Notes from the engine room.
Notes on design, engineering and the weird stuff in between.
How we stop bad line breaks on a Czech and English website
CSS text-wrap handles about half of it. The rest is no-break spaces applied in one place, plus a few rules for headlines. This is the code we run on steezr.com.
Read more- 037Engineering7 min read
StyleSmuggler and the template engine as an attack surface
CVE-2026-75650 hit Magento through its template engine. What is known about the chain, and how to keep user input out of template source in any stack.
- 036Engineering10 min read
Hardening npm installs and publishes against supply chain attacks
Lockfiles, install-script allowlists, release-age cooldowns and trusted publishing, tested on current npm and pnpm, with what each one does and doesn't stop.
- 035Engineering9 min read
Hardening GitHub Actions against supply chain attacks
What the Trivy and TanStack compromises exploited in GitHub Actions, and the workflow settings and platform controls that close each gap.
- 034Engineering8 min read
Your Postgres replication account is the blast radius
CVE-2026-6471 turned the REPLICATION attribute into code execution. How to find every replication credential, patch safely and lock down pg_hba.conf.
- 033Engineering7 min read
Your image optimizer decodes attacker-supplied files: lessons from the Next.js AVIF RCE
Who was really exposed to the Next.js AVIF bug, what the patch changed, and how to run image decoding where a memory bug can't reach your app's secrets.
- 032Engineering7 min read
When the SSO response gets to choose how it's verified
The 2026 miniOrange SAML bugs let attackers pick the signature algorithm. What went wrong, and how to configure SAML and OIDC verifiers that fail closed.
- 031Engineering10 min read
Token spend is an architecture problem: the math and the levers
Why agent bills grow faster than traffic, with worked cost math for caching, trimming, routing and metering that you can rerun on your own numbers.
- 030Engineering10 min read
Redesigning code review for teams where agents open the pull requests
Tier review by what a change touches, enforce it with rulesets and CODEOWNERS, record provenance in commit trailers, and let revert rates move the line.
- 029Engineering8 min read
Cyber Resilience Act reporting for teams that build software
Since September 2026, CRA manufacturers must report exploited vulnerabilities within 24 hours. Who is in scope, what triggers it, what to have ready.
- 028Engineering7 min read
Your LLM gateway is a credential vault, so run it like one
The LiteLLM compromise showed what an LLM proxy holds. How to isolate the gateway, pin it properly and stop a payload from phoning home.
- 027AI8 min read
A threat model for agents that read untrusted text
Two Semantic Kernel CVEs show how prompt injection reaches the host. How to model the threat: untrusted inputs, tool blast radius, controls that hold.
- 026AI8 min read
Contain coding agents with infrastructure, not prompts
Stripe's Minions and an AISI incident report show the same thing: an unattended agent is as safe as its environment. A practical containment checklist.
- 025Engineering12 min read
Auditing the permissions your AI assistant wrote
Generated code tends to over-grant: wildcard IAM, unscoped DRF querysets, half-done RLS. Tested greps, Semgrep and Rego rules, and a CI job that blocks it.
- 024Engineering9 min read
Server Actions are public endpoints: securing them in current Next.js
Ownership checks inside the action, input parsed as unknown, rate limits keyed on something you trust, and audit rows written in the same transaction.
- 023Engineering11 min read
Why AI-generated code passes CI and still breaks in production
Generated code and generated tests share the same assumptions. Four checks that don't, with tested examples in Python, Go and Envoy.
- 022Engineering7 min read
Auditing OAuth grants in Google Workspace after the Vercel breach
How one employee's grant to an AI tool led into Vercel, and how to list, revoke and restrict third-party app access in your own Google Workspace.
- 021AI7 min read
Opening a repository is now code execution
AI coding tools and editors run config files from the repo you open. How the June 2026 Microsoft repo poisoning worked and how to open untrusted code safely.
- 020AI11 min read
What the METR studies say about AI coding speed, and how to measure your own team
METR found a 19% slowdown in 2025, then called its own data unreliable in 2026. What holds up, and a small experiment to measure your team's real effect.
