feat(realism): operator-tunable planner weights via realism_config

New realism_config table (uuid PK + unique key) + two repo methods
(get/set) backs an admin-only GET/PUT /api/v1/realism/config surface.

The planner now exposes apply_payload(payload) / current_payload() /
reset_to_defaults() and reads its weights through mutable module
globals; pick() resolves the live values each call. Validation
catches negative weights, zero totals, out-of-range canary_probability,
unknown content_class names, and silently drops cross-list entries
(canary class on the user list, etc).

The orchestrator worker calls _refresh_realism_config(repo) on
startup and every 5 ticks (~5min at 60s interval). Operator changes
land within one refresh window with no bus signal — the simpler path
for a knob whose latency tolerance is minutes.
This commit is contained in:
2026-04-27 18:00:08 -04:00
parent da3c35c6a4
commit 2cc60bd677
12 changed files with 711 additions and 9 deletions

View File

@@ -39,6 +39,7 @@ from decnet.orchestrator.emailgen import (
scheduler as email_scheduler,
)
from decnet.orchestrator.emailgen.scheduler import EmailAction
from decnet.realism import planner as realism_planner
from decnet.realism.llm.circuit import LLMCircuitBreaker
from decnet.web.db.repository import BaseRepository
@@ -52,6 +53,12 @@ _PRUNE_EVERY_TICKS = 100
_PRUNE_PER_DST_CAP = 10000
_PRUNE_PER_MAIL_DECKY_CAP = 5000
# Refresh planner weights from realism_config every N ticks. Operator
# tunables drift slowly; ~minute-scale latency between PUT and effect
# is fine. No bus signal — keeps the path simple and the orchestrator
# self-contained.
_REALISM_CONFIG_REFRESH_TICKS = 5
# Action-kind weights for the per-tick roll. Email is rare because
# each LLM round-trip is expensive (~seconds) and the prior emailgen
# worker only ticked every 5 minutes. At a 60s orchestrator interval,
@@ -115,6 +122,12 @@ async def orchestrator_worker(
)
bus = None
# Initial load — pulls the operator-tuned weights from
# realism_config so the orchestrator starts ticking with the
# operator's intent rather than the baked-in defaults. A failure
# here logs and falls through; the planner already holds defaults.
await _refresh_realism_config(repo)
shutdown = asyncio.Event()
heartbeat_task = asyncio.create_task(
run_health_heartbeat(
@@ -141,6 +154,8 @@ async def orchestrator_worker(
tick_n += 1
if tick_n % _PRUNE_EVERY_TICKS == 0:
await _periodic_prune(repo)
if tick_n % _REALISM_CONFIG_REFRESH_TICKS == 0:
await _refresh_realism_config(repo)
finally:
for t in (heartbeat_task, control_task):
t.cancel()
@@ -174,6 +189,35 @@ async def _periodic_prune(repo: BaseRepository) -> None:
logger.error("orchestrator emails prune failed: %s", exc)
async def _refresh_realism_config(repo: BaseRepository) -> None:
"""Pull operator-tuned weights from realism_config into the planner.
Failure modes (DB unreachable, malformed JSON, validation reject)
log and leave the planner's current weights in place. The orchestrator
keeps ticking with whatever it had — never blocks on config.
"""
try:
row = await repo.get_realism_config("weights")
except Exception as exc: # noqa: BLE001
logger.warning("realism config refresh: DB read failed: %s", exc)
return
if row is None:
return # no overrides set; defaults stand
import json
try:
payload = json.loads(row.get("value") or "{}")
except json.JSONDecodeError as exc:
logger.warning("realism config refresh: malformed JSON: %s", exc)
return
if not isinstance(payload, dict):
logger.warning("realism config refresh: payload not an object")
return
try:
realism_planner.apply_payload(payload)
except ValueError as exc:
logger.warning("realism config refresh: rejected payload: %s", exc)
def _roll_action_kind(rng: secrets.SystemRandom) -> str:
total = sum(w for _, w in _ACTION_WEIGHTS)
target = rng.randint(1, total)

