AI Coding Tools Update — August 12, 2026: Copilot Memory, Codex Agent Plugins & Claude Code 2.1.228
Four of the most-used AI coding tools shipped meaningful updates over the past week. Claude Code moved to version 2.1.228 with a batch of session-reliability fixes, GitHub Copilot for JetBrains picked up persistent memory and local models via Ollama, OpenAI's Codex CLI reached 0.147.0 with portable Agent Plugins and a new MCP protocol version, and OpenCode kept its weekly release cadence going with session exports and right-to-left layout support.
None of these is a single blockbuster release. Taken together, though, they point at where agentic coding tools are heading: longer memory across sessions, more portable ways to extend an agent, and tighter control over what an agent is allowed to do without asking first.
Claude Code 2.1.228: cross-session messaging fixes and safer writes
Anthropic pushed three Claude Code releases in the space of a week. Version 2.1.228 (August 11) is mostly a reliability pass, but it touches areas that matter for anyone running long agent sessions:
- Fixed interactive sessions that could stop redrawing entirely after rare internal layout errors.
- Fixed Remote Control's /resume leaking conversation history into connected sessions — a real cross-session data hygiene bug.
- Fixed cross-session messaging starting without an inbox on new installations, and improved inline display and sender identification for those messages.
- Hardened skills synced from claude.ai so they can no longer shadow local commands.
- Changed the Write tool to let newer models overwrite files without reading them first, matching how the Edit tool already behaves.
The two releases just before it were more operationally focused. 2.1.227 (August 10) fixed feature flags evaluating without the user's subscription tier on an expired login token, and fixed every Bash command failing under claude-code-action when allowed_non_write_users runs on GitHub-hosted runners — a real blocker for teams using Claude Code in CI. 2.1.225 (August 8) added gateway spend-limit support to the usage warning, added a workspace-trust prompt to claude agents for untrusted directories, and fixed intermittent 401 errors on MCP OAuth servers on macOS.
GitHub Copilot for JetBrains gets memory and local models via Ollama
GitHub's August 11 update to Copilot for JetBrains is the most feature-dense of the week's releases. The headline addition is Copilot memory: a toggle in the Copilot settings portal that lets the agent retain and recall useful information across separate agent chat sessions, instead of starting cold every time.
- Ollama can now be used as a bring-your-own-key (BYOK) provider inside JetBrains IDEs, giving developers a local-model option alongside hosted models.
- New server-based enterprise controls cover plugin availability, MCP server access, permission settings, and OpenTelemetry configuration.
- Codex sessions now surface in debug logs and support custom instructions and skills, and the JetBrains plugin can auto-install the Copilot CLI from an integrated terminal on macOS, Linux, and Windows.
It follows GitHub's August 3 weekly release across the desktop app, CLI, and VS Code, which added an Auto mode indicator showing which model handled each request plus credit and cache details, a Sessions sidebar for managing multiple concurrent sessions with keyboard shortcuts, and a /side command to explore a parallel question without disrupting the main task. GitHub has also confirmed a round of model deprecations landing September 1, and resumed rolling out Kimi K3 under usage-based billing.
OpenAI Codex CLI 0.147.0: Agent Plugins and MCP 2026-07-28
OpenAI's Codex CLI reached 0.147.0 on August 7, introducing portable Agent Plugins that can be installed and searched across local, personal, workspace, and remote catalogs — a step toward treating agent extensions like a package ecosystem rather than one-off scripts.
- Added support for the MCP 2026-07-28 protocol revision, including paginated discovery, plus an SDK upgrade to MCP 3.0.0.
- Added persistent thread sections with manual ordering, and enabled cached web search plus remote conversation compaction on Amazon Bedrock.
- Introduced --approve-for-me for automatically reviewed approvals, while removing the older codex exec --full-auto flag in favor of --sandbox workspace-write. Anyone with the old flag baked into CI scripts will need to update it.
A follow-up hotfix, 0.146.1 (August 5), tightened default behavior for cyber-capable models and improved how the terminal UI explains permission changes. On August 10, OpenAI also introduced Daybreak Blue and Daybreak Red — separately approved access tiers for vetted defensive and offensive security research, layered on top of the Codex/ChatGPT stack rather than the CLI itself.
OpenCode's steady weekly cadence
OpenCode, the open-source, provider-agnostic coding agent from SST, continued its habit of shipping almost daily point releases through early August:
- v1.18.16 (Aug 10) added a right-click project menu on the home screen and a fallback search path for when certain server features are unavailable.
- v1.18.15 (Aug 7) fixed chronological message ordering and added JSON export of full session transcripts from the UI.
- v1.18.14 (Aug 5) simplified xAI login to a single device-code flow that works better in headless and remote environments.
- v1.18.13 and v1.18.12 (Aug 4) added early right-to-left layout support, fixed markdown rendering performance, and resolved Azure compatibility and composer lag when drafts include large pasted attachments.
OpenCode's changes are smaller individually, but the pace is the story: a community-driven agent is iterating on parity with the commercial tools roughly as fast as they are.
The pattern: memory, plugins, and permission control
Lined up side by side, this week's releases cluster around three themes. First, memory: Copilot's new cross-session memory and Claude Code's cross-session messaging both push these tools past the single-conversation model, toward agents that carry context forward on their own. Second, extensibility: Codex's portable Agent Plugins and the move to MCP 2026-07-28 are about making agent capabilities installable and shareable rather than hand-rolled per project. Third, permissioning: Codex's --approve-for-me flag and Claude Code's workspace-trust prompt are both attempts to let an agent move faster in a trusted context while still drawing a hard line at untrusted ones.
That combination — an agent that remembers more, does more on its own, and needs a clearer approval model — is exactly the shape you'd expect as these tools move from autocomplete-plus-chat toward semi-autonomous coding agents running in CI, on remote infrastructure, and across long-lived sessions.
What this means if you use these tools
- If you script codex exec --full-auto in CI, switch to --sandbox workspace-write before the flag is removed from your pipeline's Codex version.
- If you run Claude Code with claude-code-action and allowed_non_write_users on GitHub-hosted runners, update to 2.1.227+ — the prior versions could fail every Bash command in that configuration.
- If you build MCP servers or clients, check compatibility with the MCP 2026-07-28 revision and its paginated discovery before Codex users hit it in the wild.
- If you use JetBrains and want to keep sensitive code on local hardware, the new Ollama BYOK option is worth testing against your existing Copilot workflow.
- If you're evaluating a free, provider-agnostic alternative, OpenCode's release velocity this month makes it worth another look even if you passed on it earlier in the year.
Sources & further reading
Comments
Share your thoughts and join the conversation