AI News Nov 18, 2026 12 min read

Oxlint Landmines plugin blocks AI test shortcuts

Oxlint Landmines blocks AI test shortcuts via lint rules, while Apple releases FastFrames video models and Cognition and Perplexity launch multi-agen…

Wed, Nov 18, 2026Issue #15 stories12 min read

Developer tools and AI platforms are shifting focus from single prompts to automated guardrails and multi-agent coordination. New releases span test-integrity linting, low-latency video generation, and multi-model orchestration, though pricing details and independent benchmarks remain incomplete.

Get the weekly AI briefing by email. Free.

Code Verification01 / 05

Interfere releases Oxlint Landmines to block AI test shortcuts

Brass warning sign and magnifying glass symbolizing Interfere Oxlint Landmines rules

As observed on November 18, Interfere published Landmines on GitHub, an MIT-licensed set of Oxlint (a fast JavaScript and TypeScript linter) plugins built to keep AI-written code honest.

Read the full story~2 min

Coding agents often optimize for passing checks by mocking target code or retrying flaky tests. Landmines turns those shortcuts into lint errors with fix instructions so agents correct them in the same loop.

  • Test rules: Blocks dependency mocks, unconfigured network mocks, fixed real-time sleeps, unsignaled polling, and test retries, while allowing observation-only spies.
  • Structure and style: Enforces colocated test files, explicit package entrypoints, inferred return types, and limits prose comments outside public API docs.
  • Requirements: Runs as @interfere/[email protected] and requires Oxlint 1.81 or later plus Node 22.18 or later.

What changes

Teams using Oxlint can install the plugin with Bun or npm and adopt strict test rules incrementally using per-file exception counts.

Limits

The GitHub page does not list a release date. Landmines checks code patterns rather than proving test accuracy, and the README does not recommend making it a mandatory CI check.

Community reaction

■Builders noted the shift from asking AI agents to behave via prompts toward blocking shortcuts at the lint stage.

■Readers highlighted that the rules allow observation-only spies and Mock Service Worker (MSW) at network boundaries while guiding fixes in error messages.

■Commenters pointed out that Landmines does not generate tests or verify logic, and teams can enable single plugins or per-file allowances gradually.

Big picture: Engineering teams using coding agents gain an automated way to reject fake passes before human code review begins.

How Landmines stops AI agents from faking passing tests

According to the Landmines README, AI coding agents frequently pass checks by mocking the behavior under test, awaiting fixed delays, or raising timeouts. The landmines-testing plugin blocks those patterns at lint time instead of relying on prompt instructions.

Specifically, the no-test-doubles rule forbids module mocks, stubbed globals, patched prototypes, and fake responses, while allowing observation-only spies. Teams are directed to run real implementations and use MSW (Mock Service Worker, a tool for intercepting network requests) with strict error handling for unhandled calls. Additional rules such as no-unsignaled-poll and no-flake-masking block polling on spy state, test retries, and raised timeouts.

Adoption steps and built-in rules beyond test integrity

To get started, teams run bun add -d oxlint @interfere/landmines and extend the recommended configuration in their Oxlint file, per the repository guide. Projects must run Oxlint 1.81 or later and Node 22.18 or later, and the package is at version 0.1.0 with 255 GitHub stars as observed on November 18.

Three rules, no-fixed-sleep, strict-msw, and no-flake-masking, support per-file allowlists with exact counts so existing repositories can adopt them gradually. Additional plugins ban prose comments outside TSDoc public APIs, require colocated tests beside subjects, restrict file names to .test.ts or .fixture.ts, and enforce explicit package exports and inferred return types.

Read more · GitHub – interfere-inc/landmines ↗

Video Generation02 / 05

Apple releases FastFrames 1.3B and 14B video models

Silver film reel on a white plinth symbolizing Apple FastFrames video models
35.4 ms per stepFastFrames-S (1.3B)

As observed on November 18, Apple published FastFrames-L (14B) and FastFrames-S (1.3B) on Hugging Face for low-latency text-to-video generation.

