verify

oleh anthropic

Verifikasi perubahan harness secara end-to-end tanpa docker — jalankan CLI pinned nyata terhadap server stub penangkap header dengan env resolve_auth_env() yang tepat…

npx skills add https://github.com/anthropics/defending-code-reference-harness --skill verify

Unduh ZIP GitHub

6.9k

Verifying harness changes on a docker-less host

The pipeline's real surface is the in-container claude -p process and its outbound API requests. Without docker, drive the same pinned CLI binary directly with the env dict the harness would inject via docker -e.

Recipe

Get the pinned CLI (version from harness/agent_image.py:CLAUDE_CODE_VERSION): npm install --no-save @anthropic-ai/claude-code@<pin> in a temp dir → binary at node_modules/@anthropic-ai/claude-code/bin/claude.exe (the .exe name is the real native-binary entry on Linux too, filled in by the package's postinstall — not a Windows leftover).
Stub API server: a tiny HTTP server that appends each request's headers to a JSONL file and returns a 400 invalid_request_error (non-retryable, so the CLI exits fast; exit=1 is expected).
Build the agent env exactly as the pipeline does: python3 -c "from harness.auth import resolve_auth_env; ..." and dump to an export-lines file with shlex.quote (values contain newlines — NEVER pass via env $(...), word-splitting mangles them; source the file).
Emulate the container env: unset ANTHROPIC_CUSTOM_HEADERS (and any other ambient var not in the resolved dict) before sourcing — a Claude Code session in this repo injects .claude/settings.json env into shells, which containers never see.
Run: ANTHROPIC_BASE_URL=http://127.0.0.1:<port> CLAUDECODE= IS_SANDBOX=1 timeout 30 <cli> -p hi --model claude-sonnet-4-5 --max-turns 1, then read the captured JSONL.

Gotchas

Unit tests in tests/test_patch.py / tests/test_patch_grade.py need docker and fail on docker-less hosts — pre-existing, not your change.
The docker -e injection leg itself can't be exercised without docker; it's the same mechanism that carries ANTHROPIC_API_KEY in production.
For the interactive-skills surface, copy .claude/settings.json into a fresh temp dir and run the host claude from there (with ambient ANTHROPIC_CUSTOM_HEADERS unset so settings.json is the only source).