I find and fix defects in the tooling that agents run on — doc/contract drift, dead configuration, links that go nowhere — and I prove each one is real before opening a PR.
86 merged pull requests across 39 upstream repositories (2026-07 → 2026-11, 56 of
them in November), including pipecat-ai/pipecat (14), microsoft/agent-framework (8),
labring/FastGPT (6), vanhauser-thc/thc-hydra (6), vllm-project/llm-compressor (4),
pytorch/torchtitan (3), huggingface/peft (3).
| Every number here is generated from a ledger and reconciled against the GitHub API | gh search prs --merged |
| The defect classes are toolable, and the tools ship their false-positive cases as tests | 32 checkers with self-tests, 185 scripts total |
| I close my own PRs when the evidence goes against them | 62 stale PRs closed and archived in one pass, each with the patch kept |
| artifact | what it demonstrates | verify |
|---|---|---|
doc-drift v0.1.0 |
fourteen stdlib-only checkers for documentation that stopped being true. Every rule was validated against a real repository, and the ten false-positive classes are written down with the repository that produced each. | git clone https://github.com/lxcxjxhx/doc-drift && cd doc-drift && python tools/run_all.py --selftest → selftest ok ran=14 |
doc-drift CI |
three operating systems × two Python versions, plus a step that fails the build if a checker ever imports a third-party package | green badge on the repository |
| every merged PR | 86 merges in 39 repositories, maintainer by maintainer, straight from GitHub search | open the link, sort by repository |
Of the 136 findings these checkers produced on six repositories in three ecosystems, two were
real and both are filed upstream (llm-compressor#3526, agent-framework#10400). The other 134
changed the checker. That ratio is the point: a rule that only ever finds things is guessing.
- Contract drift — a docstring promising a parameter the signature dropped; a README link to a file that no longer exists; a constant nothing reads while the same literal is repeated in the code that should use it.
- Agent runtime harnesses — Pipecat (voice/streaming), Harbor (agent evaluation), Microsoft Agent Framework, llm-compressor.
- Local-first AI security — static scanning and eval tooling that runs without a provider key.
pipecat-ai/pipecat#6690—SUMMARY_TOKEN_BUFFERwas documented as the summary buffer but two0.8literals were used instead; both now read the constant. Merged 13 minutes after it was opened.microsoft/agent-framework#10236— CI: dropped a path filter that matched no directory.harbor-framework/harbor#4213/#4214— a trial lock that records what produced a trial's artifacts, so a regrade cannot reuse a run whose agent never ran in the target environment.
中文:做智能体运行时与工具链的安全工程。上游合并 86 条、覆盖 39 个仓库;每个结论都附可复核的方法,被证伪时主动撤回。



