samagotchi 0.51.0 → 0.52.0

data/lib/samagotchi/web/app.rb

@@ -10,6 +10,7 @@ require "rack"
 require "rack/request"
 
 require_relative "../client_id"
+require_relative "../delivery"
 require_relative "../answer_tail"
 require_relative "../bridge_client"
 require_relative "../bridge/bounded_queue"
@@ -23,6 +24,7 @@ require_relative "../host_registry"
 require_relative "../llm_context_override"
 require_relative "../model_catalog"
 require_relative "../model_profile"
+require_relative "../thinking"
 require_relative "../sampling_settings"
 require_relative "../served_model"
 require_relative "../project_scope"
@@ -476,6 +478,7 @@ module Samagotchi
           summary = SamplingSettings.summary(target, names: names, models: settings)
           model[:sampling] = summary if summary
           add_llm_context(model, target, names, settings)
+          add_thinking(model, target, names, settings)
         end
       rescue StandardError
         nil
@@ -489,6 +492,14 @@ module Samagotchi
         nil
       end
 
+      # The thinking level +model+ starts at, with where it came from (the
+      # start page's think chip; Thinking::Explained#summary, no session).
+      def add_thinking(model, target, names, settings)
+        model[:thinking] = Thinking.explain(target, names: names, models: settings).summary
+      rescue StandardError
+        nil
+      end
+
       # The served model a session's last generation reported (analytics.json,
       # SessionMetrics#persist), for a session no worker runs: shown only
       # while the name it answered is still the session's model, as
@@ -525,6 +536,19 @@ module Samagotchi
         LLMContextStrategy.for_session(session, registry: host_registry)
       end
 
+      # A session's thinking level with where it came from and its own
+      # (Thinking::Explained#summary), as its worker would resolve it; nil
+      # when it can't be. Config and the registry only: no network.
+      def thinking_for(session)
+        registry = host_registry
+        name = session.model_name.to_s.strip.empty? ? ModelProfile.required_model_name(nil) : session.model_name
+        target = registry.resolve(name)
+        names = registry.lookup_names(session.model_typed || name, resolved: name, target: target)
+        Thinking.explain(target, names: names, session: session.thinking).summary
+      rescue StandardError
+        nil
+      end
+
       # Built again when config.yml's hosts: changed (compared as read: the
       # YAML read is mtime-cached), so the model picker lists a host added
       # while chi web runs. An injected registry (specs) stays.
@@ -607,6 +631,11 @@ module Samagotchi
         llm_context, error = request_llm_context(body["llm_context"])
         return error if error
 
+        # thinking: the session's own thinking level (the start page's think
+        # chip; chi --thinking), checked the same way.
+        thinking, error = request_thinking(body["thinking"])
+        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.
@@ -617,7 +646,8 @@ 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
+        setup = SessionSetup.new(llm_context: llm_context, thinking: thinking)
+        folder[:setup] = setup unless setup.empty?
         begin
           session = @manager.spawn_session(prompt: idle ? nil : prompt.to_s, state_dir: @state_dir, **folder)
         rescue ModelProfile::MissingModel => e
@@ -635,8 +665,8 @@ module Samagotchi
 
       # POST /api/sessions with continues: <id, prefix or last:id>: the next
       # link of that session's chain (SessionManager.continue_session), in its
-      # folder, on its model and with its llm_context, so dir, model and
-      # llm_context are refused beside it. 409 continued (with next_id, the
+      # folder, on its model and with its llm_context and thinking level, so
+      # dir, model, llm_context and thinking are refused beside it. 409 continued (with next_id, the
       # link that continues it already: the page opens that one),
       # open_children (ids), folder_gone, busy, scratch or owned_by_tui.
       def handle_continue(body, prompt, idle)
@@ -644,9 +674,9 @@ module Samagotchi
         unless given.is_a?(String) && !given.strip.empty?
           return error_response(400, "invalid_continues", "continues is a session id")
         end
-        if (fields = %w[dir model model_name llm_context].select { |key| body.key?(key) }).any?
+        if (fields = %w[dir model model_name llm_context thinking].select { |key| body.key?(key) }).any?
           return error_response(400, "invalid_continues",
-                                "continues takes the previous session's folder, model and llm_context; " \
+                                "continues takes the previous session's folder, model, llm_context and thinking; " \
                                 "leave out #{fields.join(", ")}")
         end
 
