Teknium
08b140d14e
Codex OAuth caps gpt-6-astra at the same 272K window as gpt-5.4/5.5/5.6, so the global 50% trigger compacted at ~136K. Extend the existing codex_gpt55_autoraise gate to any slug containing "astra" (minus the opt-in -900k picker variants, which already unlock the wider window). Other routes (OpenAI direct, OpenRouter) keep the user threshold.
113 lines
4.6 KiB
Python
113 lines
4.6 KiB
Python
"""Behavior contracts for the GPT-5.6 (Sol/Terra/Luna) registration.
|
|
|
|
Invariant tests only — no list snapshots (per the no-change-detector-tests
|
|
policy). These pin the two behaviors that would silently regress:
|
|
|
|
1. Version/tier sorting: the flagship Sol must outrank Terra/Luna and the
|
|
whole 5.6 series must outrank 5.5, so `/model gpt` resolves to the
|
|
flagship rather than an alphabetical-first cheap tier.
|
|
2. Pricing reachability: the ("openai", <model>) official-docs pricing keys
|
|
must be reachable from BOTH the bare "openai" provider and the
|
|
"openai-api" picker slug (resolve_billing_route normalizes the latter).
|
|
"""
|
|
|
|
from decimal import Decimal
|
|
|
|
from agent.usage_pricing import (
|
|
_OFFICIAL_DOCS_PRICING,
|
|
_lookup_official_docs_pricing,
|
|
resolve_billing_route,
|
|
)
|
|
from hermes_cli.model_switch import _model_sort_key
|
|
|
|
|
|
class TestGpt56SortInvariants:
|
|
def test_sol_outranks_terra_and_luna(self):
|
|
models = ["gpt-5.6-luna", "gpt-5.6-terra", "gpt-5.6-sol"]
|
|
models.sort(key=lambda m: _model_sort_key(m, "gpt"))
|
|
assert models[0] == "gpt-5.6-sol"
|
|
|
|
|
|
def test_base_sol_outranks_sol_pro_for_alias_default(self):
|
|
# "-pro" high-effort variants parse as suffix "sol-pro" (rank 1), so
|
|
# `/model gpt` defaults to base Sol rather than the high-effort mode.
|
|
models = ["gpt-5.6-sol-pro", "gpt-5.6-sol"]
|
|
models.sort(key=lambda m: _model_sort_key(m, "gpt"))
|
|
assert models[0] == "gpt-5.6-sol"
|
|
|
|
|
|
class TestGpt56PricingRoute:
|
|
def test_official_pricing_reachable_from_openai(self):
|
|
route = resolve_billing_route("gpt-5.6-sol", provider="openai")
|
|
entry = _lookup_official_docs_pricing(route)
|
|
assert entry is not None
|
|
assert entry.input_cost_per_million == Decimal("5.00")
|
|
|
|
|
|
def test_cache_write_is_1_25x_input_for_56_series(self):
|
|
for slug in ("gpt-5.6-sol", "gpt-5.6-terra", "gpt-5.6-luna"):
|
|
entry = _OFFICIAL_DOCS_PRICING[("openai", slug)]
|
|
assert entry.input_cost_per_million is not None, slug
|
|
assert entry.cache_write_cost_per_million == (
|
|
entry.input_cost_per_million * Decimal("1.25")
|
|
), slug
|
|
assert entry.cache_read_cost_per_million == (
|
|
entry.input_cost_per_million * Decimal("0.10")
|
|
), slug
|
|
|
|
def test_pro_variants_alias_to_base_tier_pricing(self):
|
|
# -pro high-effort modes bill at the same per-token rates as their
|
|
# base tiers; the snapshot aliases them rather than duplicating rows.
|
|
for base in ("gpt-5.6-sol", "gpt-5.6-terra", "gpt-5.6-luna"):
|
|
assert (
|
|
_OFFICIAL_DOCS_PRICING[("openai", f"{base}-pro")]
|
|
is _OFFICIAL_DOCS_PRICING[("openai", base)]
|
|
), base
|
|
|
|
|
|
class TestGpt56CodexCompaction:
|
|
"""Codex OAuth caps the whole gpt-5.6 family at 272K, same as 5.4/5.5, so
|
|
the compaction auto-raise (0.85) must fire for every 5.6 variant on the
|
|
openai-codex route and NOT on the direct-API/OpenRouter routes."""
|
|
|
|
def test_autoraise_applies_to_all_56_on_codex(self):
|
|
from agent.auxiliary_client import _compression_threshold_for_model
|
|
|
|
for slug in (
|
|
"gpt-5.6-sol",
|
|
"gpt-5.6-sol-pro",
|
|
"gpt-5.6-terra",
|
|
"gpt-5.6-terra-pro",
|
|
"gpt-5.6-luna",
|
|
"gpt-5.6-luna-pro",
|
|
):
|
|
assert (
|
|
_compression_threshold_for_model(slug, provider="openai-codex")
|
|
== 0.85
|
|
), slug
|
|
|
|
def test_no_autoraise_on_direct_api_route(self):
|
|
from agent.auxiliary_client import _compression_threshold_for_model
|
|
|
|
# Direct OpenAI API / OpenRouter expose the full 1.05M window, so the
|
|
# 272K-cap override must NOT apply there.
|
|
assert (
|
|
_compression_threshold_for_model("gpt-5.6-sol", provider="openai")
|
|
is None
|
|
)
|
|
assert (
|
|
_compression_threshold_for_model(
|
|
"openai/gpt-5.6-sol", provider="openrouter"
|
|
)
|
|
is None
|
|
)
|
|
|
|
def test_astra_autoraise_on_codex_unless_900k(self):
|
|
"""Astra is 272K-capped on Codex OAuth like 5.6; the -900k opt-in variants are not."""
|
|
from agent.auxiliary_client import _compression_threshold_for_model
|
|
|
|
for slug in ("gpt-6-astra", "openai/gpt-6-astra-pro", "gpt-6-astra-2026-09-01"):
|
|
assert _compression_threshold_for_model(slug, provider="openai-codex") == 0.85, slug
|
|
assert _compression_threshold_for_model(slug, provider="openrouter") is None, slug
|
|
assert _compression_threshold_for_model("gpt-6-astra-900k", provider="openai-codex") is None
|