micro-models-agent 2.12.0 → 2.13.1

package/CHANGELOG.md

@@ -4,6 +4,35 @@ All notable changes to Micro Models Agent (MMA) will be documented in this file.
 
 The format is based on [Keep a Changelog](https://keepachangelog.com/), and this project adheres to [Semantic Versioning](https://semver.org/).
 
+## [Unreleased]
+
+## [2.13.1] - 2026-09-30
+
+### Fixed
+
+- **In-loop critic is now selective instead of always-on** (`src/core/agent/critic-gate.ts`, `loop-state.ts`, `src/core/agent.ts`): evidence is scoped to the current run (`LoopState.criticHistoryMark` + `gatherTurnToolEvents(cm, mark)`), so tool calls from previous questions no longer testify against the current task (observed: weather lookups from an earlier turn caused 4 false `critical` verdicts on a chat-only "show a React example" answer). Chat-only turns (no tool events in the run) now `pass` without an LLM call and without spending `maxCalls`; the dead mid-turn `mutationsSinceCritic` trigger branch is removed.
+- **Advice and repeat verdicts no longer retry** (`src/core/agent/critic-gate.ts`, `src/i18n/en.json`, `ru.json`): only a fresh `critical` inside `maxRepairCycles` repairs + retries; `concern`, over-budget `critical` and near-identical repeats inject `<system-summary>` (`critic.advice`, new `critic.repeat` — "change approach or finish honestly") and `pass`. Previously every non-ok verdict retried identically until `maxCalls`, looping the same answer 4 times.
+- **Critic sees chat-delivered answers** (`src/modules/critic/types.ts`, `evidence.ts`, `client.ts`, `src/core/agent.ts`): `EvidencePack.answerExcerpt` (≤2000 chars, secrets redacted) carries the final text for mutation-free turns, and the critic prompt states that a chat answer without files is not itself an issue. Previously the critic judged chat tasks blind (tools only) and demanded file artifacts the user never asked for.
+- **Honest critic evidence bits** (`src/modules/critic/evidence.ts`, `provider.ts`): successful bash no longer gets a synthesized `exitCode: 0` (`undefined` renders as `?`; `1` stays for known failures); setting `critic.provider` without `critic.model` now warns instead of silently doing nothing.
+- **Method calls no longer flagged as missing files** (`src/modules/hallucination/js-identifiers.ts`): `extractFileLikeTokens` drops tokens glued to `(` (`prev.map(`, `todos.map(` — `map` is a sourcemap extension), fixing false "mentioned files not found" warnings on code answers. Applies to all consumers (factual check, execution auditor, verifier).
+- **Rejected drafts no longer duplicate in the feed** (`src/core/agent/artifact-resolve.ts`, `hallucination-gate.ts`, `critic-gate.ts`, `audit-gate.ts`, `src/core/types.ts`, `src/modules/webui/forward.ts`, `assets/js/feed.js`, `app.js`): a gate-rejected draft was already streamed before the verdict, so the re-answer stacked as a visible double (observed: artifact-claim false positive on bare `game.ts` vs `src/game.ts`). Gates now emit `onMeta(…, "retract")` (new `MetaKind`, user line `ui.draft_retracted` en+ru); the web forwarder sends a dedicated SSE `retract` event and `feed.js onRetract` removes the rejected block (also from turn copy buffers) with a one-line note, the terminal prints the same line. Rejections are logged as `claim_retry` session entries (`SessionLogger.logClaimRetry`, new `SessionLogEntry.kind`), closing the `llm_usage`-without-`assistant` gap; critic verify exceptions now log a fail-open diagnostic instead of a silent lone `critic_check`.
+- **Artifact ground-check resolves bare basenames** (`src/core/agent/artifact-resolve.ts`): `resolveOneArtifact` goes through the auditor's `findExistingFile` (exact path, then basename/suffix subtree walk in the `baseDir` jail) instead of a strict `resolve`, so `game.ts` matches `src/game.ts` and no longer triggers a false claim-retry.
+
+## [2.13.0] - 2026-09-30
+
+### Added
+
+- **Web UI**: sidebar tabs — Session / Map / Trace. Square icon tabs with the active
+  tab name shown below them (`assets/js/sidebar.js`, `index.html`, `styles.css`).
+- **Map tab**: project-map index state — status, last-updated timestamp, watcher,
+  files/symbols/tokens, root, per-language breakdown, and a Re-index button
+  (`POST /api/index`). `IndexerModule.mapInfo()` + `ProjectMapInfo` host contract.
+- **Trace tab**: live per-session log of tool calls, phases, plan updates, errors,
+  questions and the answer text stream; clicking a row opens a detail panel (full
+  args, output, metadata) and doubles the sidebar width. Desktop only (hidden ≤760px).
+- **i18n**: `web.sidebar.tab_*`, `web.sidebar.map_*`, `web.sidebar.trace_*`,
+  `web.sidebar.detail_*` (en + ru).
+
 ## [2.12.0] - 2026-09-29
 