@@ -701,6 +731,20 @@ module Samagotchi
         [nil, error_response(400, "invalid_llm_context", e.message)]
       end
 
+      # The create request's "thinking" (a level word as /thinking takes it,
+      # "default" for none) as the session's own level: [level, nil] (nil
+      # when absent or default), or [nil, a 400 response] for anything else.
+      def request_thinking(raw)
+        return [nil, nil] if raw.nil?
+
+        word = raw.is_a?(String) ? raw.strip.downcase : nil
+        unless word && Thinking::LEVELS.map(&:to_s).include?(word)
+          return [nil, error_response(400, "invalid_thinking", "thinking is one of #{Thinking::LEVELS.join(", ")}")]
+        end
+
+        [Thinking.session_level(word), nil]
+      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"
@@ -750,6 +794,8 @@ module Samagotchi
         # The LLM context strategy the next turn runs under (the info bar's
         # chip): the worker's, else worked out here from the session file.
         session_json = session_json.merge(llm_context: (snapshot && snapshot["llm_context"]) || llm_context_for(session))
+        # The thinking level the next turn runs at (the think chip), the same way.
+        session_json = session_json.merge(thinking: (snapshot && snapshot["thinking"]) || thinking_for(session))
         raw_messages = turn_snapshot ? turn_snapshot["messages"] : session.messages
         timing = timing_payload(id, live_metrics: snapshot && snapshot["metrics"], persisted: persisted)
         json_response(200, {
@@ -1066,7 +1112,9 @@ module Samagotchi
         return error_response(400, "bad_images", images) if images.is_a?(String)
 
         delivery = body["delivery"] || body[:delivery]
-        return error_response(400, "bad_delivery", "delivery must be a string") unless delivery.nil? || delivery.is_a?(String)
+        if (refused = Delivery.refusal(delivery))
+          return error_response(400, refused[:error], refused[:detail])
+        end
 
         result = SessionManager.deliver_turn(id, prompt: prompt.to_s, client_id: client_id, images: images,
                                                  delivery: delivery, state_dir: @state_dir, manager: @manager,

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

@@ -26,14 +26,18 @@ import {
   chipHtml as llmContextChipHtml, chipModel as llmContextChipModel, choiceWords, createLlmContextControl, NEW_CHAT,
   withChoice,
 } from "./llm_context_chip.js";
+import {
+  chipHtml as thinkingChipHtml, chipModel as thinkingChipModel, createThinkingControl,
+  withChoice as withThinkingChoice,
+} from "./thinking_chip.js";
 import { renderStable } from "./stable_html.js";
 import { createShownTexts } from "./shown_texts.js";
 import { MODEL_KEY } from "./model_pick.js";
 import { emptySessionState } from "./session_state.js";
 import { createModelPicker } from "./model_picker.js";
 import { chipName, chipsHtml, imageFiles, restoredChips, thumbsHtml, turnRefs, turnText } from "./images.js";
-import { escapeHtml, memChipText, memChipTitle, noteHtml, normalize, steerRowHtml, ownerBadge, previewForCard, relativeTime, failedTurnText, modelLabel, deleteConfirmText, recapLabel, canStopSession, goneSessionNotice, withLiveStatus, withoutWorker, recapPlace, delegatedBy, childrenSummary, promptNotesChip } from "./format.js";
-import { createTurnView, turnHistoryHtml } from "./turn_view.js";
+import { escapeHtml, memChipText, memChipTitle, noteHtml, normalize, steerRowHtml, ownerBadge, previewForCard, firstPreview, relativeTime, failedTurnText, modelLabel, deleteConfirmText, recapLabel, canStopSession, goneSessionNotice, withLiveStatus, withoutWorker, recapPlace, delegatedBy, childrenSummary, promptNotesChip } from "./format.js";
+import { createTurnView, turnHistoryNodes } from "./turn_view.js";
 import { BADGES, emptyAnswerHtml, htmlElement, stateBadgeHtml, userBubbleHtml } from "./turn_html.js";
 import { createStageView } from "./stage_view.js";
 import { createFamilyPop, familyChipHtml, familyRowsHtml } from "./families_view.js";
@@ -58,12 +62,13 @@ import { applySessionEvent, batchCounts, batchTargets, batchToastText, families,
 import { createNotifications } from "./notifications.js";
 import { tabTitle, topBarTitle } from "./session_title.js";
 import { restartConfirmText, versionNotice, workerBadge } from "./update_notice.js";
+import { anchorInsertIndex, commandAnchor, isShellCommandLine, keepsCommandBubble } from "./kept_commands.js";
 import {
-  appendAboveLiveTiming, cancelLineText, elapsedSince, formatDuration, mergeTiming, normalizeTiming, STOPPED_LINE_CLASS,
+  appendAboveLiveTiming, cancelLineText, formatDuration, mergeTiming, normalizeTiming, sessionDurationNow, STOPPED_LINE_CLASS,
 } from "./timing.js";
 import { createLiveTurn } from "./live_turn.js";
 import { DONE_MS, applyInitEvent, dropDone, initRowHtml, newInitTasks, seedInitTasks } from "./init_row.js";
-import { clientLabel, commandView, isReportClient, displayWait, dropEarlyRestores, emptyRetryLine, historyHasCommandBubble, hookNoticeLabel, initWaitLine, isOwn, llmContextHover, llmContextLine, sessionCommandLine, keepEarlyRestore, newClientId, noticeLine, promptOps, reminderText, restoreAction, composerAfterAck, restoreOnAck, replaysDrawnTurn, retryStatusLine, snapshotEvents, steerCutLine, startPageReplyOrHint, turnOutput, unknownCommandHint, webLocalReply, workerGoneText } from "./turn_events.js";
+import { clientLabel, commandView, isReportClient, displayWait, dropEarlyRestores, emptyRetryLine, historyHasCommandBubble, hookNoticeLabel, initWaitLine, isOwn, llmContextHover, llmContextLine, sessionCommandLine, keepEarlyRestore, newClientId, noticeLine, promptOps, reminderText, restoreAction, composerAfterAck, restoreOnAck, replaysDrawnTurn, retryStatusLine, snapshotEvents, steerChoiceRoute, steerCutLine, startPageReplyOrHint, turnOutput, unknownCommandHint, webLocalReply, workerGoneText } from "./turn_events.js";
 
