Updates, announcements, trends and explainers from Stackness
Route a decision to a System 1 model only when it is bounded, labelled and cheap to get wrong. A 0.90 confidence gate kept 84.0% accuracy at $0.69 per 1,000.
SESergei GordeichukSenior Software engineer · The master of StacknessKeep one human on ordinary agent-written PRs and two on risky paths; drop reading only where reverts stay flat, as the Claude Code team does.
Stay on pgvector HNSW in production: pg_vectorsearch 0.1.0 is alpha and needs Postgres 18. With enough build memory, 42 h vs 49 min becomes 6.2 h vs 0.96 h.
Scion is Anthropic's in-house version control system, described on 13 October 2026. It replaced Git in one monorepo and is not available to anyone else.
Jev was faster in every outside test; Microsoft-Decision-1 has Jev's $0.042 token price but cost $0.0178 per 1,000 decisions to $0.0324 on one harness.
Linear's PRs ran 3.7x the tests by September 2026 and waited 12% less for CI. Faster runners saved 34%, tsgo 73% and syntax-only lint rules 68%.
Seven tools sit between a coding agent and the model as of October 2026. Every headline saving is self-measured, and RTK is the safest place to start.
Runaway subagent bills come from nesting no cap stops and children that inherit the parent's model. Count spawn depth in the logs, then cap it in three places.
Keep interactive coding agents local with the OS sandbox on, and send repo-scoped tasks that end in a pull request to a cloud sandbox. Cowork moved on 6 October.
OpenAI Decisions costs 1.6 times Jev per decision and answers slower, but takes images without an experimental flag and comes with OpenAI's data terms.
OSC 7501 lets a program tell the terminal it is working, blocked, done or failed. Claude Code, Pi and Crush send it; few terminals show it yet.
OpenCode 2.0 and Codex CLI 0.157 now run one background server per user that outlives your terminal. What it does, what it risks, how to turn it off.
Point ANTHROPIC_BASE_URL at DeepSeek, Kimi, GLM or Ollama, set all four model tiers, test with curl and /status. Check the cache before the bill.
Agents read 83.6% of the relevant code on turn 1 and 35.4% by turn 5. Check the transcript for a search before the write, then gate the diff with jscpd.
For a new cache in October 2026, Valkey 9.1 is the default: BSD-licensed and 20% cheaper on ElastiCache. Its forkless BGSAVE and INCREX are still rc1.
Haiku 5.5 and GPT-6 Luna list at the same price, but Haiku costs 2.5-5x more per task and 5x above 100k tokens. Which calls each one should take.
Tool output was 84% of a coding agent's context in a September 2026 study. How to tell context is the problem, and where to route the work instead.
Deno gets monthly fixes until about October 2027 and Deploy closes in six months. Node 26 is the pick for most Deno projects; here are all five options.
MXC 1.0.0 shipped on 7 October 2026 with egress denied by default, Docker Agent's microVM sandbox stays off until you ask, Agent37 leaves the internet open.
Claude Code led both Q3 2026 developer surveys, Codex grew fastest on GitHub and npm, and Copilot's vendor and survey numbers split. Data and method.