 ### Added

package/dist/assets/app.js

@@ -5,6 +5,7 @@ import { S, hooks } from "./js/store.js";
 import {
 onText, onReasoning, onMeta, onPhase, onToolStart, onToolEnd,
 onTurnEnd, onErrorEvent, onCommandOutput, onLog, onPlan, onQuestion,
+onRetract,
 showEmptyFeed, renderPlan, renderUsage, renderEnvironment, renderWorkspace,
 feedHasContent, loadHistory,
 } from "./js/feed.js";
@@ -19,6 +20,11 @@ loadMemory, loadSkills, loadPlugins, applyTheme, currentTheme, applyFontSize, se
 import {
 initRail, setPlan as railPlan, railEvent, applyRailI18n,
 } from "./js/rail.js";
+import {
+  initSidebar, sidebarSessionChanged, sidebarMapState,
+  sidebarTraceEvent, sidebarTurnEnd, sidebarBusy, sidebarMapIndicator,
+  applySidebarI18n,
+} from "./js/sidebar.js";
 
 document.querySelectorAll("[data-collapse]").forEach((head) => {
   const initArrow = head.querySelector(".arrow");
@@ -77,6 +83,9 @@ function renderState(snapshot, busy) {
     renderSelectors();
   }
   renderEnvironment(S.state.host);
+  sidebarMapState(S.state.host);
+  sidebarMapIndicator(S.state.host);
+  sidebarSessionChanged(sessionId);
   renderWorkspace(S.state.host);
   renderSessionsDropdown();
 }
@@ -131,8 +140,9 @@ const handlers = [
   ["tool_end", onToolEnd],
   ["plan", onPlan],
   ["question", onQuestion],
-  ["turn_end", onTurnEnd],
+  ["turn_end", (d) => { onTurnEnd(d); sidebarTurnEnd(); sidebarBusy(false); }],
   ["log", onLog],
+  ["retract", (d) => onRetract(d.reason || "", d.sessionId)],
 ];
 for (const [event, fn] of handlers) {
   es.addEventListener(event, (e) => {
@@ -149,6 +159,11 @@ for (const ev of ["tool_start", "tool_end", "text", "plan"]) {
     else railEvent(ev, d.name);
   });
 }
+// Sidebar Trace: live per-session event log. Feed handlers stay untouched.
+for (const ev of ["tool_start", "tool_end", "phase", "plan", "question"]) {
+  es.addEventListener(ev, (e) => { const d = parse(e); if (d) sidebarTraceEvent(ev, d); });
+}
+es.addEventListener("text", (e) => { const d = parse(e); if (d) sidebarTraceEvent("text", d); });
 es.addEventListener("command", (e) => {
   const d = parse(e);
   if (d && typeof d.text === "string") onCommandOutput(d.text, d.sessionId);
@@ -158,6 +173,8 @@ es.addEventListener("command", (e) => {
     if (d) {
       S.state.busy = !!d.busy;
       reflectBusy();
+      sidebarBusy(!!d.busy);
+      refresh();
       if (d.usage && typeof d.usage === "object") {
         S.state.host = Object.assign({}, S.state.host, d.usage);
         renderUsage(S.state.host);
@@ -167,7 +184,7 @@ es.addEventListener("command", (e) => {
   es.addEventListener("error", (e) => {
     if (!e.data) return;
     const d = parse(e);
-    if (d) onErrorEvent(d);
+    if (d) { onErrorEvent(d); sidebarTraceEvent("error", d); }
   });
 }
 
