samagotchi 0.42.0 → 0.43.0

data/lib/samagotchi/plugin/attached_context.rb

@@ -27,27 +27,71 @@ module Samagotchi
       # +name+ and +cmd+ (+hint+, +every_seconds+). A source of that name
       # already attached stays as it is. One the user removed from the
       # session (its name, or its URL) isn't attached again (#declined?)
-      # until the user adds it again.
+      # until the user adds it again, or +force+: the user's own explicit
+      # choice (a click), which clears the declined marker (by the name,
+      # and by the URL).
       # @return [String, nil] the source's name; nil when declined
       # @raise [Error]
-      def attach(url: nil, name: nil, cmd: nil, why: nil, hint: nil, every_seconds: nil)
-        own = location or raise Error, "this session has no id yet"
+      def attach(url: nil, name: nil, cmd: nil, why: nil, hint: nil, every_seconds: nil, force: false)
+        own = own_location!
         source = url ? from_url(url, name: name, why: why) : from_cmd(name: name, cmd: cmd, why: why, hint: hint, every_seconds: every_seconds)
-        return nil if own.declined?(source.name) || (url && own.declined_hints.include?(source.hint))
+        if force
+          own.undecline(source.name, hint: url && source.hint)
+        elsif own.declined?(source.name) || (url && own.declined_hints.include?(source.hint))
+          return nil
+        end
         return source.name if own.source(source.name)
 
-        own.add(source.with(scope: own.scope))
-        source.name
+        add(own, source)
       rescue ContextSources::Invalid, ContextProviders::Invalid => e
         raise Error, e.message
       end
 
-      # Whether the user removed +name+ from this session (so #attach skips it).
-      def declined?(name)
-        own = location
-        own ? own.declined?(name) : false
+      # Marks +name+ (or +url+'s source, through its provider) declined in
+      # this session, as the user's removal does: #attach skips it until the
+      # user adds it again. An attached source stays (chi context rm).
+      # @return [String] the name
+      # @raise [Error]
+      def decline(url: nil, name: nil)
+        own = own_location!
+        if url
+          resolved = ContextProviders.resolve(url) or raise Error, "no installed bundle resolves #{url}"
+          name = name ? ContextSources.check_name!(name) : resolved.name
+          hint = resolved.hint
+        else
+          raise Error, "give url: or name:" if name.to_s.empty?
+
+          name = ContextSources.check_name!(name)
+        end
+        own.decline(name, hint)
+        name
+      rescue ContextSources::Invalid, ContextProviders::Invalid => e
+        raise Error, e.message
+      end
+
+      # Whether the user removed +name+ (or +url+) from this session (so
+      # #attach skips it).
+      def declined?(name = nil, url: nil)
+        own = location or return false
+        name_declined?(own, name) || (!url.to_s.empty? && own.declined_hints.include?(url))
+      end
+
+      # Records that this plugin offered +name+ (+hint+: its URL; +why+: the
+      # reason an attach from the offer gives) here, so a worker restart
+      # doesn't offer it again.
+      # @return [ContextSources::Offer]
+      # @raise [Error]
+      def mark_offered(name, hint, why: nil)
+        own_location!.mark_offered(name, hint, why: why)
+      rescue ContextSources::Invalid => e
+        raise Error, e.message
+      end
+
+      # @return [ContextSources::Offer, nil] the offer #mark_offered recorded
+      def offered(name)
+        location&.offered(name)
       rescue ContextSources::Invalid
-        false
+        nil
       end
 
