Proposal: remove the subagent-checkpoint machinery (pre-task/post-task/post-todo) — likely unneeded

Proposal

Evaluate removing the subagent / task-checkpoint subsystem — the pre-task, post-task, and post-todo hooks and everything they feed. Investigation suggests it produces data that nothing consumes.

Raised as the deferred follow-up from #1806 (which repaired the Claude Code matchers so these hooks fire again). This issue is the "do we even want this granularity?" decision.

What this subsystem is

• pre-task — snapshots untracked files before a subagent runs (baseline helper)
• post-task — on subagent finish, writes a task checkpoint to the shadow branch bundling the subagent's file diff + its own transcript (agent-<id>.jsonl) + attribution (subagent_type, task_description, agent_id, tool_use_id)
• post-todo — cuts incremental checkpoints during a long subagent run

Backed by shared machinery: handleLifecycleSubagentStart/End (cmd/entire/cli/lifecycle.go), SaveTaskStep / TaskStepContext (cmd/entire/cli/strategy/), task-checkpoint write paths (cmd/entire/cli/checkpoint/ephemeral.go, persistent.go), and pre-task state (cmd/entire/cli/state.go).

Why it looks unneeded

The product consumes session-level data only (checked in entire.io):

• Ingestion reads full.jsonl / transcript.jsonl for session index 0 (api/src/lib/checkpoint-payload-hydrator.ts). Subagent transcripts (agent-*.jsonl) are never read.
• is_task / tool_use_id are parsed into SessionInfo once (api/src/lib/github.ts:660-661) then never queried or displayed.
• subagent_type, task_description, is_incremental, incremental_sequence, agent_id, is_task_checkpoint → 0 matches anywhere in the product. No UI renders subagents.

Local consumers are dead or cosmetic:

• rewind (the one feature using subagent-boundary checkpoints) is deprecated and slated for removal.
• Remaining use is [Task] labels in checkpoint list / explain — navigational garnish.

Cross-agent: the machinery is shared by Claude Code, Cursor, Factory AI Droid, and Copilot CLI. On Claude Code it was dead (the #1806 matcher bug); on the other three it fires today but produces the same inert task checkpoints.

What we'd lose: nothing consumed

• Code the subagent writes is still captured by the Stop / turn checkpoint (whole-worktree). No code loss in the common single-worktree flow.
• Subagent transcript, per-subagent rewind points, attribution, per-subagent tokens — none are read by the product or any live command.

One honest caveat: subagents in an isolated worktree aren't covered by Stop, and post-task was the only thing that would capture them. But that protection doesn't effectively exist today (hooks were broken/unconsumed), so removal doesn't regress vs current reality. If isolated-worktree subagent capture ever becomes a real requirement, it's a fresh, correctly-scoped feature.

Removal scope (if approved)

• Hook install + lifecycle parse across the 4 agents (subagent-start/stop, pre-tool-use/post-tool-use, Claude matchers)
• Shared dispatch cases (SubagentStart/SubagentEnd in lifecycle.go; subagent classification in hook_registry.go)
• Strategy: SaveTaskStep, TaskStepContext, task-checkpoint write paths in checkpoint/ephemeral.go + persistent.go
• pre-task state (state.go), post-todo handler
• The deprecated rewind task-checkpoint branch (already going away)
• Tests + docs across all four agents

Decision needed

Keep (bet on a future subagent-level product feature — the dead isTask parse suggests it was once intended) or remove (delete unconsumed cross-agent complexity).

If we're not going to surface subagent-level data in the product in the foreseeable future, the recommendation is remove.

/cc — deferred from #1806

Start by reading the lifecycle and strategy paths in cmd/entire/cli/lifecycle.go and cmd/entire/cli/strategy/, then inspect checkpoint/ephemeral.go, persistent.go, state.go, and api/src/lib/checkpoint-payload-hydrator.ts. Confirm the product and live commands do not consume task-level data before deciding whether the cross-agent hooks and related tests and docs can be removed. Done means the decision is implemented consistently across the four agents without losing ordinary Stop checkpoint capture.