View File

@@ -29,27 +29,31 @@ from decnet.realism.personas import EmailPersona
from decnet.realism.taxonomy import ContentClass, Plan, PlanAction # noqa: F401
# Stage-3 weighted sampling:
# Stage-3 weighted sampling defaults:
# * User content (notes/todo/draft/script) gets the bulk — those are
# the realism win when a persona "looks busy."
# * System content (cron/daemon/cache) is plausible filler.
# * Email + canary are owned by other paths and not picked here.
_USER_CLASS_WEIGHTS: tuple[tuple[ContentClass, int], ...] = (
# * Canary classes are picked rarely. Each plant materialises a real
# CanaryToken row + DNS slug + HTTP URL — flooding the fleet makes
# the dashboard noisy. ~3% of file picks land here.
#
# These are the *defaults*. Operator-tuned overrides arrive via
# :func:`apply_payload` (admin PUT /api/v1/realism/config). The
# orchestrator worker periodically refreshes the in-process state from
# the ``realism_config`` table; pick() reads the live globals each call.
_DEFAULT_USER_CLASS_WEIGHTS: tuple[tuple[ContentClass, int], ...] = (
(ContentClass.NOTE, 30),
(ContentClass.TODO, 20),
(ContentClass.DRAFT, 15),
(ContentClass.SCRIPT, 10),
)
_SYSTEM_CLASS_WEIGHTS: tuple[tuple[ContentClass, int], ...] = (
_DEFAULT_SYSTEM_CLASS_WEIGHTS: tuple[tuple[ContentClass, int], ...] = (
(ContentClass.LOG_CRON, 12),
(ContentClass.LOG_DAEMON, 8),
(ContentClass.CACHE_TMP, 5),
)
# Canary classes are picked rarely. Each plant materialises a real
# CanaryToken row + DNS slug + HTTP URL — flooding the fleet with
# canaries makes the dashboard noisy and the per-decky alert surface
# explode. ~3% of file picks land here.
_CANARY_CLASS_WEIGHTS: tuple[tuple[ContentClass, int], ...] = (
_DEFAULT_CANARY_CLASS_WEIGHTS: tuple[tuple[ContentClass, int], ...] = (
(ContentClass.CANARY_AWS_CREDS, 1),
(ContentClass.CANARY_ENV_FILE, 1),
(ContentClass.CANARY_GIT_CONFIG, 1),
@@ -59,7 +63,139 @@ _CANARY_CLASS_WEIGHTS: tuple[tuple[ContentClass, int], ...] = (
(ContentClass.CANARY_HONEYDOC_PDF, 1),
(ContentClass.CANARY_MYSQL_DUMP, 1),
)
_CANARY_PROBABILITY = 0.03
_DEFAULT_CANARY_PROBABILITY = 0.03
# Live (mutable) globals — reassigned by :func:`apply_payload`. pick()
# reads these. Reset to defaults via :func:`reset_to_defaults` (used by
# tests + the API DELETE path).
_USER_CLASS_WEIGHTS: tuple[tuple[ContentClass, int], ...] = _DEFAULT_USER_CLASS_WEIGHTS
_SYSTEM_CLASS_WEIGHTS: tuple[tuple[ContentClass, int], ...] = _DEFAULT_SYSTEM_CLASS_WEIGHTS
_CANARY_CLASS_WEIGHTS: tuple[tuple[ContentClass, int], ...] = _DEFAULT_CANARY_CLASS_WEIGHTS
_CANARY_PROBABILITY: float = _DEFAULT_CANARY_PROBABILITY
def _serialize_weights(
weights: tuple[tuple[ContentClass, int], ...],
) -> list[dict[str, Any]]:
return [{"content_class": cls.value, "weight": w} for cls, w in weights]
def _parse_weights(
raw: Any, allowed: set[ContentClass],
) -> tuple[tuple[ContentClass, int], ...]:
"""Parse ``[{"content_class": "...", "weight": N}, ...]`` into the
planner's internal tuple shape. Drops entries whose ``content_class``
isn't in *allowed* (defends against an operator pasting in a canary
class on the user list, which would skew sampling without the
canary-probability gate).
Raises ``ValueError`` on structural problems (non-list, non-int
weight, negative weight, empty result) so the API can return 400.
"""
if not isinstance(raw, list):
raise ValueError("weights must be a list")
out: list[tuple[ContentClass, int]] = []
for entry in raw:
if not isinstance(entry, dict):
raise ValueError("each weight entry must be an object")
