Updates, announcements, trends and explainers from Stackness
The Type Server Protocol lets Pylance in VS Code ask Pyright or Pyrefly for Python types as data. Version 0.4.1 changes the editor, not CI.
SESergei GordeichukSenior Software engineer · The master of StacknessOn SlopCodeBench 10/2026, xhigh cost 2.1 to 4.2 times medium. Opus 5.5 passed 28 of 196 checkpoints at medium and 25 at xhigh.
Google's Elephant Goldfish model drafts a spec in one long agent session, then has fresh-context subagents review it. v0.1.0 shipped on 27 October 2026.
When an agent's fix goes green and the bug stays, check first whether the new test can fail: run it on the parent commit, where it must fail.
A multi-user agent needs each call checked as the person it acts for, outside the agent's reach. A shared API key and a monitor in the repo fail that.
OpenBao fixed the chained Vault and OpenBao RCE 14 days before Vault. Pick OpenBao for Raft or PostgreSQL without Enterprise features; stay on Vault for them.
A 33-line architecture map loads 603 tokens per session, area notes load on demand, and a 21-line CI script fails when either names a deleted path.
Agent pull requests grow the Actions bill through repeat runs of one commit. Find them in the runs API first; cheaper runners list a minute at 33-79% less.
A superfan agent is one Claude per component, reading its telemetry in a Slack channel. Anthropic named it on 13 October 2026; alerts still page.
One worktree per agent, then the tests on a trial merge of every pair: in one study, a third of merge conflicts were merges git had called clean.
Agents raised pull requests 23% across 718 firms, but time to merge rose 49% and resolved issues stayed flat. Postgres 19 reverted 8 to 10 features.
Agents found the PHP bugs fixed on 21 October 2026, but on real apps 82-86% of agent findings were false. Three security layers, with commands.
Anthropic gave Agent Skills to the Agentic AI Foundation on 22 October 2026. The six-field SKILL.md format is unchanged; hooks and MCP wiring still don't port.
LayerFS 1.0, out 21 October 2026, gives each coding agent a writable workspace over shared history. Linux only; one official agent integration.
In a test published 13 October 2026, GPT-6 Luna in a harness passed 10 of 10 like bare Sol, at $0.273 a run against $0.406, but took 5x as long.
In a 27-company sample from October 2026, AI spend ran from a $150 cap to over $3,000 per developer a month; 12 companies spent $1,000 or more.
Strands Decider 2B is AWS's free 1.9B decision model, out 1 October 2026: about 115 ms per question on an RTX 3090, consulted before each tool call.
Deleting 168,000 lines of comments left Claude Code's output unchanged in 12 runs; removing AGENTS.md cost points. Keep why-comments, October 2026.
The old $70 Claude month now costs $100 on Cursor Premium and $39 on Copilot Pro+. Eight alternatives priced for the same work, October 2026.
A self-improving harness rewrites its own prompts, skills and hooks between runs. Harness2 lifted GLM-5 11.5 points, with Opus 4.8 editing, October 2026.
