f477c13d53
The 2^n output-budget ladder was open-coded in three places. Only the tool-call retry got the #72770 fix; the length-continuation retry (ceiling max(32768, cap)) and the Codex-incomplete retry (ceiling max(32768, base)) still re-sent the same budget once the cap reached 32768. boosted_output_cap() now serves all three sites: it ladders from the budget actually in use, its ceiling is max(32768, 2×cap), and it is clamped to the model's known output limit (_get_anthropic_max_output on anthropic_messages). When the request already sits at that limit it is not doubled, so the retry no longer earns a provider 400 (#79715). The length-continuation fix mirrors #72802. Co-authored-by: webtecnica <contato@webtecnica.com.br>
72 lines
2.8 KiB
Python
72 lines
2.8 KiB
Python
"""Truncation retries must send a larger output budget than the failed request (#72770),
|
|
without exceeding the model's known output limit (#79715)."""
|
|
from types import SimpleNamespace
|
|
|
|
import pytest
|
|
|
|
from agent.turn_iteration_prep import apply_retry_restarts
|
|
from agent.turn_retry_state import TurnRetryState
|
|
from agent.turn_truncation import _retry_truncated_tool_call
|
|
|
|
|
|
def _agent(max_tokens, requested_cap, **extra):
|
|
return SimpleNamespace(
|
|
max_tokens=max_tokens,
|
|
_ephemeral_max_output_tokens=None,
|
|
_buffer_vprint=lambda *a, **k: None,
|
|
_requested_output_cap_from_api_kwargs=lambda kw: requested_cap,
|
|
**extra,
|
|
)
|
|
|
|
|
|
def _tool_call_budgets(agent, attempts=4):
|
|
st = SimpleNamespace(agent=agent, truncated_tool_call_retries=0, is_stub=False)
|
|
st.done = lambda action, result=None: action
|
|
budgets = []
|
|
for _ in range(attempts):
|
|
assert _retry_truncated_tool_call(st, {}) == "continue"
|
|
budgets.append(agent._ephemeral_max_output_tokens)
|
|
return budgets
|
|
|
|
|
|
def _length_continuation_budgets(agent, attempts=4):
|
|
budgets = []
|
|
for n in range(1, attempts + 1):
|
|
_retry = TurnRetryState()
|
|
_retry.restart_with_length_continuation = True
|
|
verdict = apply_retry_restarts(
|
|
agent, _retry=_retry, response=None, interrupted=False, messages=[],
|
|
conversation_history=[], user_message="hi", api_kwargs={}, current_turn_user_idx=0,
|
|
final_response=None, retry_count=0, max_retries=3, api_call_count=1,
|
|
restart_count=0, length_continue_retries=n,
|
|
_preflight_compression_blocked=False, _turn_exit_reason="unknown",
|
|
)
|
|
assert verdict.action == "continue"
|
|
budgets.append(agent._ephemeral_max_output_tokens)
|
|
return budgets
|
|
|
|
|
|
SITES = [_tool_call_budgets, _length_continuation_budgets]
|
|
|
|
|
|
@pytest.mark.parametrize("site", SITES)
|
|
@pytest.mark.parametrize("requested_cap", [32768, 65536])
|
|
def test_retry_raises_budget_above_large_requested_cap(site, requested_cap):
|
|
budgets = site(_agent(None, requested_cap))
|
|
assert budgets[0] > requested_cap
|
|
assert max(budgets) <= requested_cap * 2
|
|
|
|
|
|
@pytest.mark.parametrize("site", SITES)
|
|
def test_boost_clamped_to_known_model_output_limit(site):
|
|
# claude-sonnet-4-5 outputs at most 64000: never ask for more, and don't double a
|
|
# request that already sits at the ceiling (the provider would 400).
|
|
below = site(_agent(None, 40000, api_mode="anthropic_messages", model="claude-sonnet-4-5"))
|
|
at_limit = site(_agent(None, 64000, api_mode="anthropic_messages", model="claude-sonnet-4-5"))
|
|
assert below == [64000] * 4
|
|
assert at_limit == [64000] * 4
|
|
|
|
|
|
def test_small_explicit_max_tokens_ladder_still_capped_at_floor():
|
|
assert _tool_call_budgets(_agent(4096, None)) == [8192, 16384, 32768, 32768]
|