samagotchi 0.75.0 → 0.76.0

data/lib/samagotchi/toolkit.rb

@@ -77,16 +77,17 @@ module Samagotchi
         super
       end
 
-      # What the prompt carries of it: the declared names, the reachable
-      # and removed ones (the not-loaded line, find_tools' chi line), when
-      # either list has any the kit name the line gives, and the kit's
-      # prompt text. Two equal keys build the same prompt (/toolkit's "no
-      # change", its cost).
-      def prompt_key
-        [declared.map(&:name), reachable, removed, reachable.empty? && removed.empty? ? nil : resolved.name,
-         resolved.prompt]
+      # What the prompt carries before the kit's prompt text: the declared
+      # names, reachable and removed ones (the not-loaded line, find_tools'
+      # chi line), and, when either list has any, the kit name the line gives.
+      def tools_key
+        [declared.map(&:name), reachable, removed, reachable.empty? && removed.empty? ? nil : resolved.name]
       end
 
+      # What the prompt carries of it, including the kit's prompt text. Two
+      # equal keys build the same prompt (/toolkit's "no change", its cost).
+      def prompt_key = [*tools_key, resolved.prompt]
+
       # What the UIs show (the snapshot, command_ran): nil when it declares
       # every tool on full (full, full +delegate), which has no chip.
       # +left_out+ is +reachable+ and +removed+ together; +prompt+ (the

data/lib/samagotchi/toolkit_command.rb

@@ -129,9 +129,9 @@ module Samagotchi
 
     # What the confirm says is read again: from the tools on (the front of
     # every request), or from the kit's prompt text on when only that
-    # changes (Explained#prompt_key's last part: a kit on full's tools).
+    # changes (Explained#tools_key is unchanged: only a kit's prompt text).
     def reread_line(before, after)
-      if before.prompt_key[0...-1] == after.prompt_key[0...-1]
+      if before.tools_key == after.tools_key
         "The kit's prompt text is in the system prompt: the prompt is read again from it on."
       else
         "The tools are the front of every request: the whole prompt is read again from them on."

data/lib/samagotchi/tools/builtins.rb

@@ -144,11 +144,7 @@ module Samagotchi
 
       # The XDG dirs set in +env+, made absolute (nothing for an unset one).
       def xdg_absolute(env = ENV)
-        set = ->(key) { !env.fetch(key, "").to_s.strip.empty? }
-        dirs = {}
-        dirs["XDG_STATE_HOME"] = Paths.state_home(env: env) if set.call("XDG_STATE_HOME")
-        dirs["XDG_CONFIG_HOME"] = Paths.config_home(env: env) if set.call("XDG_CONFIG_HOME")
-        dirs
+        Paths.absolute_xdg(env: env)
       end
 
       # The unsets that keep a `chi --model X` worker's default from passing

data/lib/samagotchi/version.rb

@@ -1,7 +1,7 @@
 # frozen_string_literal: true
 
 module Samagotchi
-  VERSION = "0.75.0"
+  VERSION = "0.76.0"
   # Sent as the User-Agent of every HTTP request chi makes.
   USER_AGENT = "chi/#{VERSION}"
 end

data/lib/samagotchi/web/app.rb

@@ -1794,6 +1794,9 @@ module Samagotchi
           # before the next turn streams (live over saved).
           context: source["context"],
           tokens: source["tokens"],
+          # The tokens, cost and speeds per model (SessionMetrics'
+          # model_segments; none with one model, or from an older file).
+          model_segments: Array(source["model_segments"]),
           # The memory indexes the session's prompt holds (the ctx tooltip).
           memory_index: source["memory_index"]
         }

