personal memory agent
providers+logs: cap Gemini output tokens at 65536, fix sol health logs chronicle path
- Lower default `thinking_budget` in `think/talents.py` from `8192 * 3` to `8192 * 2` (16384); new default total (16384 + 49152) equals Gemini's 65536 cap.
- Clamp outgoing `max_output_tokens` in `think/providers/google.py::_build_generate_config` to `<= GEMINI_MAX_OUTPUT_TOKENS` (65536) with a WARNING log; unit-tested at boundary and above.
- Drop `thinking_budget` / `max_output_tokens` frontmatter overrides from `talent/sense.md` so it inherits the new defaults.
- Fix `think/logs_cli.py::get_today_health_dir` to read from `journal/chronicle/<day>/health/` (missed during the 173c1773 chronicle rename); updated `tests/test_logs_cli.py::make_journal` fixture to match. Regenerated `tests/baselines/api/stats/stats.json` after the sense.md change.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
+71
-11
-2
talent/sense.md
-2
talent/sense.md
-2
tests/baselines/api/stats/stats.json
-2
tests/baselines/api/stats/stats.json
···
244
244
"talents": false,
245
245
"transcripts": true
246
246
},
247
-
"max_output_tokens": 4096,
248
247
"mtime": 0,
249
248
"output": "json",
250
249
"path": "<PROJECT>/talent/sense.md",
···
252
251
"schedule": "segment",
253
252
"schema": "sense.schema.json",
254
253
"source": "system",
255
-
"thinking_budget": 4096,
256
254
"tier": 3,
257
255
"title": "Segment Sense",
258
256
"type": "generate"
+2
-2
tests/test_logs_cli.py
+2
-2
tests/test_logs_cli.py
···
10
10
11
11
def make_journal(tmp_path, day, services, supervisor_lines=None):
12
12
"""Create a synthetic journal with health logs."""
13
-
health_dir = tmp_path / day / "health"
13
+
health_dir = tmp_path / "chronicle" / day / "health"
14
14
health_dir.mkdir(parents=True)
15
15
16
16
for name, lines in services.items():
···
25
25
26
26
for name in services:
27
27
journal_sym = journal_health / f"{name}.log"
28
-
journal_sym.symlink_to(f"../{day}/health/ref_{name}.log")
28
+
journal_sym.symlink_to(f"../chronicle/{day}/health/ref_{name}.log")
29
29
30
30
if supervisor_lines is not None:
31
31
sup = journal_health / "supervisor.log"
+52
tests/test_providers_google.py
+52
tests/test_providers_google.py
···
1
+
# SPDX-License-Identifier: AGPL-3.0-only
2
+
# Copyright (c) 2026 sol pbc
3
+
4
+
import logging
5
+
6
+
7
+
def test_build_generate_config_passes_through_at_cap(caplog):
8
+
from think.providers.google import GEMINI_MAX_OUTPUT_TOKENS, _build_generate_config
9
+
10
+
with caplog.at_level(logging.WARNING, logger="think.providers.google"):
11
+
config = _build_generate_config(
12
+
temperature=0.3,
13
+
max_output_tokens=49152,
14
+
system_instruction=None,
15
+
json_output=False,
16
+
thinking_budget=16384,
17
+
)
18
+
19
+
warnings = [
20
+
record
21
+
for record in caplog.records
22
+
if record.name == "think.providers.google" and record.levelno == logging.WARNING
23
+
]
24
+
assert config.max_output_tokens == GEMINI_MAX_OUTPUT_TOKENS
25
+
assert config.thinking_config.thinking_budget == 16384
26
+
assert warnings == []
27
+
28
+
29
+
def test_build_generate_config_clamps_and_warns_once(caplog):
30
+
from think.providers.google import GEMINI_MAX_OUTPUT_TOKENS, _build_generate_config
31
+
32
+
with caplog.at_level(logging.WARNING, logger="think.providers.google"):
33
+
config = _build_generate_config(
34
+
temperature=0.3,
35
+
max_output_tokens=49152,
36
+
system_instruction=None,
37
+
json_output=False,
38
+
thinking_budget=24576,
39
+
)
40
+
41
+
warnings = [
42
+
record
43
+
for record in caplog.records
44
+
if record.name == "think.providers.google" and record.levelno == logging.WARNING
45
+
]
46
+
assert config.max_output_tokens <= GEMINI_MAX_OUTPUT_TOKENS
47
+
assert config.max_output_tokens == GEMINI_MAX_OUTPUT_TOKENS
48
+
assert config.thinking_config.thinking_budget == 16384
49
+
assert len(warnings) == 1
50
+
assert "max_output_tokens=49152" in warnings[0].message
51
+
assert "thinking_budget=24576" in warnings[0].message
52
+
assert "clamped_thinking_budget=16384" in warnings[0].message
+2
-4
think/logs_cli.py
+2
-4
think/logs_cli.py
···
23
23
from pathlib import Path
24
24
from typing import NamedTuple
25
25
26
-
from think.utils import get_journal, setup_cli
26
+
from think.utils import day_path, get_journal, setup_cli
27
27
28
28
_DIM = "\033[2m"
29
29
_RESET = "\033[0m"
···
108
108
109
109
110
110
def get_today_health_dir() -> Path | None:
111
-
journal = Path(os.path.expanduser(get_journal()))
112
-
today = datetime.now().strftime("%Y%m%d")
113
-
health_dir = journal / today / "health"
111
+
health_dir = day_path(create=False) / "health"
114
112
return health_dir if health_dir.is_dir() else None
115
113
116
114
+14
think/providers/google.py
+14
think/providers/google.py
···
56
56
safe_raw,
57
57
)
58
58
59
+
GEMINI_MAX_OUTPUT_TOKENS = 65536
59
60
_DEFAULT_MAX_TOKENS = 8192
60
61
_DEFAULT_MODEL = GEMINI_FLASH
61
62
···
244
245
"""
245
246
# Compute total tokens: output + thinking budget
246
247
total_tokens = max_output_tokens + (thinking_budget or 0)
248
+
if total_tokens > GEMINI_MAX_OUTPUT_TOKENS:
249
+
clamped_max_output = min(max_output_tokens, GEMINI_MAX_OUTPUT_TOKENS)
250
+
clamped_thinking = max(0, GEMINI_MAX_OUTPUT_TOKENS - clamped_max_output)
251
+
logging.getLogger(__name__).warning(
252
+
"Clamping Gemini token budget: max_output_tokens=%s thinking_budget=%s "
253
+
"clamped_max_output_tokens=%s clamped_thinking_budget=%s",
254
+
max_output_tokens,
255
+
thinking_budget,
256
+
clamped_max_output,
257
+
clamped_thinking,
258
+
)
259
+
thinking_budget = clamped_thinking
260
+
total_tokens = clamped_max_output + clamped_thinking
247
261
248
262
config_args: dict[str, Any] = {
249
263
"temperature": temperature,
+1
-1
think/talents.py
+1
-1
think/talents.py
···
938
938
output_format = config.get("output")
939
939
940
940
# Get generation parameters from config (set in frontmatter)
941
-
thinking_budget = config.get("thinking_budget") or 8192 * 3
941
+
thinking_budget = config.get("thinking_budget") or 8192 * 2
942
942
max_output_tokens = config.get("max_output_tokens") or 8192 * 6
943
943
is_json_output = output_format == "json"
944
944