cls_name = entry.get("content_class")
weight = entry.get("weight")
if not isinstance(weight, int) or weight < 0:
raise ValueError(
f"weight for {cls_name!r} must be a non-negative integer"
)
try:
cls = ContentClass(cls_name)
except (ValueError, TypeError):
raise ValueError(f"unknown content_class: {cls_name!r}")
if cls not in allowed:
# Silently drop — a class that doesn't belong on this list
# (e.g. a canary class on the user list) is operator error,
# but we don't want to fail the whole save over one stray
# entry. The roundtrip in current_payload() will show the
# operator their entry didn't land.
continue
out.append((cls, weight))
if not out:
raise ValueError("weights list resolved to zero valid entries")
if sum(w for _, w in out) <= 0:
raise ValueError("weights must sum to a positive number")
return tuple(out)
_USER_CLASSES: set[ContentClass] = {
ContentClass.NOTE, ContentClass.TODO, ContentClass.DRAFT, ContentClass.SCRIPT,
}
_SYSTEM_CLASSES: set[ContentClass] = {
ContentClass.LOG_CRON, ContentClass.LOG_DAEMON, ContentClass.CACHE_TMP,
}
_CANARY_CLASSES: set[ContentClass] = {
ContentClass.CANARY_AWS_CREDS, ContentClass.CANARY_ENV_FILE,
ContentClass.CANARY_GIT_CONFIG, ContentClass.CANARY_SSH_KEY,
ContentClass.CANARY_HONEYDOC, ContentClass.CANARY_HONEYDOC_DOCX,
ContentClass.CANARY_HONEYDOC_PDF, ContentClass.CANARY_MYSQL_DUMP,
}
def current_payload() -> dict[str, Any]:
"""Export the live planner config as a JSON-safe dict.
Wire shape returned by ``GET /api/v1/realism/config``."""
return {
"user_class_weights": _serialize_weights(_USER_CLASS_WEIGHTS),
"system_class_weights": _serialize_weights(_SYSTEM_CLASS_WEIGHTS),
"canary_class_weights": _serialize_weights(_CANARY_CLASS_WEIGHTS),
"canary_probability": _CANARY_PROBABILITY,
}
def apply_payload(payload: dict[str, Any]) -> None:
"""Override the planner's live globals from a wire payload.
Validates structurally and rebinds module-level names atomically
per field — partial failures don't leave the planner in a torn
state because validation happens before any rebind.
Unknown fields are ignored (forward-compat); fields not present
leave the corresponding global untouched."""
global _USER_CLASS_WEIGHTS, _SYSTEM_CLASS_WEIGHTS
global _CANARY_CLASS_WEIGHTS, _CANARY_PROBABILITY
new_user = _USER_CLASS_WEIGHTS
new_system = _SYSTEM_CLASS_WEIGHTS
new_canary = _CANARY_CLASS_WEIGHTS
new_prob = _CANARY_PROBABILITY
if "user_class_weights" in payload:
new_user = _parse_weights(payload["user_class_weights"], _USER_CLASSES)
if "system_class_weights" in payload:
new_system = _parse_weights(
payload["system_class_weights"], _SYSTEM_CLASSES,
)
if "canary_class_weights" in payload:
new_canary = _parse_weights(
payload["canary_class_weights"], _CANARY_CLASSES,
)
if "canary_probability" in payload:
prob = payload["canary_probability"]
if not isinstance(prob, (int, float)) or not (0.0 <= prob <= 1.0):
raise ValueError("canary_probability must be in [0.0, 1.0]")
new_prob = float(prob)
_USER_CLASS_WEIGHTS = new_user
_SYSTEM_CLASS_WEIGHTS = new_system
_CANARY_CLASS_WEIGHTS = new_canary
_CANARY_PROBABILITY = new_prob
def reset_to_defaults() -> None:
"""Restore hardcoded defaults. Used by tests and the API reset path."""
global _USER_CLASS_WEIGHTS, _SYSTEM_CLASS_WEIGHTS
global _CANARY_CLASS_WEIGHTS, _CANARY_PROBABILITY
_USER_CLASS_WEIGHTS = _DEFAULT_USER_CLASS_WEIGHTS
_SYSTEM_CLASS_WEIGHTS = _DEFAULT_SYSTEM_CLASS_WEIGHTS
_CANARY_CLASS_WEIGHTS = _DEFAULT_CANARY_CLASS_WEIGHTS
_CANARY_PROBABILITY = _DEFAULT_CANARY_PROBABILITY
def _weighted_pick(

