fix(compression): re-arm same-turn attempt budget

This commit is contained in:
Henry Mascot 2026-08-13 03:51:10 +01:00 committed by Teknium
parent 5e8d25d7e7
commit 2b9da1f252
2 changed files with 216 additions and 0 deletions

View File

@ -3742,6 +3742,37 @@ def run_conversation(
)
compression_attempts = 0
# ``compression_attempts`` bounds one *continuous pressure
# episode*, not the lifetime of a tool-heavy user turn. A
# successful provider call whose real prompt usage is back
# below the compression threshold proves the previous
# episode converged. Re-arm the budget and clear its
# insufficient-progress blocker so later tool output in the
# same turn can trigger a new episode.
_compression_threshold = int(
getattr(agent.context_compressor, "threshold_tokens", 0)
or 0
)
if (
(
compression_attempts > 0
or _preflight_compression_blocked
or _last_preflight_pressure is not None
)
and prompt_tokens > 0
and _compression_threshold > 0
and prompt_tokens < _compression_threshold
):
logger.info(
"Compression pressure cleared at %s < %s tokens; "
"re-arming same-turn compression budget",
f"{prompt_tokens:,}",
f"{_compression_threshold:,}",
)
compression_attempts = 0
_preflight_compression_blocked = False
_last_preflight_pressure = None
# Stash this response's canonical usage so the post-turn
# on_turn_complete() observation hook can forward it (the
# same dict shape passed to update_from_response). A turn

View File

@ -0,0 +1,185 @@
"""Regression test for re-arming the compression budget after tool progress."""
from __future__ import annotations
from types import SimpleNamespace
from unittest.mock import MagicMock, patch
import pytest
from run_agent import AIAgent
def _tool_call():
return SimpleNamespace(
id="call_1",
type="function",
function=SimpleNamespace(name="web_search", arguments='{"query": "x"}'),
)
def _tool_response(prompt_tokens: int):
message = SimpleNamespace(
content=None,
reasoning_content=None,
reasoning=None,
tool_calls=[_tool_call()],
)
return SimpleNamespace(
choices=[SimpleNamespace(message=message, finish_reason="tool_calls")],
model="test/model",
usage=SimpleNamespace(
prompt_tokens=prompt_tokens,
completion_tokens=1,
total_tokens=prompt_tokens + 1,
),
)
def _final_response():
message = SimpleNamespace(
content="done",
reasoning_content=None,
reasoning=None,
tool_calls=None,
)
return SimpleNamespace(
choices=[SimpleNamespace(message=message, finish_reason="stop")],
model="test/model",
usage=None,
)
def _malformed_response():
return SimpleNamespace(choices=[], model="test/model", usage=None)
def _tool_definition():
return {
"type": "function",
"function": {
"name": "web_search",
"description": "Search the web",
"parameters": {
"type": "object",
"properties": {"query": {"type": "string"}},
"required": ["query"],
},
},
}
@pytest.mark.parametrize(
("prompt_tokens", "expected_compactions", "provider_recovery"),
[(50, 2, False), (150, 1, False), (50, 2, True)],
ids=[
"pressure-cleared-rearms",
"pressure-still-high-stays-capped",
"pressure-cleared-rearms-after-provider-recovery",
],
)
def test_pre_api_compression_budget_rearms_only_after_pressure_clears(
prompt_tokens: int,
expected_compactions: int,
provider_recovery: bool,
):
"""Only provider-confirmed headroom starts a new pressure episode."""
with (
patch("run_agent.get_tool_definitions", return_value=[_tool_definition()]),
patch("run_agent.check_toolset_requirements", return_value={}),
patch("run_agent.OpenAI"),
patch("agent.model_metadata.get_model_context_length", return_value=256_000),
patch("agent.context_compressor.get_model_context_length", return_value=256_000),
):
agent = AIAgent(
api_key="test-key-1234567890",
base_url="https://openrouter.ai/api/v1",
model="test/model",
quiet_mode=True,
skip_context_files=True,
skip_memory=True,
max_iterations=6,
)
agent.client = MagicMock()
responses = [_tool_response(prompt_tokens), _final_response()]
if provider_recovery:
responses.insert(0, _malformed_response())
agent._fallback_chain = [object()]
agent._try_activate_fallback = MagicMock(return_value=True)
agent.client.chat.completions.create.side_effect = responses
agent._cached_system_prompt = "You are helpful."
agent._use_prompt_caching = False
agent._disable_streaming = True
agent.tool_delay = 0
agent.save_trajectories = False
agent.max_compression_attempts = 1
compressor = MagicMock()
compressor.protect_first_n = 3
compressor.protect_last_n = 20
compressor.threshold_tokens = 100
compressor.context_length = 1_000
compressor.last_prompt_tokens = -1
compressor.should_compress.side_effect = lambda tokens: tokens >= 100
compressor.should_compress_info.return_value = (False, None)
compressor.should_compress_preflight.return_value = False
compressor.should_defer_preflight_to_real_usage.return_value = False
compressor.get_active_compression_failure_cooldown.return_value = None
compressor.select_context.return_value = None
compressor.get_automatic_compaction_status_message.return_value = ""
agent.compression_enabled = True
agent.context_compressor = compressor
estimate_values = iter([200, 190, 200, 10])
compress_calls = []
def _fake_compress(messages, _system_message, **_kwargs):
compress_calls.append(messages)
return list(messages), "compressed prompt"
def _fake_execute_tool_calls(assistant_message, messages, *_args):
tool_call = assistant_message.tool_calls[0]
messages.append(
{
"role": "tool",
"name": tool_call.function.name,
"tool_call_id": tool_call.id,
"content": "ok",
}
)
history = [
{"role": "user" if i % 2 == 0 else "assistant", "content": f"msg {i}"}
for i in range(30)
]
with (
patch(
"agent.turn_context.estimate_request_tokens_rough",
return_value=10,
),
patch(
"agent.conversation_loop.estimate_messages_tokens_rough",
side_effect=lambda *_args, **_kwargs: next(estimate_values),
),
patch(
"agent.conversation_loop._estimate_tools_tokens_rough",
return_value=0,
),
patch.object(agent, "_compress_context", side_effect=_fake_compress),
patch.object(agent, "_execute_tool_calls", side_effect=_fake_execute_tool_calls),
patch.object(agent, "_flush_messages_to_session_db", return_value=True),
patch.object(agent, "_persist_session"),
patch.object(agent, "_save_trajectory"),
patch.object(agent, "_cleanup_task_resources"),
):
result = agent.run_conversation("do a lot of tool work", conversation_history=history)
assert result["completed"] is True
assert result["final_response"] == "done"
assert len(compress_calls) == expected_compactions, (
"same-turn compression must re-arm only after the provider confirms "
f"headroom; got {len(compress_calls)} compactions for "
f"prompt_tokens={prompt_tokens}"
)