perf(backend): enable cross-user prompt caching via SystemPromptPreset

.gitignore

@@ -193,3 +193,4 @@ test.db
 .next
 # Implementation plans (generated by AI agents)
 plans/
+.claude/worktrees/

autogpt_platform/backend/backend/copilot/config.py

@@ -172,6 +172,15 @@ class ChatConfig(BaseSettings):
         description="Maximum number of retries for transient API errors "
         "(429, 5xx, ECONNRESET) before surfacing the error to the user.",
     )
+    claude_agent_cross_user_prompt_cache: bool = Field(
+        default=True,
+        description="Enable cross-user prompt caching via SystemPromptPreset. "
+        "The Claude Code default prompt becomes a cacheable prefix shared "
+        "across all users, and our custom prompt is appended after it. "
+        "Dynamic sections (working dir, git status, auto-memory) are excluded "
+        "from the prefix. Set to False to fall back to passing the system "
+        "prompt as a raw string.",
+    )
     use_openrouter: bool = Field(
         default=True,
         description="Enable routing API calls through the OpenRouter proxy. "

autogpt_platform/backend/backend/copilot/sdk/service.py

@@ -13,7 +13,23 @@
 import uuid
 from collections.abc import AsyncGenerator, AsyncIterator
 from dataclasses import dataclass
-from typing import TYPE_CHECKING, Any, NamedTuple, cast
+from typing import TYPE_CHECKING, Any, Literal, NamedTuple, TypedDict, cast
+
+
+class _SystemPromptPreset(TypedDict):
+    """Local stand-in for the SDK's SystemPromptPreset.
+
+    The production SDK (>=0.1.58) defines this type natively.  We keep a
+    local copy so that:
+      1. The code type-checks on older SDK pins (e.g. 0.1.45 on dev).
+      2. Tests can import and verify the shape without an SDK upgrade.
+    """
+
+    type: Literal["preset"]
+    preset: Literal["claude_code"]
+    append: str
+    exclude_dynamic_sections: bool
+
 
 if TYPE_CHECKING:
     from backend.copilot.permissions import CopilotPermissions
@@ -699,6 +715,28 @@ def _is_fallback_stderr(line: str) -> bool:
     return "fallback model" in line.lower()
 
 
+def _build_system_prompt_value(
+    system_prompt: str,
+    cross_user_cache: bool,
+) -> str | _SystemPromptPreset:
+    """Build the ``system_prompt`` argument for :class:`ClaudeAgentOptions`.
+
+    When *cross_user_cache* is enabled, returns a :class:`_SystemPromptPreset`
+    dict so the Claude Code default prompt becomes a cacheable prefix shared
+    across all users; our custom *system_prompt* is appended after it.
+
+    When disabled, the raw *system_prompt* string is returned unchanged.
+    """
+    if cross_user_cache:
+        return _SystemPromptPreset(
+            type="preset",
+            preset="claude_code",
+            append=system_prompt,
+            exclude_dynamic_sections=True,
+        )
+    return system_prompt
+
+
 def _make_sdk_cwd(session_id: str) -> str:
     """Create a safe, session-specific working directory path.
 
@@ -2220,8 +2258,19 @@ def _on_stderr(line: str) -> None:
                     sid,
                 )
 
+        # When cross-user prompt caching is enabled, use SystemPromptPreset
+        # so the Claude Code default prompt is a cacheable prefix shared
+        # across all users.  Our custom prompt is appended after it and
+        # dynamic sections (working dir, git status, auto-memory) are
+        # excluded from the prefix — giving us cross-user cache hits that
+        # reduce input token cost by ~90%.
+        system_prompt_value = _build_system_prompt_value(
+            system_prompt,
+            cross_user_cache=config.claude_agent_cross_user_prompt_cache,
+        )
+
         sdk_options_kwargs: dict[str, Any] = {
-            "system_prompt": system_prompt,
+            "system_prompt": system_prompt_value,
             "mcp_servers": {"copilot": mcp_server},
             "allowed_tools": allowed,
             "disallowed_tools": disallowed,

autogpt_platform/backend/backend/copilot/sdk/service_test.py

@@ -9,6 +9,7 @@
 import pytest
 
 from .service import (
+    _build_system_prompt_value,
     _is_sdk_disconnect_error,
     _normalize_model_name,
     _prepare_file_attachments,
@@ -397,6 +398,7 @@ async def test_rejects_path_outside_prefix(self, tmp_path):
     "OPENAI_BASE_URL",
     "CHAT_USE_CLAUDE_CODE_SUBSCRIPTION",
     "CHAT_USE_CLAUDE_AGENT_SDK",
+    "CHAT_CLAUDE_AGENT_CROSS_USER_PROMPT_CACHE",
 )
 
 
@@ -656,3 +658,43 @@ async def test_unrelated_value_error_propagates(self):
         client.__aexit__ = AsyncMock(side_effect=ValueError("invalid argument"))
         with pytest.raises(ValueError, match="invalid argument"):
             await _safe_close_sdk_client(client, "[test]")
+
+
+# ---------------------------------------------------------------------------
+# SystemPromptPreset — cross-user prompt caching
+# ---------------------------------------------------------------------------
+
+
+class TestSystemPromptPreset:
+    """Tests for _build_system_prompt_value — cross-user prompt caching."""
+
+    def test_preset_dict_structure_when_enabled(self):
+        """When cross_user_cache is True, returns a _SystemPromptPreset dict."""
+        custom_prompt = "You are a helpful assistant."
+        result = _build_system_prompt_value(custom_prompt, cross_user_cache=True)
+
+        assert isinstance(result, dict)
+        assert result["type"] == "preset"
+        assert result["preset"] == "claude_code"
+        assert result["append"] == custom_prompt
+        assert result["exclude_dynamic_sections"] is True
+
+    def test_raw_string_when_disabled(self):
+        """When cross_user_cache is False, returns the raw string."""
+        custom_prompt = "You are a helpful assistant."
+        result = _build_system_prompt_value(custom_prompt, cross_user_cache=False)
+
+        assert isinstance(result, str)
+        assert result == custom_prompt
+
+    def test_default_config_is_enabled(self, monkeypatch, _clean_config_env):
+        """The default value for claude_agent_cross_user_prompt_cache is True."""
+        from backend.copilot import config as cfg_mod
+
+        cfg = cfg_mod.ChatConfig(
+            use_openrouter=False,
+            api_key=None,
+            base_url=None,
+            use_claude_code_subscription=False,
+        )
+        assert cfg.claude_agent_cross_user_prompt_cache is True

autogpt_platform/backend/backend/data/platform_cost_test.py

@@ -35,7 +35,6 @@ def test_large_value(self):
         assert usd_to_microdollars(1.0) == 1_000_000
 
 
-
 class TestMaskEmail:
     def test_typical_email(self):
         assert _mask_email("user@example.com") == "us***@example.com"