Combo Test: Large Output Analysis in Fork

Combo Test: Large Output Analysis in Fork demonstrates that a forked agent can analyze persisted large outputs (1MB+) without sending raw data to the main context, preserving token budget while enabling deep inspection.

How It Works

Claude Code 2.1.2 persists oversized outputs to disk rather than embedding them in context. The test invoked a forked Explore skill (Haiku) with the file path. The fork read and analyzed the full 1MB file directly from disk, then returned only a structured summary (~500 bytes) to main context. Main context never handled the raw output.

Test Setup and Results

The test combined three primitives:

• Large outputs to disk (2.1.2 feature)
• Forked context (2.1.0)
• Agent skill Explore (Haiku runner)

Method: Earlier tests had persisted a 10,000-line synthetic file (1MB). The combo test created skill analyze-large-output with context: fork and agent: Explore, then passed the file path.

Status: VALIDATED. The fork successfully:

• Read the persisted file without network or context transfer
• Counted 10,000 lines and identified the boilerplate pattern
• Returned a 3-row metric table summarizing findings
• Main context consumed ~500 bytes instead of 1MB

Why It Matters

Context partitioning via fork enables full data preservation with cost control. Analysis runs on Haiku (cheaper tokens) in isolated context, leaving main context untouched for synthesis. Enables practical large-output workflows: log analysis, codebase scanning, data pipeline inspection.

Caveats

Test used synthetic boilerplate data (10,000 identical lines). Real-world outputs may have structure complexity Haiku misses; users should validate summarization against source when stakes are high. Fork incurs context-switch overhead (~3-5KB baseline per fork spawn).