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

@@ -1475,7 +1475,7 @@ async function handleSendTurn({ delivery: chosen = null } = {}) {
     // The chips stay (uploaded ones keep their ref, to sentTo) for another try: an
     // image refused (bad_image, too_large, bad_images) says why below.
     if (e.code === "owned_by_tui") {
-      if (sentTo !== selected) return alert(e.message);
+      if (sentTo !== selected) return showToast(cantText(`send to ${sentTo.slice(0, 8)}`, e), null, { sticky: true });
       current.owner = "tui";
       updateComposerMode();
       return;
@@ -1485,7 +1485,7 @@ async function handleSendTurn({ delivery: chosen = null } = {}) {
     if (e.code === "not_found") return leaveGoneSession(sentTo, "not_found");
     // A /queue to a worker that can't queue (it would steer): its words, the text kept.
     if (e.code === "delivery_unavailable" && sentTo === selected) return addCommandBubble({ label: null, line, text: e.body?.detail || e.message, failed: true });
-    alert(e.message);
+    showToast(cantText("send", e), null, { sticky: true });
   } finally {
     actionBtn.disabled = current.owner === "tui";
     promptEl.focus();
@@ -1502,13 +1502,14 @@ const composerCommands = createComposerCommands({
   refreshHistory: () => refreshHistory(),
   failed: (line, text) => addCommandBubble({ label: null, line, text, failed: true }),
   // As handleSendTurn says a turn's failure after a switch.
-  failedElsewhere: (line, text) => alert(`${line}: ${text}`),
+  failedElsewhere: (line, text, id) => showToast(cantText(`run ${line} in ${id.slice(0, 8)}`, text), null, { sticky: true }),
   ownedByTui: () => {
     current.owner = "tui";
     updateComposerMode();
   },
   composer: {
     busy: () => { actionBtn.disabled = true; },
+    text: () => promptEl.value,
     clear: () => clearComposer(),
     done: () => {
       actionBtn.disabled = current.owner === "tui";

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

@@ -230,7 +230,7 @@ export function chainRowsTemplate(f, { activeId = null, now = Date.now() } = {})
     const ago = relativeTime(s.updated_at || s.created_at || "", now);
     const tip = `${s.id} · ${s.status || ""}${ago ? ` · updated ${ago}` : ""}`;
     return html`<div class=${cls.join(" ")} data-id=${s.id} style=${"--depth:1"} role=${"listitem"}><button type="button" class="fr-open"
-      title=${tip}>${rowDotTemplate(s, waiting)}<span class="cr-day">#${n}</span>${
+      title=${tip}>${rowDotTemplate(s, waiting)}<span class="cr-n">#${n}</span>${
       date ? html`<span class=${date.length > 10 ? "cr-date timed" : "cr-date"}>${date}</span>` : nothing}<span class="fr-preview">${s.recap || chainTitle(s, byId)}</span>${
       rowBadgesTemplate(waiting, stopped)}${s.archived ? archivedBadge : nothing}</button></div>`;
   });

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

@@ -60,9 +60,10 @@ export function routeSessionLine(prompt, { commands = [], images = 0, delivery =
 //   followWokenWorker()  the command woke the worker: follow its stream
 //   refreshHistory()
 //   failed(line, text)   a failed command's bubble
-//   failedElsewhere(line, text)   a command to a session no longer open failed
+//   failedElsewhere(line, text, id)   a command to session id, no longer open, failed
 //   ownedByTui()         a plain terminal holds the session
-//   composer: {busy(), clear(), done()}   the composer's send button and text
+//   composer: {busy(), text(), clear(), done()}   the composer's send button
+//                        and text (text(): what it holds now)
 export function createComposerCommands(deps) {
   const { session, clientId, sendCommand, followWokenWorker, refreshHistory, failed, failedElsewhere, ownedByTui,
           composer } = deps;
@@ -72,14 +73,17 @@ export function createComposerCommands(deps) {
   // worker, owner and history are left alone then.
   async function send(line) {
     const id = session();
+    // The composer is shared: a draft typed in a session opened meanwhile
+    // stays when the answer comes; only the text sent goes.
+    const sent = composer.text();
     composer.busy();
     try {
       await sendCommand(id, line, { clientId });
-      composer.clear();
+      if (composer.text() === sent) composer.clear();
       refreshHistory();
       if (session() === id) followWokenWorker();
     } catch (e) {
-      if (session() !== id) return failedElsewhere(line, e.message);
+      if (session() !== id) return failedElsewhere(line, e.message, id);
       if (e.code === "owned_by_tui") {
         ownedByTui();
         return;
@@ -103,7 +107,7 @@ export function createComposerCommands(deps) {
       return true;
     } catch (e) {
       if (session() === id) failed(line, e.message);
-      else failedElsewhere(line, e.message);
+      else failedElsewhere(line, e.message, id);
       return false;
     }
   }

