personal memory agent
speakers: add `sol call speakers wipe` to clear legacy encoder artifacts
Adds a one-shot destructive command for removing speaker artifacts left
over from the old encoder so the new embedder starts from a clean
journal. Without --commit the command reports what would be deleted;
with --commit it removes:
- chronicle/*/*/*/*.npz (segment embeddings)
- chronicle/*/*/*/talents/
speaker_labels.json
speaker_corrections.json
- entities/*/voiceprints.npz (per-entity voiceprints)
- entities/*/owner_centroid.npz (every principal's centroid)
- awareness/owner_candidate.npz (owner-candidate snapshot)
Follows the bootstrap/backfill idioms: --commit opts into mutation,
--json emits the structured report, and the human output lists per-
category counts plus total files/bytes. The short-help and detail-help
both carry a DESTRUCTIVE warning.
The wipe is the migration — old resemblyzer-space embeddings are not
salvageable against a new encoder, and there is no compat layer for
NPZs lacking the `encoder` provenance field.
+216
+61
apps/speakers/call.py
+61
apps/speakers/call.py
···
13
13
sol call speakers resolve-names [--commit] [--json]
14
14
sol call speakers attribute-segment <day> <stream> <segment> [--commit] [--json]
15
15
sol call speakers backfill [--commit] [--json]
16
+
sol call speakers wipe [--commit] [--json]
16
17
sol call speakers discover [--json]
17
18
sol call speakers identify <cluster-id> <name> [--entity-id ID]
18
19
sol call speakers merge-names <alias> <canonical>
···
337
338
typer.echo(f" {err}", err=True)
338
339
if len(stats["errors"]) > 10:
339
340
typer.echo(f" ... and {len(stats['errors']) - 10} more", err=True)
341
+
342
+
343
+
@app.command()
344
+
def wipe(
345
+
commit: bool = typer.Option(
346
+
False,
347
+
"--commit",
348
+
help="Actually delete files. Without this flag the command only reports what would happen.",
349
+
),
350
+
json_output: bool = typer.Option(
351
+
False, "--json", help="Output full result as JSON."
352
+
),
353
+
) -> None:
354
+
"""Remove all legacy speaker artifacts from the journal (DESTRUCTIVE).
355
+
356
+
DESTRUCTIVE. Without --commit, prints a report of what would be
357
+
removed. With --commit, permanently deletes segment-embedding NPZs,
358
+
speaker labels/corrections, per-entity voiceprints, owner centroids,
359
+
and the owner-candidate snapshot.
360
+
"""
361
+
import json as json_mod
362
+
363
+
from apps.speakers.wipe import wipe_speaker_artifacts
364
+
365
+
if not commit and not json_output:
366
+
typer.echo("REPORT ONLY — pass --commit to persist.\n")
367
+
368
+
report = wipe_speaker_artifacts(dry_run=not commit)
369
+
370
+
if json_output:
371
+
typer.echo(json_mod.dumps(report.to_dict(), indent=2, default=str))
372
+
return
373
+
374
+
typer.echo(
375
+
f"segment_embeddings : {report.segment_embeddings.count} files "
376
+
f"({report.segment_embeddings.bytes} B)"
377
+
)
378
+
typer.echo(
379
+
f"speaker_labels : {report.speaker_labels.count} files "
380
+
f"({report.speaker_labels.bytes} B)"
381
+
)
382
+
typer.echo(
383
+
f"speaker_corrections: {report.speaker_corrections.count} files "
384
+
f"({report.speaker_corrections.bytes} B)"
385
+
)
386
+
typer.echo(
387
+
f"entity_voiceprints : {report.entity_voiceprints.count} files "
388
+
f"({report.entity_voiceprints.bytes} B)"
389
+
)
390
+
typer.echo(
391
+
f"owner_centroids : {report.owner_centroids.count} files "
392
+
f"({report.owner_centroids.bytes} B)"
393
+
)
394
+
typer.echo(
395
+
f"owner_candidate : {report.owner_candidate.count} files "
396
+
f"({report.owner_candidate.bytes} B)"
397
+
)
398
+
typer.echo(
399
+
f"total : {report.total_files} files ({report.total_bytes} B)"
400
+
)
340
401
341
402
342
403
@app.command()
+79
apps/speakers/tests/test_wipe.py
+79
apps/speakers/tests/test_wipe.py
···
1
+
# SPDX-License-Identifier: AGPL-3.0-only
2
+
# Copyright (c) 2026 sol pbc
3
+
4
+
"""Tests for speaker artifact wipe reporting."""
