From 32b64aef972bdc3f5c3fd39c215856e9970e8a51 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Sun, 24 May 2026 22:01:57 +0000
Subject: [PATCH] Studio: r22 fixes - prose backtick guard, take/follow steps
 verbs

- _has_unclosed_code_fence() ignores a fence run when the trailing
  text on the same line starts with a space (typical English prose
  like "Use \`\`\` to start a markdown fence."). Real fence openers
  either end the line right after the delimiters or carry an info
  string with no leading space (\`\`\`python, \`\`\`bash-session).

- _DIRECT_NUMBERED_PLAN_FRAMING accepts "take these steps",
  "follow these steps", and "perform these actions" as first-person
  intent verbs. Plans like "I'll take these steps:\n1. Open URL\n
  2. Read" still re-prompt instead of being read as final answers.
---
 studio/backend/core/inference/llama_cpp.py    | 26 ++++++++-----
 .../tests/test_llama_cpp_reprompt_guard.py    | 37 +++++++++++++++++++
 2 files changed, 54 insertions(+), 9 deletions(-)

diff --git a/studio/backend/core/inference/llama_cpp.py b/studio/backend/core/inference/llama_cpp.py
index c2652a683b..75aeaf3859 100644
--- a/studio/backend/core/inference/llama_cpp.py
+++ b/studio/backend/core/inference/llama_cpp.py
@@ -198,7 +198,11 @@ _DIRECT_NUMBERED_PLAN_FRAMING = re.compile(
     r"[^\r\n]{0,160}"
     r"\b(?:open|read|search|look (?:this |that |it |them )?up|browse|"
     r"google|find|check|verify|compare|review|inspect|examine|"
-    r"do (?:this|these|the following|it)|proceed|start|begin|"
+    r"do (?:this|these|the following|it)|"
+    r"take (?:these|the following) steps|"
+    r"follow (?:these|the following) steps|"
+    r"perform (?:these|the following) actions|"
+    r"proceed|start|begin|"
     r"create|build|implement|set up|add|calculate|compute|analy[sz]e|"
     r"parse|load|run|execute|test)\b"
     r"|"
@@ -224,13 +228,12 @@ _FENCE_RUN_RE = re.compile(
 def _has_unclosed_code_fence(text: str) -> bool:
     """True if ``text`` contains a code fence whose closer is missing.
 
-    A complete fence answer is already caught by _HAS_ANSWER_ARTIFACT.
-    This helper exists so that an OPEN fence (model still streaming
-    code, or stream cut short) does not let an embedded numbered list
-    inside the fence body masquerade as a final answer. The scan is
-    done per line and uses ``search`` (not ``match``) so an inline
-    opener such as ``First, let me write it. \\`\\`\\`python`` is also
-    tracked.
+    Each line is scanned so inline openers like ``First. \\`\\`\\`python``
+    are tracked. Prose mentions such as ``Use \\`\\`\\` to start a
+    fence.`` are filtered out by requiring the trailing characters
+    after the fence run to look like a CommonMark info string: empty,
+    or starting with a non-space character (so prose with a leading
+    space disqualifies the run).
     """
     active_char: Optional[str] = None
     active_len = 0
@@ -239,8 +242,13 @@ def _has_unclosed_code_fence(text: str) -> bool:
         if not m:
             continue
         fence = m.group("backticks") or m.group("tildes")
-        trailing = line[m.end() :].strip()
+        raw_trailing = line[m.end():]
+        trailing = raw_trailing.strip()
         ch = fence[0]
+        # Prose mention guard: a real fence line never has a space
+        # immediately after the delimiters followed by sentence text.
+        if raw_trailing and raw_trailing[0] == " " and trailing:
+            continue
         if active_char is None:
             active_char = ch
             active_len = len(fence)
diff --git a/studio/backend/tests/test_llama_cpp_reprompt_guard.py b/studio/backend/tests/test_llama_cpp_reprompt_guard.py
index b0a64bf0bc..1b408da27f 100644
--- a/studio/backend/tests/test_llama_cpp_reprompt_guard.py
+++ b/studio/backend/tests/test_llama_cpp_reprompt_guard.py
@@ -985,6 +985,43 @@ def test_no_reprompt_on_code_fence_containing_markup_literal():
         assert not _would_reprompt(content), content
 
 
+def test_reprompts_on_take_or_follow_steps_numbered_plan():
+    """``I'll take these steps:`` / ``I will follow these steps:`` +
+    numbered list of work items is a plan stall."""
+    samples = [
+        (
+            "I'll take these steps:\n"
+            "1. Open the URL.\n"
+            "2. Read the page.\n"
+            "3. Summarize the answer."
+        ),
+        (
+            "I will follow these steps:\n"
+            "1. Open the current docs.\n"
+            "2. Read the relevant section.\n"
+            "3. Answer."
+        ),
+    ]
+    for content in samples:
+        assert not _has_answer_artifact(content), content
+        assert _would_reprompt(content), content
+
+
+def test_no_reprompt_on_prose_mention_of_triple_backticks_after_code():
+    """Closed code fence followed by prose that describes triple-
+    backtick syntax (with leading space after the ticks) must NOT be
+    treated as an unclosed fence."""
+    content = (
+        "Here is the snippet:\n"
+        "```python\n"
+        "print(1)\n"
+        "```\n"
+        "Use ``` to start a markdown code fence in your reply."
+    )
+    assert _has_answer_artifact(content)
+    assert not _would_reprompt(content)
+
+
 def test_reprompts_on_direct_first_person_read_check_open_plan():
     """Direct first-person intent + open/read/check/review/inspect verbs
     + numbered list is a tool stall. The broader verb set applies to