[juliusmarminge]

Problem

When Antigravity auto-approves an invoke_subagent launch, opening its child thread shows only Running invoke_subagent and no transcript. The ACP tool call id is <sessionId>:<n>. The reader took n to be the SQLite steps.idx of the planner step that issued the call. On the Antigravity 1.4.5 runtime that only holds for a launch near the very start of a session.

The runtime's SDK builds the id as make_step_id(trajectory_id, step_index) (google/antigravity/connections/local/event_processor.py). step_index is the harness's index of the step that runs the tool. That index skips planner responses that have neither response text nor thinking.

Reproduced on the real runtime: Antigravity ACP 1.4.5 linux-x64 in auto-approve (yolo) mode, driven against a scripted stand-in for the Gemini API (GOOGLE_GEMINI_BASE_URL). Each parent read a file a few times and then called invoke_subagent. T3's reader returned no children for 3 of 4 real launches on main.

Change

findInvokeRecord now resolves <sessionId>:<n> to the INVOKE_SUBAGENT record whose runtime index is n. That index is the record's idx minus the number of lower planner responses with no text and no thinking. The reader then uses that record.

The reader still returns nothing when it cannot prove the numbering:

• a planner below the record is unreadable, oversized or in another format;
• the id belongs to another session;
• no invoke record sits at that index;
• the new planner scan bound is exceeded (2,000 rows / 32 MiB).

The launch-text path for permission prompts uses the same mapping when an auto-approved sibling claims its record.

Scope and approval

Fixes #31297

Verification

• vp test run apps/server/src/provider/acp/AntigravityChildWork.test.ts: 13 passed.
  • follows each auto-approved launch id to the step that ran it copies a real 1.4.5 session layout step for step, including the reporter's planner that issues view_file and invoke_subagent together. On main it fails with expected [] to deeply equal [ 'Test Runner' ].
  • follows a launch id past many earlier launches covers a launch after 200 earlier ones.
  • declines an auto-approved id when a planner below it is unreadable covers the case where the numbering can't be proven.
  • needs proof that the call id belongs to the parent's record was rewritten for the real numbering: another session's id, an id that ran another tool, and an id past the end all return nothing.
• vp run check:changed: format, lint, knip, typecheck (3 packages) and tests pass.
• Runtime transcript (synthetic data from the scripted endpoint). The ACP ids are as the runtime sent them; the native rows are idx:step_type:

run3 ACP: :1 :2 :3 view_file, :4 invoke_subagent (rawInput {})
run3 steps: 0:14 1:15 2:8 3:15 4:8 5:15 6:8 7:15(invoke, no text) 8:127 9:15 10:101 11:15
run4 ACP: :1 :2 :4 :5 view_file, :6 and :9 invoke_subagent
run4 steps: 0:14 1:15 2:8 3:8 4:15(text) 5:8 6:15(view+invoke) 7:8 8:127 9:15(text) 10:101 11:15 12:127
run5 ACP: :2 :3 :5 view_file, :6 invoke_subagent (planners 1 and 5 carry thinking)

readAntigravityChildWork on main 67eebaca23
  run3 :4 -> []
  run4 :6 -> [Readme Reader]   (coincidence: idx 6 is the planner that issued it)
  run4 :9 -> []
  run5 :6 -> []
readAntigravityChildWork on this branch
  run3 :4 -> [Readme Reader, 3 transcript steps]
  run4 :6 -> [Readme Reader, 3 transcript steps]
  run4 :9 -> [Second Reader, 3 transcript steps]
  run5 :6 -> [Readme Reader, 3 transcript steps]
  run4 :5 (a view_file id) -> []

A separate run on the real runtime also checked:

• two parallel invoke_subagent calls from one planner;
• launches after SYSTEM_MESSAGE and error steps;
• a launch after session/load in a new runtime process.

All resolved to the right child.

• Not checked: real Gemini output (the planner shapes came from the scripted endpoint), runtime releases other than 1.4.5, and macOS or Windows hosts. No render path changed, so there are no captures.

AI-assisted; verified with the tests and tools listed above.

