Skip to the content.

Harness verification (live-docs, mid-2026)

Verified against official docs and live tools by the harness-verify workflow. Each harness-specific fact must live behind that adapter’s targetSchema, so an upstream change is a version bump rather than a rewrite (spec §2, §8). Where a fact is UNKNOWN, the driver must degrade honestly (report UNTESTED) rather than guess.

The biggest correction to the spec: structured tool-call traces are not uniformly available. Claude/Codex/Cursor/OpenCode expose them only on a specific output mode, and Copilot has none from headless -p. trace assertions degrade to output assertions on harnesses without a trace (spec §14).

Trace availability matrix

Harness Headless command (trace mode) Structured trace?
claude claude -p <prompt> --output-format stream-json --verbose yes (NDJSON)
codex codex exec --json <prompt> yes (JSONL)
cursor cursor-agent -p <prompt> --force --output-format stream-json yes (NDJSON)
copilot copilot -p <prompt> -s (+ --allow-tool/--allow-all) no -> degrade
opencode opencode run <prompt> --format json or SDK /event SSE yes (JSONL/SSE)

Claude Code (@anthropic-ai/claude-code, v2.1.168) — confidence: high

OpenAI Codex (codex) — confidence: high

Cursor (cursor-agent) — confidence: high

GitHub Copilot CLI (@github/copilot) — confidence: high

OpenCode (opencode, @opencode-ai/sdk) — confidence: high

Official MCP Registry (federation) — confidence: high

Sources: official docs + live API per harness; full URLs in the workflow result tasks/weyjbun3r.output.