Files
hermes-agent/tests/agent/test_credits_cold_start.py
Brooklyn Nicholson ac80df1410 fix(credits): stop desktop reap/resume from re-announcing the same usage band
Desktop rebuilds the AIAgent per turn (idle reap -> next message re-mint),
and agent_init.py gives every rebuilt agent a brand-new, empty _credits_latch
via new_credits_latch(). seed_credits_at_session_start() -> _hydrate_seed_state()
then unconditionally primes latch["seen_below_90"] on that fresh latch and
evaluates once -- correct for a genuinely new session opening mid-band, but on
a reap/resume rebuild it makes evaluate_credits_notices() see
shown_band=None vs. current_band=<the same band as before>, so it re-fires
"You've used $X of your $Y cap" on every message even though the user already
saw that exact notice moments ago on the previous incarnation of the same
session (#101578).

agent.session_id is stable across these rebuilds even though the agent object
and its latch are not, so add a small process-lifetime cache
(_seen_usage_bands, bounded to 500 entries, MRU eviction) keyed by session_id
that remembers the last usage_band actually shown. _hydrate_seed_state()
restores it into the fresh latch before evaluating, so a rebuild with unchanged
usage stays quiet, while a rebuild after a genuine crossing (recorded via the
same warm-path write in rate_limit_credits._emit_credits_notices, the single
chokepoint both the seed and live-header paths share) still fires normally.
Deliberately not persisted anywhere durable -- a real process restart is a real
"session open" and should still warn immediately, matching the existing
cold-start seed behavior for a session that opens already in a band. An agent
with no session_id (plain CLI, never rebuilt) falls back to the pre-fix
behavior unchanged.

Tests (tests/agent/test_credits_cold_start.py): a rebuild with the same
session_id and unchanged usage does not re-fire; a rebuild after a genuine
band change still fires (and clears the old key); a different session_id is
never suppressed by another session's history; an agent without a session_id
degrades to the old always-prime behavior without raising.

Fixes #101578

Co-authored-by: Edizzier <umit.ediz@hotmail.com>
Co-authored-by: Claude Sonnet 5 <noreply@anthropic.com>
2026-09-26 17:59:52 -05:00

451 lines
17 KiB
Python

"""Tests for cold-start credits hydration at session open.
The L3 cold-start seed primes agent._credits_state from /api/oauth/account (or a
HERMES_DEV_CREDITS_FIXTURE) so depletion AND the 90% grant warning fire immediately
at session open, not only after the first inference header. These tests assert the
notice policy fires correctly for a seed-shaped CreditsState with the warn90 latch
primed the way conversation_loop does it.
"""
import time
from agent.credits_tracker import CreditsState, evaluate_credits_notices
def _cold_start_notices(state: CreditsState):
"""Mirror the conversation_loop seed: prime seen_below_90 when used_fraction is
computable (the snapshot IS the first observation), then evaluate once."""
latch = {"active": set(), "seen_below_90": False}
if state.used_fraction is not None:
latch["seen_below_90"] = True
show, clear = evaluate_credits_notices(state, latch)
return [n.key for n in show]
def _state(**kw) -> CreditsState:
kw.setdefault("from_header", False)
kw.setdefault("captured_at", time.time())
return CreditsState(**kw)
def test_cold_start_healthy_no_notice():
s = _state(
remaining_micros=30_340_000, subscription_micros=18_000_000,
subscription_limit_micros=20_000_000, subscription_limit_usd="20.00",
denominator_kind="subscription_cap", paid_access=True,
)
assert abs(s.used_fraction - 0.1) < 1e-9
assert _cold_start_notices(s) == []
def test_dev_fixtures_drive_cold_start():
"""Every HERMES_DEV_CREDITS_FIXTURE state produces a valid seed CreditsState."""
import os
from agent.credits_tracker import dev_fixture_credits_state
expected = {
"healthy": [],
"sub_90pct": ["credits.usage"],
"depleted": ["credits.depleted"],
}
for name, want in expected.items():
os.environ["HERMES_DEV_CREDITS"] = "1" # fixtures gate on the dev flag
os.environ["HERMES_DEV_CREDITS_FIXTURE"] = name
try:
fx = dev_fixture_credits_state()
assert fx is not None, name
assert _cold_start_notices(fx) == want, (name, _cold_start_notices(fx))
finally:
os.environ.pop("HERMES_DEV_CREDITS_FIXTURE", None)
os.environ.pop("HERMES_DEV_CREDITS", None)
# ── seed_credits_at_session_start: the shared session-open hydrator ───────────
class _FakeAgent:
"""Minimal agent surface for the seed helper: state slots + an emit that runs
the real policy against the latch (mirroring run_agent._emit_credits_notices,
including the free-model suppression flag)."""
