Skip to content

v0.8.6

Choose a tag to compare

@github-actions github-actions released this 18 Nov 23:45
· 6 commits to main since this release

Pages get information across better: the bot picks the best form for each piece of information (a chart, a calendar, a table to sort and search, a calculator), uses the width of a computer and one column on a phone, and looks at the page before it sends it. Pages have a new, calmer look and a light and dark switch. The bot hands more of its work to children and asks less of people, and long-term rules come only from people's own words. From pi: prompt cache warming, the grep and find tools, proxy settings, and the notice after a Claude subscription login.

Added

  • pots-ui.js in the page kit, optional and offline. A table with data-pots-chart="bar", "stack" or "line" becomes a bar chart (with reference lines from data-marks), one bar in parts with percentages, or a line chart, and the table stays one tap away. data-pots-tabs shows one panel at a time ("phone": only on a phone), data-pots-sort sorts a table by a click on a column header, data-pots-filter hides rows that do not match a search box, and <nav class="pots-toc"> marks the section on screen. It sets only text, never HTML, and works under the private-page CSP.
  • A light and dark switch on pages. pots-ui.js and pots.js put a small sun or moon button at the page's top corner. The page follows the system until the reader chooses; the choice is kept for every page of the host, and choosing the system's own mode again forgets it. Every colour of pots.css is now a light-dark() pair, so the switch changes all of them, and the form fields too.
  • Prompt cache warming, as in pi. A provider keeps the start of a request cached for about 5 minutes, and the bot often waits longer, for a person's answer or a child's work; the next request then writes the whole conversation to the cache again. Now Pots does what the pi command line does by default since pi 1.5.1: shortly before the cache entry of the main agent's or a child's last request expires, it sends the same request again with a one-token answer, but only when pi's numbers say the warm costs less than the miss it saves. During a run for up to an hour after the request, after it for up to 30 minutes. pi's setting cacheWarming (idle by default, streaming, off) in ~/.pi/agent/settings.json is the switch for Pots and pi alike, read again within seconds. Memory passes, compaction and the judge are not warmed, nor a stopped child. A warm is a model request in pots trace, beside the request it repeats. On three bots' requests of 12 to 18 November, pi's rules would have saved about 17% of their model cost.
  • grep and find, pi's search tools, for the main agent and its children. They search file contents and file names with ripgrep and fd, skip what .gitignore lists and cap their output; when the computer has neither program, pi downloads it to ~/.pi/agent/bin on first use. A child that ran grep -r in bash on a 6 GB folder had not finished after 17 minutes; ripgrep took a second. They run in the gateway's process, not in a systemd scope of their own. The tool list changes, so each bot's prompt cache is written once after the upgrade.
  • Pi's notice after a Claude subscription login. pots init, pots login and the login page show what the pi command line says when Claude is signed in with a subscription: a third-party program draws from extra usage, billed per token, not from the plan's limits. pi's setting warnings.anthropicExtraUsage turns it off in both.