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

@@ -133,14 +133,23 @@ export function ctxBarText(pct, windowSource = null) {
 }
 
 // The info bar's / the card's ctx tooltip: the window line under the title
-// when there's a number (tokensTipText), or why there's none — "No request
-// in this session yet" — when the page has no percent (before any request;
-// the page isn't told context.status is off, and the window always resolves,
-// so "?" never means 0).
-export function ctxBarTip(tokens, memoryIndex, window) {
-  if (num(window?.pct) === null) return "No request in this session yet";
+// when there's a number (tokensTipText), or why there's none when the page
+// has no percent: "No request in this session yet" before any request (the
+// page isn't told context.status is off, so "?" never means 0); after one,
+// "Context window unknown" over the session's tokens and per-model rows with
+// no window line (a window lookup that failed; a model list without a context
+// length still falls back to context.window_tokens or chi's default). +window.afterSwitch+ (the info
+// bar's; a /model whose new window didn't resolve, no request since:
+// Engine#announce_switched_context! sent usage null) adds "until the next
+// request", which asks for it again.
+export function ctxBarTip(tokens, memoryIndex, window, segments = null) {
+  if (num(window?.pct) === null) {
+    if (!num(tokens?.prompt_sum) && !num(tokens?.completion_sum)) return "No request in this session yet";
+    const why = window?.afterSwitch ? "Context window unknown until the next request" : "Context window unknown";
+    return tokensTipText(tokens, why, memoryIndex, null, segments);
+  }
 
-  return tokensTipText(tokens, undefined, memoryIndex, window);
+  return tokensTipText(tokens, undefined, memoryIndex, window, segments);
 }
 
 // The ctx tooltip's window line ({tokens:, source:, pct:, budget:} the
@@ -166,14 +175,67 @@ export function ctxWindowText(window) {
   return `context: ${used}${countK(tokens)} tokens${sourceText}`;
 }
 
+// ── A session's rows per model (the timing's modelSegments) ──────────────
+// The words /stats' models: block shares (Formatting#model_segment_*;
+// spec/shared/labels_matrix.json model_segments).
+
+// A row's model: its ref; the rows folded past the cap ("3 earlier
+// segments", told by folded_segments); a session's turns from before turns
+// named their model ("earlier turns").
+export function modelSegmentLabel(row) {
+  const folded = num(row?.folded_segments);
+  if (folded !== null && folded > 0) return `${folded} earlier segments`;
+  return row?.model_ref ? String(row.model_ref) : "earlier turns";
+}
+
+// "6.0k in / 40 out".
+export function modelSegmentTokensText(row) {
+  return `${countK(num(row?.prompt_sum) || 0)} in / ${countK(num(row?.completion_sum) || 0)} out`;
+}
+
+// "$0.42", "~$0.12", "$0.42 + ~$0.12"; "" for none.
+export function modelSegmentCostText(row) {
+  return [costText(row?.cost_sum), costText(row?.cost_estimate_sum, { estimate: true })].filter(Boolean).join(" + ");
+}
+
+// "switch prefill 3.0k uncached prompt ≈$0.0060": what the switch to the
+// row's model prefilled (its cost an estimate from the configured price,
+// left out without one); "" for none.
+export function modelSegmentPrefillText(row) {
+  const tokens = num(row?.prefill_tokens) || 0;
+  if (tokens <= 0) return "";
+  const cost = costText(row?.prefill_cost_estimate);
+  return `switch prefill ${countK(tokens)} uncached prompt${cost ? ` ≈${cost}` : ""}`;
+}
+
+// The tooltip's rows per model, with ≥ 2 (one model: none): "model ·
+// turns · in / out · cost", a switch's prefill under its row.
+function modelSegmentLines(segments) {
+  if (!Array.isArray(segments) || segments.length < 2) return [];
+  const lines = ["per model:"];
+  for (const row of segments) {
+    const turns = num(row.turns) || 0;
+    const parts = [modelSegmentLabel(row), `${turns} ${turns === 1 ? "turn" : "turns"}`, modelSegmentTokensText(row)];
+    const cost = modelSegmentCostText(row);
+    if (cost) parts.push(cost);
+    lines.push(`  ${parts.join(" · ")}`);
+    const prefill = modelSegmentPrefillText(row);
+    if (prefill) lines.push(`    └ ${prefill}`);
+  }
+  return lines;
+}
+
 // 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 (an estimated part with
 // "~", costLine), and the memory indexes
 // 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, window = null) {
+// +window+ (ctxWindowText's) the window line under the title; with
+// +segments+ (the timing's modelSegments, ≥ 2) the rows per model above the
+// totals. 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, window = null,
+                              segments = null) {
   const memory = memoryIndexText(memoryIndex);
   const memoryLine = memory ? `\nmemory index: ${memory}` : "";
   const windowText = ctxWindowText(window);
@@ -186,7 +248,7 @@ export function tokensTipText(tokens, title = "Context used after the last turn"
   const parts = [`in ${count(prompt)}`];
   if (cached > 0 && prompt > 0) parts.push(`cached ${count(cached)} (${Math.round((cached * 100) / prompt)}%)`);
   parts.push(`out ${count(completion)}${reasoning > 0 ? ` (reasoning ${count(reasoning)})` : ""}`);
-  const lines = [title, `tokens: ${parts.join(" · ")}`];
+  const lines = [title, ...modelSegmentLines(segments), `tokens: ${parts.join(" · ")}`];
   const cost = costLine(tokens);
   if (cost) lines.push(cost);
   lines.push("this session only, all requests");

