personal memory agent
feat(sense): close content_type enum and hydrate facet enum from runtime
Tighten talent/sense.schema.json so Gemini's response_json_schema
enforcement closes two surfaces that previously accepted garbage:
- content_type: replace 8-value enum (which included `mixed`) with the
17-value enum derived from think.activities.DEFAULT_ACTIVITIES minus
the schedule-only ids, plus the `idle` sentinel. A drift-detector
test in tests/test_sense_schema.py compares the on-disk enum to the
derivation, so adding/removing an activity will fail loudly.
- facets[].facet: replace the open `{type: string, minLength: 1}` with
a sentinel enum `[__RUNTIME_FACETS__]` that is hydrated per-call in
think.talents._execute_generate from think.facets.get_facets(),
filtered by the slug regex `^[a-z][a-z0-9_-]*$`. When the journal
has no valid facets the sentinel falls back to {type: string,
minLength: 1} so the schema remains satisfiable.
Hydration is a deep-copy, idempotent, read-only transform that lives
in think.talent. The on-disk schema and `get_talent("sense")["json_schema"]`
both still equal the raw file (existing
test_sense_loaded_json_schema_matches_on_disk_schema invariant holds);
hydration happens once before the primary provider call and is reused
for the fallback.
talent/sense.md is updated to match: the content_type field doc lists
all 17 values, the "Facet assignment rules" paragraph drops the
hardcoded pseudo-facets (`meetings`, `learning`, `technical-work`)
that were never real facet IDs, and Rule #5 stops referencing the
dropped `mixed` value.
No data migration. Existing on-disk records with content_type=mixed or
pseudo-facet entries remain valid documents; future runs simply emit
schema-conformant values.
+150
-15
+15
-9
talent/sense.md
+15
-9
talent/sense.md
···
38
38
39
39
### content_type
40
40
The dominant activity type observed:
41
-
- **meeting**: Multi-person discussion with turn-taking (video call, in-person meeting, phone call)
42
-
- **coding**: Writing or editing code, using a terminal, IDE, or code review tool
43
-
- **browsing**: Web browsing, reading articles, searching
41
+
- **meeting**: Video calls, in-person meetings, conferences with turn-taking
42
+
- **coding**: Writing or editing code, IDE work, code review, debugging
43
+
- **browsing**: Web browsing, research, reading articles online
44
44
- **email**: Reading or composing email
45
45
- **messaging**: Chat applications (Slack, Teams, Discord, iMessage)
46
+
- **ai_conversation**: Conversations with AI assistants (ChatGPT, Claude, Gemini)
47
+
- **writing**: Documents, notes, long-form writing
46
48
- **reading**: Focused reading of documents, PDFs, books
49
+
- **video**: Watching video or streaming content
50
+
- **gaming**: Playing games
51
+
- **social**: Social media browsing and interaction
52
+
- **planning**: Scheduling, calendar management, agenda setting
53
+
- **productivity**: Spreadsheets, slides, task management
54
+
- **terminal**: Command line / shell sessions
55
+
- **design**: Design tools, image editing
56
+
- **music**: Music listening
47
57
- **idle**: No meaningful activity
48
-
- **mixed**: Multiple distinct activity types with no clear dominant one
49
58
50
59
### activity_summary
51
60
Describe what $preferred did during this segment using action verbs. Be specific — name the tools, people, projects, and actions. Ban passive words: never use "reviewing", "monitoring", "tracking", "checking", "observing", "maintaining", "managing." Use instead: wrote, sent, discussed, created, switched to, typed, said, decided, asked, proposed.
···
80
89
- `activity`: 1-sentence description of what was observed for this facet
81
90
- `level`: "high" (primary focus), "medium" (significant), "low" (brief/peripheral)
82
91
83
-
**Facet assignment rules:**
84
-
- "meetings" — ONLY for live synchronous group meetings where $preferred is a participant. NOT for listening to lectures, podcasts, press conferences, or recorded media.
85
-
- "learning" — Educational content: lectures, tutorials, articles, research, Wikipedia, podcasts. Includes listening to press conferences or recorded briefings.
86
-
- "technical-work" — Hands-on technical tasks: coding, using developer tools, configuring software, installing extensions, product research for work purposes, technical document writing, online tool usage, price comparison shopping.
92
+
**Facet assignment rules:** Do not invent facet IDs that are not in the configured journal facet list. Empty array is acceptable when no configured facet matches.
87
93
88
94
### meeting_detected
89
95
`true` ONLY if you can identify distinct, named participants in a live multi-person interaction:
···
124
130
2. `entities` and `speakers` may be empty arrays `[]`.
125
131
3. `facets` may be empty array `[]` if no configured facets match.
126
132
4. Be precise with density — misclassifying active segments as idle is the worst error.
127
-
5. For content_type, choose the single best match. Use "mixed" sparingly — only when there are truly multiple equal activities.
133
+
5. For `content_type`, choose the single best match — the dominant activity in the segment. If two activities are roughly equal, pick the one with more durable continuation evidence (entities, repeated screen content); the `facets[]` array's `level` field already encodes secondary activity.