View File

@@ -78,6 +78,7 @@ from .orchestrator import (
OrchestratorEventsResponse,
)
from .realism import (
RealismConfig,
SyntheticFile,
SyntheticFilesResponse,
)
@@ -231,6 +232,7 @@ __all__ = [
"OrchestratorEvent",
"OrchestratorEventsResponse",
# realism
"RealismConfig",
"SyntheticFile",
"SyntheticFilesResponse",
# logs

View File

@@ -83,3 +83,25 @@ class SyntheticFilesResponse(BaseModel):
limit: int
offset: int
data: List[dict[str, Any]]
class RealismConfig(SQLModel, table=True):
"""Operator-tunable realism knobs.
Single-row-per-key schema: each row carries one piece of operator
config (today: ``key="weights"`` → JSON encoding the planner's
user/system/canary weights and canary probability). The planner
reads in-memory module globals; the orchestrator worker refreshes
those globals from this table on a periodic tick.
UUID PK + unique key per ``feedback_uuid_over_natural_keys.md``.
"""
__tablename__ = "realism_config"
uuid: str = Field(default_factory=lambda: str(uuid4()), primary_key=True)
key: str = Field(max_length=64, unique=True, index=True)
value: str = Field(
sa_column=Column("value", Text, nullable=False, default="{}"),
)
updated_at: datetime = Field(
default_factory=lambda: datetime.now(timezone.utc),
)

View File

@@ -1159,6 +1159,28 @@ class BaseRepository(ABC):
"""Single synthetic_files row by uuid, or ``None``."""
raise NotImplementedError
async def get_realism_config(
self, key: str,
) -> Optional[dict[str, Any]]:
"""Read one ``realism_config`` row by key.
Today only ``key="weights"`` is used; the schema is
single-row-per-key so future tunables can land without a new
table. Returns ``None`` when the key has never been set —
callers fall back to hardcoded defaults in
:mod:`decnet.realism.planner`.
"""
raise NotImplementedError
async def set_realism_config(
self, key: str, value: str,
) -> None:
"""Upsert one ``realism_config`` row. Last-write-wins.
*value* is opaque JSON text; validation belongs to the API
layer (the planner only reads what landed)."""
raise NotImplementedError
async def pick_random_synthetic_file_for_edit(
self,
decky_uuid: str,

View File

@@ -53,6 +53,7 @@ from decnet.web.db.models import (
TopologyMutation,
OrchestratorEmail,
OrchestratorEvent,
RealismConfig,
SyntheticFile,
WebhookSubscription,
CanaryBlob,
@@ -3415,6 +3416,39 @@ class SQLModelRepository(BaseRepository):
return None
return row.model_dump(mode="json")
async def get_realism_config(
self, key: str,
) -> Optional[dict[str, Any]]:
async with self._session() as session:
stmt = select(RealismConfig).where(RealismConfig.key == key)
result = await session.execute(stmt)
row = result.scalars().first()
if row is None:
return None
return row.model_dump(mode="json")
async def set_realism_config(
self, key: str, value: str,
) -> None:
"""Upsert one realism_config row. Last-write-wins."""
async with self._session() as session:
stmt = select(RealismConfig).where(RealismConfig.key == key)
result = await session.execute(stmt)
row = result.scalars().first()
if row is None:
session.add(RealismConfig(
key=key, value=value,
updated_at=datetime.now(timezone.utc),
))
else:
upd = (
update(RealismConfig)
.where(RealismConfig.uuid == row.uuid)
.values(value=value, updated_at=datetime.now(timezone.utc))
)
await session.execute(upd)
await session.commit()
async def pick_random_synthetic_file_for_edit(
self,
decky_uuid: str,

