Repository navigation
Description
Before submitting
- I searched existing issues and did not find a duplicate.
- I included enough detail to reproduce or investigate the problem.
Area
apps/server
Steps to reproduce
- Open an Antigravity ACP thread in T3 Code (e.g.,
gemini-4-argon/gemini-4-argon-high). - Ask the agent to run a command that exceeds
WaitMsBeforeAsync(e.g.,WaitMsBeforeAsync: 5000on a command that takes ~30–45 seconds, such as a deployment or test suite) so Antigravity moves the command to a background task (Tool is running as a background task with task id: <conversation-id>/task-<N>) and the agent ends its turn to wait for the background task to complete. - Observe the T3 Code thread UI while the background task is still running, and after the background task completes and wakes the agent in the harness.
Expected behavior
- While the background task is running and the agent is waiting for it:
- The open background command (
run_command/shell/schedule/run_workflow) should remain in a running/pending background state rather than being forcibly markedcompleted(Ran command) when the initial turn settles. thread.pendingBackgroundTasksshould include the running background task so the UI displays the background task indicator/banner (and keeps the thread in a waiting/working state rather than looking completely finished).
- The open background command (
- When the background task finishes and wakes the agent:
- T3 Code should start a continuation run (or stream the wakeup turn items and assistant messages live into the thread) so the user sees the agent wake up, execute follow-up tool calls, and output its final response without needing to manually send a follow-up prompt.
Actual behavior
- Thread immediately shows as Done and hides the running background task:
- As soon as the agent outputs its "waiting for
task-<N>to complete" message and ends its turn,finalizeTurnmarks the runcompletedandterminalizeOpenForegroundToolsforcibly overwrites the still-runningcommand_executionturn item fromin_progress(running) tocompleted(Ran command). - Because all turn items are now
completedandproviderThread.pendingBackgroundTasksis[],derivePendingBackgroundWorkreturns[]. The UI shows no running background tasks at all, making it impossible to tell whether the agent finished, died, or is waiting on a background task.
- As soon as the agent outputs its "waiting for
- Agent wakes up in the background when the task finishes, but the UI never shows the wakeup or response:
- When the background task finishes (e.g., 26 seconds later), the Antigravity harness delivers the
<SYSTEM_MESSAGE>task completion wakeup to the model. The agent wakes up in the background, runs follow-up tools, and generates its final completion response (verified in.system_generated/logs/transcript.jsonl), but nothing appears in the T3 Code UI until the user manually sends another prompt.
- When the background task finishes (e.g., 26 seconds later), the Antigravity harness delivers the
Root Cause Analysis (on main @ 5e9d56416f6b395095c772c597e79cd5f94e3052)
There are three interacting causes across AntigravityAdapterV2.ts, packages/provider-acp/src/server/adapter.ts, and antigravity-acp:
1. makeAntigravityAcpAdapterFlavor omits enablePostSettleContinuation: true despite AntigravityDriver wiring continuationRequests
- In
apps/server/src/provider/Drivers/AntigravityDriver.ts#L413-L428andapps/server/src/orchestration-v2/Adapters/AntigravityAdapterV2.ts#L228-L236,continuationRequests(ProviderContinuationRequests) is explicitly passed intomakeAntigravityAdapterV2and forwarded tomakeAcpAdapterV2. - However, in
packages/provider-acp/src/server/adapter.ts#L1356-L1357:const postSettleContinuationEnabled = flavor.enablePostSettleContinuation === true && continuationRequests !== undefined;
- Because
makeAntigravityAcpAdapterFlavor(apps/server/src/orchestration-v2/Adapters/AntigravityAdapterV2.ts#L154-L216) never setsenablePostSettleContinuation: true,postSettleContinuationEnabledis alwaysfalsefor Antigravity! - Consequently, when any post-settle session notification arrives while
context === null,bufferPostSettleWake(packages/provider-acp/src/server/adapter.ts#L4270-L4277) immediately returns{ buffered: false, offerContinuation: false, stopProcessing: false },offerContinuationRunis never called, and post-settleagent_message_chunk/agent_thought_chunkupdates fall through toappendLoadedHistory(adapter.ts#L4722-L4755) where they are silently written into the internalsnapshotmap without emittingmessage.updatedorturn_item.updatedevents to the UI.
