Files
hermes-agent/tests/agent/test_codex_app_server_cache_write.py
Teknium 58886fc81e fix(cache): coverage deadlines, guaranteed vs estimated ends, durable generation
The clock modelled every route as "expires = last touch + tier", which is wrong for OpenAI
GPT-5.6+: a read refreshes coverage to at least 30 minutes after it and never restarts a
12h write. It also treated documented minimums as expiries and lost or revived saved state.

- Coverage deadlines per tier with per-route refresh rules: Anthropic family restarts the tier
  on read; OpenAI keeps max(existing, read + 30m) and a 12h deadline stays at write + 12h.
- Windows carry guaranteed_until (earliest deadline, so mixed tiers in one response are only as
  warm as the shorter) and estimated_until (OpenAI: last touch + 24h retention ceiling; pre-5.6
  in_memory 1h). cache_certainly_warm / cache_certainly_cold are the two consumer helpers.
  /context: warm until / may still be warm until / expired / unknown.
- Events bind to the session's durable context generation (rows a compaction or prune archived,
  SessionDB.prompt_cache_generation), kept as saved on restore; a resumed agent restores and
  validates the saved event before folding its first response.
- Route identity is provider + normalize_route_base_url (scheme, host, port, path) + model.
- Agentless session.context_breakdown reads the saved clock (stored_cache_window).
- Codex app-server activity is recorded as an explicit unknown (it sets its own cache options).
- MoA records the resolved aggregator route and the payload the MoA client sent.
- effective_cache_ttl applies the route ceiling on the measured-provider path too.
2026-11-15 12:04:54 -08:00

42 lines
2.3 KiB
Python

"""Codex app-server usage keeps the cache-write bucket the protocol reports.
codex-rs ``TokenUsageBreakdown`` carries ``cacheWriteInputTokens`` (from the Responses
``input_tokens_details.cache_write_tokens``); like ``cachedInputTokens`` it is inside
``inputTokens``. GPT-5.6+ bills those writes at 1.25x, so dropping them under-reports cost.
"""
from __future__ import annotations
from unittest.mock import patch
import run_agent
from agent.transports.codex_app_server_session import CodexAppServerSession, TurnResult
def test_cache_write_tokens_are_read_and_not_double_counted(monkeypatch):
def fake_run_turn(self, user_input: str, **kwargs):
return TurnResult(
final_text="done", projected_messages=[{"role": "assistant", "content": "done"}],
turn_id="turn-w-1", thread_id="thread-w-1",
token_usage_last={"totalTokens": 130, "inputTokens": 100, "cachedInputTokens": 20,
"cacheWriteInputTokens": 50, "outputTokens": 30, "reasoningOutputTokens": 0},
model_context_window=200000,
)
monkeypatch.setattr(CodexAppServerSession, "run_turn", fake_run_turn)
monkeypatch.setattr(CodexAppServerSession, "ensure_started", lambda self: "thread-w-1")
agent = run_agent.AIAgent(api_key="stub", base_url="https://stub.invalid", provider="openai",
api_mode="codex_app_server", quiet_mode=True, skip_context_files=True, skip_memory=True)
with patch.object(agent, "_spawn_background_review", return_value=None):
result = agent.run_conversation("hello")
assert (result["input_tokens"], result["cache_read_tokens"], result["cache_write_tokens"]) == (30, 20, 50)
assert result["prompt_tokens"] == 100 # uncached + read + write == inputTokens, never more
assert agent.session_cache_write_tokens == 50
# The app-server sets its own cache options: the clock records the activity as an explicit
# unknown, so no older window outlives it and /context names the reason.
from agent.prompt_cache_clock import cache_window, render_prompt_cache_line
window = cache_window(agent)
assert (window.confidence, window.reason) == ("unknown", "codex_app_server")
assert "runtime sets its own cache options" in render_prompt_cache_line(window.__dict__)