 const $ = (s) => document.querySelector(s);
 const topStripEl = $("#topStrip");
@@ -117,6 +122,43 @@ const llmContextNewControl = createLlmContextControl({
   id: "llmContextNewPopover",
   anchor: "#llmContextNew",
 });
+// The open session's thinking chip, beside the llm ctx one: /thinking runs
+// in the session's worker.
+const thinkingControl = createThinkingControl({
+  el: infoTextEl,
+  escapeHtml,
+  summary: () => (selected ? current.thinking : null),
+  session: () => selected,
+  running: () => liveTurn.running,
+  run: (line) => runSessionCommand(line),
+});
+// The start page's thinking chip: the picked model's level, and a level
+// for the next new chat only (sent with its create, then dropped).
+const thinkingNewEl = $("#thinkingNew");
+let newChatThinking = null;
+const newChatThinkingSummary = () => (modelPickEl.hidden ? null : modelPicker.chosenRow()?.thinking || null);
+const thinkingNewControl = createThinkingControl({
+  el: thinkingNewEl,
+  escapeHtml,
+  summary: newChatThinkingSummary,
+  kind: NEW_CHAT,
+  choice: () => newChatThinking,
+  choose: (next) => {
+    newChatThinking = next;
+    renderNewChatThinking();
+    promptEl.focus();
+  },
+  id: "thinkingNewPopover",
+  anchor: "#thinkingNew",
+});
+function renderNewChatThinking() {
+  const chip = thinkingChipModel(withThinkingChoice(newChatThinkingSummary(), newChatThinking), { kind: NEW_CHAT });
+  thinkingNewEl.hidden = !chip;
+  if (!chip) return thinkingNewControl.close();
+  thinkingNewEl.textContent = chip.text;
+  thinkingNewEl.title = chip.title;
+  thinkingNewEl.classList.toggle("own", chip.own);
+}
 // Hidden without the model's values (the list not in, an older server).
 function renderNewChatLlmContext() {
   const chip = llmContextChipModel(withChoice(newChatLlmSummary(), newChatLlmContext), { kind: NEW_CHAT });
@@ -1057,6 +1099,11 @@ function updateComposerMode() {
       newChatLlmContext = null;
       renderNewChatLlmContext();
     }
+    thinkingNewControl.close();
+    if (newChatThinking) {
+      newChatThinking = null;
+      renderNewChatThinking();
+    }
   }
   const running = liveTurn.running && !!selected;
   cancelBtn.style.display = running ? "" : "none";
@@ -1233,8 +1280,9 @@ async function resync({ endText = null } = {}) {
       return;
     }
     if (id !== selected) return;
+    const kept = keptCommandBubbles(id);
     resetTurnView();
-    renderSessionView(data);
+    renderSessionView(data, { kept });
     // Preview is read-only: a worker is woken only by POST /turn, so attach
     // only when a live bridge exists.
     selectSeq = typeof data.last_event_seq === "number" ? (data.last_event_id || data.last_event_seq) : null;
@@ -1249,10 +1297,45 @@ async function resync({ endText = null } = {}) {
   }
 }
 