[macroscopeapp]

Approvability

Verdict: Approved at aa83aa7

Macroscope's review found this PR approvable — Updates Antigravity’s subagent view to locate auto-approved launches using runtime step indexes. The change remains confined to session-scoped, read-only display retrieval with existing transcript-path and size safeguards.

No code changes detected at ecbd4e8. Prior analysis still applies.

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

[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	5.0 KiB	5.0 KiB	0 B (0.0%)	6.8 KiB	✅
Codex	Thread snapshot wire	3.8 KiB	3.8 KiB	0 B (0.0%)	4.9 KiB	✅
Codex	Live turn WebSocket wire	1.2 KiB	1.2 KiB	0 B (0.0%)	2.0 KiB	✅
Codex	Live turn WebSocket decoded	21.5 KiB	21.5 KiB	0 B (0.0%)	24.4 KiB	✅
Codex	Live turn messages	1	1	0 (0.0%)	8	✅
Claude	Total thread wire	5.0 KiB	5.0 KiB	0 B (0.0%)	6.8 KiB	✅
Claude	Thread snapshot wire	3.8 KiB	3.8 KiB	0 B (0.0%)	4.9 KiB	✅
Claude	Live turn WebSocket wire	1.2 KiB	1.2 KiB	0 B (0.0%)	2.0 KiB	✅
Claude	Live turn WebSocket decoded	21.9 KiB	21.9 KiB	0 B (0.0%)	24.4 KiB	✅
Claude	Live turn messages	1	1	0 (0.0%)	8	✅

Baseline: 1a2a4a5 · PR result: ecbd4e8 · 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: 108.5 KiB
• Claude decoded thread snapshot: 108.8 KiB

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

[coderabbitai]

Important

Review skipped

Review was skipped as selected files did not have any reviewable changes.

Run configuration

• Source mode: Auto
• Repository: pingdotgg/t3code
• Branch: fix/issue-31297
• File: .coderabbit.config.ts
• Commit: ecbd4e8
• Layers: Repository TS config
• Review profile: CHILL
• Plan: Team
• Run ID: be7fc084-1d90-471f-8ee2-a4d2b4b6221a

Commits

Reviewing files that changed from the base of the PR and between aa83aa7 and ecbd4e8.

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• Trigger review

No actionable comments were generated in the recent review.

Walkthrough

Auto-approved Antigravity task IDs now resolve through runtime step indexes to invoke records. Planner scanning skips responses without text or thinking and stops when data is unreadable or exceeds configured limits. Permission-prompt matching excludes sibling records by resolved invoke-record indexes.

Changes: Antigravity child-work matching

Layer / File(s)	Summary
Runtime step-to-record mapping apps/server/src/provider/acp/AntigravityChildWork.ts, apps/server/src/provider/acp/AntigravityChildWork.testkit.ts, apps/server/src/provider/acp/AntigravityChildWork.test.ts	Planner responses without text or thinking are skipped when mapping runtime step indexes to invoke-record indexes. The scan returns no mapping for unreadable planner data or when it exceeds 2,000 rows or 32 MiB. Tests cover skipped and text-bearing planner steps, consecutive launches, 201 launches, and unreadable data.
Auto-approved and sibling record resolution apps/server/src/provider/acp/AntigravityChildWork.ts, apps/server/src/provider/acp/AntigravityChildWork.test.ts	Auto-approved task IDs resolve through the step mapping and a bounded invoke-record lookup. Permission-prompt matching excludes sibling records by resolved invoke-record index. Tests cover parent-session matching and unresolved steps.

Estimated code review effort: 3 (Moderate) | ~20 minutes

Change: Bug fix · Severity of issue fixed: Low

Possibly related PRs

• pingdotgg/t3code#28629: Adds the child-work reader and the planner-call-ID matching logic that this PR changes.

Suggested labels: bugfix, size:M

Suggested reviewers: t3dotgg, macroscopeapp

Merge Risk: Minimal · previous review aa83a

The child-work matching change is ready to merge after normal checks; no actionable issue was identified.