View File

@@ -31,6 +31,7 @@ from .campaigns.api_list_campaign_identities import router as campaign_identitie
from .campaigns.api_events import router as campaign_events_router
from .orchestrator.api_list_events import router as orchestrator_list_router
from .orchestrator.api_events import router as orchestrator_events_router
from .realism.api_config import router as realism_config_router
from .realism.api_personas import router as realism_personas_router
from .realism.api_synthetic_files import router as realism_synthetic_files_router
from .transcripts import transcripts_router
@@ -117,6 +118,7 @@ api_router.include_router(orchestrator_events_router)
# on-disk JSON file directly (see decnet.realism.personas_pool).
api_router.include_router(realism_personas_router)
api_router.include_router(realism_synthetic_files_router)
api_router.include_router(realism_config_router)
# Observability
api_router.include_router(stats_router)

View File

@@ -0,0 +1,115 @@
"""GET/PUT ``/api/v1/realism/config`` — operator-tunable realism knobs.
Today only the planner's content-class weights + canary probability
are exposed. The wire shape mirrors what
:func:`decnet.realism.planner.current_payload` produces and
:func:`decnet.realism.planner.apply_payload` consumes.
Reads accept viewer; writes are admin (writes mutate sampling
behaviour across the whole orchestrator fleet, same trust level as
the persona-pool surface).
The orchestrator worker periodically re-loads from the
``realism_config`` table; the API process applies overrides locally
on PUT so the GET-after-PUT round-trip reflects the change without
waiting for the orchestrator's next refresh tick.
"""
from __future__ import annotations
import json
from typing import Any
from fastapi import APIRouter, Depends, HTTPException
from decnet.logging import get_logger
from decnet.realism import planner
from decnet.telemetry import traced as _traced
from decnet.web.dependencies import repo, require_admin, require_viewer
router = APIRouter()
log = get_logger("api.realism.config")
_CONFIG_KEY = "weights"
@router.get(
"/realism/config",
tags=["Realism"],
responses={
401: {"description": "Could not validate credentials"},
403: {"description": "Insufficient permissions"},
},
)
@_traced("api.realism.get_config")
async def get_config(
user: dict = Depends(require_viewer),
) -> dict[str, Any]:
"""Return the live planner config in this API process.
Note: the API process and the orchestrator worker each carry their
own in-memory copy of the planner config. After a fresh API
restart the ``realism_config`` row is loaded into this process the
first time GET is called; subsequent reads are local.
"""
# Lazy hydration — first call after restart pulls from DB so the
# admin sees what the orchestrator is actually using, not the
# baked-in defaults.
row = await repo.get_realism_config(_CONFIG_KEY)
if row is not None:
try:
stored = json.loads(row.get("value") or "{}")
if isinstance(stored, dict):
planner.apply_payload(stored)
except (json.JSONDecodeError, ValueError) as exc:
log.warning(
"api.realism.get_config: stored payload invalid, "
"serving defaults: %s", exc,
)
return planner.current_payload()
@router.put(
"/realism/config",
tags=["Realism"],
responses={
400: {"description": "Invalid config payload"},
401: {"description": "Could not validate credentials"},
403: {"description": "Insufficient permissions"},
},
)
@_traced("api.realism.put_config")
async def put_config(
body: dict[str, Any],
user: dict = Depends(require_admin),
) -> dict[str, Any]:
"""Replace (partial) planner config and persist to ``realism_config``.
Body shape (all fields optional — unset fields keep current value):
* ``user_class_weights``: ``[{"content_class": "note", "weight": 30}, ...]``
* ``system_class_weights``: same shape
* ``canary_class_weights``: same shape
* ``canary_probability``: float in [0.0, 1.0]
Validation: any structural failure raises 400 *before* the rebind,
so the live config never goes torn.
"""
if not isinstance(body, dict):
raise HTTPException(status_code=400, detail="body must be an object")
try:
planner.apply_payload(body)
except ValueError as exc:
raise HTTPException(status_code=400, detail=str(exc)) from exc
# Persist what the planner now reflects (keeps DB in sync with the
# in-memory state — partial bodies merge into prior config).
snapshot = planner.current_payload()
await repo.set_realism_config(_CONFIG_KEY, json.dumps(snapshot))
log.info(
"api.realism.put_config user=%s canary_probability=%.4f",
user.get("username", user.get("uuid")),
snapshot["canary_probability"],
)
return snapshot