def __init__(self, provider="nous", model="", base_url="", session_id="session-1"): # session_id may be None
from agent.credits_tracker import evaluate_credits_notices, is_free_tier_model
self.provider = provider
self.model = model
self.base_url = base_url
self.session_id = session_id
self._credits_state = None
self._credits_session_start_micros = None
self._credits_latch = {"active": set(), "seen_below_90": False, "usage_band": None}
self.emitted: list = []
self._eval = evaluate_credits_notices
self._is_free = is_free_tier_model
from agent.credits_tracker import _remember_shown_band
self._remember_band = _remember_shown_band
def _emit_credits_notices(self):
if self._credits_state is None:
return
show, clear = self._eval(
self._credits_state,
self._credits_latch,
model_is_free=self._is_free(self.model, self.base_url),
)
self.emitted.append(([n.key for n in show], clear))
# Mirrors the post-fix rate_limit_credits chokepoint: the band this session
# last showed is remembered so a rebuild restores it (#101578).
self._remember_band(self.session_id, self._credits_latch.get("usage_band"))
def _seed(agent, fixture):
import os
from agent.credits_tracker import seed_credits_at_session_start
os.environ["HERMES_DEV_CREDITS"] = "1" # fixtures gate on the dev flag
os.environ["HERMES_DEV_CREDITS_FIXTURE"] = fixture
try:
return seed_credits_at_session_start(agent)
finally:
os.environ.pop("HERMES_DEV_CREDITS_FIXTURE", None)
os.environ.pop("HERMES_DEV_CREDITS", None)
def test_live_crossing_after_seed_still_fires_grant_spent():
"""The gate opens when the session observes the grant NOT yet spent — a healthy
seed followed by a grant-exhausted header is a real in-session crossing and must
still announce grant_spent once."""
a = _FakeAgent()
assert _seed(a, "healthy") is True
a.emitted = []
a._credits_state = _state( # the grant_exhausted shape, as a live header would carry it
remaining_micros=12_340_000, subscription_micros=0,
subscription_limit_micros=20_000_000, subscription_limit_usd="20.00",
purchased_micros=12_340_000, purchased_usd="12.34",
denominator_kind="subscription_cap", paid_access=True,
)
a._emit_credits_notices()
assert a.emitted == [(["credits.grant_spent"], [])]
def test_seed_is_idempotent():
a = _FakeAgent()
_seed(a, "sub_90pct")
a.emitted = []
# second call must no-op (state already populated)
assert _seed(a, "sub_90pct") is False
assert a.emitted == []
def test_seed_skips_non_nous():
from agent.credits_tracker import seed_credits_at_session_start
a = _FakeAgent(provider="openrouter")
assert seed_credits_at_session_start(a) is False
assert a._credits_state is None
# ── #101578: desktop reap/resume rebuilds must not re-announce the same band ──
#
# Desktop rebuilds the AIAgent (and its _credits_latch) on idle-reap/resume, but
# agent.session_id is stable across those rebuilds. A fresh latch that blindly
# re-primes seen_below_90 re-announces the CURRENT band on every rebuild even
# though the user already saw it moments ago on the previous incarnation.
def _band_state(used_fraction_of_cap: float) -> CreditsState:
"""A subscription-cap CreditsState at the given used fraction."""
cap = 20_000_000
return _state(
remaining_micros=int(cap * (1 - used_fraction_of_cap)),
subscription_micros=int(cap * (1 - used_fraction_of_cap)),
subscription_limit_micros=cap, subscription_limit_usd="20.00",
denominator_kind="subscription_cap", paid_access=True,
)
def test_rebuild_with_same_band_does_not_refire(monkeypatch):
import agent.credits_tracker as ct
monkeypatch.setattr(ct, "_seen_usage_bands", {})
_hydrate = ct._hydrate_seed_state
# First incarnation: session opens already at 50% — fires once (cold-start seed behavior).
a1 = _FakeAgent(session_id="desktop-session-1")
a1._credits_state = _band_state(0.5)
a1._credits_latch["seen_below_90"] = True
a1._emit_credits_notices()
assert a1.emitted == [(["credits.usage"], [])]
# Desktop idle-reap/resume: a BRAND-NEW agent, fresh empty latch, SAME session_id,
# usage unchanged (still 50%). Must NOT re-announce the band.
a2 = _FakeAgent(session_id="desktop-session-1")
_hydrate(a2, _band_state(0.5))
assert a2.emitted == [([], [])]
def test_rebuild_still_fires_on_genuine_band_change(monkeypatch):
import agent.credits_tracker as ct
monkeypatch.setattr(ct, "_seen_usage_bands", {})
_hydrate = ct._hydrate_seed_state
a1 = _FakeAgent(session_id="desktop-session-2")
a1._credits_state = _band_state(0.5)
a1._credits_latch["seen_below_90"] = True
a1._emit_credits_notices()
assert a1.emitted == [(["credits.usage"], [])]
