reasonix 0.49.0 → 0.50.1

package/dist/cli/{chunk-MQWO32ZD.js → chunk-GPUH2BNM.js}

@@ -1,36 +1,33 @@
 #!/usr/bin/env node
 import { createRequire as __cr } from 'node:module'; if (typeof globalThis.require === 'undefined') { globalThis.require = __cr(import.meta.url); }
 import {
-  MemoryStore,
-  sanitizeMemoryName
-} from "./chunk-O5LIHAMP.js";
-import {
-  countTokens,
-  countTokensBounded,
-  estimateConversationTokens,
-  estimateRequestTokens
-} from "./chunk-6OWJV3YW.js";
-import {
-  Usage
-} from "./chunk-J2IHQGPQ.js";
-import {
+  COMPACTION_SUMMARY_MARKER,
   applyEdit,
   applyMultiEdit,
   decodeFileBuffer,
   encodeFile,
   pauseGate
-} from "./chunk-ZWHSHFDP.js";
+} from "./chunk-6IUMTRFP.js";
+import {
+  MemoryStore,
+  sanitizeMemoryName
+} from "./chunk-I6FBSTTR.js";
 import {
   NEGATIVE_CLAIM_RULE,
   PROJECT_MEMORY_FILES,
   PROJECT_MEMORY_MAX_CHARS,
   TUI_FORMATTING_RULES,
   memoryEnabled
-} from "./chunk-JNTMOX7G.js";
+} from "./chunk-ENFBF6HI.js";
+import {
+  countTokens,
+  countTokensBounded,
+  estimateRequestTokens
+} from "./chunk-CAGKEGNE.js";
 import {
   formatHookOutcomeMessage,
   runHooks
-} from "./chunk-PB3MAFEI.js";
+} from "./chunk-AJIZ5KFK.js";
 import {
   ignoredByLayers,
   loadGitignoreAt,
@@ -41,17 +38,21 @@ import {
   archiveSession,
   loadSessionMessages,
   loadSessionMeta,
+  patchSessionMeta,
   rewriteSession,
   timestampSuffix
-} from "./chunk-RRXUIPWG.js";
+} from "./chunk-IK6WWRIX.js";
+import {
+  Usage
+} from "./chunk-EZ57UEZQ.js";
 import {
   DEEPSEEK_CONTEXT_TOKENS,
   DEFAULT_CONTEXT_TOKENS,
   SessionStats
-} from "./chunk-QX5TWXRZ.js";
+} from "./chunk-FQSQFCBI.js";
 import {
   t
-} from "./chunk-YEF7C4XI.js";
+} from "./chunk-PYIZZAVQ.js";
 import {
   DEFAULT_INDEX_EXCLUDES,
   ToolRateLimiter,
@@ -66,7 +67,7 @@ import {
   require_picomatch,
   webSearchEndpoint,
   webSearchEngine
-} from "./chunk-WMTMMSXU.js";
+} from "./chunk-MXWPAPZW.js";
 import {
   __commonJS,
   __esm,
@@ -6561,6 +6562,9 @@ function plainTextRejectedReason(name, result) {
   if ((name === "edit_file" || name === "write_file") && /rejected this edit/i.test(result)) {
     return "edit-gate";
   }
+  if ((name === "edit_file" || name === "write_file" || name === "multi_edit") && /queued \d+ edits? for review/i.test(result)) {
+    return "edit-gate";
+  }
   if ((name === "edit_file" || name === "multi_edit") && /read_file first/i.test(result)) {
     return "read-before-edit";
   }
@@ -6786,12 +6790,9 @@ var HISTORY_FOLD_AGGRESSIVE_THRESHOLD = 0.78;
 var HISTORY_FOLD_AGGRESSIVE_TAIL_FRACTION = 0.1;
 var HISTORY_FOLD_MIN_SAVINGS_FRACTION = 0.3;
 var FORCE_SUMMARY_THRESHOLD = 0.8;
-var PREFLIGHT_EMERGENCY_THRESHOLD = 0.95;
-var PREFLIGHT_MECHANICAL_TARGET_FRACTION = 0.7;
-var MAX_BODY_BYTES = 7e5;
-var MAX_BODY_BYTES_TARGET = 5e5;
+var TURN_START_FOLD_THRESHOLD = 0.9;
 var HISTORY_FOLD_SUMMARY_TIMEOUT_MS = 15e3;
-var HISTORY_FOLD_MARKER = "[CONVERSATION HISTORY SUMMARY \u2014 earlier turns folded for context efficiency]\n\n";
+var HISTORY_FOLD_MARKER = COMPACTION_SUMMARY_MARKER;
 var SKILL_PIN_MEMO_HEADER = "[Active skill memos \u2014 preserved verbatim across the fold:]";
 var SKILL_PIN_REGEX = /<skill-pin name="([^"]+)">\n[\s\S]*?\n<\/skill-pin>/g;
 function buildFoldSummaryInstruction(pinnedSkillNames) {
@@ -6861,28 +6862,13 @@ var ContextManager = class {
     }
     return { kind: "none", ...base };
   }
-  /** Local-side preflight before sending a request — catches oversized payloads early.
