Files
jobboard/tests/test_score.py
T
WolfandClaude Opus 5.5 7d31cca394 Two tabs and a cog: settings shell, one ranking, a script per page, the phone frame
The top bar is Roles and Applications plus health's cog. Behind the cog,
/settings/{ranking,searches,companies,pulls,messages,account} share one row
of sub-tabs, an underlined row on a desktop and pills on a phone. The old
paths answer 301 with their query string.

Ranking edits the first enabled agent, or creates one, with the intent,
the weights and the gates under Advanced. The order menu on Roles reads
Newest and Best match with one enabled agent, and names them with several.
config.yaml keeps the Copenhagen starter; the other two are commented
examples in config.example.yaml.

roles.js is split into roles.js, applications.js, ranking.js and
companies.js; count, cell, debounce, the company list and the affinity
control live once in app.js. app.css gets the phone frame after health and
newsdesk: one-row bar, edge-to-edge panels, 44px targets and 16px fields on
a coarse pointer, hover only where there is one, safe areas. The app is
installable from a manifest and roledesk-coloured icons. Every lede touched
is written for the person using the app.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
2026-09-25 21:29:06 +02:00

412 lines
16 KiB
Python

"""`embed`, `score` and the pull cascade, with the embedding backend replaced
by a deterministic fake. Nothing here downloads a model or reaches a board."""
from __future__ import annotations
import sys
from collections.abc import Iterator, Sequence
from datetime import UTC, datetime, timedelta
from pathlib import Path
from typing import Any
import numpy as np
import pytest
ROOT = Path(__file__).resolve().parents[1]
if str(ROOT) not in sys.path:
sys.path.insert(0, str(ROOT))
from desk.models import Item # noqa: E402
from pipeline import FINAL, SCHEDULES, TASKS, JobQueue, Jobs, RunContext # noqa: E402
from plugins.ats import ATSSource # noqa: E402
from app.config import AgentSeed, Config # noqa: E402
from app.jobs import cascade, embed, score # noqa: E402
from app.models import Candidate, Gate, Weights # noqa: E402
from app.storage import Store # noqa: E402
NOW = datetime(2026, 9, 23, 9, 0, tzinfo=UTC)
#: The fake's whole vocabulary: one dimension per word, plus a constant one so
#: no text is the zero vector.
WORDS = ("backend", "python", "frontend", "react", "data")
class FakeEmbedder:
model_id = "fake"
def __init__(self) -> None:
self.calls: list[list[str]] = []
def encode(self, texts: Sequence[str]) -> np.ndarray:
self.calls.append(list(texts))
rows = []
for text in texts:
lower = text.lower()
vec = np.asarray([lower.count(w) for w in WORDS] + [0.1], dtype=np.float32)
rows.append(vec / np.linalg.norm(vec))
return np.vstack(rows)
class Clock:
def __init__(self) -> None:
self.at = NOW
def __call__(self) -> datetime:
return self.at
@pytest.fixture
def fake(monkeypatch: pytest.MonkeyPatch) -> FakeEmbedder:
embedder = FakeEmbedder()
monkeypatch.setattr(embed, "load_embedder", lambda model: embedder)
return embedder
@pytest.fixture
def clock() -> Clock:
return Clock()
@pytest.fixture
def store(tmp_path: Path, clock: Clock) -> Iterator[Store]:
s = Store(tmp_path / "roledesk.db")
s.clock = clock
yield s
@pytest.fixture
def config(tmp_path: Path) -> Config:
return Config(database=str(tmp_path / "roledesk.db"), embed={"model": "fake"})
def role(store: Store, source_id: str, title: str = "Backend Engineer", *,
company: str = "Acme", body: str = "", **raw: Any) -> str:
item = Item(source_id=source_id, location=f"https://boards.example/{source_id}",
title=title, body=body, published_at=datetime(2026, 9, 20, tzinfo=UTC),
raw={"company": company, **raw})
item.source = "ats"
store.upsert_role(item)
return item.uid
def run(config: Config, store: Store, job: Any) -> dict[str, Any]:
notices: list[Any] = []
result = dict(job(config, store, context=RunContext(on_notice=notices.append)))
result["notices"] = [n.code for n in notices]
return result
# ------------------------------------------------------------------ embed
def test_embed_writes_a_vector_per_open_role_once(
store: Store, config: Config, fake: FakeEmbedder) -> None:
role(store, "1")
closed = role(store, "2")
with store.transaction() as conn:
conn.execute("UPDATE roles SET closed_at = ? WHERE uid = ?", (NOW.isoformat(), closed))
assert run(config, store, embed.run_embed)["embedded"] == 1