5
+
6
+
from __future__ import annotations
7
+
8
+
from pathlib import Path
9
+
10
+
import pytest
11
+
12
+
from apps.speakers.wipe import wipe_speaker_artifacts
13
+
14
+
FILE_BYTES = b"test-data"
15
+
16
+
17
+
@pytest.fixture
18
+
def wipe_journal(
19
+
tmp_path: Path, monkeypatch: pytest.MonkeyPatch
20
+
) -> tuple[Path, list[Path]]:
21
+
journal = tmp_path / "journal"
22
+
monkeypatch.setenv("_SOLSTONE_JOURNAL_OVERRIDE", str(journal))
23
+
24
+
import think.utils
25
+
26
+
think.utils._journal_path_cache = None
27
+
28
+
targets = [
29
+
journal / "chronicle/20240101/test/120000_300/mic_audio.npz",
30
+
journal / "chronicle/20240101/test/120000_300/talents/speaker_labels.json",
31
+
journal / "chronicle/20240101/test/120000_300/talents/speaker_corrections.json",
32
+
journal / "entities/alice/voiceprints.npz",
33
+
journal / "entities/alice/owner_centroid.npz",
34
+
journal / "entities/bob/owner_centroid.npz",
35
+
journal / "awareness/owner_candidate.npz",
36
+
]
37
+
38
+
for path in targets:
39
+
path.parent.mkdir(parents=True, exist_ok=True)
40
+
path.write_bytes(FILE_BYTES)
41
+
42
+
yield journal, targets
43
+
think.utils._journal_path_cache = None
44
+
45
+
46
+
def test_wipe_dry_run(wipe_journal: tuple[Path, list[Path]]) -> None:
47
+
_, targets = wipe_journal
48
+
49
+
report = wipe_speaker_artifacts(dry_run=True)
50
+
51
+
assert report.segment_embeddings.count == 1
52
+
assert report.speaker_labels.count == 1
53
+
assert report.speaker_corrections.count == 1
54
+
assert report.entity_voiceprints.count == 1
55
+
assert report.owner_centroids.count == 2
56
+
assert report.owner_candidate.count == 1
57
+
assert report.total_files == 7
58
+
assert report.total_bytes == 7 * len(FILE_BYTES)
59
+
60
+
for path in targets:
61
+
assert path.exists()
62
+
63
+
64
+
def test_wipe_commit(wipe_journal: tuple[Path, list[Path]]) -> None:
65
+
_, targets = wipe_journal
66
+
67
+
report = wipe_speaker_artifacts(dry_run=False)
68
+
69
+
assert report.segment_embeddings.count == 1
70
+
assert report.speaker_labels.count == 1
71
+
assert report.speaker_corrections.count == 1
72
+
assert report.entity_voiceprints.count == 1
73
+
assert report.owner_centroids.count == 2
74
+
assert report.owner_candidate.count == 1
75
+
assert report.total_files == 7
76
+
assert report.total_bytes == 7 * len(FILE_BYTES)
77
+
78
+
for path in targets:
79
+
assert not path.exists()
+76
apps/speakers/wipe.py
+76
apps/speakers/wipe.py
···
1
+
# SPDX-License-Identifier: AGPL-3.0-only
2
+
# Copyright (c) 2026 sol pbc
3
+
4
+
"""Wipe legacy speaker artifacts from the journal."""
5
+
6
+
from __future__ import annotations
7
+
8
+
from dataclasses import asdict, dataclass, field
9
+
from pathlib import Path
10
+
11
+
from think.utils import get_journal
12
+
13
+
14
+
@dataclass
15
+
class WipeCategory:
16
+
count: int = 0
17
+
bytes: int = 0
18
+
paths: list[str] = field(default_factory=list)
19
+
20
+
21
+
@dataclass
22
+
class WipeReport:
23
+
segment_embeddings: WipeCategory = field(default_factory=WipeCategory)
24
+
speaker_labels: WipeCategory = field(default_factory=WipeCategory)
25
+
speaker_corrections: WipeCategory = field(default_factory=WipeCategory)
26
+
entity_voiceprints: WipeCategory = field(default_factory=WipeCategory)
27
+
owner_centroids: WipeCategory = field(default_factory=WipeCategory)
28
+
owner_candidate: WipeCategory = field(default_factory=WipeCategory)
29
+
total_files: int = 0
30
+
total_bytes: int = 0
31
+
32
+
def to_dict(self) -> dict:
33
+
return asdict(self)
34
+
35
+
36
+
def wipe_speaker_artifacts(dry_run: bool = True) -> WipeReport:
37
+
"""Remove legacy speaker artifacts from the journal."""
38
+
journal = Path(get_journal())
39
+
report = WipeReport()
40
+
41
+
categories: list[tuple[WipeCategory, list[Path]]] = [
42
+
(report.segment_embeddings, sorted(journal.glob("chronicle/*/*/*/*.npz"))),
43
+
(
44
+
report.speaker_labels,
45
+
sorted(journal.glob("chronicle/*/*/*/talents/speaker_labels.json")),
46
+
),
47
+
(
48
+
report.speaker_corrections,
49
+
sorted(journal.glob("chronicle/*/*/*/talents/speaker_corrections.json")),
50
+
),
51
+
(report.entity_voiceprints, sorted(journal.glob("entities/*/voiceprints.npz"))),
52
+
(report.owner_centroids, sorted(journal.glob("entities/*/owner_centroid.npz"))),
53
+
(report.owner_candidate, [journal / "awareness" / "owner_candidate.npz"]),
54
+
]
55
+
56
+
for category, paths in categories:
57
+
for path in paths:
58
+
if not path.is_file():
59
+
continue
60
+
category.count += 1
61
+
category.bytes += path.stat().st_size
62
+
category.paths.append(path.relative_to(journal).as_posix())
63
+
if not dry_run:
64
+
path.unlink()
65
+
66
+
tracked_categories = [
67
+
report.segment_embeddings,
68
+
report.speaker_labels,
69
+
report.speaker_corrections,
70
+
report.entity_voiceprints,
71
+
report.owner_centroids,
72
+
report.owner_candidate,
73
+
]
74
+
report.total_files = sum(category.count for category in tracked_categories)
75
+
report.total_bytes = sum(category.bytes for category in tracked_categories)
76
+
return report