fix(compression): notify users when configured aux model fails even if main-model fallback recovers (#16775)

A misconfigured auxiliary.compression.model is a user-fixable problem that silent recovery would hide. The previous retry-on-main logic transparently swallowed aux-model failures whenever the fallback succeeded, leaving the user's broken config in place and racking up future failures.

Track the aux-model failure on the compressor alongside the existing fallback-placeholder fields:
- _last_aux_model_failure_model: str | None
- _last_aux_model_failure_error: str | None

Both are set at the moment the aux model errors (captured before summary_model is cleared for retry), regardless of whether the retry succeeds. Cleared at compress() start and on on_session_reset() so a clean run doesn't leak stale warnings.

Surface at three places:
- gateway hygiene auto-compress: ℹ note to the platform adapter (thread_id preserved)
- gateway /compress command: ℹ line appended to the reply
- CLI via _emit_warning: deduped on (model, error) so repeat compactions don't spam

Distinct from the existing ⚠️ dropped-turns warning — different severity, different emoji, explicit 'context is intact' reassurance.
This commit is contained in:
Teknium 2026-04-27 20:08:23 -07:00 committed by GitHub
parent c3e3a9c184
commit 6ea5699e3f
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
6 changed files with 367 additions and 1 deletions

View file

@ -181,3 +181,65 @@ async def test_compress_command_appends_warning_when_summary_generation_fails():
assert "historical message(s) were removed" in result
agent_instance.shutdown_memory_provider.assert_called_once()
agent_instance.close.assert_called_once()
@pytest.mark.asyncio
async def test_compress_command_surfaces_aux_model_failure_even_when_recovered():
"""When the user's configured ``auxiliary.compression.model`` errors out
but compression recovers by retrying on the main model, /compress must
STILL inform the user. Silent recovery hides broken config the user
needs to fix."""
history = _make_history()
# Compressed transcript — normal successful compression, no placeholder.
compressed = [
history[0],
{"role": "assistant", "content": "summary via main model"},
history[-1],
]
runner = _make_runner(history)
agent_instance = MagicMock()
agent_instance.shutdown_memory_provider = MagicMock()
agent_instance.close = MagicMock()
agent_instance.context_compressor.has_content_to_compress.return_value = True
# Fallback placeholder was NOT used — recovery succeeded.
agent_instance.context_compressor._last_summary_fallback_used = False
agent_instance.context_compressor._last_summary_dropped_count = 0
agent_instance.context_compressor._last_summary_error = None
# But the configured aux model DID fail before the retry succeeded.
agent_instance.context_compressor._last_aux_model_failure_model = (
"gemini-3-flash-preview"
)
agent_instance.context_compressor._last_aux_model_failure_error = (
"404 model not found: gemini-3-flash-preview"
)
agent_instance.session_id = "sess-1"
agent_instance._compress_context.return_value = (compressed, "")
def _estimate(messages):
if messages == history:
return 100
if messages == compressed:
return 60
raise AssertionError(f"unexpected transcript: {messages!r}")
with (
patch("gateway.run._resolve_runtime_agent_kwargs", return_value={"api_key": "***"}),
patch("gateway.run._resolve_gateway_model", return_value="test-model"),
patch("run_agent.AIAgent", return_value=agent_instance),
patch("agent.model_metadata.estimate_messages_tokens_rough", side_effect=_estimate),
):
result = await runner._handle_compress_command(_make_event())
# Compression succeeded
assert "Compressed:" in result
# No ⚠️ warning (that's reserved for dropped-turns case)
assert "⚠️" not in result
# But there IS an info note about the broken aux model
assert "" in result
assert "gemini-3-flash-preview" in result
assert "404" in result
assert "auxiliary.compression.model" in result
# The user's context is explicitly called out as intact
assert "intact" in result
agent_instance.shutdown_memory_provider.assert_called_once()
agent_instance.close.assert_called_once()