WHAT YOU ARE PAYING FOR
A commercial license, not the bytes. The source is public - anyone can read it, run the 19 self-tests and break it before paying. Your $49 buys the right to use it in paid client work, plus the licensing text and the 14-item pre-flight checklist we run before every handoff. Those are delivered on the page you land on right after checkout - no email, no waiting.
Your agent says "done". Sometimes it is lying.
Not maliciously - it reports success on work that never happened. A delete wrapped in a silent-error flag returns exit 0 with the files still there. A URL that 302s to a 502 gets reported as "endpoint is up". A "next steps" list gets sold as a delivery. You find out three days later, from the client.
agent-delivery-guard is one Python file, zero dependencies, that turns every claim into a command that actually runs.
WHAT IT CHECKS (5 gates)
1. Exit-code truth - swallowed failures: bare except/pass, empty catch, || true, discarded stderr, silent-error flags. Reports file:line.
2. Last-hop reachability - follows redirects up to 8 hops and judges ONLY the final status. First-hop 200 means nothing.
3. Evidence anchors - every path your report cites is stat'ed. A delivery document pointing at a file that does not exist is a failed delivery.
4. Plan-as-result - "next steps", TODOs, "waiting for your confirmation" are counted against completion evidence.
5. Claim vs reality - it RUNS the command behind each claim and compares the exit code. This is the one that catches the rest.
WHAT YOU GET
- guard.py (single file, Python 3.8+, stdlib only)
- Two runnable demo repos: one that must be blocked (exit 1), one that must pass (exit 0). Run both in 10 seconds.
- .guard.json template + .guardignore support
- Claude Code hook config (PreToolUse + Stop) - drop into .claude/settings.json
- AGENTS.md: the delivery contract you hand to your agent
- Commercial license, single team
VERIFY BEFORE YOU BUY
Full source, both demo repos and all 19 self-tests are public. Read the code, run it, then decide: https://github.com/956231studiocom/agent-delivery-guard
HONEST LIMITS
- It does not judge whether your code is correct. It judges whether your claims survive execution.
- Unreachable network is reported UNKNOWN, never FAIL. Firewalls are not bugs.
- It will produce false positives on unusual code. That is why inline guard:ignore and anchor_allow_regex exist - the tool caught six of them on itself.
WHO IT IS FOR
Agencies and freelancers shipping AI-written code to clients. If you have ever sent "fixed, please verify" and then spent a day on a follow-up, this is the 30 seconds before that message.