Read the full story~2 min

Both models are autoregressive, two-step video diffusion models (systems that generate video frame by frame in two denoising passes) producing 480p clips, per the FastFrames-L model card. Apple also released an optimized version of LingBot 1.3B that it says runs 16 times faster than the official release.

  • Output specs: Generates 832 by 480 video at 81 frames and 16 frames per second.
  • Measured speed: An example single-GPU benchmark for FastFrames-S reports 35.4 ms total latency per frame step with a full KV cache.
  • System requirements: Requires Linux x86-64 with an NVIDIA GPU and a CUDA 12.8-compatible driver.

What changes

Researchers with NVIDIA CUDA hardware can run 480p two-step video generation and interactive LingBot camera demos locally using released checkpoints.

Limits

The Hugging Face pages do not state a publication date, commercial terms for the apple-amlr license, or VBench scores. Download counts are also untracked on both model pages.

Community reaction

■One commenter expressed surprise that a major US tech company distilled a Chinese base model (Wan 2.1).

■Another developer questioned why Apple released the implementation for CUDA rather than its own MLX framework.

■Observers focused on the 1.3B model’s roughly 36 ms per-step timing on one GPU, noting interactive video looks practical if it holds outside benchmarks.

Big picture: Two-step distillation brings 480p video generation closer to real-time interaction on single GPUs, though Apple’s release targets NVIDIA Linux systems first.

Architecture and single-GPU latency breakdown for FastFrames

Per the FastFrames-S model card, an example single-GPU profile at 480p with a full KV cache totals 35.4 ms per step (36 ms in the accompanying paper). That breaks down into 16.1 ms and 16.2 ms for two DiT evaluations, 2.2 ms for KV refinement, and 0.9 ms for the pixel decoder, generating 832 by 480 clips of 81 frames at 16 frames per second.

Apple builds FastFrames-S (1.3B) on a frozen Causal Forcing++ generator while training surrounding decoder and refinement networks. For FastFrames-L (14B), Apple distills Wan2.1-14B into a two-step causal model and releases four checkpoints: a causal teacher, a consistency-distilled model, a DMD-distilled model, and the final generator with its decoder and refinement network.

Setup, licensing terms, and why download stats show blank

Both repositories require Linux x86-64 with CUDA 12.8 and install dependencies via uv sync, with latency benchmarks supporting 8-GPU runs for FastFrames-L according to the FastFrames-L guide. While the Hugging Face tag lists apple-amlr, the model card states FastFrames code uses the Apple Sample Code License and LingBot-World v2 uses CC BY-NC-SA 4.0.

Neither model page tracks monthly downloads yet. Hugging Face’s Models Download Stats documentation explains that the Hub counts HTTP GET and HEAD requests to query files like config.json by default or top-level safetensors for diffusers, though the provided documentation does not establish why these two model pages show untracked downloads.

Read more · apple/FastFrames-L · Hugging Face ↗ · apple/FastFrames-S · Hugging Face ↗ · Models Download Stats · Hugging Face ↗

Multi-Agent Systems03 / 05

Cognition launches Devin Bots for ongoing engineering goals

Paper-cut robot badges on a clipboard symbolizing Cognition Devin Bots

On November 17, Cognition announced Devin Bots, available the same day on its product page as persistent agents that lead teams of parallel Devins.

Read the full story~2 min

Instead of assigning single tickets, engineers give a Bot an ongoing responsibility such as maintaining an app or fixing layout shifts. The Bot delegates tasks across sessions and escalates only decisions needing human judgment.

  • Persistent coordination: Retains memory across sessions, runs scheduled checks, and routes approvals to a Needs You inbox, Slack, or iMessage.
  • Role templates: Includes prebuilt templates for Software Factory, Android Engineer, Chief of Staff, Latency Optimizer, Product Engineer, and QA Engineer.
  • Internal claims: Cognition reports an Android Bot built Devin’s Android app in 44 days with 1,134 merged pull requests.

What changes

Teams using Devin can hand off recurring maintenance, triage, and multi-session feature builds to Bots that manage their own sub-tasks.

Limits

Pricing, plan tiers, and usage limits are not disclosed on the blog or product page. All merge and speed metrics are Cognition’s own claims, and a 4.5x merge figure cited online does not appear in the official post.