# Usage genuinely climbed to 75% before the next rebuild — must still warn.
a2 = _FakeAgent(session_id="desktop-session-2")
_hydrate(a2, _band_state(0.75))
assert a2.emitted == [(["credits.usage"], ["credits.usage"])]
def test_unrelated_session_id_not_suppressed(monkeypatch):
import agent.credits_tracker as ct
monkeypatch.setattr(ct, "_seen_usage_bands", {})
_hydrate = ct._hydrate_seed_state
a1 = _FakeAgent(session_id="desktop-session-a")
a1._credits_state = _band_state(0.5)
a1._credits_latch["seen_below_90"] = True
a1._emit_credits_notices()
assert a1.emitted == [(["credits.usage"], [])]
# A different session at the same band has never been shown anything.
a2 = _FakeAgent(session_id="desktop-session-b")
_hydrate(a2, _band_state(0.5))
assert a2.emitted == [(["credits.usage"], [])]
def test_agent_without_session_id_falls_back_to_prior_behavior(monkeypatch):
"""CLI/plain agents may have no session_id — degrade to the pre-fix behavior
(cold-start seed warns at open), never raise."""
import agent.credits_tracker as ct
monkeypatch.setattr(ct, "_seen_usage_bands", {})
a = _FakeAgent(session_id=None)
ct._hydrate_seed_state(a, _band_state(0.5))
assert a.emitted == [(["credits.usage"], [])]
# ── background seed: the pricing warm the free-model gate depends on ─────────
_NOUS_BASE = "https://inference-api.nousresearch.com/v1"
# One subscription-billed row, keyed on the pre-/v1 root the picker caches under.
_SUBSCRIPTION_CATALOG = {
"https://inference-api.nousresearch.com": {
"openai/gpt-5.6-luna": {
"prompt": "0.0000002000", "completion": "0.0000012000", "billing_mode": "subscription",
},
}
}
class _DepletedAccount:
"""The /api/oauth/account shape _credits_state_from_account reads: access off, no money."""
class _Info:
total_usable_credits = 0.0
subscription_credits_remaining = 0.0
purchased_credits_remaining = 0.0
paid_service_access = False
paid_service_access_info = _Info()
subscription = None
def _cold_pricing_cache(monkeypatch):
"""Empty the process-wide pricing cache (and its expiry map) so the peek starts cold."""
from hermes_cli import models_pricing
monkeypatch.setattr(models_pricing, "_pricing_cache", {})
monkeypatch.setattr(models_pricing, "_pricing_cache_retry_after", {})
return models_pricing
def _run_bg_seed(monkeypatch, agent, *, warm):
"""Drive the real background branch and wait causally for its exact worker.
Hold the worker at its first side effect until the launcher returns, proving
that the seed remains fire-and-forget. Then release and join that captured
thread without making runner scheduling part of the behavior contract.
"""
import threading
import agent.memory_provider as memory_provider
import hermes_cli.nous_account as nous_account
from agent import credits_tracker
release_worker = threading.Event()
spawned: list[threading.Thread] = []
real_spawn = memory_provider.spawn_context_thread
def _gated_warm():
release_worker.wait()
warm()
def _capture_spawn(target, **kwargs):
thread = real_spawn(target, **kwargs)
spawned.append(thread)
return thread
monkeypatch.delenv("HERMES_DEV_CREDITS", raising=False) # fixtures would take the sync path
monkeypatch.setattr(credits_tracker, "_warm_nous_pricing_cache", _gated_warm)
monkeypatch.setattr(memory_provider, "spawn_context_thread", _capture_spawn)
monkeypatch.setattr(nous_account, "get_nous_portal_account_info", lambda *a, **kw: _DepletedAccount())
result = credits_tracker.seed_credits_at_session_start(agent)
try:
assert len(spawned) == 1
assert spawned[0].name == "credits-seed"
assert spawned[0].daemon is True
assert spawned[0].is_alive()
assert agent._credits_state is None
finally:
release_worker.set()
for thread in spawned:
thread.join()
return result
def test_bg_seed_warms_pricing_so_a_subscription_model_escapes_the_banner(monkeypatch):
"""The regression: a cold process peeks an EMPTY catalog, so a subscription-billed model — which
spends no credits — drew the depleted banner at session open. The seed fills the catalog first."""
models_pricing = _cold_pricing_cache(monkeypatch)
agent = _FakeAgent(model="openai/gpt-5.6-luna", base_url=_NOUS_BASE)
assert _run_bg_seed(
monkeypatch, agent, warm=lambda: models_pricing._pricing_cache.update(_SUBSCRIPTION_CATALOG)
) is True
assert agent._credits_state.depleted is True # the account really is out of credits...
assert agent.emitted == [([], [])] # ...and the session says nothing about it
def test_bg_seed_without_the_warm_still_warns(monkeypatch):
"""Guard rail for the test above: the same account and model DO warn while the catalog stays
cold, so the suppression is the warm's doing and not the model id's."""
