[eric] tools: ship the pruned built-in manifest by default, 4211 fewer prompt tokens per turn

backend/apps/agents/manager/prompt/tool_catalog.py

@@ -28,16 +28,17 @@ FULL_TOOLS = [
 def resolve_builtin_tools_option() -> Union[List[str], Dict[str, str]]:
     """The SDK `tools` option (CLI `--tools`), the base set of built-in tools.
 
-    Default: the full claude_code preset, which ships every preset built-in's schema. With
-    OSW_TOOL_MANIFEST=1, an explicit FULL_TOOLS list instead, ONLY the built-ins OpenSwarm exposes,
-    which prunes the ~9 preset extras nothing here references (Cron*/Monitor/Task*/PushNotification/
-    RemoteTrigger/etc) for ~940 schema tokens/turn, cache-stable. MCP tools ride mcp_servers, so
-    SpawnAgent/browser/schedule/skill/web/user MCPs are untouched; ToolSearch stays in FULL_TOOLS so
-    deferred loading survives (live-proven: the model still ToolSearch-loads + calls MCP tools under
-    the manifest). Flag-gated pending a real-app soak before default-on."""
-    if os.environ.get("OSW_TOOL_MANIFEST") == "1":
-        return list(FULL_TOOLS)
-    return {"type": "preset", "preset": "claude_code"}
+    Default: an explicit FULL_TOOLS list, ONLY the built-ins OpenSwarm exposes. The claude_code
+    preset also ships Monitor/PushNotification/RemoteTrigger/ScheduleWakeup/ExitWorktree, which
+    nothing in this repo references, plus a bare `Skill` that build_effective_tool_lists hard-denies
+    anyway. Measured on a live "hi" turn (Anthropic count_tokens on the captured wire request, and
+    the provider's own usage reporting, agreeing within 1 token): 31,091 -> 26,880 prompt tokens,
+    4,211 saved per turn, cache-stable. MCP tools ride mcp_servers, so SpawnAgent/browser/schedule/
+    skill/web/user MCPs are untouched; ToolSearch stays in FULL_TOOLS so deferred loading survives.
+    Kill switch: OSW_TOOL_MANIFEST=0 restores the preset."""
+    if os.environ.get("OSW_TOOL_MANIFEST") == "0":
+        return {"type": "preset", "preset": "claude_code"}
+    return list(FULL_TOOLS)
 
 
 @typechecked

backend/tests/test_tool_manifest.py

@@ -1,7 +1,7 @@
-"""OSW_TOOL_MANIFEST gate: default ships the full claude_code preset; flag-on ships an explicit
-FULL_TOOLS list that prunes dead preset built-ins WITHOUT dropping anything OpenSwarm uses. The
-manifest MUST keep every built-in in effective_allowed's source set + ToolSearch (so deferred MCP
-loading survives); MCP tools ride mcp_servers, not this list."""
+"""OSW_TOOL_MANIFEST gate: default ships an explicit FULL_TOOLS list that prunes dead preset
+built-ins WITHOUT dropping anything OpenSwarm uses; OSW_TOOL_MANIFEST=0 is the kill switch back to
+the full claude_code preset. The manifest MUST keep every built-in in effective_allowed's source set
++ ToolSearch (so deferred MCP loading survives); MCP tools ride mcp_servers, not this list."""
 
 
 from backend.apps.agents.manager.prompt.tool_catalog import (
@@ -9,14 +9,11 @@ from backend.apps.agents.manager.prompt.tool_catalog import (
     resolve_builtin_tools_option,
 )
 
+PRESET = {"type": "preset", "preset": "claude_code"}
 
-def test_default_is_the_full_preset(monkeypatch):
+
+def test_default_is_the_explicit_full_tools_manifest(monkeypatch):
     monkeypatch.delenv("OSW_TOOL_MANIFEST", raising=False)
-    assert resolve_builtin_tools_option() == {"type": "preset", "preset": "claude_code"}
-
-
-def test_flag_on_is_the_explicit_full_tools_manifest(monkeypatch):
-    monkeypatch.setenv("OSW_TOOL_MANIFEST", "1")
     out = resolve_builtin_tools_option()
     assert isinstance(out, list)
     assert out == list(FULL_TOOLS)
@@ -24,22 +21,34 @@ def test_flag_on_is_the_explicit_full_tools_manifest(monkeypatch):
     assert out is not FULL_TOOLS
 
 
+def test_kill_switch_restores_the_full_preset(monkeypatch):
+    monkeypatch.setenv("OSW_TOOL_MANIFEST", "0")
+    assert resolve_builtin_tools_option() == PRESET
+
+
 def test_manifest_keeps_toolsearch_so_deferred_mcp_loading_survives(monkeypatch):
-    monkeypatch.setenv("OSW_TOOL_MANIFEST", "1")
-    out = resolve_builtin_tools_option()
-    assert "ToolSearch" in out
+    monkeypatch.delenv("OSW_TOOL_MANIFEST", raising=False)
+    assert "ToolSearch" in resolve_builtin_tools_option()
 
 
 def test_manifest_keeps_every_core_builtin_openswarm_exposes(monkeypatch):
     # These are the built-ins the agent actually uses; none may vanish from the manifest.
-    monkeypatch.setenv("OSW_TOOL_MANIFEST", "1")
+    monkeypatch.delenv("OSW_TOOL_MANIFEST", raising=False)
     out = resolve_builtin_tools_option()
     for core in ("Read", "Edit", "Write", "Bash", "Glob", "Grep", "AskUserQuestion"):
         assert core in out
 
 
-def test_only_the_exact_flag_value_flips_it(monkeypatch):
-    monkeypatch.setenv("OSW_TOOL_MANIFEST", "true")  # not "1"
-    assert resolve_builtin_tools_option() == {"type": "preset", "preset": "claude_code"}
-    monkeypatch.setenv("OSW_TOOL_MANIFEST", "0")
-    assert resolve_builtin_tools_option() == {"type": "preset", "preset": "claude_code"}
+def test_manifest_drops_only_builtins_nothing_in_the_repo_references(monkeypatch):
+    # Live-measured saving comes from exactly these; if one ever gains a caller, it must come back.
+    monkeypatch.delenv("OSW_TOOL_MANIFEST", raising=False)
+    out = resolve_builtin_tools_option()
+    for dead in ("Monitor", "PushNotification", "RemoteTrigger", "ScheduleWakeup", "ExitWorktree", "Skill"):
+        assert dead not in out
+
+
+def test_only_the_exact_kill_switch_value_flips_it(monkeypatch):
+    monkeypatch.setenv("OSW_TOOL_MANIFEST", "false")  # not "0"
+    assert resolve_builtin_tools_option() == list(FULL_TOOLS)
+    monkeypatch.setenv("OSW_TOOL_MANIFEST", "1")
+    assert resolve_builtin_tools_option() == list(FULL_TOOLS)