feat(newos4): context window warns at 80% and says what trimming dropped (v4.6.91)

P0-5 asked what to do about a 422k-token session against a 400k threshold. There is no 400k
threshold in this tree — the budgets are maxContextTokens (12000) and maxContext*2 messages — so
rather than chase a number that does not exist, this applies the decision that is defensible
either way: warn before the cliff, never lose history without saying so.

- 80% of the token budget prints one warning per crossing (no per-turn nagging)
- the message-count trim unshifts a marker naming how many messages left the window and quoting
  the earliest one, instead of splicing them away silently
- 'max-context-tokens (default: 2000)' now reports the real code default, 12000

Verified: 900/1000 tokens -> warns once, second call silent; 1500/1000 -> compresses to 17
messages with the summary head intact; 9-message window with maxContext=2 -> 4 messages, marker
'[context trim: 6 earlier message(s) ... earliest was: "turn 0 ..."]'.

5 files changed, 75 insertions, 10 deletions

diff --git a/help.html b/help.html
index a464d5f..4eec9ef 100644
--- a/help.html
+++ b/help.html
@@ -1689,6 +1689,25 @@ listmods</pre>
     <h2>📋 Changelog</h2>
 
     <div class="version-item">
+      <div class="v-header"><span class="v-num">v4.6.91 — October 2026</span>
+        <span class="v-date">context says what it dropped</span></div>
+      <ul>
+        <li><strong>The window warns before it forgets.</strong> At 80% of the token budget the
+          terminal says so once — <code>⚠️ Context 90% full (900/1000 tokens)</code> — instead of
+          only announcing the fall after compression. It does not repeat on every turn.</li>
+        <li><strong>Trimming leaves a marker.</strong> Old turns used to vanish silently, so a long
+          task "remembered" a conversation that never started that way. The window now opens with
+          <code>[context trim: 6 earlier message(s) left this window — the earliest was: "turn 0
+          original brief text…"]</code>.</li>
+        <li><strong>The help text stopped lying.</strong> <code>max-context-tokens</code> claimed a
+          default of 2000; the code has always used 12000.</li>
+        <li><strong>On the "400k / 422k" report:</strong> no such limit exists in this codebase —
+          the only budgets are <code>maxContextTokens</code> (12 000) and <code>maxContext × 2</code>
+          messages. That figure came from outside the app; nothing here was changed to chase it.</li>
+      </ul>
+    </div>
+
+    <div class="version-item">
       <div class="v-header"><span class="v-num">v4.6.88–90 — October 2026</span>
         <span class="v-date">one phantom entry, three bugs</span></div>
       <ul>
diff --git a/main.html b/main.html
index 16b4d63..921c8a8 100644
--- a/main.html
+++ b/main.html
@@ -1349,13 +1349,13 @@
     </script>
 
     <!-- v4.5.63: A2A scratchpad fixes + agent-register arg fix -->
-<script src="modules/core/core-bundle.js?v=v4.6.90"></script>
+<script src="modules/core/core-bundle.js?v=v4.6.91"></script>
 
     <!-- Async dependency-ordered module loader now only initializes terminal and loads optional ext modules -->
-<script src="modules/core/module-loader.js?v=v4.6.90"></script>
+<script src="modules/core/module-loader.js?v=v4.6.91"></script>
 
     <!-- Critical ext bundle loads AFTER webx11 exists so registerModule calls work -->
-<script src="modules/ext/critical-ext-bundle.js?v=v4.6.90"></script>
+<script src="modules/ext/critical-ext-bundle.js?v=v4.6.91"></script>
 
