reasonix 0.5.2 → 0.5.4

package/dist/index.d.ts

@@ -903,17 +903,23 @@ declare class CacheFirstLoop {
     constructor(opts: CacheFirstLoopOptions);
     /**
      * Shrink the log by re-truncating oversized tool results to a tighter
-     * cap, and persist the result back to disk so the next launch doesn't
-     * re-inherit a fat session file. Returns a summary the TUI can
-     * display.
+     * token cap, and persist the result back to disk so the next launch
+     * doesn't re-inherit a fat session file. Returns a summary the TUI
+     * can display.
+     *
+     * The cap is in DeepSeek V3 tokens (not chars) — so CJK text gets
+     * capped at the same effective context footprint as English instead
+     * of slipping past a char cap at 2× the token cost. Default 4000
+     * tokens, matching the token-aware dispatch cap from 0.5.2.
      *
      * Only tool-role messages are touched (same rationale as
      * {@link healLoadedMessages}). User and assistant messages carry
      * authored intent we can't mechanically shrink without losing
      * meaning.
      */
-    compact(tightCapChars?: number): {
+    compact(maxTokens?: number): {
         healedCount: number;
+        tokensSaved: number;
         charsSaved: number;
     };
     private appendAndPersist;

package/dist/index.js

@@ -780,6 +780,25 @@ function encode(text) {
 function countTokens(text) {
   return encode(text).length;
 }
+function estimateConversationTokens(messages) {
+  let total = 0;
+  for (const m of messages) {
+    if (typeof m.content === "string" && m.content) {
+      total += countTokens(m.content);
+    }
+    if (m.tool_calls && Array.isArray(m.tool_calls) && m.tool_calls.length > 0) {
+      total += countTokens(JSON.stringify(m.tool_calls));
+    }
+  }
+  return total;
+}
+function estimateRequestTokens(messages, toolSpecs) {
+  let total = estimateConversationTokens(messages);
+  if (toolSpecs && toolSpecs.length > 0) {
+    total += countTokens(JSON.stringify(toolSpecs));
+  }
+  return total;
+}
 
