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
These are the steps from the session where I saw it. It happened once, and I haven't yet reproduced it in a fresh session.
- Use the Antigravity provider with Google-account sign-in, Gemini 4 Argon (High), in a full-access thread.
- The agent calls
invoke_subagentwith one research subagent, which is auto-approved. - The subagent finishes: it reads some files, sends its report to the parent with
send_message, and endsDONE. - Open the "Antigravity subagent batch" child thread.
Expected behavior
The completed subagent's work is visible in the child view.
Actual behavior
The child thread contains only Running invoke_subagent and no transcript steps. The subagent's file reads appear in the parent's timeline, and its report did reach the parent.
Cause, from reading apps/server/src/provider/acp/AntigravityChildWork.ts on main @ 935818e422f9 alongside a copy of the conversation DB:
-
The launch arrived with ACP tool call id
<sessionId>:22.findInvokeRecordtreats22as thesteps.idxof the planner step that issued the call. In this DB:steps.idxStep type Tool calls 22 PLANNER_RESPONSEview_fileonly33 PLANNER_RESPONSEview_file,invoke_subagent(the real launch)35 INVOKE_SUBAGENTnames the child id and its transcript logUriSo
callIds.length !== 1and the lookup returnsundefined. RunningfindInvokeRecordon the copied DB with the suffix:33instead resolves the child correctly, so this mismatch alone explains the empty view. -
The launch-text fallback can't recover it. The id-shaped branch returns before reaching it. And T3 stored only the generic title
Antigravity subagent batchand the promptRunning invoke_subagent, which don't match the native launch record. -
For an auto-approved call with no pending permission frame, Google's ACP server uses the SDK call id as the ACP tool call id. Its
_match_tool_call_id_to_task_iddocstring describes a typical<trajectory_id>:<step_index>form, but doesn't say that index is SQLitesteps.idx. Here the suffix is 22, while:- the issuing planner is at
steps.idx33 (generation index 15); - there are 16 tool records before it, and 18 through the invoke result at idx 35.
I haven't worked out what the suffix counts.
- the issuing planner is at
A fix needs a reliable way to tie the SDK task id to the native launch record. If you know what the SDK's step index counts, or have an approach in mind, I'm happy to put together a focused PR with a failing test on a synthetic fixture.
Impact
Minor bug or occasional failure
Version or commit
Desktop 0.1.8-nightly.20261104.4165 (commit 8b44f42b14c8); code checked on main @ 935818e422f9. Google Antigravity ACP runtime release 923923a0d7dc.
Environment
Fedora 44 (KDE), T3 Code desktop AppImage, Antigravity provider with Google-account sign-in, Gemini 4 Argon (High).
Workaround
Read the child's transcript.jsonl directly, under the Antigravity profile's antigravity-acp/brain/<childConversationId>/.system_generated/logs/.
Posted by Claude Opus 5.5 on behalf of AriGr. Researched and tested by Claude Opus 5.5 and GPT 6.1 Sol, with help from GPT 6.1 Terra, using their native harnesses (Claude Code and Codex) through T3 Code.
Activity
HandleCoding commentedon Nov 13, 2026
I verified the public regression test and patch in #31316, and backported its three-file diff locally while it awaits merge.
Using the public test fixture from that PR,
follows each auto-approved launch id to the step that ran itfails against the original reader withexpected [] to deeply equal ['Test Runner'], and passes with the patch. The focused child-work test suite passes all 13 tests. The local changed-file checks also pass: formatting, lint, unused-code checks, GraphQL validation, type checking, and 15 tests across two files.This is validation of the native child-work lookup fix, not a claim that the patch changes live child-event routing in the parent timeline. I have not deployed the patched binary or performed an end-to-end UI check.
Posting here rather than creating a duplicate issue. Thanks for the existing fix in #31316.
HandleCoding commentedon Nov 13, 2026
Follow-up on live tool-event ownership: the transcript lookup fix in #31316 does not change the parent timeline's live projection. In the current adapter, a child tool frame carried in the root ACP session is projected as parent activity even when its SDK tool ID identifies the child trajectory.
I have now implemented an opt-in routing fix locally as a separate follow-up: associate native child trajectories with their batch using the launch receipt, route tool starts/progress/content into that batch's read-only thread, buffer frames that arrive before the receipt, and retain ownership across parent turns. The synthetic regression covers multiple batches, nested children, conflicting receipts, late updates, client-file read completion, and streamed output. Unmatched activity remains visible with an explicit attribution warning rather than being discarded. Routing IDs are retained before truncating long receipts for display.
Validation: all changed-file checks pass, including type checking of 14 affected packages and 70 focused tests; another 179 generic ACP, ACP Registry and Kiro regression tests pass. This is adapter-level validation with synthetic fixtures, not an end-to-end deployed UI claim.
Posting the distinction here because the reported empty child view and the live parent-timeline tool routing have separate causes.