2. makeAntigravityAcpAdapterFlavor omits extractBackgroundTaskId, causing terminalizeOpenForegroundTools to mark still-running background commands as completed at turn end
- When an Antigravity
run_commandtool call goes to the background,antigravity-acp(server.py:5490-5495and6090-6093) intentionally keeps thetool_callinin_progressstate (moving it tosession_info.persisted_open_tool_callswithout emitting a terminaltool_call_updateuntil the background process actually exits). - However, when the turn settles in T3 Code,
finalizeTurncallsterminalizeOpenForegroundTools(packages/provider-acp/src/server/adapter.ts#L6938-L6947), which checksacpCompletedTurnShouldTerminalizeTool(adapter.ts#L1223-L1231):export function acpCompletedTurnShouldTerminalizeTool( tool: AcpToolCallState, flavor: Pick<AcpAdapterV2Flavor, "extractBackgroundTaskId" | "extractSubagentUpdate">, ): boolean { const status = toolStatus(tool.status); if (status !== "pending" && status !== "running") return false; if (flavor.extractBackgroundTaskId?.(tool) !== undefined) return false; return flavor.extractSubagentUpdate?.(tool) === undefined; }
- Because
makeAntigravityAcpAdapterFlavordoes not provideextractBackgroundTaskId(e.g., for openexecutetools that remaininProgressat turn end, seeisAntigravityOpenCommandinapps/server/src/provider/acp/AntigravityProtocol.ts#L317-L319),acpCompletedTurnShouldTerminalizeToolreturnstrueandterminalizeOpenForegroundToolsforcibly emitsstatus: "completed"(Ran command) for the still-running background command! - Because the
command_executionturn item is now markedcompleted,pendingBackgroundTurnItemsandderivePendingBackgroundWork(packages/shared/src/orchestrationV2PendingBackgroundWork.ts#L174-L243) return[], so the UI hides the running background command and marks the thread completely idle/done.
3. antigravity-acp (server.py / event_processor.py) stops draining agent.conversation.receive_steps() once session/prompt returns
- In
antigravity-acp(google3/cloud/developer_experience/antigravity_extensions/acp_server/server.py:5483), steps fromlocalharness_externalare only forwarded toself._stream_live_step(...)inside theasync for step in agent.conversation.receive_steps():loop during an activesession/promptRPC. - Meanwhile, in
localharness_external, when arun_commandexceedsWaitMsBeforeAsync, theRUN_COMMANDstep remains inSTATE_RUNNING(status = 2) while the planner receives the background task ID (<conv>/task-<N>) and finishes its turn, causinglocalharness_externalto emitTrajectoryStateUpdate(state=STATE_WAITING_FOR_TASKS)(4). - If
session/promptexits onSTATE_WAITING_FOR_TASKS, nothing drainsagent.conversation.receive_steps()when the background task later finishes and wakes the planner inlocalharness_external—so the wakeup turn (e.g., steps 253–256 below) runs insidelocalharness_externalwithout emitting ACPsession/updatenotifications until the nextsession/promptcall. Conversely, ifsession/promptstays open acrossSTATE_WAITING_FOR_TASKS(or spawns a background step drainer that triggers post-settle continuation), T3 Code needs bothextractBackgroundTaskId/ open background command tracking andenablePostSettleContinuation: trueso the UI shows the pending background task while waiting and streams the wakeup turn when it fires.
Impact
Major degradation or frequent failure
Version or commit
0.1.7-nightly.20261031.3991 (and main @ 5e9d56416f6b395095c772c597e79cd5f94e3052)
Environment
Linux x86_64 (CachyOS), T3 Code 0.1.7-nightly.20261031.3991, Provider: Antigravity ACP (gemini-4-argon / gemini-4-argon-high)
Logs or stack traces
Concrete timeline from thread 6af2808e-6d27-43fa-91d1-eda0daffacc3 (Antigravity session c7b20c6f-f402-4c58-ad7b-6f56e7fa6f9c, run ordinal 19):
-
In
.t3/userdata/statev2.sqlite(orchestration_v2_projection_runsandorchestration_v2_projection_turn_items):- At
2026-10-31T23:15:23Z, the agent launches background commandtask-251(tool item...tool:c7b20c6f-f402-4c58-ad7b-6f56e7fa6f9c:236). - At
2026-10-31T23:15:33Z, the agent outputs: "I have launchedtask-251in the background to confirmdpl_2a6G1s8uQYFF7a4QnqnxaRrvvKJEreachesREADY(PROMOTED) and will report back as soon as it finishes." - At
2026-10-31T23:15:48.720Z,terminalizeOpenForegroundToolsmarkstool:...:236asstatus = 'completed'(title = 'Ran command') even thoughtask-251is still running, and at23:15:48.749Zrunordinal: 19is markedstatus = 'completed'withpendingBackgroundTasks = [].