       # This session's sources (its own, then its project's).
@@ -69,6 +113,26 @@ module Samagotchi
 
       def state_dir = @host.state_dir&.call || Session.default_state_dir
 
+      # Two attaches at once (a double click): the one that lost the race
+      # finds the source there and returns its name too.
+      def add(own, source)
+        own.add(source.with(scope: own.scope))
+        source.name
+      rescue ContextSources::Invalid
+        raise unless own.source(source.name)
+
+        source.name
+      end
+
+      # An invalid name isn't declined (and doesn't hide the URL's check).
+      def name_declined?(own, name)
+        !name.nil? && own.declined?(name)
+      rescue ContextSources::Invalid
+        false
+      end
+
+      def own_location! = location || raise(Error, "this session has no id yet")
+
       def location
         id = @host.session_id.call
         id && ContextSources.session_location(id, state_dir: state_dir)

data/lib/samagotchi/tool_declarations.rb

@@ -3,6 +3,7 @@
 require "json"
 require_relative "tools/args"
 require_relative "llm_context_strategy"
+require_relative "tools/memory"
 
 module Samagotchi
   # Tool-declaration constants, protocol constants, and guidance text.
@@ -146,7 +147,7 @@ module Samagotchi
             },
             description: {
               type: "string",
-              description: "Optional one-line description (at most 200 characters) appended to the managed index line; without content, it changes only that line"
+              description: "Optional one-line description (at most #{Tools::MemoryWrite::DESCRIPTION_LIMIT} characters) appended to the managed index line; without content, it changes only that line"
             },
             current_model_only: {
               type: "boolean",

data/lib/samagotchi/version.rb

@@ -1,7 +1,7 @@
 # frozen_string_literal: true
 
 module Samagotchi
-  VERSION = "0.42.0"
+  VERSION = "0.43.0"
   # Sent as the User-Agent of every HTTP request chi makes.
   USER_AGENT = "chi/#{VERSION}"
 end

data/lib/samagotchi/web/app.rb

@@ -20,6 +20,7 @@ require_relative "../session_metrics"
 require_relative "../session_commands"
 require_relative "../steer"
 require_relative "../host_registry"
+require_relative "../llm_context_override"
 require_relative "../model_catalog"
 require_relative "../model_profile"
 require_relative "../sampling_settings"
@@ -458,7 +459,8 @@ module Samagotchi
       end
 
       # Each model's configured sampling as /model words it ("temperature=0.6
-      # (hosts.work)"), on the models that have some: the picker's tooltip.
+      # (hosts.work)"), on the models that have some: the picker's tooltip;
+      # and each model's LLM context (LLMContextStrategy::Explained#summary).
       # A config that can't be read leaves the list as it is.
       def add_sampling(models, registry)
         return if registry.nil? || models.empty?
@@ -466,14 +468,23 @@ module Samagotchi
         settings = ConfigFile.model_settings
         models.each do |model|
           target = registry.resolve(model[:name])
-          summary = SamplingSettings.summary(target, names: registry.lookup_names(model[:name], target: target),
-                                                     models: settings)
+          names = registry.lookup_names(model[:name], target: target)
+          summary = SamplingSettings.summary(target, names: names, models: settings)
           model[:sampling] = summary if summary
+          add_llm_context(model, target, names, settings)
         end
       rescue StandardError
         nil
       end
 
+      # The LLM context +model+ starts under (the start page's llm ctx chip);
+      # one that can't be read leaves this model without, not the others.
+      def add_llm_context(model, target, names, settings)
+        model[:llm_context] = LLMContextStrategy.explain(target, names: names, models: settings).summary(nil)
+      rescue StandardError
+        nil
+      end
+
       # A session's LLM context strategy, apply rule and budget, each with
       # where it came from (LLMContextStrategy::Explained#summary), as its
       # worker would resolve them; nil when they can't be.
@@ -563,6 +574,12 @@ module Samagotchi
         dir, error = scope_dir(body["dir"])
         return error if error
 
