Manual end-to-end test fixtures for ctxcop's harness adapters. One
subdir per scenario. These complement (don't replace) the Go unit
tests under internal/.
Use these when:
- A new Claude Code / Codex / Cursor / Pi / OpenCode release lands and you want to confirm ctxcop's hook integration still holds end-to-end.
- We're investigating an empirical question about a harness's behavior that can't be answered from docs alone (see #3).
- We need a regression fixture for an issue that surfaced only with a real harness payload (see #5).
Each subdir is self-contained and has its own README with the run procedure.
- docs/agent-egress-bench.md — how ctxcop scores against the applicable (content-DLP) slice of the external agent-egress-bench corpus, with a per-case breakdown and the content-vs-network scope boundary. Reproduce steps included; do not run AEB's network gauntlet runner against ctxcop.
- elicit-mcp/ — Claude Code
ElicitationResulthook; minimal MCP server with two elicitation tools (neutral + verbatim) for confirming whether form-submission content reaches the model's transcript / context window.