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

@@ -326,22 +326,35 @@ button.task-stop:hover:not(:disabled){background:var(--danger-bg-dim);filter:non
 .bubble.question>summary .question-said{max-width:72ch;margin:2px 0 1px 1.05em;font-size:12.5px;font-weight:400;line-height:1.5;color:var(--text-muted);white-space:normal;overflow-wrap:anywhere;display:-webkit-box;-webkit-box-orient:vertical;-webkit-line-clamp:2;line-clamp:2;overflow:hidden}
 .bubble.question[open]>summary .question-said{display:none}
 .question-said-full{margin:4px 0 6px;font-size:13px;line-height:1.55;color:var(--text-muted);white-space:pre-wrap;overflow-wrap:anywhere}
-/* Its markdown once the card is first opened (question_cards.js renderSaidOnOpen), as compact as the stage's. */
-.question-said-full.markdown{white-space:normal}
-.question-said-full.markdown :is(h1,h2,h3,h4,h5,h6){font-size:13.5px;font-weight:600;line-height:1.4;margin:8px 0 2px;color:var(--text-2)}
-.question-said-full.markdown :is(p,ul,ol,blockquote,table,pre){margin:0 0 6px}
-.question-said-full.markdown>:first-child{margin-top:0}
-.question-said-full.markdown>:last-child{margin-bottom:0}
-.question-said-full.markdown :is(ul,ol){padding-left:1.3em}
-.question-said-full.markdown li>:is(ul,ol){margin:0}
-.question-said-full.markdown hr{border:0;border-top:1px solid var(--border);margin:8px 0}
-.question-said-full.markdown blockquote{border-left:3px solid var(--border-strong);padding-left:10px}
-.question-said-full.markdown code{font-family:ui-monospace,SFMono-Regular,Menlo,monospace;font-size:.88em}
-.question-said-full.markdown :not(pre)>code{background:var(--bg-inset);border:1px solid var(--border);border-radius:4px;padding:0 4px}
-.question-said-full.markdown pre{overflow:auto;padding:6px 8px;border-radius:6px;background:var(--bg-inset);border:1px solid var(--border);font-size:12px;line-height:1.45;white-space:pre}
-.question-said-full.markdown table{border-collapse:collapse;display:block;max-width:100%;overflow:auto}
-.question-said-full.markdown :is(th,td){border:1px solid var(--border-strong);padding:3px 7px;font-size:12.5px}
-.question-said-full.markdown th{background:var(--bg-inset);font-weight:600}
+.question-said-full.md-compact{
+  --md-heading-size:13.5px;--md-heading-color:var(--text-2);
+  --md-subheading-size:13.5px;--md-subheading-color:var(--text-2);
+  --md-strong-color:currentColor;--md-blockquote-color:currentColor;--md-code-color:currentColor;
+  --md-pre-size:12px;--md-cell-size:12.5px
+}
+.ts-asking.md-compact{
+  --md-heading-size:14.5px;--md-heading-color:var(--text-bright);
+  --md-subheading-size:14px;--md-subheading-color:var(--text-2);
+  --md-strong-color:var(--text-bright);--md-blockquote-color:var(--text-muted);--md-code-color:var(--text);
+  --md-pre-size:12.5px;--md-cell-size:13px
+}
+.md-compact:is(.ts-asking,.question-said-full){white-space:normal}
+.md-compact :is(h1,h2,h3,h4,h5,h6){font-size:var(--md-heading-size);font-weight:600;line-height:1.4;margin:8px 0 2px;color:var(--md-heading-color)}
+.md-compact :is(h4,h5,h6){font-size:var(--md-subheading-size);color:var(--md-subheading-color)}
+.md-compact :is(p,ul,ol,blockquote,table,pre){margin:0 0 6px}
+.md-compact>:first-child{margin-top:0}
+.md-compact>:last-child{margin-bottom:0}
+.md-compact :is(ul,ol){padding-left:1.3em}
+.md-compact li>:is(ul,ol){margin:0}
+.md-compact hr{border:0;border-top:1px solid var(--border);margin:8px 0}
+.md-compact blockquote{border-left:3px solid var(--border-strong);padding-left:10px;color:var(--md-blockquote-color)}
+.md-compact code{font-family:ui-monospace,SFMono-Regular,Menlo,monospace;font-size:.88em}
+.md-compact :not(pre)>code{background:var(--bg-inset);border:1px solid var(--border);border-radius:4px;padding:0 4px;color:var(--md-code-color)}
+.md-compact strong{color:var(--md-strong-color)}
