Antigravity ACP (Gemini 4 Argon): Streaming text gets stuck in running/streaming state & thoughts are never streamed #28172

Summary

When using Antigravity ACP with Gemini 4 Argon (gemini-4-argon / gemini-4-argon-high via Google OAuth / CCPA), assistant text in the T3 Code UI frequently freezes mid-turn for 15–45 seconds or remains stuck in streaming: true (status: "running") for several minutes (6–14+ minutes observed in statev2.sqlite orchestration_v2_projection_turn_items) after the assistant has finished writing its response. Additionally, reasoning (agent_thought_chunk) items are never emitted (0 reasoning items across 100+ gemini-4-argon turns).

Observed Evidence (statev2.sqlite & Provider Event Logs)

From orchestration_v2_projection_turn_items in statev2.sqlite:

• assistant_message items remain in status = 'running' (streaming: true) long after the final agent_message_chunk has been received:
  • Example 1: started_at = 2026-10-31T12:54:39.643Z, completed_at = 2026-10-31T13:08:51.339Z (stuck in streaming: true for 14m 12s until the next user prompt).
  • Example 2: started_at = 2026-10-31T14:13:17.603Z, completed_at = 2026-10-31T14:19:48.160Z (stuck in streaming: true for 6m 31s after the 13,813-char response completed at 14:13:18Z).
  • Even within normal multi-step turns, assistant_message items stay in streaming: true (status: "running") for 15–45 seconds after text generation finishes while the model silently thinks on the next step.
• Across 100+ gemini-4-argon turns, SELECT count(*) FROM orchestration_v2_projection_turn_items WHERE kind = 'reasoning' is 0.

Root Causes (Interaction Between T3 Code ACP Adapter & Antigravity ACP Server)

1. T3 Code ACP Adapter (makeAcpAdapterV2) keeps context.assistant.current open in streaming: true

In makeAcpAdapterV2 (handleSessionUpdate / appendText / closeTextStream):

• Receiving an agent_message_chunk calls appendText(context, "assistant", text, update.messageId), which opens context.assistant.current with streaming: true (status: "running") and throttles flushes via ACP_TEXT_FLUSH_INTERVAL (50ms).
• closeTextStream(context, "assistant") is only called when:
  1. Another text stream (reasoning or user) starts,
  2. emitTool fires, or
  3. The entire session/prompt RPC resolves (turn.completed).
• Because gemini-4-argon on OAuth/CCPA never emits agent_thought_chunk (reasoning) between assistant text and subsequent thinking phases (see Add open-in-editor feature with Cmd+O shortcut #2 below), and because session/prompt remains open whenever the Antigravity trajectory is in STATE_WAITING_FOR_TASKS / STATE_WAITING_FOR_SUBAGENTS (see light mode #4 below), context.assistant.current never receives closeTextStream(context, "assistant") and stays stuck in streaming: true (status: "running") in the UI.
• Furthermore, in the web client projection (timelineVisibleTurnItemsAtom / quietThreadAtom), streaming assistant items (streaming: true) have their text and updatedAt stripped from the base item in quietThreadAtom and rely on assistantMessageTextAtom while the streaming indicator remains active.

2. Antigravity ACP (agy_acp_server.par + localharness_external) strips ThinkingConfig (includeThoughts: true) on OAuth/CCPA

• In google3/cloud/developer_experience/antigravity_extensions/acp_server/server.py:
  • _build_mapped_model_target (used for API key / Vertex) populates endpoint.options = GeminiModelOptions(thinking_level=thinking_level).
  • _build_literal_model_target (server.py:4025, used for AUTH_METHOD_OAUTH and AUTH_METHOD_OAUTH_BUSINESS at lines 4572 & 4588) constructs ModelTarget without setting endpoint.options (options=None).
• In localharness_external (google3/third_party/jetski_prod/localharness/geminimodelapi.(*Client).Generate):
  • BuildGenerateContentRequest initially sets genConfig.ThinkingConfig = &ThinkingConfig{IncludeThoughts: true, ...}.
  • Immediately afterward, geminimodelapi.(*Client).Generate checks if len(c.thinkingLevel) == 0 (populated from GeminiModelOptions.GetThinkingLevel()). When c.thinkingLevel == "", it overwrites genConfig.ThinkingConfig = nil!
• As a result, requests to http://127.0.0.1:<port>/v1beta/models/gemini-4-argon-high:streamGenerateContent are sent to ccpa_connection/proxy_server.py without thinkingConfig: {"includeThoughts": true}. CCPA still runs Argon with a HIGH thinking budget server-side (taking 10–45s per step), but streams zero thought: true chunks back over SSE, making the UI appear completely frozen during thinking phases and preventing makeAcpAdapterV2 from closing the previous assistant text stream when a new thinking step starts.

3. Antigravity ACP _stream_live_step (server.py:5856-5867) drops final content_delta / thinking_delta on StepStatus.DONE

In server.py:

if step.status == sdk_types_module.StepStatus.DONE:
  for tool_call in step.tool_calls or []:
    if tool_call.id:
      await self._register_live_step_tool_call(session_id, tool_call)
if step.status != sdk_types_module.StepStatus.ACTIVE:
  return
if step.thinking_delta:
  ...
if step.content_delta:
  ...

In localharness_external ((*StepToEventTranslator).translateStepLocked), text_delta and thinking_delta are computed on every StepUpdate, including the transition to CORTEX_STEP_STATUS_DONE. Returning early on step.status != StepStatus.ACTIVE before flushing step.thinking_delta and step.content_delta silently drops any final text/thought delta delivered on the DONE event.

4. Antigravity ACP LocalHarnessEventProcessor (event_processor.py:958-982) hangs session/prompt when background/orphaned tasks exist (STATE_WAITING_FOR_TASKS)

In localharness_external ((*runningTrajectory).computeCurrentStateLocked), if any background task is active (hasActiveTasks()), the trajectory transitions to STATE_WAITING_FOR_TASKS (4) rather than STATE_FULLY_IDLE (2) when the planner finishes its turn. Because LocalHarnessEventProcessor.process_event only sets self.is_idle.set() on STATE_FULLY_IDLE and STATE_CANCELLED, session/prompt never returns, leaving the last assistant_message in T3 Code stuck in streaming: true (status: "running") for minutes until the user manually interrupts or sends another prompt.

Suggested Fixes

1. In T3 Code (makeAcpAdapterV2):
   • Flush and close (or transition out of active streaming: true after an idle/step boundary) context.assistant.current when a planner step finishes or when tool_call / tool_call_update / plan updates arrive, rather than waiting exclusively for reasoning chunks or session/prompt completion.
2. In Antigravity ACP (agy_acp_server.par):
   • Populate GeminiModelOptions(thinking_level=...) in _build_literal_model_target (or ensure thinkingConfig.includeThoughts = True in ccpa_connection/proxy_server.py for Gemini thinking models).
   • Flush step.thinking_delta and step.content_delta in _stream_live_step on both StepStatus.ACTIVE and StepStatus.DONE (before registering DONE tool calls).