v2026.4.30 — The Curator — Self-curating skills

Released: April 30, 2026 Theme: Vi starts maintaining its own skill library Breaking-ish change: Secret redaction is now off by default (see Upgrade Recommendations)

This release rolls up a week of work — 1,096 commits / 550 PRs / 217k+ insertions. There’s a lot under the hood, but the headline is: Vi gets smarter the more you use it, and you don’t have to clean up after it.

---

Core Highlights

1. Autonomous Curator — Vi Maintains Itself

Starting this version, Vi runs a background “Curator” that wakes up every 7 days by default and does three things:

• Grades — figures out which skills you actually use vs. which ones are gathering dust
• Consolidates — merges overlapping skills into one
• Prunes — archives skills nothing’s been touching

You don’t have to invoke it. The gateway’s cron ticker triggers it. Each run writes a report to logs/curator/run.json and REPORT.md.

Archived skills are classified as either “consolidated into another skill” or “pruned outright” so you can see what happened. Defense-in-depth gates protect bundled and Hub skills from mutation, so the Curator can’t break your foundation.

To see the current state: hermes curator status ranks skills by usage (most-used vs. least-used).

What it feels like: Vi gets sharper the longer you use it. Skills that aren’t earning their keep retire themselves; the ones you rely on get refined. No manual cleanup.

2. Self-Improvement Loop — Substantially Upgraded

After every turn, Vi forks a background reviewer that decides what to save and which skills to update. This loop got serious work:

• Class-first rubric (instead of free-form “should this update?”) — much more stable
• Active-update bias — prefers updating skills the agent just loaded, not random ones
• Handles references/ and templates/ sub-files inside skills
• The fork inherits the parent’s live runtime — provider, model, and credentials actually propagate now (this used to silently break)
• Restricted to memory + skills toolsets — can’t accidentally escape to shell or web
• Prior-turn tool messages excluded from the review summary so the fork sees a clean context

3. 19 Messaging Platforms (Now with Microsoft Teams + Tencent Yuanbao)

Existing platforms also got attention:

• Telegram / Discord / Slack / Mattermost / Email / Signal all gain native multi-image sending
• Slack adds strict_mention (no more accidental thread engagement) and channel_skill_bindings (bind specific skills to specific channels)
• Signal gets native markdown formatting, reply quotes, and emoji reactions

4. Native Spotify Integration

7 Spotify tools, bundled out of the box (PKCE OAuth):

• Search — tracks, albums, artists, playlists
• Playback control — play / pause / next / previous
• Queue — add to current queue
• Device control — pick which device plays
• Playlist management

There’s an interactive setup wizard. Cron usage is documented.

Use case: “I’m about to enter deep work — put the focus jazz playlist on my laptop.”

5. Google Meet Integration

Google Meet plugin is bundled. Vi can:

• Join the call (with your Google account)
• Transcribe in real time (via OpenAI Realtime API)
• Speak on your behalf (through a Node bot server)
• Follow up after (summarize, email recap, push action items to Notion / Linear)

The whole pipeline ships as a single plugin.

Use case: “I have a customer review at 3pm — join, listen, and afterward send me three action items and email them to the customer.”

6. TUI Upgrades (Big Win for Admin Panel Web-Terminal Users)

This release adds a stack of TUI quality-of-life improvements:

• LaTeX rendering (math/science folks rejoice)
• /resume switches to past sessions, press d in the picker to delete old sessions inline
• /reload env hot-reload — change .env, no restart needed
• Opt-in auto-resume of last session — re-enter the TUI and pick up where you left off
• Modified mouse-wheel line scroll — long output is easier to navigate
• WSL2 ConPTY phantom mouse issue fixed (use /mouse to toggle)

If you reach Vi through the Web Terminal in the Admin Panel, you’ll feel all of these directly.

7. ~57% Faster Cold Start

Vi launches roughly half as fast as before. Under the hood:

• Lazy imports of heavy packages (OpenAI / Anthropic / Firecrawl no longer load at startup)
• Config reads cached by file mtime (no more re-parsing every time)
• Memoized tool definitions with TTL cache
• Precompiled dangerous-command patterns

What you’ll feel: hitting enter in the admin panel web terminal — Vi is ready almost instantly.

8. ⚠️ Breaking-ish: Secret Redaction Off by Default

Vi used to auto-replace anything that looked like an API key or token with *** in tool output. The detection misfired on legitimate strings (some hashes, base64 blobs) and corrupted patches and tool results.

This release flips the default to off. To restore the old behavior, add this to your config:

redaction:
  enabled: true

9. Configurable Prompt Cache TTL

Prompt cache used to expire after 5 minutes flat. You can now extend it to 1 hour via prompt_caching.cache_ttl. For long, multi-turn conversations, the 1-hour setting saves token cost.

prompt_caching:
  cache_ttl: 1h    # default is 5m

10. Other Notable Items

• ComfyUI v5 + TouchDesigner-MCP moved from optional to bundled-by-default (image generation, real-time visual art)
• Four new inference providers: GMI Cloud (first-class), Azure AI Foundry (auto-detection), MiniMax OAuth (browser PKCE flow), Tencent Tokenhub
• LM Studio promoted to first-class provider — dedicated auth, doctor checks, live /models listing
• Models Dashboard tab — switch your main and auxiliary models from the Web Dashboard
• hermes -z <prompt> — non-interactive one-shot mode for scripting
• Remote model catalog manifest — OpenRouter + Nous Portal catalogs are pulled from a remote manifest, so new models show up without a Vi update

---

Upgrade Recommendations

1. Run hermes curator status after upgrading to see your skill rankings and get a feel for the tool
2. Check secret redaction: if your environment requires log redaction, set redaction.enabled: true
3. Spotify users: run hermes tools to enable Spotify and walk through the wizard
4. Google Meet users: this is a big upgrade — worth 10 minutes to wire up the plugin
5. TUI users: try /resume + d to delete old sessions inline, and /reload for env hot-reload
6. realvco customers: enter Vi from the Admin Panel Web Terminal to feel the ~57% cold-start win
7. Long-conversation workflows: consider switching prompt cache TTL to 1h for cost savings

---

← v2026.4.23 · Back to list