.iter().take(3).cloned().collect::<Vec<_>>().join(", ")
            ));
        }

        warnings
    }

    pub fn layer_names(&self) -> Vec<String> {
        self.query_vec(
            "SELECT DISTINCT source FROM layers ORDER BY id",
            [],
            |row| row.get(0),
        )
    }

    pub fn has_layer(&self, source: &str) -> bool {
        self.conn
            .query_row(
                "SELECT COUNT(*) FROM layers WHERE source = ?1",
                params![source],
                |row| row.get::<_, i64>(0),
            )
            .unwrap_or(0)
            > 0
    }

    /// Get the layer ID to use for timeline queries (prefer nsys, fall back to torch).
    /// Returns None if no timeline layer exists.
    pub fn timeline_layer_id(&self) -> Option<i64> {
        // Prefer nsys (has real timestamps), fall back to torch
        for source in &["nsys", "torch", "proton"] {
            if let Ok(id) = self.conn.query_row(
                "SELECT id FROM layers WHERE source = ?1 ORDER BY id LIMIT 1",
                params![source],
                |row| row.get::<_, i64>(0),
            ) {
                return Some(id);
            }
        }
        None
    }

    /// SQL fragment to filter launches to the best timeline layer.
    /// Uses `launches.layer_id` to be safe in JOIN contexts where the launches
    /// table is not aliased.  Use `timeline_filter_for("alias")` when the
    /// launches table has a different alias.
    pub fn timeline_filter(&self) -> String {
        self.timeline_filter_for("launches")
    }

    /// Like `timeline_filter`, but with a custom table alias.
    pub fn timeline_filter_for(&self, alias: &str) -> String {
        match self.timeline_layer_id() {
            Some(id) => format!("{alias}.layer_id = {id}"),
            None => "1=1".to_string(),
        }
    }

    // -----------------------------------------------------------------------
    // Failures
    // -----------------------------------------------------------------------

    pub fn add_failure(&self, phase: &str, error: &str) -> Result<()> {
        self.conn.execute(
            "INSERT INTO failures (phase, error) VALUES (?1, ?2)",
            params![phase, error],
        )?;
        Ok(())
    }

    pub fn failures(&self) -> Vec<(String, String)> {
        self.query_vec(
            "SELECT phase, error FROM failures",
            [],
            |row| Ok((row.get(0)?, row.get(1)?)),
        )
    }

    // -----------------------------------------------------------------------
    // Op GPU time recomputation
    // -----------------------------------------------------------------------

    /// Re-compute `ops.gpu_time_us` against the best timeline layer.
    ///
    /// During import, `ops.gpu_time_us` is computed from the torch/proton
    /// layer's kernel launches.  When an nsys layer is also present, its
    /// kernel durations are more accurate (lower profiler overhead).  This
    /// method re-correlates every op's GPU time against whichever layer
    /// `timeline_filter` selects, so that `top-ops`, `compare-ops`, and
    /// `hotpath` stay consistent with `breakdown` and `kernels`.
    pub fn recompute_op_gpu_times(&self) {
        let Some(tl_id) = self.timeline_layer_id() else { return };

        // Check whether the timeline layer is already the op layer —
        // if so, nothing to fix.
        let op_layers: Vec<String> = self.query_vec(
            "SELECT DISTINCT source FROM layers WHERE id IN (SELECT DISTINCT layer_id FROM ops)",
            [],
            |row| row.get(0),
        );
        let tl_source: String = self.conn.query_row(
            "SELECT source FROM layers WHERE id = ?1",
            params![tl_id],
            |row| row.get(0),
        ).unwrap_or_default();

        // If the only op layer is also the timeline layer, no recomputation needed.
        if op_layers.len() == 1 && op_layers[0] == tl_source {
            return;
        }

        // Re-correlate: for each op, sum kernel durations from the timeline layer.
        let _ = self.conn.execute(
            "UPDATE ops SET gpu_time_us = (
                SELECT COALESCE(SUM(l.duration_us), 0)
                FROM op_kernel_map okm
                JOIN launches l ON l.kernel_name = okm.kernel_name AND l.layer_id = ?1
                WHERE okm.op_id = ops.id
            )",
            params![tl_id],
        );
    }