-   * Two independent signals trip mechanical truncate: token estimate above the context-window
-   * fraction, OR JSON body bytes above the gateway limit (see `MAX_BODY_BYTES`). */
-  decidePreflight(messages, toolSpecs, model) {
+  /** Turn-start estimate vs ctxMax — caller folds if the ratio crosses
+   *  TURN_START_FOLD_THRESHOLD. Replaces the old preflight/mechanical pair. */
+  estimateTurnStart(messages, toolSpecs, model) {
     const ctxMax = DEEPSEEK_CONTEXT_TOKENS[model] ?? DEFAULT_CONTEXT_TOKENS;
     const estimate = estimateRequestTokens(messages, toolSpecs ?? null, true);
-    const estimateBytes = Buffer.byteLength(JSON.stringify(messages), "utf8");
-    const tokensOver = estimate / ctxMax > PREFLIGHT_EMERGENCY_THRESHOLD;
-    const bytesOver = estimateBytes > MAX_BODY_BYTES;
-    let trigger = "none";
-    if (tokensOver && bytesOver) trigger = "both";
-    else if (tokensOver) trigger = "tokens";
-    else if (bytesOver) trigger = "bytes";
-    return {
-      needsAction: tokensOver || bytesOver,
-      estimateTokens: estimate,
-      estimateBytes,
-      ctxMax,
-      trigger
-    };
+    return { estimateTokens: estimate, ctxMax, ratio: estimate / ctxMax };
   }
-  /** Replace older turns with one summary message; keep tail within keepRecentTokens budget. */
   async fold(model, opts) {
     const ctxMax = DEEPSEEK_CONTEXT_TOKENS[model] ?? DEFAULT_CONTEXT_TOKENS;
     const tailBudget = opts?.keepRecentTokens ?? Math.floor(ctxMax * HISTORY_FOLD_TAIL_FRACTION);
@@ -6894,7 +6880,13 @@ var ContextManager = class {
       summaryChars: 0
     };
     if (all.length === 0) return noop;
-    const tokenCounts = all.map((m) => countTokensBounded(m.content ?? ""));
+    const tokenCounts = all.map((m) => {
+      let n = countTokensBounded(typeof m.content === "string" ? m.content : "");
+      if (m.role === "assistant" && Array.isArray(m.tool_calls) && m.tool_calls.length > 0) {
+        n += countTokensBounded(JSON.stringify(m.tool_calls));
+      }
+      return n;
+    });
     const totalTokens = tokenCounts.reduce((a, b) => a + b, 0);
     let cumTokens = 0;
     let boundary = all.length;
@@ -6904,6 +6896,7 @@ var ContextManager = class {
       if (all[i].role === "user") boundary = i;
     }
     if (boundary <= 0) return noop;
+    if (opts?.requireTailBoundary && boundary >= all.length) return noop;
     const head = all.slice(0, boundary);
     const tail = all.slice(boundary);
     const headTokens = totalTokens - cumTokens;
@@ -6939,58 +6932,6 @@ ${constraints}` : "";
       summaryChars: summary.content.length
     };
   }
-  /** Pure local emergency compaction for preflight: drop oldest log entries and keep a valid tail.
