diff --git a/CHANGELOG.md b/CHANGELOG.md
index 55981fba..48e579b6 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -2,6 +2,28 @@
 
 ## [1.3.1] - 2026-05-29
 
+### fix: strict chat templates reject mid-conversation system messages (#62)
+
+Qwen3 / Qwen3.5 chat templates (and other strict templates) under
+llama.cpp `--jinja` raise `System message must be at the beginning.` and
+llama.cpp returns HTTP 400 — but only when `tools` are present, since
+that's when it compiles the template to build a tool-call grammar.
+SmallCode injects system-role content mid-conversation (clarifier, plan
+request, planner injection, path-validation warnings, skill activation,
+compaction summaries), producing a messages array with `system` entries
+at positions other than 0.
+
+- New `src/session/message_normalizer.js#consolidateSystemMessages()`
+  collapses all system-role messages into a single leading system
+  message (preserving order, de-duplicating identical blocks) and emits
+  only non-system turns after it.
+- Applied in both request builders (`bin/smallcode.js` and
+  `bin/model_client.js` `chatCompletion`) right before the body is sent,
+  so it catches stray system messages regardless of which path injected
+  them. Verified end-to-end against a Qwen3 model: every tool-bearing
+  request now carries exactly one system message at index 0.
+- Test coverage: `test/message_normalizer.test.js` (9 cases).
+
 ### fix: compatibility issues #57, #58, #59
 
 Three reported environment-compatibility bugs:
diff --git a/bin/model_client.js b/bin/model_client.js
index 834b567e..a45f7542 100644
--- a/bin/model_client.js
+++ b/bin/model_client.js
@@ -35,9 +35,13 @@ async function chatCompletion(ctx) {
     });
 
     const _tools = ctx.getAllTools(config);
+    // Collapse any mid-conversation system messages into a single leading one
+    // so strict chat templates (Qwen3/Qwen3.5 under llama.cpp --jinja) don't
+    // 400 with "System message must be at the beginning." See issue #62.
+    const { consolidateSystemMessages } = require('../src/session/message_normalizer');
     const body = {
       model: target.model,
-      messages: [systemMsg, ...processedMessages],
+      messages: consolidateSystemMessages([systemMsg, ...processedMessages]),
       temperature: 0.1,
       max_tokens: 4096,
     };
diff --git a/bin/smallcode.js b/bin/smallcode.js
index 4f8d13cb..77255e12 100755
--- a/bin/smallcode.js
+++ b/bin/smallcode.js
@@ -90,6 +90,7 @@ const { PluginLoader } = require('../src/plugins/loader');
 const { SkillManager } = require('../src/plugins/skills');
 const { SessionStore } = require('../src/session/persistence');
 const { resolveReferences, formatReferencesForPrompt } = require('../src/session/references');
+const { consolidateSystemMessages } = require('../src/session/message_normalizer');
 const { TokenTracker } = require('../src/session/tokens');
 const { UndoStack } = require('../src/session/undo');
 const { shouldInjectGitContext, getGitDiffContext } = require('../src/session/git_context');
@@ -2161,9 +2162,16 @@ async function chatCompletion(config, messages) {
     });
 
