This week felt like giving our tools a trust fall and then quickly grabbing the safety net. pwning.systems shows how replacing fuzzy agent memory with Datalog-style facts turns "I think this is vulnerable" into auditable reasoning, while Johann Rehberger demonstrates that Claude Code auto mode can still be prompt-injected 60-80% of the time. Add a very real root-escalation bug in Omarchy's defaults and Drydock's npm tarball diffing for supply-chain surprises, and the message is clear: convenience is great, but verify everything.
Performance and infrastructure had a strong week too. Zod 4.5 cut schema memory dramatically, pnpm 12 moved to Rust without breaking team muscle memory, and Postgres WAIT FOR turned stale read-after-write behavior from "almost always" to "effectively never" with low latency impact. Then Sai Srirampur's POSETTE talk reminds us Postgres is often blamed for storage problems it did not create.
On the workflow side, okaneland quantified the MCP context tax, and GitHub Agentic Workflows made the bigger shift explicit: agents are becoming first-class citizens in CI, not just fancy side quests.
Enjoy!
Signup here for the newsletter to get the weekly digest right into your inbox.
Find the 10 highlighted links of weeklyfoo #153:
I accidentally turned LLM memory into program analysis
by pwning.systems
Swapping fuzzy agent memory for a Datalog fact store that supports incremental reasoning in vulnerability research
📰 Good to know, ai,security
Omarchy any user process can escalate to root
by 0xcc
A default Docker configuration let any user process become root - fixed in version 4.0.1
📰 Good to know, security,linux
pnpm 12 rewritten in Rust, same commands and lockfile
by Zoltan Kochan
The Rust rewrite is stable and nearly all v11 commands, flags and lockfile format carry over
📰 Good to know, nodejs,tooling,rust
How Zod 4.5 cut memory per schema by 9x
by Colin McDonnell
Method memoization took a bare z.string() from 7.5KB of heap down to 784 bytes
📰 Good to know, typescript,performance
Breaking Claude Code Opus 5 auto mode
by Johann Rehberger
Prompt injection against auto mode landed 60-80% of the time, so it is a speed bump and not a security boundary
📰 Good to know, ai,security,promptinjection
Read your own writes off the primary with WAIT FOR
by Radim Marek
Naive replica reads were stale 992 times out of 1000 even over loopback - WAIT FOR gets that to zero for 1-2ms
📰 Good to know, postgres,database
The MCP tax - what each server costs before your first prompt
by okaneland
Measuring the context every MCP server eats, and how schema deferral changes the bill
📰 Good to know, ai,mcp,context
GitHub Agentic Workflows
by GitHub
Event-triggered and scheduled AI jobs alongside your deterministic CI/CD, with guardrails and support for Claude Code, Copilot, Gemini and Codex
🧰 Tools, github,ci,ai
Drydock
by Jovi De Croock
Diff built npm tarballs against the last published version, flagging install scripts, network access and new binaries
🧰 Tools, npm,security,supplychain
Postgres is not slow, your storage is
by Sai Srirampur
A POSETTE 2026 talk benchmarking local NVMe against network storage and where the latency compounds
📺 Videos, postgres,performance,video
Want to read more? Check out the full article here.
To sign up for the weekly newsletter, visit weeklyfoo.com.
Top comments (0)