fix(copilot): P0 guardrails, transient retry, and security hardening

autogpt_platform/backend/backend/copilot/config.py

@@ -146,6 +146,32 @@ class ChatConfig(BaseSettings):
         description="Use --resume for multi-turn conversations instead of "
         "history compression. Falls back to compression when unavailable.",
     )
+    claude_agent_fallback_model: str = Field(
+        default="claude-sonnet-4-20250514",
+        description="Fallback model when the primary model is unavailable (e.g. 529 "
+        "overloaded). The SDK automatically retries with this cheaper model.",
+    )
+    claude_agent_max_turns: int = Field(
+        default=1000,
+        ge=1,
+        le=10000,
+        description="Maximum number of agentic turns (tool-use loops) per query. "
+        "Prevents runaway tool loops from burning budget.",
+    )
+    claude_agent_max_budget_usd: float = Field(
+        default=100.0,
+        ge=0.01,
+        le=1000.0,
+        description="Maximum spend in USD per SDK query. The CLI aborts the "
+        "request if this budget is exceeded.",
+    )
+    claude_agent_max_transient_retries: int = Field(
+        default=3,
+        ge=0,
+        le=10,
+        description="Maximum number of retries for transient API errors "
+        "(429, 5xx, ECONNRESET) before surfacing the error to the user.",
+    )
     use_openrouter: bool = Field(
         default=True,
         description="Enable routing API calls through the OpenRouter proxy. "

autogpt_platform/backend/backend/copilot/constants.py

@@ -44,12 +44,31 @@ def parse_node_id_from_exec_id(node_exec_id: str) -> str:
 # Transient Anthropic API error detection
 # ---------------------------------------------------------------------------
 # Patterns in error text that indicate a transient Anthropic API error
-# (ECONNRESET / dropped TCP connection) which is retryable.
+# which is retryable.  Covers:
+#   - Connection-level: ECONNRESET, dropped TCP connections
+#   - HTTP 429: rate-limit / too-many-requests
+#   - HTTP 5xx: server errors
+#
+# Prefer specific status-code patterns over natural-language phrases
+# (e.g. "overloaded", "bad gateway") — those phrases can appear in
+# application-level SDK messages and would trigger spurious retries.
 _TRANSIENT_ERROR_PATTERNS = (
+    # Connection-level
     "socket connection was closed unexpectedly",
     "ECONNRESET",
     "connection was forcibly closed",
     "network socket disconnected",
+    # 429 rate-limit patterns
+    "rate limit",
+    "rate_limit",
+    "too many requests",
+    "status code 429",
+    # 5xx server error patterns (status-code-specific to avoid false positives)
+    "status code 529",
+    "status code 500",
+    "status code 502",
+    "status code 503",
+    "status code 504",
 )
 
 FRIENDLY_TRANSIENT_MSG = "Anthropic connection interrupted — please retry"

autogpt_platform/backend/backend/copilot/sdk/env.py

@@ -26,14 +26,14 @@ def build_sdk_env(
 
     Three modes (checked in order):
     1. **Subscription** — clears all keys; CLI uses ``claude login`` auth.
-    2. **Direct Anthropic** — returns ``{}``; subprocess inherits
-       ``ANTHROPIC_API_KEY`` from the parent environment.
+    2. **Direct Anthropic** — subprocess inherits ``ANTHROPIC_API_KEY``
+       from the parent environment (no overrides needed).
     3. **OpenRouter** (default) — overrides base URL and auth token to
        route through the proxy, with Langfuse trace headers.
 
-    When *sdk_cwd* is provided, ``CLAUDE_CODE_TMPDIR`` is set so that
-    the CLI writes temp/sub-agent output inside the per-session workspace
-    directory rather than an inaccessible system temp path.
+    All modes receive workspace isolation (``CLAUDE_CODE_TMPDIR``) and
+    security hardening env vars to prevent .claude.md loading, prompt
+    history persistence, auto-memory writes, and non-essential traffic.
     """
     # --- Mode 1: Claude Code subscription auth ---
     if config.use_claude_code_subscription:
@@ -43,40 +43,46 @@ def build_sdk_env(
             "ANTHROPIC_AUTH_TOKEN": "",
             "ANTHROPIC_BASE_URL": "",
         }
-        if sdk_cwd:
-            env["CLAUDE_CODE_TMPDIR"] = sdk_cwd
-        return env
 
     # --- Mode 2: Direct Anthropic (no proxy hop) ---
-    if not config.openrouter_active:
+    elif not config.openrouter_active:
         env = {}
-        if sdk_cwd:
-            env["CLAUDE_CODE_TMPDIR"] = sdk_cwd
-        return env
 
     # --- Mode 3: OpenRouter proxy ---
-    base = (config.base_url or "").rstrip("/")
-    if base.endswith("/v1"):
-        base = base[:-3]
-    env = {
-        "ANTHROPIC_BASE_URL": base,
-        "ANTHROPIC_AUTH_TOKEN": config.api_key or "",
-        "ANTHROPIC_API_KEY": "",  # force CLI to use AUTH_TOKEN
-    }
+    else:
+        base = (config.base_url or "").rstrip("/")
+        if base.endswith("/v1"):
+            base = base[:-3]
+        env = {
+            "ANTHROPIC_BASE_URL": base,
+            "ANTHROPIC_AUTH_TOKEN": config.api_key or "",
+            "ANTHROPIC_API_KEY": "",  # force CLI to use AUTH_TOKEN
+        }
 
