[FEATURE] Add invocation context/provenance to PreToolUse hook payload for Task tool

Problem

When using PreToolUse hooks with the Task matcher, there is no way to determine what triggered the Task tool call. Specifically, hooks cannot distinguish between:

• A Task spawned directly by the main agent in response to user input
• A Task spawned as part of a Skill execution

This matters because hooks that enrich Task prompts (e.g., injecting session context, standards, or additional instructions) may be unnecessary or even counterproductive when the Task is spawned by a skill that already has its own well-defined prompt.

Current behavior

The PreToolUse hook payload for Task contains:

{
  "session_id": "...",
  "transcript_path": "...",
  "tool_name": "Task",
  "tool_input": {
    "prompt": "...",
    "description": "...",
    "subagent_type": "..."
  }
}

There is no field indicating why or by whom the Task was invoked.

Proposed solution

Add an invocation_context (or similar) field to the hook payload:

{
  "tool_name": "Task",
  "invocation_context": {
    "triggered_by": "skill",
    "skill_name": "sync-branches"
  },
  "tool_input": { ... }
}

This would allow hook scripts to make context-aware decisions, e.g.:

triggered_by="$(jq -r '.invocation_context.triggered_by // "agent"' <<<"$payload")"
if [[ "$triggered_by" == "skill" ]]; then
    # Skip enrichment for skill-originated tasks
    exit 0
fi

Workarounds considered

| Approach | Limitation |
|---|---|
| Parse transcript for recent Skill tool calls | Fragile, timing-dependent, expensive |
| Pattern-match Task description/prompt | Requires maintenance per skill, not reliable |
| Sentinel file via skill frontmatter hooks | Race condition (skill hooks and global hooks fire in parallel) |
| Skill frontmatter hook override | Not possible — skill hooks are additive, not replacements |

None of these are robust.

Use case

A common pattern is having a PreToolUse hook that injects session context and agent standards into every Task prompt. This is valuable for user-initiated tasks but adds unnecessary token overhead and noise for skills that already provide complete, self-contained prompts to their Task agents.