+// The prompt and command bubbles in the history, in order, by kind
+// (kept_commands.js): a prompt, a `!cmd`'s (saved, so redrawn) or another.
+function historyAnchors() {
+  const els = [...historyEl.querySelectorAll(".bubble.user, .bubble.command")];
+  const kinds = els.map((el) => {
+    if (el.classList.contains("user")) return "user";
+    const line = el.querySelector(":scope > .command-line")?.textContent;
+    return !el.classList.contains("failed") && isShellCommandLine(line) ? "shell" : "other";
+  });
+  return { els, kinds };
+}
+
+// This session's own command bubbles (/help, /model …: drawn by this page,
+// never saved) with their anchors, before a resync redraws the history.
+// In-flight ones (busy, queued) are left out: the snapshot draws them again.
+function keptCommandBubbles(id) {
+  const { els, kinds } = historyAnchors();
+  return els
+    .map((el, i) => ({ el, anchor: commandAnchor(kinds, i) }))
+    .filter(({ el }) => el.dataset.keptFor === id && !el.classList.contains("busy") && !el.classList.contains("queued"))
+    .filter(({ el }) => keepsCommandBubble(el.querySelector(":scope > .command-line")?.textContent));
+}
+
+// The kept bubbles back into the redrawn history, each after its turn.
+function restoreCommandBubbles(kept) {
+  if (!kept.length) return;
+  historyEl.querySelector(".hint")?.remove();
+  kept.forEach(({ el, anchor }) => {
+    const { els, kinds } = historyAnchors();
+    const at = anchorInsertIndex(kinds, anchor);
+    if (at < els.length) els[at].before(el);
+    else historyEl.appendChild(el);
+  });
+}
+
 // A session view: its messages, then the turn in progress and the queued
 // prompts (a live worker's snapshot) replayed through the stream handlers,
 // then a pending question.
-function renderSessionView(data) {
+function renderSessionView(data, { kept = [] } = {}) {
   const s = data.session;
   setMarkdownWarning(data.markdown_warning);
   setCapabilities(data.capabilities);
@@ -1279,20 +1362,23 @@ function renderSessionView(data) {
   current.promptNotes = Array.isArray(s.prompt_notes) ? s.prompt_notes : [];
   current.llmContext = s.llm_context || null;
   llmContextControl.refresh();
+  current.thinking = s.thinking || null;
+  thinkingControl.refresh();
   current.parentId = s.parent_id || null;
   current.continues = s.continues || null;
   current.commands = Array.isArray(data.commands) ? data.commands : [];
   if (data.messages && data.messages.length) {
     const first = data.messages.find((m) => m.role === "user");
-    if (first && first.content) current.firstPreview = normalize(first.content).slice(0,80) + (normalize(first.content).length>80?"…":"");
+    if (first && first.content) current.firstPreview = firstPreview(first.content);
     renderHistory(data.messages);
   } else {
     renderHistory(data.history || []);
     if (!current.firstPreview && data.history && data.history.length) {
-      const txt = normalize(String(data.history[0]||"")).slice(0,80);
-      if(txt) current.firstPreview = txt;
+      const txt = firstPreview(data.history[0]);
+      if (txt) current.firstPreview = txt;
     }
   }
+  restoreCommandBubbles(kept);
   // A failed turn's prompt went back to the user: shown after the
   // conversation, marked failed, as it was live (prompt_restored).
