Issue-driven era. Anchors: issue #3 (dsh investigation: open decomposition, CPU profile of the 1.3s dsh --profile acp boot — a module-resolution I/O storm inherent to dsh; NODE_COMPILE_CACHE measured and rejected) and issue #4 (opencode: open 2.5s = ~1.6s binary boot + ~0.93s per-process lazy first session/new; idle RSS ~316MB — the heaviest of the set). Daily-work orientation lives in the repo's AGENTS.md feature index.

Issues #2's per-spawn costs were gone, but opening a chat channel still paid the adapter process boot on every click, because the daemon spawns a fresh adapter per channel:

target	open (steady)	spawn+initialize	note
deepseek	~1.5s	~1.30s	the bulk
claude-code	~1.2s	~0.4s (wrapper)	its CLI child boots at session/load regardless
codex	~2.3s	~2.1s	wrapper shape
opencode (#4)	~2.5s	~1.6s	native binary boot; slowest of all

The structural fix is the reference portal's resident model: boot the adapter BEFORE the click.

Pre-warm to "initialized adapter, no session" when the user is about to need it; consume it on chat:session.open:

• machines.chat_prewarm JSON column (migration 16), written through the ordinary PATCH /api/machines/:id (chatPrewarm arm, owner-or-admin + 404-hiding — the same gates as rename/remote-chat). Redesigned same-day from the first instance-global cut: the idle processes live on the machine, so its owner decides.
• Absent = DEFAULT_CHAT_PREWARM_SETTINGS: deepseek ON; claude-code/codex/opencode OFF (opencode #4: biggest absolute saving after codex, but ~316MB idle argues for opt-in).
• The map is strict replace-semantics (all four keys; a legacy 3-key payload 400s). The web normalizes {...DEFAULT, ...machine.chatPrewarm} before sending, which heals pre-#4 rows on first toggle.
• The toggle lives in the machine page's 代理 tab — one Switch per Agent runtime card (PrewarmToggle in agents-tab.tsx; only the pi card renders none). PREWARM_ADAPTER_TARGETS (shared, re-exported by the SDK alongside DEFAULT_CHAT_PREWARM_SETTINGS) enumerates the pool targets.

• AgentSession mount → chat:adapter.prewarm {agentInstanceId} (browser→server, owner-gated like open, ack {accepted} — false when off/unsupported/offline) → chat:adapter.prewarm {target} over /ctl → the daemon pool.
• chat:session.open also stamps the current switch onto chat:session.start as an optional prewarm: true so the daemon re-arms AFTER a channel took one.

• At most ONE prewarmed adapter per target, spawned to a completed initialize (the payload carries the FULL AcpAgentConnection.start result — caps and agent identity were already paid).
• consume() on chat:session.start skips spawn+initialize and goes straight to session establishment. Pending (mid-boot) entries are NOT consumed — the channel spawns fresh, the prewarm stays for the next click.
• Idle TTL HN_PREWARM_TTL_MS (default 120s; 0 disables prewarm) kills ready entries; teardown on /ctl loss (teardownAll) and the W11 boot sweep/audit own the rest.

• The prewarm arms the W7.1 event tap AT PREWARM TIME (own listener + --patch + env), so a consumed prewarm keeps tap-grade streaming; a failed handshake closes the listener and the channel degrades to the transcript tail exactly like a tap-less spawn. The tap is dsh-ONLY (armTap no-ops elsewhere) — an opencode adoption is a plain native-ACP connection.

• Prewarms ledger under prewarm-<target> pseudo ids (the WIRE_ID regex allows no colons — hence the hyphen); adoption DELETES the pseudo file and re-ledgers under the channel's id with the same pgid; TTL/teardown kills deliberately never unlink (audit-only, per W11 A). deleteAdapterLedgerEntry exists for exactly this rename.

target	cold	prewarmed
deepseek	1.5s	59–206ms (fully warm pool vs settled boot)
codex	2.3s	199ms
claude-code	1.2s	666ms
opencode (#4)	2.5s	846–887ms

Switches off → baseline unchanged (1.48s dsh). The remaining ~0.9s on opencode is its per-process lazy first session/new — deliberately NOT warmed: a throwaway warm-up session would persist a junk row in opencode's session store (ACP has no session delete). Daemon 0.24.0-i4.

• ChatHandlersHandle.prewarmReady(target) exposes pool readiness for deterministic waits; cli/test/prewarm.test.ts covers pool semantics (dedupe, consume-once, pending-miss, TTL, dead-replace, late-resolve kill, teardownAll).
• The server suite covers the strict-map PATCH semantics (incl. the legacy 3-key 400) and each target's independent switch; smoke covers the machine-level PATCH flow end to end.