TaskOutput returns full JSONL transcript instead of agent's final response

Summary

When using TaskOutput to retrieve results from background agents (run_in_background=true), the tool returns the entire agent conversation transcript as a serialized JSONL string instead of just the agent's final response text.

Reproduction

1. Spawn a background agent:

Task(
    subagent_type="general-purpose",
    prompt="Do something and return a summary",
    run_in_background=True
)

1. Retrieve results:

TaskOutput(task_id="<agent_id>", block=True)

1. The returned task.output field contains the full JSONL conversation history (~200-500KB) instead of the agent's final text response (~500 bytes).

Expected Behavior

TaskOutput should return only the agent's final assistant message text, e.g.:

Output: outputs/my-file.md
Summary:
- Finding 1
- Finding 2
Confidence: 0.92

Actual Behavior

TaskOutput returns a serialized JSONL string containing all messages from the agent's sidechain conversation:

{"parentUuid":null,"isSidechain":true,"type":"user","message":{...}}
{"type":"assistant","message":{...}}
{"type":"user","toolUseResult":{...}}
... (40+ messages, 200-500KB total)

Impact

• Context bloat: Each background agent result consumes 200-500KB instead of ~500 bytes
• Transcript size: A session with 2 background agents had a 2.1MB transcript instead of ~200KB
• Token waste: Unnecessary context consumption affects cost and may cause truncation

Evidence

From session a97c8078-bf8e-4c09-b1e2-bda06542e91f:

• Line 107: 457KB TaskOutput result (should be ~500 bytes)
• Line 111: 514KB TaskOutput result (should be ~500 bytes)
• Both contained 42 JSONL messages each

Environment

• Claude Code version: 2.1.19
• OS: macOS (Darwin 24.6.0)

Workaround

The orchestrator can work around this by having agents write to output files and reading those files directly, ignoring the bloated task.output field.