From b5b76fe46d86ccd9c9079602df272b36dbd91826 Mon Sep 17 00:00:00 2001
From: Harshit Ojha <92013456+harshit-ojha0324@users.noreply.github.com>
Date: Tue, 13 Oct 2026 16:02:23 -0400
Subject: [PATCH] fix(skills): inject published skills when Auto-approve is off
 (#6643)

With Auto-approve OFF the agent loop asked get_relevant_skills for
min_confidence=2.0, meant as 'published only'. The gate applies the
confidence check to published skills too, and confidence is 0-1, so no
user skill was ever injected, published or not. The explicit-activation
path already treats OFF as status == published.

Add a published_only flag to get_relevant_skills and pass it from both
callers that used the sentinel (relevance injection and the skill
preflight). Auto-approve ON is unchanged: a skill still needs a passing
audit and the confidence threshold.

Fixes #6634
---
 services/memory/skills.py                | 13 ++++++++++---
 src/agent_loop.py                        | 12 ++++++++----
 tests/test_skill_index_toolset_gating.py | 17 +++++++++++++++++
 3 files changed, 35 insertions(+), 7 deletions(-)

diff --git a/services/memory/skills.py b/services/memory/skills.py
index 713ae2199..5bf583389 100644
--- a/services/memory/skills.py
+++ b/services/memory/skills.py
@@ -745,6 +745,7 @@ class SkillsManager:
         min_confidence: float = 0.0,
         available_toolsets: Optional[Iterable[str]] = None,
         platform: Optional[str] = None,
+        published_only: bool = False,
     ) -> List[Dict]:
         if skills is None:
             skills = self.load_all()
@@ -775,9 +776,15 @@ class SkillsManager:
                 if not skill.get("platforms") or platform in skill.get("platforms", [])
             ]
         # Prompt injection is fail-closed for user skills. Built-ins are
-        # shipped procedures; every other skill needs a passing audit and a
-        # confidence score at the user's current threshold.
-        if min_confidence > 0:
+        # shipped procedures. With Auto-approve OFF (published_only), the
+        # user's publish is the approval; otherwise a skill needs a passing
+        # audit and a confidence score at the user's current threshold.
+        if published_only:
+            skills = [
+                s for s in skills
+                if s.get("source") == "builtin" or s.get("status") == "published"
+            ]
+        elif min_confidence > 0:
             def _passes(s):
                 if s.get("source") == "builtin":
                     return True
diff --git a/src/agent_loop.py b/src/agent_loop.py
index 6301e86a5..c7d1f179c 100644
--- a/src/agent_loop.py
+++ b/src/agent_loop.py
@@ -10325,9 +10325,10 @@ def _has_matching_skill_for_turn(
         ))))
         if max_items == 0:
             return False
+        published_only = not prefs.get("auto_approve_skills", True)
         min_confidence = (
-            2.0
-            if not prefs.get("auto_approve_skills", True)
+            0.0
+            if published_only
             else float(prefs.get(
                 "skill_min_confidence",
                 get_setting("skill_autosave_min_confidence", 0.85),
@@ -10344,6 +10345,7 @@ def _has_matching_skill_for_turn(
             threshold=0.25,
             max_items=max_items,
             min_confidence=min_confidence,
+            published_only=published_only,
         ))
     except Exception as exc:
         logger.debug("skill preflight failed (non-fatal): %s", exc)
@@ -14952,8 +14954,9 @@ def _build_system_prompt(
                 # confidence threshold. Approve OFF → published-only (no draft
                 # passes). Approve ON → drafts at/above the chosen confidence
                 # (0 = "All"). Falls back to the global default setting.
-                if not _prefs.get("auto_approve_skills", True):
-                    _skill_min_conf = 2.0  # nothing draft clears it → published only
+                _skill_published_only = not _prefs.get("auto_approve_skills", True)
+                if _skill_published_only:
+                    _skill_min_conf = 0.0  # published_only decides eligibility
                 else:
                     try:
                         _skill_min_conf = float(_prefs.get(
@@ -14976,6 +14979,7 @@ def _build_system_prompt(
                         max_items=_skill_max_injected,
                         min_confidence=_skill_min_conf,
                         available_toolsets=relevant_tools,
+                        published_only=_skill_published_only,
                     ) if _skill_max_injected > 0 else []
                 else:
                     # Explicit client activation chooses which eligible skill
diff --git a/tests/test_skill_index_toolset_gating.py b/tests/test_skill_index_toolset_gating.py
index 92ea8f578..8a93933ae 100644
--- a/tests/test_skill_index_toolset_gating.py
+++ b/tests/test_skill_index_toolset_gating.py
@@ -133,3 +133,20 @@ def test_unaudited_or_below_threshold_published_skill_is_not_injected(tmp_path):
     assert not sm.get_relevant_skills(
         "premature skill", min_confidence=0.80,
     )
+
+
+def test_auto_approve_off_injects_published_skills_without_an_audit(tmp_path):
+    # #6634: OFF used min_confidence=2.0, which no 0-1 confidence clears, so
+    # every user skill was rejected. Publishing is the approval when OFF.
+    (tmp_path / "skills").mkdir()
+    for name, status in (("premature-skill", "published"), ("draft-skill", "draft")):
+        path = _write_skill_md(tmp_path / "skills", name)
+        text = path.read_text(encoding="utf-8").replace("audit_verdict: pass\n", "")
+        path.write_text(text.replace("status: published", f"status: {status}"), encoding="utf-8")
+    sm = SkillsManager(str(tmp_path))
+
+    query = "premature skill draft skill"
+    # threshold=0 isolates the eligibility gate from relevance scoring.
+    assert _names(sm.get_relevant_skills(query, threshold=0.0, published_only=True)) == {"premature-skill"}
+    # Auto-approve ON still needs a passing audit, as before.
+    assert not sm.get_relevant_skills(query, threshold=0.0, min_confidence=0.85)