 <script>
 // ═══ ONE TRUE VERSION SOURCE ═══
diff --git a/modules/core/core-bundle.js b/modules/core/core-bundle.js
index f72a20e..8e19189 100644
--- a/modules/core/core-bundle.js
+++ b/modules/core/core-bundle.js
@@ -9348,7 +9348,17 @@ buildSystemPrompt() {
                 // every later turn started blind and the model re-read the same files.
                 this._flushToolTrace(sess, _traceTag);
                 if (sess.context.length > this.maxContext * 2) {
-                    sess.context.splice(0, sess.context.length - this.maxContext * 2);
+                    // Silently splicing the opening turns is how a long task forgets its own brief:
+                    // the model then "remembers" a conversation that never started that way. Drop
+                    // the same messages, but leave a marker saying how many and what the first was.
+                    const cap = this.maxContext * 2;
+                    const dropped = sess.context.splice(0, sess.context.length - (cap - 1));
+                    const firstUser = dropped.find(m => m.role === 'user');
+                    const note = '[context trim: ' + dropped.length + ' earlier message(s) left this window'
+                        + (firstUser && firstUser.content ? ' — the earliest was: "' + String(firstUser.content).replace(/\s+/g, ' ').slice(0, 90) + '…"' : '')
+                        + ']';
+                    sess.context.unshift({ role: 'system', content: note });
+                    this.terminal.print('✂️ Context trimmed: ' + dropped.length + ' message(s) — the model is told what it lost', 'text-muted');
                 }
                 // Reflection: after complex tasks, run health scan and auto-skill creation
                 if (loopResult.toolTurns >= 2 && !this._cancelled) {
@@ -12197,7 +12207,7 @@ buildSystemPrompt() {
                     this.terminal.print('=== AI Config ===', 'info');
                     this.terminal.print(`max-turns: ${this.MAX_TOOL_TURNS} (default: 8)`, '');
                     this.terminal.print(`max-context: ${this.maxContext} (default: 10)`, '');
-                    this.terminal.print(`max-context-tokens: ${this.maxContextTokens} (default: 2000)`, '');
+                    this.terminal.print(`max-context-tokens: ${this.maxContextTokens} (code default: 12000)`, '');
                     this.terminal.print(`trivial-optimization: ${(this.terminal?.ollamaManager?.optimizationEnabled) ? 'on' : 'off (default)'}`, '');
                     this.terminal.print(`confirm-timeout: ${this._confirmTimeoutMs > 0 ? this._confirmTimeoutMs / 1000 + 's' : 'disabled (no auto-deny)'} (default: 120s)`, '');
                     this.terminal.print(`chat-lock-timeout: ${this._chatLockTimeoutMs / 1000}s (default: 10s)`, '');
@@ -12520,7 +12530,20 @@ buildSystemPrompt() {
             async _compressContext() {
                 if (this.context.length <= 6) return;
                 const totalTokens = this._estimateContextTokens();
-                if (totalTokens < this.maxContextTokens) return;
+                if (totalTokens < this.maxContextTokens) {
+                    // Warn before the cliff rather than only reporting the fall. Once per crossing,
+                    // so a long session is not nagged every turn.
+                    const pct = Math.round((totalTokens / this.maxContextTokens) * 100);
+                    if (pct >= 80 && !this._ctxWarned) {
+                        this._ctxWarned = true;
+                        this.terminal.print('⚠️ Context ' + pct + '% full (' + totalTokens + '/' + this.maxContextTokens +
+                            ' tokens) — the next few turns will compress earlier history into a summary', 'warning');
+                    } else if (pct < 60) {
+                        this._ctxWarned = false;
+                    }
+                    return;
+                }
+                this._ctxWarned = false;
                 const keepCount = 8;
                 const oldMsgs = this.context.slice(0, -keepCount * 2);
                 if (oldMsgs.length < 2) return;
diff --git a/modules/core/module-aiassist.js b/modules/core/module-aiassist.js
index 55dde6f..43cbb11 100644
--- a/modules/core/module-aiassist.js
+++ b/modules/core/module-aiassist.js
@@ -950,7 +950,17 @@ buildSystemPrompt() {
