v0.8.68 WhaleFlow: agent checkpoint and resume on timeout/interrupt

[Hmbown]

Problem

Sub-agents time out after 120s with uncommitted work sitting in a worktree. The agent's reasoning — why it chose that approach, what it had left to do — is lost. Manual recovery requires reading the partial diff and guessing the intent. This happened to agent_dbb22444 (#3977, 693 lines uncommitted) and agent_683da1af (#3954) today.

Scope

Add checkpoint-and-resume for sub-agents:

1. Auto-checkpoint — Every N turns (or on tool calls that mutate state), serialize the agent's reasoning state + file diffs to a checkpoint store
2. Continuation prompt — On resume, the agent receives its own checkpoint + a "continue from here" instruction
3. Filesystem durability — Uncommitted worktree changes are committed to a checkpoint branch (not the main branch), so no data is lost on timeout
4. Resume trigger — Orchestrator detects timeout → spawns a new agent with the checkpoint as input
5. Checkpoint cleanup — On successful completion, checkpoint branches are pruned

Non-goals

• Not full session replay. This is single-agent resume, not multi-agent replay.
• Not changing the 120s timeout. That's a provider constraint; checkpointing works around it.

Acceptance

• Agent times out with uncommitted changes → checkpoint is automatically saved
• Resumed agent receives the checkpoint as its initial context
• Resumed agent can continue from where the original left off
• Checkpoint branch is cleaned up after successful completion
• Tests cover: timeout checkpoint save, resume with partial work, cleanup on success

Related

• Sub-agent: crates/tui/src/tools/subagent/mod.rs
• Fleet worker runtime: crates/tui/src/fleet/worker_runtime.rs
• See v0.8.68 WhaleFlow: Conductor agent type for orchestrating agent ensembles #4010 (conductor agent)