- 019AI10 min read
pgvector 0.8 or a dedicated vector database: when Postgres is enough
What pgvector 0.8 changed for filtered search, a benchmark script you can rerun, the 2,000-dimension limit, memory sizing, and when a dedicated store wins.
- 018AI9 min read
Tool calls are proposals: an approval queue for LLM actions in Postgres
How to put a policy check, a Postgres approval queue and a server-rendered confirmation between an LLM's tool call and your production data.
- 017AI9 min read
What structured outputs don't check in your LLM data
What OpenAI and Anthropic strict schemas actually guarantee, and the validation layer, dead-letter table and drift checks you still need behind them.
- 016AI7 min read
A model lifecycle policy for AI features: pins, evals, an owner and a calendar
Hosted models retire on a schedule. What OpenAI, Anthropic and Google promise, how to detect retirements automatically, and how to switch safely.
- 015Engineering10 min read
A Postgres inbox for webhooks: dedup, retries and replay without a broker
Verify signatures, dedupe on insert, claim events with SKIP LOCKED inside the processing transaction and keep dead letters queryable.
- 014Engineering12 min read
Postgres row-level security for multi-tenant Django, done properly
ORM tenant filters leak. An RLS setup for Django 6 and Postgres that holds: roles, policies, SET LOCAL inside the transaction and PgBouncer.
- 013Engineering12 min read
What belongs in the Next.js layer and what belongs in a Go backend
Next.js 16 runs proxy.ts on Node, so the edge debate is over. The boundary still matters: auth, spoofed headers, idempotent commands and flags.
- 012Engineering9 min read
Monorepo boundaries CI can enforce: go.work, pnpm catalogs and Turborepo 2
A mixed Next.js and Go monorepo stays sane when CI checks the boundaries: GOWORK=off per module, pnpm catalogs, turbo --affected and turbo boundaries.
- 011AI10 min read
How to roll out a new embedding model in pgvector without breaking retrieval
A new encoder means new dimensions, new scores and new neighbours. A schema, a shadow-read handler in Go and a CI gate that make the swap reversible.
- 010Engineering11 min read
Tail sampling and trace retention for a Next.js and Go stack
Keep the traces that explain incidents and drop the rest: context propagation, a tested OTel Collector config, and a retention budget you can compute.
- 009Engineering9 min read
Zero-downtime Postgres migrations in Django 6
Migration outages come from the lock queue. Expand, backfill, switch and contract in Django with db_default, concurrent indexes and safe NOT NULL.
- 008Engineering9 min read
Reporting SQL in Django: files, migrations and unmanaged models
Reports as SQL files with CTEs, views shipped through RunSQL, unmanaged models, refreshes that don't block reads, and tests on real Postgres.
- 007AI11 min read
Ship prompts like code: versioned files, eval gates and alias rollback
Keep prompts in git with their model and schema, gate changes on sampled evals in CI, and roll back by moving an alias instead of redeploying.
- 006Engineering9 min read
When self-hosting LLM inference beats the API, with the math
Realtime API, provider batch API, your own GPU or a CPU box: a formula for cost per request you can redo with your own numbers.
- 005AI11 min read
Document pipelines for RAG: OCR fallback, canonical text and chunk IDs you can diff
Treat PDFs as hostile input: OCR only pages that need it, canonicalize before chunking, respect the embedder's token limit, use content-based chunk IDs.
- 004Engineering10 min read
htmx inside Next.js 16: a version that runs, and what it saves
A tested htmx setup for filtered tables in the App Router, the bugs that break naive versions, and bundle sizes measured against RSC and TanStack Query.
- 003Engineering10 min read
Pool connections before adding workers: a Postgres checklist for Django
When Django runs out of Postgres connections, more pods make it worse. Pick a pooler, size it from the database side, and read the right metrics.
- 002Engineering10 min read
Replaceable Django apps: package boundaries, an outbox and an upgrade test
How to keep a Django monolith modular: apps with a small public API, import contracts, a transactional outbox, and a CI job that upgrades real data.
- 001AI11 min read
Why RAG fails in production and how to fix it
RAG usually fails at retrieval and provenance, not in the model. Versioned chunks, hybrid search with rank fusion, claim-level citations and a verifier.