cfahlgren1/web-fetch-harness-traces

Native web fetch harness traces

Separate, lightly sanitized native JSONL traces comparing URL-fetch behavior in Claude Code and Codex CLI against:

https://huggingface.co/datasets/nyu-mll/glue

Captured on 2026-09-15. No shell HTTP client, browser automation, or MCP fetcher was used.

Files

• data/claude-code.jsonl: Claude Code's stream-json events.
• data/codex.jsonl: Codex's persisted native rollout JSONL.

Both traces are exposed together in the default subset and train split.

The files retain each harness's native event envelopes, ordering, identifiers, usage metadata, tool calls, tool results, and final responses. Codex developer and injected harness context, base instructions, turn context, world state, skill warnings, local working-directory metadata, and reasoning records were removed. No common schema or normalized trace wrapper was introduced.

Interpretation

• Claude Code asked WebFetch for the converted page verbatim and copied the returned tool result byte-for-byte as its final response. The WebFetch helper still returned a prompt-conditioned summary rather than the source document.
• Codex's Responses API web_search URL-open action exposed a transformed, line-numbered representation of the page rather than raw HTML. Its CLI event stream kept the fetched document opaque and exposed the model's report about it. These are observations from the named versions and configuration at capture time, not permanent API contracts.