Public work, open to inspection
Engineering notes.
Small failures at the boundary between systems: what broke, why the correction fits, and what the evidence actually shows.
Tencent AI-Infra-Guard · PR #680
Why AI remediation advice isn't a SARIF fix
A prose recommendation broke SARIF validation. Here is the representation change, regression evidence and migration boundary in a merged AI-Infra-Guard contribution.
Read the problem and evidence ↗IBM MCP Context Forge · PR #6049
Preserving an MCP gateway catalog after empty OAuth discovery
An empty OAuth discovery result should not erase a saved catalog. This merged contribution preserves entries and checks retained aliases within their actual visibility scope.
Read the problem and evidence ↗OpenAI Codex Security · PR #1099
Keeping Git metadata out of security scan inventories
Nested repositories, linked worktrees and scopes inside .git need explicit inventory boundaries. A merged contribution and the regressions behind it.
Read the problem and evidence ↗Try an independent project
The final answer can stay the same while an agent's execution changes.
HarnessLab records and compares lifecycle, tool and approval traces. Start with the offline quickstart or the browser demo using synthetic fixtures. The demo is not a production benchmark.
Have a similar handoff to check?
I work independently on AI workflow integrations. Start with one process and the systems it crosses.
See the review scope ↗