diff --git a/agent/prompt_builder.py b/agent/prompt_builder.py
index e53105f5f6..b433592093 100644
--- a/agent/prompt_builder.py
+++ b/agent/prompt_builder.py
@@ -506,13 +506,38 @@ OPENAI_MODEL_EXECUTION_GUIDANCE = (
 )
 
 
-def execution_guidance_text() -> str:
-    """OPENAI_MODEL_EXECUTION_GUIDANCE as injected into the system prompt.
+# Each <mandatory_tool_use>/<act_dont_ask> line above named against the tool(s) it tells the model to reach
+# for. A line is dropped when none of its named tools are in the session's valid_tool_names, so a toolset
+# without terminal/execute_code/etc. isn't told to use them (#106506). The
+# guidance names no web tool (#39797), so there is no entry for one.
+_EXECUTION_GUIDANCE_LINE_TOOLS = {
+    "- Arithmetic, math, calculations → use terminal or execute_code\n": {"terminal", "execute_code"},
+    "- Hashes, encodings, checksums → use terminal (e.g. sha256sum, base64)\n": {"terminal"},
+    "- Current time, date, timezone → use terminal (e.g. date)\n": {"terminal"},
+    "- System state: OS, CPU, memory, disk, ports, processes → use terminal\n": {"terminal"},
+    "- File contents, sizes, line counts → use read_file, search_files, or terminal\n": {
+        "read_file",
+        "search_files",
+        "terminal",
+    },
+    "- Git history, branches, diffs → use terminal\n": {"terminal"},
+    "- 'What time is it?' → run `date` (don't guess)\n": {"terminal"},
+}
 
-    The guidance names no web tool (#39797: a hard "use web_search" overrode SOUL.md and dangled in Blank Slate),
-    so the text is toolset-neutral and needs no per-session filtering.
+
+def execution_guidance_text(valid_tool_names=None) -> str:
+    """OPENAI_MODEL_EXECUTION_GUIDANCE for the session's toolset (cache-safe: the toolset is fixed per session).
+
+    Lines that tell the model to reach for a tool absent from the session's toolset would dangle, so they are
+    dropped.
     """
-    return OPENAI_MODEL_EXECUTION_GUIDANCE
+    text = OPENAI_MODEL_EXECUTION_GUIDANCE
+    if valid_tool_names is not None:
+        valid_tool_names = set(valid_tool_names)
+        for line, tools in _EXECUTION_GUIDANCE_LINE_TOOLS.items():
+            if not tools & valid_tool_names:
+                text = text.replace(line, "")
+    return text
 
 
 # Gemini/Gemma-specific operational guidance, adapted from OpenCode's gemini.txt.
diff --git a/agent/system_prompt.py b/agent/system_prompt.py
index c010d59bc4..fb2942e429 100644
--- a/agent/system_prompt.py
+++ b/agent/system_prompt.py
@@ -575,7 +575,7 @@ def _guidance_parts(agent: Any) -> List[str]:
             parts.append(GOOGLE_MODEL_OPERATIONAL_GUIDANCE)
     if _model_gate(getattr(agent, "_execution_guidance", "auto"), agent.model, EXECUTION_GUIDANCE_MODELS):
         from agent.prompt_builder import execution_guidance_text
-        parts.append(execution_guidance_text())
+        parts.append(execution_guidance_text(agent.valid_tool_names))
     # delegate_task background delivery is intentionally between turns. Put this after the generic persistence
     # blocks so their "keep working" rule cannot turn the required yield into no-op/polling activity.
     if "delegate_task" in agent.valid_tool_names:
diff --git a/tests/agent/test_phantom_tool_references.py b/tests/agent/test_phantom_tool_references.py
index b194e57f15..800a4435eb 100644
--- a/tests/agent/test_phantom_tool_references.py
+++ b/tests/agent/test_phantom_tool_references.py
@@ -4,7 +4,9 @@ session can't call (Blank Slate audit, Aug 2026).
 Covers:
   * HERMES_AGENT_HELP_GUIDANCE degrades to the docs-only variant when the
     skill tools aren't loaded.
-  * execution_guidance_text() never names a web tool (guidance is toolset-neutral).
+  * execution_guidance_text() never names a web tool, and drops the
+    terminal/execute_code/read_file/search_files lines the session's toolset
+    can't back (#106506).
   * The coding operating brief drops the `todo` sentence when the todo tool
     isn't loaded.
   * ESSENTIAL_SKILLS can't be disabled via config, and the CLI writer strips
@@ -14,6 +16,26 @@ Covers:
 from pathlib import Path
 
 
+class TestExecutionGuidanceText:
+    def test_full_text_when_toolset_has_every_named_tool_or_is_unknown(self):
+        from agent.prompt_builder import OPENAI_MODEL_EXECUTION_GUIDANCE, execution_guidance_text
+        assert execution_guidance_text(None) == OPENAI_MODEL_EXECUTION_GUIDANCE
+        assert execution_guidance_text(
+            {"terminal", "execute_code", "read_file", "search_files"}
+        ) == OPENAI_MODEL_EXECUTION_GUIDANCE
+
+    def test_lean_toolset_is_not_told_to_use_absent_tools(self):
+        # #106506: a toolset with none of terminal/execute_code/read_file/search_files must not be told to
+        # reach for them (GPT models then refused plain arithmetic as "prohibited mental computation").
+        from agent.prompt_builder import execution_guidance_text
+        text = execution_guidance_text({"memory"})
+        mandatory = text.split("<mandatory_tool_use>")[1].split("</mandatory_tool_use>")[0]
+        for tool in ("terminal", "execute_code", "read_file", "search_files", "Arithmetic"):
+            assert tool not in mandatory
+        assert "an appropriate permitted retrieval/search tool" in mandatory
+        assert "run `date`" not in text.split("<act_dont_ask>")[1].split("</act_dont_ask>")[0]
+
+
 class TestCodingBriefTodoGating:
     def _brief(self, valid_tool_names):
         from agent.coding_context import CODING_PROFILE, RuntimeMode