+.md-compact pre{overflow:auto;padding:6px 8px;border-radius:6px;background:var(--bg-inset);border:1px solid var(--border);font-size:var(--md-pre-size);line-height:1.45;white-space:pre}
+.md-compact table{border-collapse:collapse;display:block;max-width:100%;overflow:auto}
+.md-compact :is(th,td){border:1px solid var(--border-strong);padding:3px 7px;font-size:var(--md-cell-size)}
+.md-compact th{background:var(--bg-inset);font-weight:600}
 .bubble.command{border-left:3px solid var(--text-faint);background:var(--bg-panel);font-size:13px}
 .bubble.note{border-left:3px dotted var(--text-faint);background:transparent;font-size:12px;color:var(--text-muted);opacity:.85}
 .note-line{font-family:ui-monospace,SFMono-Regular,Menlo,monospace;color:var(--text-faint);margin-bottom:2px}
@@ -730,24 +743,6 @@ body.stage-view .turn-stage{position:absolute;left:0;right:0;bottom:100%;margin-
    text, then its markdown when web.markdown is on (.markdown, stage_slots.js drawAsking); capped so the card keeps
    its room, read from the top. */
 .ts-asking{max-height:calc(var(--stage-max,60vh) * .35);overflow:auto;scrollbar-width:thin;white-space:pre-wrap;overflow-wrap:anywhere;font-size:14px;line-height:1.55;color:var(--narr);border-left:2px solid var(--border-strong);padding-left:10px}
-/* Its markdown, compact: about the raw text's metrics, so the swap barely moves the slot. */
-.ts-asking.markdown{white-space:normal}
-.ts-asking.markdown :is(h1,h2,h3,h4,h5,h6){font-size:14.5px;font-weight:600;line-height:1.4;margin:8px 0 2px;color:var(--text-bright)}
-.ts-asking.markdown :is(h4,h5,h6){font-size:14px;color:var(--text-2)}
-.ts-asking.markdown :is(p,ul,ol,blockquote,table,pre){margin:0 0 6px}
-.ts-asking.markdown>:first-child{margin-top:0}
-.ts-asking.markdown>:last-child{margin-bottom:0}
-.ts-asking.markdown :is(ul,ol){padding-left:1.3em}
-.ts-asking.markdown li>:is(ul,ol){margin:0}
-.ts-asking.markdown strong{color:var(--text-bright)}
-.ts-asking.markdown hr{border:0;border-top:1px solid var(--border);margin:8px 0}
-.ts-asking.markdown blockquote{border-left:3px solid var(--border-strong);padding-left:10px;color:var(--text-muted)}
-.ts-asking.markdown code{font-family:ui-monospace,SFMono-Regular,Menlo,monospace;font-size:.88em}
-.ts-asking.markdown :not(pre)>code{background:var(--bg-inset);border:1px solid var(--border);border-radius:4px;padding:0 4px;color:var(--text)}
-.ts-asking.markdown pre{overflow:auto;padding:6px 8px;border-radius:6px;background:var(--bg-inset);border:1px solid var(--border);font-size:12.5px;line-height:1.45;white-space:pre}
-.ts-asking.markdown table{border-collapse:collapse;display:block;max-width:100%;overflow:auto}
-.ts-asking.markdown :is(th,td){border:1px solid var(--border-strong);padding:3px 7px;font-size:13px}
-.ts-asking.markdown th{background:var(--bg-inset);font-weight:600}
 .turn-stage.asking .ts-headline{display:none}
 @keyframes ts-fade{from{opacity:.35}}
 .ts-now{height:var(--ts-now-h);display:flex;align-items:center;gap:8px;min-width:0;padding:0 9px;border-radius:8px;background:var(--accent-tint);font-size:13px}
@@ -1031,7 +1026,7 @@ body.no-glass :is(#familyPop,#chainPop){background:var(--bg-panel);-webkit-backd
 .chain-chip .cc-arrow{display:inline-block;margin-right:4px}
 .chain-chip .live-dot{display:inline-block;width:6px;height:6px;border-radius:50%;background:var(--ok);margin-right:3px;vertical-align:1px}
 .chain-chip .cc-waiting{font-family:system-ui,sans-serif;font-weight:600}
-.chain-row .cr-day{font-family:ui-monospace,monospace;font-size:11px;color:var(--accent);white-space:nowrap}
+.chain-row .cr-n{font-family:ui-monospace,monospace;font-size:11px;color:var(--accent);white-space:nowrap}
 .chain-row .cr-date{font-family:ui-monospace,monospace;font-size:11px;color:var(--text-muted);white-space:nowrap;font-variant-numeric:tabular-nums}
 /* A row with its time (chainStamp): every row's date as wide, so the previews line up. */
 .chain-rows:has(.cr-date.timed) .cr-date{min-width:16ch}