-   * Bounded by tokens AND bytes — bytes matter because DeepSeek's gateway 400s on bodies past
-   * `MAX_BODY_BYTES` even when the token budget is far from exhausted. */
-  mechanicalTruncate(model, opts) {
-    const ctxMax = DEEPSEEK_CONTEXT_TOKENS[model] ?? DEFAULT_CONTEXT_TOKENS;
-    const targetTokens = opts?.targetTokens ?? Math.floor(ctxMax * PREFLIGHT_MECHANICAL_TARGET_FRACTION);
-    const targetBytes = opts?.targetBytes ?? MAX_BODY_BYTES_TARGET;
-    const all = this.deps.log.toMessages();
-    const noop = {
-      folded: false,
-      beforeMessages: all.length,
-      afterMessages: all.length,
-      summaryChars: 0
-    };
-    if (all.length === 0) return noop;
-    const tokenCounts = all.map((m) => estimateConversationTokens([m], true));
-    const byteCounts = all.map((m) => Buffer.byteLength(JSON.stringify(m), "utf8"));
-    let latestUserBoundary = -1;
-    for (let i = all.length - 1; i >= 0; i--) {
-      if (all[i].role === "user") {
-        latestUserBoundary = i;
-        break;
-      }
-    }
-    let cumTokens = 0;
-    let cumBytes = 0;
-    let boundary = all.length;
-    let foundSafeBoundary = false;
-    for (let i = all.length - 1; i >= 0; i--) {
-      const nextTokens = cumTokens + tokenCounts[i];
-      const nextBytes = cumBytes + byteCounts[i];
-      if (nextTokens > targetTokens || nextBytes > targetBytes) break;
-      cumTokens = nextTokens;
-      cumBytes = nextBytes;
-      if (all[i].role === "user") {
-        boundary = i;
-        foundSafeBoundary = true;
-      }
-    }
-    if (boundary <= 0) return noop;
-    const replacement = foundSafeBoundary ? all.slice(boundary) : opts?.allowEmpty ? [] : latestUserBoundary >= 0 ? all.slice(latestUserBoundary) : all;
-    if (replacement.length === all.length) return noop;
-    this.deps.log.compactInPlace(replacement);
-    this.persistRewrite(replacement);
-    this.deps.onLogRewrite?.();
-    return {
-      folded: true,
-      beforeMessages: all.length,
-      afterMessages: replacement.length,
-      summaryChars: 0
-    };
-  }
   /** Drop a trailing in-flight assistant-with-tool_calls before a forced summary. Tail-only mutation; prefix cache safe. */
   trimTrailingToolCalls() {
     const tail = this.deps.log.entries[this.deps.log.entries.length - 1];
@@ -7177,31 +7118,6 @@ function extractDeepSeekErrorMessage(body) {
   return trimmed;
 }
 
-// src/loop/escalation.ts
-var NEEDS_PRO_MARKER_PREFIX = "<<<NEEDS_PRO";
-var NEEDS_PRO_MARKER_RE = /^<<<NEEDS_PRO(?::\s*([^>]*))?>>>/;
-var NEEDS_PRO_BUFFER_CHARS = 256;
-function parseEscalationMarker(content) {
-  const m = NEEDS_PRO_MARKER_RE.exec(content.trimStart());
-  if (!m) return { matched: false };
-  const reason = m[1]?.trim();
-  return { matched: true, reason: reason || void 0 };
-}
-function isEscalationRequest(content) {
-  return parseEscalationMarker(content).matched;
-}
-function looksLikePartialEscalationMarker(buf) {
-  const t2 = buf.trimStart();
-  if (t2.length === 0) return true;
-  if (t2.length <= NEEDS_PRO_MARKER_PREFIX.length) {
-    return NEEDS_PRO_MARKER_PREFIX.startsWith(t2);
-  }
-  if (!t2.startsWith(NEEDS_PRO_MARKER_PREFIX)) return false;
-  const rest = t2.slice(NEEDS_PRO_MARKER_PREFIX.length);
-  if (rest[0] !== ">" && rest[0] !== ":") return false;
-  return true;
-}
-
 // src/loop/force-summary.ts
 async function* forceSummaryAfterIterLimit(ctx, opts) {
   try {
@@ -7736,7 +7652,6 @@ var ReadTracker = class _ReadTracker {
 };
 
 // src/loop.ts
-var ESCALATION_MODEL = "deepseek-v4-pro";
 var MID_TURN_STEER_WRAPPER = "[Mid-turn steer queued by the user. Do not treat this as a new task; use it only as additional guidance for the current task after completing the current step.]";
 function formatSteerUserMessage(content) {
   return [MID_TURN_STEER_WRAPPER, content].join("\n");
@@ -7756,7 +7671,6 @@ var CacheFirstLoop = class {
   model;
   stream;
   reasoningEffort;
-  autoEscalate = true;
