@ImportStatic

Working notes on AI engineering, Rust and Go, written for people who ship software and would rather read one tested example than three paragraphs of adjectives.

importstatic.com
Joined October 2026
  • Tweets40
  • Following3
  • Followers4
OpenAI's 12-hour prompt cache write costs 2.5x input, so two requests 6 hours apart cost more than with the 30m default. It pays from the third reuse. Set request-wide, a chat cost 68% more.
1
2
Full write-up: importstatic.com/ai/openai-1…

When OpenAI's 12-hour prompt cache write pays for itself

OpenAI prompt caching now has a 12-hour lifetime at 2.5x the input price. It pays from the third reuse, and set request-wide it raised a chat's input cost 68%.

importstatic.com
1
date('now','end of month') is new in SQLite 3.54. On 3.53 it returns NULL, no error, so a WHERE built on it matches zero rows. And 'weekday -0' on old builds moves forward a week.
1
3
Full write-up with the code: importstatic.com/databases/s…

Last day of month in SQLite: 'end of month' needs 3.54

SQLite 3.54 added 'end of month' and 'weekday -N'. Older builds, including the Go and Rust drivers we tried, return NULL with no error. The portable forms.

importstatic.com
2
Claude Code reads AGENTS.md only if no CLAUDE.md or CLAUDE.local.md exists in your working dir or any parent, ~/CLAUDE.md included. ~/.claude/CLAUDE.md is exempt from that check.
1
2
Full write-up with the code: importstatic.com/ai/agents-m…

AGENTS.md vs CLAUDE.md in Claude Code, Codex and Copilot

AGENTS.md vs CLAUDE.md: Claude Code skips AGENTS.md if any CLAUDE.md sits above your repo. Which files Codex, Copilot, Cursor and Gemini CLI read, in one table.

importstatic.com
1
DeepSeek V4.1 Flash takes seven reasoning_effort values. In DeepSeek's own prompt encoder, medium, high and xhigh all become the line "Reasoning Effort: 75". Low is 50, max is 100.
1
4
Full write-up: importstatic.com/ai/deepseek…

DeepSeek V4.1 Flash reasoning effort, rendered in Rust

DeepSeek V4.1 Flash accepts seven reasoning_effort values and the model sees three. We rendered requests with DeepSeek's Rust encoder to show what it gets.

importstatic.com
1
4
The $20M limit in the Muse Spark Open licence counts the gross revenue of your company and its affiliates. It only applies if you sell access to the model as an API.
1
3
Full write-up with the code: importstatic.com/ai/muse-spa…

Muse Spark Open licence: what you can ship commercially

The Muse Spark Open licence's $20M limit counts your company's total revenue, but only if you sell model access. What else it requires, and calling it from Go.

importstatic.com
3
openai-go v3.78.0 marks o4-mini deprecated, but not gpt-4o-mini, which OpenAI shuts down January 14. On our sample service, staticcheck flagged 1 of 9 doomed model IDs. How are you finding yours?
2
5
Full write-up: importstatic.com/ai/openai-m…

OpenAI Model Shutdown Dates Through January 2027

Eight OpenAI models shut down on October 23 and GPT-4o mini goes January 14. The dates, the replacements, and a Go scanner for the IDs staticcheck misses.

importstatic.com
3
Go 1.27 drains an unread HTTP/1 response body on Close, and a chunked one too. But call the same host again straight away and the drain may not be done yet, so you dial a new connection anyway.
1
5
Full write-up with the code: importstatic.com/go/go-drain…

Drain the response body in Go 1.27? Limits, chunked, tests

Go 1.27 drains an unread HTTP/1 response body on Close. We tested chunked streams, slow bodies, json.Decoder and forced closes to see which ones get reused.

importstatic.com
3
go-genai 1.73 on a 4-hop Gemini 4 Argon answer: 1 request, a quarter of the text, nil error, finish reason CONTINUATION. 1.74 sends all 4, but a 503 on hop 3 returns nil and drops hops 1 and 2.
1
6
Full write-up with the code: importstatic.com/ai/gemini-4…

Gemini 4 Argon API from Go with go-genai 1.74, tested

Long Gemini 4 Argon API answers arrive in hops. go-genai 1.73 returns the first hop with a nil error; 1.74 continues, and drops everything if one hop fails.

importstatic.com
4
PostgreSQL 18, one connection, single-row inserts: synchronous_commit = off did 47,000 commits/s and fsync = off did 36,000. With fsync off each backend still makes one WAL write per commit; async commit leaves that to the WAL writer too.
8
exhaustive v0.15.0 (Oct 9) checks Go type switches too. On a sealed interface with value receivers it asked for both Closed and *Closed as cases. With pointer receivers it asked only for *Closed.
1
7
Full write-up with the code: importstatic.com/go/golang-e…

Golang enum with iota: what compiles, what a linter catches

A Go enum is an integer type, so 7, Status(42) and a JSON 9 all get through. We ran iota, a struct wrapper and a sealed interface against the same mistakes.

importstatic.com
3
that's true
Mark Holloway@0xKetoneOct 11
Lesson from ImportStatic (10 Oct): Anthropic and OpenAI hard spend caps often return HTTP 429, and their Go SDKs retry by default. Your agent loop's "retry on 429" rule will spin until month-end if it cannot tell a spend cap from a rate limit. Classify the body before you sleep.
5
Load more