- At
-
In
antigravity-acp/brain/c7b20c6f-f402-4c58-ad7b-6f56e7fa6f9c/.system_generated/logs/transcript.jsonl(steps 251–256):23:15:23Z(step 251,RUN_COMMAND,status: "RUNNING"):Tool is running as a background task with task id: c7b20c6f-f402-4c58-ad7b-6f56e7fa6f9c/task-25123:15:33Z(step 252,PLANNER_RESPONSE,status: "DONE"): Agent stops and waits fortask-251.23:16:14Z(step 253,SYSTEM_MESSAGE,status: "DONE"):Task id "c7b20c6f-f402-4c58-ad7b-6f56e7fa6f9c/task-251" finished with result: The command exited with code 0...23:16:14Z(step 254,PLANNER_RESPONSE,status: "DONE"): Agent wakes up automatically in the background and callsrun_command(curl -sS https://text2voice-pi.vercel.app/api/health).23:16:25Z(step 255,RUN_COMMAND,status: "DONE"): Health check returns{"status":"ok","ok":true,"database":"ok"}.23:16:28Z(step 256,PLANNER_RESPONSE,status: "DONE"): Agent writes its complete 1,048-token final summary (Everything is now 100% resolved, migrated, merged, and promoted to live production...).- In T3 Code (
statev2.sqlite): Zero messages or turn items were recorded after23:15:48.749Z! Steps253–256never appeared in the UI.
Workaround
Manually sending a follow-up prompt (e.g., "where are we at?") forces a new session/prompt turn so the agent can report what happened in the background.
Activity
fix(server): keep Antigravity commands active and show wakeups
fix(server): release Antigravity commands when a prompt request fails
fix(server): keep Antigravity commands active and show wakeups
fix(server): release Antigravity commands when a prompt request fails
fix(server): forget running commands when the ACP runtime restarts
fix(server): keep Antigravity commands active and show wakeups
fix(server): release Antigravity commands when a prompt request fails
fix(server): forget running commands when the ACP runtime restarts
eliophan commentedon Nov 3, 2026
Evidence: on runtime 1.4.5,
session/promptstays open until background tasks finish, so a non-exiting server keeps the turn "Working" foreverEnvironment: Antigravity ACP runtime 1.4.5 (darwin-arm64, release sha256
886fff77…), T3 Code Nightly0.1.7-nightly.20261103.4119.I drove the runtime directly with a minimal ACP stdio client (
fs.readTextFile: false, as T3 sends for 1.4.5) and asked the agent to start a background command and end its turn. Prompt: runsleep 50 && echo BGDONEwithWaitMsBeforeAsync: 1000, reply with one sentence that it is launched, and do nothing else.tool_call/tool_call_update→in_progressfor the commandagent_message_chunk: "The background task has been launched." (the model considers its turn over)tool_call_update→completedfor the commandagent_message_chunk: "The background task has finished successfully…"session/promptreturnsstopReason: end_turnSo the prompt was held open for ~50 s after the agent's final message, until the background task ended, and the wake-up happened inside the same prompt.
This seems to differ from the assumption in #28450 that Antigravity returns
end_turnwhile the command is still running. I only tried this one scenario (a short background command), so it may not cover the cases that PR was written for.Effect in a real thread with a server that never exits (
pnpm start):assistant_messageitem stayedrunning, with nosession/promptresult logged.pnpm startas a new background task, and the turn stayed open again.Stop ends the turn but not the background task (#16442), so the server keeps running.
LOG:
C-1.4.5-bg-task.sanitized.json
fix(server): keep Antigravity commands active and show wakeups
fix(server): release Antigravity commands when a prompt request fails
fix(server): forget running commands when the ACP runtime restarts
fix(server): keep Antigravity commands active and show wakeups
fix(server): release Antigravity commands when a prompt request fails
fix(server): forget running commands when the ACP runtime restarts
fix(server): keep Antigravity commands active and show wakeups
fix(server): release Antigravity commands when a prompt request fails
fix(server): forget running commands when the ACP runtime restarts
fix(server): keep Antigravity commands active and show wakeups
fix(server): release Antigravity commands when a prompt request fails
fix(server): forget running commands when the ACP runtime restarts
fix(server): keep Antigravity commands active and show wakeups
fix(server): release Antigravity commands when a prompt request fails
fix(server): forget running commands when the ACP runtime restarts
fix(server): keep Antigravity ACP turns attached across backgrounded …
Q3yp commentedon Nov 11, 2026
Live
agy acp1.5.5 wire findings (Windows ARM64) + working reference fixSharing empirical wire behavior from real authenticated Antigravity CLI
1.5.5(agy acp,gemini-4-carbon) on T3 Code Nightly (0.1.9-nightly.20261110.4491/main@b883156e2), since #28450 noted that liveagydelivery after earlyend_turnwas still unverified.What happens on real
agy acp1.5.5Driving