   budgetUsd;
   /** One-shot 80% warning latch — cleared by setBudget so a bump re-arms at the new boundary. */
   _budgetWarned = false;
@@ -7772,6 +7686,7 @@ var CacheFirstLoop = class {
   _streamPreference;
   /** Threaded through HTTP + every tool dispatch so Esc cancels in-flight work, not after. */
   _turnAbort = new AbortController();
+  _discardAbortRequested = false;
   /** Authoritative running-id set — UI cards consult this instead of trusting end-event delivery. Insert at dispatch entry, delete in finally. */
   _inflight = new InflightSet();
   /** Typeahead steer messages set by the UI; step() consumes one at each iter boundary. */
@@ -7792,8 +7707,6 @@ var CacheFirstLoop = class {
   get steerConsumed() {
     return this._steerConsumed;
   }
-  _proArmedForNextTurn = false;
-  _escalateThisTurn = false;
   _turnSelfCorrected = false;
   _foldedThisTurn = false;
   context;
@@ -7809,8 +7722,7 @@ var CacheFirstLoop = class {
     this.prefix = opts.prefix;
     this.tools = opts.tools ?? new ToolRegistry();
     this.model = opts.model ?? "deepseek-v4-flash";
-    this.reasoningEffort = opts.reasoningEffort ?? "max";
-    if (opts.autoEscalate !== void 0) this.autoEscalate = opts.autoEscalate;
+    this.reasoningEffort = opts.reasoningEffort ?? "high";
     this.budgetUsd = typeof opts.budgetUsd === "number" && opts.budgetUsd > 0 ? opts.budgetUsd : null;
     this.hooks = opts.hooks ?? [];
     this.hookCwd = opts.hookCwd ?? process.cwd();
@@ -7850,6 +7762,7 @@ var CacheFirstLoop = class {
           turnCount: meta.turnCount,
           cacheHitTokens: meta.cacheHitTokens,
           cacheMissTokens: meta.cacheMissTokens,
+          totalCompletionTokens: meta.totalCompletionTokens,
           lastPromptTokens: meta.lastPromptTokens
         });
       }
@@ -7967,35 +7880,15 @@ var CacheFirstLoop = class {
       this.stream = opts.stream;
     }
     if (opts.reasoningEffort !== void 0) this.reasoningEffort = opts.reasoningEffort;
-    if (opts.autoEscalate !== void 0) this.autoEscalate = opts.autoEscalate;
   }
   /** `null` disables the cap; any change re-arms the 80% warning. */
   setBudget(usd) {
     this.budgetUsd = typeof usd === "number" && usd > 0 ? usd : null;
     this._budgetWarned = false;
   }
-  /** Single-turn upgrade consumed at next step() — distinct from `/preset max` (persistent). */
-  armProForNextTurn() {
-    this._proArmedForNextTurn = true;
-  }
-  /** Cancel `/pro` arming before the next turn starts. */
-  disarmPro() {
-    this._proArmedForNextTurn = false;
-  }
-  /** UI surface — true while `/pro` is queued but hasn't fired yet. */
-  get proArmed() {
-    return this._proArmedForNextTurn;
-  }
-  /** UI surface — true while the current turn is running on pro (armed or auto-escalated). */
-  get escalatedThisTurn() {
-    return this._escalateThisTurn;
-  }
-  /** UI surface — model id of the call about to run (or running) right now, including escalation. */
+  /** UI surface — model id of the call about to run (or running) right now. */
   get currentCallModel() {
-    return this.modelForCurrentCall();
-  }
-  modelForCurrentCall() {
-    return this._escalateThisTurn ? ESCALATION_MODEL : this.model;
+    return this.model;
   }
   /** A call counts as mutating when its definition reports `readOnly !== true` and any dynamic `readOnlyCheck` doesn't override that for these args. */
   isMutating(call) {
@@ -8093,9 +7986,24 @@ ${reason}`
     }
     return healed.messages;
   }
-  abort() {
+  abort(opts = {}) {
+    if (opts.discardCurrentTurn) this._discardAbortRequested = true;
     this._turnAbort.abort();
   }
+  resetAbortState() {
+    this._turnAbort = new AbortController();
+    this._discardAbortRequested = false;
+  }
+  discardLogFrom(index) {
+    const preserved = this.log.entries.slice(0, index).map((m) => ({ ...m }));
+    this.log.compactInPlace(preserved);
+    if (this.sessionName) {
