Key Signals

• GitHub open sources the Copilot for Eclipse plugin under the MIT license, exposing the full implementation of agent mode, MCP integration, and system prompts. This marks a significant transparency milestone — developers can now inspect exactly how code completion, Next Edit Suggestions, chat, custom agents, isolated subagents, and prompt files are wired inside Eclipse. The move aligns Copilot with Eclipse's open ecosystem ethos and invites community contributions to AI-powered developer tooling internals. [1]

• Claude Code renames /simplify to /code-review with inline GitHub PR comment posting, bridging the gap between terminal-based AI agents and pull request workflows. The new command accepts an effort level and the --comment flag posts correctness findings directly as inline PR comments. This positions Claude Code as a tool that not only writes code but actively participates in review workflows, potentially reducing the manual overhead of code review triage. [2][3]

• OpenAI Codex 0.133.0 enables Goals by default with dedicated storage and cross-turn progress tracking, signaling a shift toward persistent multi-session intent management in coding agents. Goals now track progress across active turns and pause continuation loops on usage limits or repeated blockers. Combined with enhanced permission profiles and richer extension lifecycle events, Codex is evolving from a single-turn tool into a persistent development companion. [4]

• Copilot CLI v1.0.51 introduces /security-review for vulnerability scanning and preMcpToolCall hooks, expanding both security posture and MCP extensibility. The stable release also adds a customizable terminal status line, --session-id for resuming sessions, secret scanning for commit messages and PR descriptions, and /chronicle cost-tips for personalized token-usage recommendations. These additions reflect a maturing CLI agent that's increasingly security-aware and cost-conscious. [5]

• Google's redesigned Android CLI reduces LLM token usage by over 70% and enables 3x faster task completion through an agent-friendly toolchain interface. The CLI is compatible with Claude Code, Codex, and Gemini agents and is complemented by "Android Skills" — modular SKILL.md instruction sets that trigger automatically on matching prompts. A built-in real-time knowledge base ensures agents access current documentation regardless of training cutoff dates. [6]

• A new architectural primitive called "plans" proposes collapsing CI inner/outer loops for coding agents, running end-to-end integration checks against real environments in seconds rather than minutes. Plans are agent-authored DAGs of typed actions that validate one user-visible behavior against ephemeral environments. This approach addresses the fundamental mismatch between agents that iterate in seconds and CI pipelines that take 10–30 minutes, catching integration bugs before PRs are opened. [7]

AI Coding News

• Google's Android CLI and Skills system makes the mobile toolchain natively agent-friendly, with community reactions noting token-usage gains but questioning verification gaps. The InfoQ coverage highlights that while the 70% token reduction and 3x speed improvement are compelling, developers observe that the biggest bottleneck remains testing and verifying agent-generated code, not project setup. The skills system uses markdown-based specifications that auto-trigger from prompts, a pattern increasingly adopted across AI coding tools. [6]

• The concept of CI "plans" redefines integration testing for the age of coding agents, proposing ephemeral environments scoped per run with agent-authored validation DAGs. The New Stack article demonstrates this with a HotROD demo where an agent catches a struct field rename breaking a frontend contract — a bug invisible to unit tests — and self-heals within a single loop. The pattern suggests CI will evolve from pipeline-triggered to library-callable, with agents invoking validation directly during authoring. [7]

Feature Update

• GitHub Copilot for Eclipse is now open source under the MIT license, with code available on GitHub. The repository exposes the implementation of code completion, Next Edit Suggestions, chat, agent mode, skills, prompt files, BYOK, custom agents, isolated subagents, and MCP integration. Community contributions have already begun from developers including @iloveeclipse, @travkin79, and @rsd-darshan. [1]

• Copilot CLI v1.0.51 ships as a stable release with /security-review, customizable status lines, and session management. Key additions include --session-id=<id> for resuming/naming sessions, /remote commands respecting organization cloud policies, preMcpToolCall hooks for controlling outgoing MCP request metadata, /chronicle cost-tips for personalized cost reduction recommendations, postToolUse hook additionalContext injection, secret scanning for commit messages and PR descriptions, and GitHub MCP web search available without tool search. The input area now grows responsively with terminal height. [5]

• Copilot CLI v1.0.52-0 adds deferred tool loading for custom agents and upgrades subagents to GPT-5.4/5.5. Custom agents can now opt into deferred tool loading via deferred-tool-loading frontmatter, enabling tool-search discovery for agents with large tool lists. /compact accepts optional focus instructions, /usage shows quota progress bars for session and weekly limits, and legacy MCP OAuth config keys are automatically migrated. [8]

• Claude Code v2.1.147 introduces /code-review with PR comment posting, pinned background sessions, and a substantially improved auto-updater. Pinned background sessions stay alive when idle and restart in place for updates. The auto-updater now retries transient network failures and reports specific error categories. Fixes address enterprise login enforcement, MCP pagination dropping items past page 1, Windows PowerShell failures, and pasted text appearing as unreadable placeholders. [2][3]

• OpenAI Codex 0.133.0 enables Goals by default and adds permission profile inheritance with Windows sandbox integration. codex remote-control now runs like a foreground command with explicit daemon-style start/stop. Plugin discovery gains marketplace-aware list output with installed versions, and extensions can observe subagent start/stop, tool execution, turn metadata, and async approval events. A new canonical Codex package archive pipeline standardizes distribution across installers, npm packages, DotSlash, and SDK runtimes. [4]

• OpenCode v1.15.7 adds Grok OAuth sign-in and a redesigned desktop interface. The release restores native OpenAI OAuth requests and reasoning streams, adds PDF attachment support for Grok, and surfaces tool schema failures as friendly errors. The desktop app gains a new home view, session entry flow, pinch zoom, and log export. V2 session APIs now return structured errors with log reference IDs instead of exposing internal config details. [9]

• Gemini CLI v0.44.0-nightly adds AgentSession invocations and Sublime Text/Emacs editor support. The nightly build introduces LocalSessionInvocation and RemoteSessionInvocation wired into the agent-tool, enabling nested agent sessions. Other fixes address snapshot recovery across sessions, proxy handling, nullable array types in MCP tools, and path traversal prevention in custom commands. Windows PowerShell 5.1 is deprioritized in favor of pwsh.exe. [10]

• Kiro CLI v2.4.1 fixes MCP environment variable expansion and Kiro Cloud adds a stop-the-agent control. The CLI fix addresses $${VAR_NAME} syntax not being recognized in MCP server configuration. Kiro Cloud gains a workspace progress indicator during clone/spin-up, a stop button for halting agents mid-task, mobile-responsive layout, and intent-based auto-scroll. [11]

• Kiro IDE v0.12.224 introduces granular shell deny rules with Full/Partial/Base scope selection. "Always deny" on a shell command now opens the same trust pattern picker as "Always allow", replacing the previous all-or-nothing blocking. Session tabs auto-title from the first message. Bug fixes address supervised mode not prompting for approval before edits, checkpoint restore not reverting changes, and subagents getting stuck in infinite retry loops on file writes. [12]