feat: write parent session context into subagent session_meta for distributed tracing

Summary

When Claude Code spawns a subagent via the Task tool, it creates a new session with a separate JSONL transcript file. The subagent's transcript marks messages with isSidechain: true and an agentId, but contains no reference to the parent session that spawned it.

This makes cross-session distributed tracing impossible for observability tools — the parent and child transcripts are two isolated islands with no shared key.

Current Behavior

Parent session transcript (~/.claude/projects/<project>/<parent-session>.jsonl):

{"type":"assistant","message":{"content":[{"type":"tool_use","id":"call-xyz","name":"Task","input":{"description":"..."}}]}}
{"type":"user","message":{"content":[{"type":"tool_result","tool_use_id":"call-xyz","content":"..."}]}}

Child session transcript (~/.claude/projects/<project>/<child-session>.jsonl):

{"type":"assistant","isSidechain":true,"agentId":"subagent-1","message":{...}}

The child session has no parent_session_id, no parent_turn_id, no parent_tool_call_id. An external tool reading both files cannot deterministically link them.

Requested Change

When Claude Code creates a subagent session (i.e. spawns via Task), write the parent context into the child session's first session_meta entry:

{
  "type": "session_meta",
  "sessionId": "<child-session-id>",
  "parentSessionId": "<parent-session-id>",
  "parentTurnId": "<turn-id-of-the-spawning-turn>",
  "parentToolCallId": "<tool_use_id-of-the-Task-call>",
  "isSidechain": true,
  ...existing fields...
}

Alternatively, if session_meta is not the right place, any structured first-line entry in the JSONL that external readers can reliably locate is acceptable.

Why This Matters

Without this link, observability tools face two bad options:

1. Timestamp guessing — scan sibling sessions for isSidechain: true entries whose start time falls within the parent's Task call window. Breaks when multiple subagents are spawned concurrently.
2. No linking at all — parent and child traces are forever disconnected.

With parentSessionId + parentToolCallId in the child's metadata, tools can:

• Reconstruct the full agent execution tree
• Aggregate token costs across the entire hierarchy
• Attribute failures to the correct spawning turn
• Build flame-graph style views of multi-agent workflows

Context

This request comes from Spanory, an open-source observability tool for Claude Code and other agentic runtimes. It emits OTel spans and Langfuse traces from Claude Code session transcripts. Parent-child linking is the last gap preventing full distributed tracing support.

The companion issue tracking Spanory's side of this work: https://github.com/Bububuger/spanory/issues/1

Proposed Fields (non-normative)

| Field | Type | Description |
|---|---|---|
| parentSessionId | string | Session ID of the spawning parent session |
| parentTurnId | string | Turn identifier within the parent session |
| parentToolCallId | string | tool_use_id of the Task call that created this session |
| isSidechain | boolean | Already exists on message entries; ideally also on session_meta |
| agentId | string | Already exists; the subagent's own identifier |

Backward Compatibility

Adding new fields to session_meta is additive. Existing readers that don't know these fields will ignore them. No breaking change.