Repository navigation
Description
Before submitting
- I searched existing issues and found [Bug]: Antigravity ACP drops delegate_task, t3_thread_launch, and schedule_task over ~40 KB MCP schema budget and lacks T3_ACP_MCP_NODE fallback #36088, which was closed as
no-reprodespite the bot's own test reproducing the missing env vars onmain. Opening this follow-up with both the live repro and the exact unit-testable repro onmain(no Antigravity binary required).
Area
apps/server
Summary
In an Antigravity ACP session on T3 Code (0.1.10-nightly.20261113.4616), cross-provider subagent delegation (delegate_task), single-thread worktree launch (t3_thread_launch), and scheduled tasks (schedule_task) are completely uncallable:
- T3 MCP schemas are dropped by Antigravity ACP's eager schema cap because
AntigravityAdapterV2does not setdeferMcpTools: true(not3-tool-exposure: deferredheader), and Antigravity ACP neither exposes aLazy:section nor registerscall_mcp_tool. - The
T3_ACP_MCP_NODE ... acp-mcp-callterminal fallback is also broken becauseAntigravityDriverdropsinput.processEnvironmentwhen spawning the ACP process, leavingT3_ACP_MCP_NODE,T3_ACP_MCP_ENDPOINT,T3_ACP_MCP_AUTHORIZATION, andT3CODE_THREAD_IDunset in the shell.
Live Reproduction (still happening on 0.1.10-nightly.20261113.4616)
- Start T3 Code Desktop (
0.1.10-nightly.20261113.4616) on Windows with Antigravity (gemini-4-hydrogen,full-access) and other enabled providers (grok,magpie,magpie-codex,pi). - Verify thread binding & capabilities:
mcp_t3-code_t3_thread_configure(nothreadId) succeeds and returns the boundthreadId.mcp_t3-code_orchestrator_capabilitiessucceeds and returnsfeatures.appOwnedSubagents: truewith cross-provider instances available.
- Inspect
<T3CODE_HOME>/userdata/providers/antigravity/<id>/antigravity-acp/mcp/t3-code/:- All 71
.jsontool schemas are present on disk, includingdelegate_task.json,t3_thread_launch.json, andschedule_task.json.
- All 71
- Inspect the
<mcp_servers>prompt block and tool declarations exposed to the model:- Only 68
mcp_t3-code_*tools are registered underEager:(e.g.create_threadsworks). - There is no
Lazy:section and nocall_mcp_tooltool declared. delegate_task,t3_thread_launch, andschedule_taskare completely missing from the tool declarations.
- Only 68
- Inspect
Get-ChildItem Env:T3_*inside Antigravity'srun_commandtool:- Empty output (
T3_ACP_MCP_NODEandT3_ACP_MCP_ENTRYPOINTare unset), so the documentedacp-mcp-callfallback cannot run either.
- Empty output (
Unit-Testable Reproduction on main (no Antigravity binary needed)
As already observed in #36088, both root causes are directly reproducible in apps/server unit tests against the ACP mock agent:
-
Dropped
input.processEnvironmentinapps/server/src/provider/Drivers/AntigravityDriver.ts:
InAntigravityDriver.makeRuntime(input),buildAntigravityAcpSpawnInputis called withbaseEnv: processEnvironment, omittinginput.processEnvironment(T3CODE_THREAD_ID,T3_ACP_MCP_ENDPOINT,T3_ACP_MCP_AUTHORIZATION,T3_ACP_MCP_NODE).- Because Antigravity executes
run_commandinside that spawned process rather than via ACPterminal/create,T3_ACP_MCP_NODEnever reaches the shell. - Fix: Merge
{ ...processEnvironment, ...input.processEnvironment }intobaseEnv(which the [Bug]: Antigravity ACP drops delegate_task, t3_thread_launch, and schedule_task over ~40 KB MCP schema budget and lacks T3_ACP_MCP_NODE fallback #36088 test already confirmed makesAntigravityDriver.test.tspass).
- Because Antigravity executes
-
Missing
deferMcpTools: trueinapps/server/src/orchestration-v2/Adapters/AntigravityAdapterV2.ts:
AntigravityAdapterV2setsmcpOverHttp: truewithoutdeferMcpTools: true(unlike Copilot CLI inpackages/provider-acp-registry/src/server/adapter.ts), sot3-tool-exposure: deferredis never sent.- Without deferred exposure, T3 sends all 71 schemas upfront, exceeding Antigravity ACP's ~40 KB eager schema cap and silently dropping the 3 largest tools (
delegate_task,t3_thread_launch,schedule_task). - Fix: Set
deferMcpTools: trueinAntigravityAdapterV2sotools/listadvertises the compact core orchestration tools +tool_searchupfront, well under the ~40 KB cap.
- Without deferred exposure, T3 sends all 71 schemas upfront, exceeding Antigravity ACP's ~40 KB eager schema cap and silently dropping the 3 largest tools (
Expected Behavior
delegate_task,t3_thread_launch, andschedule_taskare callable from Antigravity threads (viadeferMcpTools: true/tool_search).T3_ACP_MCP_NODEand relatedT3_*env vars are present in the spawned Antigravity ACP process soacp-mcp-callfallback also works.
