[breken-ai]

What

parse_messages_json decoded only \n and \" in a message's content. It did so by running string replaces over the raw JSON text. This PR moves the full decoder that json_string_field already had into read_json_string and uses it for content. json_string_field behaves the same as before.

Why

Every other escape reached the chat template verbatim, and the replace order also corrupted escaped backslashes. Parsed on main (fa094ce8):

input content:  "func main() {\n\tprint(\"C:\\temp\\new\")\n} caf\u00e9"
parsed content: func main() {
                \tprint("C:\\temp\
                ew")
                } caf\u00e9

• Tabs arrive as a backslash followed by t.
• Escaped backslashes stay doubled.
• The \n inside \\new becomes a real newline.
• \u00e9 stays 6 ASCII bytes.

Code and Windows paths are hit most, since every JSON encoder escapes tabs and backslashes.

A content string that ends in an escaped backslash was also read past its closing quote. "the path is C:\\" parsed as the path is C:\"},{.

This is the C++ side that #804 / #803 mention as follow-up hardening. That PR changes only the Python binding. This one fixes the reader for every binding (Swift, Kotlin, Flutter, and C callers too).

Tests

I added cactus-engine/tests/test_messages_json_escapes.cpp. It needs no model.

• On main, both tests fail: content_decodes_json_escapes and content_ending_in_backslash.
• With this PR, both pass.
• All engine test targets build on macOS arm64. test_bucket (10/10) and test_kv_compress (29/29) still pass. The model-backed suites were not run: they need a bundle.

AI disclosure

I found and fixed this with help from an AI coding assistant. I checked the code path by hand and reproduced the failure on main before writing the fix. The commit is signed off for the DCO.

🤖 Generated with Claude Code