_cold_pricing_cache(monkeypatch)
agent = _FakeAgent(model="openai/gpt-5.6-luna", base_url=_NOUS_BASE)
_run_bg_seed(monkeypatch, agent, warm=lambda: None)
assert agent.emitted == [(["credits.depleted"], [])]
def test_bg_seed_reruns_the_policy_when_a_header_beat_it(monkeypatch):
"""A live inference header can land while the warm is still in flight; it evaluates against the
cold catalog and shows the banner. The seed must keep that state but re-run the policy."""
models_pricing = _cold_pricing_cache(monkeypatch)
agent = _FakeAgent(model="openai/gpt-5.6-luna", base_url=_NOUS_BASE)
header_state = _state(paid_access=False)
def _warm():
agent._credits_state = header_state # the header lands mid-fetch...
agent._emit_credits_notices() # ...and evaluates against the cold catalog
models_pricing._pricing_cache.update(_SUBSCRIPTION_CATALOG)
_run_bg_seed(monkeypatch, agent, warm=_warm)
assert agent._credits_state is header_state # the seed never clobbers a live header
assert agent.emitted[0] == (["credits.depleted"], []) # cold: shown
assert agent.emitted[-1] == ([], ["credits.depleted"]) # warm re-run: cleared, no "restored"
# ── inference-header path: the catalog goes cold again after the Nous TTL ────
def _mixin_agent(model="openai/gpt-5.6-luna"):
"""The real notice path (RateLimitCreditsMixin._emit_credits_notices) with a recording driver."""
from agent.rate_limit_credits import RateLimitCreditsMixin
from agent.status_output import StatusOutputMixin
class _Agent(RateLimitCreditsMixin, StatusOutputMixin):
provider = "nous"
base_url = _NOUS_BASE
def __init__(self):
self.model = model
self.shown: list = []
self.cleared: list = []
self.notice_callback = lambda n: self.shown.append(n.key)
self.notice_clear_callback = self.cleared.append
self._credits_state = _state(paid_access=False)
return _Agent()
def _join_pricing_warm(agent):
thread = getattr(agent, "_credits_pricing_warm", None)
if thread is not None:
thread.join(timeout=10)
assert not thread.is_alive(), "pricing warm hung"
def test_header_after_ttl_expiry_rewarms_instead_of_flashing_the_banner(monkeypatch):
"""The session-start warm is one-shot; a Nous catalog expires after _NOUS_CATALOG_TTL_SECONDS. A
header landing after that saw a cold peek and brought the depleted banner back for a
subscription-billed model. Now the cold peek starts a re-warm and the banner never shows."""
from agent import credits_tracker
models_pricing = _cold_pricing_cache(monkeypatch)
models_pricing._cache_catalog(
_NOUS_BASE[:-3] + models_pricing._PRICING_AUTH_KEY_PREFIX + "abc",
_SUBSCRIPTION_CATALOG[_NOUS_BASE[:-3]], models_pricing._NOUS_CATALOG_TTL_SECONDS)
agent = _mixin_agent()
agent._emit_credits_notices()
assert agent.shown == [] # warm catalog: suppressed
models_pricing._pricing_cache_retry_after = { # ...then the TTL passes
k: v - models_pricing._NOUS_CATALOG_TTL_SECONDS - 1 for k, v in models_pricing._pricing_cache_retry_after.items()}
monkeypatch.setattr(credits_tracker, "_warm_nous_pricing_cache",
lambda: models_pricing._pricing_cache.update(_SUBSCRIPTION_CATALOG))
agent._emit_credits_notices() # the next inference header
_join_pricing_warm(agent)
assert agent.shown == []
assert agent.cleared == []
def test_header_on_a_cold_catalog_still_warns_when_the_warm_fails(monkeypatch):
"""Fail-open guard rail: a warm whose fetch fails (the cache keeps the empty result for its retry
window) decides against the cold peek — the banner shows — and neither the warm's own re-run nor
the next header inside that window starts another warm."""
from agent import credits_tracker
models_pricing = _cold_pricing_cache(monkeypatch)
warms: list = []
def _failed_fetch():
warms.append(1)
models_pricing._cache_catalog(_NOUS_BASE[:-3] + models_pricing._PRICING_AUTH_KEY_PREFIX + "abc", {})
monkeypatch.setattr(credits_tracker, "_warm_nous_pricing_cache", _failed_fetch)
agent = _mixin_agent()
agent._emit_credits_notices()
_join_pricing_warm(agent)
assert agent.shown == ["credits.depleted"]
agent._emit_credits_notices() # next header, still inside the failed-fetch window
_join_pricing_warm(agent)
assert warms == [1]