Pre-merge checks: 4

Check name	Status	Explanation
Title check		The title "fix(server): Antigravity subagent view finds auto-approved launches" follows conventional commit format and names the primary change. The raw summary and objectives show the change maps aut…
Description check		The description covers the problem (auto-approved invoke_subagent launches show no transcript because the ACP call ID suffix was read as steps.idx), the change (findInvokeRecord maps the runtime index…
Linked Issues check		Issue #31297 requires completed auto-approved Antigravity subagent work to appear in the child view. AntigravityChildWork.ts now maps the ACP <sessionId>:<n> suffix to the invoke record at the run…
Out of Scope Changes check		The changes remain within issue #31297. AntigravityChildWork.ts contains the lookup and permission-prompt updates. AntigravityChildWork.test.ts and AntigravityChildWork.testkit.ts add focused fi…

Full details: Title check

Explanation

The title "fix(server): Antigravity subagent view finds auto-approved launches" follows conventional commit format and names the primary change. The raw summary and objectives show the change maps auto-approved tool call IDs to invoke records by runtime step index, so the child view can find launches it previously missed. The title is specific and related to the main change.

Full details: Description check

Explanation

The description covers the problem (auto-approved invoke_subagent launches show no transcript because the ACP call ID suffix was read as steps.idx), the change (findInvokeRecord maps the runtime index to the INVOKE_SUBAGENT record, with bounds and fallbacks), and verification (named focused tests, a check:changed run, and a runtime transcript). It also lists what was not checked. The Scope and approval section links issue #31297 but does not include the maintainer approval comment or explain why a prior discussion is not needed. The template asks for that, so this is a gap. The description also names the model and harness only implicitly, and the template asks for that at the end when an agent was used. The description is mostly complete, so the check passes, with the approval gap noted for the author to address.

Full details: Linked Issues check

Explanation

Issue #31297 requires completed auto-approved Antigravity subagent work to appear in the child view. AntigravityChildWork.ts now maps the ACP <sessionId>:<n> suffix to the invoke record at the runtime step index, while rejecting unproven mappings. The same mapping supports sibling exclusion for launch-text matching. Tests cover skipped planner indexes, later launches, wrong sessions, non-invoke steps, out-of-range indexes, and unreadable planners.

Full details: Out of Scope Changes check

Explanation

The changes remain within issue #31297. AntigravityChildWork.ts contains the lookup and permission-prompt updates. AntigravityChildWork.test.ts and AntigravityChildWork.testkit.ts add focused fixtures and regression coverage for that lookup. No unrelated product area or independent behavior appears in the reviewed diff.

Finishing Touches

Generate docstrings

• Commit to this branch
• Create a new PR

Generate unit tests (beta)

• Commit to this branch
• Create a new PR

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Recent review info

Run configuration

• Source mode: Auto
• Repository: pingdotgg/t3code
• Branch: fix/issue-31297
• File: .coderabbit.config.ts
• Commit: aa83aa7
• Layers: Repository TS config
• Review profile: CHILL
• Plan: Team
• Run ID: 292026f6-f770-456b-a9cc-9620865510eb

Commits

Reviewing files that changed from the base of the PR and between 67eebac and aa83aa7.

Files selected for processing: 3

• apps/server/src/provider/acp/AntigravityChildWork.test.ts
• apps/server/src/provider/acp/AntigravityChildWork.testkit.ts
• apps/server/src/provider/acp/AntigravityChildWork.ts

Included review availability: This review used your included allowance. Your plan provides up to 10 included reviews per hour; 9 remain after this review.

[juliusmarminge]

The automated pipeline stopped on this pull request and will not merge it:

the proof has no capture anyone can open:

• Permanent blocker: Antigravity is unavailable on this VM, preventing the required runtime 1.4.5 Google-authenticated, full-access reproduction.
• The base showing only Running invoke_subagent was not observed.
• The head displaying the completed child's correct role, prompt, and transcript was not observed.
• No captures or driver scripts were produced, so no ghdrop upload was attempted and no public evidence URLs are available.

It stays open for a person to finish or close.