[Juliusolsson05]

Problem

ensureSessionLive treated "the backend did not report input-ready within 30 s" as "the backend failed to start": it killed the process and marked the pane failed (session.ts readiness branch). On a restart with a dozen CLIs starting at once, opencode TUIs (readiness = first PTY byte + 250 ms) and Codex resumes of large rollouts take longer than that. Lifecycle journal, run 75117: kill.request cause=live-entry at 35.6–35.8 s ×3 and at 69.5 s in run 95737 — against processes that were alive (no exit in between); the same sessions re-woken by hand became ready in 1.7–4.1 s. The pane then sat dead until the user retried it, observed 94 s, 217 s and 556 s after boot. This is the "resuming a terminal on boot is like a 3 min process" and "opencode terminal resume is fucked up" reports (#772; investigations in the boot audit and the opencode report).

Two amplifiers: AgentTerminalLeaf wakes on mount while the rehydrate recovery is already in flight, main joined that recovery but still reported disposition: 'spawned', so the leaf's wake inherited the kill timer; and the leaf retried its attach only after that wake resolved, so a raw terminal pane was blank for the whole TUI boot even when it succeeded.

Behaviour

• A readiness timeout never kills a live backend. waitForSessionInputReady already rejects immediately for a backend that exited or failed, so reaching the deadline means the process is alive and slow. The wake now resolves ok, the pane keeps its status and stays not-ready (readiness arrives whenever the composer does), and the journal records wake.result ok:true code:'ready-timeout'. An observed exit or start failure still fails the wake and kills a spawned backend.
• Joining is not spawning. SessionManager.recover returns disposition: 'joined' when it dedupes onto an in-flight claim (SessionRecoverResult widened); the renderer treats joined + live like adopted and skips the wait. The leaf therefore attaches as soon as the recovery it joined resolves.
• The raw terminal never waits. ensureSessionLive takes { awaitInputReady?: boolean }; AgentTerminalLeaf passes false at both wake sites. providerRuntime === 'terminal' sessions skip the wait everywhere.
• rehydrate's own 30 s deadline on the recover RPC is unchanged: its consequence is a cancelled claim and a retryable pane, not a kill.

Design decisions

• Not raising the constant. A longer deadline still kills at some load; the destructive consequence was the bug, not the number. The "accepted cost" note already present in the code (a wedged provider now sits not-ready instead of being self-healed) applies to this class too and is stated there.
• joined over renderer-side detection: only main knows whether a recover call started the process; the disposition is the honest signal, and it makes the leaf's attach follow backend liveness for free.
• Wake single-flight ignores the option: a terminal pane joining a composer's wake resolves when that one does; since the wait can no longer kill, joining a waiting wake costs latency, never the backend.

Linked issues

Fixes #772. Refs #774 (spawn herd), #775 (boot-path I/O), #596, #548.

Verification

• sessionRecovery.renderer.test.tsx (+3): a spawned backend alive but not ready past the deadline → wake resolves, killOwnedSession not called, pane started/not-ready; a backend that exits before ready → wake rejects, kill issued, pane failed; terminal-runtime and awaitInputReady: false wakes resolve without the wait.
• sessionManager.wake.test.ts: the second wake that joins an in-flight recovery gets disposition: 'joined'.
• AgentTerminalLeaf.dimensionOwnership.renderer.test.tsx (+1): a spawning backend is woken with { awaitInputReady: false }.
• --project renderer src/renderer/src/workspace/hook/actions/ + the leaf suite — 9 files, 41 tests passed; --project unit wake + rehydrate — 6 passed; npx tsc -b — clean.

Limitations / follow-ups

• A provider that registers and then wedges before its composer now sits not-ready with no in-app self-heal (pre-existing trade for adopted backends, now also for spawned ones); the pane's readiness readout shows the elapsed time.
• The boot herd itself (perf(boot): bound and prioritise rehydrate spawns instead of launching every pane at once #774) and boot-path I/O (perf(boot): keep retention pruning, feed-debug flushing, workflow service start and setup checks off the boot critical path #775) are separate; this PR removes the destructive consequence, not the slowness.
• Not verified in the running app; the journal fingerprint (code:'ready-timeout' on an ok wake) is there to confirm on the next restart.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Kk16MNVJtWAnCeGxGRuHqa

Current integration verification (2026-09-04)

Integrated merged main feb6690 without discarding the dispatcher/WebGL/engagement/resize tests. Head 87ca7e9 passes both full host CI gates. Focused checks: 14 renderer tests plus six session-manager wake tests. Original live app and Bringdown were not restarted. Ready for owner merge confirmation.