Index of every page on this site: llms.txt
All skills
Run it yourself

Verify harness changes end-to-end without docker — drive the real pinned CLI against a header-capturing stub server with the exact env resolve_auth_env() produces.

Use this Skill: https://skilld.dev/gh/anthropics/defending-code-reference-harness/verify

Nothing lands on disk. Nothing to clean up.

Fork this Skill

Edit a local copy. It keeps the author and licence.

SKILL.md

≈43 tokens for metadata: the name and description. ≈527 when used: this file.

Description uses 2.1% of example budget

Before choosing a Skill, your Agent reads its name and description. All available Skills share that space.

  • A shorter description leaves more room for other Skills. This entry exceeds our 1% size suggestion.

In our Claude Code example, all Skill names and descriptions share 8,000 characters. This Skill uses ≈172 characters, or 2.1%.

The 1% threshold is a size suggestion. Longer descriptions can still fit.

Your model, settings, and other Skills decide how much text your Agent can read.

Example settings and source

The example uses a 200k-token context and default Claude Code settings. The count includes the name, description, separators, and when_to_use when present. Codex also counts local file paths.

Skit's source and limits: Codex 0.160.1, Claude Code 2.1.292.

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

  1. 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).
  2. 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).
  3. 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).
  4. 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.
  5. 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).

Source: SKILL.md on GitHub

skilld matched fixed text patterns in SKILL.md and file names. Patterns miss obfuscated code.

skilld run checks every file with the same patterns. It asks for approval before it loads a Skill with a behavior marked Needs approval.

1 warning4mo3 checks · Risk SAFE
  • Gen Agent Trust Hub4mo

    This skill provides a procedure for verifying harness changes in a local environment. It utilizes vendor-specific tools and standard development practices, such as package installation and environment variable configuration, to emulate a containerized environment for testing purposes.

  • Socket4mo

    1 alert: gptAnomaly

  • Snyk4mo

    Risk: LOW · No issues

Signed by skilld at d3bea6b. This ties the file your Agent reads to that commit on GitHub. It does not review the instructions.

Last checked against GitHub 4 weeks ago.

Steadyupdated 4 months ago

README badge

README badge for anthropics/defending-code-reference-harness/verify