assert run(config, store, embed.run_embed)["embedded"] == 0
assert store.vector_counts() == {"fake": 1}
def test_embed_without_sentence_transformers_says_so_and_does_nothing(
store: Store, config: Config, monkeypatch: pytest.MonkeyPatch) -> None:
def missing(model: str) -> Any:
raise ImportError("no sentence_transformers")
monkeypatch.setattr(embed, "load_embedder", missing)
role(store, "1")
result = run(config, store, embed.run_embed)
assert result["embedded"] == 0 and result["notices"] == ["embed.unavailable"]
assert store.vector_counts() == {}
# ------------------------------------------------------------------ terms
def test_recency_halves_every_fourteen_days() -> None:
assert score.recency(NOW.isoformat(), NOW) == 1.0
assert score.recency((NOW - timedelta(days=14)).isoformat(), NOW) == pytest.approx(0.5)
assert score.recency((NOW - timedelta(days=28)).isoformat(), NOW) == pytest.approx(0.25)
def test_a_role_without_a_vector_is_scored_without_fit_unless_fit_is_all() -> None:
candidate = Candidate(uid="u", first_seen=NOW.isoformat(), affinity=0)
value, reason = score.score_one(candidate, [], Weights(), None, NOW)
assert reason == "" and value == pytest.approx((0.2 * 1.0 + 0.2 / 3) / 0.4)
only_fit = Weights(fit=1, recency=0, affinity=0)
assert score.score_one(candidate, [], only_fit, None, NOW) == (0.0, "no vector")
# ------------------------------------------------------------------ score
def test_a_require_gate_keeps_the_row_with_a_reason(
store: Store, config: Config, fake: FakeEmbedder) -> None:
berlin = role(store, "1", place="Berlin, Germany")
munich = role(store, "2", place="Munich")
nowhere = role(store, "3")
agent = store.save_agent("Berlin", "python backend",
[Gate(field="place", values=["Berlin"])])
run(config, store, score.run_score)
scores = store.scores(agent.id)
assert scores[berlin][1] == "" and scores[berlin][0] > 0
assert scores[munich] == (0.0, "place not Berlin")
assert scores[nowhere] == (0.0, "place unknown")
shown = store.roles(agent=agent.id)
assert [r.uid for r in shown] == [berlin]
assert {r.uid for r in store.roles(agent=agent.id, gated=True)} == {berlin, munich, nowhere}
def test_a_prefer_gate_adds_its_weight_to_the_score(
store: Store, config: Config, fake: FakeEmbedder) -> None:
remote = role(store, "1", remote="remote")
office = role(store, "2", remote="onsite")
agent = store.save_agent("Remote", "", [Gate(field="remote", values=["remote"],
mode="prefer", weight=0.5)])
run(config, store, score.run_score)
scores = store.scores(agent.id)
# No intent, so fit is out: recency 0.2 + affinity 0.2 + the gate's 0.5.
assert scores[remote][0] - scores[office][0] == pytest.approx(0.5 / 0.9, abs=1e-5)
assert scores[office][1] == ""
def test_a_favourite_skips_the_gates_and_is_floored(
store: Store, config: Config, fake: FakeEmbedder) -> None:
store.add_company("Fave", affinity=2)
fave = role(store, "1", company="Fave", place="Munich")
agent = store.save_agent("Berlin", "", [Gate(field="place", values=["Berlin"])])
run(config, store, score.run_score)
value, reason = store.scores(agent.id)[fave]
assert reason == "" and value >= 0.9
def test_a_muted_company_is_never_scored(
store: Store, config: Config, fake: FakeEmbedder) -> None:
store.add_company("Muted", affinity=-1)
muted = role(store, "1", company="Muted")
kept = role(store, "2")
agent = store.save_agent("Any")
run(config, store, score.run_score)
assert set(store.scores(agent.id)) == {kept}
assert muted not in {r.uid for r in store.roles(agent=agent.id, gated=True)}
def test_an_older_role_scores_lower(
store: Store, config: Config, clock: Clock, fake: FakeEmbedder) -> None:
clock.at = NOW - timedelta(days=14)
old = role(store, "1")
clock.at = NOW
new = role(store, "2")
agent = store.save_agent("Any")
run(config, store, score.run_score)
scores = store.scores(agent.id)
assert scores[new][0] > scores[old][0]
def test_two_agents_rank_the_same_roles_the_other_way_round(
store: Store, config: Config, fake: FakeEmbedder) -> None:
back = role(store, "1", "Backend Engineer", body="Python services")
front = role(store, "2", "Frontend Engineer", body="React apps")
only_fit = Weights(fit=1, recency=0, affinity=0)
backend = store.save_agent("Backend", "python backend developer", weights=only_fit)
frontend = store.save_agent("Frontend", "react frontend developer", weights=only_fit)
run(config, store, embed.run_embed)
result = run(config, store, score.run_score)
