Tech Radar - 2026-10-25 06:00
Priority Today
Claude Code 2.1.310 fixes gateway failures and opens Monitor on Bedrock, Vertex, Foundry
Anthropic released Claude Code 2.1.310 on October 23, bringing the Monitor tool and plugin monitors to Bedrock, Vertex, and Foundry (with telemetry or nonessential traffic disabled), and blocking long foreground sleeps in those setups. The release also retries failed requests behind some gateways that reject the effort-2025-11-24 beta value or mid-conversation system messages, stops background cleanup from deleting recent files like pasted images when cleanupPeriodDays is large, and fixes compaction after "Prompt is too long" with dense text.
Why it matters: Gateway compatibility and cleanup/compaction bugs can break long-running agent sessions; the Monitor expansion gives enterprise multi-cloud users observability hooks.
What changed: Monitor + plugin monitors available on Bedrock/Vertex/Foundry; retry logic for gateway incompatibilities; multiple bug fixes including Windows PreToolUse hooks with missing Git Bash, symlink permission prompt clarity, and compaction behavior.
For me: Upgrade if you run Claude Code on Bedrock/Vertex/Foundry or hit gateway/proxy failures; test hooks on Windows if using Git Bash dependencies.
Tags: #topic/ai-coding #topic/security #entity/anthropic #entity/claude-code #technology/agent-security #event/tool-release #relevance/developer #learning/understand
Sources:
Codex CLI 0.164.0 adds presets, inline args, and credential masking
OpenAI released Codex CLI v0.164.0 as stable on October 23. It adds ten-slot /presets for reusable prompts/slash commands, inline arguments for /model and /permissions, image previews in Kitty/Sixel terminals, GA enterprise MCP auth (account grants, dedicated sign-in, or OAuth with secrets), credential masking for remote Bash and zsh snapshot sessions (including custom providers), and Bedrock setup defaulting on. Sandbox checks were hardened for case-variant paths and replaceable symlinks.
Why it matters: Reusable presets and inline args reduce friction for repeat workflows; credential masking reduces secret leakage in agent transcripts/snapshots.
What changed: Feature additions plus sandbox hardening and fixes (failed compaction, Windows proxy auth for HTTPS/WebSocket, etc.).
For me: Upgrade if you use Codex CLI, save common roles as presets, and verify credential masking in your environment.
Tags: #topic/ai-coding #topic/security #entity/openai #entity/codex #technology/mcp #technology/agent-security #event/tool-release #relevance/developer
Sources:
Claude Code sandbox/permission fixes continue in 2.1.309
Claude Code 2.1.309 (October 22) fixed multiple sandbox and permission bypasses: sandboxed Bash ran startup files outside sandbox on macOS/Linux, macOS denyWrite missed symlinked paths, sandboxed commands/auto-accepted edits could modify --mcp-config, redirect/compound command bypasses, worktree isolation gaps, and Windows PowerShell permission auto-approvals. It also made sed -i run as normal Bash (so /rewind won't restore), added claude_code.safety_event telemetry, and sandbox.allowNestedUserNamespaces for Linux.
Why it matters: These are boundary enforcement fixes; incorrect sandbox/permission behavior is a direct agent security risk.
What changed: Broad hardening across sandbox startup, symlink handling, permission checks, and hooks; operational changes to sed -i behavior and trust requirements.
For me: Upgrade to at least 2.1.309/2.1.310 and recheck deny rules, automation using sed -i, and sandbox assumptions.
Tags: #topic/ai-coding #topic/security #entity/anthropic #entity/claude-code #technology/agent-security #event/tool-release #relevance/developer #learning/understand
Sources:
OpenCode 2.0 moves to background service with API/SDK
OpenCode 2.0 shipped October 12, moving from a one-shot CLI invocation to a background service with its own API and SDK, and introducing Code Mode for scriptable tool execution. The architectural shift allows other programs and scripts to call into OpenCode rather than relying on interactive terminal sessions.
Why it matters: Turning the agent into a service enables deeper integration into workflows and automation beyond interactive use.
