Find inputs that break a program’s rules
Using Codex CLI where available: Supply authorized code and approved rules to generate test inputs and record reproducible failures.
Execution environment and usage
AI / software
Codex CLI (suggested exercise route; original stack differs)
Where to operate
Codex CLI in a terminal with an isolated authorized project copy. This suggested route is not the same installable system as the original research setup.
Getting started
- Follow the linked official Codex setup guide, sign in and start Codex in the test project folder.
- Open an authorized code copy and configured test environment.
- Supply approved invariants, input domain and resource limits, then use the case prompt.
- Retain real runner output, seeds and minimized failures for maintainer review.
Setup and plan requirements
Compatible property-test framework, runnable code and explicit approved rules. Requires Codex access and compatible project tools/versions. Authorize only necessary files and commands, without production-service connections.
Prompt portability limits
Prompt explicitly requests a run; plain chat can draft properties but cannot truthfully provide execution evidence.
Usage limitations
This review performed no installation, hands-on test or regional-availability check. Bluesky page failed retrieval; environment inference is from supplied record and prompt, not renewed primary verification.
Environment sources
適用情境 / When to use
Use the failures to narrow a debugging task. Running them requires a configured test framework and resource limit. Preserve seeds and minimized examples, then have a maintainer check whether the code or the rule is wrong. hikikomorphism reports finding software bugs using agent swarms for property-based testing.
步驟 / Steps
- Prepare an authorized code copy, a working test environment and explainable invariants.
- Ask for properties and input generators for a bounded function; have its maintainer approve the rules.
- Record seeds and shrink failures, then decide whether the implementation or property is wrong.
提示詞 / Prompt
Design property tests for [function in test copy] against [approved specification]. Explain each invariant and its domain. Run within [resource limit] and report seeds, minimal counterexamples and reproduction steps. Do not edit implementation or specification, or automatically call a failure a vulnerability.
留意事項 / Cautions
- Limit runtime and resources; do not test unauthorized services.
- Never silently weaken properties or edit implementations merely to pass.
來源限制 / Evidence limits
The short account supplies no full evaluation; codex is a tested project, not evidence of the agent tool used.
流程說明 / Workflow note
These steps and prompts are original, untested exercises by this site, not the author’s verbatim workflow. Reported results do not guarantee exercise outcomes.
原始工具組合 / Original tool stack
Unspecified agent swarms; property-based testing
Source publication date: 2026-10-14
Source checked: 2026-10-19