The harness-specific half of a run trace: how a real model's raw stdout maps
to the common fields. Claude Code's parseClaudeRun reads its stream-json; a
second harness (Codex) supplies its own parser of codex exec --json JSONL, so
the eval tier (measureTriggerRate/runEval) isn't bound to Claude's format.
The non-harness fields (cwd/exitCode/stdout/file/sh) stay in makeContext.
The harness-specific half of a run trace: how a real model's raw stdout maps to the common fields. Claude Code's
parseClaudeRunreads its stream-json; a second harness (Codex) supplies its own parser ofcodex exec --jsonJSONL, so the eval tier (measureTriggerRate/runEval) isn't bound to Claude's format. The non-harness fields (cwd/exitCode/stdout/file/sh) stay inmakeContext.