                 // every later turn started blind and the model re-read the same files.
                 this._flushToolTrace(sess, _traceTag);
                 if (sess.context.length > this.maxContext * 2) {
-                    sess.context.splice(0, sess.context.length - this.maxContext * 2);
+                    // Silently splicing the opening turns is how a long task forgets its own brief:
+                    // the model then "remembers" a conversation that never started that way. Drop
+                    // the same messages, but leave a marker saying how many and what the first was.
+                    const cap = this.maxContext * 2;
+                    const dropped = sess.context.splice(0, sess.context.length - (cap - 1));
+                    const firstUser = dropped.find(m => m.role === 'user');
+                    const note = '[context trim: ' + dropped.length + ' earlier message(s) left this window'
+                        + (firstUser && firstUser.content ? ' — the earliest was: "' + String(firstUser.content).replace(/\s+/g, ' ').slice(0, 90) + '…"' : '')
+                        + ']';
+                    sess.context.unshift({ role: 'system', content: note });
+                    this.terminal.print('✂️ Context trimmed: ' + dropped.length + ' message(s) — the model is told what it lost', 'text-muted');
                 }
                 // Reflection: after complex tasks, run health scan and auto-skill creation
                 if (loopResult.toolTurns >= 2 && !this._cancelled) {
@@ -3799,7 +3809,7 @@ buildSystemPrompt() {
                     this.terminal.print('=== AI Config ===', 'info');
                     this.terminal.print(`max-turns: ${this.MAX_TOOL_TURNS} (default: 8)`, '');
                     this.terminal.print(`max-context: ${this.maxContext} (default: 10)`, '');
-                    this.terminal.print(`max-context-tokens: ${this.maxContextTokens} (default: 2000)`, '');
+                    this.terminal.print(`max-context-tokens: ${this.maxContextTokens} (code default: 12000)`, '');
                     this.terminal.print(`trivial-optimization: ${(this.terminal?.ollamaManager?.optimizationEnabled) ? 'on' : 'off (default)'}`, '');
                     this.terminal.print(`confirm-timeout: ${this._confirmTimeoutMs > 0 ? this._confirmTimeoutMs / 1000 + 's' : 'disabled (no auto-deny)'} (default: 120s)`, '');
                     this.terminal.print(`chat-lock-timeout: ${this._chatLockTimeoutMs / 1000}s (default: 10s)`, '');
@@ -4122,7 +4132,20 @@ buildSystemPrompt() {
             async _compressContext() {
                 if (this.context.length <= 6) return;
                 const totalTokens = this._estimateContextTokens();
-                if (totalTokens < this.maxContextTokens) return;
+                if (totalTokens < this.maxContextTokens) {
+                    // Warn before the cliff rather than only reporting the fall. Once per crossing,
+                    // so a long session is not nagged every turn.
+                    const pct = Math.round((totalTokens / this.maxContextTokens) * 100);
+                    if (pct >= 80 && !this._ctxWarned) {
+                        this._ctxWarned = true;
+                        this.terminal.print('⚠️ Context ' + pct + '% full (' + totalTokens + '/' + this.maxContextTokens +
+                            ' tokens) — the next few turns will compress earlier history into a summary', 'warning');
+                    } else if (pct < 60) {
+                        this._ctxWarned = false;
+                    }
+                    return;
+                }
+                this._ctxWarned = false;
                 const keepCount = 8;
                 const oldMsgs = this.context.slice(0, -keepCount * 2);
                 if (oldMsgs.length < 2) return;
diff --git a/modules/core/module-loader.js b/modules/core/module-loader.js
index 95176d9..3ab51d7 100644
--- a/modules/core/module-loader.js
+++ b/modules/core/module-loader.js
@@ -4,7 +4,7 @@
     // Core module bundle loaded synchronously in main.html.
     // This loader only handles version cache-busting, splash, terminal init, and optional ext modules.
 
-    const LATEST_VER = 'v4.6.90';
+    const LATEST_VER = 'v4.6.91';
     const CACHE = '?v=' + LATEST_VER;
     const stored = localStorage.getItem('webx11_cache_ver');
     if (stored !== LATEST_VER) {