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 (packages/provider-acp)
Steps to reproduce
- Start a thread using the Antigravity provider (ACP, e.g.
gemini-4-neon). - In a single run, have the agent launch a long-running background shell command (so it continues running asynchronously), write a substantive user-facing reply/update, and end its turn waiting for the reactive wakeup.
- Let the background task finish and wake the agent inside the same T3 Code run, causing the agent to emit a follow-up reply (or multiple wakeups producing
assistant:0,assistant:1, ...,assistant:Nunder the samerunId). - Inspect the thread timeline in T3 Code once the run finishes.
Expected behavior
Earlier completed assistant replies produced before background-task wakeups should remain visible in the thread timeline (not silently hidden behind the collapsed Worked for ... group), matching how phase: "final_answer" is preserved for OpenCode (packages/provider-opencode/src/server/v2/adapter.ts:1276-1283) and proposed for Claude in #38562.
Actual behavior
Every earlier assistant_message in the run (assistant:0, assistant:1, ...) is collapsed into Worked for <duration> >, leaving only the final assistant_message (assistant:N) visible in the timeline.
If the final wakeup reply is a brief wrap-up referencing an earlier reply from the same run, the user never sees the substantive earlier update unless they manually expand Worked for ....
Root cause
- Web (
apps/web/src/components/chat/MessagesTimeline.logic.ts:796-804) and mobile (apps/mobile/src/lib/threadActivity.ts) keep an earlierassistant_messagevisible across continuations only whenitem.phase === "final_answer":if ( entry.kind === "message" && entry.projectedItem?.item.type === "assistant_message" && entry.projectedItem.item.phase === "final_answer" ) replyEntryIds.add(entry.id);
packages/provider-acp/src/server/adapter.ts(lines 3222-3244) emitsassistant_messageturn_item.updateditems without ever settingphase: "final_answer"anywhere in the adapter.- By contrast:
packages/provider-opencode/src/server/v2/adapter.ts:1276-1283explicitly promotes the closing assistant text tophase: "final_answer"before a background-work report continues the turn (// A report can land right after the answer; the reply to it must not fold the answer away.).- fix(claude): preserve final-answer phases across continuations #38562 adds the equivalent
phase: "final_answer"tagging forClaudeAdapterV2.ts.
Impact
Major degradation or frequent failure (any Antigravity run with background-task wakeups hides earlier replies inside Worked for ...).
Version or commit
0.1.12-nightly.20261121.5024
Environment
- OS: macOS 26.6.2 (darwin/arm64)
- Provider / Model: Antigravity (
antigravity-acp,gemini-4-neon)
Workaround
Instruct the agent to repeat its full, self-contained summary in the final wakeup reply instead of referring to anything written prior to the background task completing.
Related
- fix(claude): preserve final-answer phases across continuations #38562 (same root issue for Claude — preserves
phase: "final_answer"across continuations) - Turn folding hides the substantive answer when a provider ends a turn with a short progress note (Grok) #8879, fix(chat): keep final replies visible after steering #34501, fix(mobile): a final answer followed by more work in the same run stays visible #36116 (related hidden/folded assistant replies)
- Distinct from fix(server): Antigravity send_to_user messages now show and T3 MCP tool labels resolve #38956 (
send_to_usermissingrawOutputon Antigravity) and Antigravity ACP: background-task wakeups silently deny shell commands in full-access mode #40032 (Antigravity background-task wakeup permission denials)
Activity
rowbradley commentedon Nov 21, 2026
Seeing this on 0.1.12-nightly.20261121.5038 with Antigravity 1.5.14 (
gemini-4-neon, desktop + iOS mobile): mid-run replies to steers disappear once later background-task wakes emit another message in the same run. Looking forward to #40559.