sharness-hello

You are about to run two experiments from https://sharness.ai/code/experiments/index.html in this
session. They are small, read-only or confined to child processes, and they are the first two
rungs of the SHarness ladder: nothing here is a SHarness component.

Safety rules (from https://sharness.ai/design/process-tree/index.html#4-safety-rules-for-testing-inside-a-live-agent-session)

1. Never touch the orchestrator (the Claude Code process) or the session bridge.
2. Never add firewall rules, never change sysctls, never create users for this.
3. Everything runs as a child subtree that ends when the command ends. sharness-ai run
   and sharness-ai hello fork a child, apply the policy to the child, and exit. They never
   restrict the shell you run them from.
4. Leave nothing behind. hello writes only a temporary policy file and deletes it.

Steps

1. Install from PyPI, or from the repository checkout if that is where you are:
   pip install sharness-ai (or pip install -e .). Check: sharness-ai --version.
   If the installed release does not have the probe and hello commands yet, install from
   the repository at a commit you name, never a moving branch:
   pip install git+https://github.com/SHarness-AI/SHarness-AI__CLI@<commit> (the current dev
   head is on https://github.com/SHarness-AI/SHarness-AI__CLI/commits/dev).
2. sharness-ai probe and sharness-ai probe --json. This is E01: the grant. It prints the
   names of proxy variables only, never values. Do not add values yourself.
3. sharness-ai hello and sharness-ai hello --json. This is E04. Exit 0 means every
   expectation held. If the kernel has no Landlock, it says so and nothing is enforced.
4. Report, in this order and with these words:
   • Environment: OS and kernel, Landlock ABI, uid, Python version (from probe).
   • Measured, as of <today's date>: the hello result table and the policy hash.
   • The caveat, verbatim from the output. A policy applied by this session to its own
     child is a Boundary for the child and only a Setting for the session.
   • The not covered line, verbatim. Unix-domain sockets by path, the inherited
     environment and UDP without the namespace are not enforced by anything here.
   • Any probe marked ? is inconclusive: the environment refused it too, so the policy was
     not tested there. Say so; do not count it as held.
   • What was not measured: seccomp, uids, cgroups, mutual TLS, the process tree.
5. If asked to publish: the JSON belongs on the Evidence page of the site with method and
   date (https://sharness.ai/evidence/index.html), not in a chat message alone.

Never

• Never present hello passing as "the agent is sandboxed". It shows one child process
  under one policy. Say so.
• Never write a policy that includes a secret, and never print the value of an environment
  variable that may hold one.