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

@@ -67,7 +67,8 @@ export function infoBarModel({ id, state, sessions, chain, cardTitle, badge, res
   const tokens = state.timing.tokens;
   const ctxTip = ctxBarTip(tokens, state.timing.memoryIndex,
                            { tokens: state.ctxWindow, source: state.ctxWindowSource, pct: state.ctxPct,
-                             budget: state.llmContext?.budget_tokens });
+                             budget: state.llmContext?.budget_tokens, afterSwitch: state.ctxUnknownAfterSwitch },
+                           state.timing.modelSegments);
   const speedTextNow = lastSpeedText(tokens);
   const avgSpeed = speedText(tokens?.avg_decode_tps, tokens?.tps_source);
   const sessionDuration = sessionDurationNow(state.timing, running);

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

@@ -450,7 +450,7 @@ export function createQuestionCards({
           state = "done";
           // Sanitised on the server (POST /api/render).
           full.innerHTML = html;
-          full.classList.add("markdown");
+          full.classList.add("markdown", "md-compact");
           linkRefs(full);
         });
     };

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

@@ -31,6 +31,9 @@ export function emptySessionState() {
     ctxWindow: null,
     // Where it came from ("server", "config", … "default": chi's guess).
     ctxWindowSource: null,
+    // A /model's new window didn't resolve and no request has started since
+    // (stream_handlers.js context_status): the ctx tooltip says it waits.
+    ctxUnknownAfterSwitch: false,
     status: "",
     model: "",
     // What the worker's server said it served for a model name: [served, asked].

