fix(cloud): every cloud chat hears where heavy work belongs, not only brand-new ones

[andrewcai8]

A disk census of 11 awake cloud chats (73 GB disks) found the 4 recent chats at about 31 GB with one worktree each. The 7 long-running chats that started before the cloud machine note existed sat at 36 to 61 GB, with 5 to 20 worktrees, up to 17 GB under ~/.t3/worktrees, up to 13 GB in /tmp, and 20 to 71 node_modules directories. They run a build that has the note, but it went only to the first turn of a brand-new native conversation, so they never got it. A long chat can also lose turn-1 context to compaction.

The note now goes on the first turn of each provider session for a native thread. That covers a new chat, a wake, an upgrade restart, an account switch, and a session reopened after 30 idle minutes. cloudMachineNoteFor records which native threads each live session runtime has told, in a WeakMap keyed by the runtime object that ProviderSessionManager.open returns. A new session gets a new runtime object and hears the note again, and the record is dropped along with the session. The other gates are unchanged: cloud box only, never subagents, and the forks sentence only with MCP tools off darwin. The note also gains one sentence asking the agent to remove worktrees, installs and /tmp data it created once it no longer needs them. lowDiskNote is unchanged.

Tests: DiskHeadroom.test.ts checks that a session hears the note on its first turn but not its second, that the session's other native threads and a new session hear it again, and that subagents never do. ProviderTurnStartService.test.ts checks that a native thread that already has turns hears the note when a new session starts. Both failed before the change.

Claude Opus 5.5 (1M context), Claude Code

🤖 Generated with Claude Code

[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	4.9 KiB	5.0 KiB	+23 B (+0.5%)	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	+23 B (+1.9%)	2.0 KiB	✅
Codex	Live turn WebSocket decoded	20.8 KiB	20.9 KiB	+41 B (+0.2%)	29.3 KiB	✅
Codex	Live turn messages	1	2	+1 (+100.0%)	8	✅
Claude	Total thread wire	5.0 KiB	5.0 KiB	−24 B (−0.5%)	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	−24 B (−2.0%)	2.0 KiB	✅
Claude	Live turn WebSocket decoded	21.2 KiB	21.2 KiB	−41 B (−0.2%)	29.3 KiB	✅
Claude	Live turn messages	2	1	−1 (−50.0%)	8	✅

Baseline: bc068c7 · PR result: 33f9226 · 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.