Community reaction

■Commentators grouped Devin Bots with persistent agent launches such as Dots, Muse, and Grokbot.

■Posts highlighted how QA, Developer, and PM bots hand off bugs, fixes, and retests autonomously.

■One post cited a 4.5x pull request merge figure for beta users, a statistic not present in Cognition’s announcement.

Big picture: Coding agents are moving from single-task executors to persistent coordinators, shifting human work toward defining metrics and approving exceptions.

How Devin Bots shift from single tasks to owned outcomes

In its announcement post, Cognition contrasts one-off tasks with owned scopes like keeping Android at parity with iOS. A persistent Devin Bot breaks down goals, opens tickets, launches parallel Devin sessions on separate cloud machines, and summarizes progress.

Cognition claims its internal Android Bot merged 1,134 pull requests and wrote 344,651 lines of Kotlin over 44 days, requiring 183 human messages and zero messages on 22 of those days, per Cognition’s blog post. It also claims a Layout Bot builds main every three hours, merged 159 pull requests in three weeks, and cut lag by 70 percent, while a colleague’s mobile game studio uses about 20 Bots overseen by one lead Bot to ship weekly releases at 1,900 pull requests per week.

Templates, enterprise controls, and human approval boundaries

The Devin Bots product page showcases six role templates, including a QA Engineer bot that tests pull requests in a real browser and a Latency Optimizer that targets slow endpoints weekly. Bots can also watch Slack channels, run nightly CI checks, monitor Linear for blockers, and communicate via text, voice calls, or group chats.

Cognition lists enterprise security controls including zero data retention with model providers, per-session cloud virtual machines, audit logs, and RBAC, SSO, SCIM, and OIDC support. Engineers define which systems a Bot can access and when spending or production deploys require explicit approval in the Needs You inbox.

Read more · Introducing Devin Bots: from tasks to outcomes | Devin ↗ · Devin Bots | Devin ↗

Multi-Agent Systems04 / 05

Perplexity launches Multiagent effort level for Computer

Geometric paper cubes and compass symbolizing Perplexity Multiagent mode
Enterprise Max at $271 per seat per month

On November 17, Perplexity launched Multiagent, a new effort setting above Ultra in Perplexity Computer (its multistep research and app-building agent platform).

Read the full story~2 min

Unlike Ultra, where supporting agents report only to one lead, Multiagent lets agents on different models talk directly, check assumptions, and edit shared files. It is available to all Perplexity Max subscribers on the web.

  • Team size and models: A Claude Fable 5.2 lead assigns up to 16 agents using Claude Opus 5.5, GPT 6.1 Sol, and Gemini 4 Carbon by default.
  • Model overrides: Users can specify models in their query, while Fable 5.2 and GPT 6 Astra join as supporting agents only when requested or needed.
  • Target workflows: Built for interconnected tasks like financial forecasting, annual report review, data room diligence, and tax planning.

What changes

Perplexity Max users on the web can select Multiagent in Computer’s effort slider to run cross-checking model teams from one conversation thread.

Limits

Consumer Max pricing, credit burn rates for Multiagent runs, and regional availability (such as Korea) are not disclosed. Perplexity presents Model Council as a separate mode rather than calling Multiagent its successor.

Community reaction

■One news account framed Multiagent as a successor to Model Council, though Perplexity’s post treats them as distinct options.

■Users praised assigning different frontier models to different agents within a single collaborative task.

■Commenters joked that Multiagent is finally a group project where no teammate ghosts the chat.

Big picture: Multi-model orchestration is evolving from comparing independent answers side by side to letting specialized models debate and edit shared work.

How Multiagent differs from Ultra and Model Council

In its Multiagent announcement, Perplexity outlines three distinct modes for Computer. Ultra delegates subtasks that report back only to the lead, while Model Council runs models independently on the same question to show where they agree or disagree.

Multiagent sits above Ultra, which Perplexity introduced alongside Light, Standard, and High in its September 17 effort controls post. Under Multiagent, up to 16 supporting agents exchange evidence, flag discrepancies in annual reports or deal data rooms, and coordinate edits to shared financial models while the Fable 5.2 lead manages follow-ups.