    // -----------------------------------------------------------------------
    // Scalar query helpers
    // -----------------------------------------------------------------------

    /// Execute a SQL query that returns a single integer, defaulting to 0.
    pub fn count(&self, sql: &str) -> usize {
        self.conn
            .query_row(sql, [], |row| row.get::<_, i64>(0))
            .unwrap_or(0) as usize
    }

    /// Execute a SQL query that returns a single float, defaulting to 0.0.
    pub fn scalar_f64(&self, sql: &str) -> f64 {
        self.conn
            .query_row(sql, [], |row| row.get(0))
            .unwrap_or(0.0)
    }

    // -----------------------------------------------------------------------
    // Counts
    // -----------------------------------------------------------------------

    pub fn unique_kernel_count(&self) -> usize {
        let tl = self.timeline_filter();
        self.count(&format!("SELECT COUNT(DISTINCT kernel_name) FROM launches WHERE {tl}"))
    }

    pub fn total_launch_count(&self) -> usize {
        let tl = self.timeline_filter();
        self.count(&format!("SELECT COUNT(*) FROM launches WHERE {tl}"))
    }

    pub fn total_gpu_time_us(&self) -> f64 {
        let tl = self.timeline_filter();
        self.scalar_f64(&format!("SELECT COALESCE(SUM(duration_us), 0) FROM launches WHERE {tl}"))
    }

    pub fn transfer_count(&self) -> usize {
        self.count("SELECT COUNT(*) FROM transfers")
    }

    pub fn stream_count(&self) -> usize {
        let tl = self.timeline_filter();
        self.count(&format!("SELECT COUNT(DISTINCT stream_id) FROM launches WHERE stream_id IS NOT NULL AND {tl}"))
    }

    pub fn kernels_with_metrics(&self) -> usize {
        self.count("SELECT COUNT(*) FROM metrics")
    }

    pub fn kernels_with_ops(&self) -> usize {
        self.count("SELECT COUNT(DISTINCT kernel_name) FROM op_kernel_map")
    }

    // -----------------------------------------------------------------------
    // Filter helpers — builds WHERE clause fragments
    // -----------------------------------------------------------------------

    pub fn kernel_filter(&self) -> String {
        let mut clauses = Vec::new();
        if let Some(ref f) = self.focus {
            clauses.push(format!(r"launches.kernel_name LIKE '%{}%' ESCAPE '\'", escape_sql_like(f)));
        }
        if let Some(ref ig) = self.ignore {
            clauses.push(format!(r"launches.kernel_name NOT LIKE '%{}%' ESCAPE '\'", escape_sql_like(ig)));
        }
        if let Some(ref r) = self.region_filter {
            // Only include launches whose start_us falls within a matching region.
            clauses.push(format!(
                r"start_us IS NOT NULL AND EXISTS (
                   SELECT 1 FROM regions
                   WHERE name LIKE '%{}%' ESCAPE '\'
                     AND launches.start_us >= regions.start_us
                     AND launches.start_us <= regions.start_us + regions.duration_us
                 )",
                escape_sql_like(r)
            ));
        }
        if clauses.is_empty() {
            "1=1".to_string()
        } else {
            clauses.join(" AND ")
        }
    }

    // -----------------------------------------------------------------------
    // Attach another DB for diff
    // -----------------------------------------------------------------------

    pub fn attach(&self, path: &str, alias: &str) -> Result<()> {
        self.conn.execute_batch(&format!(
            "ATTACH DATABASE '{}' AS {alias}",
            path.replace('\'', "''")
        ))?;
        Ok(())
    }

    pub fn detach(&self, alias: &str) -> Result<()> {
        self.conn
            .execute_batch(&format!("DETACH DATABASE {alias}"))?;
        Ok(())
    }
}

// ---------------------------------------------------------------------------
// Schema
// ---------------------------------------------------------------------------

fn init_schema(conn: &Connection) -> Result<()> {
    conn.execute_batch(
        "
        CREATE TABLE IF NOT EXISTS meta (
            key   TEXT PRIMARY KEY,
            value TEXT
        );

