fix(antigravity): keep subagent batches active after launch

[t3dotgg]

Antigravity marked subagents done as soon as start_subagent returned, even while the children were still running. One call can launch several children, and its output is a launch description.

Show each launch as an Antigravity subagent batch. Keep it active for the parent turn, then mark it idle with an explanation that individual agent status is unavailable. Preserve launch failures, cancellation, disconnect handling, and history replay. The shared client fold now displays that status explanation.

Google's current ACP stream omits individual child status, names, and reply ownership. This fixes the false completion and batch label. It cannot provide separate child entries or separate their replies from the parent conversation. The provider docs describe these limits.

Validation:

• 195 focused tests passed across the adapter, protocol, shared subagent state, and mobile work log.
• Server and shared-client typechecks, targeted lint, and formatting passed.
• An isolated run through the official ACP binary launched two agents. The launch returned in 3 ms. The updated adapter kept the batch running through three later child tool completions, then marked it idle when the parent turn ended. It emitted no subagent completion event.
• Browser replay verification confirmed the batch lifecycle. It exposed remaining summary labels, fixed in #9616, which includes before/after screenshots.

Created with GPT-6 Astra (preview) in Codex.

---

Note

Medium Risk
Changes Antigravity subagent event semantics (running vs completed vs idle) that drive Agents UI and work log; behavior is heavily tested but affects user-visible task state.

Overview
Fixes Antigravity treating start_subagent as finished when the launch tool returns, even though one launch can start several agents and its output is only a launch description—not child results.

Antigravity adapter now models each launch as Antigravity subagent batch: it stays running through launch completion and sibling child tool activity, emits task.progress with the launch description, and only emits task.completed on real failed launches (including replayed history). Normal turn end settles open batches to idle via task.updated with an explanation and timelineBypass, instead of interrupted or a fake completion. Historical launches without an active turn get a similar idle update; late updates after settlement are ignored. antigravitySubagentResult is renamed to antigravitySubagentOutput.

Shared client foldSubagentActivities applies detail from task.updated to agent progress, so the idle message shows in Agents / work log. User docs describe batch behavior and ACP limits (no per-child status or separate entries).

Tests were rewritten/added across the adapter, protocol, and subagent fold to match this lifecycle.

^{Reviewed by Cursor Bugbot for commit 6b05620. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Keep Antigravity subagent launches active as idle batches after turn end

• Active native start-subagent calls are now treated as batch launches that stay running during the parent turn instead of being reported as completed child results. Failed launches still complete with bounded output.
• When a parent turn ends normally, finishSubagents in AntigravityAdapter.ts settles unfinished batch tasks as idle rather than interrupted, adds an "individual agent status unavailable" message, and marks the update for timeline bypass. Cancelled and failed settlements retain their statuses.
• The turn settlement handler no longer synthesizes a missing-result error for normally completed turns with unfinished subagent work; only provider-supplied turn errors are forwarded.
• foldSubagentActivities in subagentRuntime.ts now reads a non-empty detail field from task.updated activities and stores it as bounded progress text.
• Behavioral Change: subagentLinkage task titles changed from the individual-subagent label to the batch label; normal turn-end status changed from interrupted to idle; completed historical launches now emit task.updated (idle) instead of task.completed.

^{Macroscope summarized 6b05620.}

[github-actions]

Thread transfer impact

✅ Thread transfer remains within every enforced ceiling.

Provider	Metric	Main baseline	This PR	Impact	PR ceiling
Codex	Total thread wire	13.8 KiB	13.8 KiB	+3 B (+0.0%)	15.1 KiB	✅
Codex	Thread snapshot wire	7.0 KiB	7.0 KiB	−5 B (−0.1%)	7.3 KiB	✅
Codex	Live turn WebSocket wire	6.7 KiB	6.7 KiB	+8 B (+0.1%)	7.8 KiB	✅
Codex	Live turn WebSocket decoded	58.5 KiB	58.5 KiB	0 B (0.0%)	66.4 KiB	✅
Codex	Live turn messages	10	10	0 (0.0%)	21	✅
Claude	Total thread wire	13.6 KiB	13.7 KiB	+161 B (+1.2%)	15.1 KiB	✅
Claude	Thread snapshot wire	7.0 KiB	7.0 KiB	−4 B (−0.1%)	7.3 KiB	✅
Claude	Live turn WebSocket wire	6.5 KiB	6.7 KiB	+165 B (+2.5%)	7.8 KiB	✅
Claude	Live turn WebSocket decoded	57.9 KiB	59.3 KiB	+1.4 KiB (+2.5%)	66.4 KiB	✅
Claude	Live turn messages	10	10	0 (0.0%)	21	✅

Baseline: 0cb02ab · PR result: 6b05620 · Source CI: success

Scenario and decoded snapshot size

10 historical turns, 5 command tools per turn, 878.9 KiB retained MCP result per historical turn, and a 1.05 MiB retained result in the measured turn.

• Codex decoded thread snapshot: 113.8 KiB
• Claude decoded thread snapshot: 114.5 KiB

Updated in place by a trusted workflow. PR artifacts are strictly validated and never executed.

[macroscopeapp]

Approvability

Verdict: Approved at 6b05620

Macroscope's review found this PR approvable — This is a narrowly scoped Antigravity lifecycle fix that keeps batch entries active until the parent turn ends, accurately represents unavailable child status, and ignores stale updates. The production impact is confined to subagent task-state presentation, with focused tests covering the revised behavior.

^{You can add or adjust custom eligibility rules. Learn more.}

[t3dotgg]

Note

🤖 GPT-6 Astra (preview) responding on behalf of Theo

Browser verification confirmed that batches stay active after launch and become idle when the parent turn ends. It found a remaining mismatch: the conversation summary still said completed while Agents showed idle.

PR #9616 fixes that summary, counts launches as batches, and preserves the inactive status explanation on mobile. It includes before/after screenshots from isolated fixtures generated by replaying recorded ACP traffic through the adapters.