From 4911962ae7f9455f64229c59fb5f5c72b2e814e2 Mon Sep 17 00:00:00 2001
From: Junjun Zhang <topcheer@me.com>
Date: Fri, 25 Sep 2026 16:08:27 +0800
Subject: [PATCH] fix(agent): subgoal_track no longer matches keywords against
 tool names (#2758)

---
 internal/agent/subgoal_track.go     | 10 +++++--
 internal/agent/zz_issue2758_test.go | 41 +++++++++++++++++++++++++++++
 2 files changed, 49 insertions(+), 2 deletions(-)
 create mode 100644 internal/agent/zz_issue2758_test.go

diff --git a/internal/agent/subgoal_track.go b/internal/agent/subgoal_track.go
index 8732cfd72..905830169 100644
--- a/internal/agent/subgoal_track.go
+++ b/internal/agent/subgoal_track.go
@@ -187,13 +187,19 @@ func (s *subgoalState) recordToolCall(toolName, args string) {
 		return
 	}
 	argLower := strings.ToLower(args)
-	toolLower := strings.ToLower(toolName)
 	for i := range s.subgoals {
 		if s.subgoals[i].addressed {
 			continue
 		}
 		for _, kw := range s.subgoals[i].keywords {
-			if strings.Contains(argLower, kw) || strings.Contains(toolLower, kw) {
+			// #2758: match keywords against the tool ARGUMENTS only. The old
+			// `|| strings.Contains(toolLower, kw)` matched tool NAMES - any
+			// read_file whitewashed a "file" keyword, search_files a "search"
+			// keyword - so unrelated calls marked subgoals addressed and the
+			// un-addressed-subgoal warning went silent (systematic
+			// under-reporting). Tool names are generic verbs/nouns by design
+			// and carry no evidence about WHICH subgoal was addressed.
+			if strings.Contains(argLower, kw) {
 				s.subgoals[i].addressed = true
 				break
 			}
diff --git a/internal/agent/zz_issue2758_test.go b/internal/agent/zz_issue2758_test.go
new file mode 100644
index 000000000..b8f016ea4
--- /dev/null
+++ b/internal/agent/zz_issue2758_test.go
@@ -0,0 +1,41 @@
+package agent
+
+// Issue #2758 probe: recordToolCall matched subgoal keywords against the
+// TOOL NAME with substring Contains - "read_file" whitewashed a "file"
+// keyword, "search_files" whitewashed "search"... any unrelated call
+// marked the subgoal addressed and the un-addressed-subgoal warning went
+// silent (systematic under-reporting). The fix: keywords may only match in
+// the tool ARGUMENTS; tool-name matching is removed entirely.
+
+import "testing"
+
+func TestIssue2758ToolNameSubstringNoLongerWhitewashes(t *testing.T) {
+	s := newSubgoalState()
+	s.recordAssistantText("Plan:\n1. Modify the config file\n2. Update database schema\n3. Refactor payment service", 1)
+	// Same reproduction as the issue: an unrelated read of /tmp/README.md
+	// used to mark subgoal 1 ("file") addressed via the tool name
+	// "read_file".
+	s.recordToolCall("read_file", `{"path": "/tmp/README.md"}`)
+	for i, sg := range s.subgoals {
+		if sg.addressed {
+			t.Fatalf("subgoal %d whitewashed by tool-name substring (kw=%v)", i+1, sg.keywords)
+		}
+	}
+	// And the warning must still fire: neither subgoal was truly addressed.
+	if w := s.maybeWarn(10); w == "" {
+		t.Fatal("unaddressed-subgoal warning silenced by tool-name whitewash")
+	}
+}
+
+func TestIssue2758ArgumentSideStillMatches(t *testing.T) {
+	s := newSubgoalState()
+	s.recordAssistantText("Plan:\n1. Modify the config file\n2. Update database schema\n3. Refactor payment service", 1)
+	// A real edit whose arguments name the config file still counts.
+	s.recordToolCall("edit_file", `{"file_path": "/app/config.go", "old_text": "x", "new_text": "y"}`)
+	if !s.subgoals[0].addressed {
+		t.Fatal("argument-side match must keep marking the subgoal addressed")
+	}
+	if s.subgoals[1].addressed || s.subgoals[2].addressed {
+		t.Fatal("subgoals 2/3 must stay unaddressed (nothing touched the schema or payment)")
+	}
+}