        CREATE TABLE IF NOT EXISTS layers (
            id              INTEGER PRIMARY KEY,
            source          TEXT NOT NULL,
            file            TEXT,
            collected_at    TEXT,
            command_used    TEXT,
            collection_secs REAL,
            target_hash     TEXT
        );

        CREATE TABLE IF NOT EXISTS launches (
            id             INTEGER PRIMARY KEY,
            kernel_name    TEXT NOT NULL,
            duration_us    REAL NOT NULL,
            grid_x         INTEGER,
            grid_y         INTEGER,
            grid_z         INTEGER,
            block_x        INTEGER,
            block_y        INTEGER,
            block_z        INTEGER,
            stream_id      INTEGER,
            start_us       REAL,
            correlation_id INTEGER,
            layer_id       INTEGER REFERENCES layers(id)
        );

        CREATE TABLE IF NOT EXISTS metrics (
            kernel_name              TEXT PRIMARY KEY,
            occupancy_pct            REAL,
            compute_throughput_pct   REAL,
            memory_throughput_pct    REAL,
            registers_per_thread     INTEGER,
            shared_mem_static_bytes  INTEGER,
            shared_mem_dynamic_bytes INTEGER,
            l2_hit_rate_pct          REAL,
            achieved_bandwidth_gb_s  REAL,
            peak_bandwidth_gb_s      REAL,
            boundedness              TEXT,
            layer_id                 INTEGER REFERENCES layers(id)
        );

        CREATE TABLE IF NOT EXISTS transfers (
            id          INTEGER PRIMARY KEY,
            kind        TEXT NOT NULL,
            bytes       INTEGER,
            duration_us REAL,
            start_us    REAL,
            stream_id   INTEGER,
            layer_id    INTEGER REFERENCES layers(id)
        );

        CREATE TABLE IF NOT EXISTS ops (
            id           INTEGER PRIMARY KEY,
            name         TEXT NOT NULL,
            module_path  TEXT,
            cpu_time_us  REAL,
            gpu_time_us  REAL,
            input_shapes TEXT,
            layer_id     INTEGER REFERENCES layers(id)
        );

        CREATE TABLE IF NOT EXISTS op_kernel_map (
            op_id       INTEGER REFERENCES ops(id),
            kernel_name TEXT,
            PRIMARY KEY (op_id, kernel_name)
        );

        CREATE TABLE IF NOT EXISTS allocations (
            id        INTEGER PRIMARY KEY,
            op        TEXT NOT NULL,        -- 'alloc' or 'free'
            address   INTEGER NOT NULL,
            bytes     INTEGER NOT NULL,     -- 0 for frees when size unknown
            start_us  REAL,
            stream_id INTEGER,
            layer_id  INTEGER REFERENCES layers(id)
        );

        CREATE INDEX IF NOT EXISTS idx_alloc_addr ON allocations(address);
        CREATE INDEX IF NOT EXISTS idx_alloc_time ON allocations(start_us);

        CREATE TABLE IF NOT EXISTS regions (
            id          INTEGER PRIMARY KEY,
            name        TEXT NOT NULL,
            start_us    REAL,
            duration_us REAL,
            layer_id    INTEGER REFERENCES layers(id)
        );

        CREATE TABLE IF NOT EXISTS failures (
            phase TEXT,
            error TEXT
        );

        CREATE INDEX IF NOT EXISTS idx_launches_kernel ON launches(kernel_name);
        CREATE INDEX IF NOT EXISTS idx_launches_start ON launches(start_us);
        CREATE INDEX IF NOT EXISTS idx_launches_stream ON launches(stream_id);
        CREATE INDEX IF NOT EXISTS idx_transfers_start ON transfers(start_us);
        ",
    )?;
    Ok(())
}

// ---------------------------------------------------------------------------
// Project root detection
// ---------------------------------------------------------------------------

