[FEATURE] Task-level lifecycle hooks (PreTask / PostTask) for per-task analytics

Summary

Current hooks operate at session level (SessionStart) or tool level (PreToolUse/PostToolUse). There is nothing in between — no signal when a logical unit of work starts or ends within a session.

A single session often contains multiple tasks: "fix the bug", "add tests", "refactor the helper". Without task-level boundaries, plugins cannot compute per-task metrics.

Current hook granularity

Session level:    SessionStart ─────────────────────── Stop
                       │
Task level:        (nothing)                            ← THE GAP
                       │
Turn level:        UserPromptSubmit ─── UserPromptSubmit
                       │                      │
Tool level:     Pre/PostToolUse         Pre/PostToolUse

Proposed: PreTask / PostTask hooks

TaskStart

Fires when Claude begins working on a distinct task.

{
  "event": "TaskStart",
  "task_id": "...",
  "session_id": "...",
  "prompt_preview": "Fix the null pointer in parser.rs"
}

TaskEnd

Fires when a task reaches a natural conclusion.

{
  "event": "TaskEnd",
  "task_id": "...",
  "outcome": "completed",
  "turn_count": 8,
  "tool_calls": 5,
  "duration_seconds": 180
}

What defines a "task"?

Possible heuristics (any combination):

• User explicitly labels it: "Now fix the login bug"
• Significant topic shift in user prompt
• User invokes a workflow command (slash command = new task boundary)
• Plugin declares it via a hook response field

A pragmatic first step: let plugins signal task boundaries, and fire TaskEnd when the next TaskStart occurs or the session ends.

Why this matters

Per-task cost: "Fixing the bug cost $0.35 / 8 turns. Adding tests cost $0.12 / 3 turns."

Per-task FTR: Session-level FTR is meaningless with 5 tasks. Need: "3 of 5 tasks first-try."

Pattern detection: "Tasks starting with analysis complete 22% faster. Tasks that skip it have 3x corrections."

Related

• #47612 — Pre-exit hook (session-level, different scope)
• #11008 — Token usage in hooks (complementary — tokens + task boundaries = per-task cost)