@@ -239,6 +256,7 @@ function applyStaticI18n() {
   renderSelectors();
   renderSessionsDropdown();
   applyRailI18n();
+  applySidebarI18n();
   if (msgs.querySelector(".empty")) showEmptyFeed();
   if (connStatus && !connStatus.hidden) setConnStatus(true);
 }
@@ -265,6 +283,8 @@ async function boot() {
   renderSettings();
   await loadConfig();
   initRail(S.state.config && S.state.config.ui ? S.state.config.ui.activityRail : undefined);
+  initSidebar();
+  applySidebarI18n();
   await loadModels();
   await loadProviders();
   await loadSessions();

package/dist/assets/index.html

@@ -96,19 +96,49 @@
   </div>
 
   <div class="sb">
-    <div class="sec plan-sec">
-      <button class="sec-h" type="button" data-collapse aria-expanded="true"><span class="arrow">▾</span> <span id="planHeaderLabel">plan</span> <span class="n" id="planCount"></span></button>
-      <div class="sec-body" id="plan"></div>
+    <div class="sb-tabs" role="tablist" aria-label="sidebar">
+      <button class="sb-tab on" type="button" role="tab" data-tab="session" aria-selected="true" aria-controls="pane-session" title="Session">
+        <span class="sb-ico" data-icon="session" aria-hidden="true"></span>
+        <span class="sb-dot" aria-hidden="true"></span>
+      </button>
+      <button class="sb-tab" type="button" role="tab" data-tab="map" aria-selected="false" aria-controls="pane-map" title="Map">
+        <span class="sb-ico" data-icon="map" aria-hidden="true"></span>
+        <span class="sb-dot" aria-hidden="true"></span>
+      </button>
+      <button class="sb-tab" type="button" role="tab" data-tab="trace" aria-selected="false" aria-controls="pane-trace" title="Trace">
+        <span class="sb-ico" data-icon="trace" aria-hidden="true"></span>
+        <span class="sb-dot" aria-hidden="true"></span>
+      </button>
     </div>
+    <div class="sb-name" id="sbName"><span class="sb-ico" data-icon="session" aria-hidden="true"></span><span id="sbNameText">Session</span></div>
 
-    <div class="sec usage-sec">
-      <button class="sec-h" type="button" data-collapse aria-expanded="true"><span class="arrow">▾</span> <span id="usageHeaderLabel">session · usage</span></button>
-      <div class="sec-body" id="usage"></div>
+    <div class="sb-pane" id="pane-session" role="tabpanel" data-pane="session">
+      <div class="sec plan-sec">
+        <button class="sec-h" type="button" data-collapse aria-expanded="true"><span class="arrow">▾</span> <span id="planHeaderLabel">plan</span> <span class="n" id="planCount"></span></button>
+        <div class="sec-body" id="plan"></div>
+      </div>
+      <div class="sec usage-sec">
+        <button class="sec-h" type="button" data-collapse aria-expanded="true"><span class="arrow">▾</span> <span id="usageHeaderLabel">session · usage</span></button>
+        <div class="sec-body" id="usage"></div>
+      </div>
+      <div class="sec env-sec">
+        <button class="sec-h" type="button" data-collapse aria-expanded="true"><span class="arrow">▾</span> <span id="envHeaderLabel">environment</span></button>
+        <div class="sec-body" id="env"></div>
+      </div>
     </div>
 
