fix: double-paren bug in dropped-tools prefix + add empty-response nudge sibling
Fix: _LENGTH_CONTINUATION_DROPPED_TOOLS_PREFIX ended with '(' but
_get_continuation_prompt still had f'({tool_list})', producing
'((write_file)' instead of '(write_file)'. Removed the '(' from
the prefix constant — the parenthesis belongs in the interpolation.
Widened: promoted the empty-response nudge (line 6993,
'You just executed tool calls but returned an empty response...')
to _EMPTY_TOOL_RESPONSE_NUDGE constant and added it to the
classifier's recognition set. Same bug class — its
_empty_recovery_synthetic metadata flag doesn't survive SessionDB
projection either.
Test: added parametrize case for the empty-response nudge (7→8 cases).
E2E: verified byte-for-byte string equivalence for all nudge constants.
This commit is contained in:
@@ -4601,6 +4601,7 @@ This compaction should PRIORITISE preserving all information related to the focu
|
||||
_CODEX_ACK_CONTINUATION_NUDGE,
|
||||
_CODEX_INCOMPLETE_NUDGE,
|
||||
_DROPPED_TOOLCALL_NUDGE_CONTENT,
|
||||
_EMPTY_TOOL_RESPONSE_NUDGE,
|
||||
_LENGTH_CONTINUATION_DROPPED_TOOLS_PREFIX,
|
||||
_LENGTH_CONTINUATION_NETWORK_STUB,
|
||||
_LENGTH_CONTINUATION_OUTPUT_LIMIT,
|
||||
@@ -4613,6 +4614,7 @@ This compaction should PRIORITISE preserving all information related to the focu
|
||||
_CODEX_INCOMPLETE_NUDGE,
|
||||
_CODEX_ACK_CONTINUATION_NUDGE,
|
||||
_DROPPED_TOOLCALL_NUDGE_CONTENT,
|
||||
_EMPTY_TOOL_RESPONSE_NUDGE,
|
||||
_LENGTH_CONTINUATION_NETWORK_STUB,
|
||||
_LENGTH_CONTINUATION_OUTPUT_LIMIT,
|
||||
} or text.startswith(
|
||||
|
||||
@@ -779,7 +779,7 @@ _LENGTH_CONTINUATION_OUTPUT_LIMIT = (
|
||||
# The dropped-tools variant interpolates the tool name list right after this
|
||||
# prefix, so it can't be exact-matched — this stable prefix is what
|
||||
# _is_synthetic_compression_user_turn checks with str.startswith instead.
|
||||
_LENGTH_CONTINUATION_DROPPED_TOOLS_PREFIX = "[System: Your previous tool call ("
|
||||
_LENGTH_CONTINUATION_DROPPED_TOOLS_PREFIX = "[System: Your previous tool call "
|
||||
|
||||
|
||||
def _get_continuation_prompt(is_partial_stub: bool, dropped_tools: Optional[List[str]] = None) -> str:
|
||||
@@ -839,6 +839,15 @@ _DROPPED_TOOLCALL_NUDGE_CONTENT = (
|
||||
"the actual tool call now to continue the task."
|
||||
)
|
||||
|
||||
# Re-prompt sent when the model returns an empty response after executing tool
|
||||
# calls (#9400). Named for the same reason as the nudges above — its
|
||||
# _empty_recovery_synthetic metadata flag doesn't survive SessionDB projection.
|
||||
_EMPTY_TOOL_RESPONSE_NUDGE = (
|
||||
"You just executed tool calls but returned an "
|
||||
"empty response. Please process the tool "
|
||||
"results above and continue with the task."
|
||||
)
|
||||
|
||||
|
||||
# Shared recovery hint appended to every content-policy refusal message. Both
|
||||
# the HTTP-200 refusal path (``finish_reason=content_filter``) and the
|
||||
@@ -7070,11 +7079,7 @@ def run_conversation(
|
||||
messages.append(_nudge_msg)
|
||||
messages.append({
|
||||
"role": "user",
|
||||
"content": (
|
||||
"You just executed tool calls but returned an "
|
||||
"empty response. Please process the tool "
|
||||
"results above and continue with the task."
|
||||
),
|
||||
"content": _EMPTY_TOOL_RESPONSE_NUDGE,
|
||||
"_empty_recovery_synthetic": True,
|
||||
})
|
||||
continue
|
||||
|
||||
@@ -280,6 +280,12 @@ def test_real_task_wins_over_trailing_max_iterations_nudge(compressor):
|
||||
"the actual tool call now to continue the task.",
|
||||
id="dropped_toolcall_nudge",
|
||||
),
|
||||
pytest.param(
|
||||
"You just executed tool calls but returned an "
|
||||
"empty response. Please process the tool "
|
||||
"results above and continue with the task.",
|
||||
id="empty_tool_response_nudge",
|
||||
),
|
||||
],
|
||||
)
|
||||
def test_conversation_loop_retry_nudges_are_synthetic(content):
|
||||
|
||||
Reference in New Issue
Block a user