   if (data.failed_turn) {
@@ -1352,7 +1438,7 @@ function renderHistory(items) {
     return;
   }
   if (typeof items[0] === "object" && items[0].role) {
-    historyEl.innerHTML = turnHistoryHtml(items, current.timing, { thumbs: (images) => thumbsHtml(selected, images) });
+    historyEl.replaceChildren(turnHistoryNodes(items, current.timing, { thumbs: (images) => thumbsHtml(selected, images) }));
   } else {
     historyEl.innerHTML = items.map((c) => `<div class="bubble output">${escapeHtml(c)}</div>`).join("");
   }
@@ -1965,7 +2051,7 @@ function updateInfoBar() {
   const attachCmd = `chi --attach ${selected}`;
   const copied = Date.now() - attachCopiedAt < 2000;
   const copyHtml = `<button type="button" class="copy-attach" title="Copy &quot;${escapeHtml(attachCmd)}&quot; to attach a terminal">${copied ? "copied ✓" : "copy chi --attach"}</button>`;
-  const metaHtml = `<div class="meta"><span class="id" title="${escapeHtml(selected)}">${escapeHtml(idShort)}</span>${copyHtml}<span class="${statusClass}">${escapeHtml(current.status||"")}</span><span class="model${label.mismatch ? " served-mismatch" : ""}" title="${escapeHtml(label.title)}">${escapeHtml(label.text)}</span>${chainHtml}${parentHtml}${kidsHtml}${llmContextChipHtml(current.llmContext, escapeHtml)}${notesHtml}${current.usedMemories.length ? `<span class="model mem-chip" title="${escapeHtml(memChipTitle(current.usedMemories))}">${escapeHtml(memChipText(current.usedMemories))}</span>` : ""}<span class="dir" title="${escapeHtml(current.dir||"")}"><bdi>${escapeHtml(current.dir||"")}</bdi></span>${ctxText ? `<span class="model ctx" title="${escapeHtml(ctxTip)}">${escapeHtml(ctxText)}</span>` : ""}${speedTextNow ? `<span class="model speed" title="${escapeHtml(speedTip)}">${escapeHtml(speedTextNow)}</span>` : ""}${timingText ? `<span class="model session-time"></span>` : ""}${workerBadgeHtml()}</div>`;
+  const metaHtml = `<div class="meta"><span class="id" title="${escapeHtml(selected)}">${escapeHtml(idShort)}</span>${copyHtml}<span class="${statusClass}">${escapeHtml(current.status||"")}</span><span class="model${label.mismatch ? " served-mismatch" : ""}" title="${escapeHtml(label.title)}">${escapeHtml(label.text)}</span>${chainHtml}${parentHtml}${kidsHtml}${llmContextChipHtml(current.llmContext, escapeHtml)}${thinkingChipHtml(current.thinking, escapeHtml)}${notesHtml}${current.usedMemories.length ? `<span class="model mem-chip" title="${escapeHtml(memChipTitle(current.usedMemories))}">${escapeHtml(memChipText(current.usedMemories))}</span>` : ""}<span class="dir" title="${escapeHtml(current.dir||"")}"><bdi>${escapeHtml(current.dir||"")}</bdi></span>${ctxText ? `<span class="model ctx" title="${escapeHtml(ctxTip)}">${escapeHtml(ctxText)}</span>` : ""}${speedTextNow ? `<span class="model speed" title="${escapeHtml(speedTip)}">${escapeHtml(speedTextNow)}</span>` : ""}${timingText ? `<span class="model session-time"></span>` : ""}${workerBadgeHtml()}</div>`;
   // The session clock ticks every second while a turn runs: only its text
   // changes, the bar's other nodes (a tooltip, a button mid-press) stay.
   renderStable(infoTextEl, metaHtml, { ".session-time": timingText });
@@ -2020,10 +2106,7 @@ function memoriesTooltip(used, preloaded, muted) {
 }
 