agy acp1.5.5 directly over stdio JSON-RPC with a minimal harness reproduces two behaviors that prevent passive post-settle listeners alone from receiving the wakeup:session/updatenotifications are emitted aftersession/promptreturnsend_turn: When a foregroundrun_command(IsDaemon: false, such as chainednpm run typecheck; npm run lint; npx vitest run ...) exceedsWaitMsBeforeAsync(capped at 10s),agyemitstool_callwithkind: "execute", status: "in_progress"(notool_call_update) and instructs the planner to end its turn immediately (<messaging>reactive wakeup). As soon as the planner ends its step,agy acpresponds tosession/promptwith{"stopReason": "end_turn"}whileconversation_summaries.dbremainsnot_fully_idle = 1. When the backgrounded shell process exits seconds later with no activesession/promptRPC attached,agywakes the planner internally in its transcript (SYSTEM_MESSAGE->PLANNER_RESPONSE), but sends0session/updateframes over ACP stdout. A client waiting passively for post-settlesession/updateframes never receives the command completion or the wakeup reply.session/request_permissionhangsagyindefinitely on follow-up commands indefaultmode:agy acpalso binds itssession/request_permissionbroker strictly to an activesession/promptcall. Indefaultmode, when the internal wakeup triggers a follow-up tool call (such as a secondrun_command) while nosession/promptis open,agysends0session/request_permissionframes over ACP and blocks forever internally waiting for permission (not_fully_idle = 1permanently).agy1.4.5 (Antigravity: a command still running after the reply is marked completed, and Stop cannot end it #16442): Unlike 1.4.5 (which heldsession/promptopen across all background tasks, causingIsDaemon: truedev servers to hang the turn forever), 1.5.5 returnsend_turnimmediately for bothIsDaemon: falseandIsDaemon: true.Verified behavior when
session/promptstays attached or re-attachesTesting both cases against live
agy acp1.5.5:session/promptopen via an MCP tool (t3_background_wait): Whilesession/promptremains active,agystreamstool_call_update(status: "completed", rawOutput)live the instant the backgrounded command exits, delivers<SYSTEM_MESSAGE>in the next step of the same turn, and routes follow-upsession/request_permissionprompts normally.session/promptimmediately after earlyend_turn: Sending a continuationsession/promptwhile theIsDaemon: falsecommand is stillin_progressre-attaches bothsession/updateandsession/request_permission(and replays gap completions if the command finished between the two RPCs).Working reference implementation (live-tested on Windows 11 ARM64)
Reference branch against
main(b883156e2):Summary of the reference approach (for maintainer design reference; not opened as a PR per
CONTRIBUTING.md):IsDaemoninAntigravityProtocol.tssoIsDaemon: trueservers/watchers never hold turns (Antigravity: a command still running after the reply is marked completed, and Stop cannot end it #16442).awaitBackgroundTasks(t3_background_wait) onAcpAdapterV2flavors opting intoforegroundCommandWait(plus MCP guidance whenhasT3Mcpis active) so foregroundrun_commandtasks hold the activesession/promptstream open untiltool_call_update(completed)arrives (orwaiting: falsereleases them).<t3_background_continuation>prompt loop inAcpAdapterV2as a fallback if the model still ends its turn early while an unreleased foreground command (IsDaemon: false) isinProgress, and drain the Antigravity update semaphore before resolvingruntime.prompt.Environment tested: Windows 11 ARM64, Node
v24.22.1, Antigravity CLI1.5.5(gemini-4-carbon), T3 Code desktop0.1.9-nightlybuild.(Investigated and drafted with
gemini-4-carbonvia Antigravity in T3 Code.)juliusmarminge commentedon Nov 12, 2026
This is the same problem as #16442, so I'm closing it in favour of that one. If it is different, say how and it will be reopened.
furkansandal commentedon Nov 16, 2026
Confirming this also reproduces on both MacOS and Windows with the nightly build and CLI 1.5.9.
mahmoodnizamani94-png commentedon Nov 18, 2026
@juliusmarminge following up on how this differs from #16442.
#16442 and PRs #35073 and #36585 keep the turn open while a command runs so pressing Stop can kill it. #36585 says in its scope section that automatic background completion delivery is still unresolved. Holding the turn open fixes Stop, yet T3 still drops the agent's reply when a finite background command finishes.
Google fixed the CLI side in agy 1.5.10 so it sends the wakeup reply and tool calls right after the command exits instead of waiting for the next prompt. Testing on current T3 with managed Antigravity 1.5.10, asking the agent to run
sleep 15 && echo "Task finished"in the background without polling and tell me when it finishes still fails the same way. Nine seconds into the sleep, T3 marks the turn done and resets the composer even though the agent just said it is waiting for the command to finish. Six seconds later the sleep exits and the agent writes its finished reply to the transcript on disk, while T3 stays idle and never shows that reply until you send another message.Several other users hit the same thing on 1.5.7 through 1.5.10 in #36105. Could we reopen this or track the wakeup delivery side alongside #16442?