assert result["agents"] == 2
# Fit alone, scaled per agent: each agent's closer role is 1, the other 0.
assert store.scores(backend.id)[back][0] == 1.0
assert store.scores(backend.id)[front][0] == 0.0
assert store.scores(frontend.id)[front][0] == 1.0
assert store.scores(frontend.id)[back][0] == 0.0
# ------------------------------------------------------------ fit scaling
def test_scale_maps_the_closest_to_one_and_the_farthest_to_zero() -> None:
# The spread the last real run showed under `(cos + 1) / 2`.
assert score.scale({"a": 0.55, "b": 0.79, "c": 0.67}) == pytest.approx(
{"a": 0.0, "b": 1.0, "c": 0.5})
assert score.scale({"only": 0.3}) == {"only": 1.0}
assert score.scale({"a": 0.4, "b": 0.4}) == {"a": 1.0, "b": 1.0}
assert score.scale({}) == {}
def test_a_single_candidate_gets_full_fit(
store: Store, config: Config, fake: FakeEmbedder) -> None:
uid = role(store, "1", "Frontend Engineer", body="React apps")
agent = store.save_agent("Backend", "python backend developer",
weights=Weights(fit=1, recency=0, affinity=0))
run(config, store, embed.run_embed)
run(config, store, score.run_score)
assert store.scores(agent.id)[uid] == (1.0, "")
def test_a_gated_role_does_not_stretch_the_fit_range(
store: Store, config: Config, fake: FakeEmbedder) -> None:
close = role(store, "1", "Backend Engineer", body="Python backend", place="Berlin")
far = role(store, "2", "Data Engineer", body="Data", place="Berlin")
# Farther from the intent than either, but gated out by place.
role(store, "3", "Frontend Engineer", body="React react", place="Munich")
agent = store.save_agent("Berlin", "python backend",
[Gate(field="place", values=["Berlin"])],
weights=Weights(fit=1, recency=0, affinity=0))
run(config, store, embed.run_embed)
run(config, store, score.run_score)
scores = store.scores(agent.id)
assert scores[close][0] == 1.0 and scores[far][0] == 0.0
def test_a_small_prefer_gate_no_longer_outweighs_fit(
store: Store, config: Config, fake: FakeEmbedder) -> None:
"""The flaw the last real run showed: real cosines sit close together, so
under `(cos + 1) / 2` a prefer gate decided the order over fit. Here the
cosines are about 0.95 and 0.87: the old fit terms differ by 0.04, which
at weight 0.6 is less than the gate's 0.05, so the remote role won. Scaled,
the closer role wins."""
close = role(store, "1", "Backend Engineer", body="Python backend", remote="onsite")
far = role(store, "2", "Backend Engineer", body="Python backend data", remote="remote")
agent = store.save_agent("Backend", "python backend",
[Gate(field="remote", values=["remote"], mode="prefer",
weight=0.05)])
run(config, store, embed.run_embed)
run(config, store, score.run_score)
scores = store.scores(agent.id)
assert scores[close][0] > scores[far][0]
def test_the_intent_vector_is_cached_until_the_intent_changes(
store: Store, config: Config, fake: FakeEmbedder) -> None:
role(store, "1")
agent = store.save_agent("Backend", "python backend")
run(config, store, score.run_score)
run(config, store, score.run_score)
assert fake.calls == [["python backend"]]
store.save_agent("Backend", "react frontend", agent_id=agent.id)
run(config, store, score.run_score)
assert fake.calls[-1] == ["react frontend"]
def test_a_disabled_agent_is_not_scored(
store: Store, config: Config, fake: FakeEmbedder) -> None:
role(store, "1")
agent = store.save_agent("Off", enabled=False)
assert run(config, store, score.run_score)["agents"] == 0
assert store.scores(agent.id) == {}
def test_starter_agents_are_imported_once(store: Store, tmp_path: Path) -> None:
config = Config(database=str(tmp_path / "roledesk.db"),
agents=[AgentSeed(name="A"), AgentSeed(name="B")])
assert score.seed_agents(config, store) == ["A", "B"]
for agent in store.agents():
store.delete_agent(agent.id)
assert score.seed_agents(config, store) == []
assert store.agents() == []
def test_the_shipped_starter_agents_validate() -> None:
from app import config as config_mod
agents = config_mod.load(ROOT / "config.yaml").agents
assert [a.name for a in agents] == ["Copenhagen engineering"]