-    <div class="sec env-sec">
-      <button class="sec-h" type="button" data-collapse aria-expanded="true"><span class="arrow">▾</span> <span id="envHeaderLabel">environment</span></button>
-      <div class="sec-body" id="env"></div>
+    <div class="sb-pane" id="pane-map" role="tabpanel" data-pane="map" hidden>
+      <div class="sec map-sec">
+        <div class="sec-body" id="mapBody"></div>
+      </div>
+    </div>
+
+    <div class="sb-pane" id="pane-trace" role="tabpanel" data-pane="trace" hidden>
+      <div class="trace-list" id="traceList"></div>
+      <div class="trc-detail" id="traceDetail" hidden>
+        <div class="detail-h"><span class="g" id="detailGlyph">◇</span><span id="detailTitle"></span><button class="x linkbtn" type="button" id="detailClose">✕</button></div>
+        <div class="detail-b" id="detailBody"></div>
+      </div>
     </div>
   </div>
 </div>

package/dist/assets/js/dom.js

@@ -74,6 +74,21 @@ search: [
   ["circle", { cx: 11, cy: 11, r: 8 }],
   ["path", { d: "m21 21-4.3-4.3" }],
 ],
+"list-checks": [
+  ["path", { d: "m3 17 2 2 4-4" }],
+  ["path", { d: "m3 7 2 2 4-4" }],
+  ["path", { d: "M13 6h8" }],
+  ["path", { d: "M13 12h8" }],
+  ["path", { d: "M13 18h8" }],
+],
+map: [
+  ["path", { d: "M14.106 5.553a2 2 0 0 0 1.788 0l3.659-1.83A1 1 0 0 1 21 4.619v12.764a1 1 0 0 1-.553.894l-4.553 2.277a2 2 0 0 1-1.788 0l-4.212-2.106a2 2 0 0 0-1.788 0l-3.659 1.83A1 1 0 0 1 3 19.381V6.618a1 1 0 0 1 .553-.894l4.553-2.277a2 2 0 0 1 1.788 0z" }],
+  ["path", { d: "M15 5.764v15" }],
+  ["path", { d: "M9 3.236v15" }],
+],
+activity: [
+  ["path", { d: "M22 12h-2.48a2 2 0 0 0-1.93 1.46l-2.35 8.36a.25.25 0 0 1-.48 0L9.24 2.18a.25.25 0 0 0-.48 0l-2.35 8.36A2 2 0 0 1 4.49 12H2" }],
+],
 "arrow-up": [["path", { d: "M12 19V5" }], ["path", { d: "m5 12 7-7 7 7" }]],
 panel: [
   ["rect", { width: 18, height: 18, x: 3, y: 3, rx: 2, ry: 2 }],

package/dist/assets/js/feed.js

@@ -36,12 +36,16 @@ export function makeFeed() {
 
 export function showEmptyFeed() {
   msgs.replaceChildren(emptyView());
+  follow = true;
+  hideToLatest();
   syncUrl();
 }
 
 export function clearFeed(feed) {
   if (!feed) return;
   feed.replaceChildren(emptyView());
+  follow = true;
+  hideToLatest();
   const tab = S.tabs.find((x) => x.feed === feed);
   if (tab) tab.planCard = null;
   if (feed === activeFeed()) scroll();
@@ -71,10 +75,80 @@ function nearBottom() {
   return msgs.scrollHeight - msgs.scrollTop - msgs.clientHeight < 140;
 }
 
-function scroll() {
-  if (nearBottom()) msgs.scrollTop = msgs.scrollHeight;
+// Sticky follow: while the reader is at the live edge we pin to the bottom
+// on every append; the moment they scroll up we leave them alone (the
+// "jump to latest" pill offers the trip back). A new user turn always
+// force-scrolls — otherwise one missed pin cascades into "no autoscroll"
+// for the whole turn, since every later `scroll()` would see `nearBottom()`
+// as false and skip.
+let follow = true;
+
+function scroll(force) {
+  if (force || follow || nearBottom()) {
+    msgs.scrollTop = msgs.scrollHeight;
+    follow = true;
+    hideToLatest();
+  } else {
+    maybeShowToLatest();
+  }
+}
+
+// Force-pin to the live edge (send / history load / pill click).
+export function scrollToBottom(smooth) {
+  follow = true;
+  hideToLatest();
+  if (smooth && typeof msgs.scrollTo === "function") {
+    try {
+      msgs.scrollTo({ top: msgs.scrollHeight, behavior: "smooth" });
+      return;
+    } catch (err) {
+      /* fall through to instant */
+    }
+  }
+  msgs.scrollTop = msgs.scrollHeight;
+}
+
+export function isFollowing() {
+  return follow;
+}
+
+// Floating "jump to latest" pill: appears only when the reader scrolled up
+// and new content landed below. Created lazily so tests/SSR without `.chat`
+// never break.
+let toLatestBtn = null;
+function ensureToLatest() {
+  if (toLatestBtn) return toLatestBtn;
+  const chat = document.querySelector(".chat");
+  if (!chat) return null;
+  toLatestBtn = el("button", "to-latest");
+  toLatestBtn.type = "button";
+  toLatestBtn.hidden = true;
+  toLatestBtn.addEventListener("click", () => scrollToBottom(true));
+  chat.appendChild(toLatestBtn);
+  return toLatestBtn;
+}
+
+function maybeShowToLatest() {
+  const btn = ensureToLatest();
+  if (!btn) return;
+  btn.textContent = "↓ " + t("web.scroll.latest");
+  btn.hidden = false;
+  msgs.classList.add("unfollowed");
+}
+
+function hideToLatest() {
+  if (toLatestBtn) toLatestBtn.hidden = true;
+  msgs.classList.remove("unfollowed");
 }
 
