[jake-molnia]

This sweep fixes native Antigravity routing, stale gateway discovery, misleading harness configuration, and recovery paths that could leave chats or workspace progress stuck.

• Allow native Antigravity authentication and model discovery alongside the gateway, including credentials across worker replacement. Gateway Google sign-in remains separate from native Antigravity sign-in.
• Refresh compatible gateway models centrally while utility workers sleep. Exclude models explicitly marked as image-output-only, while retaining mixed-output models and unknown metadata. Rebuild both canonical images to publish the new proxy metadata.
• Show bundled harness configuration rows without install/enable switches in the list. Preserve explicit session preferences and native Antigravity configuration.
• Open the composer before passive project-default reads settle. Persist pending defaults across reload, preserve typing and explicit overrides, and block Send before side effects until defaults resolve. Optimistic edits remain immediate.
• Allow provider switching after a failed first startup without native session history. Preserve active-run and negotiated capability checks.
• Order workspace setup snapshots by attempt before sequence, so tracker retention cannot leave the previous attempt visible.
• Return empty preview sessions for unbound drafts while preserving real controller and transport failures.
• Backport upstream 5e5c52b to prevent colliding workflow/check names from hiding a failed CI check.

Validation: integrated focused batches pass with 194, 347, and 104 tests, plus 310 after the immediate-composer follow-up. Affected server, web, client-runtime, and source-control-core typechecks pass. Owners demonstrated failing regressions before fixes. The pinned proxy patch applies cleanly and its focused Go tests pass. Independent review found no blocker in the native/gateway boundaries.

Browser verification uses isolated copied state, not live user data. Desktop comparison confirms five old rows with five toggles become eight configuration rows without list toggles. Native Antigravity fields remain visible; its disabled preference survives reload; the 390px mobile viewport has no document overflow. Download the browser evidence HTML to view embedded before/after screenshots, mobile layout, and the setup-stream replay video. The replay confirms a new sequence-0 attempt replaces an old sequence-9 attempt and rejects a late old event. Delayed-default browser acceptance passes: the mounted draft opens while pending, retains typing, resolves to worktree mode, and preserves an explicit override against a late result.

Live runtime diagnosis measured a controlled fresh cross-worker path at 21.427s and retained seeded resume at 5.4s. T3 uses direct Sandbox resources, not the unrelated devbox warm pool. No runtime/capacity change is included. The original user attempt, duplicate-loading report, and exact active-chat screenshot state were not reproduced. Live Google OAuth and model-turn acceptance remain rollout checks.

Implemented and reviewed with GPT-6 Astra through the Codex harness in T3 Code.

Rollout configuration must add Antigravity to both execution-pool driver allowlists. Native sign-in uses a utility path, but starting a chat enforces the pool list.

[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	20.9 KiB	20.9 KiB	0 B (0.0%)	29.3 KiB	✅
Codex	Live turn messages	2	2	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	21.2 KiB	21.2 KiB	0 B (0.0%)	29.3 KiB	✅
Claude	Live turn messages	2	2	0 (0.0%)	8	✅

Baseline: 5f32ff3 · PR result: 589b65e · 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.