← 2.1.2 Test inconclusive · runtime-test

Combo Test: Large Output Analysis in Fork — runtime test

Hands-on runtime battle-test of Combo Test: Large Output Analysis in Fork. Result: INCONCLUSIVE.

Combo Test: Large Output Analysis in Fork

A forked Explore skill successfully analyzed a 1MB persisted output file while keeping main context clean, validating the combination of large-output persistence, fork isolation, and delegation.

Test Setup

The test combined three Claude Code 2.1.x features: Large outputs to disk (2.1.2), fork context isolation (2.1.0), and the Explore agent skill (2.1.0). A 10,000-line (~1MB) test file was persisted to disk by an earlier test. The combo test created a new skill, analyze-large-output, configured with context: fork and agent: Explore, then invoked it with the path to the persisted file.

What Happened

The forked Explore skill read the entire 1MB file and analyzed its content without polling back to the parent. It counted lines (10,000), identified the pattern (identical repeated boilerplate across all lines), and returned a structured markdown summary table. The key finding: the parent context received only the summary (~500 bytes), not the full file or intermediate reads.

Why It Matters

This validates a critical cost pattern: large data processing can be offloaded to a forked, cheaper model (Haiku running Explore) while the parent thread (typically Opus) stays focused on synthesis. The full original file remains available on disk for follow-up deep dives. The pattern enables log analysis, codebase scanning, and large dataset processing without context pollution.

Caveats

The test used synthetic test data (identical boilerplate across 10,000 lines), not real-world outputs. Explore's analysis capability on heterogeneous or high-entropy data remains untested in this combo scenario.

Primary source
⎘ 2.1.2/tests/combo-01-large-output-fork/TEST-RESULTS.mdverbatim from the corpus

Combo Test: Large Output Analysis in Fork

Hypothesis

Can a forked Explore skill analyze a persisted large output file without polluting main context?

Features Combined

  • Large outputs to disk (2.1.2)
  • context: fork (2.1.0)
  • agent: Explore (2.1.0)

Test Method

  1. Earlier test persisted 1MB (10,000 line) output to disk
  2. Created skill analyze-large-output with context: fork and agent: Explore
  3. Invoked skill with path to persisted file
  4. Verified fork could read and analyze the file

Results

Status: VALIDATED ✅

The forked Explore skill:

  • Successfully read the 1MB persisted file
  • Analyzed content (counted lines, identified patterns)
  • Returned structured summary
  • Main context received only the summary (~500 bytes)

Skill Output

| Metric | Result |
|--------|--------|
| Total Lines | 10,000 |
| First 3 Lines | Lines 1-3: Identical repeated test content |
| Last 3 Lines | Lines 9998-10000: Same pattern continues |
| Pattern | 10,000 lines of identical boilerplate |
| Summary | Synthetic test data for persistence feature |

Implications

  1. Context efficiency - Fork analyzes 1MB, main sees ~500 bytes
  2. Cost savings - Explore (Haiku) does the heavy reading
  3. Full data preserved - Original file available for deep dives
  4. Pattern enables: Large log analysis, codebase scanning, data processing

Test Skill Created

.claude/skills/analyze-large-output/SKILL.md

Status: VALIDATED ✅

Evidence & receipt
◇ ed25519 receipt
idtest_122a849dc3e4a9e896db066e
alged25519
pubkey9b87705613b1e2fd064d57fa75a6b679d2856ceafad6b1daa8f982493871b6dd
sig60f2d2c3926e1ebe7d9998f7c38c660c14d24198d6ed8ad6da025ae29d25f22794aff44e6a242d7ed7337daa60b01c81b762ab70d00e8d039838338dbbb33904

Signed with an ed25519 key held off the repo. Anyone can verify against the published public key; nobody without the secret key can forge it. Click verify: it recomputes the signature in your browser. The signature proves integrity and authorship of this exact content — not a third-party timestamp or that the underlying claim is objectively true. signedAt is when the @f3/attest pipeline ran, not when the work happened; the evidence refs carry the source dates.

Connected