     const _tools = getAllTools(config, currentToolCategory);
+    // Consolidate any mid-conversation system messages into a single leading
+    // system message. Strict chat templates (Qwen3/Qwen3.5 under llama.cpp
+    // --jinja) reject a `system` role anywhere but index 0 and return HTTP 400
+    // when tools are present. SmallCode injects system content mid-stream
+    // (clarifier, plan, planner, path warnings, skills, compaction), so we
+    // normalize here, right before the request is built. See issue #62.
+    const normalizedMessages = consolidateSystemMessages([systemMsg, ...processedWithImages]);
     const body = {
       model: target.model,
-      messages: [systemMsg, ...processedWithImages],
+      messages: normalizedMessages,
       temperature: 0.1,
       max_tokens: parseInt(process.env.SMALLCODE_MAX_OUTPUT_TOKENS) || 8192,
     };
diff --git a/src/session/message_normalizer.js b/src/session/message_normalizer.js
new file mode 100644
index 00000000..cc44b259
--- /dev/null
+++ b/src/session/message_normalizer.js
@@ -0,0 +1,75 @@
+// SmallCode — Message Normalizer
+//
+// Some chat templates (notably Qwen3 / Qwen3.5 under llama.cpp with --jinja)
+// enforce that a `system` role message may only appear at index 0 of the
+// messages array. Their Jinja template raises:
+//
+//   raise_exception('System message must be at the beginning.')
+//
+// …and llama.cpp returns HTTP 400 BEFORE the request is processed — but only
+// when `tools` are present, because that's when it compiles the template to
+// build a tool-call grammar. (See issue #62.)
+//
+// SmallCode legitimately injects system-role content mid-conversation in
+// several places: clarification instructions, plan requests, planner
+// injection, path-validation warnings, skill activation, and compaction
+// summaries. Each of those pushes a `{ role: 'system', content }` object into
+// the live conversation history, so by the time we assemble the request the
+// array can look like:
+//
+//   [system(prompt), user, assistant, system(plan), user, system(warning), ...]
+//
+// This module collapses any such array into a single leading system message
+// followed by only non-system turns — satisfying strict templates while
+// preserving the injected instructions (they're merged into the lead system
+// message, not dropped).
+//
+// Design notes:
+//   - Order is preserved: stray system messages are appended to the lead
+//     system content in the order they appeared, so later instructions still
+//     come after earlier ones.
+//   - Non-string content (multimodal image arrays on user turns) is never
+//     touched — only `role: 'system'` entries are merged, and those are
+//     always plain strings in this codebase.
+//   - Idempotent: running it on an already-normalized array is a no-op.
+
+'use strict';
+
+/**
+ * Collapse all system-role messages into a single leading system message.
+ *
+ * @param {Array<{role:string, content:any}>} messages  OpenAI-style messages.
+ * @returns {Array} A new array with exactly one system message at index 0
+ *   (when any system content exists), followed by all non-system messages in
+ *   their original order. The input array is not mutated.
+ */
+function consolidateSystemMessages(messages) {
+  if (!Array.isArray(messages) || messages.length === 0) return messages;
+
+  const systemParts = [];
+  const rest = [];
+
+  for (const msg of messages) {
+    if (msg && msg.role === 'system' && typeof msg.content === 'string') {
+      const trimmed = msg.content.trim();
+      if (trimmed) systemParts.push(trimmed);
+    } else {
+      rest.push(msg);
+    }
+  }
+
+  // No system content at all → return the non-system messages unchanged.
+  if (systemParts.length === 0) return rest.length === messages.length ? messages : rest;
+
+  // De-duplicate consecutive identical blocks (the same instruction can be
+  // re-injected across turns; collapsing avoids ballooning the lead prompt).
+  const deduped = [];
+  for (const part of systemParts) {
+    if (deduped[deduped.length - 1] !== part) deduped.push(part);
+  }
+
+  const merged = { role: 'system', content: deduped.join('\n\n') };
+  return [merged, ...rest];
+}
+
+module.exports = { consolidateSystemMessages };
diff --git a/test/message_normalizer.test.js b/test/message_normalizer.test.js
new file mode 100644
index 00000000..e2b9f529
--- /dev/null
+++ b/test/message_normalizer.test.js
@@ -0,0 +1,100 @@
+// SmallCode — message normalizer tests (issue #62)
+//
+// Strict chat templates (Qwen3/Qwen3.5 under llama.cpp --jinja) raise
+// "System message must be at the beginning." and llama.cpp 400s when a
+// system-role message appears anywhere but index 0 AND tools are present.
+// consolidateSystemMessages() must guarantee exactly one leading system
+// message.
+
+'use strict';
+
+const { test } = require('node:test');
+const assert = require('node:assert');
+
+const { consolidateSystemMessages } = require('../src/session/message_normalizer');
+
+test('merges a mid-conversation system message into the leading one', () => {
+  const out = consolidateSystemMessages([
+    { role: 'system', content: 'base prompt' },
+    { role: 'user', content: 'hi' },
+    { role: 'assistant', content: 'hello' },
+    { role: 'system', content: 'PLAN: do the thing' },
+    { role: 'user', content: 'go' },
+  ]);
+  // Exactly one system message, at index 0.
+  assert.equal(out.filter(m => m.role === 'system').length, 1);
+  assert.equal(out[0].role, 'system');
+  assert.match(out[0].content, /base prompt/);
+  assert.match(out[0].content, /PLAN: do the thing/);
+  // Non-system turns preserved in order.
+  assert.deepEqual(out.slice(1).map(m => m.role), ['user', 'assistant', 'user']);
+});
+
+test('order of merged system parts is preserved', () => {
+  const out = consolidateSystemMessages([
+    { role: 'system', content: 'first' },
+    { role: 'user', content: 'a' },
+    { role: 'system', content: 'second' },
+    { role: 'system', content: 'third' },
+  ]);
+  assert.equal(out[0].content, 'first\n\nsecond\n\nthird');
+});
+
+test('no system messages → array returned unchanged in content', () => {
+  const input = [
+    { role: 'user', content: 'a' },
+    { role: 'assistant', content: 'b' },
+  ];
+  const out = consolidateSystemMessages(input);
+  assert.deepEqual(out.map(m => m.role), ['user', 'assistant']);
+});
+
+test('idempotent on an already-normalized array', () => {
+  const once = consolidateSystemMessages([
+    { role: 'system', content: 'sys' },
+    { role: 'user', content: 'u' },
+  ]);
+  const twice = consolidateSystemMessages(once);
+  assert.deepEqual(twice, once);
+});
+
+test('deduplicates identical consecutive system blocks', () => {
+  const out = consolidateSystemMessages([
+    { role: 'system', content: 'same' },
+    { role: 'user', content: 'u' },
+    { role: 'system', content: 'same' },
+  ]);
+  assert.equal(out[0].content, 'same');
+});
+
+test('drops empty/whitespace-only system messages', () => {
+  const out = consolidateSystemMessages([
+    { role: 'system', content: 'real' },
+    { role: 'system', content: '   ' },
+    { role: 'user', content: 'u' },
+  ]);
+  assert.equal(out[0].content, 'real');
+  assert.equal(out.filter(m => m.role === 'system').length, 1);
+});
+
+test('preserves multimodal user content untouched', () => {
+  const img = { role: 'user', content: [{ type: 'text', text: 'hi' }, { type: 'image_url', image_url: { url: 'data:...' } }] };
+  const out = consolidateSystemMessages([
+    { role: 'system', content: 'sys' },
+    img,
+  ]);
+  assert.equal(out[1], img); // same reference, unmodified
+});
+
+test('handles only-system input (collapses to one)', () => {
+  const out = consolidateSystemMessages([
+    { role: 'system', content: 'a' },
+    { role: 'system', content: 'b' },
+  ]);
+  assert.equal(out.length, 1);
+  assert.equal(out[0].content, 'a\n\nb');
+});
+
+test('empty array is a no-op', () => {
+  assert.deepEqual(consolidateSystemMessages([]), []);
+});