+      try {
+        rewriteSession(this.sessionName, preserved);
+      } catch {
+      }
+    }
+  }
   /** Drop the last user message + everything after; caller re-sends. Persists to session file. */
   retryLastUser() {
     const entries = this.log.entries;
@@ -8178,38 +8086,54 @@ ${reason}`
     this.scratch.reset();
     this.repair.resetStorm();
     this._turnSelfCorrected = false;
-    this._escalateThisTurn = false;
     this._foldedThisTurn = false;
-    let armedConsumed = false;
-    if (this._proArmedForNextTurn) {
-      this._escalateThisTurn = true;
-      this._proArmedForNextTurn = false;
-      armedConsumed = true;
-    }
     const carryAbort = this._turnAbort.signal.aborted;
     this._turnAbort = new AbortController();
     if (carryAbort) this._turnAbort.abort();
     const signal = this._turnAbort.signal;
-    if (armedConsumed) {
-      yield {
-        turn: this._turn,
-        role: "warning",
-        content: t("loop.proArmed")
-      };
-    }
+    const turnStartLogIndex = this.log.length;
     this.appendAndPersist({ role: "user", content: userInput });
     const toolSpecs = this.prefix.tools();
     let rateLimitWarningShown = false;
-    for (let iter = 0; ; iter++) {
-      if (signal.aborted) {
-        try {
+    {
+      const turnStart = this.context.estimateTurnStart(
+        this.buildMessages(),
+        this.prefix.toolSpecs,
+        this.model
+      );
+      if (turnStart.ratio > TURN_START_FOLD_THRESHOLD) {
+        yield {
+          turn: this._turn,
+          role: "status",
+          content: t("loop.turnStartFoldStatus")
+        };
+        const result = await this.context.fold(this.model, { requireTailBoundary: true });
+        if (result.folded) {
+          this._foldedThisTurn = true;
           yield {
             turn: this._turn,
             role: "warning",
-            content: t("loop.abortedAtIter", { iter })
+            content: t("loop.turnStartFolded", {
+              estimate: turnStart.estimateTokens.toLocaleString(),
+              ctxMax: turnStart.ctxMax.toLocaleString(),
+              pct: Math.round(turnStart.ratio * 100),
+              beforeMessages: result.beforeMessages,
+              afterMessages: result.afterMessages
+            })
           };
-          const stoppedMsg = "[aborted by user (Esc) \u2014 no summary produced. Ask again or /retry when ready; prior tool output is still in the log.]";
-          this.appendAndPersist(buildSyntheticAssistantMessage(stoppedMsg, this.model));
+        }
+      }
+    }
+    for (let iter = 0; ; iter++) {
+      if (signal.aborted) {
+        try {
+          const discardTurn = this._discardAbortRequested;
+          const stoppedMsg = discardTurn ? "[aborted by user (Esc) \u2014 interrupted turn discarded. Ask again when ready.]" : "[aborted by user (Esc) \u2014 no summary produced. Ask again or /retry when ready; prior tool output is still in the log.]";
+          if (discardTurn) {
+            this.discardLogFrom(turnStartLogIndex);
+          } else {
+            this.appendAndPersist(buildSyntheticAssistantMessage(stoppedMsg, this.model));
+          }
           yield {
             turn: this._turn,
             role: "assistant_final",
@@ -8218,7 +8142,7 @@ ${reason}`
           };
           yield { turn: this._turn, role: "done", content: stoppedMsg };
         } finally {
-          this._turnAbort = new AbortController();
+          this.resetAbortState();
         }
         this._steerQueue.length = 0;
         return;
@@ -8245,51 +8169,6 @@ ${reason}`
           content: steer
         };
       }
-      {
-        const decision2 = this.context.decidePreflight(messages, this.prefix.toolSpecs, this.model);
-        if (decision2.needsAction) {
-          const { estimateTokens: estimate, estimateBytes, ctxMax } = decision2;
-          yield {
-            turn: this._turn,
-            role: "status",
