f7567a62af
Three consecutive Codex Responses answers that carry only (encrypted) reasoning — no visible text, no tool call — used to exhaust the 3-continuation budget and end the turn on "Codex response remained incomplete after 3 continuation attempts", never touching configured fallback_providers (#67321). Encrypted reasoning items replay byte-for-byte, so a bare retry deterministically repeats the stall. - Track a per-turn `_codex_reasoning_only_streak` apart from the aggregate `_codex_incomplete_retries`: a visible partial resets the streak, so the mixed partial-then-stall variant still reaches its own recovery threshold while the turn-wide iteration budget stays the hard bound. - At streak 3, `continue_codex_incomplete` activates the next fallback with the semantic `FailoverReason.incomplete_response`, grants exactly one grace call when the trigger consumed the last iteration, and returns `CODEX_FALLBACK_ACTIVATED`; the intake re-syncs the Model:/Provider: identity on the system prompt. - Off the Codex wire the synthetic continuation nudge is stripped alongside the opaque replay state (`drop_nudge_marker`) so the Chat Completions payload keeps valid role ordering and no Codex-only control text. - No fallback configured: unchanged terminal sentinel, still bounded at 3 calls. Ported from PR #67336 by @PRATHAMESH75 onto the decomposed agent/turn_*.py siblings.
80 lines
3.5 KiB
Python
80 lines
3.5 KiB
Python
"""Responses-wire ``status=incomplete`` continuation must change the request when reasoning
|
|
consumed the whole output budget (#90393): a retry with the same ``max_output_tokens`` and the
|
|
same effort re-burns the budget identically and the turn can never converge. A fragment that
|
|
already carries visible text is a normal partial answer and keeps today's bare replay; with no
|
|
configured cap the escalation seeds from the ceiling the provider reported."""
|
|
|
|
from __future__ import annotations
|
|
|
|
from types import SimpleNamespace
|
|
from unittest.mock import MagicMock
|
|
|
|
from agent.turn_truncation import continue_codex_incomplete
|
|
|
|
|
|
def _agent(max_tokens: int | None = 2000):
|
|
agent = MagicMock()
|
|
agent.max_tokens = max_tokens
|
|
agent.quiet_mode = True
|
|
agent.log_prefix = ""
|
|
agent._codex_incomplete_retries = 0
|
|
agent._codex_reasoning_only_streak = 0
|
|
agent._ephemeral_reasoning_off = False
|
|
agent._ephemeral_max_output_tokens = None
|
|
agent._build_assistant_message.side_effect = lambda msg, fr: {
|
|
"role": "assistant", "content": msg.content or "", "finish_reason": fr,
|
|
"codex_reasoning_items": msg.codex_reasoning_items,
|
|
}
|
|
agent._interim_assistant_visible_text.return_value = ""
|
|
return agent
|
|
|
|
|
|
def _reasoning_only_message():
|
|
return SimpleNamespace(
|
|
content="", tool_calls=None, reasoning=None,
|
|
codex_reasoning_items=[{"type": "reasoning", "id": "rs_1", "encrypted_content": "x"}],
|
|
)
|
|
|
|
|
|
def _run(agent, response):
|
|
messages = [{"role": "user", "content": "write an essay"}]
|
|
return continue_codex_incomplete(
|
|
agent, _reasoning_only_message(), "incomplete", messages=messages,
|
|
conversation_history=None, api_call_count=1, response=response,
|
|
)
|
|
|
|
|
|
def test_budget_exhausted_empty_fragment_raises_cap_and_drops_reasoning():
|
|
agent = _agent()
|
|
exhausted = SimpleNamespace(
|
|
status="incomplete", incomplete_details={"reason": "max_output_tokens"},
|
|
usage=SimpleNamespace(output_tokens=2000, output_tokens_details={"reasoning_tokens": 1997}),
|
|
)
|
|
assert _run(agent, exhausted) is None
|
|
assert agent._ephemeral_reasoning_off is True
|
|
first_boost = agent._ephemeral_max_output_tokens
|
|
assert first_boost > 2000
|
|
agent._ephemeral_max_output_tokens = None # request builder consumes it
|
|
assert _run(agent, exhausted) is None
|
|
assert agent._ephemeral_max_output_tokens > first_boost
|
|
|
|
|
|
def test_no_configured_cap_seeds_escalation_from_observed_ceiling():
|
|
agent = _agent(max_tokens=None)
|
|
exhausted = SimpleNamespace(
|
|
status="incomplete", incomplete_details={"reason": "max_output_tokens"},
|
|
usage=SimpleNamespace(output_tokens=6000, output_tokens_details={"reasoning_tokens": 5997}),
|
|
)
|
|
assert _run(agent, exhausted) is None
|
|
assert agent._ephemeral_max_output_tokens > 6000
|
|
# A fragment that carries visible text is a normal partial answer: no override, today's replay.
|
|
agent = _agent()
|
|
agent._interim_assistant_visible_text.return_value = "First paragraph of the essay"
|
|
messages = [{"role": "user", "content": "write an essay"}]
|
|
partial = SimpleNamespace(content="First paragraph of the essay", tool_calls=None, reasoning=None,
|
|
codex_reasoning_items=None)
|
|
continue_codex_incomplete(agent, partial, "incomplete", messages=messages,
|
|
conversation_history=None, api_call_count=1, response=exhausted)
|
|
assert agent._ephemeral_reasoning_off is False
|
|
assert agent._ephemeral_max_output_tokens is None
|