# The two commented examples in config.example.yaml, uncommented, so they
# still validate when somebody copies them.
import yaml
lines = (ROOT / "config.example.yaml").read_text(encoding="utf-8").splitlines()
start = lines.index("agents: []") + 1
block = [line[2:] for line in lines[start:] if line.startswith("# ")]
examples = [AgentSeed.model_validate(a) for a in yaml.safe_load("\n".join(block))]
assert [a.name for a in examples] == ["Remote from Europe", "Europe hybrid"]
# Small enough that a prefer gate breaks ties rather than overrides fit.
assert {g.weight for a in [*agents, *examples]
for g in a.gates if g.mode == "prefer"} == {0.05}
def test_the_shipped_starter_searches_validate() -> None:
from app import config as config_mod
searches = config_mod.load(ROOT / "config.yaml").searches
assert searches and {s.provider for s in searches} == {"jobindex"}
assert all(s.params.get("q") for s in searches)
# ---------------------------------------------------------------- cascade
def test_the_pull_job_runs_pull_extract_embed_then_score(
store: Store, config: Config, fake: FakeEmbedder, tmp_path: Path,
monkeypatch: pytest.MonkeyPatch) -> None:
store.add_subscription("ats", "greenhouse:acme", label="Acme", company="Acme",
board="greenhouse")
agent = store.save_agent("Backend", "python backend")
def fake_fetch(self: ATSSource) -> Iterator[Item]:
item = Item(source_id="greenhouse:acme:1", location="https://x.example/1",
title="Backend Engineer", body="Python",
published_at=datetime(2026, 9, 20, tzinfo=UTC),
raw={"company": "Acme", "board": "greenhouse", "slug": "acme"})
yield item
self.state["ats:ok:greenhouse:acme"] = self.state.get("ats:run", "")
self.state["ats:ids:greenhouse:acme"] = '["greenhouse:acme:1"]'
monkeypatch.setattr(ATSSource, "fetch", fake_fetch)
queue = JobQueue(Jobs(tmp_path / "roledesk.db"), config, store,
workers={"cpu": 1, "http": 1, "llm": 1})
# The `pull` job, which is what the schedule fires: its own work is to ask
# for the chain, which then runs to its end in the same drain.
started = queue.enqueue("pull")
queue.drain(timeout=60)
assert started is not None and queue.jobs.get(started.id).state == "done" # type: ignore[union-attr]
batch = queue.jobs.last("pull_batch")
assert batch is not None
jobs = sorted(queue.jobs.of_cascade(batch.cascade_id), key=lambda j: j.id)
stages = [j.result.get("task") for j in jobs if j.kind == FINAL]
assert stages == list(cascade.CASCADES["pull"])
work = [j.kind for j in jobs if j.kind != FINAL]
# The `pull` job and the chain it asked for are one cascade, so the Jobs
# tab shows a scheduled pull as one run. Extract is off and finds nothing.
assert work == ["pull", "pull_batch", "embed", "score"]
assert all(j.state == "done" for j in jobs)
assert len(store.scores(agent.id)) == 1
def test_pull_is_a_job_on_a_six_hour_schedule_and_the_stages_are_tasks() -> None:
"""A schedule enqueues a job kind, so `pull` is a job; were it a task too,
the Jobs tab would draw two `pull` buttons."""
assert "pull" not in TASKS
assert {"extract", "embed", "score"} <= set(TASKS)
assert SCHEDULES["pull"].kind == "pull"
assert SCHEDULES["pull"].cadence == "every 6h"
def test_follows_walks_the_cascade_and_stops_at_its_end() -> None:
assert cascade.follows("pull", "pull") == "extract"
assert cascade.follows("pull", "extract") == "embed"
assert cascade.follows("pull", "embed") == "score"
assert cascade.follows("pull", "score") == ""
assert cascade.follows("pull", "nope") == ""