/// Walk up from cwd to find a `.git` directory. Returns that parent, or cwd.
fn find_project_root() -> PathBuf {
    let cwd = std::env::current_dir().unwrap_or_else(|_| PathBuf::from("."));
    let mut dir = cwd.as_path();
    loop {
        if dir.join(".git").exists() {
            return dir.to_path_buf();
        }
        match dir.parent() {
            Some(parent) => dir = parent,
            None => return cwd,
        }
    }
}

// ---------------------------------------------------------------------------
// SQL safety
// ---------------------------------------------------------------------------

/// Escape a value for safe interpolation into a SQL LIKE pattern.
/// Doubles single quotes and escapes LIKE wildcards.
/// Escape a user pattern for use in SQL LIKE.
///
/// - Quotes are doubled for SQL string safety.
/// - `%` is escaped with backslash (the wildcard meaning is reserved internally).
/// - `_` is NOT escaped: kernel names contain many underscores and users
///   typing "vector_add" expect a literal match, not a wildcard.  Allowing
///   `_` as a single-char wildcard is harmless in practice.
///
/// Callers using this helper must append `ESCAPE '\'` to their LIKE clause
/// so the backslash-escaped `%` is recognized.
pub fn escape_sql_like(s: &str) -> String {
    s.replace('\'', "''").replace('%', "\\%")
}

// ---------------------------------------------------------------------------
// Saved session info
// ---------------------------------------------------------------------------

#[derive(Debug)]
pub struct SavedSession {
    pub name: String,
    pub _target: String,
    pub device: String,
    pub kernel_count: usize,
    pub _launch_count: usize,
    pub layers: Vec<String>,
    pub created: String,
}

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

    fn temp_db() -> GpuDb {
        let dir = tempfile::tempdir().unwrap();
        let path = dir.keep().join("test.gpu.db");
        GpuDb::create(&path).unwrap()
    }

    #[test]
    fn meta_roundtrip() {
        let db = temp_db();
        db.set_meta("target", "train.py").unwrap();
        db.set_meta("device", "A100").unwrap();
        assert_eq!(db.meta("target"), "train.py");
        assert_eq!(db.meta("device"), "A100");
        assert_eq!(db.meta("missing"), "");
    }

    #[test]
    fn add_layer() {
        let db = temp_db();
        let id = db.add_layer("nsys", "/tmp/trace.nsys-rep", Some("nsys profile"), Some(12.5), None).unwrap();
        assert_eq!(id, 1);
        assert!(db.has_layer("nsys"));
        assert!(!db.has_layer("ncu"));
        assert_eq!(db.layer_names(), vec!["nsys"]);
    }

    /// Regression: when the nsys import failed (nsys 2023 schema
    /// mismatch, for example), the layer row inserted before the
    /// import persisted and `has_layer("nsys")` returned true for an
    /// empty layer — causing the session summary to claim `Layers:
    /// nsys + ncu` with zero data. `remove_layer` lets the collector
    /// roll back the row on error.
    #[test]
    fn remove_layer_restores_has_layer_to_false() {
        let db = temp_db();
        let id = db
            .add_layer("nsys", "/tmp/trace.nsys-rep", None, None, None)
            .unwrap();
        assert!(db.has_layer("nsys"));
        db.remove_layer(id).unwrap();
        assert!(!db.has_layer("nsys"), "layer row must be gone after remove");
        assert!(db.layer_names().is_empty());
    }

    #[test]
    fn kernel_counts() {
        let db = temp_db();
        let lid = db.add_layer("nsys", "test", None, None, None).unwrap();
        db.conn.execute(
            "INSERT INTO launches (kernel_name, duration_us, layer_id) VALUES ('k1', 100.0, ?1)",
            params![lid],
        ).unwrap();
        db.conn.execute(
            "INSERT INTO launches (kernel_name, duration_us, layer_id) VALUES ('k1', 200.0, ?1)",
            params![lid],
        ).unwrap();
        db.conn.execute(
            "INSERT INTO launches (kernel_name, duration_us, layer_id) VALUES ('k2', 50.0, ?1)",
            params![lid],
        ).unwrap();

        assert_eq!(db.unique_kernel_count(), 2);
        assert_eq!(db.total_launch_count(), 3);
        assert!((db.total_gpu_time_us() - 350.0).abs() < 0.1);
    }

