Skip to content
View in the app

A better way to browse. Learn more.

Universal Devices Forum

A full-screen app on your home screen with push notifications, badges and more.

To install this app on iOS and iPadOS
  1. Tap the Share icon in Safari
  2. Scroll the menu and tap Add to Home Screen.
  3. Tap Add in the top-right corner.
To install this app on Android
  1. Tap the 3-dot menu (⋮) in the top-right corner of the browser.
  2. Tap Add to Home screen or Install app.
  3. Confirm by tapping Install.

Michel Kohanim

Administrators
  • Joined

  • Last visited

  1. @DennisC Not entirely true. You cannot publish if you're not setup as a developer. So, a) verify identity b) set you up as a developer. So, right now, non-developer customers do NOT get the Publish button. With kind regards, Michel
  2. @Guy Lavoie , Thank you! Please submit a ticket and we'll walk you through it next week. With kind regards, Michel
  3. Hello everyone, Version 2.4.2 is released and it's a major improvement for token management for Claude with more than 90% cache hit rate in majority of cases. If you wish, you can also increase token_history_budget to 250K. The only charge will be after 1+ hour of inactivity for a LONG session. Otherwise, it should be smooth sailing. Most interesting thing is /session-stats that gives you a breakdown of everything you need to know regarding your tokens. Here are the release notes: - Compact a Claude session by forking the live conversation. The summarizing request now reuses the turn's own system prompt, history and tools (so the already-cached history is read at the cache-read price) with a final instruction to summarize, and the model sees every tool result; on success the entire history is replaced by that one summary. Any other provider, or a fork that fails, keeps the standalone summarizer (oldest turns summarized, newest half-budget kept), which now also sees one line per tool call with a short result excerpt. `/compact-session` passes the connection context through so it can fork too. - Fix a cache regression on every idle gap: Claude cache breakpoints now all use a fixed 1-hour TTL (tools, every system block, and the conversation, which was hardcoded to 5 minutes), so a pause of up to an hour between turns reads the cache instead of rewriting the whole history. **Upgrade note:** the `cache_ttl` profile option is removed; a leftover `"cache_ttl"` (the example configs set `"5m"`) is silently ignored, so those profiles now cache for 1 hour, where writes cost 2x instead of 1.25x (reads are unchanged at 0.1x). - Remove the installed-plugins list from the plugin-authoring system prompt. It was the last block before the conversation, so any plugin install/start/stop rewrote the whole cached conversation (about 25K tokens in one logged session). The model calls `list_installed_plugins` for live state instead. - Add `/session-stats`: shows what the next request is made of (each system section, the tools, and history split into compacted summary, conversation text and tool calls/results, scaled to the last real request's input total when comparable), the last request's cache reads/writes and hit rate, session totals with an estimated cost for known models, compactions (fork/summary/truncated, before and after), and the last five requests. It makes no model call. Per-request token usage is now recorded for every provider and saved with the session instead of only appearing in the prompt log. New optional profile setting `context_window` (shown as the share of the window used) for models the built-in table doesn't know. - Log (without compacting) when a single turn grows past twice `history_token_budget`: a `would compact mid-turn` warning, a `mid_turn_over_budget` prompt-log line and a counter in `/session-stats`. Compaction is still only evaluated at the start of a turn. - Scrub API keys and tokens out of every log record, including `debug.log`: udi_interface's own handler logged each Polyglot message with the plugin's custom parameters, provider API keys included, which the existing redaction never reached. A logging record factory now removes keys by shape (`sk-ant-`, `sk-`, `xai-`, `gsk_`, `AIza`, `Bearer`) and quoted values after fields such as `api_key`/`client_secret`/`password`. Session persistence is unaffected. Lines written by earlier versions stay in old logs, so rotate any key that appeared in one. - Add the `mdns_search` plugin-authoring tool (a stdlib-only mDNS browser that queries from an ephemeral port so it never competes with the eisy's own responder) and mDNS/Bonjour support in generated plugins (`server_entry.mdns_enabled` wires a handler forwarding to the required `handle_mdns_results` override). - Put a blank line after the history-budget and compaction notices so the reply no longer runs onto the notice's last word. With kind regards, Michel
  4. No, there's an API for moves. Next is to have it make a UI ... With kind regards, Michel
  5. While testing long context, multi-turn, token-gobbling scenarios today, I had it create a chess plugin! And it won! With kind regards, Michel
  6. Well, you need to first make it fail by going back and forth between plugins in the same session! With kind regards, Michel
  7. @DennisC , Thank you so much for the feedback. A minor issue is that the compaction doesn't summarize tool calls, so you will be seeing something like this after compaction: I told you x, but I was lying ... do you want to verify? Another minor issue is that the compaction happens way too frequently (for me). We are working on both of these issues while keeping the costs down. I shall keep you posted. With kind regards, Michel
  8. @DennisC , Good questions: max_turns is no longer used. It does not hurt if you keep it. My suggestions: Customer mode: max_tokens: 20000 history_token_budget: 60000 Developer mode: max_tokens: 64000 history_token_budget: 100000 You can always adjust history_token_budgets if you your conversations are compacted too frequently. With kind regards, Michel
  9. Hello everyone, Initial analysis of the new version showed 70% improvement in token/context usage. Added a little more housekeeping, and removes things that made processing more difficult (max_turns), and in one case we got 96% cache hit developing a plugin for Rheem Econet! So, if you have not done so yet, please upgrade to the latest to save on tokens/costs. With kind regards, Michel
  10. @Guy Lavoie thank you for the report! Regarding the issue, please submit a ticket. With kind regards, Michel
  11. Hello everyone, please upgrade to 2.4.0 ... major improvements in cache management and token usage. With kind regards, Michel
  12. @SJK , Need more information to understand what exactly you mean. With kind regards, Michel
  13. I cannot reproduce at all. So, it's always giving your program numbers instead of names???? Always? With kind regards, Michel
  14. Do these have names? So glad to hear! Ticket site should be fixed now. With kind regards, Michel
  15. It would be great if you can paste an example. This is a bug that needs to be fixed. With kind regards, Michel

Account

Navigation

Search

Search

Configure browser push notifications

Chrome (Android)
  1. Tap the lock icon next to the address bar.
  2. Tap Permissions → Notifications.
  3. Adjust your preference.
Chrome (Desktop)
  1. Click the padlock icon in the address bar.
  2. Select Site settings.
  3. Find Notifications and adjust your preference.