-    # Inject broadcast headers so OpenRouter forwards traces to Langfuse.
-    def _safe(v: str) -> str:
-        return v.replace("\r", "").replace("\n", "").strip()[:128]
+        # Inject broadcast headers so OpenRouter forwards traces to Langfuse.
+        def _safe(v: str) -> str:
+            return v.replace("\r", "").replace("\n", "").strip()[:128]
 
-    parts = []
-    if session_id:
-        parts.append(f"x-session-id: {_safe(session_id)}")
-    if user_id:
-        parts.append(f"x-user-id: {_safe(user_id)}")
-    if parts:
-        env["ANTHROPIC_CUSTOM_HEADERS"] = "\n".join(parts)
+        parts = []
+        if session_id:
+            parts.append(f"x-session-id: {_safe(session_id)}")
+        if user_id:
+            parts.append(f"x-user-id: {_safe(user_id)}")
+        if parts:
+            env["ANTHROPIC_CUSTOM_HEADERS"] = "\n".join(parts)
 
+    # --- Common: workspace isolation + security hardening (all modes) ---
+    # Route subagent temp files into the per-session workspace so output
+    # files are accessible (fixes /tmp/claude-0/ permission errors in E2B).
     if sdk_cwd:
         env["CLAUDE_CODE_TMPDIR"] = sdk_cwd
 
+    # Harden multi-tenant deployment: prevent loading untrusted workspace
+    # .claude.md files, persisting prompt history, writing auto-memory,
+    # and sending non-essential telemetry traffic.
+    env["CLAUDE_CODE_DISABLE_CLAUDE_MDS"] = "1"
+    env["CLAUDE_CODE_SKIP_PROMPT_HISTORY"] = "1"
+    env["CLAUDE_CODE_DISABLE_AUTO_MEMORY"] = "1"
+    env["CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC"] = "1"
+
     return env

autogpt_platform/backend/backend/copilot/sdk/env_test.py

@@ -41,11 +41,9 @@ def test_returns_blanked_keys(self, mock_validate):
 
             result = build_sdk_env()
 
-        assert result == {
-            "ANTHROPIC_API_KEY": "",
-            "ANTHROPIC_AUTH_TOKEN": "",
-            "ANTHROPIC_BASE_URL": "",
-        }
+        assert result["ANTHROPIC_API_KEY"] == ""
+        assert result["ANTHROPIC_AUTH_TOKEN"] == ""
+        assert result["ANTHROPIC_BASE_URL"] == ""
         mock_validate.assert_called_once()
 
     @patch(
@@ -68,18 +66,20 @@ def test_propagates_validation_error(self, mock_validate):
 
 
 class TestBuildSdkEnvDirectAnthropic:
-    """When OpenRouter is inactive, return empty dict (inherit parent env)."""
+    """When OpenRouter is inactive, no ANTHROPIC_* overrides (inherit parent env)."""
 
-    def test_returns_empty_dict_when_openrouter_inactive(self):
+    def test_no_anthropic_key_overrides_when_openrouter_inactive(self):
         cfg = _make_config(use_openrouter=False)
         with patch("backend.copilot.sdk.env.config", cfg):
             from backend.copilot.sdk.env import build_sdk_env
 
             result = build_sdk_env()
 
-        assert result == {}
+        assert "ANTHROPIC_API_KEY" not in result
+        assert "ANTHROPIC_AUTH_TOKEN" not in result
+        assert "ANTHROPIC_BASE_URL" not in result
 
-    def test_returns_empty_dict_when_openrouter_flag_true_but_no_key(self):
+    def test_no_anthropic_key_overrides_when_openrouter_flag_true_but_no_key(self):
         """OpenRouter flag is True but no api_key => openrouter_active is False."""
         cfg = _make_config(use_openrouter=True, base_url="https://openrouter.ai/api/v1")
         # Force api_key to None after construction (field_validator may pick up env vars)
@@ -90,7 +90,9 @@ def test_returns_empty_dict_when_openrouter_flag_true_but_no_key(self):
 
             result = build_sdk_env()
 
-        assert result == {}
+        assert "ANTHROPIC_API_KEY" not in result
+        assert "ANTHROPIC_AUTH_TOKEN" not in result
+        assert "ANTHROPIC_BASE_URL" not in result
 
 
 # ---------------------------------------------------------------------------
@@ -234,12 +236,12 @@ def test_subscription_overrides_openrouter(self, mock_validate):
 
             result = build_sdk_env()
 
-        # Should get subscription result, not OpenRouter
-        assert result == {
-            "ANTHROPIC_API_KEY": "",
-            "ANTHROPIC_AUTH_TOKEN": "",
-            "ANTHROPIC_BASE_URL": "",
-        }
+        # Should get subscription result (blanked keys), not OpenRouter proxy
+        assert result["ANTHROPIC_API_KEY"] == ""
+        assert result["ANTHROPIC_AUTH_TOKEN"] == ""
+        assert result["ANTHROPIC_BASE_URL"] == ""
+        # OpenRouter-specific key must NOT be present
+        assert "ANTHROPIC_CUSTOM_HEADERS" not in result
 
 
 # ---------------------------------------------------------------------------