+msgs.addEventListener("scroll", () => {
+  const nb = nearBottom();
+  follow = nb;
+  msgs.classList.toggle("scrolled", msgs.scrollTop > 24);
+  if (nb) hideToLatest();
+  else if (liveFeed() && feedHasContent(liveFeed())) maybeShowToLatest();
+});
+
 export function appendTo(feed, node) {
   if (!feed) return node;
   feed.appendChild(node);
@@ -109,6 +183,9 @@ function clearBusy(feed) {
 export function showBusy(feed) {
   feed = feed || liveFeed();
   if (!feed) return;
+  // The delayed post-send call must not flash a spinner over a stream that
+  // already started (fast first token).
+  if (S.curMd || S.curThink) return;
   clearEmpty(feed);
   if (feed.querySelector(".busy")) return;
   if (S.busyTimer) {
@@ -142,7 +219,10 @@ function flushRun() {
   // replacing the per-action lines (mirrors `Renderer.flushRecon`). `normal`
   // and `verbose` keep the individual action lines and add nothing.
   if (verbosity() === "quiet" && S.runReconOnly && S.runFails === 0 && S.runCount > 1) {
-    S.runNodes.forEach((node) => node.remove());
+    // Fade the folded lines out instead of snapping them away.
+    const folding = S.runNodes.slice();
+    folding.forEach((node) => node.classList.add("fold-out"));
+    window.setTimeout(() => folding.forEach((node) => node.remove()), 160);
     const line = el("div", "grouped");
     line.appendChild(el("span", "pipe", "│"));
     const dur = formatMs(S.runDuration);
@@ -177,6 +257,7 @@ function mdOpts() {
 // Close the active Markdown segment (if any) and clear the text S.state.
 export function finishMd() {
   if (S.curMd) {
+    S.curMd.root.classList.remove("streaming");
     S.curMd.end();
     S.curMd = null;
   }
@@ -195,6 +276,7 @@ export function onText(chunk, sid) {
     flushRun();
     S.curText = "";
     S.curMd = mountMarkdown(feed, mdOpts());
+    S.curMd.root.classList.add("streaming");
     S.turnRoots.push(S.curMd.root);
   }
   S.curText += chunk;
@@ -204,17 +286,21 @@ export function onText(chunk, sid) {
 }
 
 export function finishThink() {
-  if (S.curThink && S.curThink.startedAt) {
-    const dur = S.curThink.head.querySelector(".dur");
-    if (dur) dur.textContent = "· " + ((Date.now() - S.curThink.startedAt) / 1000).toFixed(1) + "s";
+  if (S.curThink) {
+    if (S.curThink.wrap) S.curThink.wrap.classList.remove("live");
+    if (S.curThink.startedAt) {
+      const dur = S.curThink.head.querySelector(".dur");
+      if (dur) dur.textContent = "· " + ((Date.now() - S.curThink.startedAt) / 1000).toFixed(1) + "s";
+    }
   }
   S.curThink = null;
 }
 