Availability, enterprise pricing context, and open questions

Perplexity states that Multiagent is live for all Perplexity Max subscribers on the web via the omnibar effort slider. Stopping the main conversation thread halts the entire agent team, and users can view which models each running agent uses as the task runs.

While consumer Max pricing is not listed on the blog, Perplexity’s Enterprise pricing page lists Enterprise Max at $271 per seat per month billed annually, including 15,000 monthly Computer credits and 40,000 bonus credits. Enterprise Pro is $34 per seat per month with 500 monthly credits, and both plans guarantee customer data is not used for training under SOC 2 Type II, HIPAA, GDPR, and PCI DSS compliance, though Multiagent credit burn rates remain unspecified.

Read more · Introducing Computer’s Multiagent effort level ↗ · Computer adds effort mode for model selection ↗ · Perplexity Enterprise Pricing ↗

Developer CLI05 / 05

Claude Code v2.1.327 adds setup confirmations and stdin MCP input

Metallic keycap and paper checklist symbolizing Claude Code v2.1.327 updates

As observed on November 18, Anthropic published Claude Code v2.1.327 on GitHub with safety prompts, stdin-based MCP (Model Context Protocol, a standard for connecting AI tools to external apps) setup, and expanded diagnostics.

Read the full story~2 min

The release focuses on preventing accidental file overwrites or leaked tokens during configuration while fixing dozens of session resume, hook, and proxy issues.

  • Safer configuration: Adds a confirmation prompt before /install-github-app modifies a repository and before /export overwrites an existing file.
  • Secret-free MCP: Adds claude mcp add-json with a trailing dash to read server JSON from stdin (outside an active session), keeping tokens off command-line arguments.
  • Diagnostics: Expands /status to list broken agent, skill, command, and output style files, keybinding errors, unavailable sandboxes, and unreachable hooks.

What changes

Developers can pipe MCP server JSON securely from stdin and catch broken local configs or ignored permissions directly in /status.

Limits

The GitHub release page does not display a release date, and the bottom of the captured release notes is truncated.

Community reaction

■No verified community posts or replies were captured for v2.1.327 in this cycle’s social collection.

■Public discussion around the release notes was not observed on X or Threads before publication.

■Reader feedback on the new stdin MCP input and status diagnostics remains to be seen.

Big picture: Frequent patch releases continue to tighten CLI safety defaults and background session reliability for teams running Claude Code in automated workflows.

Safer MCP configuration, file exports, and diagnostics

Building on Claude Code v2.1.326, which added Code Mode support for agent teams and tool option warnings, the v2.1.327 release notes focus on safer configuration commands. Running claude mcp add-json with a dash reads JSON from standard input so API tokens stay out of shell history, while project-scoped add and remove commands now preserve other top-level keys in .mcp.json.

The update also stops /export from silently overwriting files, adds notes in /config when other settings override a row, and introduces /tui classic for the non-fullscreen renderer. Meanwhile, /status now flags broken extension files, ignored permission settings, and hook matchers that never trigger, and attributionProject lets gateway admins name a session’s billing group.

Network resilience, proxy fixes, and session reliability

For remote environments, the v2.1.327 changelog notes that most remote MCP servers on other machines reached by hostname over HTTPS now retry in the background for about a day after outages such as VPN drops. In addition, claude -p continues after 10 seconds if fetching model settings stalls at a proxy, and CLAUDE_CODE_DISABLE_SHELL_HEAD_START=1 disables pre-running shell start-up files before the first Bash command approval.

Other guardrails limit file lookups to the first 100 @ mentions per message and block WebFetch URLs over 2,000 characters before permission prompts appear. Anthropic also fixed worktree Remote Control sessions deleting git-ignored files like .env, ensured in-process agent teammates compact at their own model’s limit and rate, and improved transcript retries when resuming long sessions.

Read more · Release v2.1.327 · anthropics/claude-code ↗

Fluxaivory desk

Built from primary sources with AI assistance and published after automated checks. AI Use Policy · Corrections

Follow Fluxaivory

Get the briefing by email

A weekly roundup of the top AI news in plain English, from primary sources. Free. Unsubscribe anytime.