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

@@ -58,7 +58,7 @@ export function createStageSlots({
     asking.hidden = !text;
     if (shown.asking === text) return;
     shown.asking = text;
-    asking.classList.remove("markdown");
+    asking.classList.remove("markdown", "md-compact");
     asking.textContent = text;
     asking.scrollTop = 0;
     if (!text) return;
@@ -76,7 +76,7 @@ export function createStageSlots({
       if (!html || shown.asking !== text || asking.classList.contains("markdown")) return;
       // Sanitised on the server (POST /api/render), as an answer's HTML is.
       asking.innerHTML = html;
-      asking.classList.add("markdown");
+      asking.classList.add("markdown", "md-compact");
       asking.scrollTop = 0;
       linkRefs(asking);
     });
@@ -154,7 +154,7 @@ export function createStageSlots({
   function reset() {
     for (const slot of [headline, asking, now, trail, ticks]) slot.replaceChildren();
     asking.hidden = true;
-    asking.classList.remove("markdown");
+    asking.classList.remove("markdown", "md-compact");
     shown.headline = shown.now = shown.trail = shown.asking = null;
   }
 

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

@@ -73,6 +73,9 @@ export function createStreamHandlers(deps) {
       }
     },
     generation_started: (data) => {
+      // The request after a /model's unresolved window started: whatever it
+      // brings, the ctx tooltip no longer waits for it (ctxBarTip).
+      state().ctxUnknownAfterSwitch = false;
       if (data?.context_window_tokens) {
         const current = state();
         current.ctxWindow = data.context_window_tokens;
@@ -91,6 +94,10 @@ export function createStreamHandlers(deps) {
       if (data?.served_model) current.served = [data.served_model, data.requested_model, data.served_expected_by || null];
       // The generation's speed and the session's token totals, from the server.
       current.timing.tokens = generationTokens(current.timing.tokens, data);
+      // The rows per model, the server's whole list (the open one with this
+      // generation in); an event without them (a replay, an older worker)
+      // keeps the page's.
+      if (Array.isArray(data?.model_segments)) current.timing.modelSegments = data.model_segments;
       if (data?.served_model || data?.tokens) updateInfoBar();
       turnDom.generationCompleted();
     },
@@ -185,6 +192,7 @@ export function createStreamHandlers(deps) {
     // then show the prompt once more, marked failed (with why). This tab's own
     // prompt goes back into an empty composer.
     prompt_restored: async (data) => {
+      const id = selected();
       keepEarlyRestore(data, earlyRestores, { myId: clientId, sentIds });
       const { refill, label, images: restoredImages } = restoreAction(data, { myId: clientId, sentIds, composerEmpty: composer.empty() });
       if (refill) composer.refill(refill);
@@ -192,6 +200,9 @@ export function createStreamHandlers(deps) {
       const failedText = liveTurn.failedText;
       liveTurn.failedText = null;
       await resync();
+      // Another session was opened during the re-read: its history isn't
+      // this one's (the re-read drew nothing either).
+      if (selected() !== id) return;
       // The re-render drew it from the session (failed_turn): drawn again
       // below with this event's label and images.
       const drawn = historyEl.querySelector(".bubble.user[data-failed-turn]");
@@ -207,6 +218,7 @@ export function createStreamHandlers(deps) {
     // A session command ran in the worker (any client's). One that changed
     // the conversation (!rollback, !cmd, a continue answered no) re-reads it.
     command_ran: async (data) => {
+      const id = selected();
       const view = commandView(data, clientId);
       if (view.modelName && view.modelName !== state().model) {
         state().model = view.modelName;
@@ -240,7 +252,11 @@ export function createStreamHandlers(deps) {
         state().toolkit = data.toolkit || null;
         updateInfoBar();
       }
-      if (view.resync) await resync();
+      if (view.resync) {
+        await resync();
+        // Another session was opened during the re-read: the bubble isn't its.
+        if (selected() !== id) return;
+      }
       // A card's action that isn't anytime (the step-limit question's
       // /continue) has no bubble: nothing to remove. An anytime or queued
       // one's bubble was drawn at its command_queued: filled in here. One
@@ -286,6 +302,21 @@ export function createStreamHandlers(deps) {
     },
     context_status: (data) => {
       const current = state();
+      // After /model (Engine#announce_switched_context!): the new model's
+      // window, and usage null when it didn't resolve (the % clears, "ctx ?").
+      // Only that event carries the key: a null drops the old model's window,
+      // so the chunks before the next window count no % against it.
+      if (data && "context_window_tokens" in data) {
+        current.ctxWindow = data.context_window_tokens || null;
+        current.ctxWindowSource = data.context_window_tokens ? data.context_window_source || null : null;
+        // The ctx tooltip's "until the next request" (ctxBarTip).
+        current.ctxUnknownAfterSwitch = !data.context_window_tokens;
+      }
+      if (data && "usage" in data && data.usage === null) {
+        current.ctxPct = null;
+        updateInfoBar();
+        return;
+      }
       const pct = extractCtxPct(data, current.ctxWindow, current.llmContext?.budget_tokens);
       if (pct !== null) {
         current.ctxPct = pct;