What changed: Background service architecture with API/SDK, plus Code Mode for programmatic/scriptable execution.
For me: Worth watching if you want to embed OpenCode into automation or build tools on top of it.
Tags: #topic/ai-coding #topic/developer-tools #entity/opencode #technology/agent-orchestration #event/tool-release #relevance/developer #learning/understand
Sources:
Kimi Code 2.2.0 adds mid-turn notifications and desktop app entry
Kimi Code 2.2.0 added opt-in NotifyUser mid-turn progress, a kimi app [path] desktop chat entry, and capped in-turn waits for background tasks at 90s (longer timeouts no longer accepted; new messages end the wait). The changes affect how background work is surfaced and bounded during agent runs.
Why it matters: Better progress signaling and bounded waits help manage long-running tasks without indefinite blocking.
What changed: NotifyUser (opt-in), desktop app entry, 90s cap on in-turn waits for background tasks.
For me: Worth trying if you use Kimi Code and want clearer progress or a desktop entry point.
Tags: #topic/ai-coding #topic/developer-tools #entity/moonshot #technology/agent-orchestration #event/tool-release #relevance/developer #learning/awareness
Sources:
Cortex CLI brings governed agentic AI to the terminal
Pervaziv AI announced Cortex CLI, bringing the Cortex Enterprise AI platform to the terminal with natural language engineering, project context, action validation, and user control across macOS, Windows, and Linux. It connects investigation, implementation, testing, review, and validation without leaving the terminal environment.
Why it matters: Enterprise-focused terminal agent with governance/validation built in addresses needs around control and traceability.
What changed: New terminal CLI surface for Cortex platform with integrated workflow across engineering phases.
For me: Relevant if you work in environments where governed agentic tooling is required.
Tags: #topic/ai-coding #topic/ai-agents #entity/pervaziv #technology/agent-orchestration #event/agent-release #relevance/practical #learning/awareness
Sources:
Kiro adds sandbox controls and autonomous mode updates
Kiro updates added /sandbox for session-scoped sandboxing without global changes, with sandbox allowing local network by default, plus autonomous mode adjustments. The updates focus on safer unattended runs and clearer sandbox controls.
Why it matters: Session-scoped sandboxing gives tighter control without changing global defaults for all runs.
What changed: Session sandbox control, local network allowed by default in sandbox, autonomous mode behavior updates.
For me: Worth noting if you use Kiro for unattended agent runs.
Tags: #topic/ai-coding #topic/security #entity/kiro #technology/agent-security #event/tool-release #relevance/developer #learning/awareness
Sources:
AI Coding and Agents
GitHub Copilot CLI updates add subagent hooks, marketplace/git-subdir, and billing visibility
GitHub Copilot CLI saw multiple updates in mid-October (notably 1.0.97-1.0.103). Changes include agent-scoped declarative hooks for custom agents, remote plugin marketplace with git-subdir sources and retry with signed-in token, autopilot denying preToolUse "ask" decisions, discovery of all custom agents even with >100 or nested files, managed permission rules read-only/hidden unless org-set, billed AI Credits shown in /usage, skills/custom agents from --plugin-dir reach autocomplete, and ACP sandbox settings honored on resume. Inline images render by default on supported terminals.
Why it matters: Tighter control surface for custom agents/plugins and better visibility into billing/permissions in multi-agent workflows.
What changed: Feature additions and fixes across hooks, plugins/marketplace, permissions, usage visibility, and terminal behavior.
For me: If you use custom agents/plugins or org-managed permissions, test upgrades carefully against your config.
Tags: #topic/ai-coding #topic/developer-tools #entity/github #technology/mcp #event/tool-release #relevance/developer #learning/awareness
Sources:
Tag Index
- #entity/anthropic
- #entity/claude-code
- #entity/codex
- #entity/github
- #entity/openai
- #event/tool-release
- #learning/awareness
- #learning/hands-on
- #learning/understand
- #relevance/developer
- #technology/agent-security
- #technology/mcp
- #topic/ai-coding
- #topic/developer-tools
- #topic/security