AI Coding News

March 29, 2026

Key Signals

  • OpenCode ships three releases in a single day, accelerating its Effect-based architecture overhaul. Versions 1.3.4 through 1.3.6 landed on March 29, with v1.3.4 delivering TUI plugin support, AI SDK v6 compatibility, and a dedicated GPT system prompt for non-Codex models — then v1.3.5 and v1.3.6 followed quickly with plugin hook and token-counting fixes. The rapid cadence signals that OpenCode is aggressively iterating on its plugin ecosystem and multi-provider architecture, positioning itself as a serious contender in the open-source AI coding tool space. [1][2][3]

  • AI-generated "slop" is overwhelming open source maintainers, threatening the ecosystem that 96% of commercial codebases depend on. Low-quality AI-generated pull requests now take reviewers an estimated 12× longer to evaluate than they take to generate, and the Jazzband project has already shut down entirely due to the flood. Emerging countermeasures — reputation systems like vouch, contributor AI policies across 63 projects, and cryptographic identity proofs — remain nascent, meaning this pressure on maintainers will likely intensify before it eases. [4]

  • Claude Code patches a Cowork Dispatch messaging bug, keeping multi-agent collaboration on track. Version 2.1.87 is a targeted single-fix release addressing undelivered messages in Cowork Dispatch, Anthropic's multi-agent coordination feature. While minor in scope, the quick turnaround reflects how central reliable inter-agent messaging has become to Claude Code's agentic workflow model. [5]

AI Coding News

  • AI-generated low-quality contributions are creating an existential crisis for open source software. The flood of AI-generated PRs — described as "Slopmageddon" — is draining unpaid maintainers who form 60% of the open-source workforce. Projects like Jazzband have shut down, and the Godot engine maintainers describe the experience as "draining and demoralizing." Red Monk research has identified 63 formal AI contribution policies across open-source foundations and projects, with approaches ranging from mandatory disclosure to outright bans. GitHub has introduced tools to help and even suggested disabling PRs entirely, while community-driven solutions include Mitchell Hashimoto's vouch trust system and the good-egg contributor scoring project. Security risks compound the problem — AI-generated contributions can introduce subtle vulnerabilities and poorly understood dependencies that expand the attack surface. [4]

Feature Update

  • OpenCode v1.3.4 delivers a major architectural overhaul with Effect-based internals, TUI plugin support, and AI SDK v6 compatibility. The release refactors core services — session processing, compaction, config, LSP, plugin, and skill services — to use the Effect library for structured concurrency and error handling. A new model variant selection dialog replaces the previous cycling UI, and a dedicated GPT system prompt has been added for non-Codex GPT models. Desktop improvements include Azure Artifact Signing for Windows releases, improved startup efficiency, and reduced markdown rendering jank during streaming. MCP transport now properly closes on failed or timed-out connections, and app startup performance has been improved. [1]

  • OpenCode v1.3.5 fixes async plugin hooks and refines the GPT system prompt. Plugin hooks now properly handle async operations, resolving a regression introduced alongside the plugin architecture in v1.3.4. The GPT prompt has been adjusted to be more minimal, and file reference annoyances that could confuse the model have been fixed. [2]

  • OpenCode v1.3.6 fixes token usage double-counting for Anthropic and Amazon Bedrock providers. Session metrics were reporting inflated token counts due to a double-counting bug specific to Anthropic and Amazon Bedrock backends. The TUI variant dialog search has also been fixed so typing now properly filters the list of available model variants. [3]

  • Claude Code v2.1.87 fixes Cowork Dispatch message delivery. Messages sent through Cowork Dispatch — the coordination layer for multi-agent workflows — were not being delivered, potentially causing silent failures in delegated tasks. This single-fix patch release resolves the issue. [5]