 function currentSessionDuration() {
-  if (liveTurn.running && current.timing.startedAt) {
-    return elapsedSince(current.timing.startedAt);
-  }
-  return current.timing.sessionDurationMs;
+  return sessionDurationNow(current.timing, liveTurn.running);
 }
 
 // The stage takes a running turn's rows (its extras); else the history.
@@ -2281,6 +2364,12 @@ const streamHandlers = {
       updateInfoBar();
       llmContextControl.refresh();
     }
+    // /thinking (or /model) changed the level the next turn runs at.
+    if (data.thinking) {
+      current.thinking = data.thinking;
+      updateInfoBar();
+      thinkingControl.refresh();
+    }
     if (view.resync) await resync();
     // A card's action that isn't anytime (the step-limit question's
     // /continue) has no bubble: nothing to remove. An anytime or queued
@@ -2433,10 +2522,27 @@ function handleCreate() {
 
 async function handleSendTurn() {
   if (current.owner === "tui") return;
-  const prompt = turnText(promptEl.value, chips);
+  let prompt = turnText(promptEl.value, chips);
   if (!prompt) return;
   // An image-only send's text is its placeholders: not for ↑.
   const typed = promptEl.value.trim() !== "";
+  // /cut TEXT and /queue TEXT (the terminals' words for a delivery): the
+  // text as a cut, or the page's own answer (/queue, the bare word).
+  let delivery = null;
+  const choice = steerChoiceRoute(prompt, { running: liveTurn.running && !!selected });
+  if (choice?.reply) {
+    if (selected) addCommandBubble({ label: null, line: prompt, text: choice.reply });
+    else showNotice(choice.reply);
+    promptEl.value = "";
+    fitPrompt();
+    return;
+  }
+  // The composer holds the line as typed until the ack (composerAfterAck).
+  const line = prompt;
+  if (choice) {
+    prompt = choice.send;
+    delivery = choice.delivery || null;
+  }
   if (!selected) {
     // The start page's own reply (a web-local command) or the hint for a
     // typo: neither needs a session, so none is made for it (2.29).
@@ -2450,10 +2556,13 @@ async function handleSendTurn() {
     actionBtn.disabled = true;
     try {
       const s = await createIdleSession({ dir: scope.dir, model: chosenModel(), preview: prompt,
-                                          llm_context: choiceWords(newChatLlmContext) });
+                                          llm_context: choiceWords(newChatLlmContext),
+                                          thinking: newChatThinking || undefined });
       // The choice was for this chat only.
       newChatLlmContext = null;
       renderNewChatLlmContext();
+      newChatThinking = null;
+      renderNewChatThinking();
       addCreatedSession(s);
       await select(s.id);
     } catch (e) {
@@ -2485,7 +2594,7 @@ async function handleSendTurn() {
   const echo = makeUserBubble(prompt, { state: "queued", own: true, images: sending });
   if (!turnDom.holds?.()) historyEl.scrollTop = historyEl.scrollHeight;
   if (!current.firstPreview || current.firstPreview==="—") {
-    current.firstPreview = normalize(prompt).slice(0,80) + (normalize(prompt).length>80?"…":"");
+    current.firstPreview = firstPreview(prompt);
     updateInfoBar();
   }
   actionBtn.disabled = true;
@@ -2493,7 +2602,7 @@ async function handleSendTurn() {
     for (const chip of sending) {
       if (!chip.ref) chip.ref = await uploadImage(selected, chip.file, chip.name);
     }
-    const ack = await sendTurn(selected, prompt, { clientId, images: turnRefs(sending), history: typed });
+    const ack = await sendTurn(selected, prompt, { clientId, images: turnRefs(sending), history: typed, delivery });
     if (ack?.command_id) {
       // The worker knew it as a command (its list was newer than ours):
       // its command_ran shows it, not a prompt bubble.
@@ -2509,7 +2618,7 @@ async function handleSendTurn() {
     // (and images) go back into the composer, without replacing text typed
     // since (4.13). The composer still holds the text just sent: it goes.
     const restored = restoreOnAck(ack?.enqueued_id, earlyRestores, { myId: clientId, sentIds });
-    promptEl.value = composerAfterAck({ sent: prompt, current: promptEl.value, refill: restored?.refill || "" });
+    promptEl.value = composerAfterAck({ sent: line, current: promptEl.value, refill: restored?.refill || "" });
     fitPrompt();
     clearChips();
     if (restored?.images?.length) setChips(restoredChips(selected, restored.images));
@@ -2589,6 +2698,8 @@ async function runSessionCommand(line) {
 function addCommandBubble(view) {
   removeHintIfEmpty();
   const div = document.createElement("div");
+  // Page-only: kept across a resync of this session (keptCommandBubbles).
+  if (selected) div.dataset.keptFor = selected;
   drawCommandBubble(div, view);
   const follow = isNearBottom();
   appendToHistory(div);
@@ -2667,6 +2778,7 @@ const modelPicker = createModelPicker({
   center: $("#center"),
   onPick: () => {
     renderNewChatLlmContext();
+    renderNewChatThinking();
     promptEl.focus();
   },
 });
@@ -2683,6 +2795,7 @@ async function loadModels() {
   modelPickEl.title = payload.warning ? `The model the new chat starts on · ${payload.warning}` : "The model the new chat starts on";
   modelPickEl.hidden = false;
   renderNewChatLlmContext();
+  renderNewChatThinking();
 }
 // The picker's choice for a create request; nothing (the server's default)
 // while the list is not in.