-            content: t("loop.preflightTruncateStatus")
-          };
-          const result = this.context.mechanicalTruncate(this.model, {
-            allowEmpty: false
-          });
-          if (result.folded) {
-            messages = this.buildMessages();
-            const after = this.context.decidePreflight(messages, this.prefix.toolSpecs, this.model);
-            const stillFull = after.needsAction;
-            yield {
-              turn: this._turn,
-              role: "warning",
-              content: t(
-                stillFull ? "loop.preflightTruncatedStillFull" : "loop.preflightTruncated",
-                {
-                  estimate: after.estimateTokens.toLocaleString(),
-                  ctxMax: after.ctxMax.toLocaleString(),
-                  pct: Math.round(after.estimateTokens / after.ctxMax * 100),
-                  bodyKB: Math.round(after.estimateBytes / 1024).toLocaleString(),
-                  beforeMessages: result.beforeMessages,
-                  afterMessages: result.afterMessages
-                }
-              )
-            };
-          } else {
-            yield {
-              turn: this._turn,
-              role: "warning",
-              content: t("loop.preflightNoFold", {
-                estimate: estimate.toLocaleString(),
-                ctxMax: ctxMax.toLocaleString(),
-                pct: Math.round(estimate / ctxMax * 100),
-                bodyKB: Math.round(estimateBytes / 1024).toLocaleString()
-              })
-            };
-          }
-        }
-      }
       let assistantContent = "";
       let reasoningContent = "";
       let toolCalls = [];
@@ -8298,10 +8177,7 @@ ${reason}`
         if (this.stream) {
           const callBuf = /* @__PURE__ */ new Map();
           const readyIndices = /* @__PURE__ */ new Set();
-          const callModel = this.modelForCurrentCall();
-          const bufferForEscalation = this.autoEscalate && callModel !== ESCALATION_MODEL;
-          let escalationBuf = "";
-          let escalationBufFlushed = false;
+          const callModel = this.model;
           for await (const chunk of this.client.stream({
             model: callModel,
             messages,
@@ -8321,27 +8197,11 @@ ${reason}`
             }
             if (chunk.contentDelta) {
               assistantContent += chunk.contentDelta;
-              if (bufferForEscalation && !escalationBufFlushed) {
-                escalationBuf += chunk.contentDelta;
-                if (isEscalationRequest(escalationBuf)) {
-                  break;
-                }
-                if (escalationBuf.length >= NEEDS_PRO_BUFFER_CHARS || !looksLikePartialEscalationMarker(escalationBuf)) {
-                  escalationBufFlushed = true;
-                  yield {
-                    turn: this._turn,
-                    role: "assistant_delta",
-                    content: escalationBuf
-                  };
-                  escalationBuf = "";
-                }
-              } else {
-                yield {
-                  turn: this._turn,
-                  role: "assistant_delta",
-                  content: chunk.contentDelta
-                };
-              }
+              yield {
+                turn: this._turn,
+                role: "assistant_delta",
+                content: chunk.contentDelta
+              };
             }
             if (chunk.toolCallDelta) {
               const d = chunk.toolCallDelta;
@@ -8373,17 +8233,8 @@ ${reason}`
             if (chunk.usage) usage = chunk.usage;
           }
           toolCalls = [...callBuf.values()];
-          if (bufferForEscalation && !escalationBufFlushed && escalationBuf.length > 0) {
-            if (!isEscalationRequest(escalationBuf)) {
-              yield {
-                turn: this._turn,
-                role: "assistant_delta",
-                content: escalationBuf
-              };
-            }