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

@@ -81,6 +81,10 @@ export function normalizeTiming(data = {}) {
     // The memory indexes the session's prompt holds (SessionMetrics'
     // memory_index), for the ctx tooltip.
     memoryIndex: validRecord(data.memory_index) ? data.memory_index : null,
+    // The session's tokens, cost and speeds per model after a /model
+    // (SessionMetrics' model_segments; [] with one model), for the ctx
+    // tooltip.
+    modelSegments: Array.isArray(data.model_segments) ? data.model_segments.filter(validRecord) : [],
   };
 }
 
@@ -91,7 +95,7 @@ export function normalizeTiming(data = {}) {
 // dates keep arrival order), an unknown tool record at the end (they come
 // grouped by turn, in call order). startedAt, sessionDurationMs and
 // activeTurn are the reply's, and so are its tokens (a reply without them
-// keeps the page's). `short`: the reply's turn_count says the page
+// keeps the page's) and its model_segments (present: authoritative). `short`: the reply's turn_count says the page
 // misses a turn (it then reads the whole timing). A reply without
 // turn_count (an older server: the whole timing) replaces, as before; more
 // records than the count (a worker died before saving a turn the page saw)
@@ -117,7 +121,12 @@ export function mergeTiming(timing, data = {}) {
   });
   const tokens = fresh.tokens || timing?.tokens || null;
   const memoryIndex = fresh.memoryIndex || timing?.memoryIndex || null;
-  return { timing: { ...fresh, turnRecords, toolRecords, tokens, memoryIndex }, short: turnRecords.length < count };
+  // A reply that has the rows is the answer, an empty list too (one model
+  // now): a stale list never survives it; one without them (an older
+  // server) keeps the page's.
+  const modelSegments = data && "model_segments" in data ? fresh.modelSegments : (timing?.modelSegments || []);
+  return { timing: { ...fresh, turnRecords, toolRecords, tokens, memoryIndex, modelSegments },
+           short: turnRecords.length < count };
 }
 
 function replaceById(records, record) {

metadata

@@ -1,7 +1,7 @@
 --- !ruby/object:Gem::Specification
 name: samagotchi
 version: !ruby/object:Gem::Version
-  version: 0.75.0
+  version: 0.76.0
 platform: ruby
 authors:
 - Dmitry Dedov
@@ -461,6 +461,8 @@ files:
 - lib/samagotchi/session_inbox.rb
 - lib/samagotchi/session_manager.rb
 - lib/samagotchi/session_metrics.rb
+- lib/samagotchi/session_metrics/model_segment.rb
+- lib/samagotchi/session_metrics/totals.rb
 - lib/samagotchi/session_model_settings.rb
 - lib/samagotchi/session_observer.rb
 - lib/samagotchi/session_retention.rb