+        # llm_context: the session's own LLM context values before its first
+        # turn (the start page's llm ctx chip; chi --llm-context), read
+        # before anything spawns so a bad one spawns nothing.
+        llm_context, error = request_llm_context(body["llm_context"])
+        return error if error
+
         folder = dir ? { working_directory: dir } : {}
         # model: the model the session starts on (the start page's picker,
         # spelled as GET /api/models lists it); blank means the default.
@@ -573,6 +590,7 @@ module Samagotchi
         # until that turn is saved.
         preview = idle ? body["preview"].to_s.strip : ""
         folder[:title] = preview unless preview.empty?
+        folder[:llm_context] = llm_context if llm_context
         begin
           session = @manager.spawn_session(prompt: idle ? nil : prompt.to_s, state_dir: @state_dir, **folder)
         rescue ModelProfile::MissingModel => e
@@ -588,6 +606,25 @@ module Samagotchi
         json_response(201, session_to_json(session).merge(bridge_port: port))
       end
 
+      # The create request's "llm_context" ({strategy:, apply:, budget:},
+      # each a word as /llm-context takes it, "default" for unset) as an
+      # override: [override, nil] (nil when absent or all default), or
+      # [nil, a 400 response] for anything else.
+      def request_llm_context(raw)
+        return [nil, nil] if raw.nil?
+
+        fields = LLMContextOverride::COMMAND_WORDS.invert
+        unless raw.is_a?(Hash) && raw.all? { |key, word| fields.key?(key) && word.is_a?(String) }
+          return [nil, error_response(400, "invalid_llm_context",
+                                      "llm_context is an object of #{fields.keys.join(", ")} words")]
+        end
+
+        override = LLMContextOverride.update(nil, raw.to_h { |key, word| [fields.fetch(key), word] })
+        [override.empty? ? nil : override, nil]
+      rescue ArgumentError => e
+        [nil, error_response(400, "invalid_llm_context", e.message)]
+      end
+
       def handle_show(req, id)
         # The page's light re-reads: never the whole conversation.
         return handle_cards_read(id) if req.params["cards"] == "1"