-          }
         } else {
-          const callModel = this.modelForCurrentCall();
+          const callModel = this.model;
           const resp = await this.client.chat({
             model: callModel,
             messages,
@@ -8399,10 +8250,11 @@ ${reason}`
         }
       } catch (err) {
         if (signal.aborted) {
+          if (this._discardAbortRequested) this.discardLogFrom(turnStartLogIndex);
           try {
             yield { turn: this._turn, role: "done", content: "" };
           } finally {
-            this._turnAbort = new AbortController();
+            this.resetAbortState();
           }
           this._steerQueue.length = 0;
           return;
@@ -8417,27 +8269,20 @@ ${reason}`
         this._steerQueue.length = 0;
         return;
       }
-      if (this.autoEscalate && this.modelForCurrentCall() !== ESCALATION_MODEL && isEscalationRequest(assistantContent)) {
-        const { reason } = parseEscalationMarker(assistantContent);
-        this._escalateThisTurn = true;
-        const reasonSuffix = reason ? ` \u2014 ${reason}` : "";
-        yield {
-          turn: this._turn,
-          role: "warning",
-          content: t("loop.flashEscalation", { model: ESCALATION_MODEL, reasonSuffix })
-        };
-        assistantContent = "";
-        reasoningContent = "";
-        toolCalls = [];
-        usage = null;
-        iter--;
-        continue;
+      const turnStats = this.stats.record(this._turn, this.model, usage ?? new Usage());
+      if (this.sessionName) {
+        try {
+          const last = this.stats.turns.length > 0 ? this.stats.turns[this.stats.turns.length - 1] : null;
+          patchSessionMeta(this.sessionName, {
+            totalCostUsd: this.stats.totalCost,
+            cacheHitTokens: this.stats.cumulativeCacheHitTokens,
+            cacheMissTokens: this.stats.cumulativeCacheMissTokens,
+            totalCompletionTokens: this.stats.cumulativeCompletionTokens,
+            lastPromptTokens: last?.usage.promptTokens
+          });
+        } catch {
+        }
       }
-      const turnStats = this.stats.record(
-        this._turn,
-        this.modelForCurrentCall(),
-        usage ?? new Usage()
-      );
       this.scratch.reasoning = reasoningContent || null;
       const { calls: repairedCalls, report } = this.repair.process(
         toolCalls,
@@ -8445,12 +8290,7 @@ ${reason}`
         assistantContent || null
       );
       this.appendAndPersist(
-        buildAssistantMessage(
-          assistantContent,
-          repairedCalls,
-          this.modelForCurrentCall(),
-          reasoningContent
-        )
+        buildAssistantMessage(assistantContent, repairedCalls, this.model, reasoningContent)
       );
       yield {
         turn: this._turn,
@@ -8463,12 +8303,7 @@ ${reason}`
       if (allSuppressed && !this._turnSelfCorrected) {
         this._turnSelfCorrected = true;
         this.replaceTailAssistantMessage(
-          buildAssistantMessage(
-            assistantContent,
-            toolCalls,
-            this.modelForCurrentCall(),
-            reasoningContent
-          )
+          buildAssistantMessage(assistantContent, toolCalls, this.model, reasoningContent)
         );
         for (const call of toolCalls) {
           this.appendAndPersist({
@@ -8481,6 +8316,7 @@ ${reason}`
         yield {
           turn: this._turn,
           role: "warning",
+          severity: "low",
           content: t("loop.repeatToolCallWarning")
         };
         continue;
@@ -8491,6 +8327,7 @@ ${reason}`
         yield {
           turn: this._turn,
           role: "warning",
+          severity: allSuppressed ? "high" : "low",
           content: `${phrase}${noteTail}`
         };
       }
@@ -12431,4 +12268,4 @@ export {
 he/he.js:
   (*! https://mths.be/he v1.2.0 by @mathias | MIT license *)
 */
-//# sourceMappingURL=chunk-MQWO32ZD.js.map
+//# sourceMappingURL=chunk-GPUH2BNM.js.map