Version & Environment
- T3 Code:
0.1.10-nightly.20261113.4616(Desktop, Windows) - Provider: Antigravity (
gemini-4-hydrogen,full-access) - Related closed issue: [Bug]: Antigravity ACP drops delegate_task, t3_thread_launch, and schedule_task over ~40 KB MCP schema budget and lacks T3_ACP_MCP_NODE fallback #36088
Activity
juliusmarminge commentedon Nov 13, 2026
I tried to reproduce this on the current build and could not.
Antigravity cannot run on the VM, so no live run. On origin/main 2d8a843 I added two tests to apps/server/src/provider/Drivers/AntigravityDriver.test.ts that open a thread session through AntigravityDriver's orchestrationAdapter against the ACP mock agent, with an MCP provider session registered for the thread. (1) The spawned agent process's env has T3CODE_THREAD_ID undefined (and no T3_ACP_MCP_NODE): 'expected undefined to be thread-antigravity-thread-env'. (2) With the mock advertising HTTP MCP (T3_ACP_MCP_HTTP=1), the t3-code http server in session/new carries only the Authorization header, with no t3-tool-exposure: deferred. Both fail on main. Temporarily applying baseEnv: { ...processEnvironment, ...input.processEnvironment } in AntigravityDriver.ts and deferMcpTools: true in AntigravityAdapterV2.ts made all 8 tests in the file pass; I reverted those fixes and committed only the tests.
Closing as not planned. If it still happens for you, reply with the exact steps and the version you are on and it will be looked at again.
Fonsi479 commentedon Nov 14, 2026
Independent macOS / CLI 1.5.8 live confirmation on 2026-11-14. Please reconsider this issue, or track the report against #30083, which is currently still open / unmerged.
Environment: macOS arm64 host; T3 Code desktop with an iPhone client; official external
agy acp1.5.8;gemini-pro-orchestrator(Gemini 4 Carbon),thought_level: high. The installed ASAR version is 0.1.9-zh.2, commit 620970e8a23f, a downstream stable-0.1.9 build containing a background-command mitigation. This is not a claim that the current PR head fails.Observed in the live session
orchestrator_capabilitiestool call was recorded, and the agent reported Codex modelsgpt-6.1-solandgpt-6-lunaas available for cross-provider child tasks.delegate_taskabsent and nocall_mcp_toolfallback. The transcript shows the agent stopping to explain that it could not delegate, rather than recording a successfuldelegate_taskcall. The eager-list count is the agent's report; I have not independently extracted the full model request.run_commandshows noT3_ACP_MCP_NODEor otherT3_ACP_MCP_*variables. Only the unrelated desktop CLI path variable was present. Thus the documentedacp-mcp-callroute was unavailable too.t3_project_create,create_threads,preview_navigate,preview_open,t3_project_update,preview_act,html_render,delegate_task,t3_thread_launch,schedule_task.Their sizes range from 1,163 B to 3,359 B;
delegate_taskis 1,939 B. The smallest 58 total 28,375 B. These are measured file sizes, not a measurement of the model's exact schema budget.AntigravityDriver.makeRuntimepassesbaseEnv: processEnvironmentwithout merginginput.processEnvironment, while the Antigravity flavor setsmcpOverHttp: truewithoutdeferMcpTools: true.The agent's user-facing explanation blamed
anyOf / additionalProperties. That explanation is not established by this inspection and should not be treated as the root cause. The size-ranked omission pattern is consistent with #36088, and the missing process environment is directly visible in the recorded shell output and the driver source.Steps / expected behavior
Enable Antigravity and Codex in the project, start the Antigravity model above, and request a child task using a Codex model. The parent should be able to invoke
delegate_taskthrough the T3 MCP route (including deferred tool discovery if needed), or through T3's documented authenticated ACP fallback. Listing the model successfully while leaving both invocation paths unavailable blocks the requested delegation.No local repair or fresh synthetic delegation attempt was made for this report. Please distinguish this existing released-build failure from validation of the proposed PR fix. Private account data, workspace paths, and raw screenshots/transcripts are omitted.
JendrikLutz commentedon Nov 17, 2026
Confirming this still reproduces on T3 Code
v0.1.10(2026-11-17T01:44:04Z, macOS arm64, Antigravity ACPgemini-4-hydrogen/full-access):v0.1.10(and also in brand-new Antigravity threads),delegate_task.jsonis present on disk underantigravity-acp/mcp/t3-code/delegate_task.json, andorchestrator_capabilitiessucceeds (features.appOwnedSubagents: true).delegate_taskis missing from# t3-codeEager:tool declarations (noLazy:section exists), and calling it viacall_mcp_toolfails with:tool delegate_task is not enabled for server t3-codeenv | grep -E 'T3_ACP|T3_'inrun_commandreturns empty (T3_ACP_MCP_NODEandT3_ACP_MCP_ENDPOINTare unset), so the documentedacp-mcp-callfallback is also unavailable.Could this be reopened/revisited for
v0.1.10+?