v0.3.1

Full Changelog: v0.3.0...v0.3.1

v0.3.0

🛡️ AgentSeed v0.3.0 — the anti-hallucination gate just grew teeth

Your AI coding agent invented an API, claimed "all tests pass" without running anything, and suggested a Python package that doesn't exist. AgentSeed v0.3.0 catches all three — in 16 languages, before it ships.

What's new in 0.3.0

🧬 Multi-language hallucination detection (16 languages)

verify_code now runs a config-driven generic lexical engine across Go, Rust, Java, C, C++, C#, PHP, Ruby, Kotlin, Swift, Dart, Lua, R, and Zig — on top of Python (full AST) and TypeScript/JavaScript. Same rule everywhere: a bare call to a symbol that is never defined is a hallucination, whatever the syntax. Live-tested: every language flags its invented call; clean code reports zero false positives.

Dart, Lua, R, and Zig landed on main immediately after this tag — they ship with the next release (v0.3.1).

📜 check_contract — code against a written spec

Declare a contract ({"requires": ["run"], "prohibits": ["stub"]}) and verify the source satisfies it. New MCP tool + guard_cli contract subcommand.

📦 check_imports — slopsquatting guard (USENIX Security 2025)

LLMs hallucinate non-existent package names in 5.2–21.7% of generated code, and ~58% of those names recur — attackers pre-register them to ship malware ("slopsquatting"). check_imports flags any top-level import that is neither Python stdlib nor in your known-package set, so a phantom package can't reach your lockfile silently.

⚡ sandbox_run — bounded-memory streaming

Output truncation is now streamed (tail ring buffers, 8 KB / 4 KB): a child that floods output can no longer balloon the server's memory.

🧩 PROMPT-POOL → your agent configs

scripts/export_prompt_pool.py renders the 25-entry guardrail prompt pool as CLAUDE.md, AGENTS.md, and Cursor rules — so the gates apply even outside plugin-aware clients.

🎓 Research-backed hallucination arsenal

Token pools expanded (+15 EN / +10 CJK low-false-positive signals); the pattern library and prompt pool now incorporate the 2025 code-hallucination taxonomy (arXiv:2504.20799), scaffolding hallucination / phantom symbols (arXiv:2604.20202), and the "assisted developers are more confident and less secure" finding (Perry et al., CCS 2023).

The numbers (measured, not promised)

Metric	Value
Detection benchmark (5 defect classes, seeded corpus)	precision 1.0 · recall 1.0 (tp=100, fp=0, fn=0)
Test suite	176 tests, all green (unittest, zero required deps)
Languages verified	16 (14 via config-driven registry + Python AST + TS/JS lexical)
Required dependencies	0 — pure Python standard library

What it is

AgentSeed is a drop-in Agent Plugins 1.0.0 plugin — Skill + 8 MCP tools + optional client enforcement hook + CI gate — that makes "done" mean observed fact, not self-report.

Skill teaches → MCP verifies → Hook enforces → CI exits nonzero

Quick start

git clone https://github.com/Morningstar202604/AgentSeed.git   # mirrors: GitCode · Gitee
bash install.sh --client auto --hooks        # or: ./install.ps1 -Client auto -Hooks
python3 server/guard_cli.py gate --root .    # CI-equivalent hard gate
npm install -g agentseed-mcp                 # or install the MCP server from npm

Honest scope

Attribute calls (obj.m()), macros, and cross-file symbols are not type-checked — that's a type checker's job. check_imports is a report, not a hard gate: verify flagged packages in the registry. Bash/PowerShell are not in the registry by design — their command namespaces are open, so "undefined symbol" cannot be decided statically. Details in docs/BENCHMARK.md.

⭐ If AgentSeed saves you from shipping hallucinated code, star the repo.