128
134
6. Activity summary must describe observable actions, not inferred states.
129
135
130
136
Return ONLY the JSON object, no other text or explanation.
+2
-2
talent/sense.schema.json
+2
-2
talent/sense.schema.json
···
5
5
"required": ["density","content_type","activity_summary","entities","facets","meeting_detected","speakers","recommend","emotional_register"],
6
6
"properties": {
7
7
"density": {"type": "string", "enum": ["active","low_change","idle"]},
8
-
"content_type": {"type": "string", "enum": ["meeting","coding","browsing","email","messaging","reading","idle","mixed"]},
8
+
"content_type": {"type": "string", "enum": ["meeting","coding","browsing","email","messaging","ai_conversation","writing","reading","video","gaming","social","planning","productivity","terminal","design","music","idle"]},
9
9
"activity_summary": {"type": "string", "minLength": 1},
10
10
"entities": {
11
11
"type": "array",
···
29
29
"additionalProperties": false,
30
30
"required": ["facet","activity","level"],
31
31
"properties": {
32
-
"facet": {"type": "string", "minLength": 1},
32
+
"facet": {"type": "string", "enum": ["__RUNTIME_FACETS__"]},
33
33
"activity": {"type": "string", "minLength": 1},
34
34
"level": {"type": "string", "enum": ["high","medium","low"]}
35
35
}
+72
-1
tests/test_sense_schema.py
+72
-1
tests/test_sense_schema.py
···
1
1
# SPDX-License-Identifier: AGPL-3.0-only
2
2
# Copyright (c) 2026 sol pbc
3
3
4
+
import copy
4
5
import json
5
6
from pathlib import Path
6
7
7
8
import frontmatter
9
+
from jsonschema import Draft202012Validator
8
10
9
-
from think.talent import get_talent
11
+
from think.activities import DEFAULT_ACTIVITIES
12
+
from think.talent import (
13
+
RUNTIME_FACETS_SENTINEL,
14
+
get_talent,
15
+
hydrate_runtime_enums,
16
+
)
10
17
11
18
SENSE_PATH = Path(__file__).resolve().parents[1] / "talent" / "sense.md"
12
19
SENSE_SCHEMA_PATH = SENSE_PATH.with_suffix(".schema.json")
···
51
58
on_disk = json.loads(SENSE_SCHEMA_PATH.read_text(encoding="utf-8"))
52
59
53
60
assert get_talent("sense")["json_schema"] == on_disk
61
+
62
+
63
+
def test_content_type_enum_matches_default_activities_drift_detector():
64
+
schedule_only = "Scheduled events emitted by talent/schedule.md"
65
+
expected = [
66
+
a["id"]
67
+
for a in DEFAULT_ACTIVITIES
68
+
if schedule_only not in a.get("instructions", "")
69
+
] + ["idle"]
70
+
schema = json.loads(SENSE_SCHEMA_PATH.read_text(encoding="utf-8"))
71
+
72
+
assert schema["properties"]["content_type"]["enum"] == expected
73
+
74
+
75
+
def test_sense_schema_facet_uses_runtime_sentinel_constant():
76
+
schema = json.loads(SENSE_SCHEMA_PATH.read_text(encoding="utf-8"))
77
+
facet_schema = schema["properties"]["facets"]["items"]["properties"]["facet"]
78
+
79
+
assert facet_schema["enum"] == [RUNTIME_FACETS_SENTINEL]
80
+
81
+
82
+
def test_hydrate_runtime_enums_replaces_facet_sentinel(monkeypatch):
83
+
monkeypatch.setattr(
84
+
"think.talent.get_facets",
85
+
lambda: {"alpha": {}, "Beta": {}, "weird,name": {}, "valid_one": {}},
86
+
)
87
+
schema = {
88
+
"properties": {"facet": {"type": "string", "enum": [RUNTIME_FACETS_SENTINEL]}}
89
+
}
90
+
91
+
hydrated = hydrate_runtime_enums(schema)
92
+
93
+
assert hydrated["properties"]["facet"]["enum"] == ["alpha", "valid_one"]
94
+
95
+
96
+
def test_hydrate_runtime_enums_empty_facets_fallback(monkeypatch):
97
+
monkeypatch.setattr("think.talent.get_facets", lambda: {})
98
+
schema = {
99
+
"type": "object",
100
+
"properties": {"facet": {"type": "string", "enum": [RUNTIME_FACETS_SENTINEL]}},
101
+
}
102
+
103
+
hydrated = hydrate_runtime_enums(schema)
104
+
facet_schema = hydrated["properties"]["facet"]
105
+
106
+
assert facet_schema == {"type": "string", "minLength": 1}
107
+
Draft202012Validator.check_schema(hydrated)
108
+
109
+
110
+
def test_hydrate_runtime_enums_idempotent_and_pure(monkeypatch):
111
+
monkeypatch.setattr("think.talent.get_facets", lambda: {})
112
+
original = {"type": "object", "properties": {"x": {"type": "string"}}}
113
+
saved_copy = copy.deepcopy(original)
114
+
115
+
hydrated = hydrate_runtime_enums(original)
116
+
117
+
assert hydrated == original
118
+
assert original == saved_copy
119
+
assert hydrated is not original
120
+