 // Build the collapsible reasoning block. The global `.think-h` click handler
 // depends on this exact structure: wrap > [head, body], body carries `hidden`.
-function buildThinkBlock(withDur) {
-  const wrap = el("div", "think");
+// `live` marks an in-progress block (pulsing label); history renders static.
+function buildThinkBlock(withDur, live) {
+  const wrap = el("div", "think" + (live === false ? "" : " live"));
   const head = el("button", "think-h");
   head.type = "button";
   head.setAttribute("aria-expanded", "false");
@@ -264,6 +350,31 @@ export function onMeta(chunk, sid, kind) {
   if (feed === activeFeed()) scroll();
 }
 
+// Гейт забраковал уже отстримленный черновик (SSE `retract`): убираем его
+// из ленты до переответа, иначе финал встанет дублем следом
+// (ses_munvrcfz_na9ze3). Копипаст хода тоже чистим — отклонённого текста
+// в «copy / copy as markdown» быть не должно. Без видимого черновика
+// (пустой ответ, только тулы) — тихо, без шумовой строки.
+export function onRetract(reason, sid) {
+  const feed = feedForEvent(sid);
+  if (!feed) return;
+  clearBusy(feed);
+  const draftRoot = S.curMd ? S.curMd.root : null;
+  const hadDraft = !!draftRoot || (S.curText != null && S.curText !== "");
+  const draftLen = S.curText ? S.curText.length : 0;
+  finishMd();
+  finishThink();
+  if (draftRoot) {
+    const idx = S.turnRoots.indexOf(draftRoot);
+    if (idx !== -1) S.turnRoots.splice(idx, 1);
+    draftRoot.remove();
+  }
+  if (draftLen > 0) S.turnMd = S.turnMd.slice(0, Math.max(0, S.turnMd.length - draftLen));
+  if (!hadDraft) return;
+  if (reason) appendTo(feed, el("div", "metaln", reason));
+  if (feed === activeFeed()) scroll();
+}
+
 export function onPhase(phase, sid) {
   const feed = feedForEvent(sid);
   if (phase === "thinking") {
@@ -291,6 +402,7 @@ export function onToolStart(data) {
   }
   const status = el("span", "ok", "");
   line.appendChild(status);
+  line.classList.add("running");
   appendTo(feed, line);
   S.toolNodes.set(data.toolCallId, { status, line });
   if (categoryOf(data.name) !== "recon") S.runReconOnly = false;
@@ -419,6 +531,7 @@ export function onToolEnd(data) {
   const node = S.toolNodes.get(data.toolCallId);
   const line = node ? node.line : null;
   if (node) {
+    node.line.classList.remove("running");
     const dur = formatMs(data.durationMs);
     node.status.textContent = data.success ? " ✓" + (dur ? " " + dur : "") : " ✗" + (dur ? " " + dur : "");
     if (!data.success) {
@@ -456,6 +569,11 @@ export function onTurnEnd(data) {
   finishThink();
   finishMd();
   flushRun();
+  // Interrupted tools never get `tool_end`: drop their pulse so nothing
+  // looks alive after the turn is over.
+  S.toolNodes.forEach((n) => {
+    if (n && n.line && feed && n.line.parentNode === feed) n.line.classList.remove("running");
+  });
   const owner = S.tabs.find((x) => x.id === data.sessionId);
   if (owner) owner.planCard = null;
   if (!feed) {
@@ -562,7 +680,7 @@ function renderAssistantText(feed, text) {
 }
 
 function renderThinkBlock(feed, text) {
-  const { wrap, body } = buildThinkBlock(false);
+  const { wrap, body } = buildThinkBlock(false, false);
   body.textContent = text || "";
   feed.appendChild(wrap);
 }
@@ -659,7 +777,8 @@ export async function loadHistory(tab) {
   if (feedHasContent(tab.feed)) return;
   renderHistory(tab.feed, r.data.history);
   if (!r.data.history.length) tab.feed.appendChild(el("div", "detail", t("web.history.empty")));
-  if (tab.feed === activeFeed()) scroll();
+  if (tab.feed === activeFeed()) scrollToBottom();
+  else if (tab.feed === liveFeed()) scrollToBottom();
 }
 
 export function syncUrl() {