The Mash · three stories, one thread
AI builders face a shared interface question
Claude Code fixes, VELA’s findings, and Ollama’s changes put attention on how tools expose operations and handle inputs.
Explore
Compare the operations your agents can access with the interface restrictions VELA tested. Check whether late-connecting MCP servers work for Claude Code subagents in your setup, and whether Ollama preserves the image detail your workflow needs.[1][2][3]
Create
Build a small test suite for an agent workflow this week. Include a late-connecting MCP server, an over-200K-token Claude Code session, restricted web operations, and small images processed with Ollama.[1][2][3][4]
Theorize
Hypothesis: builders may need to test agent interfaces as carefully as the models behind them. VELA found that some restrictions prevent attacks, while the Claude Code and Ollama releases address different points where an interface can affect a workflow. These claims do not establish how any particular deployment will perform.[1][2][3]
In the mash
Anthropic fixes Claude Code stalls and session errors
Claude Code v2.1.311 adds runner controls and fixes several stalls and session errors for developers.
Vela finds web vulnerabilities persist through agent interfaces
AI agents can exploit existing web vulnerabilities when service interfaces expose the operations and inputs needed to do so.
Ollama adds DFlash2 support in v0.40.12 release notes
The listed changes add DFlash2 support and adjust MLX image handling for compact Gemma 4 vision towers.