data/lib/samagotchi/web/public/app.js

@@ -21,7 +21,10 @@ import {
   addContext,
 } from "./data.js";
 import { createContextChips } from "./context_chips.js";
-import { chipHtml as llmContextChipHtml, createLlmContextControl } from "./llm_context_chip.js";
+import {
+  chipHtml as llmContextChipHtml, chipModel as llmContextChipModel, choiceWords, createLlmContextControl, NEW_CHAT,
+  withChoice,
+} from "./llm_context_chip.js";
 import { renderStable } from "./stable_html.js";
 import { createShownTexts } from "./shown_texts.js";
 import { MODEL_KEY } from "./model_pick.js";
@@ -39,7 +42,7 @@ import { parsePresets } from "./annotate_presets.js";
 import { createAnnotateBar } from "./annotate_bar.js";
 import { COPY_SOURCE_ATTR, copySource, copyText, installCopy, showsAnswer } from "./copy.js";
 import { routeChunk } from "./chunk_router.js";
-import { extractCtxPct, savedCtxPct, cardCtxText, generationTokens, lastSpeedText, speedText, tokensTipText } from "./ctx.js";
+import { extractCtxPct, savedCtxPct, cardCtxText, ctxBarText, generationTokens, lastSpeedText, speedText, tokensTipText } from "./ctx.js";
 import { createQuestionCards } from "./question_cards.js";
 import { cardClass, cardInnerHtml, cardPlace, isCard, isNoticeCard, noticeInnerHtml, shownElsewhere, splitSnapshotCards, turnNoticePlace } from "./card.js";
 import { PROMPT_LINE_PX, PROMPT_MIN_PX, gripFloor, promptHeight } from "./composer_size.js";
@@ -86,6 +89,36 @@ const llmContextControl = createLlmContextControl({
   running: () => liveTurn.running,
   run: (line) => runSessionCommand(line),
 });
+// The start page's LLM context chip, beside the model picker: the picked
+// model's values, and a choice for the next new chat only (sent with its
+// create, then dropped; never remembered: a reload or opening a session
+// drops it too, a failed create keeps it).
+const llmContextNewEl = $("#llmContextNew");
+let newChatLlmContext = null;
+const newChatLlmSummary = () => (modelPickEl.hidden ? null : modelPicker.chosenRow()?.llm_context || null);
+const llmContextNewControl = createLlmContextControl({
+  el: llmContextNewEl,
+  escapeHtml,
+  summary: newChatLlmSummary,
+  kind: NEW_CHAT,
+  choice: () => newChatLlmContext,
+  choose: (next) => {
+    newChatLlmContext = next;
+    renderNewChatLlmContext();
+    promptEl.focus();
+  },
+  id: "llmContextNewPopover",
+  anchor: "#llmContextNew",
+});
+// Hidden without the model's values (the list not in, an older server).
+function renderNewChatLlmContext() {
+  const chip = llmContextChipModel(withChoice(newChatLlmSummary(), newChatLlmContext), { kind: NEW_CHAT });
+  llmContextNewEl.hidden = !chip;
+  if (!chip) return llmContextNewControl.close();
+  llmContextNewEl.textContent = chip.text;
+  llmContextNewEl.title = chip.title;
+  llmContextNewEl.classList.toggle("own", chip.own);
+}
 const markdownWarningEl = $("#markdownWarning");
 // What the page renders that a notice may stand in for (the server's
 // Web::Capabilities, next to markdown_warning); a notice it names isn't
@@ -908,7 +941,15 @@ function updateComposerMode() {
     hadSession = !hadSession;
     fitPrompt();
   }
-  if (selected) modelPicker.close();
+  if (selected) {
+    modelPicker.close();
+    llmContextNewControl.close();
+    // The start page's choice was for a new chat: opening a session drops it.
+    if (newChatLlmContext) {
+      newChatLlmContext = null;
+      renderNewChatLlmContext();
+    }
+  }
   const running = liveTurn.running && !!selected;
   cancelBtn.style.display = running ? "" : "none";
   if (running) {
@@ -1121,7 +1162,10 @@ function renderSessionView(data) {
   // next turn (a running turn's replay below streams over it).
   const savedPct = savedCtxPct(data.timing?.context);
   if (savedPct !== null) current.ctxPct = savedPct;
-  if (data.timing?.context?.window_tokens) current.ctxWindow = data.timing.context.window_tokens;
+  if (data.timing?.context?.window_tokens) {
+    current.ctxWindow = data.timing.context.window_tokens;
+    current.ctxWindowSource = data.timing.context.window_source || null;
+  }
   current.firstPreview = s.first_preview || previewForCard(s);
   current.usedMemories = s.used_memory_names || [];
   current.preloadedMemories = s.preloaded_memory_names || [];
@@ -1766,9 +1810,10 @@ function updateInfoBar() {
   if (stopBtn) stopBtn.style.display = canStopSession({ owner: current.owner, streaming: !!streamControl }) ? "" : "none";
   const idShort = selected.slice(0,8);
   const memText = current.usedMemories.length ? current.usedMemories.join(", ") : "—";
-  const ctxText = current.ctxPct !== null ? `ctx ${Math.round(current.ctxPct)}%` : "";
+  const ctxText = ctxBarText(current.ctxPct, current.ctxWindowSource);
   const tokens = current.timing.tokens;
-  const ctxTip = tokensTipText(tokens, undefined, current.timing.memoryIndex);
+  const ctxTip = tokensTipText(tokens, undefined, current.timing.memoryIndex,
+                               { tokens: current.ctxWindow, source: current.ctxWindowSource, pct: current.ctxPct });
   const speedTextNow = lastSpeedText(tokens);
   const avgSpeed = speedText(tokens?.avg_decode_tps, tokens?.tps_source);
   const speedTip = `Decode speed of the last generation${avgSpeed ? `; session average ${avgSpeed}` : ""}`;
@@ -1939,7 +1984,10 @@ const streamHandlers = {
     }
   },
   generation_started: (data) => {
-    if (data?.context_window_tokens) current.ctxWindow = data.context_window_tokens;
+    if (data?.context_window_tokens) {
+      current.ctxWindow = data.context_window_tokens;
+      current.ctxWindowSource = data.context_window_source || null;
+    }
     liveTurn.timing.setNote("");
     turnDom.generationStarted(data || {});
   },
@@ -2226,7 +2274,11 @@ async function handleSendTurn() {
     }
     actionBtn.disabled = true;
     try {
-      const s = await createIdleSession({ dir: scope.dir, model: chosenModel(), preview: prompt });
+      const s = await createIdleSession({ dir: scope.dir, model: chosenModel(), preview: prompt,
+                                          llm_context: choiceWords(newChatLlmContext) });
+      // The choice was for this chat only.
+      newChatLlmContext = null;
+      renderNewChatLlmContext();
       addCreatedSession(s);
       await select(s.id);
     } catch (e) {
@@ -2439,7 +2491,10 @@ const modelPicker = createModelPicker({
   list: $("#modelList"),
   composer: $("#composer"),
   center: $("#center"),
-  onPick: () => promptEl.focus(),
+  onPick: () => {
+    renderNewChatLlmContext();
+    promptEl.focus();
+  },
 });
 async function loadModels() {
   let payload;
@@ -2453,6 +2508,7 @@ async function loadModels() {
   if (!modelPicker.setModels(payload, stored)) return;
   modelPickEl.title = payload.warning ? `The model the new chat starts on · ${payload.warning}` : "The model the new chat starts on";
   modelPickEl.hidden = false;
+  renderNewChatLlmContext();
 }
 // The picker's choice for a create request; nothing (the server's default)
 // while the list is not in.