    #[test]
    fn failures() {
        let db = temp_db();
        db.add_failure("ncu", "ncu not found").unwrap();
        let f = db.failures();
        assert_eq!(f.len(), 1);
        assert_eq!(f[0].0, "ncu");
        assert_eq!(f[0].1, "ncu not found");
    }

    #[test]
    fn kernel_filter() {
        let mut db = temp_db();
        assert_eq!(db.kernel_filter(), "1=1");
        db.focus = Some("sgemm".into());
        assert!(db.kernel_filter().contains("launches.kernel_name LIKE '%sgemm%'"));
        db.ignore = Some("copy".into());
        assert!(db.kernel_filter().contains("NOT LIKE '%copy%'"));
        // Verify table-qualified to avoid ambiguity in JOINs
        assert!(db.kernel_filter().contains("launches.kernel_name"));
    }

    #[test]
    fn save_and_load() {
        let db = temp_db();
        db.set_meta("target", "test.py").unwrap();
        let lid = db.add_layer("nsys", "test", None, None, None).unwrap();
        db.conn.execute(
            "INSERT INTO launches (kernel_name, duration_us, layer_id) VALUES ('k1', 100.0, ?1)",
            params![lid],
        ).unwrap();

        let dir = tempfile::tempdir().unwrap();
        let dest = dir.path().join("saved.gpu.db");
        // Save via file copy (simpler test than going through .dbg/gpu/)
        {
            let mut dest_conn = Connection::open(&dest).unwrap();
            let backup = rusqlite::backup::Backup::new(&db.conn, &mut dest_conn).unwrap();
            backup.run_to_completion(100, std::time::Duration::from_millis(10), None).unwrap();
        }

        let loaded = GpuDb::open(&dest).unwrap();
        assert_eq!(loaded.meta("target"), "test.py");
        assert_eq!(loaded.unique_kernel_count(), 1);
    }

    #[test]
    fn create_stamps_schema_version() {
        let db = temp_db();
        let v: i64 = db
            .conn
            .query_row("PRAGMA user_version", [], |r| r.get(0))
            .unwrap();
        assert_eq!(v, GDBG_SCHEMA_VERSION);
    }

    #[test]
    fn open_refuses_unstamped_old_format() {
        // Simulate a pre-versioning `.gpu.db`: the tables + schema are
        // there but PRAGMA user_version is 0 (SQLite's default).
        let dir = tempfile::tempdir().unwrap();
        let path = dir.path().join("legacy.gpu.db");
        {
            let conn = Connection::open(&path).unwrap();
            init_schema(&conn).unwrap();
            // deliberately NOT stamping user_version
        }
        let err = GpuDb::open(&path).unwrap_err().to_string();
        assert!(err.contains("schema_version=0"), "{err}");
        assert!(err.contains("No migration path"), "{err}");
        assert!(err.contains("re-run `gdbg"), "{err}");
    }

    #[test]
    fn open_refuses_future_version() {
        let dir = tempfile::tempdir().unwrap();
        let path = dir.path().join("future.gpu.db");
        {
            let conn = Connection::open(&path).unwrap();
            init_schema(&conn).unwrap();
            conn.execute("PRAGMA user_version = 99", []).unwrap();
        }
        let err = GpuDb::open(&path).unwrap_err().to_string();
        assert!(err.contains("schema_version=99"));
    }

    #[test]
    fn save_preserves_version_through_backup() {
        let db = temp_db();
        db.set_meta("marker", "present").unwrap();
        let dir = tempfile::tempdir().unwrap();
        let dest = dir.path().join("backed_up.gpu.db");
        {
            let mut dest_conn = Connection::open(&dest).unwrap();
            let backup =
                rusqlite::backup::Backup::new(&db.conn, &mut dest_conn).unwrap();
            backup
                .run_to_completion(100, std::time::Duration::from_millis(10), None)
                .unwrap();
        }
        // The backup path must pass the version gate cleanly.
        let loaded = GpuDb::open(&dest).unwrap();
        assert_eq!(loaded.meta("marker"), "present");
    }
}