View File

@@ -0,0 +1,109 @@
"""GET/PUT /api/v1/realism/config — operator-tunable weights."""
from __future__ import annotations
import json
from unittest.mock import AsyncMock, patch
import pytest
from fastapi import HTTPException
from decnet.realism import planner
@pytest.fixture(autouse=True)
def _reset_planner():
yield
planner.reset_to_defaults()
@pytest.mark.asyncio
async def test_get_returns_defaults_when_no_row():
from decnet.web.router.realism.api_config import get_config
with patch("decnet.web.router.realism.api_config.repo") as mock_repo:
mock_repo.get_realism_config = AsyncMock(return_value=None)
result = await get_config(user={"uuid": "u", "role": "viewer"})
assert result["canary_probability"] == pytest.approx(0.03)
assert result["user_class_weights"]
@pytest.mark.asyncio
async def test_get_hydrates_from_db_row():
from decnet.web.router.realism.api_config import get_config
stored = json.dumps({"canary_probability": 0.10})
with patch("decnet.web.router.realism.api_config.repo") as mock_repo:
mock_repo.get_realism_config = AsyncMock(
return_value={"key": "weights", "value": stored},
)
result = await get_config(user={"uuid": "u", "role": "viewer"})
assert result["canary_probability"] == pytest.approx(0.10)
@pytest.mark.asyncio
async def test_get_serves_defaults_when_stored_payload_invalid():
"""Stored JSON parsed but failed planner validation: log + serve
defaults rather than 500."""
from decnet.web.router.realism.api_config import get_config
stored = json.dumps({"canary_probability": 9.0})
with patch("decnet.web.router.realism.api_config.repo") as mock_repo:
mock_repo.get_realism_config = AsyncMock(
return_value={"key": "weights", "value": stored},
)
result = await get_config(user={"uuid": "u", "role": "viewer"})
assert result["canary_probability"] == pytest.approx(0.03)
@pytest.mark.asyncio
async def test_put_persists_and_returns_snapshot():
from decnet.web.router.realism.api_config import put_config
with patch("decnet.web.router.realism.api_config.repo") as mock_repo:
mock_repo.set_realism_config = AsyncMock()
result = await put_config(
body={"canary_probability": 0.20},
user={"uuid": "u", "role": "admin", "username": "anti"},
)
assert result["canary_probability"] == pytest.approx(0.20)
mock_repo.set_realism_config.assert_awaited_once()
args, _ = mock_repo.set_realism_config.call_args
assert args[0] == "weights"
persisted = json.loads(args[1])
assert persisted["canary_probability"] == pytest.approx(0.20)
@pytest.mark.asyncio
async def test_put_returns_400_on_invalid_payload():
from decnet.web.router.realism.api_config import put_config
with patch("decnet.web.router.realism.api_config.repo") as mock_repo:
mock_repo.set_realism_config = AsyncMock()
with pytest.raises(HTTPException) as exc:
await put_config(
body={"canary_probability": 9.0},
user={"uuid": "u", "role": "admin", "username": "anti"},
)
assert exc.value.status_code == 400
# No DB write on validation failure.
mock_repo.set_realism_config.assert_not_called()
@pytest.mark.asyncio
async def test_put_rejects_non_dict_body():
from decnet.web.router.realism.api_config import put_config
with pytest.raises(HTTPException) as exc:
await put_config(
body=[1, 2, 3], # type: ignore[arg-type]
user={"uuid": "u", "role": "admin", "username": "anti"},
)
assert exc.value.status_code == 400