data/lib/samagotchi/web/public/ctx.js

@@ -91,14 +91,39 @@ export function memoryIndexText(memoryIndex) {
   return `~${countK(total)} tokens in this session's prompt (${scopes.map(([scope, tokens]) => `${scope} ${countK(tokens)}`).join(", ")})`;
 }
 
+// The info bar's ctx: "ctx 12%", "ctx ~12%" when the window is chi's
+// default (no server, model list or setting gave one, so the percent is of
+// a guess); "" without a pct.
+export function ctxBarText(pct, windowSource = null) {
+  if (num(pct) === null) return "";
+  return `ctx ${windowSource === "default" ? "~" : ""}${Math.round(pct)}%`;
+}
+
+// The ctx tooltip's window line ({tokens:, source:, pct:} the session's
+// window, where it came from and the meter's value): "context: ~41.0k of
+// 128.0k tokens (server)", as /stats names the source; "" without a window.
+export function ctxWindowText(window) {
+  const tokens = num(window?.tokens);
+  if (!tokens || tokens <= 0) return "";
+  const pct = num(window.pct);
+  const used = pct === null ? "" : `~${countK((pct / 100) * tokens)} of `;
+  const source = window.source === "default"
+    ? "chi's default, a guess: set window_tokens for this model"
+    : window.source;
+  return `context: ${used}${countK(tokens)} tokens${source ? ` (${source})` : ""}`;
+}
+
 // The ctx tooltip (info bar and card): +title+, then the session's tokens
 // (in summed over every request, cached with its share, out with the
 // reasoning part) and its cost, when it has any, and the memory indexes
-// its prompt holds (+memoryIndex+, the snapshot's memory_index). A
+// its prompt holds (+memoryIndex+, the snapshot's memory_index); with
+// +window+ (ctxWindowText's) the window line under the title. A
 // delegated session keeps its own metrics, so these are this session's only.
-export function tokensTipText(tokens, title = "Context used after the last turn", memoryIndex = null) {
+export function tokensTipText(tokens, title = "Context used after the last turn", memoryIndex = null, window = null) {
   const memory = memoryIndexText(memoryIndex);
   const memoryLine = memory ? `\nmemory index: ${memory}` : "";
+  const windowText = ctxWindowText(window);
+  if (windowText) title = `${title}\n${windowText}`;
   const prompt = num(tokens?.prompt_sum) || 0;
   const completion = num(tokens?.completion_sum) || 0;
   if (!prompt && !completion) return `${title}${memoryLine}`;