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

@@ -35,10 +35,11 @@ export function listSessions(sort, order, { dir, ...opts } = {}) {
 // it); either left out means the server's default. preview: that message,
 // 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 } = {}) {
+// start page's chip set some; thinking: its own level ("low"), when the
+// start page's think chip set one.
+export function createIdleSession({ dir, model, preview, llm_context: llmContext, thinking, ...opts } = {}) {
   const body = { idle: true, ...(dir ? { dir } : {}), ...(model ? { model } : {}), ...(preview ? { preview } : {}),
-                 ...(llmContext ? { llm_context: llmContext } : {}) };
+                 ...(llmContext ? { llm_context: llmContext } : {}), ...(thinking ? { thinking } : {}) };
   return api("/api/sessions", { method: "POST", body: JSON.stringify(body), ...opts });
 }
 
@@ -90,9 +91,12 @@ export function getSession(id, { tail, recent, turnId, timing, parts, cards, ...
 // opts.images: refs ({file, name}) to images uploaded into the session.
 // history: false keeps the prompt out of the shared prompt history (an
 // image-only send, whose text is only its [image: …] placeholders).
-export function sendTurn(id, prompt, { clientId, images, history = true, ...opts } = {}) {
+// delivery: what the message does to a running turn ("cut": /cut TEXT);
+// none = the server's default (the next step).
+export function sendTurn(id, prompt, { clientId, images, history = true, delivery = null, ...opts } = {}) {
   const body = clientId ? { prompt, client_id: clientId } : { prompt };
   if (images?.length) body.images = images;
+  if (delivery) body.delivery = delivery;
   if (history === false) body.history = false;
   return api(`/api/sessions/${id}/turn`, {
     method: "POST",

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

@@ -94,13 +94,17 @@ export function previewOf(lastPrompt) {
   return normalized;
 }
 
+// A session's first message as the info bar's title: whitespace collapsed,
+// cut at FIRST_PREVIEW_CHARS with an ellipsis ("" for none).
+export function firstPreview(text) {
+  const n = normalize(text);
+  return n.length > FIRST_PREVIEW_CHARS ? `${n.slice(0, FIRST_PREVIEW_CHARS)}\u2026` : n;
+}
+
 export function previewForCard(session) {
   // Unified card preview: first message (80) else previewOf(last_prompt)
   const raw = session.first_preview || session.last_prompt || "";
-  const n = normalize(raw);
-  if (!n) return "\u2014";
-  if (n.length > FIRST_PREVIEW_CHARS) return `${n.slice(0, FIRST_PREVIEW_CHARS)}\u2026`;
-  return n;
+  return firstPreview(raw) || "\u2014";
 }
 
 const DELETE_PREVIEW_CHARS = 40;

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

@@ -502,8 +502,10 @@ body.has-session #composer .actions-col{flex-wrap:nowrap;flex-shrink:0}
 #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}
+/* The new chat's thinking chip, after it: shorter text, so a smaller share of the row. */
+#composer .actions-col .think-new{max-width:min(25%,140px)}
 /* 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}}
+@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}#composer .actions-col .think-new{max-width:30%}}
 /* 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
@@ -694,10 +696,10 @@ body.stage-view #dock,body.stage-view #dock:focus-within{background:transparent;
 #infoBar #infoArchiveBtn:hover{color:var(--text);background:var(--bg-hover)}
 #infoBar #infoDeleteBtn{background:transparent;color:var(--text-muted);border:1px solid transparent}
 #infoBar #infoDeleteBtn:hover{color:var(--on-solid);background:var(--danger-bg);border-color:var(--danger-bg)}
-/* The LLM context chip (llm_context_chip.js): a model-like chip that opens the strategy popover; accent when the session set its own. */
-#infoBar .meta .llmctx-chip{padding:2px 6px;border-radius:6px;font-size:11px;font-weight:400;border:1px solid transparent;cursor:pointer;font-family:inherit;white-space:nowrap}
-#infoBar .meta .llmctx-chip:hover{color:var(--text);border-color:var(--accent-border);background:var(--accent-bg)}
-#infoBar .meta .llmctx-chip.own{color:var(--accent);border-color:var(--accent-border)}
+/* The LLM context and thinking chips (llm_context_chip.js, thinking_chip.js): model-like chips that open their popovers; accent when the session set its own. */
+#infoBar .meta .llmctx-chip,#infoBar .meta .think-chip{padding:2px 6px;border-radius:6px;font-size:11px;font-weight:400;border:1px solid transparent;cursor:pointer;font-family:inherit;white-space:nowrap}
+#infoBar .meta .llmctx-chip:hover,#infoBar .meta .think-chip:hover{color:var(--text);border-color:var(--accent-border);background:var(--accent-bg)}
+#infoBar .meta .llmctx-chip.own,#infoBar .meta .think-chip.own{color:var(--accent);border-color:var(--accent-border)}
 .llmctx-form{display:flex;flex-direction:column;gap:8px;margin:0}
 .llmctx-form .ctx-pop-head strong{font-family:inherit}
 .llmctx-field{display:flex;align-items:center;gap:8px}
@@ -1046,6 +1048,7 @@ a{color:var(--accent)}
         <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>
+          <button type="button" id="thinkingNew" class="llmctx-new think-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">Stop turn</button>
           <div class="split">