View File

@@ -0,0 +1,74 @@
"""The orchestrator pulls operator-tuned weights from realism_config.
§3c contract: the planner reads in-memory module globals, but the
operator's tuning lives in the DB (admin PUT /api/v1/realism/config).
The orchestrator worker bridges the two by calling
``_refresh_realism_config(repo)`` at startup and every Nth tick.
"""
from __future__ import annotations
import json
from unittest.mock import AsyncMock
import pytest
from decnet.orchestrator.worker import _refresh_realism_config
from decnet.realism import planner
@pytest.fixture(autouse=True)
def _reset_planner():
yield
planner.reset_to_defaults()
@pytest.mark.asyncio
async def test_refresh_no_row_keeps_defaults():
repo = AsyncMock()
repo.get_realism_config = AsyncMock(return_value=None)
await _refresh_realism_config(repo)
assert planner.current_payload()["canary_probability"] == pytest.approx(0.03)
@pytest.mark.asyncio
async def test_refresh_applies_stored_payload():
repo = AsyncMock()
repo.get_realism_config = AsyncMock(return_value={
"key": "weights",
"value": json.dumps({"canary_probability": 0.12}),
})
await _refresh_realism_config(repo)
assert planner.current_payload()["canary_probability"] == pytest.approx(0.12)
@pytest.mark.asyncio
async def test_refresh_swallows_db_error():
"""A wedged DB must not bring down the orchestrator's tick loop."""
repo = AsyncMock()
repo.get_realism_config = AsyncMock(side_effect=RuntimeError("boom"))
await _refresh_realism_config(repo) # does not raise
# planner unchanged
assert planner.current_payload()["canary_probability"] == pytest.approx(0.03)
@pytest.mark.asyncio
async def test_refresh_swallows_malformed_json():
repo = AsyncMock()
repo.get_realism_config = AsyncMock(return_value={
"key": "weights",
"value": "not-json",
})
await _refresh_realism_config(repo)
assert planner.current_payload()["canary_probability"] == pytest.approx(0.03)
@pytest.mark.asyncio
async def test_refresh_swallows_invalid_payload():
repo = AsyncMock()
repo.get_realism_config = AsyncMock(return_value={
"key": "weights",
"value": json.dumps({"canary_probability": 9.0}),
})
await _refresh_realism_config(repo)
# Planner config not corrupted by a bad refresh.
assert planner.current_payload()["canary_probability"] == pytest.approx(0.03)

