Claude Code v2.1.296 adds allow_large to the Read tool so huge files can be read in a single pass, and loosens the limits on MCP descriptions and retry waits. Sonnet 5.5's cache reads are halved from $0.20 to $0.10. New knobs for subagent auto-compaction and model alignment were added as well.
Codex CLI 0.162 adds a tool that lets the agent itself create "managed git worktrees." It makes parallel tasks pinnable in the command center, pulling the groundwork for unattended, parallel operation inside the agent. The challenge of where to draw trust boundaries remains.
Fail-open—where a broken safety device silently waves operations through—was the default blind spot. Claude Code v2.1.294/295 fixed an instruction-based hook bug and made it possible to choose "stop on failure" via onFailure:"block." Here are the key points, the differences from other agents, and th
Anthropic has released Claude Haiku 5.5, a new version of its small model. Priced at roughly one-tenth of the previous generation and expanded to a 1M-token context, it jumps from 0% to 39.2% on the agentic coding benchmark Terminal-Bench. Adopted as Claude Code's default Haiku, it settles into the
OpenAI's Codex lead pledged that for 28 days the team will either ship one improvement that helps most users every day or reset everyone's usage limits. Day one boosted generation speed by about 1.5x. We lay out the aims—and the criticism that it turns usage into a gamble.
Claude's Max and Team plans now include monthly API credits. Max 20x gets $200, Team up to $500. The credits go toward running your own agents via the Agent SDK or claude -p, but the limits are clear: interactive Claude Code and rollover are both out of scope.
Cryptographic Context Injection (CCI), disclosed by Adversa AI, is a technique that smuggles encrypted instructions past content filters and gets GitHub Copilot CLI to exfiltrate local secrets on its own. It depends on autopilot mode and having the agent browse an external URL, and GitHub has drawn
Claude Code v2.1.292 (released October 6) fixes the "vanishing jobs" of unattended operation—/loop stopping on restart, scheduled tasks that fail to fire—and more. It adds effort specification for subagents and closes a batch of mid-read file-swap security holes.
Claude Code v2.1.291 (October 6) fixes, same-day, a regression brought in by the major v2.1.290 release the day before. It plugs a bug that dropped replies to permission prompts in cloud sessions, along with a bug that lost the final message on exit. The day after approvals were reinforced, the "pat
Claude Code v2.1.290 (October 5) adds ceiling to indicate an organization's approval level, agentId to identify subagents, and serverToolUses to record server-side tool runs to its approval hooks. It also rolls up fixes for "silent failures" in unattended operation, such as scheduled tasks stopping
On October 1, GitHub released "Computer Use" in public preview, letting Copilot directly operate the screens of desktop apps. It opens the door to automating legacy software without an API and to E2E testing, while this piece lays out the risks of misoperation and information leaks and the key point
Claude Code v2.1.289 shipped on October 3. It introduces isolation so that "Mods"—extensions that run inside the core—no longer take the whole session down when they break mid-render, and it fixes a gap where dangerous commands hidden behind environment variables slipped past permission rules. It ad
