v2.3.0 - October 13, 2026

This release covers version 2.3.0 of the @yugabytedb/meko-mcp installer and the public Meko agent skills plugin.

Published skill versions: meko-mcp-tools 3.2.0, meko-mcp-tools-desktop 2.9.0, meko-select-datapack 1.1.3, and meko-select-datapack-desktop 1.2.2.

Install or upgrade

npx @yugabytedb/meko-mcp@2.3.0

Claude Desktop can also install the MCP server from meko-2.3.0.mcpb. That file installs the MCP server only. It doesn't install the Desktop skills. See Claude Desktop.

Claude Code users can also update the skills plugin directly.

The public plugin source is in yugabyte/meko-skills.

Changes

New features

• OpenCode CLI and TUI support - --client opencode configures OpenCode 1.18.x at user or project scope, installs the Meko skill, migrates existing local sessions, and captures new conversations. Restart OpenCode after installation.
• Kiro power package - the public meko-skills repository now includes a Kiro Agent Plugin package for registry submission. It uses Kiro's OAuth sign-in and includes Meko memory guidance and connection troubleshooting. It doesn't capture conversations automatically.
• Capture cleanup commands - --capture-cleanup status inspects capture processes and unsent work. Use stop to park retries, drain to send parked work in the foreground, resume to return it to normal retries, repair to clear failure fields on sessions that are already fully delivered, and discard only when you intend to delete unsent exchanges. repair also removes leftover timer, lease, and job files for processes that are gone. Parking works with the capture hooks from 2.2.0 onward, so npx @yugabytedb/meko-mcp@latest --capture-cleanup stop stops retries without upgrading.
• Broader context search - the updated skills use context_search for open-ended questions such as "what do you know about X?". It searches personal memory, shared knowledge, and past conversations and returns each source in a separate result bucket. The Claude Code installer pre-allows mcp__meko__context_search, and the SessionStart text sends that question to context_search.

Improvements

• Stable agent IDs from the repository root - sessions started in a subdirectory now use the repository name instead of names such as docs, src, or webapp. Sessions outside a repository use the client name. A linked git worktree uses the worktree directory name, so worktrees of the same repository keep separate agent IDs. Existing sessions started in a subdirectory write new turns under the repository-based agent ID after upgrading.
• Clear migration outcomes - --migrate-status reports whether a run succeeded, failed, was cancelled, or crashed, and points to the failure log.
• Codex app authentication - API-key installs now set MEKO_API_KEY in the macOS launchd environment. Quit every running Codex process and relaunch Codex after installation so it reads the new value. A process that stays open keeps the environment it had when it started.
• Current tool guidance - the public agent skills, meko-mcp-tools and meko-mcp-tools-desktop, list all 25 tools, including context_search and knowledgebase_delete_document, and use the current required arguments and response fields.

Fixes

• Capture continues after one rejected exchange - a transcript payload rejected by the edge proxy is recorded as dropped instead of blocking every later exchange in that session.
• Capture reports the server's reason - a quota rejection returned as HTTP 402 or 403 is reported as a quota problem. Status and SessionStart notices carry that reason.
• A missing datapack is an error - the server returns datapack_id_required when no datapack resolves. It used to report success for a conversation it never stored. The SessionStart hook tells the agent to pin a datapack, and --migrate marks the session failed so a rerun retries it.
• Denied datapack access holds the queue - datapack_access_denied means the target datapack isn't owned by or shared with the account. Queued turns stay in the outbox until access is restored.
• Long writes aren't sent again - capture now reads server-sent event keepalives correctly while a conversation turn takes more than 10 seconds to store.
• Deleted conversations recover - after repeated not-found responses, capture creates a new conversation for the undelivered turns without replaying turns already stored.
• Claude Code captures the user's turn - loaded skill text, background task notifications, and a subagent's final report stay with the turn that produced them instead of being stored as new user requests. --migrate applies the same rules to imported Claude Code history.
• SessionStart memory preload works again - Claude Code now loads the recent-memory window using the server's current request and response fields.
• Claude Desktop reports the Meko failure - when a Meko call fails, Claude Desktop reports the failure. It doesn't read a local file as a fallback.
• A conversation that is too large to load stays usable - when the trace is too large, conversation_get and conversation_update return not_found or time out. Agents accept that result and keep working. They don't retry the read or recreate the conversation.

Known limitations

• OpenCode support covers CLI and TUI 1.18.x. It doesn't cover OpenCode Desktop, V2 preview APIs, headless or remote sessions, or child-session migration.
• OpenCode capture stores visible user text and the first visible assistant message. It doesn't store tool calls, hidden reasoning, files, attachments, or streaming output. Compaction summaries aren't excluded from capture.
• context_search reports each source separately. An empty bucket can mean that source failed, so agents confirm with memory_search or knowledgebase_search before reporting that nothing is stored.
• Claude Code capture can skip a prompt that includes a pasted image. The image caption is written as its own user entry, so the prompt's turn has no assistant reply to store.