data/lib/samagotchi/web/public/data.js

@@ -30,9 +30,12 @@ export function listSessions(sort, order, { dir, ...opts } = {}) {
 // as a turn (with its images, once they are uploaded). dir: the folder the
 // chat starts in; model: the model it starts on (as GET /api/models spells
 // it); either left out means the server's default. preview: that message,
-// which names the session until its turn is saved.
-export function createIdleSession({ dir, model, preview, ...opts } = {}) {
-  const body = { idle: true, ...(dir ? { dir } : {}), ...(model ? { model } : {}), ...(preview ? { preview } : {}) };
+// which names the session until its turn is saved. llm_context: the
+// session's own LLM context words ({strategy, apply, budget}), when the
+// start page's chip set some.
+export function createIdleSession({ dir, model, preview, llm_context: llmContext, ...opts } = {}) {
+  const body = { idle: true, ...(dir ? { dir } : {}), ...(model ? { model } : {}), ...(preview ? { preview } : {}),
+                 ...(llmContext ? { llm_context: llmContext } : {}) };
   return api("/api/sessions", { method: "POST", body: JSON.stringify(body), ...opts });
 }
 

data/lib/samagotchi/web/public/index.html

@@ -491,6 +491,13 @@ body.has-session #composer .actions-col{flex-wrap:nowrap;flex-shrink:0}
 .model-pick::after{content:"";position:absolute;right:9px;top:50%;width:5px;height:5px;border-right:1.5px solid currentColor;border-bottom:1.5px solid currentColor;transform:translateY(-70%) rotate(45deg);pointer-events:none;transition:transform .12s ease}
 .model-pick[aria-expanded="true"]::after{transform:translateY(-20%) rotate(225deg)}
 @media(max-width:600px){#composer .actions-col .model-pick{max-width:60%}}
+/* The new chat's LLM context chip, after the picker: as quiet as it; accent once a choice is set (the info bar chip's .own). */
+#composer .actions-col .llmctx-new{order:0;flex:0 1 auto;min-width:0;max-width:min(40%,260px);height:24px;min-height:24px;line-height:22px;padding:0 6px;border-radius:6px;border:1px solid transparent;background:transparent;color:var(--text-muted);font-size:11px;font-weight:400;overflow:hidden;text-overflow:ellipsis}
+#composer .actions-col .llmctx-new:hover{filter:none;color:var(--text);border-color:var(--accent-border);background:var(--accent-bg)}
+#composer .actions-col .llmctx-new.own{color:var(--accent);border-color:var(--accent-border)}
+#composer .actions-col .llmctx-new[hidden],body.has-session #composer .actions-col .llmctx-new{display:none}
+/* Phones: picker · chip · Start on one row (the key hint is gone there); names cut short rather than wrapping, the chip first. */
+@media(max-width:600px){#composer .actions-col:has(.llmctx-new:not([hidden])){flex-wrap:nowrap}#composer .actions-col .model-pick{flex:0 1 auto;min-width:0}#composer .actions-col .llmctx-new{max-width:45%;flex-shrink:4}}
 /* The model search: opaque like #commandList (a list over the conversation
    must stay readable), above the composer, or below it (.down) when the
    start page leaves no room above, or over the composer (.over) when
@@ -1005,6 +1012,7 @@ a{color:var(--accent)}
         </div>
         <div class="actions-col">
           <button type="button" id="modelPick" class="model-pick" title="The model the new chat starts on" aria-label="Model for the new chat" aria-haspopup="listbox" aria-expanded="false" hidden><span class="model-pick-name"></span></button>
+          <button type="button" id="llmContextNew" class="llmctx-new" aria-haspopup="dialog" hidden></button>
           <div id="composeHint">Enter to start · Shift+Enter for a newline</div>
           <button id="cancelBtn" title="Cancel the running turn, whoever sent it" style="display:none">Cancel</button>
           <div class="split">