 // src/repair/flatten.ts
 function analyzeSchema(schema) {
@@ -1737,20 +1756,26 @@ var CacheFirstLoop = class {
   }
   /**
    * Shrink the log by re-truncating oversized tool results to a tighter
-   * cap, and persist the result back to disk so the next launch doesn't
-   * re-inherit a fat session file. Returns a summary the TUI can
-   * display.
+   * token cap, and persist the result back to disk so the next launch
+   * doesn't re-inherit a fat session file. Returns a summary the TUI
+   * can display.
+   *
+   * The cap is in DeepSeek V3 tokens (not chars) — so CJK text gets
+   * capped at the same effective context footprint as English instead
+   * of slipping past a char cap at 2× the token cost. Default 4000
+   * tokens, matching the token-aware dispatch cap from 0.5.2.
    *
    * Only tool-role messages are touched (same rationale as
    * {@link healLoadedMessages}). User and assistant messages carry
    * authored intent we can't mechanically shrink without losing
    * meaning.
    */
-  compact(tightCapChars = 4e3) {
+  compact(maxTokens = 4e3) {
     const before = this.log.toMessages();
-    const { messages, healedCount, healedFrom } = shrinkOversizedToolResults(before, tightCapChars);
-    const afterBytes = messages.filter((m) => m.role === "tool").reduce((s, m) => s + (typeof m.content === "string" ? m.content.length : 0), 0);
-    const charsSaved = healedFrom - afterBytes;
+    const { messages, healedCount, tokensSaved, charsSaved } = shrinkOversizedToolResultsByTokens(
+      before,
+      maxTokens
+    );
     if (healedCount > 0) {
       this.log.compactInPlace(messages);
       if (this.sessionName) {
@@ -1760,7 +1785,7 @@ var CacheFirstLoop = class {
         }
       }
     }
-    return { healedCount, charsSaved };
+    return { healedCount, tokensSaved, charsSaved };
   }
   appendAndPersist(message) {
     this.log.append(message);
@@ -1917,7 +1942,32 @@ var CacheFirstLoop = class {
           content: `${iter}/${this.maxToolIters} tool calls used \u2014 approaching budget. Press Esc to force a summary now.`
         };
       }
-      const messages = this.buildMessages(pendingUser);
+      let messages = this.buildMessages(pendingUser);
+      {
+        const ctxMax2 = DEEPSEEK_CONTEXT_TOKENS[this.model] ?? DEFAULT_CONTEXT_TOKENS;
+        const estimate = estimateRequestTokens(messages, this.prefix.toolSpecs);
+        if (estimate / ctxMax2 > 0.95) {
+          const result = this.compact(1e3);
+          if (result.healedCount > 0) {
+            yield {
+              turn: this._turn,
+              role: "warning",
+              content: `preflight: request ~${estimate.toLocaleString()}/${ctxMax2.toLocaleString()} tokens (${Math.round(
+                estimate / ctxMax2 * 100
+              )}%) \u2014 pre-compacted ${result.healedCount} tool result(s), saved ${result.tokensSaved.toLocaleString()} tokens. Sending.`
+            };
+            messages = this.buildMessages(pendingUser);
+          } else {
+            yield {
+              turn: this._turn,
+              role: "warning",
+              content: `preflight: request ~${estimate.toLocaleString()}/${ctxMax2.toLocaleString()} tokens (${Math.round(
+                estimate / ctxMax2 * 100
+              )}%) and nothing to auto-compact \u2014 DeepSeek will likely 400. Run /forget or /clear to start fresh.`
+            };
+          }
+        }
+      }
       let assistantContent = "";
       let reasoningContent = "";
       let toolCalls = [];
@@ -2124,30 +2174,28 @@ var CacheFirstLoop = class {
         const ratio = usage.promptTokens / ctxMax;
         if (ratio > 0.6 && ratio <= 0.8) {
           const before = usage.promptTokens;
-          const soft = this.compact(16e3);
+          const soft = this.compact(4e3);
           if (soft.healedCount > 0) {
-            const approxSaved = Math.round(soft.charsSaved / 4);
-            const after = Math.max(0, before - approxSaved);
+            const after = Math.max(0, before - soft.tokensSaved);
             yield {
               turn: this._turn,
               role: "warning",
               content: `context ${before.toLocaleString()}/${ctxMax.toLocaleString()} (${Math.round(
                 ratio * 100
-              )}%) \u2014 proactively compacted ${soft.healedCount} tool result(s) to 16k, saved ~${approxSaved.toLocaleString()} tokens (now ~${after.toLocaleString()}). Staying ahead of the 80% guard.`
+              )}%) \u2014 proactively compacted ${soft.healedCount} tool result(s) to 4k tokens, saved ${soft.tokensSaved.toLocaleString()} tokens (now ~${after.toLocaleString()}). Staying ahead of the 80% guard.`
             };
           }
         }
       }
       if (usage && usage.promptTokens / ctxMax > 0.8) {
         const before = usage.promptTokens;
-        const compactResult = this.compact(4e3);
+        const compactResult = this.compact(1e3);
         if (compactResult.healedCount > 0) {
-          const approxSaved = Math.round(compactResult.charsSaved / 4);
-          const after = before - approxSaved;
+          const after = Math.max(0, before - compactResult.tokensSaved);
           yield {
             turn: this._turn,
             role: "warning",
-            content: `context ${before.toLocaleString()}/${ctxMax.toLocaleString()} \u2014 auto-compacted ${compactResult.healedCount} oversized tool result(s), saved ~${approxSaved.toLocaleString()} tokens (now ~${after.toLocaleString()}). Continuing.`
+            content: `context ${before.toLocaleString()}/${ctxMax.toLocaleString()} \u2014 auto-compacted ${compactResult.healedCount} oversized tool result(s), saved ${compactResult.tokensSaved.toLocaleString()} tokens (now ~${after.toLocaleString()}). Continuing.`
           };
         } else {
           yield {
@@ -2348,6 +2396,25 @@ function shrinkOversizedToolResults(messages, maxChars) {
   });
   return { messages: out, healedCount, healedFrom };
 }
+function shrinkOversizedToolResultsByTokens(messages, maxTokens) {
+  let healedCount = 0;
+  let tokensSaved = 0;
+  let charsSaved = 0;
+  const out = messages.map((msg) => {
+    if (msg.role !== "tool") return msg;
+    const content = typeof msg.content === "string" ? msg.content : "";
+    if (content.length <= maxTokens) return msg;
+    const beforeTokens = countTokens(content);
+    if (beforeTokens <= maxTokens) return msg;
+    const truncated = truncateForModelByTokens(content, maxTokens);
+    const afterTokens = countTokens(truncated);
+    healedCount += 1;
+    tokensSaved += Math.max(0, beforeTokens - afterTokens);
+    charsSaved += Math.max(0, content.length - truncated.length);
+    return { ...msg, content: truncated };
+  });
+  return { messages: out, healedCount, tokensSaved, charsSaved };
+}
 function healLoadedMessages(messages, maxChars) {
   const shrunk = shrinkOversizedToolResults(messages, maxChars);
   let healedCount = shrunk.healedCount;