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

@@ -0,0 +1,52 @@
+// Command bubbles (/help, /model, /stats, /llm-context …) live only in the
+// page: the session saves just a `!cmd`'s output (a message, redrawn as its
+// command bubble). A resync (a `!cmd`, a `!rollback`, a dropped stream)
+// redraws the history from the server, so the page keeps its own command
+// bubbles across it: each one's anchor is the prompts and saved shell
+// commands before it, and it goes back after them in the redrawn history.
+// A reload still drops them.
+//
+// +kinds+: the history's children in order, each "user" (a prompt bubble),
+// "shell" (a `!cmd`'s bubble, which the redrawn history has too) or
+// anything else.
+
+// @return {{users: number, shells: number}} the anchor of kinds[index]
+export function commandAnchor(kinds, index) {
+  const before = kinds.slice(0, index);
+  return {
+    users: before.filter((k) => k === "user").length,
+    shells: before.filter((k) => k === "shell").length,
+  };
+}
+
+// The index to insert a kept bubble at: before the first prompt or shell
+// command past its anchor (it ran after its turn), else at the end (a
+// history a !rollback made shorter too).
+export function anchorInsertIndex(kinds, { users = 0, shells = 0 } = {}) {
+  let seenUsers = 0;
+  let seenShells = 0;
+  for (let i = 0; i < kinds.length; i += 1) {
+    if (kinds[i] === "user") {
+      if (seenUsers >= users) return i;
+      seenUsers += 1;
+    } else if (kinds[i] === "shell") {
+      if (seenShells >= shells) return i;
+      seenShells += 1;
+    }
+  }
+  return kinds.length;
+}
+
+const SHELL_LINE = /^!\s*\S/;
+const ROLLBACK_LINE = /^!rollback(\s|$)/;
+
+// Whether a command bubble's line is kept across a resync: a `!cmd`'s is
+// redrawn from its saved message (keeping it too would show it twice).
+export function isShellCommandLine(line) {
+  const text = String(line ?? "");
+  return SHELL_LINE.test(text) && !ROLLBACK_LINE.test(text);
+}
+
+export function keepsCommandBubble(line) {
+  return !isShellCommandLine(line);
+}

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

@@ -22,7 +22,8 @@ export const RECENT_KEY = "chi_model_recent";
 
 // The payload's models as rows: {name, host, id, isDefaultHost}, and
 // sampling (the configured request parameters as /model words them) when
-// the model has some, and llm_context (the LLM context it starts under,
+// the model has some, thinking (the level it starts at, Thinking::Explained#summary),
+// and llm_context (the LLM context it starts under,
 // LLMContextStrategy::Explained#summary) when the server sent one, and
 // configured when the host declares the id (hosts.<name>.models), and
 // unavailable when that host failed to list its models. The
@@ -42,6 +43,7 @@ export function modelRows(payload) {
       if (m.configured === true) row.configured = true;
       if (m.unavailable === true) row.unavailable = true;
       if (m.llm_context && typeof m.llm_context === "object") row.llm_context = m.llm_context;
+      if (m.thinking && typeof m.thinking === "object") row.thinking = m.thinking;
       return row;
     });
 }