This release covers version 2.2.0 of the `@yugabytedb/meko-mcp` installer and the public Meko agent skills plugin. The installer, public plugin, and Meko MCP server now release independently, so client wiring and agent guidance no longer wait for a server release.

Published skill versions: `meko-mcp-tools` 3.1.0, `meko-mcp-tools-desktop` 2.8.0, `meko-select-datapack` 1.1.2, and `meko-select-datapack-desktop` 1.2.1.

## Install or upgrade

```bash
npx @yugabytedb/meko-mcp@2.2.0
```

Claude Code users can also [update the skills plugin directly](../../../integrations/connect-to-ai-agent/#updating-the-skills-plugin).

See the [installer 2.2.0 release](https://github.com/yugabyte/meko-mcp-server/releases/tag/installer-v2.2.0) for the Desktop extension and release audit evidence. The public plugin source is in [`yugabyte/meko-skills`](https://github.com/yugabyte/meko-skills).

## Changes

### New features

- **Kiro IDE and CLI support** — `--client kiro` configures the shared Kiro MCP and skills locations, installs durable hooks for both Kiro hook formats, and supports `$KIRO_HOME`. The `kiro-cli` and `kiro-ide` aliases select the same client.
- **Durable capture with visible status** — Claude Code, Cursor, Codex, and Kiro keep failed exchanges in a local outbox and retry them without marking the transcript as saved. Run `npx @yugabytedb/meko-mcp@2.2.0 --capture-status` or add `--json` for machine-readable health.

### Improvements

- **Independent release versions** — the npm installer and Desktop extension advance together. The public skills plugin and Meko MCP server keep their own versions.
- **Safer agent guidance** — the skills distinguish failed searches from empty results, treat `memory_get_all` as a recent window instead of an export, limit the evidence passed into model context, and verify writes before reporting success.
- **Production-safe retrieval guidance** — agents use `memory_search` for personal memory and `knowledgebase_search` for shared datapack knowledge. The published skills don't direct agents to unreleased retrieval tools.
- **Current memory contracts** — `memory_search` uses semantic, keyword, and entity signals. `run_id` filters memory reads and deletes by conversation; writes still use `conversation_id` for trace scope.
- **Working Codex capture and migration** — Codex transcripts now use the `codex:<repo>` agent ID, hold incomplete turns for retry, and migrate historical rollout files.

### Fixes

- **Backspace works in the API-key prompt** — interactive entry now treats both terminal erase byte forms as Backspace instead of saving them in the key.
- **Claude Desktop startup** — automated installs require Node.js 20.18.1 or newer and pin `mcp-remote@0.1.38` to prevent “Server disconnected” failures.
- **Migration dry runs stay read-only** — `--migrate --dry-run` no longer creates migration progress or run state.
- **Claude Desktop uninstall instructions** — config-file installs, OAuth connectors, and imported Desktop skills now have separate removal steps.

### Known limitations

- Kiro hooks expose the user prompt and final assistant response. Kiro doesn't expose tool calls or hidden reasoning through these hooks, so Meko can't capture them automatically.
- Claude Desktop has no session hooks. Automatic capture still depends on the selected model following the installed skill and profile instructions.
- `memory_get_all` returns a recent window and a total count. MCP doesn't currently provide a complete memory export.