Changed

  • A new stylesheet for pages (pots.css). Calm and plain, made for reading: Inter (in the package, under the SIL Open Font License) for Latin text and the system's font for Chinese, Japanese and Korean, near-black greys, hairlines instead of boxes, and one muted teal accent for links, key numbers, the chosen and the recommended choice, and focus. A page can set its own accent with one line. On a computer, contents at the side and blocks that use the width; on a phone, one column, and a wide table scrolls inside itself, never the page. New parts: an opening with a line above the title and a line of facts (pots-eyebrow, pots-meta), columns and grids (pots-cols, pots-grid, pots-wide), key numbers and their change (pots-stats, pots-delta), key points (pots-points), steps (pots-steps), name and value rows (pots-kv), flows with forks (pots-flow, pots-lanes), terms (pots-terms), pull quotes (pots-quote), 35 line icons with data-icon, a contents list (pots-toc), tables that become one card per row on a phone (table.pots-stack), and chart colours --pots-c1 to --pots-c7. Notes and questions get an icon; tables take the width inside a scroll box (pots-scroll, added by pots-ui.js) with numbers right-aligned; buttons, tabs, choices and fields have hover and focus states. Every class and colour variable of the first stylesheet still works, so pages need no change.
  • The pages skill on display pages. It starts from the goal: a page gets information across better than plain text can, so use a better form where it fits, text where text fits, and never a component for show. A table that pairs kinds of information with their best form, when and how to write a small component of the page's own (the content complete in the HTML, the script in a file), a layout that adapts to the device, an example page, and how to take screenshots as a phone, as a computer and in dark mode before sending, with the skill's script shot.mjs. Each page chooses a muted accent colour that fits its content; the kit's teal is only the fallback. Page text is in plain words, in the style of ASD-STE100 Simplified Technical English in the reader's language, with the reader's own terms.
  • The bot hands most of its work to children. Its prompt said only that longer work goes to children, so the main agent, which thinks less deeply, still did much of the work and answered questions that needed thought itself. Now it hands children as much of the work as it can, also questions that need careful thought.
  • The bot asks less of people. The more it takes on, the less it asks of the person: it does what it can itself and hands the person only what really needs them. Taking over its screen is the last step.
  • What the bot writes for an agent gives the task, not rules on how to do it. The prompt said this only of a child's brief. Now it holds for whatever the bot writes for an agent to act on: detailed rules cannot cover every case, and they limit what the agent can do.
  • Long-term rules come only from people's own words. The memory format, taken from Claude Code, asked for a "How to apply" line under each correction, so a bot wrote one mistake down as a method of its own and read it at every request; the daily consolidation deleted only memories that were stale, wrong or repeated, so such methods stayed and grew. Now a feedback memory holds what happened, the person's own words with the date, and why. The consolidation also rewrites a method that no person gave back into the fact behind it. AGENTS.md is said to hold facts, and what people asked of the bot in their own words.
  • Shorter guidance, without set steps or set words. Three places gave set steps or a text to say, which a strong model works out itself and which can fit the case badly. The prompt's line on coding agents' command lines now says only that, unlike a child, one tells the bot nothing while it runs or when it ends, so the bot looks in on it now and then. The pages skill's Quick Tunnel note says what to tell the person, without a text to say or a memory to save. Step 4 of docs/for-agents.md says what the person signs in to and why, and names the menu's choices, without a text to paste.
  • The network as pi sets it up. The gateway, pots init and pots login now use a proxy from HTTP_PROXY, HTTPS_PROXY and NO_PROXY, or from httpProxy in pi's settings, and pi's httpIdleTimeoutMs, and give each address of a host 2 seconds to connect, as pi does. Before, Node ignored the proxy variables, so on a network that needs a proxy the pi command line worked and Pots did not reach its model. Model requests, MCP servers over HTTP and the chat apps that use fetch go through the proxy; Slack's Socket Mode and Discord's gateway connection do not.
  • Requests name Pots, not pi. OpenRouter, NVIDIA NIM and OpenCode now see Pots (https://pots.sh) as the app; OpenRouter lists apps by this name. Subscription logins send no such name.

Fixed

  • A choice with an explanation on an answer page. A label with a title and a line under it, such as <label><input type="radio"> Test <small>Nobody sees it</small></label>, was squeezed into columns; the explanation now goes under the title.

What's Changed

  • Trim three released texts to what was asked: no set steps or texts to say by @Crokily in #63
  • Use what pi has: prompt cache warming, grep and find, pi's network settings, the Claude subscription notice, and Pots as the app name by @Crokily in #64
  • The bot hands most work to children, asks people only for what needs them, gives agents the task, and keeps lessons as facts by @Crokily in #61
  • Pages that get information across: new kit stylesheet, pots-ui.js, display-page guidance by @Crokily in #60

Full Changelog: v0.8.5...v0.8.6