assert hydrate_runtime_enums(hydrated) == hydrated
121
+
122
+
123
+
def test_hydrate_runtime_enums_none_passthrough():
124
+
assert hydrate_runtime_enums(None) is None
54
125
55
126
56
127
def test_role_and_source_do_not_leak_into_other_sense_sections():
+57
think/talent.py
+57
think/talent.py
···
17
17
18
18
from __future__ import annotations
19
19
20
+
import copy
20
21
import importlib.util
21
22
import json
23
+
import logging
22
24
import os
25
+
import re
23
26
from pathlib import Path
24
27
from typing import Any, Callable
25
28
26
29
import frontmatter
27
30
from jsonschema import Draft202012Validator, SchemaError
28
31
32
+
from think.facets import get_facets
33
+
29
34
# Import core prompt utilities from think.prompts
30
35
from think.prompts import _load_prompt_metadata, load_prompt
31
36
···
35
40
36
41
TALENT_DIR = Path(__file__).parent.parent / "talent"
37
42
APPS_DIR = Path(__file__).parent.parent / "apps"
43
+
RUNTIME_FACETS_SENTINEL = "__RUNTIME_FACETS__"
44
+
SLUG_RE = re.compile(r"^[a-z][a-z0-9_-]*$")
45
+
LOG = logging.getLogger(__name__)
38
46
39
47
40
48
# ---------------------------------------------------------------------------
···
438
446
if value is False:
439
447
return {} # No talents
440
448
return None # All talents (True or "required")
449
+
450
+
451
+
def _valid_runtime_facets() -> list[str]:
452
+
"""Return sorted list of facet directory names matching SLUG_RE."""
453
+
return sorted(slug for slug in get_facets() if SLUG_RE.fullmatch(slug))
454
+
455
+
456
+
def hydrate_runtime_enums(schema: Any) -> Any:
457
+
"""Replace runtime sentinels in schema enums with current journal state.
458
+
459
+
Walks the schema; wherever an `enum` is exactly [RUNTIME_FACETS_SENTINEL],
460
+
replaces it with the sorted list of valid runtime facet slugs. If no
461
+
valid facets exist, drops the `enum` key and sets `minLength: 1` on
462
+
that node so the schema remains satisfiable.
463
+
464
+
Returns None when given None. Deep-copies non-None input. Idempotent
465
+
for already-hydrated schemas (sentinel is gone after first call).
466
+
"""
467
+
if schema is None:
468
+
return None
469
+
470
+
hydrated = copy.deepcopy(schema)
471
+
facets = _valid_runtime_facets()
472
+
used_empty_fallback = False
473
+
474
+
def _walk(node: Any) -> None:
475
+
nonlocal used_empty_fallback
476
+
if isinstance(node, dict):
477
+
if node.get("enum") == [RUNTIME_FACETS_SENTINEL]:
478
+
if facets:
479
+
node["enum"] = list(facets)
480
+
else:
481
+
node.pop("enum", None)
482
+
node["minLength"] = 1
483
+
used_empty_fallback = True
484
+
for value in node.values():
485
+
_walk(value)
486
+
elif isinstance(node, list):
487
+
for item in node:
488
+
_walk(item)
489
+
490
+
_walk(hydrated)
491
+
492
+
if used_empty_fallback:
493
+
LOG.info(
494
+
"hydrate_runtime_enums: no valid runtime facets; using minLength fallback"
495
+
)
496
+
497
+
return hydrated
441
498
442
499
443
500
# ---------------------------------------------------------------------------
+4
-3
think/talents.py
+4
-3
think/talents.py
···
975
975
emit_event: Event emission callback
976
976
"""
977
977
from think.models import generate_with_result
978
-
from think.talent import key_to_context
978
+
from think.talent import hydrate_runtime_enums, key_to_context
979
979
980
980
name = config["name"]
981
981
messages = config.get("messages")
···
1014
1014
contents = ["No input provided."]
1015
1015
1016
1016
context = key_to_context(name)
1017
+
runtime_json_schema = hydrate_runtime_enums(config.get("json_schema"))
1017
1018
try:
1018
1019
gen_result = generate_with_result(
1019
1020
contents=contents,
···
1023
1024
thinking_budget=thinking_budget,
1024
1025
system_instruction=system_instruction,
1025
1026
json_output=is_json_output,
1026
-
json_schema=config.get("json_schema"),
1027
+
json_schema=runtime_json_schema,
1027
1028
timeout_s=timeout_s,
1028
1029
)
1029
1030
except Exception as exc:
···
1069
1070
thinking_budget=thinking_budget,
1070
1071
system_instruction=system_instruction,
1071
1072
json_output=is_json_output,
1072
-
json_schema=config.get("json_schema"),
1073
+
json_schema=runtime_json_schema,
1073
1074
timeout_s=timeout_s,
1074
1075
provider=backup,
1075
1076
model=backup_model,