Teknium
2a95791992
run_agent.py: delete the `# noqa: F401` re-export block (agent.process_bootstrap
OpenAI/_SafeWriter/_get_proxy_*, model_tools get_tool_definitions/
handle_function_call/check_toolset_requirements, FailoverReason,
_qwen_portal_headers/_routermint_headers, session_persistence names,
estimate_request_tokens_rough, ContextCompressor + friends, jittered_backoff,
prompt_builder names, message_sanitization names, tool_dispatch_helpers
names) — 41 names run_agent never used itself — and the `_STREAM_DIAG_HEADERS`
back-compat class alias (no in-tree reader). run_agent now imports only what
it uses (get_toolset_for_tool, is_local_endpoint, coalesce/uniquify tool-call
ids, cleanup_vm/get_active_env from terminal_tool_lifecycle).
agent/*: `_ra().X` late-binds that only reached a re-export now import the
defining module directly (agent_runtime_helpers -> process_bootstrap.OpenAI,
model_tools.handle_function_call, session_persistence._safe_session_filename_component;
agent_init -> model_tools.get_tool_definitions/check_toolset_requirements,
_lazy_headers("agent.client_lifecycle", ...) for qwen/routermint;
system_prompt -> agent.prompt_builder / model_tools directly, dropping its
own _ra() shim and the `_r` parameter threading). `_ra()` stays for
run_agent-resident names (logger, AIAgent, _hermes_home, _set_interrupt, ...).
toolsets.py: remove resolve_multiple_toolsets (shim-only, restored by
34abf954bd); tests/test_toolsets.py pins the same union behavior via
resolve_toolset over each name.
providers/__init__.py: drop the OMIT_TEMPERATURE re-export (no callers via the
package); ProviderProfile stays because __init__ uses it for annotations —
2 tests repointed to providers.base.
agent/iteration_budget.py: drop the "run_agent re-exports the class"
docstring pointer; 4 tests import IterationBudget from its home.
model_tools.py (arg_coercion names), agent/tool_executor.py, and
hermes_cli/cli_session_mixin.py repoints landed via a sibling commit on this
shared worktree.
Callers repointed: gateway/run.py, hermes_cli/cli_chat_turn_mixin.py,
hermes_cli/cli_tui_mixin.py, tui_gateway/session_workdir.py,
agent/transports/codex.py (one-line imports) + comment pointers in
tools/file_state.py, tools/schema_sanitizer.py, scripts/tool_search_livetest.py.
Tests: patch("run_agent.X") / monkeypatch.setattr(run_agent, "X") /
`from run_agent import X` -> defining module across 99 test files.
97 lines
3.3 KiB
Python
97 lines
3.3 KiB
Python
"""Native Gemini output-token reservation at agent init (#57275 claim 4).
|
||
|
||
When ``model.max_tokens`` is unset, the native generateContent adapter still
|
||
sends ``maxOutputTokens=65,535`` (GEMINI_DEFAULT_MAX_OUTPUT_TOKENS) — Gemini
|
||
treats an omitted cap as a low internal default, not "unlimited", so the
|
||
adapter always sends an explicit cap. The compressor's trigger is
|
||
``pct × (window − max_tokens)``; constructing it with ``max_tokens=None``
|
||
reserved 0 while the wire reserved 65,535: on a 128K window the trigger
|
||
landed at ~96K against a real safe input budget of ~65K, and the provider
|
||
400'd before compaction ever fired.
|
||
|
||
These tests assert the compressor's reservation mirrors the adapter default
|
||
on the native Gemini path, and ONLY there.
|
||
"""
|
||
|
||
|
||
|
||
|
||
from unittest.mock import patch
|
||
|
||
|
||
|
||
import agent.context_compressor as cc_mod
|
||
from agent.gemini_native_adapter import GEMINI_DEFAULT_MAX_OUTPUT_TOKENS
|
||
|
||
|
||
CFG = {"agent": {}}
|
||
|
||
|
||
def _build_agent(model, base_url, provider="", max_tokens=None, window=131072):
|
||
with (
|
||
patch("model_tools.get_tool_definitions", return_value=[]),
|
||
patch("model_tools.check_toolset_requirements", return_value={}),
|
||
patch("agent.process_bootstrap.OpenAI"),
|
||
patch("hermes_cli.config.load_config", return_value=CFG),
|
||
patch("hermes_cli.config.load_config_readonly", return_value=CFG),
|
||
patch(
|
||
"agent.model_metadata.get_model_context_length", return_value=window,
|
||
),
|
||
patch.object(cc_mod, "get_model_context_length", return_value=window),
|
||
):
|
||
from run_agent import AIAgent
|
||
|
||
return AIAgent(
|
||
model=model,
|
||
api_key="test-key-1234567890",
|
||
base_url=base_url,
|
||
provider=provider,
|
||
max_tokens=max_tokens,
|
||
quiet_mode=True,
|
||
skip_context_files=True,
|
||
skip_memory=True,
|
||
)
|
||
|
||
|
||
def test_native_gemini_unset_max_tokens_reserves_adapter_default():
|
||
agent = _build_agent(
|
||
"gemma-3-27b-it",
|
||
"https://generativelanguage.googleapis.com/v1beta",
|
||
)
|
||
cc = agent.context_compressor
|
||
assert cc.max_tokens == GEMINI_DEFAULT_MAX_OUTPUT_TOKENS
|
||
# Trigger must sit at/below the real safe input budget the wire leaves.
|
||
assert cc.threshold_tokens <= cc.context_length - GEMINI_DEFAULT_MAX_OUTPUT_TOKENS
|
||
|
||
|
||
def test_gemini_provider_name_also_reserves_default():
|
||
agent = _build_agent(
|
||
"gemini-3.7-flash", "https://example-proxy.invalid/v1", provider="google",
|
||
)
|
||
assert agent.context_compressor.max_tokens == GEMINI_DEFAULT_MAX_OUTPUT_TOKENS
|
||
|
||
|
||
def test_explicit_max_tokens_wins_over_adapter_default():
|
||
agent = _build_agent(
|
||
"gemma-3-27b-it",
|
||
"https://generativelanguage.googleapis.com/v1beta",
|
||
max_tokens=8192,
|
||
)
|
||
assert agent.context_compressor.max_tokens == 8192
|
||
|
||
|
||
def test_non_gemini_paths_keep_no_reservation():
|
||
agent = _build_agent(
|
||
"openai/gpt-4.1", "https://openrouter.ai/api/v1",
|
||
)
|
||
assert agent.context_compressor.max_tokens is None
|
||
|
||
|
||
def test_gemini_openai_compat_endpoint_not_treated_as_native():
|
||
# The /openai compatibility endpoint does not use the native adapter.
|
||
agent = _build_agent(
|
||
"gemma-3-27b-it",
|
||
"https://generativelanguage.googleapis.com/v1beta/openai",
|
||
)
|
||
assert agent.context_compressor.max_tokens is None
|