View File

@@ -0,0 +1,119 @@
"""Operator-tunable planner knobs (apply_payload / current_payload).
§3c of the realism handoff: the planner reads mutable module globals
that an admin can override via PUT /api/v1/realism/config. These tests
pin the validation surface and the payload roundtrip so a regression
that breaks operator tunables surfaces here, not on a live fleet.
"""
from __future__ import annotations
import pytest
from decnet.realism import planner
from decnet.realism.taxonomy import ContentClass
@pytest.fixture(autouse=True)
def _reset_after_each_test():
yield
planner.reset_to_defaults()
def test_current_payload_returns_defaults_after_reset():
planner.reset_to_defaults()
payload = planner.current_payload()
assert payload["canary_probability"] == pytest.approx(0.03)
user = {e["content_class"]: e["weight"] for e in payload["user_class_weights"]}
assert user[ContentClass.NOTE.value] == 30
assert user[ContentClass.TODO.value] == 20
def test_apply_payload_overrides_user_weights():
planner.apply_payload({
"user_class_weights": [
{"content_class": "note", "weight": 5},
{"content_class": "todo", "weight": 95},
],
})
payload = planner.current_payload()
weights = {e["content_class"]: e["weight"] for e in payload["user_class_weights"]}
assert weights == {"note": 5, "todo": 95}
# System weights left untouched by a partial body.
assert payload["system_class_weights"]
def test_apply_payload_overrides_canary_probability():
planner.apply_payload({"canary_probability": 0.15})
assert planner.current_payload()["canary_probability"] == pytest.approx(0.15)
def test_apply_payload_rejects_bad_canary_probability():
with pytest.raises(ValueError, match="canary_probability"):
planner.apply_payload({"canary_probability": 1.5})
with pytest.raises(ValueError, match="canary_probability"):
planner.apply_payload({"canary_probability": -0.1})
with pytest.raises(ValueError, match="canary_probability"):
planner.apply_payload({"canary_probability": "high"})
def test_apply_payload_rejects_negative_weight():
with pytest.raises(ValueError, match="non-negative integer"):
planner.apply_payload({
"user_class_weights": [{"content_class": "note", "weight": -1}],
})
def test_apply_payload_rejects_unknown_content_class():
with pytest.raises(ValueError, match="unknown content_class"):
planner.apply_payload({
"user_class_weights": [{"content_class": "vibes", "weight": 1}],
})
def test_apply_payload_drops_class_from_wrong_list():
"""A canary class on the user list is silently dropped (operator
error), not raised — the partial save still applies the legit
entries. Roundtrip shows the operator their entry didn't land."""
planner.apply_payload({
"user_class_weights": [
{"content_class": "note", "weight": 10},
{"content_class": "canary_aws_creds", "weight": 100},
],
})
weights = {
e["content_class"]: e["weight"]
for e in planner.current_payload()["user_class_weights"]
}
assert weights == {"note": 10}
# canary class did NOT bleed onto the user list.
assert "canary_aws_creds" not in weights
def test_apply_payload_rejects_zero_total_weight():
with pytest.raises(ValueError, match="positive number"):
planner.apply_payload({
"user_class_weights": [{"content_class": "note", "weight": 0}],
})
def test_apply_payload_partial_failure_leaves_state_intact():
"""If validation rejects part of a payload, the planner's other
fields must not have been silently rebound."""
planner.apply_payload({"canary_probability": 0.10})
pre = planner.current_payload()
with pytest.raises(ValueError):
planner.apply_payload({
"user_class_weights": [{"content_class": "note", "weight": 5}],
"canary_probability": 9.0, # invalid
})
post = planner.current_payload()
assert post == pre # nothing rebound on failure
def test_apply_payload_ignores_unknown_keys():
"""Forward-compat: future fields land without breaking older clients."""
planner.apply_payload({"future_knob": "ignored"})
# Nothing changed.
assert planner.current_payload()["canary_probability"] == pytest.approx(0.03)

View File

@@ -152,6 +152,29 @@ async def test_get_synthetic_file_returns_none_when_missing(repo):
assert await repo.get_synthetic_file("does-not-exist") is None
@pytest.mark.asyncio
async def test_realism_config_get_returns_none_when_unset(repo):
assert await repo.get_realism_config("weights") is None
@pytest.mark.asyncio
async def test_realism_config_set_then_get_roundtrips(repo):
await repo.set_realism_config("weights", '{"canary_probability": 0.07}')
row = await repo.get_realism_config("weights")
assert row is not None
assert row["key"] == "weights"
assert row["value"] == '{"canary_probability": 0.07}'
@pytest.mark.asyncio
async def test_realism_config_set_is_upsert(repo):
await repo.set_realism_config("weights", '{"a": 1}')
await repo.set_realism_config("weights", '{"a": 2}')
row = await repo.get_realism_config("weights")
assert row is not None
assert row["value"] == '{"a": 2}'
def test_path_max_length_fits_mysql_utf8mb4_index_limit():
"""The unique (decky_uuid, path) index has to fit MySQL's 3072-byte
utf8mb4 cap: (decky_uuid_len + path_len) * 4 <= 3072. A regression